Do not scare the user with error messages when everything is ok.
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file h264.c
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "dsputil.h"
29 #include "avcodec.h"
30 #include "mpegvideo.h"
31 #include "h264.h"
32 #include "h264data.h"
33 #include "h264_parser.h"
34 #include "golomb.h"
35 #include "rectangle.h"
36
37 #include "cabac.h"
38 #ifdef ARCH_X86
39 #include "i386/h264_i386.h"
40 #endif
41
42 //#undef NDEBUG
43 #include <assert.h>
44
45 /**
46  * Value of Picture.reference when Picture is not a reference picture, but
47  * is held for delayed output.
48  */
49 #define DELAYED_PIC_REF 4
50
51 static VLC coeff_token_vlc[4];
52 static VLC chroma_dc_coeff_token_vlc;
53
54 static VLC total_zeros_vlc[15];
55 static VLC chroma_dc_total_zeros_vlc[3];
56
57 static VLC run_vlc[6];
58 static VLC run7_vlc;
59
60 static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
61 static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
62 static void filter_mb( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
63 static void filter_mb_fast( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
64 static Picture * remove_long(H264Context *h, int i, int ref_mask);
65
66 static av_always_inline uint32_t pack16to32(int a, int b){
67 #ifdef WORDS_BIGENDIAN
68    return (b&0xFFFF) + (a<<16);
69 #else
70    return (a&0xFFFF) + (b<<16);
71 #endif
72 }
73
74 const uint8_t ff_rem6[52]={
75 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,
76 };
77
78 const uint8_t ff_div6[52]={
79 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,
80 };
81
82
83 static void fill_caches(H264Context *h, int mb_type, int for_deblock){
84     MpegEncContext * const s = &h->s;
85     const int mb_xy= h->mb_xy;
86     int topleft_xy, top_xy, topright_xy, left_xy[2];
87     int topleft_type, top_type, topright_type, left_type[2];
88     int left_block[8];
89     int topleft_partition= -1;
90     int i;
91
92     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
93
94     //FIXME deblocking could skip the intra and nnz parts.
95     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
96         return;
97
98     /* Wow, what a mess, why didn't they simplify the interlacing & intra
99      * stuff, I can't imagine that these complex rules are worth it. */
100
101     topleft_xy = top_xy - 1;
102     topright_xy= top_xy + 1;
103     left_xy[1] = left_xy[0] = mb_xy-1;
104     left_block[0]= 0;
105     left_block[1]= 1;
106     left_block[2]= 2;
107     left_block[3]= 3;
108     left_block[4]= 7;
109     left_block[5]= 10;
110     left_block[6]= 8;
111     left_block[7]= 11;
112     if(FRAME_MBAFF){
113         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
114         const int top_pair_xy      = pair_xy     - s->mb_stride;
115         const int topleft_pair_xy  = top_pair_xy - 1;
116         const int topright_pair_xy = top_pair_xy + 1;
117         const int topleft_mb_frame_flag  = !IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
118         const int top_mb_frame_flag      = !IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
119         const int topright_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
120         const int left_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
121         const int curr_mb_frame_flag = !IS_INTERLACED(mb_type);
122         const int bottom = (s->mb_y & 1);
123         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);
124         if (bottom
125                 ? !curr_mb_frame_flag // bottom macroblock
126                 : (!curr_mb_frame_flag && !top_mb_frame_flag) // top macroblock
127                 ) {
128             top_xy -= s->mb_stride;
129         }
130         if (bottom
131                 ? !curr_mb_frame_flag // bottom macroblock
132                 : (!curr_mb_frame_flag && !topleft_mb_frame_flag) // top macroblock
133                 ) {
134             topleft_xy -= s->mb_stride;
135         } else if(bottom && curr_mb_frame_flag && !left_mb_frame_flag) {
136             topleft_xy += s->mb_stride;
137             // take top left mv from the middle of the mb, as opposed to all other modes which use the bottom right partition
138             topleft_partition = 0;
139         }
140         if (bottom
141                 ? !curr_mb_frame_flag // bottom macroblock
142                 : (!curr_mb_frame_flag && !topright_mb_frame_flag) // top macroblock
143                 ) {
144             topright_xy -= s->mb_stride;
145         }
146         if (left_mb_frame_flag != curr_mb_frame_flag) {
147             left_xy[1] = left_xy[0] = pair_xy - 1;
148             if (curr_mb_frame_flag) {
149                 if (bottom) {
150                     left_block[0]= 2;
151                     left_block[1]= 2;
152                     left_block[2]= 3;
153                     left_block[3]= 3;
154                     left_block[4]= 8;
155                     left_block[5]= 11;
156                     left_block[6]= 8;
157                     left_block[7]= 11;
158                 } else {
159                     left_block[0]= 0;
160                     left_block[1]= 0;
161                     left_block[2]= 1;
162                     left_block[3]= 1;
163                     left_block[4]= 7;
164                     left_block[5]= 10;
165                     left_block[6]= 7;
166                     left_block[7]= 10;
167                 }
168             } else {
169                 left_xy[1] += s->mb_stride;
170                 //left_block[0]= 0;
171                 left_block[1]= 2;
172                 left_block[2]= 0;
173                 left_block[3]= 2;
174                 //left_block[4]= 7;
175                 left_block[5]= 10;
176                 left_block[6]= 7;
177                 left_block[7]= 10;
178             }
179         }
180     }
181
182     h->top_mb_xy = top_xy;
183     h->left_mb_xy[0] = left_xy[0];
184     h->left_mb_xy[1] = left_xy[1];
185     if(for_deblock){
186         topleft_type = 0;
187         topright_type = 0;
188         top_type     = h->slice_table[top_xy     ] < 255 ? s->current_picture.mb_type[top_xy]     : 0;
189         left_type[0] = h->slice_table[left_xy[0] ] < 255 ? s->current_picture.mb_type[left_xy[0]] : 0;
190         left_type[1] = h->slice_table[left_xy[1] ] < 255 ? s->current_picture.mb_type[left_xy[1]] : 0;
191
192         if(FRAME_MBAFF && !IS_INTRA(mb_type)){
193             int list;
194             int v = *(uint16_t*)&h->non_zero_count[mb_xy][14];
195             for(i=0; i<16; i++)
196                 h->non_zero_count_cache[scan8[i]] = (v>>i)&1;
197             for(list=0; list<h->list_count; list++){
198                 if(USES_LIST(mb_type,list)){
199                     uint32_t *src = (uint32_t*)s->current_picture.motion_val[list][h->mb2b_xy[mb_xy]];
200                     uint32_t *dst = (uint32_t*)h->mv_cache[list][scan8[0]];
201                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
202                     for(i=0; i<4; i++, dst+=8, src+=h->b_stride){
203                         dst[0] = src[0];
204                         dst[1] = src[1];
205                         dst[2] = src[2];
206                         dst[3] = src[3];
207                     }
208                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
209                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = pack16to32(ref[0],ref[1])*0x0101;
210                     ref += h->b8_stride;
211                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
212                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = pack16to32(ref[0],ref[1])*0x0101;
213                 }else{
214                     fill_rectangle(&h-> mv_cache[list][scan8[ 0]], 4, 4, 8, 0, 4);
215                     fill_rectangle(&h->ref_cache[list][scan8[ 0]], 4, 4, 8, (uint8_t)LIST_NOT_USED, 1);
216                 }
217             }
218         }
219     }else{
220         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
221         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
222         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
223         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
224         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
225     }
226
227     if(IS_INTRA(mb_type)){
228         h->topleft_samples_available=
229         h->top_samples_available=
230         h->left_samples_available= 0xFFFF;
231         h->topright_samples_available= 0xEEEA;
232
233         if(!IS_INTRA(top_type) && (top_type==0 || h->pps.constrained_intra_pred)){
234             h->topleft_samples_available= 0xB3FF;
235             h->top_samples_available= 0x33FF;
236             h->topright_samples_available= 0x26EA;
237         }
238         for(i=0; i<2; i++){
239             if(!IS_INTRA(left_type[i]) && (left_type[i]==0 || h->pps.constrained_intra_pred)){
240                 h->topleft_samples_available&= 0xDF5F;
241                 h->left_samples_available&= 0x5F5F;
242             }
243         }
244
245         if(!IS_INTRA(topleft_type) && (topleft_type==0 || h->pps.constrained_intra_pred))
246             h->topleft_samples_available&= 0x7FFF;
247
248         if(!IS_INTRA(topright_type) && (topright_type==0 || h->pps.constrained_intra_pred))
249             h->topright_samples_available&= 0xFBFF;
250
251         if(IS_INTRA4x4(mb_type)){
252             if(IS_INTRA4x4(top_type)){
253                 h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
254                 h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
255                 h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
256                 h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
257             }else{
258                 int pred;
259                 if(!top_type || (IS_INTER(top_type) && h->pps.constrained_intra_pred))
260                     pred= -1;
261                 else{
262                     pred= 2;
263                 }
264                 h->intra4x4_pred_mode_cache[4+8*0]=
265                 h->intra4x4_pred_mode_cache[5+8*0]=
266                 h->intra4x4_pred_mode_cache[6+8*0]=
267                 h->intra4x4_pred_mode_cache[7+8*0]= pred;
268             }
269             for(i=0; i<2; i++){
270                 if(IS_INTRA4x4(left_type[i])){
271                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
272                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
273                 }else{
274                     int pred;
275                     if(!left_type[i] || (IS_INTER(left_type[i]) && h->pps.constrained_intra_pred))
276                         pred= -1;
277                     else{
278                         pred= 2;
279                     }
280                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
281                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
282                 }
283             }
284         }
285     }
286
287
288 /*
289 0 . T T. T T T T
290 1 L . .L . . . .
291 2 L . .L . . . .
292 3 . T TL . . . .
293 4 L . .L . . . .
294 5 L . .. . . . .
295 */
296 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
297     if(top_type){
298         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
299         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
300         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
301         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
302
303         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
304         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
305
306         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
307         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
308
309     }else{
310         h->non_zero_count_cache[4+8*0]=
311         h->non_zero_count_cache[5+8*0]=
312         h->non_zero_count_cache[6+8*0]=
313         h->non_zero_count_cache[7+8*0]=
314
315         h->non_zero_count_cache[1+8*0]=
316         h->non_zero_count_cache[2+8*0]=
317
318         h->non_zero_count_cache[1+8*3]=
319         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
320
321     }
322
323     for (i=0; i<2; i++) {
324         if(left_type[i]){
325             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
326             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
327             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
328             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
329         }else{
330             h->non_zero_count_cache[3+8*1 + 2*8*i]=
331             h->non_zero_count_cache[3+8*2 + 2*8*i]=
332             h->non_zero_count_cache[0+8*1 +   8*i]=
333             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
334         }
335     }
336
337     if( h->pps.cabac ) {
338         // top_cbp
339         if(top_type) {
340             h->top_cbp = h->cbp_table[top_xy];
341         } else if(IS_INTRA(mb_type)) {
342             h->top_cbp = 0x1C0;
343         } else {
344             h->top_cbp = 0;
345         }
346         // left_cbp
347         if (left_type[0]) {
348             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
349         } else if(IS_INTRA(mb_type)) {
350             h->left_cbp = 0x1C0;
351         } else {
352             h->left_cbp = 0;
353         }
354         if (left_type[0]) {
355             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
356         }
357         if (left_type[1]) {
358             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
359         }
360     }
361
362 #if 1
363     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
364         int list;
365         for(list=0; list<h->list_count; list++){
366             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
367                 /*if(!h->mv_cache_clean[list]){
368                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
369                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
370                     h->mv_cache_clean[list]= 1;
371                 }*/
372                 continue;
373             }
374             h->mv_cache_clean[list]= 0;
375
376             if(USES_LIST(top_type, list)){
377                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
378                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
379                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
380                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
381                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
382                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
383                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
384                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
385                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
386                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
387             }else{
388                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
389                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
390                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
391                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
392                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
393             }
394
395             for(i=0; i<2; i++){
396                 int cache_idx = scan8[0] - 1 + i*2*8;
397                 if(USES_LIST(left_type[i], list)){
398                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
399                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
400                     *(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]];
401                     *(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]];
402                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
403                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
404                 }else{
405                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
406                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
407                     h->ref_cache[list][cache_idx  ]=
408                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
409                 }
410             }
411
412             if((for_deblock || (IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred)) && !FRAME_MBAFF)
413                 continue;
414
415             if(USES_LIST(topleft_type, list)){
416                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
417                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
418                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
419                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
420             }else{
421                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
422                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
423             }
424
425             if(USES_LIST(topright_type, list)){
426                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
427                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
428                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
429                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
430             }else{
431                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
432                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
433             }
434
435             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
436                 continue;
437
438             h->ref_cache[list][scan8[5 ]+1] =
439             h->ref_cache[list][scan8[7 ]+1] =
440             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
441             h->ref_cache[list][scan8[4 ]] =
442             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
443             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
444             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
445             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
446             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
447             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
448
449             if( h->pps.cabac ) {
450                 /* XXX beurk, Load mvd */
451                 if(USES_LIST(top_type, list)){
452                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
453                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
454                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
455                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
456                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
457                 }else{
458                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
459                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
460                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
461                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
462                 }
463                 if(USES_LIST(left_type[0], list)){
464                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
465                     *(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]];
466                     *(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]];
467                 }else{
468                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
469                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
470                 }
471                 if(USES_LIST(left_type[1], list)){
472                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
473                     *(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]];
474                     *(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]];
475                 }else{
476                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
477                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
478                 }
479                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
480                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
481                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
482                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
483                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
484
485                 if(h->slice_type_nos == FF_B_TYPE){
486                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
487
488                     if(IS_DIRECT(top_type)){
489                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
490                     }else if(IS_8X8(top_type)){
491                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
492                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
493                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
494                     }else{
495                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
496                     }
497
498                     if(IS_DIRECT(left_type[0]))
499                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
500                     else if(IS_8X8(left_type[0]))
501                         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)];
502                     else
503                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
504
505                     if(IS_DIRECT(left_type[1]))
506                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
507                     else if(IS_8X8(left_type[1]))
508                         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)];
509                     else
510                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
511                 }
512             }
513
514             if(FRAME_MBAFF){
515 #define MAP_MVS\
516                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
517                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
518                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
519                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
520                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
521                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
522                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
523                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
524                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
525                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
526                 if(MB_FIELD){
527 #define MAP_F2F(idx, mb_type)\
528                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
529                         h->ref_cache[list][idx] <<= 1;\
530                         h->mv_cache[list][idx][1] /= 2;\
531                         h->mvd_cache[list][idx][1] /= 2;\
532                     }
533                     MAP_MVS
534 #undef MAP_F2F
535                 }else{
536 #define MAP_F2F(idx, mb_type)\
537                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
538                         h->ref_cache[list][idx] >>= 1;\
539                         h->mv_cache[list][idx][1] <<= 1;\
540                         h->mvd_cache[list][idx][1] <<= 1;\
541                     }
542                     MAP_MVS
543 #undef MAP_F2F
544                 }
545             }
546         }
547     }
548 #endif
549
550     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
551 }
552
553 static inline void write_back_intra_pred_mode(H264Context *h){
554     const int mb_xy= h->mb_xy;
555
556     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
557     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
558     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
559     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
560     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
561     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
562     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
563 }
564
565 /**
566  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
567  */
568 static inline int check_intra4x4_pred_mode(H264Context *h){
569     MpegEncContext * const s = &h->s;
570     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
571     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
572     int i;
573
574     if(!(h->top_samples_available&0x8000)){
575         for(i=0; i<4; i++){
576             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
577             if(status<0){
578                 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);
579                 return -1;
580             } else if(status){
581                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
582             }
583         }
584     }
585
586     if(!(h->left_samples_available&0x8000)){
587         for(i=0; i<4; i++){
588             int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
589             if(status<0){
590                 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);
591                 return -1;
592             } else if(status){
593                 h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
594             }
595         }
596     }
597
598     return 0;
599 } //FIXME cleanup like next
600
601 /**
602  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
603  */
604 static inline int check_intra_pred_mode(H264Context *h, int mode){
605     MpegEncContext * const s = &h->s;
606     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
607     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
608
609     if(mode > 6U) {
610         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);
611         return -1;
612     }
613
614     if(!(h->top_samples_available&0x8000)){
615         mode= top[ mode ];
616         if(mode<0){
617             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);
618             return -1;
619         }
620     }
621
622     if(!(h->left_samples_available&0x8000)){
623         mode= left[ mode ];
624         if(mode<0){
625             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);
626             return -1;
627         }
628     }
629
630     return mode;
631 }
632
633 /**
634  * gets the predicted intra4x4 prediction mode.
