ecd4fc742ad0adc6534af521e3d9bfb2980e53d4
[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_TYPE coeff_token_vlc_tables[520+332+280+256][2];
53 static const int coeff_token_vlc_tables_size[4]={520,332,280,256};
54
55 static VLC chroma_dc_coeff_token_vlc;
56 static VLC_TYPE chroma_dc_coeff_token_vlc_table[256][2];
57 static const int chroma_dc_coeff_token_vlc_table_size = 256;
58
59 static VLC total_zeros_vlc[15];
60 static VLC_TYPE total_zeros_vlc_tables[15][512][2];
61 static const int total_zeros_vlc_tables_size = 512;
62
63 static VLC chroma_dc_total_zeros_vlc[3];
64 static VLC_TYPE chroma_dc_total_zeros_vlc_tables[3][8][2];
65 static const int chroma_dc_total_zeros_vlc_tables_size = 8;
66
67 static VLC run_vlc[6];
68 static VLC_TYPE run_vlc_tables[6][8][2];
69 static const int run_vlc_tables_size = 8;
70
71 static VLC run7_vlc;
72 static VLC_TYPE run7_vlc_table[96][2];
73 static const int run7_vlc_table_size = 96;
74
75 static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
76 static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
77 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);
78 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);
79 static Picture * remove_long(H264Context *h, int i, int ref_mask);
80
81 static av_always_inline uint32_t pack16to32(int a, int b){
82 #ifdef WORDS_BIGENDIAN
83    return (b&0xFFFF) + (a<<16);
84 #else
85    return (a&0xFFFF) + (b<<16);
86 #endif
87 }
88
89 static const uint8_t rem6[52]={
90 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,
91 };
92
93 static const uint8_t div6[52]={
94 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,
95 };
96
97 static const int left_block_options[4][8]={
98     {0,1,2,3,7,10,8,11},
99     {2,2,3,3,8,11,8,11},
100     {0,0,1,1,7,10,7,10},
101     {0,2,0,2,7,10,7,10}
102 };
103
104 static void fill_caches(H264Context *h, int mb_type, int for_deblock){
105     MpegEncContext * const s = &h->s;
106     const int mb_xy= h->mb_xy;
107     int topleft_xy, top_xy, topright_xy, left_xy[2];
108     int topleft_type, top_type, topright_type, left_type[2];
109     int * left_block;
110     int topleft_partition= -1;
111     int i;
112
113     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
114
115     //FIXME deblocking could skip the intra and nnz parts.
116     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
117         return;
118
119     /* Wow, what a mess, why didn't they simplify the interlacing & intra
120      * stuff, I can't imagine that these complex rules are worth it. */
121
122     topleft_xy = top_xy - 1;
123     topright_xy= top_xy + 1;
124     left_xy[1] = left_xy[0] = mb_xy-1;
125     left_block = left_block_options[0];
126     if(FRAME_MBAFF){
127         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
128         const int top_pair_xy      = pair_xy     - s->mb_stride;
129         const int topleft_pair_xy  = top_pair_xy - 1;
130         const int topright_pair_xy = top_pair_xy + 1;
131         const int topleft_mb_frame_flag  = !IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
132         const int top_mb_frame_flag      = !IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
133         const int topright_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
134         const int left_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
135         const int curr_mb_frame_flag = !IS_INTERLACED(mb_type);
136         const int bottom = (s->mb_y & 1);
137         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);
138         if (bottom
139                 ? !curr_mb_frame_flag // bottom macroblock
140                 : (!curr_mb_frame_flag && !top_mb_frame_flag) // top macroblock
141                 ) {
142             top_xy -= s->mb_stride;
143         }
144         if (bottom
145                 ? !curr_mb_frame_flag // bottom macroblock
146                 : (!curr_mb_frame_flag && !topleft_mb_frame_flag) // top macroblock
147                 ) {
148             topleft_xy -= s->mb_stride;
149         } else if(bottom && curr_mb_frame_flag && !left_mb_frame_flag) {
150             topleft_xy += s->mb_stride;
151             // take top left mv from the middle of the mb, as opposed to all other modes which use the bottom right partition
152             topleft_partition = 0;
153         }
154         if (bottom
155                 ? !curr_mb_frame_flag // bottom macroblock
156                 : (!curr_mb_frame_flag && !topright_mb_frame_flag) // top macroblock
157                 ) {
158             topright_xy -= s->mb_stride;
159         }
160         if (left_mb_frame_flag != curr_mb_frame_flag) {
161             left_xy[1] = left_xy[0] = pair_xy - 1;
162             if (curr_mb_frame_flag) {
163                 if (bottom) {
164                     left_block = left_block_options[1];
165                 } else {
166                     left_block= left_block_options[2];
167                 }
168             } else {
169                 left_xy[1] += s->mb_stride;
170                 left_block = left_block_options[3];
171             }
172         }
173     }
174
175     h->top_mb_xy = top_xy;
176     h->left_mb_xy[0] = left_xy[0];
177     h->left_mb_xy[1] = left_xy[1];
178     if(for_deblock){
179         topleft_type = 0;
180         topright_type = 0;
181         top_type     = h->slice_table[top_xy     ] < 255 ? s->current_picture.mb_type[top_xy]     : 0;
182         left_type[0] = h->slice_table[left_xy[0] ] < 255 ? s->current_picture.mb_type[left_xy[0]] : 0;
183         left_type[1] = h->slice_table[left_xy[1] ] < 255 ? s->current_picture.mb_type[left_xy[1]] : 0;
184
185         if(MB_MBAFF && !IS_INTRA(mb_type)){
186             int list;
187             for(list=0; list<h->list_count; list++){
188                 //These values where changed for ease of performing MC, we need to change them back
189                 //FIXME maybe we can make MC and loop filter use the same values or prevent
190                 //the MC code from changing ref_cache and rather use a temporary array.
191                 if(USES_LIST(mb_type,list)){
192                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
193                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
194                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
195                     ref += h->b8_stride;
196                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
197                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
198                 }
199             }
200         }
201     }else{
202         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
203         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
204         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
205         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
206         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
207
208     if(IS_INTRA(mb_type)){
209         int type_mask= h->pps.constrained_intra_pred ? IS_INTRA(-1) : -1;
210         h->topleft_samples_available=
211         h->top_samples_available=
212         h->left_samples_available= 0xFFFF;
213         h->topright_samples_available= 0xEEEA;
214
215         if(!(top_type & type_mask)){
216             h->topleft_samples_available= 0xB3FF;
217             h->top_samples_available= 0x33FF;
218             h->topright_samples_available= 0x26EA;
219         }
220         if(IS_INTERLACED(mb_type) != IS_INTERLACED(left_type[0])){
221             if(IS_INTERLACED(mb_type)){
222                 if(!(left_type[0] & type_mask)){
223                     h->topleft_samples_available&= 0xDFFF;
224                     h->left_samples_available&= 0x5FFF;
225                 }
226                 if(!(left_type[1] & type_mask)){
227                     h->topleft_samples_available&= 0xFF5F;
228                     h->left_samples_available&= 0xFF5F;
229                 }
230             }else{
231                 int left_typei = h->slice_table[left_xy[0] + s->mb_stride ] == h->slice_num
232                                 ? s->current_picture.mb_type[left_xy[0] + s->mb_stride] : 0;
233                 assert(left_xy[0] == left_xy[1]);
234                 if(!((left_typei & type_mask) && (left_type[0] & type_mask))){
235                     h->topleft_samples_available&= 0xDF5F;
236                     h->left_samples_available&= 0x5F5F;
237                 }
238             }
239         }else{
240             if(!(left_type[0] & type_mask)){
241                 h->topleft_samples_available&= 0xDF5F;
242                 h->left_samples_available&= 0x5F5F;
243             }
244         }
245
246         if(!(topleft_type & type_mask))
247             h->topleft_samples_available&= 0x7FFF;
248
249         if(!(topright_type & type_mask))
250             h->topright_samples_available&= 0xFBFF;
251
252         if(IS_INTRA4x4(mb_type)){
253             if(IS_INTRA4x4(top_type)){
254                 h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
255                 h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
256                 h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
257                 h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
258             }else{
259                 int pred;
260                 if(!(top_type & type_mask))
261                     pred= -1;
262                 else{
263                     pred= 2;
264                 }
265                 h->intra4x4_pred_mode_cache[4+8*0]=
266                 h->intra4x4_pred_mode_cache[5+8*0]=
267                 h->intra4x4_pred_mode_cache[6+8*0]=
268                 h->intra4x4_pred_mode_cache[7+8*0]= pred;
269             }
270             for(i=0; i<2; i++){
271                 if(IS_INTRA4x4(left_type[i])){
272                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
273                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
274                 }else{
275                     int pred;
276                     if(!(left_type[i] & type_mask))
277                         pred= -1;
278                     else{
279                         pred= 2;
280                     }
281                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
282                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
283                 }
284             }
285         }
286     }
287     }
288
289
290 /*
291 0 . T T. T T T T
292 1 L . .L . . . .
293 2 L . .L . . . .
294 3 . T TL . . . .
295 4 L . .L . . . .
296 5 L . .. . . . .
