Fix brain-dead parts of r11216, specifically:
[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     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
175
176     //FIXME deblocking could skip the intra and nnz parts.
177     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
178         return;
179
180     //wow what a mess, why didn't they simplify the interlacing&intra stuff, i can't imagine that these complex rules are worth it
181
182     topleft_xy = top_xy - 1;
183     topright_xy= top_xy + 1;
184     left_xy[1] = left_xy[0] = mb_xy-1;
185     left_block[0]= 0;
186     left_block[1]= 1;
187     left_block[2]= 2;
188     left_block[3]= 3;
189     left_block[4]= 7;
190     left_block[5]= 10;
191     left_block[6]= 8;
192     left_block[7]= 11;
193     if(FRAME_MBAFF){
194         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
195         const int top_pair_xy      = pair_xy     - s->mb_stride;
196         const int topleft_pair_xy  = top_pair_xy - 1;
197         const int topright_pair_xy = top_pair_xy + 1;
198         const int topleft_mb_frame_flag  = !IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
199         const int top_mb_frame_flag      = !IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
200         const int topright_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
201         const int left_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
202         const int curr_mb_frame_flag = !IS_INTERLACED(mb_type);
203         const int bottom = (s->mb_y & 1);
204         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);
205         if (bottom
206                 ? !curr_mb_frame_flag // bottom macroblock
207                 : (!curr_mb_frame_flag && !top_mb_frame_flag) // top macroblock
208                 ) {
209             top_xy -= s->mb_stride;
210         }
211         if (bottom
212                 ? !curr_mb_frame_flag // bottom macroblock
213                 : (!curr_mb_frame_flag && !topleft_mb_frame_flag) // top macroblock
214                 ) {
215             topleft_xy -= s->mb_stride;
216         }
217         if (bottom
218                 ? !curr_mb_frame_flag // bottom macroblock
219                 : (!curr_mb_frame_flag && !topright_mb_frame_flag) // top macroblock
220                 ) {
221             topright_xy -= s->mb_stride;
222         }
223         if (left_mb_frame_flag != curr_mb_frame_flag) {
224             left_xy[1] = left_xy[0] = pair_xy - 1;
225             if (curr_mb_frame_flag) {
226                 if (bottom) {
227                     left_block[0]= 2;
228                     left_block[1]= 2;
229                     left_block[2]= 3;
230                     left_block[3]= 3;
231                     left_block[4]= 8;
232                     left_block[5]= 11;
233                     left_block[6]= 8;
234                     left_block[7]= 11;
235                 } else {
236                     left_block[0]= 0;
237                     left_block[1]= 0;
238                     left_block[2]= 1;
239                     left_block[3]= 1;
240                     left_block[4]= 7;
241                     left_block[5]= 10;
242                     left_block[6]= 7;
243                     left_block[7]= 10;
244                 }
245             } else {
246                 left_xy[1] += s->mb_stride;
247                 //left_block[0]= 0;
248                 left_block[1]= 2;
249                 left_block[2]= 0;
250                 left_block[3]= 2;
251                 //left_block[4]= 7;
252                 left_block[5]= 10;
253                 left_block[6]= 7;
254                 left_block[7]= 10;
255             }
256         }
257     }
258
259     h->top_mb_xy = top_xy;
260     h->left_mb_xy[0] = left_xy[0];
261     h->left_mb_xy[1] = left_xy[1];
262     if(for_deblock){
263         topleft_type = 0;
264         topright_type = 0;
265         top_type     = h->slice_table[top_xy     ] < 255 ? s->current_picture.mb_type[top_xy]     : 0;
266         left_type[0] = h->slice_table[left_xy[0] ] < 255 ? s->current_picture.mb_type[left_xy[0]] : 0;
267         left_type[1] = h->slice_table[left_xy[1] ] < 255 ? s->current_picture.mb_type[left_xy[1]] : 0;
268
269         if(FRAME_MBAFF && !IS_INTRA(mb_type)){
270             int list;
271             int v = *(uint16_t*)&h->non_zero_count[mb_xy][14];
272             for(i=0; i<16; i++)
273                 h->non_zero_count_cache[scan8[i]] = (v>>i)&1;
274             for(list=0; list<h->list_count; list++){
275                 if(USES_LIST(mb_type,list)){
276                     uint32_t *src = (uint32_t*)s->current_picture.motion_val[list][h->mb2b_xy[mb_xy]];
277                     uint32_t *dst = (uint32_t*)h->mv_cache[list][scan8[0]];
278                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
279                     for(i=0; i<4; i++, dst+=8, src+=h->b_stride){
280                         dst[0] = src[0];
281                         dst[1] = src[1];
282                         dst[2] = src[2];
283                         dst[3] = src[3];
284                     }
285                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
286                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = pack16to32(ref[0],ref[1])*0x0101;
287                     ref += h->b8_stride;
288                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
289                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = pack16to32(ref[0],ref[1])*0x0101;
290                 }else{
291                     fill_rectangle(&h-> mv_cache[list][scan8[ 0]], 4, 4, 8, 0, 4);
292                     fill_rectangle(&h->ref_cache[list][scan8[ 0]], 4, 4, 8, (uint8_t)LIST_NOT_USED, 1);
293                 }
294             }
295         }
296     }else{
297         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
298         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
299         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
300         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
301         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
302     }
303
304     if(IS_INTRA(mb_type)){
305         h->topleft_samples_available=
306         h->top_samples_available=
307         h->left_samples_available= 0xFFFF;
308         h->topright_samples_available= 0xEEEA;
309
310         if(!IS_INTRA(top_type) && (top_type==0 || h->pps.constrained_intra_pred)){
311             h->topleft_samples_available= 0xB3FF;
312             h->top_samples_available= 0x33FF;
313             h->topright_samples_available= 0x26EA;
314         }
315         for(i=0; i<2; i++){
316             if(!IS_INTRA(left_type[i]) && (left_type[i]==0 || h->pps.constrained_intra_pred)){
317                 h->topleft_samples_available&= 0xDF5F;
318                 h->left_samples_available&= 0x5F5F;
319             }
320         }
321
322         if(!IS_INTRA(topleft_type) && (topleft_type==0 || h->pps.constrained_intra_pred))
323             h->topleft_samples_available&= 0x7FFF;
324
325         if(!IS_INTRA(topright_type) && (topright_type==0 || h->pps.constrained_intra_pred))
326             h->topright_samples_available&= 0xFBFF;
327
328         if(IS_INTRA4x4(mb_type)){
329             if(IS_INTRA4x4(top_type)){
330                 h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
331                 h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
332                 h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
333                 h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
334             }else{
335                 int pred;
336                 if(!top_type || (IS_INTER(top_type) && h->pps.constrained_intra_pred))
337                     pred= -1;
338                 else{
339                     pred= 2;
340                 }
341                 h->intra4x4_pred_mode_cache[4+8*0]=
342                 h->intra4x4_pred_mode_cache[5+8*0]=
343                 h->intra4x4_pred_mode_cache[6+8*0]=
344                 h->intra4x4_pred_mode_cache[7+8*0]= pred;
345             }
346             for(i=0; i<2; i++){
347                 if(IS_INTRA4x4(left_type[i])){
348                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
349                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
350                 }else{
351                     int pred;
352                     if(!left_type[i] || (IS_INTER(left_type[i]) && h->pps.constrained_intra_pred))
353                         pred= -1;
354                     else{
355                         pred= 2;
356                     }
357                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
358                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
359                 }
360             }
361         }
362     }
363
364
365 /*
366 0 . T T. T T T T
367 1 L . .L . . . .
368 2 L . .L . . . .
369 3 . T TL . . . .
370 4 L . .L . . . .
371 5 L . .. . . . .
372 */
373 //FIXME constraint_intra_pred & partitioning & nnz (lets hope this is just a typo in the spec)
374     if(top_type){
375         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
376         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
377         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
378         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
379
380         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
381         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
382
383         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
384         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
385
386     }else{
387         h->non_zero_count_cache[4+8*0]=
388         h->non_zero_count_cache[5+8*0]=
389         h->non_zero_count_cache[6+8*0]=
390         h->non_zero_count_cache[7+8*0]=
391
392         h->non_zero_count_cache[1+8*0]=
393         h->non_zero_count_cache[2+8*0]=
394
395         h->non_zero_count_cache[1+8*3]=
396         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
397
398     }
399
400     for (i=0; i<2; i++) {
401         if(left_type[i]){
402             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
403             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
404             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
405             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
406         }else{
407             h->non_zero_count_cache[3+8*1 + 2*8*i]=
408             h->non_zero_count_cache[3+8*2 + 2*8*i]=
409             h->non_zero_count_cache[0+8*1 +   8*i]=
410             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
411         }
412     }
413
414     if( h->pps.cabac ) {
415         // top_cbp
416         if(top_type) {
417             h->top_cbp = h->cbp_table[top_xy];
418         } else if(IS_INTRA(mb_type)) {
419             h->top_cbp = 0x1C0;
420         } else {
421             h->top_cbp = 0;
422         }
423         // left_cbp
424         if (left_type[0]) {
425             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
426         } else if(IS_INTRA(mb_type)) {
427             h->left_cbp = 0x1C0;
428         } else {
429             h->left_cbp = 0;
430         }
431         if (left_type[0]) {
432             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
433         }
434         if (left_type[1]) {
435             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
436         }
437     }
438
439 #if 1
440     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
441         int list;
442         for(list=0; list<h->list_count; list++){
443             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
444                 /*if(!h->mv_cache_clean[list]){
445                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
446                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
447                     h->mv_cache_clean[list]= 1;
448                 }*/
449                 continue;
450             }
451             h->mv_cache_clean[list]= 0;
452
453             if(USES_LIST(top_type, list)){
454                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
455                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
456                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
457                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
458                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
459                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
460                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
461                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
462                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
463                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
464             }else{
465                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
466                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
467                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
468                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
469                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
470             }
471
472             for(i=0; i<2; i++){
473                 int cache_idx = scan8[0] - 1 + i*2*8;
474                 if(USES_LIST(left_type[i], list)){
475                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
476                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
477                     *(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]];
478                     *(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]];
479                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
480                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
481                 }else{
482                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
483                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
484                     h->ref_cache[list][cache_idx  ]=
485                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
486                 }
487             }
488
489             if((for_deblock || (IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred)) && !FRAME_MBAFF)
490                 continue;
491
492             if(USES_LIST(topleft_type, list)){
493                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + 3*h->b_stride;
494                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + h->b8_stride;
495                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
496                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
497             }else{
498                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
499                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
500             }
501
502             if(USES_LIST(topright_type, list)){
503                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
504                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
505                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
506                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
507             }else{
508                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
509                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
510             }
511
512             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
513                 continue;
514
515             h->ref_cache[list][scan8[5 ]+1] =
516             h->ref_cache[list][scan8[7 ]+1] =
517             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
518             h->ref_cache[list][scan8[4 ]] =
519             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
520             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
521             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
522             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
523             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
524             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
525
526             if( h->pps.cabac ) {
527                 /* XXX beurk, Load mvd */
528                 if(USES_LIST(top_type, list)){
529                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
530                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
531                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
532                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
533                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
534                 }else{
535                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
536                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
537                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
538                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
539                 }
540                 if(USES_LIST(left_type[0], list)){
541                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
542                     *(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]];
543                     *(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]];
544                 }else{
545                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
546                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
547                 }
548                 if(USES_LIST(left_type[1], list)){
549                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
550                     *(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]];
551                     *(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]];
552                 }else{
553                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
554                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
555                 }
556                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
557                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
558                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
559                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
560                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
561
562                 if(h->slice_type == B_TYPE){
563                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
564
565                     if(IS_DIRECT(top_type)){
566                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
567                     }else if(IS_8X8(top_type)){
568                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
569                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
570                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
571                     }else{
572                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
573                     }
574
575                     if(IS_DIRECT(left_type[0]))
576                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
577                     else if(IS_8X8(left_type[0]))
578                         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)];
579                     else
580                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
581
582                     if(IS_DIRECT(left_type[1]))
583                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
584                     else if(IS_8X8(left_type[1]))
585                         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)];
586                     else
587                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
588                 }
589             }
590
591             if(FRAME_MBAFF){
592 #define MAP_MVS\
593                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
594                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
595                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
596                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
597                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
598                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
599                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
600                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
601                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
602                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
603                 if(MB_FIELD){
604 #define MAP_F2F(idx, mb_type)\
605                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
606                         h->ref_cache[list][idx] <<= 1;\
607                         h->mv_cache[list][idx][1] /= 2;\
608                         h->mvd_cache[list][idx][1] /= 2;\
609                     }
610                     MAP_MVS
611 #undef MAP_F2F
612                 }else{
613 #define MAP_F2F(idx, mb_type)\
614                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
615                         h->ref_cache[list][idx] >>= 1;\
616                         h->mv_cache[list][idx][1] <<= 1;\
617                         h->mvd_cache[list][idx][1] <<= 1;\
618                     }
619                     MAP_MVS
620 #undef MAP_F2F
621                 }
622             }
623         }
624     }
625 #endif
626
627     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
628 }
629
630 static inline void write_back_intra_pred_mode(H264Context *h){
631     MpegEncContext * const s = &h->s;
632     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
633
634     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
635     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
636     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
637     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
638     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
639     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
640     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
641 }
642
643 /**
644  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
645  */
646 static inline int check_intra4x4_pred_mode(H264Context *h){
647     MpegEncContext * const s = &h->s;
648     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
649     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
650     int i;
651
652     if(!(h->top_samples_available&0x8000)){
653         for(i=0; i<4; i++){
654             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
655             if(status<0){
656                 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);
657                 return -1;
658             } else if(status){
659                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
660             }
661         }
662     }
663
664     if(!(h->left_samples_available&0x8000)){
665         for(i=0; i<4; i++){
666             int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
667             if(status<0){
668                 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);
669                 return -1;
670             } else if(status){
671                 h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
672             }
673         }
674     }
675
676     return 0;
677 } //FIXME cleanup like next
678
679 /**
680  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
681  */
682 static inline int check_intra_pred_mode(H264Context *h, int mode){
683     MpegEncContext * const s = &h->s;
684     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
685     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
686
687     if(mode > 6U) {
688         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);
689         return -1;
690     }
691
692     if(!(h->top_samples_available&0x8000)){
693         mode= top[ mode ];
694         if(mode<0){
695             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);
696             return -1;
697         }
698     }
699
700     if(!(h->left_samples_available&0x8000)){
701         mode= left[ mode ];
702         if(mode<0){
703             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);
704             return -1;
705         }
706     }
707
708     return mode;
709 }
710
711 /**
712  * gets the predicted intra4x4 prediction mode.