635  */
636 static inline int pred_intra_mode(H264Context *h, int n){
637     const int index8= scan8[n];
638     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
639     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
640     const int min= FFMIN(left, top);
641
642     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
643
644     if(min<0) return DC_PRED;
645     else      return min;
646 }
647
648 static inline void write_back_non_zero_count(H264Context *h){
649     const int mb_xy= h->mb_xy;
650
651     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
652     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
653     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
654     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
655     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
656     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
657     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
658
659     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
660     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
661     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
662
663     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
664     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
665     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
666
667     if(FRAME_MBAFF){
668         // store all luma nnzs, for deblocking
669         int v = 0, i;
670         for(i=0; i<16; i++)
671             v += (!!h->non_zero_count_cache[scan8[i]]) << i;
672         *(uint16_t*)&h->non_zero_count[mb_xy][14] = v;
673     }
674 }
675
676 /**
677  * gets the predicted number of non-zero coefficients.
678  * @param n block index
679  */
680 static inline int pred_non_zero_count(H264Context *h, int n){
681     const int index8= scan8[n];
682     const int left= h->non_zero_count_cache[index8 - 1];
683     const int top = h->non_zero_count_cache[index8 - 8];
684     int i= left + top;
685
686     if(i<64) i= (i+1)>>1;
687
688     tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);
689
690     return i&31;
691 }
692
693 static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
694     const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
695     MpegEncContext *s = &h->s;
696
697     /* there is no consistent mapping of mvs to neighboring locations that will
698      * make mbaff happy, so we can't move all this logic to fill_caches */
699     if(FRAME_MBAFF){
700         const uint32_t *mb_types = s->current_picture_ptr->mb_type;
701         const int16_t *mv;
702         *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
703         *C = h->mv_cache[list][scan8[0]-2];
704
705         if(!MB_FIELD
706            && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
707             int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
708             if(IS_INTERLACED(mb_types[topright_xy])){
709 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
710                 const int x4 = X4, y4 = Y4;\
711                 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
712                 if(!USES_LIST(mb_type,list))\
713                     return LIST_NOT_USED;\
714                 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
715                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
716                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
717                 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
718
719                 SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
720             }
721         }
722         if(topright_ref == PART_NOT_AVAILABLE
723            && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
724            && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
725             if(!MB_FIELD
726                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
727                 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
728             }
729             if(MB_FIELD
730                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
731                && i >= scan8[0]+8){
732                 // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
733                 SET_DIAG_MV(/2, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
734             }
735         }
736 #undef SET_DIAG_MV
737     }
738
739     if(topright_ref != PART_NOT_AVAILABLE){
740         *C= h->mv_cache[list][ i - 8 + part_width ];
741         return topright_ref;
742     }else{
743         tprintf(s->avctx, "topright MV not available\n");
744
745         *C= h->mv_cache[list][ i - 8 - 1 ];
746         return h->ref_cache[list][ i - 8 - 1 ];
747     }
748 }
749
750 /**
751  * gets the predicted MV.
752  * @param n the block index
753  * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
754  * @param mx the x component of the predicted motion vector
755  * @param my the y component of the predicted motion vector
756  */
757 static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
758     const int index8= scan8[n];
759     const int top_ref=      h->ref_cache[list][ index8 - 8 ];
760     const int left_ref=     h->ref_cache[list][ index8 - 1 ];
761     const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
762     const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
763     const int16_t * C;
764     int diagonal_ref, match_count;
765
766     assert(part_width==1 || part_width==2 || part_width==4);
767
768 /* mv_cache
769   B . . A T T T T
770   U . . L . . , .
771   U . . L . . . .
772   U . . L . . , .
773   . . . L . . . .
774 */
775
776     diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
777     match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
778     tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
779     if(match_count > 1){ //most common
780         *mx= mid_pred(A[0], B[0], C[0]);
781         *my= mid_pred(A[1], B[1], C[1]);
782     }else if(match_count==1){
783         if(left_ref==ref){
784             *mx= A[0];
785             *my= A[1];
786         }else if(top_ref==ref){
787             *mx= B[0];
788             *my= B[1];
789         }else{
790             *mx= C[0];
791             *my= C[1];
792         }
793     }else{
794         if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
795             *mx= A[0];
796             *my= A[1];
797         }else{
798             *mx= mid_pred(A[0], B[0], C[0]);
799             *my= mid_pred(A[1], B[1], C[1]);
800         }
801     }
802
803     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);
804 }
805
806 /**
807  * gets the directionally predicted 16x8 MV.
808  * @param n the block index
809  * @param mx the x component of the predicted motion vector
810  * @param my the y component of the predicted motion vector
811  */
812 static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
813     if(n==0){
814         const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
815         const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
816
817         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);
818
819         if(top_ref == ref){
820             *mx= B[0];
821             *my= B[1];
822             return;
823         }
824     }else{
825         const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
826         const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];
827
828         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);
829
830         if(left_ref == ref){
831             *mx= A[0];
832             *my= A[1];
833             return;
834         }
835     }
836
837     //RARE
838     pred_motion(h, n, 4, list, ref, mx, my);
839 }
840
841 /**
842  * gets the directionally predicted 8x16 MV.
843  * @param n the block index
844  * @param mx the x component of the predicted motion vector
845  * @param my the y component of the predicted motion vector
846  */
847 static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
848     if(n==0){
849         const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
850         const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];
851
852         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);
853
854         if(left_ref == ref){
855             *mx= A[0];
856             *my= A[1];
857             return;
858         }
859     }else{
860         const int16_t * C;
861         int diagonal_ref;
862
863         diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);
864
865         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);
866
867         if(diagonal_ref == ref){
868             *mx= C[0];
869             *my= C[1];
870             return;
871         }
872     }
873
874     //RARE
875     pred_motion(h, n, 2, list, ref, mx, my);
876 }
877
878 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
879     const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
880     const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];
881
882     tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);
883
884     if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
885        || (top_ref == 0  && *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ] == 0)
886        || (left_ref == 0 && *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ] == 0)){
887
888         *mx = *my = 0;
889         return;
890     }
891
892     pred_motion(h, 0, 4, 0, 0, mx, my);
893
894     return;
895 }
896
897 static inline void direct_dist_scale_factor(H264Context * const h){
898     const int poc = h->s.current_picture_ptr->poc;
899     const int poc1 = h->ref_list[1][0].poc;
900     int i;
901     for(i=0; i<h->ref_count[0]; i++){
902         int poc0 = h->ref_list[0][i].poc;
903         int td = av_clip(poc1 - poc0, -128, 127);
904         if(td == 0 /* FIXME || pic0 is a long-term ref */){
905             h->dist_scale_factor[i] = 256;
906         }else{
907             int tb = av_clip(poc - poc0, -128, 127);
908             int tx = (16384 + (FFABS(td) >> 1)) / td;
909             h->dist_scale_factor[i] = av_clip((tb*tx + 32) >> 6, -1024, 1023);
910         }
911     }
912     if(FRAME_MBAFF){
913         for(i=0; i<h->ref_count[0]; i++){
914             h->dist_scale_factor_field[2*i] =
915             h->dist_scale_factor_field[2*i+1] = h->dist_scale_factor[i];
916         }
917     }
918 }
919 static inline void direct_ref_list_init(H264Context * const h){
920     MpegEncContext * const s = &h->s;
921     Picture * const ref1 = &h->ref_list[1][0];
922     Picture * const cur = s->current_picture_ptr;
923     int list, i, j;
924     if(cur->pict_type == FF_I_TYPE)
925         cur->ref_count[0] = 0;
926     if(cur->pict_type != FF_B_TYPE)
927         cur->ref_count[1] = 0;
928     for(list=0; list<2; list++){
929         cur->ref_count[list] = h->ref_count[list];
930         for(j=0; j<h->ref_count[list]; j++)
931             cur->ref_poc[list][j] = h->ref_list[list][j].poc;
932     }
933     if(cur->pict_type != FF_B_TYPE || h->direct_spatial_mv_pred)
934         return;
935     for(list=0; list<2; list++){
936         for(i=0; i<ref1->ref_count[list]; i++){
937             const int poc = ref1->ref_poc[list][i];
938             h->map_col_to_list0[list][i] = 0; /* bogus; fills in for missing frames */
939             for(j=0; j<h->ref_count[list]; j++)
940                 if(h->ref_list[list][j].poc == poc){
941                     h->map_col_to_list0[list][i] = j;
942                     break;
943                 }
944         }
945     }
946     if(FRAME_MBAFF){
947         for(list=0; list<2; list++){
948             for(i=0; i<ref1->ref_count[list]; i++){
949                 j = h->map_col_to_list0[list][i];
950                 h->map_col_to_list0_field[list][2*i] = 2*j;
951                 h->map_col_to_list0_field[list][2*i+1] = 2*j+1;
952             }
953         }
954     }
955 }
956
957 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
958     MpegEncContext * const s = &h->s;
959     const int mb_xy =   h->mb_xy;
960     const int b8_xy = 2*s->mb_x + 2*s->mb_y*h->b8_stride;
961     const int b4_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
962     const int mb_type_col = h->ref_list[1][0].mb_type[mb_xy];
963     const int16_t (*l1mv0)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[0][b4_xy];
964     const int16_t (*l1mv1)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[1][b4_xy];
965     const int8_t *l1ref0 = &h->ref_list[1][0].ref_index[0][b8_xy];
966     const int8_t *l1ref1 = &h->ref_list[1][0].ref_index[1][b8_xy];
967     const int is_b8x8 = IS_8X8(*mb_type);
968     unsigned int sub_mb_type;
969     int i8, i4;
970
971 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
972     if(IS_8X8(mb_type_col) && !h->sps.direct_8x8_inference_flag){
973         /* FIXME save sub mb types from previous frames (or derive from MVs)
974          * so we know exactly what block size to use */
975         sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
976         *mb_type =    MB_TYPE_8x8|MB_TYPE_L0L1;
977     }else if(!is_b8x8 && (mb_type_col & MB_TYPE_16x16_OR_INTRA)){
978         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
979         *mb_type =    MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
980     }else{
981         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
982         *mb_type =    MB_TYPE_8x8|MB_TYPE_L0L1;
983     }
984     if(!is_b8x8)
985         *mb_type |= MB_TYPE_DIRECT2;
986     if(MB_FIELD)
987         *mb_type |= MB_TYPE_INTERLACED;
988
989     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);
990
991     if(h->direct_spatial_mv_pred){
992         int ref[2];
993         int mv[2][2];
994         int list;
995
996         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
997
998         /* ref = min(neighbors) */
999         for(list=0; list<2; list++){
1000             int refa = h->ref_cache[list][scan8[0] - 1];
1001             int refb = h->ref_cache[list][scan8[0] - 8];
1002             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1003             if(refc == -2)
1004                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1005             ref[list] = FFMIN3((unsigned)refa, (unsigned)refb, (unsigned)refc);
1006             if(ref[list] < 0)
1007                 ref[list] = -1;
1008         }
1009
1010         if(ref[0] < 0 && ref[1] < 0){
1011             ref[0] = ref[1] = 0;
1012             mv[0][0] = mv[0][1] =
1013             mv[1][0] = mv[1][1] = 0;
1014         }else{
1015             for(list=0; list<2; list++){
1016                 if(ref[list] >= 0)
1017                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1018                 else
1019                     mv[list][0] = mv[list][1] = 0;
1020             }
1021         }
1022
1023         if(ref[1] < 0){
1024             if(!is_b8x8)
1025                 *mb_type &= ~MB_TYPE_L1;
1026             sub_mb_type &= ~MB_TYPE_L1;
1027         }else if(ref[0] < 0){
1028             if(!is_b8x8)
1029                 *mb_type &= ~MB_TYPE_L0;
1030             sub_mb_type &= ~MB_TYPE_L0;
1031         }
1032
1033         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col)){
1034             int pair_xy = s->mb_x + (s->mb_y&~1)*s->mb_stride;
1035             int mb_types_col[2];
1036             int b8_stride = h->b8_stride;
1037             int b4_stride = h->b_stride;
1038
1039             *mb_type = (*mb_type & ~MB_TYPE_16x16) | MB_TYPE_8x8;
1040
1041             if(IS_INTERLACED(*mb_type)){
1042                 mb_types_col[0] = h->ref_list[1][0].mb_type[pair_xy];
1043                 mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1044                 if(s->mb_y&1){
1045                     l1ref0 -= 2*b8_stride;