297 */
298 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
299     if(top_type){
300         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
301         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
302         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
303         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
304
305         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
306         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
307
308         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
309         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
310
311     }else{
312         h->non_zero_count_cache[4+8*0]=
313         h->non_zero_count_cache[5+8*0]=
314         h->non_zero_count_cache[6+8*0]=
315         h->non_zero_count_cache[7+8*0]=
316
317         h->non_zero_count_cache[1+8*0]=
318         h->non_zero_count_cache[2+8*0]=
319
320         h->non_zero_count_cache[1+8*3]=
321         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
322
323     }
324
325     for (i=0; i<2; i++) {
326         if(left_type[i]){
327             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
328             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
329             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
330             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
331         }else{
332             h->non_zero_count_cache[3+8*1 + 2*8*i]=
333             h->non_zero_count_cache[3+8*2 + 2*8*i]=
334             h->non_zero_count_cache[0+8*1 +   8*i]=
335             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
336         }
337     }
338
339     if( h->pps.cabac ) {
340         // top_cbp
341         if(top_type) {
342             h->top_cbp = h->cbp_table[top_xy];
343         } else if(IS_INTRA(mb_type)) {
344             h->top_cbp = 0x1C0;
345         } else {
346             h->top_cbp = 0;
347         }
348         // left_cbp
349         if (left_type[0]) {
350             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
351         } else if(IS_INTRA(mb_type)) {
352             h->left_cbp = 0x1C0;
353         } else {
354             h->left_cbp = 0;
355         }
356         if (left_type[0]) {
357             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
358         }
359         if (left_type[1]) {
360             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
361         }
362     }
363
364 #if 1
365     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
366         int list;
367         for(list=0; list<h->list_count; list++){
368             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
369                 /*if(!h->mv_cache_clean[list]){
370                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
371                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
372                     h->mv_cache_clean[list]= 1;
373                 }*/
374                 continue;
375             }
376             h->mv_cache_clean[list]= 0;
377
378             if(USES_LIST(top_type, list)){
379                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
380                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
381                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
382                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
383                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
384                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
385                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
386                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
387                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
388                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
389             }else{
390                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
391                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
392                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
393                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
394                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
395             }
396
397             for(i=0; i<2; i++){
398                 int cache_idx = scan8[0] - 1 + i*2*8;
399                 if(USES_LIST(left_type[i], list)){
400                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
401                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
402                     *(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]];
403                     *(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]];
404                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
405                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
406                 }else{
407                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
408                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
409                     h->ref_cache[list][cache_idx  ]=
410                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
411                 }
412             }
413
414             if(for_deblock || ((IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred) && !FRAME_MBAFF))
415                 continue;
416
417             if(USES_LIST(topleft_type, list)){
418                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
419                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
420                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
421                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
422             }else{
423                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
424                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
425             }
426
427             if(USES_LIST(topright_type, list)){
428                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
429                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
430                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
431                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
432             }else{
433                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
434                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
435             }
436
437             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
438                 continue;
439
440             h->ref_cache[list][scan8[5 ]+1] =
441             h->ref_cache[list][scan8[7 ]+1] =
442             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
443             h->ref_cache[list][scan8[4 ]] =
444             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
445             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
446             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
447             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
448             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
449             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
450
451             if( h->pps.cabac ) {
452                 /* XXX beurk, Load mvd */
453                 if(USES_LIST(top_type, list)){
454                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
455                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
456                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
457                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
458                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
459                 }else{
460                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
461                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
462                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
463                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
464                 }
465                 if(USES_LIST(left_type[0], list)){
466                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
467                     *(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]];
468                     *(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]];
469                 }else{
470                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
471                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
472                 }
473                 if(USES_LIST(left_type[1], list)){
474                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
475                     *(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]];
476                     *(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]];
477                 }else{
478                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
479                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
480                 }
481                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
482                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
483                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
484                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
485                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
486
487                 if(h->slice_type_nos == FF_B_TYPE){
488                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
489
490                     if(IS_DIRECT(top_type)){
491                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
492                     }else if(IS_8X8(top_type)){
493                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
494                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
495                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
496                     }else{
497                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
498                     }
499
500                     if(IS_DIRECT(left_type[0]))
501                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
502                     else if(IS_8X8(left_type[0]))
503                         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)];
504                     else
505                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
506
507                     if(IS_DIRECT(left_type[1]))
508                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
509                     else if(IS_8X8(left_type[1]))
510                         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)];
511                     else
512                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
513                 }
514             }
515
516             if(FRAME_MBAFF){
517 #define MAP_MVS\
518                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
519                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
520                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
521                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
522                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
523                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
524                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
525                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
526                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
527                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
528                 if(MB_FIELD){
529 #define MAP_F2F(idx, mb_type)\
530                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
531                         h->ref_cache[list][idx] <<= 1;\
532                         h->mv_cache[list][idx][1] /= 2;\
533                         h->mvd_cache[list][idx][1] /= 2;\
534                     }
535                     MAP_MVS
536 #undef MAP_F2F
537                 }else{
538 #define MAP_F2F(idx, mb_type)\
539                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
540                         h->ref_cache[list][idx] >>= 1;\
541                         h->mv_cache[list][idx][1] <<= 1;\
542                         h->mvd_cache[list][idx][1] <<= 1;\
543                     }
544                     MAP_MVS
545 #undef MAP_F2F
546                 }
547             }
548         }
549     }
550 #endif
551
552     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
553 }
554
555 static inline void write_back_intra_pred_mode(H264Context *h){
556     const int mb_xy= h->mb_xy;
557
558     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
559     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
560     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
561     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
562     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
563     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
564     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
565 }
566
567 /**
568  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
569  */
570 static inline int check_intra4x4_pred_mode(H264Context *h){
571     MpegEncContext * const s = &h->s;
572     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
573     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
574     int i;
575
576     if(!(h->top_samples_available&0x8000)){
577         for(i=0; i<4; i++){
578             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
579             if(status<0){
580                 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);
581                 return -1;
582             } else if(status){
583                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
584             }
585         }
586     }
587
588     if((h->left_samples_available&0x8888)!=0x8888){
589         static const int mask[4]={0x8000,0x2000,0x80,0x20};
590         for(i=0; i<4; i++){
591             if(!(h->left_samples_available&mask[i])){
592             int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
593             if(status<0){
594                 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);
595                 return -1;
596             } else if(status){
597                 h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
598             }
599             }
600         }
601     }
602
603     return 0;
604 } //FIXME cleanup like next
605
606 /**
607  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
608  */
609 static inline int check_intra_pred_mode(H264Context *h, int mode){
610     MpegEncContext * const s = &h->s;
611     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
612     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
613
614     if(mode > 6U) {
615         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);
616         return -1;
617     }
618
619     if(!(h->top_samples_available&0x8000)){
620         mode= top[ mode ];
621         if(mode<0){
622             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);
623             return -1;
624         }
625     }
626
627     if((h->left_samples_available&0x8080) != 0x8080){
628         mode= left[ mode ];
629         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
630             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
631         }
632         if(mode<0){
633             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);
634             return -1;
635         }
636     }
637
638     return mode;
639 }
640
641 /**
642  * gets the predicted intra4x4 prediction mode.
643  */
644 static inline int pred_intra_mode(H264Context *h, int n){
645     const int index8= scan8[n];
646     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
647     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
648     const int min= FFMIN(left, top);
649
650     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
651
652     if(min<0) return DC_PRED;
653     else      return min;
654 }
655
656 static inline void write_back_non_zero_count(H264Context *h){
657     const int mb_xy= h->mb_xy;
658
659     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
660     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
661     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
662     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
663     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
664     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
665     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
666
667     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
668     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
669     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
670
671     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
672     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
673     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
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 int get_scale_factor(H264Context * const h, int poc, int poc1, int i){
898     int poc0 = h->ref_list[0][i].poc;
899     int td = av_clip(poc1 - poc0, -128, 127);
900     if(td == 0 || h->ref_list[0][i].long_ref){
901         return 256;
902     }else{
903         int tb = av_clip(poc - poc0, -128, 127);
904         int tx = (16384 + (FFABS(td) >> 1)) / td;
905         return av_clip((tb*tx + 32) >> 6, -1024, 1023);
906     }
907 }
908
909 static inline void direct_dist_scale_factor(H264Context * const h){
910     MpegEncContext * const s = &h->s;
911     const int poc = h->s.current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
912     const int poc1 = h->ref_list[1][0].poc;
913     int i, field;
914     for(field=0; field<2; field++){
915         const int poc  = h->s.current_picture_ptr->field_poc[field];
916         const int poc1 = h->ref_list[1][0].field_poc[field];
917         for(i=0; i < 2*h->ref_count[0]; i++)
918             h->dist_scale_factor_field[field][i^field] = get_scale_factor(h, poc, poc1, i+16);
919     }
920
921     for(i=0; i<h->ref_count[0]; i++){
922         h->dist_scale_factor[i] = get_scale_factor(h, poc, poc1, i);
923     }
924 }
925
926 static void fill_colmap(H264Context *h, int map[2][16+32], int list, int field, int colfield, int mbafi){
927     MpegEncContext * const s = &h->s;
928     Picture * const ref1 = &h->ref_list[1][0];
929     int j, old_ref, rfield;
930     int start= mbafi ? 16                      : 0;
931     int end  = mbafi ? 16+2*h->ref_count[list] : h->ref_count[list];
932     int interl= mbafi || s->picture_structure != PICT_FRAME;
933
934     /* bogus; fills in for missing frames */
935     memset(map[list], 0, sizeof(map[list]));
936
937     for(rfield=0; rfield<2; rfield++){
938         for(old_ref=0; old_ref<ref1->ref_count[colfield][list]; old_ref++){
939             int poc = ref1->ref_poc[colfield][list][old_ref];
940
941             if     (!interl)
942                 poc |= 3;
943             else if( interl && (poc&3) == 3) //FIXME store all MBAFF references so this isnt needed
944                 poc= (poc&~3) + rfield + 1;
945
946             for(j=start; j<end; j++){
947                 if(4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3) == poc){
948                     int cur_ref= mbafi ? (j-16)^field : j;
949                     map[list][2*old_ref + (rfield^field) + 16] = cur_ref;
950                     if(rfield == field)
951                         map[list][old_ref] = cur_ref;
952                     break;
953                 }
954             }
955         }
956     }
957 }
958
959 static inline void direct_ref_list_init(H264Context * const h){
960     MpegEncContext * const s = &h->s;
961     Picture * const ref1 = &h->ref_list[1][0];
962     Picture * const cur = s->current_picture_ptr;
963     int list, j, field;
964     int sidx= (s->picture_structure&1)^1;
965     int ref1sidx= (ref1->reference&1)^1;
966
967     for(list=0; list<2; list++){
