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