1046                     l1ref1 -= 2*b8_stride;
1047                     l1mv0 -= 4*b4_stride;
1048                     l1mv1 -= 4*b4_stride;
1049                 }
1050                 b8_stride *= 3;
1051                 b4_stride *= 6;
1052             }else{
1053                 int cur_poc = s->current_picture_ptr->poc;
1054                 int *col_poc = h->ref_list[1]->field_poc;
1055                 int col_parity = FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc);
1056                 int dy = 2*col_parity - (s->mb_y&1);
1057                 mb_types_col[0] =
1058                 mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy + col_parity*s->mb_stride];
1059                 l1ref0 += dy*b8_stride;
1060                 l1ref1 += dy*b8_stride;
1061                 l1mv0 += 2*dy*b4_stride;
1062                 l1mv1 += 2*dy*b4_stride;
1063                 b8_stride = 0;
1064             }
1065
1066             for(i8=0; i8<4; i8++){
1067                 int x8 = i8&1;
1068                 int y8 = i8>>1;
1069                 int xy8 = x8+y8*b8_stride;
1070                 int xy4 = 3*x8+y8*b4_stride;
1071                 int a=0, b=0;
1072
1073                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1074                     continue;
1075                 h->sub_mb_type[i8] = sub_mb_type;
1076
1077                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1078                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1079                 if(!IS_INTRA(mb_types_col[y8])
1080                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
1081                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
1082                     if(ref[0] > 0)
1083                         a= pack16to32(mv[0][0],mv[0][1]);
1084                     if(ref[1] > 0)
1085                         b= pack16to32(mv[1][0],mv[1][1]);
1086                 }else{
1087                     a= pack16to32(mv[0][0],mv[0][1]);
1088                     b= pack16to32(mv[1][0],mv[1][1]);
1089                 }
1090                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
1091                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
1092             }
1093         }else if(IS_16X16(*mb_type)){
1094             int a=0, b=0;
1095
1096             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1097             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1098             if(!IS_INTRA(mb_type_col)
1099                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1100                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1101                        && (h->x264_build>33 || !h->x264_build)))){
1102                 if(ref[0] > 0)
1103                     a= pack16to32(mv[0][0],mv[0][1]);
1104                 if(ref[1] > 0)
1105                     b= pack16to32(mv[1][0],mv[1][1]);
1106             }else{
1107                 a= pack16to32(mv[0][0],mv[0][1]);
1108                 b= pack16to32(mv[1][0],mv[1][1]);
1109             }
1110             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1111             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1112         }else{
1113             for(i8=0; i8<4; i8++){
1114                 const int x8 = i8&1;
1115                 const int y8 = i8>>1;
1116
1117                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1118                     continue;
1119                 h->sub_mb_type[i8] = sub_mb_type;
1120
1121                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1122                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1123                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1124                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1125
1126                 /* col_zero_flag */
1127                 if(!IS_INTRA(mb_type_col) && (   l1ref0[x8 + y8*h->b8_stride] == 0
1128                                               || (l1ref0[x8 + y8*h->b8_stride] < 0 && l1ref1[x8 + y8*h->b8_stride] == 0
1129                                                   && (h->x264_build>33 || !h->x264_build)))){
1130                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*h->b8_stride] == 0 ? l1mv0 : l1mv1;
1131                     if(IS_SUB_8X8(sub_mb_type)){
1132                         const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1133                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1134                             if(ref[0] == 0)
1135                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1136                             if(ref[1] == 0)
1137                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1138                         }
1139                     }else
1140                     for(i4=0; i4<4; i4++){
1141                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1142                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1143                             if(ref[0] == 0)
1144                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1145                             if(ref[1] == 0)
1146                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1147                         }
1148                     }
1149                 }
1150             }
1151         }
1152     }else{ /* direct temporal mv pred */
1153         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1154         const int *dist_scale_factor = h->dist_scale_factor;
1155
1156         if(FRAME_MBAFF){
1157             if(IS_INTERLACED(*mb_type)){
1158                 map_col_to_list0[0] = h->map_col_to_list0_field[0];
1159                 map_col_to_list0[1] = h->map_col_to_list0_field[1];
1160                 dist_scale_factor = h->dist_scale_factor_field;
1161             }
1162             if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col)){
1163                 /* FIXME assumes direct_8x8_inference == 1 */
1164                 const int pair_xy = s->mb_x + (s->mb_y&~1)*s->mb_stride;
1165                 int mb_types_col[2];
1166                 int y_shift;
1167
1168                 *mb_type = MB_TYPE_8x8|MB_TYPE_L0L1
1169                          | (is_b8x8 ? 0 : MB_TYPE_DIRECT2)
1170                          | (*mb_type & MB_TYPE_INTERLACED);
1171                 sub_mb_type = MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2|MB_TYPE_16x16;
1172
1173                 if(IS_INTERLACED(*mb_type)){
1174                     /* frame to field scaling */
1175                     mb_types_col[0] = h->ref_list[1][0].mb_type[pair_xy];
1176                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1177                     if(s->mb_y&1){
1178                         l1ref0 -= 2*h->b8_stride;
1179                         l1ref1 -= 2*h->b8_stride;
1180                         l1mv0 -= 4*h->b_stride;
1181                         l1mv1 -= 4*h->b_stride;
1182                     }
1183                     y_shift = 0;
1184
1185                     if(   (mb_types_col[0] & MB_TYPE_16x16_OR_INTRA)
1186                        && (mb_types_col[1] & MB_TYPE_16x16_OR_INTRA)
1187                        && !is_b8x8)
1188                         *mb_type |= MB_TYPE_16x8;
1189                     else
1190                         *mb_type |= MB_TYPE_8x8;
1191                 }else{
1192                     /* field to frame scaling */
1193                     /* col_mb_y = (mb_y&~1) + (topAbsDiffPOC < bottomAbsDiffPOC ? 0 : 1)
1194                      * but in MBAFF, top and bottom POC are equal */
1195                     int dy = (s->mb_y&1) ? 1 : 2;
1196                     mb_types_col[0] =
1197                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1198                     l1ref0 += dy*h->b8_stride;
1199                     l1ref1 += dy*h->b8_stride;
1200                     l1mv0 += 2*dy*h->b_stride;
1201                     l1mv1 += 2*dy*h->b_stride;
1202                     y_shift = 2;
1203
1204                     if((mb_types_col[0] & (MB_TYPE_16x16_OR_INTRA|MB_TYPE_16x8))
1205                        && !is_b8x8)
1206                         *mb_type |= MB_TYPE_16x16;
1207                     else
1208                         *mb_type |= MB_TYPE_8x8;
1209                 }
1210
1211                 for(i8=0; i8<4; i8++){
1212                     const int x8 = i8&1;
1213                     const int y8 = i8>>1;
1214                     int ref0, scale;
1215                     const int16_t (*l1mv)[2]= l1mv0;
1216
1217                     if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1218                         continue;
1219                     h->sub_mb_type[i8] = sub_mb_type;
1220
1221                     fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1222                     if(IS_INTRA(mb_types_col[y8])){
1223                         fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1224                         fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1225                         fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1226                         continue;
1227                     }
1228
1229                     ref0 = l1ref0[x8 + (y8*2>>y_shift)*h->b8_stride];
1230                     if(ref0 >= 0)
1231                         ref0 = map_col_to_list0[0][ref0*2>>y_shift];
1232                     else{
1233                         ref0 = map_col_to_list0[1][l1ref1[x8 + (y8*2>>y_shift)*h->b8_stride]*2>>y_shift];
1234                         l1mv= l1mv1;
1235                     }
1236                     scale = dist_scale_factor[ref0];
1237                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1238
1239                     {
1240                         const int16_t *mv_col = l1mv[x8*3 + (y8*6>>y_shift)*h->b_stride];
1241                         int my_col = (mv_col[1]<<y_shift)/2;
1242                         int mx = (scale * mv_col[0] + 128) >> 8;
1243                         int my = (scale * my_col + 128) >> 8;
1244                         fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1245                         fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1246                     }
1247                 }
1248                 return;
1249             }
1250         }
1251
1252         /* one-to-one mv scaling */
1253
1254         if(IS_16X16(*mb_type)){
1255             int ref, mv0, mv1;
1256
1257             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1258             if(IS_INTRA(mb_type_col)){
1259                 ref=mv0=mv1=0;
1260             }else{
1261                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0]]
1262                                                 : map_col_to_list0[1][l1ref1[0]];
1263                 const int scale = dist_scale_factor[ref0];
1264                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1265                 int mv_l0[2];
1266                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1267                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1268                 ref= ref0;
1269                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1270                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1271             }
1272             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1273             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1274             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1275         }else{
1276             for(i8=0; i8<4; i8++){
1277                 const int x8 = i8&1;
1278                 const int y8 = i8>>1;
1279                 int ref0, scale;
1280                 const int16_t (*l1mv)[2]= l1mv0;
1281
1282                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1283                     continue;
1284                 h->sub_mb_type[i8] = sub_mb_type;
1285                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1286                 if(IS_INTRA(mb_type_col)){
1287                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1288                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1289                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1290                     continue;
1291                 }
1292
1293                 ref0 = l1ref0[x8 + y8*h->b8_stride];
1294                 if(ref0 >= 0)
1295                     ref0 = map_col_to_list0[0][ref0];
1296                 else{
1297                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*h->b8_stride]];
1298                     l1mv= l1mv1;
1299                 }
1300                 scale = dist_scale_factor[ref0];
1301
1302                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1303                 if(IS_SUB_8X8(sub_mb_type)){
1304                     const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1305                     int mx = (scale * mv_col[0] + 128) >> 8;
1306                     int my = (scale * mv_col[1] + 128) >> 8;
1307                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1308                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1309                 }else
1310                 for(i4=0; i4<4; i4++){
1311                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1312                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1313                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1314                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1315                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1316                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1317                 }
1318             }
1319         }
1320     }
1321 }
1322
1323 static inline void write_back_motion(H264Context *h, int mb_type){
1324     MpegEncContext * const s = &h->s;
1325     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1326     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1327     int list;
1328
1329     if(!USES_LIST(mb_type, 0))
1330         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1331
1332     for(list=0; list<h->list_count; list++){
1333         int y;
1334         if(!USES_LIST(mb_type, list))
1335             continue;
1336
1337         for(y=0; y<4; y++){
1338             *(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];
1339             *(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];
1340         }
1341         if( h->pps.cabac ) {
1342             if(IS_SKIP(mb_type))
1343                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1344             else
1345             for(y=0; y<4; y++){
1346                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1347                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1348             }
1349         }
1350
1351         {
1352             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1353             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1354             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1355             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1356             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1357         }
1358     }
1359
1360     if(h->slice_type_nos == FF_B_TYPE && h->pps.cabac){
1361         if(IS_8X8(mb_type)){
1362             uint8_t *direct_table = &h->direct_table[b8_xy];
1363             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1364             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1365             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1366         }
1367     }
1368 }
1369
1370 /**
1371  * Decodes a network abstraction layer unit.