713  */
714 static inline int pred_intra_mode(H264Context *h, int n){
715     const int index8= scan8[n];
716     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
717     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
718     const int min= FFMIN(left, top);
719
720     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
721
722     if(min<0) return DC_PRED;
723     else      return min;
724 }
725
726 static inline void write_back_non_zero_count(H264Context *h){
727     MpegEncContext * const s = &h->s;
728     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
729
730     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
731     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
732     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
733     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
734     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
735     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
736     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
737
738     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
739     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
740     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
741
742     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
743     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
744     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
745
746     if(FRAME_MBAFF){
747         // store all luma nnzs, for deblocking
748         int v = 0, i;
749         for(i=0; i<16; i++)
750             v += (!!h->non_zero_count_cache[scan8[i]]) << i;
751         *(uint16_t*)&h->non_zero_count[mb_xy][14] = v;
752     }
753 }
754
755 /**
756  * gets the predicted number of non zero coefficients.
757  * @param n block index
758  */
759 static inline int pred_non_zero_count(H264Context *h, int n){
760     const int index8= scan8[n];
761     const int left= h->non_zero_count_cache[index8 - 1];
762     const int top = h->non_zero_count_cache[index8 - 8];
763     int i= left + top;
764
765     if(i<64) i= (i+1)>>1;
766
767     tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);
768
769     return i&31;
770 }
771
772 static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
773     const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
774     MpegEncContext *s = &h->s;
775
776     /* there is no consistent mapping of mvs to neighboring locations that will
777      * make mbaff happy, so we can't move all this logic to fill_caches */
778     if(FRAME_MBAFF){
779         const uint32_t *mb_types = s->current_picture_ptr->mb_type;
780         const int16_t *mv;
781         *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
782         *C = h->mv_cache[list][scan8[0]-2];
783
784         if(!MB_FIELD
785            && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
786             int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
787             if(IS_INTERLACED(mb_types[topright_xy])){
788 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
789                 const int x4 = X4, y4 = Y4;\
790                 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
791                 if(!USES_LIST(mb_type,list) && !IS_8X8(mb_type))\
792                     return LIST_NOT_USED;\
793                 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
794                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
795                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
796                 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
797
798                 SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
799             }
800         }
801         if(topright_ref == PART_NOT_AVAILABLE
802            && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
803            && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
804             if(!MB_FIELD
805                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
806                 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
807             }
808             if(MB_FIELD
809                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
810                && i >= scan8[0]+8){
811                 // leftshift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's ok.
812                 SET_DIAG_MV(>>1, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
813             }
814         }
815 #undef SET_DIAG_MV
816     }
817
818     if(topright_ref != PART_NOT_AVAILABLE){
819         *C= h->mv_cache[list][ i - 8 + part_width ];
820         return topright_ref;
821     }else{
822         tprintf(s->avctx, "topright MV not available\n");
823
824         *C= h->mv_cache[list][ i - 8 - 1 ];
825         return h->ref_cache[list][ i - 8 - 1 ];
826     }
827 }
828
829 /**
830  * gets the predicted MV.
831  * @param n the block index
832  * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
833  * @param mx the x component of the predicted motion vector
834  * @param my the y component of the predicted motion vector
835  */
836 static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
837     const int index8= scan8[n];
838     const int top_ref=      h->ref_cache[list][ index8 - 8 ];
839     const int left_ref=     h->ref_cache[list][ index8 - 1 ];
840     const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
841     const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
842     const int16_t * C;
843     int diagonal_ref, match_count;
844
845     assert(part_width==1 || part_width==2 || part_width==4);
846
847 /* mv_cache
848   B . . A T T T T
849   U . . L . . , .
850   U . . L . . . .
851   U . . L . . , .
852   . . . L . . . .
853 */
854
855     diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
856     match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
857     tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
858     if(match_count > 1){ //most common
859         *mx= mid_pred(A[0], B[0], C[0]);
860         *my= mid_pred(A[1], B[1], C[1]);
861     }else if(match_count==1){
862         if(left_ref==ref){
863             *mx= A[0];
864             *my= A[1];
865         }else if(top_ref==ref){
866             *mx= B[0];
867             *my= B[1];
868         }else{
869             *mx= C[0];
870             *my= C[1];
871         }
872     }else{
873         if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
874             *mx= A[0];
875             *my= A[1];
876         }else{
877             *mx= mid_pred(A[0], B[0], C[0]);
878             *my= mid_pred(A[1], B[1], C[1]);
879         }
880     }
881
882     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);
883 }
884
885 /**
886  * gets the directionally predicted 16x8 MV.
887  * @param n the block index
888  * @param mx the x component of the predicted motion vector
889  * @param my the y component of the predicted motion vector
890  */
891 static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
892     if(n==0){
893         const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
894         const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
895
896         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);
897
898         if(top_ref == ref){
899             *mx= B[0];
900             *my= B[1];
901             return;
902         }
903     }else{
904         const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
905         const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];
906
907         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);
908
909         if(left_ref == ref){
910             *mx= A[0];
911             *my= A[1];
912             return;
913         }
914     }
915
916     //RARE
917     pred_motion(h, n, 4, list, ref, mx, my);
918 }
919
920 /**
921  * gets the directionally predicted 8x16 MV.
922  * @param n the block index
923  * @param mx the x component of the predicted motion vector
924  * @param my the y component of the predicted motion vector
925  */
926 static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
927     if(n==0){
928         const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
929         const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];
930
931         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);
932
933         if(left_ref == ref){
934             *mx= A[0];
935             *my= A[1];
936             return;
937         }
938     }else{
939         const int16_t * C;
940         int diagonal_ref;
941
942         diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);
943
944         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);
945
946         if(diagonal_ref == ref){
947             *mx= C[0];
948             *my= C[1];
949             return;
950         }
951     }
952
953     //RARE
954     pred_motion(h, n, 2, list, ref, mx, my);
955 }
956
957 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
958     const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
959     const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];
960
961     tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);
962
963     if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
964        || (top_ref == 0  && *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ] == 0)
965        || (left_ref == 0 && *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ] == 0)){
966
967         *mx = *my = 0;
968         return;
969     }
970
971     pred_motion(h, 0, 4, 0, 0, mx, my);
972
973     return;
974 }
975
976 static inline void direct_dist_scale_factor(H264Context * const h){
977     const int poc = h->s.current_picture_ptr->poc;
978     const int poc1 = h->ref_list[1][0].poc;
979     int i;
980     for(i=0; i<h->ref_count[0]; i++){
981         int poc0 = h->ref_list[0][i].poc;
982         int td = av_clip(poc1 - poc0, -128, 127);
983         if(td == 0 /* FIXME || pic0 is a long-term ref */){
984             h->dist_scale_factor[i] = 256;
985         }else{
986             int tb = av_clip(poc - poc0, -128, 127);
987             int tx = (16384 + (FFABS(td) >> 1)) / td;
988             h->dist_scale_factor[i] = av_clip((tb*tx + 32) >> 6, -1024, 1023);
989         }
990     }
991     if(FRAME_MBAFF){
992         for(i=0; i<h->ref_count[0]; i++){
993             h->dist_scale_factor_field[2*i] =
994             h->dist_scale_factor_field[2*i+1] = h->dist_scale_factor[i];
995         }
996     }
997 }
998 static inline void direct_ref_list_init(H264Context * const h){
999     MpegEncContext * const s = &h->s;
1000     Picture * const ref1 = &h->ref_list[1][0];
1001     Picture * const cur = s->current_picture_ptr;
1002     int list, i, j;
1003     if(cur->pict_type == I_TYPE)
1004         cur->ref_count[0] = 0;
1005     if(cur->pict_type != B_TYPE)
1006         cur->ref_count[1] = 0;
1007     for(list=0; list<2; list++){
1008         cur->ref_count[list] = h->ref_count[list];
1009         for(j=0; j<h->ref_count[list]; j++)
1010             cur->ref_poc[list][j] = h->ref_list[list][j].poc;
1011     }
1012     if(cur->pict_type != B_TYPE || h->direct_spatial_mv_pred)
1013         return;
1014     for(list=0; list<2; list++){
1015         for(i=0; i<ref1->ref_count[list]; i++){
1016             const int poc = ref1->ref_poc[list][i];
1017             h->map_col_to_list0[list][i] = 0; /* bogus; fills in for missing frames */
1018             for(j=0; j<h->ref_count[list]; j++)
1019                 if(h->ref_list[list][j].poc == poc){
1020                     h->map_col_to_list0[list][i] = j;
1021                     break;
1022                 }
1023         }
1024     }
1025     if(FRAME_MBAFF){
1026         for(list=0; list<2; list++){
1027             for(i=0; i<ref1->ref_count[list]; i++){
1028                 j = h->map_col_to_list0[list][i];
1029                 h->map_col_to_list0_field[list][2*i] = 2*j;
1030                 h->map_col_to_list0_field[list][2*i+1] = 2*j+1;
1031             }
1032         }
1033     }
1034 }
1035
1036 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
1037     MpegEncContext * const s = &h->s;
1038     const int mb_xy =   s->mb_x +   s->mb_y*s->mb_stride;
1039     const int b8_xy = 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1040     const int b4_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1041     const int mb_type_col = h->ref_list[1][0].mb_type[mb_xy];
1042     const int16_t (*l1mv0)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[0][b4_xy];
1043     const int16_t (*l1mv1)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[1][b4_xy];
1044     const int8_t *l1ref0 = &h->ref_list[1][0].ref_index[0][b8_xy];
1045     const int8_t *l1ref1 = &h->ref_list[1][0].ref_index[1][b8_xy];
1046     const int is_b8x8 = IS_8X8(*mb_type);
1047     unsigned int sub_mb_type;
1048     int i8, i4;
1049
1050 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
1051     if(IS_8X8(mb_type_col) && !h->sps.direct_8x8_inference_flag){
1052         /* FIXME save sub mb types from previous frames (or derive from MVs)
1053          * so we know exactly what block size to use */
1054         sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
1055         *mb_type =    MB_TYPE_8x8|MB_TYPE_L0L1;
1056     }else if(!is_b8x8 && (mb_type_col & MB_TYPE_16x16_OR_INTRA)){
1057         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1058         *mb_type =    MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
1059     }else{
1060         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1061         *mb_type =    MB_TYPE_8x8|MB_TYPE_L0L1;
1062     }
1063     if(!is_b8x8)
1064         *mb_type |= MB_TYPE_DIRECT2;
1065     if(MB_FIELD)
1066         *mb_type |= MB_TYPE_INTERLACED;
1067
1068     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);
1069
1070     if(h->direct_spatial_mv_pred){
1071         int ref[2];
1072         int mv[2][2];
1073         int list;
1074
1075         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
1076
1077         /* ref = min(neighbors) */
1078         for(list=0; list<2; list++){
1079             int refa = h->ref_cache[list][scan8[0] - 1];
1080             int refb = h->ref_cache[list][scan8[0] - 8];
1081             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1082             if(refc == -2)
1083                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1084             ref[list] = refa;
1085             if(ref[list] < 0 || (refb < ref[list] && refb >= 0))
1086                 ref[list] = refb;
1087             if(ref[list] < 0 || (refc < ref[list] && refc >= 0))
1088                 ref[list] = refc;
1089             if(ref[list] < 0)
1090                 ref[list] = -1;
1091         }
1092
1093         if(ref[0] < 0 && ref[1] < 0){
1094             ref[0] = ref[1] = 0;
1095             mv[0][0] = mv[0][1] =
1096             mv[1][0] = mv[1][1] = 0;
1097         }else{
1098             for(list=0; list<2; list++){
1099                 if(ref[list] >= 0)
1100                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1101                 else
1102                     mv[list][0] = mv[list][1] = 0;
1103             }
1104         }
1105
1106         if(ref[1] < 0){
1107             *mb_type &= ~MB_TYPE_P0L1;
1108             sub_mb_type &= ~MB_TYPE_P0L1;
1109         }else if(ref[0] < 0){
1110             *mb_type &= ~MB_TYPE_P0L0;
1111             sub_mb_type &= ~MB_TYPE_P0L0;
1112         }
1113
1114         if(IS_16X16(*mb_type)){
1115             int a=0, b=0;
1116
1117             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1118             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1119             if(!IS_INTRA(mb_type_col)