968         cur->ref_count[sidx][list] = h->ref_count[list];
969         for(j=0; j<h->ref_count[list]; j++)
970             cur->ref_poc[sidx][list][j] = 4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3);
971     }
972
973     if(s->picture_structure == PICT_FRAME){
974         memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0]));
975         memcpy(cur->ref_poc  [1], cur->ref_poc  [0], sizeof(cur->ref_poc  [0]));
976     }
977
978     cur->mbaff= FRAME_MBAFF;
979
980     if(cur->pict_type != FF_B_TYPE || h->direct_spatial_mv_pred)
981         return;
982
983     for(list=0; list<2; list++){
984         fill_colmap(h, h->map_col_to_list0, list, sidx, ref1sidx, 0);
985         for(field=0; field<2; field++)
986             fill_colmap(h, h->map_col_to_list0_field[field], list, field, field, 1);
987     }
988 }
989
990 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
991     MpegEncContext * const s = &h->s;
992     int b8_stride = h->b8_stride;
993     int b4_stride = h->b_stride;
994     int mb_xy = h->mb_xy;
995     int mb_type_col[2];
996     const int16_t (*l1mv0)[2], (*l1mv1)[2];
997     const int8_t *l1ref0, *l1ref1;
998     const int is_b8x8 = IS_8X8(*mb_type);
999     unsigned int sub_mb_type;
1000     int i8, i4;
1001
1002 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
1003
1004     if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL
1005         if(!IS_INTERLACED(*mb_type)){                    //     AFR/FR    -> AFL/FL
1006             int cur_poc = s->current_picture_ptr->poc;
1007             int *col_poc = h->ref_list[1]->field_poc;
1008             int col_parity = FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc);
1009             mb_xy= s->mb_x + ((s->mb_y&~1) + col_parity)*s->mb_stride;
1010             b8_stride = 0;
1011         }else if(!(s->picture_structure & h->ref_list[1][0].reference) && !h->ref_list[1][0].mbaff){// FL -> FL & differ parity
1012             int fieldoff= 2*(h->ref_list[1][0].reference)-3;
1013             mb_xy += s->mb_stride*fieldoff;
1014         }
1015         goto single_col;
1016     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
1017         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
1018             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
1019             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
1020             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
1021             b8_stride *= 3;
1022             b4_stride *= 6;
1023             //FIXME IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag
1024             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
1025                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
1026                 && !is_b8x8){
1027                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1028                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
1029             }else{
1030                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1031                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1032             }
1033         }else{                                           //     AFR/FR    -> AFR/FR
1034 single_col:
1035             mb_type_col[0] =
1036             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
1037             if(IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag){
1038                 /* FIXME save sub mb types from previous frames (or derive from MVs)
1039                 * so we know exactly what block size to use */
1040                 sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
1041                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1042             }else if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
1043                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1044                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
1045             }else{
1046                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1047                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1048             }
1049         }
1050     }
1051
1052     l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
1053     l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
1054     l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
1055     l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
1056     if(!b8_stride){
1057         if(s->mb_y&1){
1058             l1ref0 += h->b8_stride;
1059             l1ref1 += h->b8_stride;
1060             l1mv0  +=  2*b4_stride;
1061             l1mv1  +=  2*b4_stride;
1062         }
1063     }
1064
1065     if(h->direct_spatial_mv_pred){
1066         int ref[2];
1067         int mv[2][2];
1068         int list;
1069
1070         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
1071
1072         /* ref = min(neighbors) */
1073         for(list=0; list<2; list++){
1074             int refa = h->ref_cache[list][scan8[0] - 1];
1075             int refb = h->ref_cache[list][scan8[0] - 8];
1076             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1077             if(refc == PART_NOT_AVAILABLE)
1078                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1079             ref[list] = FFMIN3((unsigned)refa, (unsigned)refb, (unsigned)refc);
1080             if(ref[list] < 0)
1081                 ref[list] = -1;
1082         }
1083
1084         if(ref[0] < 0 && ref[1] < 0){
1085             ref[0] = ref[1] = 0;
1086             mv[0][0] = mv[0][1] =
1087             mv[1][0] = mv[1][1] = 0;
1088         }else{
1089             for(list=0; list<2; list++){
1090                 if(ref[list] >= 0)
1091                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1092                 else
1093                     mv[list][0] = mv[list][1] = 0;
1094             }
1095         }
1096
1097         if(ref[1] < 0){
1098             if(!is_b8x8)
1099                 *mb_type &= ~MB_TYPE_L1;
1100             sub_mb_type &= ~MB_TYPE_L1;
1101         }else if(ref[0] < 0){
1102             if(!is_b8x8)
1103                 *mb_type &= ~MB_TYPE_L0;
1104             sub_mb_type &= ~MB_TYPE_L0;
1105         }
1106
1107         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1108             for(i8=0; i8<4; i8++){
1109                 int x8 = i8&1;
1110                 int y8 = i8>>1;
1111                 int xy8 = x8+y8*b8_stride;
1112                 int xy4 = 3*x8+y8*b4_stride;
1113                 int a=0, b=0;
1114
1115                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1116                     continue;
1117                 h->sub_mb_type[i8] = sub_mb_type;
1118
1119                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1120                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1121                 if(!IS_INTRA(mb_type_col[y8])
1122                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
1123                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
1124                     if(ref[0] > 0)
1125                         a= pack16to32(mv[0][0],mv[0][1]);
1126                     if(ref[1] > 0)
1127                         b= pack16to32(mv[1][0],mv[1][1]);
1128                 }else{
1129                     a= pack16to32(mv[0][0],mv[0][1]);
1130                     b= pack16to32(mv[1][0],mv[1][1]);
1131                 }
1132                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
1133                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
1134             }
1135         }else if(IS_16X16(*mb_type)){
1136             int a=0, b=0;
1137
1138             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1139             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1140             if(!IS_INTRA(mb_type_col[0])
1141                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1142                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1143                        && (h->x264_build>33 || !h->x264_build)))){
1144                 if(ref[0] > 0)
1145                     a= pack16to32(mv[0][0],mv[0][1]);
1146                 if(ref[1] > 0)
1147                     b= pack16to32(mv[1][0],mv[1][1]);
1148             }else{
1149                 a= pack16to32(mv[0][0],mv[0][1]);
1150                 b= pack16to32(mv[1][0],mv[1][1]);
1151             }
1152             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1153             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1154         }else{
1155             for(i8=0; i8<4; i8++){
1156                 const int x8 = i8&1;
1157                 const int y8 = i8>>1;
1158
1159                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1160                     continue;
1161                 h->sub_mb_type[i8] = sub_mb_type;
1162
1163                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1164                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1165                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1166                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1167
1168                 /* col_zero_flag */
1169                 if(!IS_INTRA(mb_type_col[0]) && (   l1ref0[x8 + y8*b8_stride] == 0
1170                                               || (l1ref0[x8 + y8*b8_stride] < 0 && l1ref1[x8 + y8*b8_stride] == 0
1171                                                   && (h->x264_build>33 || !h->x264_build)))){
1172                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*b8_stride] == 0 ? l1mv0 : l1mv1;
1173                     if(IS_SUB_8X8(sub_mb_type)){
1174                         const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1175                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1176                             if(ref[0] == 0)
1177                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1178                             if(ref[1] == 0)
1179                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1180                         }
1181                     }else
1182                     for(i4=0; i4<4; i4++){
1183                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1184                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1185                             if(ref[0] == 0)
1186                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1187                             if(ref[1] == 0)
1188                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1189                         }
1190                     }
1191                 }
1192             }
1193         }
1194     }else{ /* direct temporal mv pred */
1195         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1196         const int *dist_scale_factor = h->dist_scale_factor;
1197         int ref_offset= 0;
1198
1199         if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
1200             map_col_to_list0[0] = h->map_col_to_list0_field[s->mb_y&1][0];
1201             map_col_to_list0[1] = h->map_col_to_list0_field[s->mb_y&1][1];
1202             dist_scale_factor   =h->dist_scale_factor_field[s->mb_y&1];
1203         }
1204         if(h->ref_list[1][0].mbaff && IS_INTERLACED(mb_type_col[0]))
1205             ref_offset += 16;
1206
1207         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1208             /* FIXME assumes direct_8x8_inference == 1 */
1209             int y_shift  = 2*!IS_INTERLACED(*mb_type);
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_type_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*b8_stride];
1230                 if(ref0 >= 0)
1231                     ref0 = map_col_to_list0[0][ref0 + ref_offset];
1232                 else{
1233                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
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*b4_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         /* one-to-one mv scaling */
1252
1253         if(IS_16X16(*mb_type)){
1254             int ref, mv0, mv1;
1255
1256             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1257             if(IS_INTRA(mb_type_col[0])){
1258                 ref=mv0=mv1=0;
1259             }else{
1260                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset]
1261                                                 : map_col_to_list0[1][l1ref1[0] + ref_offset];
1262                 const int scale = dist_scale_factor[ref0];
1263                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1264                 int mv_l0[2];
1265                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1266                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1267                 ref= ref0;
1268                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1269                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1270             }
1271             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1272             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1273             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1274         }else{
1275             for(i8=0; i8<4; i8++){
1276                 const int x8 = i8&1;
1277                 const int y8 = i8>>1;
1278                 int ref0, scale;
1279                 const int16_t (*l1mv)[2]= l1mv0;
1280
1281                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1282                     continue;
1283                 h->sub_mb_type[i8] = sub_mb_type;
1284                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1285                 if(IS_INTRA(mb_type_col[0])){
1286                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1287                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1288                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1289                     continue;
1290                 }
1291
1292                 ref0 = l1ref0[x8 + y8*b8_stride] + ref_offset;
1293                 if(ref0 >= 0)
1294                     ref0 = map_col_to_list0[0][ref0];
1295                 else{
1296                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
1297                     l1mv= l1mv1;
1298                 }
1299                 scale = dist_scale_factor[ref0];
1300
1301                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1302                 if(IS_SUB_8X8(sub_mb_type)){
1303                     const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1304                     int mx = (scale * mv_col[0] + 128) >> 8;
1305                     int my = (scale * mv_col[1] + 128) >> 8;
1306                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1307                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1308                 }else
1309                 for(i4=0; i4<4; i4++){
1310                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1311                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1312                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1313                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1314                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1315                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1316                 }
1317             }
1318         }
1319     }
1320 }
1321
1322 static inline void write_back_motion(H264Context *h, int mb_type){
1323     MpegEncContext * const s = &h->s;
1324     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1325     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1326     int list;
1327
1328     if(!USES_LIST(mb_type, 0))
1329         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1330
1331     for(list=0; list<h->list_count; list++){
1332         int y;
1333         if(!USES_LIST(mb_type, list))
1334             continue;
1335
1336         for(y=0; y<4; y++){
1337             *(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];
1338             *(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];
1339         }
1340         if( h->pps.cabac ) {
1341             if(IS_SKIP(mb_type))
1342                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1343             else
1344             for(y=0; y<4; y++){
1345                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1346                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1347             }
1348         }
1349
1350         {
1351             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1352             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1353             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1354             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1355             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1356         }
1357     }
1358
1359     if(h->slice_type_nos == FF_B_TYPE && h->pps.cabac){
1360         if(IS_8X8(mb_type)){
1361             uint8_t *direct_table = &h->direct_table[b8_xy];
1362             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1363             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1364             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1365         }
1366     }
1367 }
1368
1369 /**
1370  * Decodes a network abstraction layer unit.
1371  * @param consumed is the number of bytes used as input
1372  * @param length is the length of the array
1373  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1374  * @returns decoded bytes, might be src+1 if no escapes
1375  */
1376 static const uint8_t *decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
1377     int i, si, di;
1378     uint8_t *dst;
1379     int bufidx;
1380
1381 //    src[0]&0x80;                //forbidden bit
1382     h->nal_ref_idc= src[0]>>5;
1383     h->nal_unit_type= src[0]&0x1F;
1384
1385     src++; length--;
1386 #if 0
1387     for(i=0; i<length; i++)
1388         printf("%2X ", src[i]);
1389 #endif
1390     for(i=0; i+1<length; i+=2){
1391         if(src[i]) continue;
1392         if(i>0 && src[i-1]==0) i--;
1393         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1394             if(src[i+2]!=3){
1395                 /* startcode, so we must be past the end */
1396                 length=i;
1397             }
1398             break;
1399         }
1400     }
1401
1402     if(i>=length-1){ //no escaped 0
1403         *dst_length= length;
1404         *consumed= length+1; //+1 for the header
1405         return src;
1406     }
1407
1408     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1409     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length);
1410     dst= h->rbsp_buffer[bufidx];
1411
1412     if (dst == NULL){
1413         return NULL;
1414     }
1415
1416 //printf("decoding esc\n");
1417     si=di=0;
1418     while(si<length){
1419         //remove escapes (very rare 1:2^22)
1420         if(si+2<length && src[si]==0 && src[si+1]==0 && src[si+2]<=3){
1421             if(src[si+2]==3){ //escape
1422                 dst[di++]= 0;
1423                 dst[di++]= 0;
1424                 si+=3;
1425                 continue;
1426             }else //next start code
1427                 break;
1428         }
1429
1430         dst[di++]= src[si++];
1431     }
1432
1433     *dst_length= di;
1434     *consumed= si + 1;//+1 for the header
1435 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1436     return dst;
1437 }
1438
1439 /**
1440  * identifies the exact end of the bitstream
1441  * @return the length of the trailing, or 0 if damaged
1442  */
1443 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src){
1444     int v= *src;
1445     int r;
1446
1447     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1448
1449     for(r=1; r<9; r++){
1450         if(v&1) return r;
1451         v>>=1;
1452     }
1453     return 0;
1454 }
1455
1456 /**
1457  * IDCT transforms the 16 dc values and dequantizes them.