1372  * @param consumed is the number of bytes used as input
1373  * @param length is the length of the array
1374  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1375  * @returns decoded bytes, might be src+1 if no escapes
1376  */
1377 static const uint8_t *decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
1378     int i, si, di;
1379     uint8_t *dst;
1380     int bufidx;
1381
1382 //    src[0]&0x80;                //forbidden bit
1383     h->nal_ref_idc= src[0]>>5;
1384     h->nal_unit_type= src[0]&0x1F;
1385
1386     src++; length--;
1387 #if 0
1388     for(i=0; i<length; i++)
1389         printf("%2X ", src[i]);
1390 #endif
1391     for(i=0; i+1<length; i+=2){
1392         if(src[i]) continue;
1393         if(i>0 && src[i-1]==0) i--;
1394         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1395             if(src[i+2]!=3){
1396                 /* startcode, so we must be past the end */
1397                 length=i;
1398             }
1399             break;
1400         }
1401     }
1402
1403     if(i>=length-1){ //no escaped 0
1404         *dst_length= length;
1405         *consumed= length+1; //+1 for the header
1406         return src;
1407     }
1408
1409     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1410     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length);
1411     dst= h->rbsp_buffer[bufidx];
1412
1413     if (dst == NULL){
1414         return NULL;
1415     }
1416
1417 //printf("decoding esc\n");
1418     si=di=0;
1419     while(si<length){
1420         //remove escapes (very rare 1:2^22)
1421         if(si+2<length && src[si]==0 && src[si+1]==0 && src[si+2]<=3){
1422             if(src[si+2]==3){ //escape
1423                 dst[di++]= 0;
1424                 dst[di++]= 0;
1425                 si+=3;
1426                 continue;
1427             }else //next start code
1428                 break;
1429         }
1430
1431         dst[di++]= src[si++];
1432     }
1433
1434     *dst_length= di;
1435     *consumed= si + 1;//+1 for the header
1436 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1437     return dst;
1438 }
1439
1440 /**
1441  * identifies the exact end of the bitstream
1442  * @return the length of the trailing, or 0 if damaged
1443  */
1444 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src){
1445     int v= *src;
1446     int r;
1447
1448     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1449
1450     for(r=1; r<9; r++){
1451         if(v&1) return r;
1452         v>>=1;
1453     }
1454     return 0;
1455 }
1456
1457 /**
1458  * IDCT transforms the 16 dc values and dequantizes them.
1459  * @param qp quantization parameter
1460  */
1461 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1462 #define stride 16
1463     int i;
1464     int temp[16]; //FIXME check if this is a good idea
1465     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1466     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1467
1468 //memset(block, 64, 2*256);
1469 //return;
1470     for(i=0; i<4; i++){
1471         const int offset= y_offset[i];
1472         const int z0= block[offset+stride*0] + block[offset+stride*4];
1473         const int z1= block[offset+stride*0] - block[offset+stride*4];
1474         const int z2= block[offset+stride*1] - block[offset+stride*5];
1475         const int z3= block[offset+stride*1] + block[offset+stride*5];
1476
1477         temp[4*i+0]= z0+z3;
1478         temp[4*i+1]= z1+z2;
1479         temp[4*i+2]= z1-z2;
1480         temp[4*i+3]= z0-z3;
1481     }
1482
1483     for(i=0; i<4; i++){
1484         const int offset= x_offset[i];
1485         const int z0= temp[4*0+i] + temp[4*2+i];
1486         const int z1= temp[4*0+i] - temp[4*2+i];
1487         const int z2= temp[4*1+i] - temp[4*3+i];
1488         const int z3= temp[4*1+i] + temp[4*3+i];
1489
1490         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
1491         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1492         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1493         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1494     }
1495 }
1496
1497 #if 0
1498 /**
1499  * DCT transforms the 16 dc values.
1500  * @param qp quantization parameter ??? FIXME
1501  */
1502 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1503 //    const int qmul= dequant_coeff[qp][0];
1504     int i;
1505     int temp[16]; //FIXME check if this is a good idea
1506     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1507     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1508
1509     for(i=0; i<4; i++){
1510         const int offset= y_offset[i];
1511         const int z0= block[offset+stride*0] + block[offset+stride*4];
1512         const int z1= block[offset+stride*0] - block[offset+stride*4];
1513         const int z2= block[offset+stride*1] - block[offset+stride*5];
1514         const int z3= block[offset+stride*1] + block[offset+stride*5];
1515
1516         temp[4*i+0]= z0+z3;
1517         temp[4*i+1]= z1+z2;
1518         temp[4*i+2]= z1-z2;
1519         temp[4*i+3]= z0-z3;
1520     }
1521
1522     for(i=0; i<4; i++){
1523         const int offset= x_offset[i];
1524         const int z0= temp[4*0+i] + temp[4*2+i];
1525         const int z1= temp[4*0+i] - temp[4*2+i];
1526         const int z2= temp[4*1+i] - temp[4*3+i];
1527         const int z3= temp[4*1+i] + temp[4*3+i];
1528
1529         block[stride*0 +offset]= (z0 + z3)>>1;
1530         block[stride*2 +offset]= (z1 + z2)>>1;
1531         block[stride*8 +offset]= (z1 - z2)>>1;
1532         block[stride*10+offset]= (z0 - z3)>>1;
1533     }
1534 }
1535 #endif
1536
1537 #undef xStride
1538 #undef stride
1539
1540 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1541     const int stride= 16*2;
1542     const int xStride= 16;
1543     int a,b,c,d,e;
1544
1545     a= block[stride*0 + xStride*0];
1546     b= block[stride*0 + xStride*1];
1547     c= block[stride*1 + xStride*0];
1548     d= block[stride*1 + xStride*1];
1549
1550     e= a-b;
1551     a= a+b;
1552     b= c-d;
1553     c= c+d;
1554
1555     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1556     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1557     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1558     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1559 }
1560
1561 #if 0
1562 static void chroma_dc_dct_c(DCTELEM *block){
1563     const int stride= 16*2;
1564     const int xStride= 16;
1565     int a,b,c,d,e;
1566
1567     a= block[stride*0 + xStride*0];
1568     b= block[stride*0 + xStride*1];
1569     c= block[stride*1 + xStride*0];
1570     d= block[stride*1 + xStride*1];
1571
1572     e= a-b;
1573     a= a+b;
1574     b= c-d;
1575     c= c+d;
1576
1577     block[stride*0 + xStride*0]= (a+c);
1578     block[stride*0 + xStride*1]= (e+b);
1579     block[stride*1 + xStride*0]= (a-c);
1580     block[stride*1 + xStride*1]= (e-b);
1581 }
1582 #endif
1583
1584 /**
1585  * gets the chroma qp.
1586  */
1587 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1588     return h->pps.chroma_qp_table[t][qscale];
1589 }
1590
1591 //FIXME need to check that this does not overflow signed 32 bit for low qp, I am not sure, it's very close
1592 //FIXME check that gcc inlines this (and optimizes intra & separate_dc stuff away)
1593 static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int separate_dc){
1594     int i;
1595     const int * const quant_table= quant_coeff[qscale];
1596     const int bias= intra ? (1<<QUANT_SHIFT)/3 : (1<<QUANT_SHIFT)/6;
1597     const unsigned int threshold1= (1<<QUANT_SHIFT) - bias - 1;
1598     const unsigned int threshold2= (threshold1<<1);
1599     int last_non_zero;
1600
1601     if(separate_dc){
1602         if(qscale<=18){
1603             //avoid overflows
1604             const int dc_bias= intra ? (1<<(QUANT_SHIFT-2))/3 : (1<<(QUANT_SHIFT-2))/6;
1605             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT-2)) - dc_bias - 1;
1606             const unsigned int dc_threshold2= (dc_threshold1<<1);
1607
1608             int level= block[0]*quant_coeff[qscale+18][0];
1609             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1610                 if(level>0){
1611                     level= (dc_bias + level)>>(QUANT_SHIFT-2);
1612                     block[0]= level;
1613                 }else{
1614                     level= (dc_bias - level)>>(QUANT_SHIFT-2);
1615                     block[0]= -level;
1616                 }
1617 //                last_non_zero = i;
1618             }else{
1619                 block[0]=0;
1620             }
1621         }else{
1622             const int dc_bias= intra ? (1<<(QUANT_SHIFT+1))/3 : (1<<(QUANT_SHIFT+1))/6;
1623             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT+1)) - dc_bias - 1;
1624             const unsigned int dc_threshold2= (dc_threshold1<<1);
1625
1626             int level= block[0]*quant_table[0];
1627             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1628                 if(level>0){
1629                     level= (dc_bias + level)>>(QUANT_SHIFT+1);
1630                     block[0]= level;
1631                 }else{
1632                     level= (dc_bias - level)>>(QUANT_SHIFT+1);
1633                     block[0]= -level;
1634                 }
1635 //                last_non_zero = i;
1636             }else{
1637                 block[0]=0;
1638             }
1639         }
1640         last_non_zero= 0;
1641         i=1;
1642     }else{
1643         last_non_zero= -1;
1644         i=0;
1645     }
1646
1647     for(; i<16; i++){
1648         const int j= scantable[i];
1649         int level= block[j]*quant_table[j];
1650
1651 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
1652 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
1653         if(((unsigned)(level+threshold1))>threshold2){
1654             if(level>0){
1655                 level= (bias + level)>>QUANT_SHIFT;
1656                 block[j]= level;
1657             }else{
1658                 level= (bias - level)>>QUANT_SHIFT;
1659                 block[j]= -level;
1660             }
1661             last_non_zero = i;
1662         }else{
1663             block[j]=0;
1664         }
1665     }
1666
1667     return last_non_zero;
1668 }
1669
1670 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1671                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1672                            int src_x_offset, int src_y_offset,
1673                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1674     MpegEncContext * const s = &h->s;
1675     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1676     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1677     const int luma_xy= (mx&3) + ((my&3)<<2);
1678     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1679     uint8_t * src_cb, * src_cr;
1680     int extra_width= h->emu_edge_width;
1681     int extra_height= h->emu_edge_height;
1682     int emu=0;
1683     const int full_mx= mx>>2;
1684     const int full_my= my>>2;
1685     const int pic_width  = 16*s->mb_width;
1686     const int pic_height = 16*s->mb_height >> MB_FIELD;
1687
1688     if(!pic->data[0]) //FIXME this is unacceptable, some sensible error concealment must be done for missing reference frames
1689         return;
1690
1691     if(mx&7) extra_width -= 3;
1692     if(my&7) extra_height -= 3;
1693
1694     if(   full_mx < 0-extra_width
1695        || full_my < 0-extra_height
1696        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1697        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1698         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);
1699             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1700         emu=1;
1701     }
1702
1703     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1704     if(!square){
1705         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1706     }
1707
1708     if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1709
1710     if(MB_FIELD){
1711         // chroma offset when predicting from a field of opposite parity
1712         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1713         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1714     }
1715     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1716     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1717
1718     if(emu){
1719         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);
1720             src_cb= s->edge_emu_buffer;
1721     }
1722     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1723
1724     if(emu){
1725         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);
1726             src_cr= s->edge_emu_buffer;
1727     }
1728     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1729 }
1730
1731 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1732                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1733                            int x_offset, int y_offset,
1734                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1735                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1736                            int list0, int list1){
1737     MpegEncContext * const s = &h->s;
1738     qpel_mc_func *qpix_op=  qpix_put;
1739     h264_chroma_mc_func chroma_op= chroma_put;
1740
1741     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1742     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1743     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1744     x_offset += 8*s->mb_x;
1745     y_offset += 8*(s->mb_y >> MB_FIELD);
1746
1747     if(list0){
1748         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1749         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1750                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1751                            qpix_op, chroma_op);
1752
1753         qpix_op=  qpix_avg;
1754         chroma_op= chroma_avg;
1755     }
1756
1757     if(list1){
1758         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1759         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1760                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1761                            qpix_op, chroma_op);
1762     }
1763 }
1764
1765 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1766                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1767                            int x_offset, int y_offset,
1768                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1769                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1770                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1771                            int list0, int list1){
1772     MpegEncContext * const s = &h->s;
1773
1774     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1775     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1776     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1777     x_offset += 8*s->mb_x;
1778     y_offset += 8*(s->mb_y >> MB_FIELD);
1779
1780     if(list0 && list1){
1781         /* don't optimize for luma-only case, since B-frames usually
1782          * use implicit weights => chroma too. */
1783         uint8_t *tmp_cb = s->obmc_scratchpad;
1784         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1785         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1786         int refn0 = h->ref_cache[0][ scan8[n] ];
1787         int refn1 = h->ref_cache[1][ scan8[n] ];
1788
1789         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1790                     dest_y, dest_cb, dest_cr,
1791                     x_offset, y_offset, qpix_put, chroma_put);
1792         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1793                     tmp_y, tmp_cb, tmp_cr,
1794                     x_offset, y_offset, qpix_put, chroma_put);
1795
1796         if(h->use_weight == 2){
1797             int weight0 = h->implicit_weight[refn0][refn1];
1798             int weight1 = 64 - weight0;
1799             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1800             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1801             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1802         }else{
1803             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1804                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1805                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1806             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1807                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1808                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1809             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1810                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1811                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1812         }
1813     }else{
1814         int list = list1 ? 1 : 0;
1815         int refn = h->ref_cache[list][ scan8[n] ];
1816         Picture *ref= &h->ref_list[list][refn];
1817         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1818                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1819                     qpix_put, chroma_put);
1820
1821         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1822                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1823         if(h->use_weight_chroma){
1824             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1825                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1826             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1827                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1828         }