1120                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1121                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1122                        && (h->x264_build>33 || !h->x264_build)))){
1123                 if(ref[0] > 0)
1124                     a= pack16to32(mv[0][0],mv[0][1]);
1125                 if(ref[1] > 0)
1126                     b= pack16to32(mv[1][0],mv[1][1]);
1127             }else{
1128                 a= pack16to32(mv[0][0],mv[0][1]);
1129                 b= pack16to32(mv[1][0],mv[1][1]);
1130             }
1131             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1132             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1133         }else{
1134             for(i8=0; i8<4; i8++){
1135                 const int x8 = i8&1;
1136                 const int y8 = i8>>1;
1137
1138                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1139                     continue;
1140                 h->sub_mb_type[i8] = sub_mb_type;
1141
1142                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1143                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1144                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1145                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1146
1147                 /* col_zero_flag */
1148                 if(!IS_INTRA(mb_type_col) && (   l1ref0[x8 + y8*h->b8_stride] == 0
1149                                               || (l1ref0[x8 + y8*h->b8_stride] < 0 && l1ref1[x8 + y8*h->b8_stride] == 0
1150                                                   && (h->x264_build>33 || !h->x264_build)))){
1151                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*h->b8_stride] == 0 ? l1mv0 : l1mv1;
1152                     if(IS_SUB_8X8(sub_mb_type)){
1153                         const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1154                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1155                             if(ref[0] == 0)
1156                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1157                             if(ref[1] == 0)
1158                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1159                         }
1160                     }else
1161                     for(i4=0; i4<4; i4++){
1162                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1163                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1164                             if(ref[0] == 0)
1165                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1166                             if(ref[1] == 0)
1167                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1168                         }
1169                     }
1170                 }
1171             }
1172         }
1173     }else{ /* direct temporal mv pred */
1174         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1175         const int *dist_scale_factor = h->dist_scale_factor;
1176
1177         if(FRAME_MBAFF){
1178             if(IS_INTERLACED(*mb_type)){
1179                 map_col_to_list0[0] = h->map_col_to_list0_field[0];
1180                 map_col_to_list0[1] = h->map_col_to_list0_field[1];
1181                 dist_scale_factor = h->dist_scale_factor_field;
1182             }
1183             if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col)){
1184                 /* FIXME assumes direct_8x8_inference == 1 */
1185                 const int pair_xy = s->mb_x + (s->mb_y&~1)*s->mb_stride;
1186                 int mb_types_col[2];
1187                 int y_shift;
1188
1189                 *mb_type = MB_TYPE_8x8|MB_TYPE_L0L1
1190                          | (is_b8x8 ? 0 : MB_TYPE_DIRECT2)
1191                          | (*mb_type & MB_TYPE_INTERLACED);
1192                 sub_mb_type = MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2|MB_TYPE_16x16;
1193
1194                 if(IS_INTERLACED(*mb_type)){
1195                     /* frame to field scaling */
1196                     mb_types_col[0] = h->ref_list[1][0].mb_type[pair_xy];
1197                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1198                     if(s->mb_y&1){
1199                         l1ref0 -= 2*h->b8_stride;
1200                         l1ref1 -= 2*h->b8_stride;
1201                         l1mv0 -= 4*h->b_stride;
1202                         l1mv1 -= 4*h->b_stride;
1203                     }
1204                     y_shift = 0;
1205
1206                     if(   (mb_types_col[0] & MB_TYPE_16x16_OR_INTRA)
1207                        && (mb_types_col[1] & MB_TYPE_16x16_OR_INTRA)
1208                        && !is_b8x8)
1209                         *mb_type |= MB_TYPE_16x8;
1210                     else
1211                         *mb_type |= MB_TYPE_8x8;
1212                 }else{
1213                     /* field to frame scaling */
1214                     /* col_mb_y = (mb_y&~1) + (topAbsDiffPOC < bottomAbsDiffPOC ? 0 : 1)
1215                      * but in MBAFF, top and bottom POC are equal */
1216                     int dy = (s->mb_y&1) ? 1 : 2;
1217                     mb_types_col[0] =
1218                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1219                     l1ref0 += dy*h->b8_stride;
1220                     l1ref1 += dy*h->b8_stride;
1221                     l1mv0 += 2*dy*h->b_stride;
1222                     l1mv1 += 2*dy*h->b_stride;
1223                     y_shift = 2;
1224
1225                     if((mb_types_col[0] & (MB_TYPE_16x16_OR_INTRA|MB_TYPE_16x8))
1226                        && !is_b8x8)
1227                         *mb_type |= MB_TYPE_16x16;
1228                     else
1229                         *mb_type |= MB_TYPE_8x8;
1230                 }
1231
1232                 for(i8=0; i8<4; i8++){
1233                     const int x8 = i8&1;
1234                     const int y8 = i8>>1;
1235                     int ref0, scale;
1236                     const int16_t (*l1mv)[2]= l1mv0;
1237
1238                     if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1239                         continue;
1240                     h->sub_mb_type[i8] = sub_mb_type;
1241
1242                     fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1243                     if(IS_INTRA(mb_types_col[y8])){
1244                         fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1245                         fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1246                         fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1247                         continue;
1248                     }
1249
1250                     ref0 = l1ref0[x8 + (y8*2>>y_shift)*h->b8_stride];
1251                     if(ref0 >= 0)
1252                         ref0 = map_col_to_list0[0][ref0*2>>y_shift];
1253                     else{
1254                         ref0 = map_col_to_list0[1][l1ref1[x8 + (y8*2>>y_shift)*h->b8_stride]*2>>y_shift];
1255                         l1mv= l1mv1;
1256                     }
1257                     scale = dist_scale_factor[ref0];
1258                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1259
1260                     {
1261                         const int16_t *mv_col = l1mv[x8*3 + (y8*6>>y_shift)*h->b_stride];
1262                         int my_col = (mv_col[1]<<y_shift)/2;
1263                         int mx = (scale * mv_col[0] + 128) >> 8;
1264                         int my = (scale * my_col + 128) >> 8;
1265                         fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1266                         fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1267                     }
1268                 }
1269                 return;
1270             }
1271         }
1272
1273         /* one-to-one mv scaling */
1274
1275         if(IS_16X16(*mb_type)){
1276             int ref, mv0, mv1;
1277
1278             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1279             if(IS_INTRA(mb_type_col)){
1280                 ref=mv0=mv1=0;
1281             }else{
1282                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0]]
1283                                                 : map_col_to_list0[1][l1ref1[0]];
1284                 const int scale = dist_scale_factor[ref0];
1285                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1286                 int mv_l0[2];
1287                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1288                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1289                 ref= ref0;
1290                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1291                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1292             }
1293             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1294             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1295             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1296         }else{
1297             for(i8=0; i8<4; i8++){
1298                 const int x8 = i8&1;
1299                 const int y8 = i8>>1;
1300                 int ref0, scale;
1301                 const int16_t (*l1mv)[2]= l1mv0;
1302
1303                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1304                     continue;
1305                 h->sub_mb_type[i8] = sub_mb_type;
1306                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1307                 if(IS_INTRA(mb_type_col)){
1308                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1309                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1310                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1311                     continue;
1312                 }
1313
1314                 ref0 = l1ref0[x8 + y8*h->b8_stride];
1315                 if(ref0 >= 0)
1316                     ref0 = map_col_to_list0[0][ref0];
1317                 else{
1318                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*h->b8_stride]];
1319                     l1mv= l1mv1;
1320                 }
1321                 scale = dist_scale_factor[ref0];
1322
1323                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1324                 if(IS_SUB_8X8(sub_mb_type)){
1325                     const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1326                     int mx = (scale * mv_col[0] + 128) >> 8;
1327                     int my = (scale * mv_col[1] + 128) >> 8;
1328                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1329                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1330                 }else
1331                 for(i4=0; i4<4; i4++){
1332                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1333                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1334                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1335                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1336                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1337                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1338                 }
1339             }
1340         }
1341     }
1342 }
1343
1344 static inline void write_back_motion(H264Context *h, int mb_type){
1345     MpegEncContext * const s = &h->s;
1346     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1347     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1348     int list;
1349
1350     if(!USES_LIST(mb_type, 0))
1351         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1352
1353     for(list=0; list<h->list_count; list++){
1354         int y;
1355         if(!USES_LIST(mb_type, list))
1356             continue;
1357
1358         for(y=0; y<4; y++){
1359             *(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];
1360             *(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];
1361         }
1362         if( h->pps.cabac ) {
1363             if(IS_SKIP(mb_type))
1364                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1365             else
1366             for(y=0; y<4; y++){
1367                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1368                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1369             }
1370         }
1371
1372         {
1373             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1374             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1375             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1376             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1377             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1378         }
1379     }
1380
1381     if(h->slice_type == B_TYPE && h->pps.cabac){
1382         if(IS_8X8(mb_type)){
1383             uint8_t *direct_table = &h->direct_table[b8_xy];
1384             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1385             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1386             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1387         }
1388     }
1389 }
1390
1391 /**
1392  * Decodes a network abstraction layer unit.
1393  * @param consumed is the number of bytes used as input
1394  * @param length is the length of the array
1395  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1396  * @returns decoded bytes, might be src+1 if no escapes
1397  */
1398 static uint8_t *decode_nal(H264Context *h, uint8_t *src, int *dst_length, int *consumed, int length){
1399     int i, si, di;
1400     uint8_t *dst;
1401     int bufidx;
1402
1403 //    src[0]&0x80;                //forbidden bit
1404     h->nal_ref_idc= src[0]>>5;
1405     h->nal_unit_type= src[0]&0x1F;
1406
1407     src++; length--;
1408 #if 0
1409     for(i=0; i<length; i++)
1410         printf("%2X ", src[i]);
1411 #endif
1412     for(i=0; i+1<length; i+=2){
1413         if(src[i]) continue;
1414         if(i>0 && src[i-1]==0) i--;
1415         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1416             if(src[i+2]!=3){
1417                 /* startcode, so we must be past the end */
1418                 length=i;
1419             }
1420             break;
1421         }
1422     }
1423
1424     if(i>=length-1){ //no escaped 0
1425         *dst_length= length;
1426         *consumed= length+1; //+1 for the header
1427         return src;
1428     }
1429
1430     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1431     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length);
1432     dst= h->rbsp_buffer[bufidx];
1433
1434     if (dst == NULL){
1435         return NULL;
1436     }
1437
1438 //printf("decoding esc\n");
1439     si=di=0;
1440     while(si<length){
1441         //remove escapes (very rare 1:2^22)
1442         if(si+2<length && src[si]==0 && src[si+1]==0 && src[si+2]<=3){
1443             if(src[si+2]==3){ //escape
1444                 dst[di++]= 0;
1445                 dst[di++]= 0;
1446                 si+=3;
1447                 continue;
1448             }else //next start code
1449                 break;
1450         }
1451
1452         dst[di++]= src[si++];
1453     }
1454
1455     *dst_length= di;
1456     *consumed= si + 1;//+1 for the header
1457 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1458     return dst;
1459 }
1460
1461 /**
1462  * identifies the exact end of the bitstream
1463  * @return the length of the trailing, or 0 if damaged
1464  */
1465 static int decode_rbsp_trailing(H264Context *h, uint8_t *src){
1466     int v= *src;
1467     int r;
1468
1469     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1470
1471     for(r=1; r<9; r++){
1472         if(v&1) return r;
1473         v>>=1;
1474     }
1475     return 0;
1476 }
1477
1478 /**
1479  * idct tranforms the 16 dc values and dequantize them.