1458  * @param qp quantization parameter
1459  */
1460 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1461 #define stride 16
1462     int i;
1463     int temp[16]; //FIXME check if this is a good idea
1464     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1465     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1466
1467 //memset(block, 64, 2*256);
1468 //return;
1469     for(i=0; i<4; i++){
1470         const int offset= y_offset[i];
1471         const int z0= block[offset+stride*0] + block[offset+stride*4];
1472         const int z1= block[offset+stride*0] - block[offset+stride*4];
1473         const int z2= block[offset+stride*1] - block[offset+stride*5];
1474         const int z3= block[offset+stride*1] + block[offset+stride*5];
1475
1476         temp[4*i+0]= z0+z3;
1477         temp[4*i+1]= z1+z2;
1478         temp[4*i+2]= z1-z2;
1479         temp[4*i+3]= z0-z3;
1480     }
1481
1482     for(i=0; i<4; i++){
1483         const int offset= x_offset[i];
1484         const int z0= temp[4*0+i] + temp[4*2+i];
1485         const int z1= temp[4*0+i] - temp[4*2+i];
1486         const int z2= temp[4*1+i] - temp[4*3+i];
1487         const int z3= temp[4*1+i] + temp[4*3+i];
1488
1489         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
1490         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1491         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1492         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1493     }
1494 }
1495
1496 #if 0
1497 /**
1498  * DCT transforms the 16 dc values.
1499  * @param qp quantization parameter ??? FIXME
1500  */
1501 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1502 //    const int qmul= dequant_coeff[qp][0];
1503     int i;
1504     int temp[16]; //FIXME check if this is a good idea
1505     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1506     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1507
1508     for(i=0; i<4; i++){
1509         const int offset= y_offset[i];
1510         const int z0= block[offset+stride*0] + block[offset+stride*4];
1511         const int z1= block[offset+stride*0] - block[offset+stride*4];
1512         const int z2= block[offset+stride*1] - block[offset+stride*5];
1513         const int z3= block[offset+stride*1] + block[offset+stride*5];
1514
1515         temp[4*i+0]= z0+z3;
1516         temp[4*i+1]= z1+z2;
1517         temp[4*i+2]= z1-z2;
1518         temp[4*i+3]= z0-z3;
1519     }
1520
1521     for(i=0; i<4; i++){
1522         const int offset= x_offset[i];
1523         const int z0= temp[4*0+i] + temp[4*2+i];
1524         const int z1= temp[4*0+i] - temp[4*2+i];
1525         const int z2= temp[4*1+i] - temp[4*3+i];
1526         const int z3= temp[4*1+i] + temp[4*3+i];
1527
1528         block[stride*0 +offset]= (z0 + z3)>>1;
1529         block[stride*2 +offset]= (z1 + z2)>>1;
1530         block[stride*8 +offset]= (z1 - z2)>>1;
1531         block[stride*10+offset]= (z0 - z3)>>1;
1532     }
1533 }
1534 #endif
1535
1536 #undef xStride
1537 #undef stride
1538
1539 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1540     const int stride= 16*2;
1541     const int xStride= 16;
1542     int a,b,c,d,e;
1543
1544     a= block[stride*0 + xStride*0];
1545     b= block[stride*0 + xStride*1];
1546     c= block[stride*1 + xStride*0];
1547     d= block[stride*1 + xStride*1];
1548
1549     e= a-b;
1550     a= a+b;
1551     b= c-d;
1552     c= c+d;
1553
1554     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1555     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1556     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1557     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1558 }
1559
1560 #if 0
1561 static void chroma_dc_dct_c(DCTELEM *block){
1562     const int stride= 16*2;
1563     const int xStride= 16;
1564     int a,b,c,d,e;
1565
1566     a= block[stride*0 + xStride*0];
1567     b= block[stride*0 + xStride*1];
1568     c= block[stride*1 + xStride*0];
1569     d= block[stride*1 + xStride*1];
1570
1571     e= a-b;
1572     a= a+b;
1573     b= c-d;
1574     c= c+d;
1575
1576     block[stride*0 + xStride*0]= (a+c);
1577     block[stride*0 + xStride*1]= (e+b);
1578     block[stride*1 + xStride*0]= (a-c);
1579     block[stride*1 + xStride*1]= (e-b);
1580 }
1581 #endif
1582
1583 /**
1584  * gets the chroma qp.
1585  */
1586 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1587     return h->pps.chroma_qp_table[t][qscale];
1588 }
1589
1590 //FIXME need to check that this does not overflow signed 32 bit for low qp, I am not sure, it's very close
1591 //FIXME check that gcc inlines this (and optimizes intra & separate_dc stuff away)
1592 static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int separate_dc){
1593     int i;
1594     const int * const quant_table= quant_coeff[qscale];
1595     const int bias= intra ? (1<<QUANT_SHIFT)/3 : (1<<QUANT_SHIFT)/6;
1596     const unsigned int threshold1= (1<<QUANT_SHIFT) - bias - 1;
1597     const unsigned int threshold2= (threshold1<<1);
1598     int last_non_zero;
1599
1600     if(separate_dc){
1601         if(qscale<=18){
1602             //avoid overflows
1603             const int dc_bias= intra ? (1<<(QUANT_SHIFT-2))/3 : (1<<(QUANT_SHIFT-2))/6;
1604             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT-2)) - dc_bias - 1;
1605             const unsigned int dc_threshold2= (dc_threshold1<<1);
1606
1607             int level= block[0]*quant_coeff[qscale+18][0];
1608             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1609                 if(level>0){
1610                     level= (dc_bias + level)>>(QUANT_SHIFT-2);
1611                     block[0]= level;
1612                 }else{
1613                     level= (dc_bias - level)>>(QUANT_SHIFT-2);
1614                     block[0]= -level;
1615                 }
1616 //                last_non_zero = i;
1617             }else{
1618                 block[0]=0;
1619             }
1620         }else{
1621             const int dc_bias= intra ? (1<<(QUANT_SHIFT+1))/3 : (1<<(QUANT_SHIFT+1))/6;
1622             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT+1)) - dc_bias - 1;
1623             const unsigned int dc_threshold2= (dc_threshold1<<1);
1624
1625             int level= block[0]*quant_table[0];
1626             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1627                 if(level>0){
1628                     level= (dc_bias + level)>>(QUANT_SHIFT+1);
1629                     block[0]= level;
1630                 }else{
1631                     level= (dc_bias - level)>>(QUANT_SHIFT+1);
1632                     block[0]= -level;
1633                 }
1634 //                last_non_zero = i;
1635             }else{
1636                 block[0]=0;
1637             }
1638         }
1639         last_non_zero= 0;
1640         i=1;
1641     }else{
1642         last_non_zero= -1;
1643         i=0;
1644     }
1645
1646     for(; i<16; i++){
1647         const int j= scantable[i];
1648         int level= block[j]*quant_table[j];
1649
1650 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
1651 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
1652         if(((unsigned)(level+threshold1))>threshold2){
1653             if(level>0){
1654                 level= (bias + level)>>QUANT_SHIFT;
1655                 block[j]= level;
1656             }else{
1657                 level= (bias - level)>>QUANT_SHIFT;
1658                 block[j]= -level;
1659             }
1660             last_non_zero = i;
1661         }else{
1662             block[j]=0;
1663         }
1664     }
1665
1666     return last_non_zero;
1667 }
1668
1669 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1670                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1671                            int src_x_offset, int src_y_offset,
1672                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1673     MpegEncContext * const s = &h->s;
1674     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1675     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1676     const int luma_xy= (mx&3) + ((my&3)<<2);
1677     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1678     uint8_t * src_cb, * src_cr;
1679     int extra_width= h->emu_edge_width;
1680     int extra_height= h->emu_edge_height;
1681     int emu=0;
1682     const int full_mx= mx>>2;
1683     const int full_my= my>>2;
1684     const int pic_width  = 16*s->mb_width;
1685     const int pic_height = 16*s->mb_height >> MB_FIELD;
1686
1687     if(!pic->data[0]) //FIXME this is unacceptable, some sensible error concealment must be done for missing reference frames
1688         return;
1689
1690     if(mx&7) extra_width -= 3;
1691     if(my&7) extra_height -= 3;
1692
1693     if(   full_mx < 0-extra_width
1694        || full_my < 0-extra_height
1695        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1696        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1697         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);
1698             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1699         emu=1;
1700     }
1701
1702     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1703     if(!square){
1704         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1705     }
1706
1707     if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1708
1709     if(MB_FIELD){
1710         // chroma offset when predicting from a field of opposite parity
1711         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1712         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1713     }
1714     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1715     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1716
1717     if(emu){
1718         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);
1719             src_cb= s->edge_emu_buffer;
1720     }
1721     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1722
1723     if(emu){
1724         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);
1725             src_cr= s->edge_emu_buffer;
1726     }
1727     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1728 }
1729
1730 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1731                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1732                            int x_offset, int y_offset,
1733                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1734                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1735                            int list0, int list1){
1736     MpegEncContext * const s = &h->s;
1737     qpel_mc_func *qpix_op=  qpix_put;
1738     h264_chroma_mc_func chroma_op= chroma_put;
1739
1740     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1741     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1742     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1743     x_offset += 8*s->mb_x;
1744     y_offset += 8*(s->mb_y >> MB_FIELD);
1745
1746     if(list0){
1747         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1748         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1749                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1750                            qpix_op, chroma_op);
1751
1752         qpix_op=  qpix_avg;
1753         chroma_op= chroma_avg;
1754     }
1755
1756     if(list1){
1757         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1758         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1759                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1760                            qpix_op, chroma_op);
1761     }
1762 }
1763
1764 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1765                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1766                            int x_offset, int y_offset,
1767                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1768                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1769                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1770                            int list0, int list1){
1771     MpegEncContext * const s = &h->s;
1772
1773     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1774     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1775     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1776     x_offset += 8*s->mb_x;
1777     y_offset += 8*(s->mb_y >> MB_FIELD);
1778
1779     if(list0 && list1){
1780         /* don't optimize for luma-only case, since B-frames usually
1781          * use implicit weights => chroma too. */
1782         uint8_t *tmp_cb = s->obmc_scratchpad;
1783         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1784         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1785         int refn0 = h->ref_cache[0][ scan8[n] ];
1786         int refn1 = h->ref_cache[1][ scan8[n] ];
1787
1788         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1789                     dest_y, dest_cb, dest_cr,
1790                     x_offset, y_offset, qpix_put, chroma_put);
1791         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1792                     tmp_y, tmp_cb, tmp_cr,
1793                     x_offset, y_offset, qpix_put, chroma_put);
1794
1795         if(h->use_weight == 2){
1796             int weight0 = h->implicit_weight[refn0][refn1];
1797             int weight1 = 64 - weight0;
1798             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1799             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1800             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1801         }else{