1829     }
1830 }
1831
1832 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1833                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1834                            int x_offset, int y_offset,
1835                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1836                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1837                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1838                            int list0, int list1){
1839     if((h->use_weight==2 && list0 && list1
1840         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1841        || h->use_weight==1)
1842         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1843                          x_offset, y_offset, qpix_put, chroma_put,
1844                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1845     else
1846         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1847                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1848 }
1849
1850 static inline void prefetch_motion(H264Context *h, int list){
1851     /* fetch pixels for estimated mv 4 macroblocks ahead
1852      * optimized for 64byte cache lines */
1853     MpegEncContext * const s = &h->s;
1854     const int refn = h->ref_cache[list][scan8[0]];
1855     if(refn >= 0){
1856         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1857         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1858         uint8_t **src= h->ref_list[list][refn].data;
1859         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1860         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1861         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1862         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1863     }
1864 }
1865
1866 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1867                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1868                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1869                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1870     MpegEncContext * const s = &h->s;
1871     const int mb_xy= h->mb_xy;
1872     const int mb_type= s->current_picture.mb_type[mb_xy];
1873
1874     assert(IS_INTER(mb_type));
1875
1876     prefetch_motion(h, 0);
1877
1878     if(IS_16X16(mb_type)){
1879         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1880                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1881                 &weight_op[0], &weight_avg[0],
1882                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1883     }else if(IS_16X8(mb_type)){
1884         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1885                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1886                 &weight_op[1], &weight_avg[1],
1887                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1888         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1889                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1890                 &weight_op[1], &weight_avg[1],
1891                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1892     }else if(IS_8X16(mb_type)){
1893         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1894                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1895                 &weight_op[2], &weight_avg[2],
1896                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1897         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1898                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1899                 &weight_op[2], &weight_avg[2],
1900                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1901     }else{
1902         int i;
1903
1904         assert(IS_8X8(mb_type));
1905
1906         for(i=0; i<4; i++){
1907             const int sub_mb_type= h->sub_mb_type[i];
1908             const int n= 4*i;
1909             int x_offset= (i&1)<<2;
1910             int y_offset= (i&2)<<1;
1911
1912             if(IS_SUB_8X8(sub_mb_type)){
1913                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1914                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1915                     &weight_op[3], &weight_avg[3],
1916                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1917             }else if(IS_SUB_8X4(sub_mb_type)){
1918                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1919                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1920                     &weight_op[4], &weight_avg[4],
1921                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1922                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1923                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1924                     &weight_op[4], &weight_avg[4],
1925                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1926             }else if(IS_SUB_4X8(sub_mb_type)){
1927                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1928                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1929                     &weight_op[5], &weight_avg[5],
1930                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1931                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1932                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1933                     &weight_op[5], &weight_avg[5],
1934                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1935             }else{
1936                 int j;
1937                 assert(IS_SUB_4X4(sub_mb_type));
1938                 for(j=0; j<4; j++){
1939                     int sub_x_offset= x_offset + 2*(j&1);
1940                     int sub_y_offset= y_offset +   (j&2);
1941                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1942                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1943                         &weight_op[6], &weight_avg[6],
1944                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1945                 }
1946             }
1947         }
1948     }
1949
1950     prefetch_motion(h, 1);
1951 }
1952
1953 static av_cold void decode_init_vlc(void){
1954     static int done = 0;
1955
1956     if (!done) {
1957         int i;
1958         done = 1;
1959
1960         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1961                  &chroma_dc_coeff_token_len [0], 1, 1,
1962                  &chroma_dc_coeff_token_bits[0], 1, 1, 1);
1963
1964         for(i=0; i<4; i++){
1965             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1966                      &coeff_token_len [i][0], 1, 1,
1967                      &coeff_token_bits[i][0], 1, 1, 1);
1968         }
1969
1970         for(i=0; i<3; i++){
1971             init_vlc(&chroma_dc_total_zeros_vlc[i], CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1972                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1973                      &chroma_dc_total_zeros_bits[i][0], 1, 1, 1);
1974         }
1975         for(i=0; i<15; i++){
1976             init_vlc(&total_zeros_vlc[i], TOTAL_ZEROS_VLC_BITS, 16,
1977                      &total_zeros_len [i][0], 1, 1,
1978                      &total_zeros_bits[i][0], 1, 1, 1);
1979         }
1980
1981         for(i=0; i<6; i++){
1982             init_vlc(&run_vlc[i], RUN_VLC_BITS, 7,
1983                      &run_len [i][0], 1, 1,
1984                      &run_bits[i][0], 1, 1, 1);
1985         }
1986         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
1987                  &run_len [6][0], 1, 1,
1988                  &run_bits[6][0], 1, 1, 1);
1989     }
1990 }
1991
1992 static void free_tables(H264Context *h){
1993     int i;
1994     H264Context *hx;
1995     av_freep(&h->intra4x4_pred_mode);
1996     av_freep(&h->chroma_pred_mode_table);
1997     av_freep(&h->cbp_table);
1998     av_freep(&h->mvd_table[0]);
1999     av_freep(&h->mvd_table[1]);
2000     av_freep(&h->direct_table);
2001     av_freep(&h->non_zero_count);
2002     av_freep(&h->slice_table_base);
2003     h->slice_table= NULL;
2004
2005     av_freep(&h->mb2b_xy);
2006     av_freep(&h->mb2b8_xy);
2007
2008     for(i = 0; i < MAX_SPS_COUNT; i++)
2009         av_freep(h->sps_buffers + i);
2010
2011     for(i = 0; i < MAX_PPS_COUNT; i++)
2012         av_freep(h->pps_buffers + i);
2013
2014     for(i = 0; i < h->s.avctx->thread_count; i++) {
2015         hx = h->thread_context[i];
2016         if(!hx) continue;
2017         av_freep(&hx->top_borders[1]);
2018         av_freep(&hx->top_borders[0]);
2019         av_freep(&hx->s.obmc_scratchpad);
2020     }
2021 }
2022
2023 static void init_dequant8_coeff_table(H264Context *h){
2024     int i,q,x;
2025     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2026     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2027     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2028
2029     for(i=0; i<2; i++ ){
2030         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2031             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2032             break;
2033         }
2034
2035         for(q=0; q<52; q++){
2036             int shift = ff_div6[q];
2037             int idx = ff_rem6[q];
2038             for(x=0; x<64; x++)
2039                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2040                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2041                     h->pps.scaling_matrix8[i][x]) << shift;
2042         }
2043     }
2044 }
2045
2046 static void init_dequant4_coeff_table(H264Context *h){
2047     int i,j,q,x;
2048     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2049     for(i=0; i<6; i++ ){
2050         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2051         for(j=0; j<i; j++){
2052             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2053                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2054                 break;
2055             }
2056         }
2057         if(j<i)
2058             continue;
2059
2060         for(q=0; q<52; q++){
2061             int shift = ff_div6[q] + 2;
2062             int idx = ff_rem6[q];
2063             for(x=0; x<16; x++)
2064                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2065                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2066                     h->pps.scaling_matrix4[i][x]) << shift;
2067         }
2068     }
2069 }
2070
2071 static void init_dequant_tables(H264Context *h){
2072     int i,x;
2073     init_dequant4_coeff_table(h);
2074     if(h->pps.transform_8x8_mode)
2075         init_dequant8_coeff_table(h);
2076     if(h->sps.transform_bypass){
2077         for(i=0; i<6; i++)
2078             for(x=0; x<16; x++)
2079                 h->dequant4_coeff[i][0][x] = 1<<6;
2080         if(h->pps.transform_8x8_mode)
2081             for(i=0; i<2; i++)
2082                 for(x=0; x<64; x++)
2083                     h->dequant8_coeff[i][0][x] = 1<<6;
2084     }
2085 }
2086
2087
2088 /**
2089  * allocates tables.
2090  * needs width/height
2091  */
2092 static int alloc_tables(H264Context *h){
2093     MpegEncContext * const s = &h->s;
2094     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2095     int x,y;
2096
2097     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2098
2099     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2100     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(uint8_t))
2101     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2102
2103     CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2104     CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2105     CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2106     CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2107
2108     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(uint8_t));
2109     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2110
2111     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2112     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2113     for(y=0; y<s->mb_height; y++){
2114         for(x=0; x<s->mb_width; x++){
2115             const int mb_xy= x + y*s->mb_stride;
2116             const int b_xy = 4*x + 4*y*h->b_stride;
2117             const int b8_xy= 2*x + 2*y*h->b8_stride;
2118
2119             h->mb2b_xy [mb_xy]= b_xy;
2120             h->mb2b8_xy[mb_xy]= b8_xy;
2121         }
2122     }
2123
2124     s->obmc_scratchpad = NULL;
2125
2126     if(!h->dequant4_coeff[0])
2127         init_dequant_tables(h);
2128
2129     return 0;
2130 fail:
2131     free_tables(h);
2132     return -1;
2133 }
2134
2135 /**
2136  * Mimic alloc_tables(), but for every context thread.
2137  */
2138 static void clone_tables(H264Context *dst, H264Context *src){
2139     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2140     dst->non_zero_count           = src->non_zero_count;
2141     dst->slice_table              = src->slice_table;
2142     dst->cbp_table                = src->cbp_table;
2143     dst->mb2b_xy                  = src->mb2b_xy;
2144     dst->mb2b8_xy                 = src->mb2b8_xy;
2145     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2146     dst->mvd_table[0]             = src->mvd_table[0];
2147     dst->mvd_table[1]             = src->mvd_table[1];
2148     dst->direct_table             = src->direct_table;
2149
2150     dst->s.obmc_scratchpad = NULL;
2151     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2152 }
2153
2154 /**
2155  * Init context
2156  * Allocate buffers which are not shared amongst multiple threads.
2157  */
2158 static int context_init(H264Context *h){
2159     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2160     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2161
2162     return 0;
2163 fail:
2164     return -1; // free_tables will clean up for us
2165 }
2166
2167 static av_cold void common_init(H264Context *h){
2168     MpegEncContext * const s = &h->s;
2169
2170     s->width = s->avctx->width;
2171     s->height = s->avctx->height;
2172     s->codec_id= s->avctx->codec->id;
2173
2174     ff_h264_pred_init(&h->hpc, s->codec_id);
2175
2176     h->dequant_coeff_pps= -1;
2177     s->unrestricted_mv=1;
2178     s->decode=1; //FIXME
2179
2180     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2181     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2182 }
2183
2184 static av_cold int decode_init(AVCodecContext *avctx){
2185     H264Context *h= avctx->priv_data;
2186     MpegEncContext * const s = &h->s;
2187
2188     MPV_decode_defaults(s);
2189
2190     s->avctx = avctx;
2191     common_init(h);
2192
2193     s->out_format = FMT_H264;
2194     s->workaround_bugs= avctx->workaround_bugs;
2195
2196     // set defaults
2197 //    s->decode_mb= ff_h263_decode_mb;
2198     s->quarter_sample = 1;
2199     s->low_delay= 1;
2200
2201     if(avctx->codec_id == CODEC_ID_SVQ3)
2202         avctx->pix_fmt= PIX_FMT_YUVJ420P;
2203     else
2204         avctx->pix_fmt= PIX_FMT_YUV420P;
2205
2206     decode_init_vlc();
2207
2208     if(avctx->extradata_size > 0 && avctx->extradata &&
2209        *(char *)avctx->extradata == 1){
2210         h->is_avc = 1;
2211         h->got_avcC = 0;
2212     } else {
2213         h->is_avc = 0;
2214     }
2215
2216     h->thread_context[0] = h;
2217     return 0;
2218 }
2219
2220 static int frame_start(H264Context *h){
2221     MpegEncContext * const s = &h->s;
2222     int i;
2223
2224     if(MPV_frame_start(s, s->avctx) < 0)
2225         return -1;
2226     ff_er_frame_start(s);
2227     /*
2228      * MPV_frame_start uses pict_type to derive key_frame.
2229      * This is incorrect for H.264; IDR markings must be used.
2230      * Zero here; IDR markings per slice in frame or fields are ORed in later.
2231      * See decode_nal_units().
2232      */
2233     s->current_picture_ptr->key_frame= 0;
2234
2235     assert(s->linesize && s->uvlinesize);
2236
2237     for(i=0; i<16; i++){
2238         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2239         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2240     }
2241     for(i=0; i<4; i++){
2242         h->block_offset[16+i]=
2243         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2244         h->block_offset[24+16+i]=
2245         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2246     }
2247
2248     /* can't be in alloc_tables because linesize isn't known there.
2249      * FIXME: redo bipred weight to not require extra buffer? */
2250     for(i = 0; i < s->avctx->thread_count; i++)
2251         if(!h->thread_context[i]->s.obmc_scratchpad)
2252             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2253
2254     /* some macroblocks will be accessed before they're available */
2255     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2256         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(uint8_t));
2257
2258 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2259
2260     // We mark the current picture as non-reference after allocating it, so
2261     // that if we break out due to an error it can be released automatically
2262     // in the next MPV_frame_start().
2263     // SVQ3 as well as most other codecs have only last/next/current and thus
2264     // get released even with set reference, besides SVQ3 and others do not
2265     // mark frames as reference later "naturally".