1480  * @param qp quantization parameter
1481  */
1482 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1483 #define stride 16
1484     int i;
1485     int temp[16]; //FIXME check if this is a good idea
1486     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1487     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1488
1489 //memset(block, 64, 2*256);
1490 //return;
1491     for(i=0; i<4; i++){
1492         const int offset= y_offset[i];
1493         const int z0= block[offset+stride*0] + block[offset+stride*4];
1494         const int z1= block[offset+stride*0] - block[offset+stride*4];
1495         const int z2= block[offset+stride*1] - block[offset+stride*5];
1496         const int z3= block[offset+stride*1] + block[offset+stride*5];
1497
1498         temp[4*i+0]= z0+z3;
1499         temp[4*i+1]= z1+z2;
1500         temp[4*i+2]= z1-z2;
1501         temp[4*i+3]= z0-z3;
1502     }
1503
1504     for(i=0; i<4; i++){
1505         const int offset= x_offset[i];
1506         const int z0= temp[4*0+i] + temp[4*2+i];
1507         const int z1= temp[4*0+i] - temp[4*2+i];
1508         const int z2= temp[4*1+i] - temp[4*3+i];
1509         const int z3= temp[4*1+i] + temp[4*3+i];
1510
1511         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_resdual
1512         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1513         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1514         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1515     }
1516 }
1517
1518 #if 0
1519 /**
1520  * dct tranforms the 16 dc values.
1521  * @param qp quantization parameter ??? FIXME
1522  */
1523 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1524 //    const int qmul= dequant_coeff[qp][0];
1525     int i;
1526     int temp[16]; //FIXME check if this is a good idea
1527     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1528     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1529
1530     for(i=0; i<4; i++){
1531         const int offset= y_offset[i];
1532         const int z0= block[offset+stride*0] + block[offset+stride*4];
1533         const int z1= block[offset+stride*0] - block[offset+stride*4];
1534         const int z2= block[offset+stride*1] - block[offset+stride*5];
1535         const int z3= block[offset+stride*1] + block[offset+stride*5];
1536
1537         temp[4*i+0]= z0+z3;
1538         temp[4*i+1]= z1+z2;
1539         temp[4*i+2]= z1-z2;
1540         temp[4*i+3]= z0-z3;
1541     }
1542
1543     for(i=0; i<4; i++){
1544         const int offset= x_offset[i];
1545         const int z0= temp[4*0+i] + temp[4*2+i];
1546         const int z1= temp[4*0+i] - temp[4*2+i];
1547         const int z2= temp[4*1+i] - temp[4*3+i];
1548         const int z3= temp[4*1+i] + temp[4*3+i];
1549
1550         block[stride*0 +offset]= (z0 + z3)>>1;
1551         block[stride*2 +offset]= (z1 + z2)>>1;
1552         block[stride*8 +offset]= (z1 - z2)>>1;
1553         block[stride*10+offset]= (z0 - z3)>>1;
1554     }
1555 }
1556 #endif
1557
1558 #undef xStride
1559 #undef stride
1560
1561 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1562     const int stride= 16*2;
1563     const int xStride= 16;
1564     int a,b,c,d,e;
1565
1566     a= block[stride*0 + xStride*0];
1567     b= block[stride*0 + xStride*1];
1568     c= block[stride*1 + xStride*0];
1569     d= block[stride*1 + xStride*1];
1570
1571     e= a-b;
1572     a= a+b;
1573     b= c-d;
1574     c= c+d;
1575
1576     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1577     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1578     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1579     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1580 }
1581
1582 #if 0
1583 static void chroma_dc_dct_c(DCTELEM *block){
1584     const int stride= 16*2;
1585     const int xStride= 16;
1586     int a,b,c,d,e;
1587
1588     a= block[stride*0 + xStride*0];
1589     b= block[stride*0 + xStride*1];
1590     c= block[stride*1 + xStride*0];
1591     d= block[stride*1 + xStride*1];
1592
1593     e= a-b;
1594     a= a+b;
1595     b= c-d;
1596     c= c+d;
1597
1598     block[stride*0 + xStride*0]= (a+c);
1599     block[stride*0 + xStride*1]= (e+b);
1600     block[stride*1 + xStride*0]= (a-c);
1601     block[stride*1 + xStride*1]= (e-b);
1602 }
1603 #endif
1604
1605 /**
1606  * gets the chroma qp.
1607  */
1608 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1609     return h->pps.chroma_qp_table[t][qscale & 0xff];
1610 }
1611
1612 //FIXME need to check that this does not overflow signed 32 bit for low qp, i am not sure, it's very close
1613 //FIXME check that gcc inlines this (and optimizes intra & separate_dc stuff away)
1614 static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int separate_dc){
1615     int i;
1616     const int * const quant_table= quant_coeff[qscale];
1617     const int bias= intra ? (1<<QUANT_SHIFT)/3 : (1<<QUANT_SHIFT)/6;
1618     const unsigned int threshold1= (1<<QUANT_SHIFT) - bias - 1;
1619     const unsigned int threshold2= (threshold1<<1);
1620     int last_non_zero;
1621
1622     if(separate_dc){
1623         if(qscale<=18){
1624             //avoid overflows
1625             const int dc_bias= intra ? (1<<(QUANT_SHIFT-2))/3 : (1<<(QUANT_SHIFT-2))/6;
1626             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT-2)) - dc_bias - 1;
1627             const unsigned int dc_threshold2= (dc_threshold1<<1);
1628
1629             int level= block[0]*quant_coeff[qscale+18][0];
1630             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1631                 if(level>0){
1632                     level= (dc_bias + level)>>(QUANT_SHIFT-2);
1633                     block[0]= level;
1634                 }else{
1635                     level= (dc_bias - level)>>(QUANT_SHIFT-2);
1636                     block[0]= -level;
1637                 }
1638 //                last_non_zero = i;
1639             }else{
1640                 block[0]=0;
1641             }
1642         }else{
1643             const int dc_bias= intra ? (1<<(QUANT_SHIFT+1))/3 : (1<<(QUANT_SHIFT+1))/6;
1644             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT+1)) - dc_bias - 1;
1645             const unsigned int dc_threshold2= (dc_threshold1<<1);
1646
1647             int level= block[0]*quant_table[0];
1648             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1649                 if(level>0){
1650                     level= (dc_bias + level)>>(QUANT_SHIFT+1);
1651                     block[0]= level;
1652                 }else{
1653                     level= (dc_bias - level)>>(QUANT_SHIFT+1);
1654                     block[0]= -level;
1655                 }
1656 //                last_non_zero = i;
1657             }else{
1658                 block[0]=0;
1659             }
1660         }
1661         last_non_zero= 0;
1662         i=1;
1663     }else{
1664         last_non_zero= -1;
1665         i=0;
1666     }
1667
1668     for(; i<16; i++){
1669         const int j= scantable[i];
1670         int level= block[j]*quant_table[j];
1671
1672 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
1673 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
1674         if(((unsigned)(level+threshold1))>threshold2){
1675             if(level>0){
1676                 level= (bias + level)>>QUANT_SHIFT;
1677                 block[j]= level;
1678             }else{
1679                 level= (bias - level)>>QUANT_SHIFT;
1680                 block[j]= -level;
1681             }
1682             last_non_zero = i;
1683         }else{
1684             block[j]=0;
1685         }
1686     }
1687
1688     return last_non_zero;
1689 }
1690
1691 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1692                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1693                            int src_x_offset, int src_y_offset,
1694                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1695     MpegEncContext * const s = &h->s;
1696     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1697     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1698     const int luma_xy= (mx&3) + ((my&3)<<2);
1699     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1700     uint8_t * src_cb, * src_cr;
1701     int extra_width= h->emu_edge_width;
1702     int extra_height= h->emu_edge_height;
1703     int emu=0;
1704     const int full_mx= mx>>2;
1705     const int full_my= my>>2;
1706     const int pic_width  = 16*s->mb_width;
1707     const int pic_height = 16*s->mb_height >> MB_FIELD;
1708
1709     if(!pic->data[0]) //FIXME this is unacceptable, some senseable error concealment must be done for missing reference frames
1710         return;
1711
1712     if(mx&7) extra_width -= 3;
1713     if(my&7) extra_height -= 3;
1714
1715     if(   full_mx < 0-extra_width
1716        || full_my < 0-extra_height
1717        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1718        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1719         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);
1720             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1721         emu=1;
1722     }
1723
1724     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1725     if(!square){
1726         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1727     }
1728
1729     if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1730
1731     if(MB_FIELD){
1732         // chroma offset when predicting from a field of opposite parity
1733         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1734         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1735     }
1736     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1737     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1738
1739     if(emu){
1740         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);
1741             src_cb= s->edge_emu_buffer;
1742     }
1743     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1744
1745     if(emu){
1746         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);
1747             src_cr= s->edge_emu_buffer;
1748     }
1749     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1750 }
1751
1752 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1753                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1754                            int x_offset, int y_offset,
1755                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1756                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1757                            int list0, int list1){
1758     MpegEncContext * const s = &h->s;
1759     qpel_mc_func *qpix_op=  qpix_put;
1760     h264_chroma_mc_func chroma_op= chroma_put;
1761
1762     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1763     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1764     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1765     x_offset += 8*s->mb_x;
1766     y_offset += 8*(s->mb_y >> MB_FIELD);
1767
1768     if(list0){
1769         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1770         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1771                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1772                            qpix_op, chroma_op);
1773
1774         qpix_op=  qpix_avg;
1775         chroma_op= chroma_avg;
1776     }
1777
1778     if(list1){
1779         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1780         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1781                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1782                            qpix_op, chroma_op);
1783     }
1784 }
1785
1786 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1787                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1788                            int x_offset, int y_offset,
1789                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1790                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1791                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1792                            int list0, int list1){
1793     MpegEncContext * const s = &h->s;
1794
1795     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1796     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1797     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1798     x_offset += 8*s->mb_x;
1799     y_offset += 8*(s->mb_y >> MB_FIELD);
1800
1801     if(list0 && list1){
1802         /* don't optimize for luma-only case, since B-frames usually
1803          * use implicit weights => chroma too. */
1804         uint8_t *tmp_cb = s->obmc_scratchpad;
1805         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1806         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1807         int refn0 = h->ref_cache[0][ scan8[n] ];
1808         int refn1 = h->ref_cache[1][ scan8[n] ];
1809
1810         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1811                     dest_y, dest_cb, dest_cr,
1812                     x_offset, y_offset, qpix_put, chroma_put);
1813         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1814                     tmp_y, tmp_cb, tmp_cr,
1815                     x_offset, y_offset, qpix_put, chroma_put);
1816
1817         if(h->use_weight == 2){
1818             int weight0 = h->implicit_weight[refn0][refn1];
1819             int weight1 = 64 - weight0;
1820             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1821             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1822             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1823         }else{
1824             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1825                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1826                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1827             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1828                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1829                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1830             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1831                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1832                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1833         }
1834     }else{
1835         int list = list1 ? 1 : 0;
1836         int refn = h->ref_cache[list][ scan8[n] ];
1837         Picture *ref= &h->ref_list[list][refn];
1838         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1839                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1840                     qpix_put, chroma_put);
1841
1842         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1843                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1844         if(h->use_weight_chroma){
1845             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1846                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1847             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1848                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1849         }
1850     }
1851 }
1852
1853 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1854                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1855                            int x_offset, int y_offset,
1856                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1857                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1858                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1859                            int list0, int list1){
1860     if((h->use_weight==2 && list0 && list1
1861         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1862        || h->use_weight==1)
1863         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1864                          x_offset, y_offset, qpix_put, chroma_put,
1865                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1866     else