1802             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1803                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1804                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1805             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1806                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1807                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1808             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1809                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1810                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1811         }
1812     }else{
1813         int list = list1 ? 1 : 0;
1814         int refn = h->ref_cache[list][ scan8[n] ];
1815         Picture *ref= &h->ref_list[list][refn];
1816         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1817                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1818                     qpix_put, chroma_put);
1819
1820         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1821                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1822         if(h->use_weight_chroma){
1823             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1824                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1825             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1826                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1827         }
1828     }
1829 }
1830
1831 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1832                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1833                            int x_offset, int y_offset,
1834                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1835                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1836                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1837                            int list0, int list1){
1838     if((h->use_weight==2 && list0 && list1
1839         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1840        || h->use_weight==1)
1841         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1842                          x_offset, y_offset, qpix_put, chroma_put,
1843                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1844     else
1845         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1846                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1847 }
1848
1849 static inline void prefetch_motion(H264Context *h, int list){
1850     /* fetch pixels for estimated mv 4 macroblocks ahead
1851      * optimized for 64byte cache lines */
1852     MpegEncContext * const s = &h->s;
1853     const int refn = h->ref_cache[list][scan8[0]];
1854     if(refn >= 0){
1855         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1856         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1857         uint8_t **src= h->ref_list[list][refn].data;
1858         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1859         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1860         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1861         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1862     }
1863 }
1864
1865 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1866                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1867                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1868                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1869     MpegEncContext * const s = &h->s;
1870     const int mb_xy= h->mb_xy;
1871     const int mb_type= s->current_picture.mb_type[mb_xy];
1872
1873     assert(IS_INTER(mb_type));
1874
1875     prefetch_motion(h, 0);
1876
1877     if(IS_16X16(mb_type)){
1878         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1879                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1880                 &weight_op[0], &weight_avg[0],
1881                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1882     }else if(IS_16X8(mb_type)){
1883         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1884                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1885                 &weight_op[1], &weight_avg[1],
1886                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1887         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1888                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1889                 &weight_op[1], &weight_avg[1],
1890                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1891     }else if(IS_8X16(mb_type)){
1892         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1893                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1894                 &weight_op[2], &weight_avg[2],
1895                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1896         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1897                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1898                 &weight_op[2], &weight_avg[2],
1899                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1900     }else{
1901         int i;
1902
1903         assert(IS_8X8(mb_type));
1904
1905         for(i=0; i<4; i++){
1906             const int sub_mb_type= h->sub_mb_type[i];
1907             const int n= 4*i;
1908             int x_offset= (i&1)<<2;
1909             int y_offset= (i&2)<<1;
1910
1911             if(IS_SUB_8X8(sub_mb_type)){
1912                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1913                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1914                     &weight_op[3], &weight_avg[3],
1915                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1916             }else if(IS_SUB_8X4(sub_mb_type)){
1917                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1918                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1919                     &weight_op[4], &weight_avg[4],
1920                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1921                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1922                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1923                     &weight_op[4], &weight_avg[4],
1924                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1925             }else if(IS_SUB_4X8(sub_mb_type)){
1926                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1927                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1928                     &weight_op[5], &weight_avg[5],
1929                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1930                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1931                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1932                     &weight_op[5], &weight_avg[5],
1933                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1934             }else{
1935                 int j;
1936                 assert(IS_SUB_4X4(sub_mb_type));
1937                 for(j=0; j<4; j++){
1938                     int sub_x_offset= x_offset + 2*(j&1);
1939                     int sub_y_offset= y_offset +   (j&2);
1940                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1941                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1942                         &weight_op[6], &weight_avg[6],
1943                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1944                 }
1945             }
1946         }
1947     }
1948
1949     prefetch_motion(h, 1);
1950 }
1951
1952 static av_cold void decode_init_vlc(void){
1953     static int done = 0;
1954
1955     if (!done) {
1956         int i;
1957         int offset;
1958         done = 1;
1959
1960         chroma_dc_coeff_token_vlc.table = chroma_dc_coeff_token_vlc_table;
1961         chroma_dc_coeff_token_vlc.table_allocated = chroma_dc_coeff_token_vlc_table_size;
1962         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1963                  &chroma_dc_coeff_token_len [0], 1, 1,
1964                  &chroma_dc_coeff_token_bits[0], 1, 1,
1965                  INIT_VLC_USE_NEW_STATIC);
1966
1967         offset = 0;
1968         for(i=0; i<4; i++){
1969             coeff_token_vlc[i].table = coeff_token_vlc_tables+offset;
1970             coeff_token_vlc[i].table_allocated = coeff_token_vlc_tables_size[i];
1971             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1972                      &coeff_token_len [i][0], 1, 1,
1973                      &coeff_token_bits[i][0], 1, 1,
1974                      INIT_VLC_USE_NEW_STATIC);
1975             offset += coeff_token_vlc_tables_size[i];
1976         }
1977         /*
1978          * This is a one time safety check to make sure that
1979          * the packed static coeff_token_vlc table sizes
1980          * were initialized correctly.
1981          */
1982         assert(offset == FF_ARRAY_ELEMS(coeff_token_vlc_tables));
1983
1984         for(i=0; i<3; i++){
1985             chroma_dc_total_zeros_vlc[i].table = chroma_dc_total_zeros_vlc_tables[i];
1986             chroma_dc_total_zeros_vlc[i].table_allocated = chroma_dc_total_zeros_vlc_tables_size;
1987             init_vlc(&chroma_dc_total_zeros_vlc[i],
1988                      CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1989                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1990                      &chroma_dc_total_zeros_bits[i][0], 1, 1,
1991                      INIT_VLC_USE_NEW_STATIC);
1992         }
1993         for(i=0; i<15; i++){
1994             total_zeros_vlc[i].table = total_zeros_vlc_tables[i];
1995             total_zeros_vlc[i].table_allocated = total_zeros_vlc_tables_size;
1996             init_vlc(&total_zeros_vlc[i],
1997                      TOTAL_ZEROS_VLC_BITS, 16,
1998                      &total_zeros_len [i][0], 1, 1,
1999                      &total_zeros_bits[i][0], 1, 1,
2000                      INIT_VLC_USE_NEW_STATIC);
2001         }
2002
2003         for(i=0; i<6; i++){
2004             run_vlc[i].table = run_vlc_tables[i];
2005             run_vlc[i].table_allocated = run_vlc_tables_size;
2006             init_vlc(&run_vlc[i],
2007                      RUN_VLC_BITS, 7,
2008                      &run_len [i][0], 1, 1,
2009                      &run_bits[i][0], 1, 1,
2010                      INIT_VLC_USE_NEW_STATIC);
2011         }
2012         run7_vlc.table = run7_vlc_table,
2013         run7_vlc.table_allocated = run7_vlc_table_size;
2014         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
2015                  &run_len [6][0], 1, 1,
2016                  &run_bits[6][0], 1, 1,
2017                  INIT_VLC_USE_NEW_STATIC);
2018     }
2019 }
2020
2021 static void free_tables(H264Context *h){
2022     int i;
2023     H264Context *hx;
2024     av_freep(&h->intra4x4_pred_mode);
2025     av_freep(&h->chroma_pred_mode_table);
2026     av_freep(&h->cbp_table);
2027     av_freep(&h->mvd_table[0]);
2028     av_freep(&h->mvd_table[1]);
2029     av_freep(&h->direct_table);
2030     av_freep(&h->non_zero_count);
2031     av_freep(&h->slice_table_base);
2032     h->slice_table= NULL;
2033
2034     av_freep(&h->mb2b_xy);
2035     av_freep(&h->mb2b8_xy);
2036
2037     for(i = 0; i < h->s.avctx->thread_count; i++) {
2038         hx = h->thread_context[i];
2039         if(!hx) continue;
2040         av_freep(&hx->top_borders[1]);
2041         av_freep(&hx->top_borders[0]);
2042         av_freep(&hx->s.obmc_scratchpad);
2043     }
2044 }
2045
2046 static void init_dequant8_coeff_table(H264Context *h){
2047     int i,q,x;
2048     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2049     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2050     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2051
2052     for(i=0; i<2; i++ ){
2053         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2054             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2055             break;
2056         }
2057
2058         for(q=0; q<52; q++){
2059             int shift = div6[q];
2060             int idx = rem6[q];
2061             for(x=0; x<64; x++)
2062                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2063                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2064                     h->pps.scaling_matrix8[i][x]) << shift;
2065         }
2066     }
2067 }
2068
2069 static void init_dequant4_coeff_table(H264Context *h){
2070     int i,j,q,x;
2071     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2072     for(i=0; i<6; i++ ){
2073         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2074         for(j=0; j<i; j++){
2075             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2076                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2077                 break;
2078             }
2079         }
2080         if(j<i)
2081             continue;
2082
2083         for(q=0; q<52; q++){
2084             int shift = div6[q] + 2;
2085             int idx = rem6[q];
2086             for(x=0; x<16; x++)
2087                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2088                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2089                     h->pps.scaling_matrix4[i][x]) << shift;
2090         }
2091     }
2092 }
2093
2094 static void init_dequant_tables(H264Context *h){
2095     int i,x;
2096     init_dequant4_coeff_table(h);
2097     if(h->pps.transform_8x8_mode)
2098         init_dequant8_coeff_table(h);
2099     if(h->sps.transform_bypass){
2100         for(i=0; i<6; i++)
2101             for(x=0; x<16; x++)
2102                 h->dequant4_coeff[i][0][x] = 1<<6;
2103         if(h->pps.transform_8x8_mode)
2104             for(i=0; i<2; i++)
2105                 for(x=0; x<64; x++)
2106                     h->dequant8_coeff[i][0][x] = 1<<6;
2107     }
2108 }
2109
2110
2111 /**
2112  * allocates tables.