2266     if(s->codec_id != CODEC_ID_SVQ3)
2267         s->current_picture_ptr->reference= 0;
2268
2269     s->current_picture_ptr->field_poc[0]=
2270     s->current_picture_ptr->field_poc[1]= INT_MAX;
2271     assert(s->current_picture_ptr->long_ref==0);
2272
2273     return 0;
2274 }
2275
2276 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){
2277     MpegEncContext * const s = &h->s;
2278     int i;
2279
2280     src_y  -=   linesize;
2281     src_cb -= uvlinesize;
2282     src_cr -= uvlinesize;
2283
2284     // There are two lines saved, the line above the the top macroblock of a pair,
2285     // and the line above the bottom macroblock
2286     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2287     for(i=1; i<17; i++){
2288         h->left_border[i]= src_y[15+i*  linesize];
2289     }
2290
2291     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2292     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2293
2294     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2295         h->left_border[17  ]= h->top_borders[0][s->mb_x][16+7];
2296         h->left_border[17+9]= h->top_borders[0][s->mb_x][24+7];
2297         for(i=1; i<9; i++){
2298             h->left_border[i+17  ]= src_cb[7+i*uvlinesize];
2299             h->left_border[i+17+9]= src_cr[7+i*uvlinesize];
2300         }
2301         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2302         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2303     }
2304 }
2305
2306 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){
2307     MpegEncContext * const s = &h->s;
2308     int temp8, i;
2309     uint64_t temp64;
2310     int deblock_left;
2311     int deblock_top;
2312     int mb_xy;
2313
2314     if(h->deblocking_filter == 2) {
2315         mb_xy = h->mb_xy;
2316         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2317         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2318     } else {
2319         deblock_left = (s->mb_x > 0);
2320         deblock_top =  (s->mb_y > 0);
2321     }
2322
2323     src_y  -=   linesize + 1;
2324     src_cb -= uvlinesize + 1;
2325     src_cr -= uvlinesize + 1;
2326
2327 #define XCHG(a,b,t,xchg)\
2328 t= a;\
2329 if(xchg)\
2330     a= b;\
2331 b= t;
2332
2333     if(deblock_left){
2334         for(i = !deblock_top; i<17; i++){
2335             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2336         }
2337     }
2338
2339     if(deblock_top){
2340         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2341         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2342         if(s->mb_x+1 < s->mb_width){
2343             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2344         }
2345     }
2346
2347     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2348         if(deblock_left){
2349             for(i = !deblock_top; i<9; i++){
2350                 XCHG(h->left_border[i+17  ], src_cb[i*uvlinesize], temp8, xchg);
2351                 XCHG(h->left_border[i+17+9], src_cr[i*uvlinesize], temp8, xchg);
2352             }
2353         }
2354         if(deblock_top){
2355             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2356             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2357         }
2358     }
2359 }
2360
2361 static inline void backup_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize){
2362     MpegEncContext * const s = &h->s;
2363     int i;
2364
2365     src_y  -= 2 *   linesize;
2366     src_cb -= 2 * uvlinesize;
2367     src_cr -= 2 * uvlinesize;
2368
2369     // There are two lines saved, the line above the the top macroblock of a pair,
2370     // and the line above the bottom macroblock
2371     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2372     h->left_border[1]= h->top_borders[1][s->mb_x][15];
2373     for(i=2; i<34; i++){
2374         h->left_border[i]= src_y[15+i*  linesize];
2375     }
2376
2377     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  32*linesize);
2378     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+32*linesize);
2379     *(uint64_t*)(h->top_borders[1][s->mb_x]+0)= *(uint64_t*)(src_y +  33*linesize);
2380     *(uint64_t*)(h->top_borders[1][s->mb_x]+8)= *(uint64_t*)(src_y +8+33*linesize);
2381
2382     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2383         h->left_border[34     ]= h->top_borders[0][s->mb_x][16+7];
2384         h->left_border[34+   1]= h->top_borders[1][s->mb_x][16+7];
2385         h->left_border[34+18  ]= h->top_borders[0][s->mb_x][24+7];
2386         h->left_border[34+18+1]= h->top_borders[1][s->mb_x][24+7];
2387         for(i=2; i<18; i++){
2388             h->left_border[i+34   ]= src_cb[7+i*uvlinesize];
2389             h->left_border[i+34+18]= src_cr[7+i*uvlinesize];
2390         }
2391         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+16*uvlinesize);
2392         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+16*uvlinesize);
2393         *(uint64_t*)(h->top_borders[1][s->mb_x]+16)= *(uint64_t*)(src_cb+17*uvlinesize);
2394         *(uint64_t*)(h->top_borders[1][s->mb_x]+24)= *(uint64_t*)(src_cr+17*uvlinesize);
2395     }
2396 }
2397
2398 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){
2399     MpegEncContext * const s = &h->s;
2400     int temp8, i;
2401     uint64_t temp64;
2402     int deblock_left = (s->mb_x > 0);
2403     int deblock_top  = (s->mb_y > 1);
2404
2405     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);
2406
2407     src_y  -= 2 *   linesize + 1;
2408     src_cb -= 2 * uvlinesize + 1;
2409     src_cr -= 2 * uvlinesize + 1;
2410
2411 #define XCHG(a,b,t,xchg)\
2412 t= a;\
2413 if(xchg)\
2414     a= b;\
2415 b= t;
2416
2417     if(deblock_left){
2418         for(i = (!deblock_top)<<1; i<34; i++){
2419             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2420         }
2421     }
2422
2423     if(deblock_top){
2424         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2425         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2426         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+0), *(uint64_t*)(src_y +1 +linesize), temp64, xchg);
2427         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+8), *(uint64_t*)(src_y +9 +linesize), temp64, 1);
2428         if(s->mb_x+1 < s->mb_width){
2429             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2430             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x+1]), *(uint64_t*)(src_y +17 +linesize), temp64, 1);
2431         }
2432     }
2433
2434     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2435         if(deblock_left){
2436             for(i = (!deblock_top) << 1; i<18; i++){
2437                 XCHG(h->left_border[i+34   ], src_cb[i*uvlinesize], temp8, xchg);
2438                 XCHG(h->left_border[i+34+18], src_cr[i*uvlinesize], temp8, xchg);
2439             }
2440         }
2441         if(deblock_top){
2442             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2443             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2444             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+16), *(uint64_t*)(src_cb+1 +uvlinesize), temp64, 1);
2445             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+24), *(uint64_t*)(src_cr+1 +uvlinesize), temp64, 1);
2446         }
2447     }
2448 }
2449
2450 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2451     MpegEncContext * const s = &h->s;
2452     const int mb_x= s->mb_x;
2453     const int mb_y= s->mb_y;
2454     const int mb_xy= h->mb_xy;
2455     const int mb_type= s->current_picture.mb_type[mb_xy];
2456     uint8_t  *dest_y, *dest_cb, *dest_cr;
2457     int linesize, uvlinesize /*dct_offset*/;
2458     int i;
2459     int *block_offset = &h->block_offset[0];
2460     const unsigned int bottom = mb_y & 1;
2461     const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass), is_h264 = (simple || s->codec_id == CODEC_ID_H264);
2462     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2463     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2464
2465     dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2466     dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2467     dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2468
2469     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2470     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2471
2472     if (!simple && MB_FIELD) {
2473         linesize   = h->mb_linesize   = s->linesize * 2;
2474         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2475         block_offset = &h->block_offset[24];
2476         if(mb_y&1){ //FIXME move out of this function?
2477             dest_y -= s->linesize*15;
2478             dest_cb-= s->uvlinesize*7;
2479             dest_cr-= s->uvlinesize*7;
2480         }
2481         if(FRAME_MBAFF) {
2482             int list;
2483             for(list=0; list<h->list_count; list++){
2484                 if(!USES_LIST(mb_type, list))
2485                     continue;
2486                 if(IS_16X16(mb_type)){
2487                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2488                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2489                 }else{
2490                     for(i=0; i<16; i+=4){
2491                         //FIXME can refs be smaller than 8x8 when !direct_8x8_inference ?
2492                         int ref = h->ref_cache[list][scan8[i]];
2493                         if(ref >= 0)
2494                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2495                     }
2496                 }
2497             }
2498         }
2499     } else {
2500         linesize   = h->mb_linesize   = s->linesize;
2501         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2502 //        dct_offset = s->linesize * 16;
2503     }
2504
2505     if(transform_bypass){
2506         idct_dc_add =
2507         idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2508     }else if(IS_8x8DCT(mb_type)){
2509         idct_dc_add = s->dsp.h264_idct8_dc_add;
2510         idct_add = s->dsp.h264_idct8_add;
2511     }else{
2512         idct_dc_add = s->dsp.h264_idct_dc_add;
2513         idct_add = s->dsp.h264_idct_add;
2514     }
2515
2516     if(!simple && FRAME_MBAFF && h->deblocking_filter && IS_INTRA(mb_type)
2517        && (!bottom || !IS_INTRA(s->current_picture.mb_type[mb_xy-s->mb_stride]))){
2518         int mbt_y = mb_y&~1;
2519         uint8_t *top_y  = s->current_picture.data[0] + (mbt_y * 16* s->linesize  ) + mb_x * 16;
2520         uint8_t *top_cb = s->current_picture.data[1] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2521         uint8_t *top_cr = s->current_picture.data[2] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2522         xchg_pair_border(h, top_y, top_cb, top_cr, s->linesize, s->uvlinesize, 1);
2523     }
2524
2525     if (!simple && IS_INTRA_PCM(mb_type)) {
2526         unsigned int x, y;
2527
2528         // The pixels are stored in h->mb array in the same order as levels,
2529         // copy them in output in the correct order.
2530         for(i=0; i<16; i++) {
2531             for (y=0; y<4; y++) {
2532                 for (x=0; x<4; x++) {
2533                     *(dest_y + block_offset[i] + y*linesize + x) = h->mb[i*16+y*4+x];
2534                 }
2535             }
2536         }
2537         for(i=16; i<16+4; i++) {
2538             for (y=0; y<4; y++) {
2539                 for (x=0; x<4; x++) {
2540                     *(dest_cb + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2541                 }
2542             }
2543         }
2544         for(i=20; i<20+4; i++) {
2545             for (y=0; y<4; y++) {
2546                 for (x=0; x<4; x++) {
2547                     *(dest_cr + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2548                 }
2549             }
2550         }
2551     } else {
2552         if(IS_INTRA(mb_type)){
2553             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2554                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2555
2556             if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2557                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2558                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2559             }
2560
2561             if(IS_INTRA4x4(mb_type)){
2562                 if(simple || !s->encoding){
2563                     if(IS_8x8DCT(mb_type)){
2564                         for(i=0; i<16; i+=4){
2565                             uint8_t * const ptr= dest_y + block_offset[i];
2566                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2567                             const int nnz = h->non_zero_count_cache[ scan8[i] ];
2568                             h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2569                                                    (h->topright_samples_available<<i)&0x4000, linesize);
2570                             if(nnz){
2571                                 if(nnz == 1 && h->mb[i*16])
2572                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2573                                 else
2574                                     idct_add(ptr, h->mb + i*16, linesize);
2575                             }
2576                         }
2577                     }else
2578                     for(i=0; i<16; i++){
2579                         uint8_t * const ptr= dest_y + block_offset[i];
2580                         uint8_t *topright;
2581                         const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2582                         int nnz, tr;
2583
2584                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2585                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2586                             assert(mb_y || linesize <= block_offset[i]);
2587                             if(!topright_avail){
2588                                 tr= ptr[3 - linesize]*0x01010101;
2589                                 topright= (uint8_t*) &tr;
2590                             }else
2591                                 topright= ptr + 4 - linesize;
2592                         }else
2593                             topright= NULL;
2594
2595                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2596                         nnz = h->non_zero_count_cache[ scan8[i] ];
2597                         if(nnz){
2598                             if(is_h264){
2599                                 if(nnz == 1 && h->mb[i*16])
2600                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2601                                 else
2602                                     idct_add(ptr, h->mb + i*16, linesize);
2603                             }else
2604                                 svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2605                         }
2606                     }
2607                 }
2608             }else{
2609                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2610                 if(is_h264){
2611                     if(!transform_bypass)
2612                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2613                 }else
2614                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2615             }
2616             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2617                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2618         }else if(is_h264){
2619             hl_motion(h, dest_y, dest_cb, dest_cr,
2620                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2621                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2622                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2623         }
2624
2625
2626         if(!IS_INTRA4x4(mb_type)){
2627             if(is_h264){
2628                 if(IS_INTRA16x16(mb_type)){
2629                     for(i=0; i<16; i++){
2630                         if(h->non_zero_count_cache[ scan8[i] ])
2631                             idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2632                         else if(h->mb[i*16])
2633                             idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2634                     }
2635                 }else{
2636                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2637                     for(i=0; i<16; i+=di){
2638                         int nnz = h->non_zero_count_cache[ scan8[i] ];
2639                         if(nnz){
2640                             if(nnz==1 && h->mb[i*16])
2641                                 idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2642                             else
2643                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2644                         }
2645                     }
2646                 }
2647             }else{
2648                 for(i=0; i<16; i++){
2649                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2650                         uint8_t * const ptr= dest_y + block_offset[i];
2651                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2652                     }
2653                 }
2654             }
2655         }
2656
2657         if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2658             uint8_t *dest[2] = {dest_cb, dest_cr};
2659             if(transform_bypass){
2660                 idct_add = idct_dc_add = s->dsp.add_pixels4;
2661             }else{
2662                 idct_add = s->dsp.h264_idct_add;
2663                 idct_dc_add = s->dsp.h264_idct_dc_add;
2664                 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]);
2665                 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]);
2666             }
2667             if(is_h264){
2668                 for(i=16; i<16+8; i++){
2669                     if(h->non_zero_count_cache[ scan8[i] ])
2670                         idct_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2671                     else if(h->mb[i*16])
2672                         idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2673                 }
2674             }else{
2675                 for(i=16; i<16+8; i++){
2676                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2677                         uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2678                         svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2679                     }
2680                 }
2681             }
2682         }
2683     }
2684     if(h->deblocking_filter) {
2685         if (!simple && FRAME_MBAFF) {
2686             //FIXME try deblocking one mb at a time?