1867         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1868                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1869 }
1870
1871 static inline void prefetch_motion(H264Context *h, int list){
1872     /* fetch pixels for estimated mv 4 macroblocks ahead
1873      * optimized for 64byte cache lines */
1874     MpegEncContext * const s = &h->s;
1875     const int refn = h->ref_cache[list][scan8[0]];
1876     if(refn >= 0){
1877         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1878         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1879         uint8_t **src= h->ref_list[list][refn].data;
1880         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1881         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1882         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1883         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1884     }
1885 }
1886
1887 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1888                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1889                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1890                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1891     MpegEncContext * const s = &h->s;
1892     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
1893     const int mb_type= s->current_picture.mb_type[mb_xy];
1894
1895     assert(IS_INTER(mb_type));
1896
1897     prefetch_motion(h, 0);
1898
1899     if(IS_16X16(mb_type)){
1900         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1901                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1902                 &weight_op[0], &weight_avg[0],
1903                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1904     }else if(IS_16X8(mb_type)){
1905         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1906                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1907                 &weight_op[1], &weight_avg[1],
1908                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1909         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1910                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1911                 &weight_op[1], &weight_avg[1],
1912                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1913     }else if(IS_8X16(mb_type)){
1914         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1915                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1916                 &weight_op[2], &weight_avg[2],
1917                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1918         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1919                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1920                 &weight_op[2], &weight_avg[2],
1921                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1922     }else{
1923         int i;
1924
1925         assert(IS_8X8(mb_type));
1926
1927         for(i=0; i<4; i++){
1928             const int sub_mb_type= h->sub_mb_type[i];
1929             const int n= 4*i;
1930             int x_offset= (i&1)<<2;
1931             int y_offset= (i&2)<<1;
1932
1933             if(IS_SUB_8X8(sub_mb_type)){
1934                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1935                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1936                     &weight_op[3], &weight_avg[3],
1937                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1938             }else if(IS_SUB_8X4(sub_mb_type)){
1939                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1940                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1941                     &weight_op[4], &weight_avg[4],
1942                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1943                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1944                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1945                     &weight_op[4], &weight_avg[4],
1946                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1947             }else if(IS_SUB_4X8(sub_mb_type)){
1948                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1949                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1950                     &weight_op[5], &weight_avg[5],
1951                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1952                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1953                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1954                     &weight_op[5], &weight_avg[5],
1955                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1956             }else{
1957                 int j;
1958                 assert(IS_SUB_4X4(sub_mb_type));
1959                 for(j=0; j<4; j++){
1960                     int sub_x_offset= x_offset + 2*(j&1);
1961                     int sub_y_offset= y_offset +   (j&2);
1962                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1963                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1964                         &weight_op[6], &weight_avg[6],
1965                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1966                 }
1967             }
1968         }
1969     }
1970
1971     prefetch_motion(h, 1);
1972 }
1973
1974 static void decode_init_vlc(void){
1975     static int done = 0;
1976
1977     if (!done) {
1978         int i;
1979         done = 1;
1980
1981         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1982                  &chroma_dc_coeff_token_len [0], 1, 1,
1983                  &chroma_dc_coeff_token_bits[0], 1, 1, 1);
1984
1985         for(i=0; i<4; i++){
1986             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1987                      &coeff_token_len [i][0], 1, 1,
1988                      &coeff_token_bits[i][0], 1, 1, 1);
1989         }
1990
1991         for(i=0; i<3; i++){
1992             init_vlc(&chroma_dc_total_zeros_vlc[i], CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1993                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1994                      &chroma_dc_total_zeros_bits[i][0], 1, 1, 1);
1995         }
1996         for(i=0; i<15; i++){
1997             init_vlc(&total_zeros_vlc[i], TOTAL_ZEROS_VLC_BITS, 16,
1998                      &total_zeros_len [i][0], 1, 1,
1999                      &total_zeros_bits[i][0], 1, 1, 1);
2000         }
2001
2002         for(i=0; i<6; i++){
2003             init_vlc(&run_vlc[i], RUN_VLC_BITS, 7,
2004                      &run_len [i][0], 1, 1,
2005                      &run_bits[i][0], 1, 1, 1);
2006         }
2007         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
2008                  &run_len [6][0], 1, 1,
2009                  &run_bits[6][0], 1, 1, 1);
2010     }
2011 }
2012
2013 static void free_tables(H264Context *h){
2014     int i;
2015     H264Context *hx;
2016     av_freep(&h->intra4x4_pred_mode);
2017     av_freep(&h->chroma_pred_mode_table);
2018     av_freep(&h->cbp_table);
2019     av_freep(&h->mvd_table[0]);
2020     av_freep(&h->mvd_table[1]);
2021     av_freep(&h->direct_table);
2022     av_freep(&h->non_zero_count);
2023     av_freep(&h->slice_table_base);
2024     h->slice_table= NULL;
2025
2026     av_freep(&h->mb2b_xy);
2027     av_freep(&h->mb2b8_xy);
2028
2029     for(i = 0; i < MAX_SPS_COUNT; i++)
2030         av_freep(h->sps_buffers + i);
2031
2032     for(i = 0; i < MAX_PPS_COUNT; i++)
2033         av_freep(h->pps_buffers + i);
2034
2035     for(i = 0; i < h->s.avctx->thread_count; i++) {
2036         hx = h->thread_context[i];
2037         if(!hx) continue;
2038         av_freep(&hx->top_borders[1]);
2039         av_freep(&hx->top_borders[0]);
2040         av_freep(&hx->s.obmc_scratchpad);
2041         av_freep(&hx->s.allocated_edge_emu_buffer);
2042     }
2043 }
2044
2045 static void init_dequant8_coeff_table(H264Context *h){
2046     int i,q,x;
2047     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2048     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2049     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2050
2051     for(i=0; i<2; i++ ){
2052         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2053             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2054             break;
2055         }
2056
2057         for(q=0; q<52; q++){
2058             int shift = ff_div6[q];
2059             int idx = ff_rem6[q];
2060             for(x=0; x<64; x++)
2061                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2062                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2063                     h->pps.scaling_matrix8[i][x]) << shift;
2064         }
2065     }
2066 }
2067
2068 static void init_dequant4_coeff_table(H264Context *h){
2069     int i,j,q,x;
2070     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2071     for(i=0; i<6; i++ ){
2072         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2073         for(j=0; j<i; j++){
2074             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2075                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2076                 break;
2077             }
2078         }
2079         if(j<i)
2080             continue;
2081
2082         for(q=0; q<52; q++){
2083             int shift = ff_div6[q] + 2;
2084             int idx = ff_rem6[q];
2085             for(x=0; x<16; x++)
2086                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2087                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2088                     h->pps.scaling_matrix4[i][x]) << shift;
2089         }
2090     }
2091 }
2092
2093 static void init_dequant_tables(H264Context *h){
2094     int i,x;
2095     init_dequant4_coeff_table(h);
2096     if(h->pps.transform_8x8_mode)
2097         init_dequant8_coeff_table(h);
2098     if(h->sps.transform_bypass){
2099         for(i=0; i<6; i++)
2100             for(x=0; x<16; x++)
2101                 h->dequant4_coeff[i][0][x] = 1<<6;
2102         if(h->pps.transform_8x8_mode)
2103             for(i=0; i<2; i++)
2104                 for(x=0; x<64; x++)
2105                     h->dequant8_coeff[i][0][x] = 1<<6;
2106     }
2107 }
2108
2109
2110 /**
2111  * allocates tables.
2112  * needs width/height
2113  */
2114 static int alloc_tables(H264Context *h){
2115     MpegEncContext * const s = &h->s;
2116     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2117     int x,y;
2118
2119     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2120
2121     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2122     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(uint8_t))
2123     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2124
2125     if( h->pps.cabac ) {
2126         CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2127         CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2128         CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2129         CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2130     }
2131
2132     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(uint8_t));
2133     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2134
2135     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2136     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2137     for(y=0; y<s->mb_height; y++){
2138         for(x=0; x<s->mb_width; x++){
2139             const int mb_xy= x + y*s->mb_stride;
2140             const int b_xy = 4*x + 4*y*h->b_stride;
2141             const int b8_xy= 2*x + 2*y*h->b8_stride;
2142
2143             h->mb2b_xy [mb_xy]= b_xy;
2144             h->mb2b8_xy[mb_xy]= b8_xy;
2145         }
2146     }
2147
2148     s->obmc_scratchpad = NULL;
2149
2150     if(!h->dequant4_coeff[0])
2151         init_dequant_tables(h);
2152
2153     return 0;
2154 fail:
2155     free_tables(h);
2156     return -1;
2157 }
2158
2159 /**
2160  * Mimic alloc_tables(), but for every context thread.
2161  */
2162 static void clone_tables(H264Context *dst, H264Context *src){
2163     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2164     dst->non_zero_count           = src->non_zero_count;
2165     dst->slice_table              = src->slice_table;
2166     dst->cbp_table                = src->cbp_table;
2167     dst->mb2b_xy                  = src->mb2b_xy;
2168     dst->mb2b8_xy                 = src->mb2b8_xy;
2169     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2170     dst->mvd_table[0]             = src->mvd_table[0];
2171     dst->mvd_table[1]             = src->mvd_table[1];
2172     dst->direct_table             = src->direct_table;
2173
2174     dst->s.obmc_scratchpad = NULL;
2175     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2176 }
2177
2178 /**
2179  * Init context
2180  * Allocate buffers which are not shared amongst multiple threads.
2181  */
2182 static int context_init(H264Context *h){
2183     MpegEncContext * const s = &h->s;
2184
2185     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2186     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2187
2188     // edge emu needs blocksize + filter length - 1 (=17x17 for halfpel / 21x21 for h264)
2189     CHECKED_ALLOCZ(s->allocated_edge_emu_buffer,
2190                    (s->width+64)*2*21*2); //(width + edge + align)*interlaced*MBsize*tolerance
2191     s->edge_emu_buffer= s->allocated_edge_emu_buffer + (s->width+64)*2*21;
2192     return 0;
2193 fail:
2194     return -1; // free_tables will clean up for us
2195 }
2196
2197 static void common_init(H264Context *h){
2198     MpegEncContext * const s = &h->s;
2199
2200     s->width = s->avctx->width;
2201     s->height = s->avctx->height;
2202     s->codec_id= s->avctx->codec->id;
2203
2204     ff_h264_pred_init(&h->hpc, s->codec_id);
2205
2206     h->dequant_coeff_pps= -1;
2207     s->unrestricted_mv=1;
2208     s->decode=1; //FIXME
2209
2210     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2211     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2212 }
2213
2214 static int decode_init(AVCodecContext *avctx){
2215     H264Context *h= avctx->priv_data;
2216     MpegEncContext * const s = &h->s;
2217
2218     MPV_decode_defaults(s);
2219
2220     s->avctx = avctx;
2221     common_init(h);
2222
2223     s->out_format = FMT_H264;
2224     s->workaround_bugs= avctx->workaround_bugs;
2225
2226     // set defaults
2227 //    s->decode_mb= ff_h263_decode_mb;
2228     s->quarter_sample = 1;
2229     s->low_delay= 1;
2230     avctx->pix_fmt= PIX_FMT_YUV420P;
2231
2232     decode_init_vlc();
2233
2234     if(avctx->extradata_size > 0 && avctx->extradata &&
2235        *(char *)avctx->extradata == 1){
2236         h->is_avc = 1;
2237         h->got_avcC = 0;
2238     } else {
2239         h->is_avc = 0;
2240     }
2241
2242     h->thread_context[0] = h;
2243     return 0;
2244 }
2245
2246 static int frame_start(H264Context *h){
2247     MpegEncContext * const s = &h->s;
2248     int i;
2249
2250     if(MPV_frame_start(s, s->avctx) < 0)
2251         return -1;
2252     ff_er_frame_start(s);
2253     /*
2254      * MPV_frame_start uses pict_type to derive key_frame.
2255      * This is incorrect for H.264; IDR markings must be used.
2256      * Zero here; IDR markings per slice in frame or fields are OR'd in later.
2257      * See decode_nal_units().
2258      */
2259     s->current_picture_ptr->key_frame= 0;
2260
2261     assert(s->linesize && s->uvlinesize);
2262
2263     for(i=0; i<16; i++){
2264         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2265         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2266     }
2267     for(i=0; i<4; i++){
2268         h->block_offset[16+i]=
2269         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2270         h->block_offset[24+16+i]=
2271         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2272     }
2273
2274     /* can't be in alloc_tables because linesize isn't known there.