2113  * needs width/height
2114  */
2115 static int alloc_tables(H264Context *h){
2116     MpegEncContext * const s = &h->s;
2117     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2118     int x,y;
2119
2120     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2121
2122     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2123     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(uint8_t))
2124     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2125
2126     CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2127     CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2128     CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2129     CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2130
2131     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(uint8_t));
2132     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2133
2134     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2135     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2136     for(y=0; y<s->mb_height; y++){
2137         for(x=0; x<s->mb_width; x++){
2138             const int mb_xy= x + y*s->mb_stride;
2139             const int b_xy = 4*x + 4*y*h->b_stride;
2140             const int b8_xy= 2*x + 2*y*h->b8_stride;
2141
2142             h->mb2b_xy [mb_xy]= b_xy;
2143             h->mb2b8_xy[mb_xy]= b8_xy;
2144         }
2145     }
2146
2147     s->obmc_scratchpad = NULL;
2148
2149     if(!h->dequant4_coeff[0])
2150         init_dequant_tables(h);
2151
2152     return 0;
2153 fail:
2154     free_tables(h);
2155     return -1;
2156 }
2157
2158 /**
2159  * Mimic alloc_tables(), but for every context thread.
2160  */
2161 static void clone_tables(H264Context *dst, H264Context *src){
2162     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2163     dst->non_zero_count           = src->non_zero_count;
2164     dst->slice_table              = src->slice_table;
2165     dst->cbp_table                = src->cbp_table;
2166     dst->mb2b_xy                  = src->mb2b_xy;
2167     dst->mb2b8_xy                 = src->mb2b8_xy;
2168     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2169     dst->mvd_table[0]             = src->mvd_table[0];
2170     dst->mvd_table[1]             = src->mvd_table[1];
2171     dst->direct_table             = src->direct_table;
2172
2173     dst->s.obmc_scratchpad = NULL;
2174     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2175 }
2176
2177 /**
2178  * Init context
2179  * Allocate buffers which are not shared amongst multiple threads.
2180  */
2181 static int context_init(H264Context *h){
2182     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2183     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2184
2185     return 0;
2186 fail:
2187     return -1; // free_tables will clean up for us
2188 }
2189
2190 static av_cold void common_init(H264Context *h){
2191     MpegEncContext * const s = &h->s;
2192
2193     s->width = s->avctx->width;
2194     s->height = s->avctx->height;
2195     s->codec_id= s->avctx->codec->id;
2196
2197     ff_h264_pred_init(&h->hpc, s->codec_id);
2198
2199     h->dequant_coeff_pps= -1;
2200     s->unrestricted_mv=1;
2201     s->decode=1; //FIXME
2202
2203     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2204     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2205 }
2206
2207 static av_cold int decode_init(AVCodecContext *avctx){
2208     H264Context *h= avctx->priv_data;
2209     MpegEncContext * const s = &h->s;
2210
2211     MPV_decode_defaults(s);
2212
2213     s->avctx = avctx;
2214     common_init(h);
2215
2216     s->out_format = FMT_H264;
2217     s->workaround_bugs= avctx->workaround_bugs;
2218
2219     // set defaults
2220 //    s->decode_mb= ff_h263_decode_mb;
2221     s->quarter_sample = 1;
2222     s->low_delay= 1;
2223
2224     if(avctx->codec_id == CODEC_ID_SVQ3)
2225         avctx->pix_fmt= PIX_FMT_YUVJ420P;
2226     else
2227         avctx->pix_fmt= PIX_FMT_YUV420P;
2228
2229     decode_init_vlc();
2230
2231     if(avctx->extradata_size > 0 && avctx->extradata &&
2232        *(char *)avctx->extradata == 1){
2233         h->is_avc = 1;
2234         h->got_avcC = 0;
2235     } else {
2236         h->is_avc = 0;
2237     }
2238
2239     h->thread_context[0] = h;
2240     h->outputed_poc = INT_MIN;
2241     h->prev_poc_msb= 1<<16;
2242     return 0;
2243 }
2244
2245 static int frame_start(H264Context *h){
2246     MpegEncContext * const s = &h->s;
2247     int i;
2248
2249     if(MPV_frame_start(s, s->avctx) < 0)
2250         return -1;
2251     ff_er_frame_start(s);
2252     /*
2253      * MPV_frame_start uses pict_type to derive key_frame.
2254      * This is incorrect for H.264; IDR markings must be used.
2255      * Zero here; IDR markings per slice in frame or fields are ORed in later.
2256      * See decode_nal_units().
2257      */
2258     s->current_picture_ptr->key_frame= 0;
2259
2260     assert(s->linesize && s->uvlinesize);
2261
2262     for(i=0; i<16; i++){
2263         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2264         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2265     }
2266     for(i=0; i<4; i++){
2267         h->block_offset[16+i]=
2268         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2269         h->block_offset[24+16+i]=
2270         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2271     }
2272
2273     /* can't be in alloc_tables because linesize isn't known there.
2274      * FIXME: redo bipred weight to not require extra buffer? */
2275     for(i = 0; i < s->avctx->thread_count; i++)
2276         if(!h->thread_context[i]->s.obmc_scratchpad)
2277             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2278
2279     /* some macroblocks will be accessed before they're available */
2280     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2281         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(uint8_t));
2282
2283 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2284
2285     // We mark the current picture as non-reference after allocating it, so
2286     // that if we break out due to an error it can be released automatically
2287     // in the next MPV_frame_start().
2288     // SVQ3 as well as most other codecs have only last/next/current and thus
2289     // get released even with set reference, besides SVQ3 and others do not
2290     // mark frames as reference later "naturally".
2291     if(s->codec_id != CODEC_ID_SVQ3)
2292         s->current_picture_ptr->reference= 0;
2293
2294     s->current_picture_ptr->field_poc[0]=
2295     s->current_picture_ptr->field_poc[1]= INT_MAX;
2296     assert(s->current_picture_ptr->long_ref==0);
2297
2298     return 0;
2299 }
2300
2301 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){
2302     MpegEncContext * const s = &h->s;
2303     int i;
2304     int step    = 1;
2305     int offset  = 1;
2306     int uvoffset= 1;
2307     int top_idx = 1;
2308     int skiplast= 0;
2309
2310     src_y  -=   linesize;
2311     src_cb -= uvlinesize;
2312     src_cr -= uvlinesize;
2313
2314     if(!simple && FRAME_MBAFF){
2315         if(s->mb_y&1){
2316             offset  = MB_MBAFF ? 1 : 17;
2317             uvoffset= MB_MBAFF ? 1 : 9;
2318             if(!MB_MBAFF){
2319                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 0)= *(uint64_t*)(src_y +  15*linesize);
2320                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 8)= *(uint64_t*)(src_y +8+15*linesize);
2321                 if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2322                     *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+7*uvlinesize);
2323                     *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+7*uvlinesize);
2324                 }
2325             }
2326         }else{
2327             if(!MB_MBAFF){
2328                 h->left_border[0]= h->top_borders[0][s->mb_x][15];
2329                 if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2330                     h->left_border[34   ]= h->top_borders[0][s->mb_x][16+7  ];
2331                     h->left_border[34+18]= h->top_borders[0][s->mb_x][16+8+7];
2332                 }
2333                 skiplast= 1;
2334             }
2335             offset  =
2336             uvoffset=
2337             top_idx = MB_MBAFF ? 0 : 1;
2338         }
2339         step= MB_MBAFF ? 2 : 1;
2340     }
2341
2342     // There are two lines saved, the line above the the top macroblock of a pair,
2343     // and the line above the bottom macroblock
2344     h->left_border[offset]= h->top_borders[top_idx][s->mb_x][15];
2345     for(i=1; i<17 - skiplast; i++){
2346         h->left_border[offset+i*step]= src_y[15+i*  linesize];
2347     }
2348
2349     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2350     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2351
2352     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2353         h->left_border[uvoffset+34   ]= h->top_borders[top_idx][s->mb_x][16+7];
2354         h->left_border[uvoffset+34+18]= h->top_borders[top_idx][s->mb_x][24+7];
2355         for(i=1; i<9 - skiplast; i++){
2356             h->left_border[uvoffset+34   +i*step]= src_cb[7+i*uvlinesize];
2357             h->left_border[uvoffset+34+18+i*step]= src_cr[7+i*uvlinesize];
2358         }
2359         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2360         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2361     }
2362 }
2363
2364 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){
2365     MpegEncContext * const s = &h->s;
2366     int temp8, i;
2367     uint64_t temp64;
2368     int deblock_left;
2369     int deblock_top;
2370     int mb_xy;
2371     int step    = 1;
2372     int offset  = 1;
2373     int uvoffset= 1;
2374     int top_idx = 1;
2375
2376     if(!simple && FRAME_MBAFF){
2377         if(s->mb_y&1){
2378             offset  = MB_MBAFF ? 1 : 17;
2379             uvoffset= MB_MBAFF ? 1 : 9;
2380         }else{
2381             offset  =
2382             uvoffset=
2383             top_idx = MB_MBAFF ? 0 : 1;
2384         }
2385         step= MB_MBAFF ? 2 : 1;
2386     }
2387
2388     if(h->deblocking_filter == 2) {
2389         mb_xy = h->mb_xy;
2390         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2391         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2392     } else {
2393         deblock_left = (s->mb_x > 0);
2394         deblock_top =  (s->mb_y > !!MB_FIELD);
2395     }
2396
2397     src_y  -=   linesize + 1;
2398     src_cb -= uvlinesize + 1;
2399     src_cr -= uvlinesize + 1;
2400
2401 #define XCHG(a,b,t,xchg)\
2402 t= a;\
2403 if(xchg)\
2404     a= b;\
2405 b= t;
2406
2407     if(deblock_left){
2408         for(i = !deblock_top; i<16; i++){
2409             XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, xchg);
2410         }
2411         XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, 1);
2412     }
2413
2414     if(deblock_top){
2415         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2416         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2417         if(s->mb_x+1 < s->mb_width){
2418             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2419         }
2420     }
2421
2422     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2423         if(deblock_left){
2424             for(i = !deblock_top; i<8; i++){
2425                 XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, xchg);
2426                 XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, xchg);
2427             }
2428             XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, 1);
2429             XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, 1);
2430         }
2431         if(deblock_top){
2432             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2433             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2434         }
2435     }
2436 }
2437
2438 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2439     MpegEncContext * const s = &h->s;
2440     const int mb_x= s->mb_x;
2441     const int mb_y= s->mb_y;
2442     const int mb_xy= h->mb_xy;
2443     const int mb_type= s->current_picture.mb_type[mb_xy];
2444     uint8_t  *dest_y, *dest_cb, *dest_cr;
2445     int linesize, uvlinesize /*dct_offset*/;
2446     int i;
2447     int *block_offset = &h->block_offset[0];
2448     const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass), is_h264 = (simple || s->codec_id == CODEC_ID_H264);
2449     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2450     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2451
2452     dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2453     dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2454     dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2455
2456     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2457     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2458
2459     if (!simple && MB_FIELD) {
2460         linesize   = h->mb_linesize   = s->linesize * 2;
2461         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2462         block_offset = &h->block_offset[24];
2463         if(mb_y&1){ //FIXME move out of this function?