2687             // the reduction in load/storing mvs and such might outweigh the extra backup/xchg_border
2688             const int mb_y = s->mb_y - 1;
2689             uint8_t  *pair_dest_y, *pair_dest_cb, *pair_dest_cr;
2690             const int mb_xy= mb_x + mb_y*s->mb_stride;
2691             const int mb_type_top   = s->current_picture.mb_type[mb_xy];
2692             const int mb_type_bottom= s->current_picture.mb_type[mb_xy+s->mb_stride];
2693             if (!bottom) return;
2694             pair_dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2695             pair_dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2696             pair_dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2697
2698             if(IS_INTRA(mb_type_top | mb_type_bottom))
2699                 xchg_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize, 0);
2700
2701             backup_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize);
2702             // deblock a pair
2703             // top
2704             s->mb_y--; h->mb_xy -= s->mb_stride;
2705             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);
2706             fill_caches(h, mb_type_top, 1); //FIXME don't fill stuff which isn't used by filter_mb
2707             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2708             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2709             filter_mb(h, mb_x, mb_y, pair_dest_y, pair_dest_cb, pair_dest_cr, linesize, uvlinesize);
2710             // bottom
2711             s->mb_y++; h->mb_xy += s->mb_stride;
2712             tprintf(h->s.avctx, "call mbaff filter_mb\n");
2713             fill_caches(h, mb_type_bottom, 1); //FIXME don't fill stuff which isn't used by filter_mb
2714             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2715             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2716             filter_mb(h, mb_x, mb_y+1, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2717         } else {
2718             tprintf(h->s.avctx, "call filter_mb\n");
2719             backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2720             fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2721             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2722             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2723             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2724         }
2725     }
2726 }
2727
2728 /**
2729  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2730  */
2731 static void hl_decode_mb_simple(H264Context *h){
2732     hl_decode_mb_internal(h, 1);
2733 }
2734
2735 /**
2736  * Process a macroblock; this handles edge cases, such as interlacing.
2737  */
2738 static void av_noinline hl_decode_mb_complex(H264Context *h){
2739     hl_decode_mb_internal(h, 0);
2740 }
2741
2742 static void hl_decode_mb(H264Context *h){
2743     MpegEncContext * const s = &h->s;
2744     const int mb_xy= h->mb_xy;
2745     const int mb_type= s->current_picture.mb_type[mb_xy];
2746     int is_complex = FRAME_MBAFF || MB_FIELD || IS_INTRA_PCM(mb_type) || s->codec_id != CODEC_ID_H264 ||
2747                     (ENABLE_GRAY && (s->flags&CODEC_FLAG_GRAY)) || (ENABLE_H264_ENCODER && s->encoding) || ENABLE_SMALL;
2748
2749     if(ENABLE_H264_ENCODER && !s->decode)
2750         return;
2751
2752     if (is_complex)
2753         hl_decode_mb_complex(h);
2754     else hl_decode_mb_simple(h);
2755 }
2756
2757 static void pic_as_field(Picture *pic, const int parity){
2758     int i;
2759     for (i = 0; i < 4; ++i) {
2760         if (parity == PICT_BOTTOM_FIELD)
2761             pic->data[i] += pic->linesize[i];
2762         pic->reference = parity;
2763         pic->linesize[i] *= 2;
2764     }
2765 }
2766
2767 static int split_field_copy(Picture *dest, Picture *src,
2768                             int parity, int id_add){
2769     int match = !!(src->reference & parity);
2770
2771     if (match) {
2772         *dest = *src;
2773         pic_as_field(dest, parity);
2774         dest->pic_id *= 2;
2775         dest->pic_id += id_add;
2776     }
2777
2778     return match;
2779 }
2780
2781 /**
2782  * Split one reference list into field parts, interleaving by parity
2783  * as per H.264 spec section 8.2.4.2.5. Output fields have their data pointers
2784  * set to look at the actual start of data for that field.
2785  *
2786  * @param dest output list
2787  * @param dest_len maximum number of fields to put in dest
2788  * @param src the source reference list containing fields and/or field pairs
2789  *            (aka short_ref/long_ref, or
2790  *             refFrameListXShortTerm/refFrameListLongTerm in spec-speak)
2791  * @param src_len number of Picture's in source (pairs and unmatched fields)
2792  * @param parity the parity of the picture being decoded/needing
2793  *        these ref pics (PICT_{TOP,BOTTOM}_FIELD)
2794  * @return number of fields placed in dest
2795  */
2796 static int split_field_half_ref_list(Picture *dest, int dest_len,
2797                                      Picture *src,  int src_len,  int parity){
2798     int same_parity   = 1;
2799     int same_i        = 0;
2800     int opp_i         = 0;
2801     int out_i;
2802     int field_output;
2803
2804     for (out_i = 0; out_i < dest_len; out_i += field_output) {
2805         if (same_parity && same_i < src_len) {
2806             field_output = split_field_copy(dest + out_i, src + same_i,
2807                                             parity, 1);
2808             same_parity = !field_output;
2809             same_i++;
2810
2811         } else if (opp_i < src_len) {
2812             field_output = split_field_copy(dest + out_i, src + opp_i,
2813                                             PICT_FRAME - parity, 0);
2814             same_parity = field_output;
2815             opp_i++;
2816
2817         } else {
2818             break;
2819         }
2820     }
2821
2822     return out_i;
2823 }
2824
2825 /**
2826  * Split the reference frame list into a reference field list.
2827  * This implements H.264 spec 8.2.4.2.5 for a combined input list.
2828  * The input list contains both reference field pairs and
2829  * unmatched reference fields; it is ordered as spec describes
2830  * RefPicListX for frames in 8.2.4.2.1 and 8.2.4.2.3, except that
2831  * unmatched field pairs are also present. Conceptually this is equivalent
2832  * to concatenation of refFrameListXShortTerm with refFrameListLongTerm.
2833  *
2834  * @param dest output reference list where ordered fields are to be placed
2835  * @param dest_len max number of fields to place at dest
2836  * @param src source reference list, as described above
2837  * @param src_len number of pictures (pairs and unmatched fields) in src
2838  * @param parity parity of field being currently decoded
2839  *        (one of PICT_{TOP,BOTTOM}_FIELD)
2840  * @param long_i index into src array that holds first long reference picture,
2841  *        or src_len if no long refs present.
2842  */
2843 static int split_field_ref_list(Picture *dest, int dest_len,
2844                                 Picture *src,  int src_len,
2845                                 int parity,    int long_i){
2846
2847     int i = split_field_half_ref_list(dest, dest_len, src, long_i, parity);
2848     dest += i;
2849     dest_len -= i;
2850
2851     i += split_field_half_ref_list(dest, dest_len, src + long_i,
2852                                    src_len - long_i, parity);
2853     return i;
2854 }
2855
2856 /**
2857  * fills the default_ref_list.
2858  */
2859 static int fill_default_ref_list(H264Context *h){
2860     MpegEncContext * const s = &h->s;
2861     int i;
2862     int smallest_poc_greater_than_current = -1;
2863     int structure_sel;
2864     Picture sorted_short_ref[32];
2865     Picture field_entry_list[2][32];
2866     Picture *frame_list[2];
2867
2868     if (FIELD_PICTURE) {
2869         structure_sel = PICT_FRAME;
2870         frame_list[0] = field_entry_list[0];
2871         frame_list[1] = field_entry_list[1];
2872     } else {
2873         structure_sel = 0;
2874         frame_list[0] = h->default_ref_list[0];
2875         frame_list[1] = h->default_ref_list[1];
2876     }
2877
2878     if(h->slice_type_nos==FF_B_TYPE){
2879         int list;
2880         int len[2];
2881         int short_len[2];
2882         int out_i;
2883         int limit= INT_MIN;
2884
2885         /* sort frame according to POC in B slice */
2886         for(out_i=0; out_i<h->short_ref_count; out_i++){
2887             int best_i=INT_MIN;
2888             int best_poc=INT_MAX;
2889
2890             for(i=0; i<h->short_ref_count; i++){
2891                 const int poc= h->short_ref[i]->poc;
2892                 if(poc > limit && poc < best_poc){
2893                     best_poc= poc;
2894                     best_i= i;
2895                 }
2896             }
2897
2898             assert(best_i != INT_MIN);
2899
2900             limit= best_poc;
2901             sorted_short_ref[out_i]= *h->short_ref[best_i];
2902             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);
2903             if (-1 == smallest_poc_greater_than_current) {
2904                 if (h->short_ref[best_i]->poc >= s->current_picture_ptr->poc) {
2905                     smallest_poc_greater_than_current = out_i;
2906                 }
2907             }
2908         }
2909
2910         tprintf(h->s.avctx, "current poc: %d, smallest_poc_greater_than_current: %d\n", s->current_picture_ptr->poc, smallest_poc_greater_than_current);
2911
2912         // find the largest POC
2913         for(list=0; list<2; list++){
2914             int index = 0;
2915             int j= -99;
2916             int step= list ? -1 : 1;
2917
2918             for(i=0; i<h->short_ref_count && index < h->ref_count[list]; i++, j+=step) {
2919                 int sel;
2920                 while(j<0 || j>= h->short_ref_count){
2921                     if(j != -99 && step == (list ? -1 : 1))
2922                         return -1;
2923                     step = -step;
2924                     j= smallest_poc_greater_than_current + (step>>1);
2925                 }
2926                 sel = sorted_short_ref[j].reference | structure_sel;
2927                 if(sel != PICT_FRAME) continue;
2928                 frame_list[list][index  ]= sorted_short_ref[j];
2929                 frame_list[list][index++].pic_id= sorted_short_ref[j].frame_num;
2930             }
2931             short_len[list] = index;
2932
2933             for(i = 0; i < 16 && index < h->ref_count[ list ]; i++){
2934                 int sel;
2935                 if(h->long_ref[i] == NULL) continue;
2936                 sel = h->long_ref[i]->reference | structure_sel;
2937                 if(sel != PICT_FRAME) continue;
2938
2939                 frame_list[ list ][index  ]= *h->long_ref[i];
2940                 frame_list[ list ][index++].pic_id= i;
2941             }
2942             len[list] = index;
2943         }
2944
2945         for(list=0; list<2; list++){
2946             if (FIELD_PICTURE)
2947                 len[list] = split_field_ref_list(h->default_ref_list[list],
2948                                                  h->ref_count[list],
2949                                                  frame_list[list],
2950                                                  len[list],
2951                                                  s->picture_structure,
2952                                                  short_len[list]);
2953
2954             // swap the two first elements of L1 when L0 and L1 are identical
2955             if(list && len[0] > 1 && len[0] == len[1])
2956                 for(i=0; h->default_ref_list[0][i].data[0] == h->default_ref_list[1][i].data[0]; i++)
2957                     if(i == len[0]){
2958                         FFSWAP(Picture, h->default_ref_list[1][0], h->default_ref_list[1][1]);
2959                         break;
2960                     }
2961
2962             if(len[list] < h->ref_count[ list ])
2963                 memset(&h->default_ref_list[list][len[list]], 0, sizeof(Picture)*(h->ref_count[ list ] - len[list]));
2964         }
2965
2966
2967     }else{
2968         int index=0;
2969         int short_len;
2970         for(i=0; i<h->short_ref_count; i++){
2971             int sel;
2972             sel = h->short_ref[i]->reference | structure_sel;
2973             if(sel != PICT_FRAME) continue;
2974             frame_list[0][index  ]= *h->short_ref[i];
2975             frame_list[0][index++].pic_id= h->short_ref[i]->frame_num;
2976         }
2977         short_len = index;
2978         for(i = 0; i < 16; i++){
2979             int sel;
2980             if(h->long_ref[i] == NULL) continue;
2981             sel = h->long_ref[i]->reference | structure_sel;
2982             if(sel != PICT_FRAME) continue;
2983             frame_list[0][index  ]= *h->long_ref[i];
2984             frame_list[0][index++].pic_id= i;
2985         }
2986
2987         if (FIELD_PICTURE)
2988             index = split_field_ref_list(h->default_ref_list[0],
2989                                          h->ref_count[0], frame_list[0],
2990                                          index, s->picture_structure,
2991                                          short_len);
2992
2993         if(index < h->ref_count[0])
2994             memset(&h->default_ref_list[0][index], 0, sizeof(Picture)*(h->ref_count[0] - index));
2995     }
2996 #ifdef TRACE
2997     for (i=0; i<h->ref_count[0]; i++) {
2998         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]);
2999     }
3000     if(h->slice_type_nos==FF_B_TYPE){
3001         for (i=0; i<h->ref_count[1]; i++) {
3002             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]);
3003         }
3004     }
3005 #endif
3006     return 0;
3007 }
3008
3009 static void print_short_term(H264Context *h);
3010 static void print_long_term(H264Context *h);
3011
3012 /**
3013  * Extract structure information about the picture described by pic_num in
3014  * the current decoding context (frame or field). Note that pic_num is
3015  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
3016  * @param pic_num picture number for which to extract structure information
3017  * @param structure one of PICT_XXX describing structure of picture
3018  *                      with pic_num
3019  * @return frame number (short term) or long term index of picture
3020  *         described by pic_num
3021  */
3022 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
3023     MpegEncContext * const s = &h->s;
3024
3025     *structure = s->picture_structure;
3026     if(FIELD_PICTURE){
3027         if (!(pic_num & 1))
3028             /* opposite field */
3029             *structure ^= PICT_FRAME;
3030         pic_num >>= 1;
3031     }
3032
3033     return pic_num;
3034 }
3035
3036 static int decode_ref_pic_list_reordering(H264Context *h){
3037     MpegEncContext * const s = &h->s;