2275      * FIXME: redo bipred weight to not require extra buffer? */
2276     for(i = 0; i < s->avctx->thread_count; i++)
2277         if(!h->thread_context[i]->s.obmc_scratchpad)
2278             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2279
2280     /* some macroblocks will be accessed before they're available */
2281     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2282         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(uint8_t));
2283
2284 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2285     return 0;
2286 }
2287
2288 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){
2289     MpegEncContext * const s = &h->s;
2290     int i;
2291
2292     src_y  -=   linesize;
2293     src_cb -= uvlinesize;
2294     src_cr -= uvlinesize;
2295
2296     // There are two lines saved, the line above the the top macroblock of a pair,
2297     // and the line above the bottom macroblock
2298     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2299     for(i=1; i<17; i++){
2300         h->left_border[i]= src_y[15+i*  linesize];
2301     }
2302
2303     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2304     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2305
2306     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2307         h->left_border[17  ]= h->top_borders[0][s->mb_x][16+7];
2308         h->left_border[17+9]= h->top_borders[0][s->mb_x][24+7];
2309         for(i=1; i<9; i++){
2310             h->left_border[i+17  ]= src_cb[7+i*uvlinesize];
2311             h->left_border[i+17+9]= src_cr[7+i*uvlinesize];
2312         }
2313         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2314         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2315     }
2316 }
2317
2318 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){
2319     MpegEncContext * const s = &h->s;
2320     int temp8, i;
2321     uint64_t temp64;
2322     int deblock_left;
2323     int deblock_top;
2324     int mb_xy;
2325
2326     if(h->deblocking_filter == 2) {
2327         mb_xy = s->mb_x + s->mb_y*s->mb_stride;
2328         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2329         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2330     } else {
2331         deblock_left = (s->mb_x > 0);
2332         deblock_top =  (s->mb_y > 0);
2333     }
2334
2335     src_y  -=   linesize + 1;
2336     src_cb -= uvlinesize + 1;
2337     src_cr -= uvlinesize + 1;
2338
2339 #define XCHG(a,b,t,xchg)\
2340 t= a;\
2341 if(xchg)\
2342     a= b;\
2343 b= t;
2344
2345     if(deblock_left){
2346         for(i = !deblock_top; i<17; i++){
2347             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2348         }
2349     }
2350
2351     if(deblock_top){
2352         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2353         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2354         if(s->mb_x+1 < s->mb_width){
2355             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2356         }
2357     }
2358
2359     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2360         if(deblock_left){
2361             for(i = !deblock_top; i<9; i++){
2362                 XCHG(h->left_border[i+17  ], src_cb[i*uvlinesize], temp8, xchg);
2363                 XCHG(h->left_border[i+17+9], src_cr[i*uvlinesize], temp8, xchg);
2364             }
2365         }
2366         if(deblock_top){
2367             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2368             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2369         }
2370     }
2371 }
2372
2373 static inline void backup_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize){
2374     MpegEncContext * const s = &h->s;
2375     int i;
2376
2377     src_y  -= 2 *   linesize;
2378     src_cb -= 2 * uvlinesize;
2379     src_cr -= 2 * uvlinesize;
2380
2381     // There are two lines saved, the line above the the top macroblock of a pair,
2382     // and the line above the bottom macroblock
2383     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2384     h->left_border[1]= h->top_borders[1][s->mb_x][15];
2385     for(i=2; i<34; i++){
2386         h->left_border[i]= src_y[15+i*  linesize];
2387     }
2388
2389     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  32*linesize);
2390     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+32*linesize);
2391     *(uint64_t*)(h->top_borders[1][s->mb_x]+0)= *(uint64_t*)(src_y +  33*linesize);
2392     *(uint64_t*)(h->top_borders[1][s->mb_x]+8)= *(uint64_t*)(src_y +8+33*linesize);
2393
2394     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2395         h->left_border[34     ]= h->top_borders[0][s->mb_x][16+7];
2396         h->left_border[34+   1]= h->top_borders[1][s->mb_x][16+7];
2397         h->left_border[34+18  ]= h->top_borders[0][s->mb_x][24+7];
2398         h->left_border[34+18+1]= h->top_borders[1][s->mb_x][24+7];
2399         for(i=2; i<18; i++){
2400             h->left_border[i+34   ]= src_cb[7+i*uvlinesize];
2401             h->left_border[i+34+18]= src_cr[7+i*uvlinesize];
2402         }
2403         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+16*uvlinesize);
2404         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+16*uvlinesize);
2405         *(uint64_t*)(h->top_borders[1][s->mb_x]+16)= *(uint64_t*)(src_cb+17*uvlinesize);
2406         *(uint64_t*)(h->top_borders[1][s->mb_x]+24)= *(uint64_t*)(src_cr+17*uvlinesize);
2407     }
2408 }
2409
2410 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){
2411     MpegEncContext * const s = &h->s;
2412     int temp8, i;
2413     uint64_t temp64;
2414     int deblock_left = (s->mb_x > 0);
2415     int deblock_top  = (s->mb_y > 1);
2416
2417     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);
2418
2419     src_y  -= 2 *   linesize + 1;
2420     src_cb -= 2 * uvlinesize + 1;
2421     src_cr -= 2 * uvlinesize + 1;
2422
2423 #define XCHG(a,b,t,xchg)\
2424 t= a;\
2425 if(xchg)\
2426     a= b;\
2427 b= t;
2428
2429     if(deblock_left){
2430         for(i = (!deblock_top)<<1; i<34; i++){
2431             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2432         }
2433     }
2434
2435     if(deblock_top){
2436         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2437         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2438         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+0), *(uint64_t*)(src_y +1 +linesize), temp64, xchg);
2439         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+8), *(uint64_t*)(src_y +9 +linesize), temp64, 1);
2440         if(s->mb_x+1 < s->mb_width){
2441             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2442             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x+1]), *(uint64_t*)(src_y +17 +linesize), temp64, 1);
2443         }
2444     }
2445
2446     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2447         if(deblock_left){
2448             for(i = (!deblock_top) << 1; i<18; i++){
2449                 XCHG(h->left_border[i+34   ], src_cb[i*uvlinesize], temp8, xchg);
2450                 XCHG(h->left_border[i+34+18], src_cr[i*uvlinesize], temp8, xchg);
2451             }
2452         }
2453         if(deblock_top){
2454             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2455             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2456             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+16), *(uint64_t*)(src_cb+1 +uvlinesize), temp64, 1);
2457             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+24), *(uint64_t*)(src_cr+1 +uvlinesize), temp64, 1);
2458         }
2459     }
2460 }
2461
2462 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2463     MpegEncContext * const s = &h->s;
2464     const int mb_x= s->mb_x;
2465     const int mb_y= s->mb_y;
2466     const int mb_xy= mb_x + mb_y*s->mb_stride;
2467     const int mb_type= s->current_picture.mb_type[mb_xy];
2468     uint8_t  *dest_y, *dest_cb, *dest_cr;
2469     int linesize, uvlinesize /*dct_offset*/;
2470     int i;
2471     int *block_offset = &h->block_offset[0];
2472     const unsigned int bottom = mb_y & 1;
2473     const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass), is_h264 = (simple || s->codec_id == CODEC_ID_H264);
2474     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2475     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2476
2477     dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2478     dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2479     dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2480
2481     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2482     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2483
2484     if (!simple && MB_FIELD) {
2485         linesize   = h->mb_linesize   = s->linesize * 2;
2486         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2487         block_offset = &h->block_offset[24];
2488         if(mb_y&1){ //FIXME move out of this func?
2489             dest_y -= s->linesize*15;
2490             dest_cb-= s->uvlinesize*7;
2491             dest_cr-= s->uvlinesize*7;
2492         }
2493         if(FRAME_MBAFF) {
2494             int list;
2495             for(list=0; list<h->list_count; list++){
2496                 if(!USES_LIST(mb_type, list))
2497                     continue;
2498                 if(IS_16X16(mb_type)){
2499                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2500                     fill_rectangle(ref, 4, 4, 8, 16+*ref^(s->mb_y&1), 1);
2501                 }else{
2502                     for(i=0; i<16; i+=4){
2503                         //FIXME can refs be smaller than 8x8 when !direct_8x8_inference ?
2504                         int ref = h->ref_cache[list][scan8[i]];
2505                         if(ref >= 0)
2506                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, 16+ref^(s->mb_y&1), 1);
2507                     }
2508                 }
2509             }
2510         }
2511     } else {
2512         linesize   = h->mb_linesize   = s->linesize;
2513         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2514 //        dct_offset = s->linesize * 16;
2515     }
2516
2517     if(transform_bypass){
2518         idct_dc_add =
2519         idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2520     }else if(IS_8x8DCT(mb_type)){
2521         idct_dc_add = s->dsp.h264_idct8_dc_add;
2522         idct_add = s->dsp.h264_idct8_add;
2523     }else{
2524         idct_dc_add = s->dsp.h264_idct_dc_add;
2525         idct_add = s->dsp.h264_idct_add;
2526     }
2527
2528     if(!simple && FRAME_MBAFF && h->deblocking_filter && IS_INTRA(mb_type)
2529        && (!bottom || !IS_INTRA(s->current_picture.mb_type[mb_xy-s->mb_stride]))){
2530         int mbt_y = mb_y&~1;
2531         uint8_t *top_y  = s->current_picture.data[0] + (mbt_y * 16* s->linesize  ) + mb_x * 16;
2532         uint8_t *top_cb = s->current_picture.data[1] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2533         uint8_t *top_cr = s->current_picture.data[2] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2534         xchg_pair_border(h, top_y, top_cb, top_cr, s->linesize, s->uvlinesize, 1);
2535     }
2536
2537     if (!simple && IS_INTRA_PCM(mb_type)) {
2538         unsigned int x, y;
2539
2540         // The pixels are stored in h->mb array in the same order as levels,
2541         // copy them in output in the correct order.
2542         for(i=0; i<16; i++) {
2543             for (y=0; y<4; y++) {
2544                 for (x=0; x<4; x++) {
2545                     *(dest_y + block_offset[i] + y*linesize + x) = h->mb[i*16+y*4+x];
2546                 }
2547             }
2548         }
2549         for(i=16; i<16+4; i++) {
2550             for (y=0; y<4; y++) {
2551                 for (x=0; x<4; x++) {
2552                     *(dest_cb + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2553                 }
2554             }
2555         }
2556         for(i=20; i<20+4; i++) {
2557             for (y=0; y<4; y++) {
2558                 for (x=0; x<4; x++) {
2559                     *(dest_cr + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2560                 }
2561             }
2562         }
2563     } else {
2564         if(IS_INTRA(mb_type)){
2565             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2566                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2567
2568             if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2569                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2570                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2571             }
2572
2573             if(IS_INTRA4x4(mb_type)){
2574                 if(simple || !s->encoding){
2575                     if(IS_8x8DCT(mb_type)){
2576                         for(i=0; i<16; i+=4){
2577                             uint8_t * const ptr= dest_y + block_offset[i];
2578                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2579                             const int nnz = h->non_zero_count_cache[ scan8[i] ];
2580                             h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2581                                                    (h->topright_samples_available<<i)&0x4000, linesize);
2582                             if(nnz){
2583                                 if(nnz == 1 && h->mb[i*16])
2584                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2585                                 else
2586                                     idct_add(ptr, h->mb + i*16, linesize);
2587                             }
2588                         }
2589                     }else
2590                     for(i=0; i<16; i++){
2591                         uint8_t * const ptr= dest_y + block_offset[i];
2592                         uint8_t *topright;
2593                         const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2594                         int nnz, tr;
2595
2596                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2597                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2598                             assert(mb_y || linesize <= block_offset[i]);
2599                             if(!topright_avail){
2600                                 tr= ptr[3 - linesize]*0x01010101;
2601                                 topright= (uint8_t*) &tr;
2602                             }else
2603                                 topright= ptr + 4 - linesize;
2604                         }else
2605                             topright= NULL;
2606
2607                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2608                         nnz = h->non_zero_count_cache[ scan8[i] ];
2609                         if(nnz){
2610                             if(is_h264){
2611                                 if(nnz == 1 && h->mb[i*16])
2612                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2613                                 else
2614                                     idct_add(ptr, h->mb + i*16, linesize);
2615                             }else
2616                                 svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2617                         }
2618                     }
2619                 }
2620             }else{
2621                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2622                 if(is_h264){
2623                     if(!transform_bypass)
2624                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2625                 }else
2626                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2627             }
2628             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2629                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2630         }else if(is_h264){
2631             hl_motion(h, dest_y, dest_cb, dest_cr,
2632                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2633                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2634                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2635         }
2636
2637
2638         if(!IS_INTRA4x4(mb_type)){
2639             if(is_h264){
2640                 if(IS_INTRA16x16(mb_type)){
2641                     for(i=0; i<16; i++){
2642                         if(h->non_zero_count_cache[ scan8[i] ])
2643                             idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2644                         else if(h->mb[i*16])
2645                             idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2646                     }
2647                 }else{
2648                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2649                     for(i=0; i<16; i+=di){
2650                         int nnz = h->non_zero_count_cache[ scan8[i] ];
2651                         if(nnz){
2652                             if(nnz==1 && h->mb[i*16])
2653                                 idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2654                             else
2655                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2656                         }
2657                     }
2658                 }
2659             }else{
2660                 for(i=0; i<16; i++){
2661                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2662                         uint8_t * const ptr= dest_y + block_offset[i];
2663                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2664                     }
2665                 }
2666             }
2667         }
2668
2669         if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2670             uint8_t *dest[2] = {dest_cb, dest_cr};
2671             if(transform_bypass){
2672                 idct_add = idct_dc_add = s->dsp.add_pixels4;
2673             }else{
2674                 idct_add = s->dsp.h264_idct_add;
2675                 idct_dc_add = s->dsp.h264_idct_dc_add;
2676                 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]);
2677                 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]);
2678             }
2679             if(is_h264){
2680                 for(i=16; i<16+8; i++){
2681                     if(h->non_zero_count_cache[ scan8[i] ])
2682                         idct_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2683                     else if(h->mb[i*16])
2684                         idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2685                 }
2686             }else{
2687                 for(i=16; i<16+8; i++){
2688                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2689                         uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2690                         svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2691                     }
2692                 }
2693             }
2694         }
2695     }
2696     if(h->deblocking_filter) {
2697         if (!simple && FRAME_MBAFF) {
2698             //FIXME try deblocking one mb at a time?