2464             dest_y -= s->linesize*15;
2465             dest_cb-= s->uvlinesize*7;
2466             dest_cr-= s->uvlinesize*7;
2467         }
2468         if(FRAME_MBAFF) {
2469             int list;
2470             for(list=0; list<h->list_count; list++){
2471                 if(!USES_LIST(mb_type, list))
2472                     continue;
2473                 if(IS_16X16(mb_type)){
2474                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2475                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2476                 }else{
2477                     for(i=0; i<16; i+=4){
2478                         int ref = h->ref_cache[list][scan8[i]];
2479                         if(ref >= 0)
2480                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2481                     }
2482                 }
2483             }
2484         }
2485     } else {
2486         linesize   = h->mb_linesize   = s->linesize;
2487         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2488 //        dct_offset = s->linesize * 16;
2489     }
2490
2491     if(transform_bypass){
2492         idct_dc_add =
2493         idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2494     }else if(IS_8x8DCT(mb_type)){
2495         idct_dc_add = s->dsp.h264_idct8_dc_add;
2496         idct_add = s->dsp.h264_idct8_add;
2497     }else{
2498         idct_dc_add = s->dsp.h264_idct_dc_add;
2499         idct_add = s->dsp.h264_idct_add;
2500     }
2501
2502     if (!simple && IS_INTRA_PCM(mb_type)) {
2503         for (i=0; i<16; i++) {
2504             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
2505         }
2506         for (i=0; i<8; i++) {
2507             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
2508             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
2509         }
2510     } else {
2511         if(IS_INTRA(mb_type)){
2512             if(h->deblocking_filter)
2513                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2514
2515             if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2516                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2517                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2518             }
2519
2520             if(IS_INTRA4x4(mb_type)){
2521                 if(simple || !s->encoding){
2522                     if(IS_8x8DCT(mb_type)){
2523                         for(i=0; i<16; i+=4){
2524                             uint8_t * const ptr= dest_y + block_offset[i];
2525                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2526                             const int nnz = h->non_zero_count_cache[ scan8[i] ];
2527                             h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2528                                                    (h->topright_samples_available<<i)&0x4000, linesize);
2529                             if(nnz){
2530                                 if(nnz == 1 && h->mb[i*16])
2531                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2532                                 else
2533                                     idct_add(ptr, h->mb + i*16, linesize);
2534                             }
2535                         }
2536                     }else
2537                     for(i=0; i<16; i++){
2538                         uint8_t * const ptr= dest_y + block_offset[i];
2539                         uint8_t *topright;
2540                         const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2541                         int nnz, tr;
2542
2543                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2544                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2545                             assert(mb_y || linesize <= block_offset[i]);
2546                             if(!topright_avail){
2547                                 tr= ptr[3 - linesize]*0x01010101;
2548                                 topright= (uint8_t*) &tr;
2549                             }else
2550                                 topright= ptr + 4 - linesize;
2551                         }else
2552                             topright= NULL;
2553
2554                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2555                         nnz = h->non_zero_count_cache[ scan8[i] ];
2556                         if(nnz){
2557                             if(is_h264){
2558                                 if(nnz == 1 && h->mb[i*16])
2559                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2560                                 else
2561                                     idct_add(ptr, h->mb + i*16, linesize);
2562                             }else
2563                                 svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2564                         }
2565                     }
2566                 }
2567             }else{
2568                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2569                 if(is_h264){
2570                     if(!transform_bypass)
2571                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2572                 }else
2573                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2574             }
2575             if(h->deblocking_filter)
2576                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2577         }else if(is_h264){
2578             hl_motion(h, dest_y, dest_cb, dest_cr,
2579                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2580                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2581                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2582         }
2583
2584
2585         if(!IS_INTRA4x4(mb_type)){
2586             if(is_h264){
2587                 if(IS_INTRA16x16(mb_type)){
2588                     for(i=0; i<16; i++){
2589                         if(h->non_zero_count_cache[ scan8[i] ])
2590                             idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2591                         else if(h->mb[i*16])
2592                             idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2593                     }
2594                 }else{
2595                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2596                     for(i=0; i<16; i+=di){
2597                         int nnz = h->non_zero_count_cache[ scan8[i] ];
2598                         if(nnz){
2599                             if(nnz==1 && h->mb[i*16])
2600                                 idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2601                             else
2602                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2603                         }
2604                     }
2605                 }
2606             }else{
2607                 for(i=0; i<16; i++){
2608                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2609                         uint8_t * const ptr= dest_y + block_offset[i];
2610                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2611                     }
2612                 }
2613             }
2614         }
2615
2616         if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2617             uint8_t *dest[2] = {dest_cb, dest_cr};
2618             if(transform_bypass){
2619                 idct_add = idct_dc_add = s->dsp.add_pixels4;
2620             }else{
2621                 idct_add = s->dsp.h264_idct_add;
2622                 idct_dc_add = s->dsp.h264_idct_dc_add;
2623                 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]);
2624                 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]);
2625             }
2626             if(is_h264){
2627                 for(i=16; i<16+8; i++){
2628                     if(h->non_zero_count_cache[ scan8[i] ])
2629                         idct_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2630                     else if(h->mb[i*16])
2631                         idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2632                 }
2633             }else{
2634                 for(i=16; i<16+8; i++){
2635                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2636                         uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2637                         svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2638                     }
2639                 }
2640             }
2641         }
2642     }
2643     if(h->deblocking_filter) {
2644         backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2645         fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2646         h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2647         h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2648         if (!simple && FRAME_MBAFF) {
2649             filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2650         } else {
2651             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2652         }
2653     }
2654 }
2655
2656 /**
2657  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2658  */
2659 static void hl_decode_mb_simple(H264Context *h){
2660     hl_decode_mb_internal(h, 1);
2661 }
2662
2663 /**
2664  * Process a macroblock; this handles edge cases, such as interlacing.
2665  */
2666 static void av_noinline hl_decode_mb_complex(H264Context *h){
2667     hl_decode_mb_internal(h, 0);
2668 }
2669
2670 static void hl_decode_mb(H264Context *h){
2671     MpegEncContext * const s = &h->s;
2672     const int mb_xy= h->mb_xy;
2673     const int mb_type= s->current_picture.mb_type[mb_xy];
2674     int is_complex = FRAME_MBAFF || MB_FIELD || IS_INTRA_PCM(mb_type) || s->codec_id != CODEC_ID_H264 ||
2675                     (ENABLE_GRAY && (s->flags&CODEC_FLAG_GRAY)) || (ENABLE_H264_ENCODER && s->encoding) || ENABLE_SMALL;
2676
2677     if(ENABLE_H264_ENCODER && !s->decode)
2678         return;
2679
2680     if (is_complex)
2681         hl_decode_mb_complex(h);
2682     else hl_decode_mb_simple(h);
2683 }
2684
2685 static void pic_as_field(Picture *pic, const int parity){
2686     int i;
2687     for (i = 0; i < 4; ++i) {
2688         if (parity == PICT_BOTTOM_FIELD)
2689             pic->data[i] += pic->linesize[i];
2690         pic->reference = parity;
2691         pic->linesize[i] *= 2;
2692     }
2693     pic->poc= pic->field_poc[parity == PICT_BOTTOM_FIELD];
2694 }
2695
2696 static int split_field_copy(Picture *dest, Picture *src,
2697                             int parity, int id_add){
2698     int match = !!(src->reference & parity);
2699
2700     if (match) {
2701         *dest = *src;
2702         if(parity != PICT_FRAME){
2703             pic_as_field(dest, parity);
2704             dest->pic_id *= 2;
2705             dest->pic_id += id_add;
2706         }
2707     }
2708
2709     return match;
2710 }
2711
2712 static int build_def_list(Picture *def, Picture **in, int len, int is_long, int sel){
2713     int i[2]={0};
2714     int index=0;
2715
2716     while(i[0]<len || i[1]<len){
2717         while(i[0]<len && !(in[ i[0] ] && (in[ i[0] ]->reference & sel)))
2718             i[0]++;
2719         while(i[1]<len && !(in[ i[1] ] && (in[ i[1] ]->reference & (sel^3))))
2720             i[1]++;
2721         if(i[0] < len){
2722             in[ i[0] ]->pic_id= is_long ? i[0] : in[ i[0] ]->frame_num;
2723             split_field_copy(&def[index++], in[ i[0]++ ], sel  , 1);
2724         }
2725         if(i[1] < len){
2726             in[ i[1] ]->pic_id= is_long ? i[1] : in[ i[1] ]->frame_num;
2727             split_field_copy(&def[index++], in[ i[1]++ ], sel^3, 0);
2728         }
2729     }
2730
2731     return index;
2732 }
2733
2734 static int add_sorted(Picture **sorted, Picture **src, int len, int limit, int dir){
2735     int i, best_poc;
2736     int out_i= 0;
2737
2738     for(;;){
2739         best_poc= dir ? INT_MIN : INT_MAX;
2740
2741         for(i=0; i<len; i++){
2742             const int poc= src[i]->poc;
2743             if(((poc > limit) ^ dir) && ((poc < best_poc) ^ dir)){
2744                 best_poc= poc;
2745                 sorted[out_i]= src[i];
2746             }
2747         }
2748         if(best_poc == (dir ? INT_MIN : INT_MAX))
2749             break;
2750         limit= sorted[out_i++]->poc - dir;
2751     }
2752     return out_i;
2753 }
2754
2755 /**
2756  * fills the default_ref_list.