3038     int list, index, pic_structure;
3039
3040     print_short_term(h);
3041     print_long_term(h);
3042     if(h->slice_type_nos==FF_I_TYPE) return 0; //FIXME move before function
3043
3044     for(list=0; list<h->list_count; list++){
3045         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
3046
3047         if(get_bits1(&s->gb)){
3048             int pred= h->curr_pic_num;
3049
3050             for(index=0; ; index++){
3051                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb(&s->gb);
3052                 unsigned int pic_id;
3053                 int i;
3054                 Picture *ref = NULL;
3055
3056                 if(reordering_of_pic_nums_idc==3)
3057                     break;
3058
3059                 if(index >= h->ref_count[list]){
3060                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
3061                     return -1;
3062                 }
3063
3064                 if(reordering_of_pic_nums_idc<3){
3065                     if(reordering_of_pic_nums_idc<2){
3066                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
3067                         int frame_num;
3068
3069                         if(abs_diff_pic_num > h->max_pic_num){
3070                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
3071                             return -1;
3072                         }
3073
3074                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
3075                         else                                pred+= abs_diff_pic_num;
3076                         pred &= h->max_pic_num - 1;
3077
3078                         frame_num = pic_num_extract(h, pred, &pic_structure);
3079
3080                         for(i= h->short_ref_count-1; i>=0; i--){
3081                             ref = h->short_ref[i];
3082                             assert(ref->reference);
3083                             assert(!ref->long_ref);
3084                             if(ref->data[0] != NULL &&
3085                                    ref->frame_num == frame_num &&
3086                                    (ref->reference & pic_structure) &&
3087                                    ref->long_ref == 0) // ignore non-existing pictures by testing data[0] pointer
3088                                 break;
3089                         }
3090                         if(i>=0)
3091                             ref->pic_id= pred;
3092                     }else{
3093                         int long_idx;
3094                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
3095
3096                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
3097
3098                         if(long_idx>31){
3099                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
3100                             return -1;
3101                         }
3102                         ref = h->long_ref[long_idx];
3103                         assert(!(ref && !ref->reference));
3104                         if(ref && (ref->reference & pic_structure)){
3105                             ref->pic_id= pic_id;
3106                             assert(ref->long_ref);
3107                             i=0;
3108                         }else{
3109                             i=-1;
3110                         }
3111                     }
3112
3113                     if (i < 0) {
3114                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
3115                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
3116                     } else {
3117                         for(i=index; i+1<h->ref_count[list]; i++){
3118                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
3119                                 break;
3120                         }
3121                         for(; i > index; i--){
3122                             h->ref_list[list][i]= h->ref_list[list][i-1];
3123                         }
3124                         h->ref_list[list][index]= *ref;
3125                         if (FIELD_PICTURE){
3126                             pic_as_field(&h->ref_list[list][index], pic_structure);
3127                         }
3128                     }
3129                 }else{
3130                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
3131                     return -1;
3132                 }
3133             }
3134         }
3135     }
3136     for(list=0; list<h->list_count; list++){
3137         for(index= 0; index < h->ref_count[list]; index++){
3138             if(!h->ref_list[list][index].data[0])
3139                 h->ref_list[list][index]= s->current_picture;
3140         }
3141     }
3142
3143     if(h->slice_type_nos==FF_B_TYPE && !h->direct_spatial_mv_pred)
3144         direct_dist_scale_factor(h);
3145     direct_ref_list_init(h);
3146     return 0;
3147 }
3148
3149 static void fill_mbaff_ref_list(H264Context *h){
3150     int list, i, j;
3151     for(list=0; list<2; list++){ //FIXME try list_count
3152         for(i=0; i<h->ref_count[list]; i++){
3153             Picture *frame = &h->ref_list[list][i];
3154             Picture *field = &h->ref_list[list][16+2*i];
3155             field[0] = *frame;
3156             for(j=0; j<3; j++)
3157                 field[0].linesize[j] <<= 1;
3158             field[0].reference = PICT_TOP_FIELD;
3159             field[1] = field[0];
3160             for(j=0; j<3; j++)
3161                 field[1].data[j] += frame->linesize[j];
3162             field[1].reference = PICT_BOTTOM_FIELD;
3163
3164             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
3165             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
3166             for(j=0; j<2; j++){
3167                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
3168                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
3169             }
3170         }
3171     }
3172     for(j=0; j<h->ref_count[1]; j++){
3173         for(i=0; i<h->ref_count[0]; i++)
3174             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
3175         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
3176         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
3177     }
3178 }
3179
3180 static int pred_weight_table(H264Context *h){
3181     MpegEncContext * const s = &h->s;
3182     int list, i;
3183     int luma_def, chroma_def;
3184
3185     h->use_weight= 0;
3186     h->use_weight_chroma= 0;
3187     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
3188     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
3189     luma_def = 1<<h->luma_log2_weight_denom;
3190     chroma_def = 1<<h->chroma_log2_weight_denom;
3191
3192     for(list=0; list<2; list++){
3193         for(i=0; i<h->ref_count[list]; i++){
3194             int luma_weight_flag, chroma_weight_flag;
3195
3196             luma_weight_flag= get_bits1(&s->gb);
3197             if(luma_weight_flag){
3198                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3199                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3200                 if(   h->luma_weight[list][i] != luma_def
3201                    || h->luma_offset[list][i] != 0)
3202                     h->use_weight= 1;
3203             }else{
3204                 h->luma_weight[list][i]= luma_def;
3205                 h->luma_offset[list][i]= 0;
3206             }
3207
3208             chroma_weight_flag= get_bits1(&s->gb);
3209             if(chroma_weight_flag){
3210                 int j;
3211                 for(j=0; j<2; j++){
3212                     h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3213                     h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3214                     if(   h->chroma_weight[list][i][j] != chroma_def
3215                        || h->chroma_offset[list][i][j] != 0)
3216                         h->use_weight_chroma= 1;
3217                 }
3218             }else{
3219                 int j;
3220                 for(j=0; j<2; j++){
3221                     h->chroma_weight[list][i][j]= chroma_def;
3222                     h->chroma_offset[list][i][j]= 0;
3223                 }
3224             }
3225         }
3226         if(h->slice_type_nos != FF_B_TYPE) break;
3227     }
3228     h->use_weight= h->use_weight || h->use_weight_chroma;
3229     return 0;
3230 }
3231
3232 static void implicit_weight_table(H264Context *h){
3233     MpegEncContext * const s = &h->s;
3234     int ref0, ref1;
3235     int cur_poc = s->current_picture_ptr->poc;
3236
3237     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3238        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3239         h->use_weight= 0;
3240         h->use_weight_chroma= 0;
3241         return;
3242     }
3243
3244     h->use_weight= 2;
3245     h->use_weight_chroma= 2;
3246     h->luma_log2_weight_denom= 5;
3247     h->chroma_log2_weight_denom= 5;
3248
3249     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3250         int poc0 = h->ref_list[0][ref0].poc;
3251         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3252             int poc1 = h->ref_list[1][ref1].poc;
3253             int td = av_clip(poc1 - poc0, -128, 127);
3254             if(td){
3255                 int tb = av_clip(cur_poc - poc0, -128, 127);
3256                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3257                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3258                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3259                     h->implicit_weight[ref0][ref1] = 32;
3260                 else
3261                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3262             }else
3263                 h->implicit_weight[ref0][ref1] = 32;
3264         }
3265     }
3266 }
3267
3268 /**
3269  * Mark a picture as no longer needed for reference. The refmask
3270  * argument allows unreferencing of individual fields or the whole frame.
3271  * If the picture becomes entirely unreferenced, but is being held for
3272  * display purposes, it is marked as such.
3273  * @param refmask mask of fields to unreference; the mask is bitwise
3274  *                anded with the reference marking of pic
3275  * @return non-zero if pic becomes entirely unreferenced (except possibly
3276  *         for display purposes) zero if one of the fields remains in
3277  *         reference
3278  */
3279 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3280     int i;
3281     if (pic->reference &= refmask) {
3282         return 0;
3283     } else {
3284         for(i = 0; h->delayed_pic[i]; i++)
3285             if(pic == h->delayed_pic[i]){
3286                 pic->reference=DELAYED_PIC_REF;
3287                 break;
3288             }
3289         return 1;
3290     }
3291 }
3292
3293 /**
3294  * instantaneous decoder refresh.
3295  */
3296 static void idr(H264Context *h){
3297     int i;
3298
3299     for(i=0; i<16; i++){
3300         remove_long(h, i, 0);
3301     }
3302     assert(h->long_ref_count==0);
3303
3304     for(i=0; i<h->short_ref_count; i++){
3305         unreference_pic(h, h->short_ref[i], 0);
3306         h->short_ref[i]= NULL;
3307     }
3308     h->short_ref_count=0;
3309     h->prev_frame_num= 0;
3310     h->prev_frame_num_offset= 0;
3311     h->prev_poc_msb=
3312     h->prev_poc_lsb= 0;
3313 }
3314
3315 /* forget old pics after a seek */
3316 static void flush_dpb(AVCodecContext *avctx){
3317     H264Context *h= avctx->priv_data;
3318     int i;
3319     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
3320         if(h->delayed_pic[i])
3321             h->delayed_pic[i]->reference= 0;
3322         h->delayed_pic[i]= NULL;
3323     }
3324     h->outputed_poc= INT_MIN;
3325     idr(h);
3326     if(h->s.current_picture_ptr)
3327         h->s.current_picture_ptr->reference= 0;
3328     h->s.first_field= 0;
3329     ff_mpeg_flush(avctx);
3330 }
3331
3332 /**
3333  * Find a Picture in the short term reference list by frame number.
3334  * @param frame_num frame number to search for
3335  * @param idx the index into h->short_ref where returned picture is found
3336  *            undefined if no picture found.
3337  * @return pointer to the found picture, or NULL if no pic with the provided
3338  *                 frame number is found
3339  */
3340 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3341     MpegEncContext * const s = &h->s;
3342     int i;
3343
3344     for(i=0; i<h->short_ref_count; i++){
3345         Picture *pic= h->short_ref[i];
3346         if(s->avctx->debug&FF_DEBUG_MMCO)
3347             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3348         if(pic->frame_num == frame_num) {
3349             *idx = i;
3350             return pic;
3351         }
3352     }
3353     return NULL;
3354 }
3355
3356 /**
3357  * Remove a picture from the short term reference list by its index in
3358  * that list.  This does no checking on the provided index; it is assumed
3359  * to be valid. Other list entries are shifted down.
3360  * @param i index into h->short_ref of picture to remove.
3361  */
3362 static void remove_short_at_index(H264Context *h, int i){
3363     assert(i >= 0 && i < h->short_ref_count);
3364     h->short_ref[i]= NULL;
3365     if (--h->short_ref_count)
3366         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3367 }
3368
3369 /**
3370  *
3371  * @return the removed picture or NULL if an error occurs
3372  */
3373 static Picture * remove_short(H264Context *h, int frame_num, int ref_mask){
3374     MpegEncContext * const s = &h->s;
3375     Picture *pic;
3376     int i;
3377
3378     if(s->avctx->debug&FF_DEBUG_MMCO)
3379         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3380
3381     pic = find_short(h, frame_num, &i);
3382     if (pic){
3383         if(unreference_pic(h, pic, ref_mask))
3384         remove_short_at_index(h, i);
3385     }
3386
3387     return pic;
3388 }
3389
3390 /**
3391  * Remove a picture from the long term reference list by its index in
3392  * that list.
3393  * @return the removed picture or NULL if an error occurs
3394  */
3395 static Picture * remove_long(H264Context *h, int i, int ref_mask){
3396     Picture *pic;
3397
3398     pic= h->long_ref[i];
3399     if (pic){
3400         if(unreference_pic(h, pic, ref_mask)){
3401             assert(h->long_ref[i]->long_ref == 1);
3402             h->long_ref[i]->long_ref= 0;
3403             h->long_ref[i]= NULL;
3404             h->long_ref_count--;
3405         }
3406     }
3407
3408     return pic;
3409 }
3410
3411 /**
3412  * print short term list
3413  */
3414 static void print_short_term(H264Context *h) {
3415     uint32_t i;
3416     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3417         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3418         for(i=0; i<h->short_ref_count; i++){
3419             Picture *pic= h->short_ref[i];
3420             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3421         }
3422     }
3423 }
3424