2699             // the reduction in load/storing mvs and such might outweigh the extra backup/xchg_border
2700             const int mb_y = s->mb_y - 1;
2701             uint8_t  *pair_dest_y, *pair_dest_cb, *pair_dest_cr;
2702             const int mb_xy= mb_x + mb_y*s->mb_stride;
2703             const int mb_type_top   = s->current_picture.mb_type[mb_xy];
2704             const int mb_type_bottom= s->current_picture.mb_type[mb_xy+s->mb_stride];
2705             if (!bottom) return;
2706             pair_dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2707             pair_dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2708             pair_dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2709
2710             if(IS_INTRA(mb_type_top | mb_type_bottom))
2711                 xchg_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize, 0);
2712
2713             backup_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize);
2714             // deblock a pair
2715             // top
2716             s->mb_y--;
2717             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);
2718             fill_caches(h, mb_type_top, 1); //FIXME don't fill stuff which isn't used by filter_mb
2719             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2720             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2721             filter_mb(h, mb_x, mb_y, pair_dest_y, pair_dest_cb, pair_dest_cr, linesize, uvlinesize);
2722             // bottom
2723             s->mb_y++;
2724             tprintf(h->s.avctx, "call mbaff filter_mb\n");
2725             fill_caches(h, mb_type_bottom, 1); //FIXME don't fill stuff which isn't used by filter_mb
2726             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2727             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2728             filter_mb(h, mb_x, mb_y+1, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2729         } else {
2730             tprintf(h->s.avctx, "call filter_mb\n");
2731             backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2732             fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2733             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2734         }
2735     }
2736 }
2737
2738 /**
2739  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2740  */
2741 static void hl_decode_mb_simple(H264Context *h){
2742     hl_decode_mb_internal(h, 1);
2743 }
2744
2745 /**
2746  * Process a macroblock; this handles edge cases, such as interlacing.
2747  */
2748 static void av_noinline hl_decode_mb_complex(H264Context *h){
2749     hl_decode_mb_internal(h, 0);
2750 }
2751
2752 static void hl_decode_mb(H264Context *h){
2753     MpegEncContext * const s = &h->s;
2754     const int mb_x= s->mb_x;
2755     const int mb_y= s->mb_y;
2756     const int mb_xy= mb_x + mb_y*s->mb_stride;
2757     const int mb_type= s->current_picture.mb_type[mb_xy];
2758     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;
2759
2760     if(!s->decode)
2761         return;
2762
2763     if (is_complex)
2764         hl_decode_mb_complex(h);
2765     else hl_decode_mb_simple(h);
2766 }
2767
2768 static void pic_as_field(Picture *pic, const int parity){
2769     int i;
2770     for (i = 0; i < 4; ++i) {
2771         if (parity == PICT_BOTTOM_FIELD)
2772             pic->data[i] += pic->linesize[i];
2773         pic->reference = parity;
2774         pic->linesize[i] *= 2;
2775     }
2776 }
2777
2778 static int split_field_copy(Picture *dest, Picture *src,
2779                             int parity, int id_add){
2780     int match = !!(src->reference & parity);
2781
2782     if (match) {
2783         *dest = *src;
2784         pic_as_field(dest, parity);
2785         dest->pic_id *= 2;
2786         dest->pic_id += id_add;
2787     }
2788
2789     return match;
2790 }
2791
2792 /**
2793  * Split one reference list into field parts, interleaving by parity
2794  * as per H.264 spec section 8.2.4.2.5. Output fields have their data pointers
2795  * set to look at the actual start of data for that field.
2796  *
2797  * @param dest output list
2798  * @param dest_len maximum number of fields to put in dest
2799  * @param src the source reference list containing fields and/or field pairs
2800  *            (aka short_ref/long_ref, or
2801  *             refFrameListXShortTerm/refFrameListLongTerm in spec-speak)
2802  * @param src_len number of Picture's in source (pairs and unmatched fields)
2803  * @param parity the parity of the picture being decoded/needing
2804  *        these ref pics (PICT_{TOP,BOTTOM}_FIELD)
2805  * @return number of fields placed in dest
2806  */
2807 static int split_field_half_ref_list(Picture *dest, int dest_len,
2808                                      Picture *src,  int src_len,  int parity){
2809     int same_parity   = 1;
2810     int same_i        = 0;
2811     int opp_i         = 0;
2812     int out_i;
2813     int field_output;
2814
2815     for (out_i = 0; out_i < dest_len; out_i += field_output) {
2816         if (same_parity && same_i < src_len) {
2817             field_output = split_field_copy(dest + out_i, src + same_i,
2818                                             parity, 1);
2819             same_parity = !field_output;
2820             same_i++;
2821
2822         } else if (opp_i < src_len) {
2823             field_output = split_field_copy(dest + out_i, src + opp_i,
2824                                             PICT_FRAME - parity, 0);
2825             same_parity = field_output;
2826             opp_i++;
2827
2828         } else {
2829             break;
2830         }
2831     }
2832
2833     return out_i;
2834 }
2835
2836 /**
2837  * Split the reference frame list into a reference field list.
2838  * This implements H.264 spec 8.2.4.2.5 for a combined input list.
2839  * The input list contains both reference field pairs and
2840  * unmatched reference fields; it is ordered as spec describes
2841  * RefPicListX for frames in 8.2.4.2.1 and 8.2.4.2.3, except that
2842  * unmatched field pairs are also present. Conceptually this is equivalent
2843  * to concatenation of refFrameListXShortTerm with refFrameListLongTerm.
2844  *
2845  * @param dest output reference list where ordered fields are to be placed
2846  * @param dest_len max number of fields to place at dest
2847  * @param src source reference list, as described above
2848  * @param src_len number of pictures (pairs and unmatched fields) in src
2849  * @param parity parity of field being currently decoded
2850  *        (one of PICT_{TOP,BOTTOM}_FIELD)
2851  * @param long_i index into src array that holds first long reference picture,
2852  *        or src_len if no long refs present.
2853  */
2854 static int split_field_ref_list(Picture *dest, int dest_len,
2855                                 Picture *src,  int src_len,
2856                                 int parity,    int long_i){
2857
2858     int i = split_field_half_ref_list(dest, dest_len, src, long_i, parity);
2859     dest += i;
2860     dest_len -= i;
2861
2862     i += split_field_half_ref_list(dest, dest_len, src + long_i,
2863                                    src_len - long_i, parity);
2864     return i;
2865 }
2866
2867 /**
2868  * fills the default_ref_list.
2869  */
2870 static int fill_default_ref_list(H264Context *h){
2871     MpegEncContext * const s = &h->s;
2872     int i;
2873     int smallest_poc_greater_than_current = -1;
2874     int structure_sel;
2875     Picture sorted_short_ref[32];
2876     Picture field_entry_list[2][32];
2877     Picture *frame_list[2];
2878
2879     if (FIELD_PICTURE) {
2880         structure_sel = PICT_FRAME;
2881         frame_list[0] = field_entry_list[0];
2882         frame_list[1] = field_entry_list[1];
2883     } else {
2884         structure_sel = 0;
2885         frame_list[0] = h->default_ref_list[0];
2886         frame_list[1] = h->default_ref_list[1];
2887     }
2888
2889     if(h->slice_type==B_TYPE){
2890         int list;
2891         int len[2];
2892         int short_len[2];
2893         int out_i;
2894         int limit= INT_MIN;
2895
2896         /* sort frame according to poc in B slice */
2897         for(out_i=0; out_i<h->short_ref_count; out_i++){
2898             int best_i=INT_MIN;
2899             int best_poc=INT_MAX;
2900
2901             for(i=0; i<h->short_ref_count; i++){
2902                 const int poc= h->short_ref[i]->poc;
2903                 if(poc > limit && poc < best_poc){
2904                     best_poc= poc;
2905                     best_i= i;
2906                 }
2907             }
2908
2909             assert(best_i != INT_MIN);
2910
2911             limit= best_poc;
2912             sorted_short_ref[out_i]= *h->short_ref[best_i];
2913             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);
2914             if (-1 == smallest_poc_greater_than_current) {
2915                 if (h->short_ref[best_i]->poc >= s->current_picture_ptr->poc) {
2916                     smallest_poc_greater_than_current = out_i;
2917                 }
2918             }
2919         }
2920
2921         tprintf(h->s.avctx, "current poc: %d, smallest_poc_greater_than_current: %d\n", s->current_picture_ptr->poc, smallest_poc_greater_than_current);
2922
2923         // find the largest poc
2924         for(list=0; list<2; list++){
2925             int index = 0;
2926             int j= -99;
2927             int step= list ? -1 : 1;
2928
2929             for(i=0; i<h->short_ref_count && index < h->ref_count[list]; i++, j+=step) {
2930                 int sel;
2931                 while(j<0 || j>= h->short_ref_count){
2932                     if(j != -99 && step == (list ? -1 : 1))
2933                         return -1;
2934                     step = -step;
2935                     j= smallest_poc_greater_than_current + (step>>1);
2936                 }
2937                 sel = sorted_short_ref[j].reference | structure_sel;
2938                 if(sel != PICT_FRAME) continue;
2939                 frame_list[list][index  ]= sorted_short_ref[j];
2940                 frame_list[list][index++].pic_id= sorted_short_ref[j].frame_num;
2941             }
2942             short_len[list] = index;
2943
2944             for(i = 0; i < 16 && index < h->ref_count[ list ]; i++){
2945                 int sel;
2946                 if(h->long_ref[i] == NULL) continue;
2947                 sel = h->long_ref[i]->reference | structure_sel;
2948                 if(sel != PICT_FRAME) continue;
2949
2950                 frame_list[ list ][index  ]= *h->long_ref[i];
2951                 frame_list[ list ][index++].pic_id= i;;
2952             }
2953             len[list] = index;
2954
2955             if(list && (smallest_poc_greater_than_current<=0 || smallest_poc_greater_than_current>=h->short_ref_count) && (1 < index)){
2956                 // swap the two first elements of L1 when
2957                 // L0 and L1 are identical
2958                 Picture temp= frame_list[1][0];
2959                 frame_list[1][0] = frame_list[1][1];
2960                 frame_list[1][1] = temp;
2961             }
2962
2963         }
2964
2965         for(list=0; list<2; list++){
2966             if (FIELD_PICTURE)
2967                 len[list] = split_field_ref_list(h->default_ref_list[list],
2968                                                  h->ref_count[list],
2969                                                  frame_list[list],
2970                                                  len[list],
2971                                                  s->picture_structure,
2972                                                  short_len[list]);
2973
2974             if(len[list] < h->ref_count[ list ])
2975                 memset(&h->default_ref_list[list][len[list]], 0, sizeof(Picture)*(h->ref_count[ list ] - len[list]));
2976         }
2977
2978
2979     }else{
2980         int index=0;
2981         int short_len;
2982         for(i=0; i<h->short_ref_count; i++){
2983             int sel;
2984             sel = h->short_ref[i]->reference | structure_sel;
2985             if(sel != PICT_FRAME) continue;
2986             frame_list[0][index  ]= *h->short_ref[i];
2987             frame_list[0][index++].pic_id= h->short_ref[i]->frame_num;
2988         }
2989         short_len = index;
2990         for(i = 0; i < 16; i++){
2991             int sel;
2992             if(h->long_ref[i] == NULL) continue;
2993             sel = h->long_ref[i]->reference | structure_sel;
2994             if(sel != PICT_FRAME) continue;
2995             frame_list[0][index  ]= *h->long_ref[i];
2996             frame_list[0][index++].pic_id= i;;
2997         }
2998
2999         if (FIELD_PICTURE)
3000             index = split_field_ref_list(h->default_ref_list[0],
3001                                          h->ref_count[0], frame_list[0],
3002                                          index, s->picture_structure,
3003                                          short_len);
3004
3005         if(index < h->ref_count[0])
3006             memset(&h->default_ref_list[0][index], 0, sizeof(Picture)*(h->ref_count[0] - index));
3007     }
3008 #ifdef TRACE
3009     for (i=0; i<h->ref_count[0]; i++) {
3010         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]);
3011     }
3012     if(h->slice_type==B_TYPE){
3013         for (i=0; i<h->ref_count[1]; i++) {
3014             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]);
3015         }
3016     }
3017 #endif
3018     return 0;
3019 }
3020
3021 static void print_short_term(H264Context *h);
3022 static void print_long_term(H264Context *h);
3023
3024 /**
3025  * Extract structure information about the picture described by pic_num in
3026  * the current decoding context (frame or field). Note that pic_num is
3027  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
3028  * @param pic_num picture number for which to extract structure information
3029  * @param structure one of PICT_XXX describing structure of picture
3030  *                      with pic_num
3031  * @return frame number (short term) or long term index of picture
3032  *         described by pic_num
3033  */
3034 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
3035     MpegEncContext * const s = &h->s;
3036
3037     *structure = s->picture_structure;
3038     if(FIELD_PICTURE){
3039         if (!(pic_num & 1))
3040             /* opposite field */
3041             *structure ^= PICT_FRAME;
3042         pic_num >>= 1;
3043     }
3044
3045     return pic_num;
3046 }
3047
3048 static int decode_ref_pic_list_reordering(H264Context *h){
3049     MpegEncContext * const s = &h->s;
3050     int list, index, pic_structure;
3051
3052     print_short_term(h);
3053     print_long_term(h);
3054     if(h->slice_type==I_TYPE || h->slice_type==SI_TYPE) return 0; //FIXME move before func
3055
3056     for(list=0; list<h->list_count; list++){
3057         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
3058
3059         if(get_bits1(&s->gb)){
3060             int pred= h->curr_pic_num;
3061
3062             for(index=0; ; index++){
3063                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb(&s->gb);
3064                 unsigned int pic_id;
3065                 int i;
3066                 Picture *ref = NULL;
3067
3068                 if(reordering_of_pic_nums_idc==3)
3069                     break;
3070
3071                 if(index >= h->ref_count[list]){
3072                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
3073                     return -1;
3074                 }
3075
3076                 if(reordering_of_pic_nums_idc<3){
3077                     if(reordering_of_pic_nums_idc<2){
3078                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
3079                         int frame_num;
3080
3081                         if(abs_diff_pic_num > h->max_pic_num){
3082                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
3083                             return -1;
3084                         }
3085
3086                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
3087                         else                                pred+= abs_diff_pic_num;
3088                         pred &= h->max_pic_num - 1;
3089
3090                         frame_num = pic_num_extract(h, pred, &pic_structure);
3091
3092                         for(i= h->short_ref_count-1; i>=0; i--){
3093                             ref = h->short_ref[i];
3094                             assert(ref->reference);
3095                             assert(!ref->long_ref);
3096                             if(ref->data[0] != NULL &&
3097                                    ref->frame_num == frame_num &&
3098                                    (ref->reference & pic_structure) &&
3099                                    ref->long_ref == 0) // ignore non existing pictures by testing data[0] pointer
3100                                 break;
3101                         }
3102                         if(i>=0)
3103                             ref->pic_id= pred;
3104                     }else{
3105                         int long_idx;
3106                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
3107
3108                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
3109
3110                         if(long_idx>31){
3111                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
3112                             return -1;
3113                         }
3114                         ref = h->long_ref[long_idx];
3115                         assert(!(ref && !ref->reference));
3116                         if(ref && (ref->reference & pic_structure)){
3117                             ref->pic_id= pic_id;
3118                             assert(ref->long_ref);
3119                             i=0;
3120                         }else{
3121                             i=-1;
3122                         }
3123                     }
3124
3125                     if (i < 0) {
3126                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
3127                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
3128                     } else {
3129                         for(i=index; i+1<h->ref_count[list]; i++){
3130                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
3131                                 break;
3132                         }
3133                         for(; i > index; i--){
3134                             h->ref_list[list][i]= h->ref_list[list][i-1];
3135                         }
3136                         h->ref_list[list][index]= *ref;
3137                         if (FIELD_PICTURE){
3138                             pic_as_field(&h->ref_list[list][index], pic_structure);
3139                         }
3140                     }
3141                 }else{
3142                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
3143                     return -1;
3144                 }
3145             }
3146         }
3147     }
3148     for(list=0; list<h->list_count; list++){
3149         for(index= 0; index < h->ref_count[list]; index++){
3150             if(!h->ref_list[list][index].data[0])
3151                 h->ref_list[list][index]= s->current_picture;
3152         }
3153     }
3154
3155     if(h->slice_type==B_TYPE && !h->direct_spatial_mv_pred)
3156         direct_dist_scale_factor(h);
3157     direct_ref_list_init(h);
3158     return 0;
3159 }
3160
3161 static void fill_mbaff_ref_list(H264Context *h){
3162     int list, i, j;
3163     for(list=0; list<2; list++){ //FIXME try list_count
3164         for(i=0; i<h->ref_count[list]; i++){
3165             Picture *frame = &h->ref_list[list][i];
3166             Picture *field = &h->ref_list[list][16+2*i];
3167             field[0] = *frame;
3168             for(j=0; j<3; j++)
3169                 field[0].linesize[j] <<= 1;
3170             field[0].reference = PICT_TOP_FIELD;
3171             field[1] = field[0];
3172             for(j=0; j<3; j++)
3173                 field[1].data[j] += frame->linesize[j];
3174             field[1].reference = PICT_BOTTOM_FIELD;
3175
3176             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
3177             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
3178             for(j=0; j<2; j++){
3179                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
3180                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
3181             }
3182         }
3183     }
3184     for(j=0; j<h->ref_count[1]; j++){
3185         for(i=0; i<h->ref_count[0]; i++)
3186             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
3187         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
3188         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
3189     }
3190 }
3191
3192 static int pred_weight_table(H264Context *h){
3193     MpegEncContext * const s = &h->s;
3194     int list, i;
3195     int luma_def, chroma_def;
3196
3197     h->use_weight= 0;
3198     h->use_weight_chroma= 0;
3199     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
3200     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
3201     luma_def = 1<<h->luma_log2_weight_denom;
3202     chroma_def = 1<<h->chroma_log2_weight_denom;
3203
3204     for(list=0; list<2; list++){
3205         for(i=0; i<h->ref_count[list]; i++){
3206             int luma_weight_flag, chroma_weight_flag;
3207
3208             luma_weight_flag= get_bits1(&s->gb);
3209             if(luma_weight_flag){
3210                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3211                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3212                 if(   h->luma_weight[list][i] != luma_def
3213                    || h->luma_offset[list][i] != 0)
3214                     h->use_weight= 1;
3215             }else{
3216                 h->luma_weight[list][i]= luma_def;
3217                 h->luma_offset[list][i]= 0;
3218             }
3219
3220             chroma_weight_flag= get_bits1(&s->gb);
3221             if(chroma_weight_flag){
3222                 int j;
3223                 for(j=0; j<2; j++){
3224                     h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3225                     h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3226                     if(   h->chroma_weight[list][i][j] != chroma_def
3227                        || h->chroma_offset[list][i][j] != 0)
3228                         h->use_weight_chroma= 1;
3229                 }
3230             }else{
3231                 int j;
3232                 for(j=0; j<2; j++){
3233                     h->chroma_weight[list][i][j]= chroma_def;
3234                     h->chroma_offset[list][i][j]= 0;
3235                 }
3236             }
3237         }
3238         if(h->slice_type != B_TYPE) break;
3239     }
3240     h->use_weight= h->use_weight || h->use_weight_chroma;
3241     return 0;
3242 }
3243
3244 static void implicit_weight_table(H264Context *h){
3245     MpegEncContext * const s = &h->s;
3246     int ref0, ref1;
3247     int cur_poc = s->current_picture_ptr->poc;
3248
3249     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3250        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3251         h->use_weight= 0;
3252         h->use_weight_chroma= 0;
3253         return;
3254     }
3255
3256     h->use_weight= 2;
3257     h->use_weight_chroma= 2;
3258     h->luma_log2_weight_denom= 5;
3259     h->chroma_log2_weight_denom= 5;
3260
3261     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3262         int poc0 = h->ref_list[0][ref0].poc;
3263         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3264             int poc1 = h->ref_list[1][ref1].poc;
3265             int td = av_clip(poc1 - poc0, -128, 127);
3266             if(td){
3267                 int tb = av_clip(cur_poc - poc0, -128, 127);
3268                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3269                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3270                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3271                     h->implicit_weight[ref0][ref1] = 32;
3272                 else
3273                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3274             }else
3275                 h->implicit_weight[ref0][ref1] = 32;
3276         }
3277     }
3278 }
3279
3280 /**
3281  * Mark a picture as no longer needed for reference. The refmask
3282  * argument allows unreferencing of individual fields or the whole frame.
3283  * If the picture becomes entirely unreferenced, but is being held for
3284  * display purposes, it is marked as such.
3285  * @param refmask mask of fields to unreference; the mask is bitwise
3286  *                anded with the reference marking of pic
3287  * @return non-zero if pic becomes entirely unreferenced (except possibly
3288  *         for display purposes) zero if one of the fields remains in
3289  *         reference
3290  */
3291 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3292     int i;
3293     if (pic->reference &= refmask) {
3294         return 0;
3295     } else {
3296         if(pic == h->delayed_output_pic)
3297             pic->reference=DELAYED_PIC_REF;
3298         else{
3299             for(i = 0; h->delayed_pic[i]; i++)
3300                 if(pic == h->delayed_pic[i]){
3301                     pic->reference=DELAYED_PIC_REF;
3302                     break;
3303                 }
3304         }
3305         return 1;
3306     }
3307 }
3308
3309 /**
3310  * instantaneous decoder refresh.
3311  */
3312 static void idr(H264Context *h){
3313     int i;
3314
3315     for(i=0; i<16; i++){
3316         if (h->long_ref[i] != NULL) {
3317             unreference_pic(h, h->long_ref[i], 0);
3318             h->long_ref[i]= NULL;
3319         }
3320     }
3321     h->long_ref_count=0;
3322
3323     for(i=0; i<h->short_ref_count; i++){
3324         unreference_pic(h, h->short_ref[i], 0);
3325         h->short_ref[i]= NULL;
3326     }
3327     h->short_ref_count=0;
3328 }
3329
3330 /* forget old pics after a seek */
3331 static void flush_dpb(AVCodecContext *avctx){
3332     H264Context *h= avctx->priv_data;
3333     int i;
3334     for(i=0; i<16; i++) {
3335         if(h->delayed_pic[i])
3336             h->delayed_pic[i]->reference= 0;
3337         h->delayed_pic[i]= NULL;
3338     }
3339     if(h->delayed_output_pic)
3340         h->delayed_output_pic->reference= 0;
3341     h->delayed_output_pic= NULL;
3342     idr(h);
3343     if(h->s.current_picture_ptr)
3344         h->s.current_picture_ptr->reference= 0;
3345     h->s.first_field= 0;
3346     ff_mpeg_flush(avctx);
3347 }
3348
3349 /**
3350  * Find a Picture in the short term reference list by frame number.
3351  * @param frame_num frame number to search for
3352  * @param idx the index into h->short_ref where returned picture is found
3353  *            undefined if no picture found.
3354  * @return pointer to the found picture, or NULL if no pic with the provided
3355  *                 frame number is found
3356  */
3357 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3358     MpegEncContext * const s = &h->s;
3359     int i;
3360
3361     for(i=0; i<h->short_ref_count; i++){
3362         Picture *pic= h->short_ref[i];
3363         if(s->avctx->debug&FF_DEBUG_MMCO)
3364             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3365         if(pic->frame_num == frame_num) {
3366             *idx = i;
3367             return pic;
3368         }
3369     }
3370     return NULL;
3371 }
3372
3373 /**
3374  * Remove a picture from the short term reference list by its index in
3375  * that list.  This does no checking on the provided index; it is assumed
3376  * to be valid. Other list entries are shifted down.
3377  * @param i index into h->short_ref of picture to remove.
3378  */
3379 static void remove_short_at_index(H264Context *h, int i){
3380     assert(i > 0 && i < h->short_ref_count);
3381     h->short_ref[i]= NULL;
3382     if (--h->short_ref_count)
3383         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3384 }
3385
3386 /**
3387  *
3388  * @return the removed picture or NULL if an error occurs
3389  */
3390 static Picture * remove_short(H264Context *h, int frame_num){
3391     MpegEncContext * const s = &h->s;
3392     Picture *pic;
3393     int i;
3394
3395     if(s->avctx->debug&FF_DEBUG_MMCO)
3396         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3397
3398     pic = find_short(h, frame_num, &i);
3399     if (pic)
3400         remove_short_at_index(h, i);
3401
3402     return pic;
3403 }
3404
3405 /**
3406  * Remove a picture from the long term reference list by its index in
3407  * that list.  This does no checking on the provided index; it is assumed
3408  * to be valid. The removed entry is set to NULL. Other entries are unaffected.
3409  * @param i index into h->long_ref of picture to remove.
3410  */
3411 static void remove_long_at_index(H264Context *h, int i){
3412     h->long_ref[i]= NULL;
3413     h->long_ref_count--;
3414 }
3415
3416 /**
3417  *
3418  * @return the removed picture or NULL if an error occurs
3419  */
3420 static Picture * remove_long(H264Context *h, int i){
3421     Picture *pic;
3422
3423     pic= h->long_ref[i];
3424     if (pic)
3425         remove_long_at_index(h, i);
3426
3427     return pic;
3428 }
3429
3430 /**
3431  * print short term list
3432  */
3433 static void print_short_term(H264Context *h) {
3434     uint32_t i;
3435     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3436         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3437         for(i=0; i<h->sh