2757  */
2758 static int fill_default_ref_list(H264Context *h){
2759     MpegEncContext * const s = &h->s;
2760     int i, len;
2761
2762     if(h->slice_type_nos==FF_B_TYPE){
2763         Picture *sorted[32];
2764         int cur_poc, list;
2765         int lens[2];
2766
2767         if(FIELD_PICTURE)
2768             cur_poc= s->current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
2769         else
2770             cur_poc= s->current_picture_ptr->poc;
2771
2772         for(list= 0; list<2; list++){
2773             len= add_sorted(sorted    , h->short_ref, h->short_ref_count, cur_poc, 1^list);
2774             len+=add_sorted(sorted+len, h->short_ref, h->short_ref_count, cur_poc, 0^list);
2775             assert(len<=32);
2776             len= build_def_list(h->default_ref_list[list]    , sorted     , len, 0, s->picture_structure);
2777             len+=build_def_list(h->default_ref_list[list]+len, h->long_ref, 16 , 1, s->picture_structure);
2778             assert(len<=32);
2779
2780             if(len < h->ref_count[list])
2781                 memset(&h->default_ref_list[list][len], 0, sizeof(Picture)*(h->ref_count[list] - len));
2782             lens[list]= len;
2783         }
2784
2785         if(lens[0] == lens[1] && lens[1] > 1){
2786             for(i=0; h->default_ref_list[0][i].data[0] == h->default_ref_list[1][i].data[0] && i<lens[0]; i++);
2787             if(i == lens[0])
2788                 FFSWAP(Picture, h->default_ref_list[1][0], h->default_ref_list[1][1]);
2789         }
2790     }else{
2791         len = build_def_list(h->default_ref_list[0]    , h->short_ref, h->short_ref_count, 0, s->picture_structure);
2792         len+= build_def_list(h->default_ref_list[0]+len, h-> long_ref, 16                , 1, s->picture_structure);
2793         assert(len <= 32);
2794         if(len < h->ref_count[0])
2795             memset(&h->default_ref_list[0][len], 0, sizeof(Picture)*(h->ref_count[0] - len));
2796     }
2797 #ifdef TRACE
2798     for (i=0; i<h->ref_count[0]; i++) {
2799         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]);
2800     }
2801     if(h->slice_type_nos==FF_B_TYPE){
2802         for (i=0; i<h->ref_count[1]; i++) {
2803             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]);
2804         }
2805     }
2806 #endif
2807     return 0;
2808 }
2809
2810 static void print_short_term(H264Context *h);
2811 static void print_long_term(H264Context *h);
2812
2813 /**
2814  * Extract structure information about the picture described by pic_num in
2815  * the current decoding context (frame or field). Note that pic_num is
2816  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
2817  * @param pic_num picture number for which to extract structure information
2818  * @param structure one of PICT_XXX describing structure of picture
2819  *                      with pic_num
2820  * @return frame number (short term) or long term index of picture
2821  *         described by pic_num
2822  */
2823 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
2824     MpegEncContext * const s = &h->s;
2825
2826     *structure = s->picture_structure;
2827     if(FIELD_PICTURE){
2828         if (!(pic_num & 1))
2829             /* opposite field */
2830             *structure ^= PICT_FRAME;
2831         pic_num >>= 1;
2832     }
2833
2834     return pic_num;
2835 }
2836
2837 static int decode_ref_pic_list_reordering(H264Context *h){
2838     MpegEncContext * const s = &h->s;
2839     int list, index, pic_structure;
2840
2841     print_short_term(h);
2842     print_long_term(h);
2843
2844     for(list=0; list<h->list_count; list++){
2845         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
2846
2847         if(get_bits1(&s->gb)){
2848             int pred= h->curr_pic_num;
2849
2850             for(index=0; ; index++){
2851                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb(&s->gb);
2852                 unsigned int pic_id;
2853                 int i;
2854                 Picture *ref = NULL;
2855
2856                 if(reordering_of_pic_nums_idc==3)
2857                     break;
2858
2859                 if(index >= h->ref_count[list]){
2860                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
2861                     return -1;
2862                 }
2863
2864                 if(reordering_of_pic_nums_idc<3){
2865                     if(reordering_of_pic_nums_idc<2){
2866                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
2867                         int frame_num;
2868
2869                         if(abs_diff_pic_num > h->max_pic_num){
2870                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
2871                             return -1;
2872                         }
2873
2874                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
2875                         else                                pred+= abs_diff_pic_num;
2876                         pred &= h->max_pic_num - 1;
2877
2878                         frame_num = pic_num_extract(h, pred, &pic_structure);
2879
2880                         for(i= h->short_ref_count-1; i>=0; i--){
2881                             ref = h->short_ref[i];
2882                             assert(ref->reference);
2883                             assert(!ref->long_ref);
2884                             if(
2885                                    ref->frame_num == frame_num &&
2886                                    (ref->reference & pic_structure)
2887                               )
2888                                 break;
2889                         }
2890                         if(i>=0)
2891                             ref->pic_id= pred;
2892                     }else{
2893                         int long_idx;
2894                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
2895
2896                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
2897
2898                         if(long_idx>31){
2899                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
2900                             return -1;
2901                         }
2902                         ref = h->long_ref[long_idx];
2903                         assert(!(ref && !ref->reference));
2904                         if(ref && (ref->reference & pic_structure)){
2905                             ref->pic_id= pic_id;
2906                             assert(ref->long_ref);
2907                             i=0;
2908                         }else{
2909                             i=-1;
2910                         }
2911                     }
2912
2913                     if (i < 0) {
2914                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
2915                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
2916                     } else {
2917                         for(i=index; i+1<h->ref_count[list]; i++){
2918                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
2919                                 break;
2920                         }
2921                         for(; i > index; i--){
2922                             h->ref_list[list][i]= h->ref_list[list][i-1];
2923                         }
2924                         h->ref_list[list][index]= *ref;
2925                         if (FIELD_PICTURE){
2926                             pic_as_field(&h->ref_list[list][index], pic_structure);
2927                         }
2928                     }
2929                 }else{
2930                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
2931                     return -1;
2932                 }
2933             }
2934         }
2935     }
2936     for(list=0; list<h->list_count; list++){
2937         for(index= 0; index < h->ref_count[list]; index++){
2938             if(!h->ref_list[list][index].data[0]){
2939                 av_log(h->s.avctx, AV_LOG_ERROR, "Missing reference picture\n");
2940                 h->ref_list[list][index]= s->current_picture; //FIXME this is not a sensible solution
2941             }
2942         }
2943     }
2944
2945     return 0;
2946 }
2947
2948 static void fill_mbaff_ref_list(H264Context *h){
2949     int list, i, j;
2950     for(list=0; list<2; list++){ //FIXME try list_count
2951         for(i=0; i<h->ref_count[list]; i++){
2952             Picture *frame = &h->ref_list[list][i];
2953             Picture *field = &h->ref_list[list][16+2*i];
2954             field[0] = *frame;
2955             for(j=0; j<3; j++)
2956                 field[0].linesize[j] <<= 1;
2957             field[0].reference = PICT_TOP_FIELD;
2958             field[0].poc= field[0].field_poc[0];
2959             field[1] = field[0];
2960             for(j=0; j<3; j++)
2961                 field[1].data[j] += frame->linesize[j];
2962             field[1].reference = PICT_BOTTOM_FIELD;
2963             field[1].poc= field[1].field_poc[1];
2964
2965             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
2966             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
2967             for(j=0; j<2; j++){
2968                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
2969                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
2970             }
2971         }
2972     }
2973     for(j=0; j<h->ref_count[1]; j++){
2974         for(i=0; i<h->ref_count[0]; i++)
2975             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
2976         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
2977         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
2978     }
2979 }
2980
2981 static int pred_weight_table(H264Context *h){
2982     MpegEncContext * const s = &h->s;
2983     int list, i;
2984     int luma_def, chroma_def;
2985
2986     h->use_weight= 0;
2987     h->use_weight_chroma= 0;
2988     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2989     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2990     luma_def = 1<<h->luma_log2_weight_denom;
2991     chroma_def = 1<<h->chroma_log2_weight_denom;
2992
2993     for(list=0; list<2; list++){
2994         for(i=0; i<h->ref_count[list]; i++){
2995             int luma_weight_flag, chroma_weight_flag;
2996
2997             luma_weight_flag= get_bits1(&s->gb);
2998             if(luma_weight_flag){
2999                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3000                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3001                 if(   h->luma_weight[list][i] != luma_def
3002                    || h->luma_offset[list][i] != 0)
3003                     h->use_weight= 1;
3004             }else{
3005                 h->luma_weight[list][i]= luma_def;
3006                 h->luma_offset[list][i]= 0;
3007             }
3008
3009             if(CHROMA){
3010                 chroma_weight_flag= get_bits1(&s->gb);
3011                 if(chroma_weight_flag){
3012                     int j;
3013                     for(j=0; j<2; j++){
3014                         h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3015                         h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3016                         if(   h->chroma_weight[list][i][j] != chroma_def
3017                         || h->chroma_offset[list][i][j] != 0)
3018                             h->use_weight_chroma= 1;
3019                     }
3020                 }else{
3021                     int j;
3022                     for(j=0; j<2; j++){
3023                         h->chroma_weight[list][i][j]= chroma_def;
3024                         h->chroma_offset[list][i][j]= 0;
3025                     }
3026                 }
3027             }
3028         }
3029         if(h->slice_type_nos != FF_B_TYPE) break;
3030     }
3031     h->use_weight= h->use_weight || h->use_weight_chroma;
3032     return 0;
3033 }
3034
3035 static void implicit_weight_table(H264Context *h){
3036     MpegEncContext * const s = &h->s;
3037     int ref0, ref1;
3038     int cur_poc = s->current_picture_ptr->poc;
3039
3040     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3041        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3042         h->use_weight= 0;
3043         h->use_weight_chroma= 0;
3044         return;
3045     }
3046
3047     h->use_weight= 2;
3048     h->use_weight_chroma= 2;
3049     h->luma_log2_weight_denom= 5;
3050     h->chroma_log2_weight_denom= 5;
3051
3052     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3053         int poc0 = h->ref_list[0][ref0].poc;
3054         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3055             int poc1 = h->ref_list[1][ref1].poc;
3056             int td = av_clip(poc1 - poc0, -128, 127);
3057             if(td){
3058                 int tb = av_clip(cur_poc - poc0, -128, 127);
3059                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3060                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3061                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3062                     h->implicit_weight[ref0][ref1] = 32;
3063                 else
3064                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3065             }else
3066                 h->implicit_weight[ref0][ref1] = 32;
3067         }
3068     }
3069 }
3070
3071 /**
3072  * Mark a picture as no longer needed for reference. The refmask
3073  * argument allows unreferencing of individual fields or the whole frame.
3074  * If the picture becomes entirely unreferenced, but is being held for
3075  * display purposes, it is marked as such.
3076  * @param refmask mask of fields to unreference; the mask is bitwise
3077  *                anded with the reference marking of pic
3078  * @return non-zero if pic becomes entirely unreferenced (except possibly
3079  *         for display purposes) zero if one of the fields remains in
3080  *         reference
3081  */
3082 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3083     int i;
3084     if (pic->reference &= refmask) {
3085         return 0;
3086     } else {
3087         for(i = 0; h->delayed_pic[i]; i++)
3088             if(pic == h->delayed_pic[i]){
3089                 pic->reference=DELAYED_PIC_REF;
3090                 break;
3091             }
3092         return 1;
3093     }
3094 }
3095
3096 /**
3097  * instantaneous decoder refresh.
3098  */
3099 static void idr(H264Context *h){
3100     int i;
3101
3102     for(i=0; i<16; i++){
3103         remove_long(h, i, 0);
3104     }
3105     assert(h->long_ref_count==0);
3106
3107     for(i=0; i<h->short_ref_count; i++){
3108         unreference_pic(h, h->short_ref[i], 0);
3109         h->short_ref[i]= NULL;
3110     }
3111     h->short_ref_count=0;
3112     h->prev_frame_num= 0;
3113     h->prev_frame_num_offset= 0;
3114     h->prev_poc_msb=
3115     h->prev_poc_lsb= 0;
3116 }
3117
3118 /* forget old pics after a seek */
3119 static void flush_dpb(AVCodecContext *avctx){
3120     H264Context *h= avctx->priv_data;
3121     int i;
3122     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
3123         if(h->delayed_pic[i])
3124             h->delayed_pic[i]->reference= 0;
3125         h->delayed_pic[i]= NULL;
3126     }
3127     h->outputed_poc= INT_MIN;
3128     idr(h);
3129     if(h->s.current_picture_ptr)
3130         h->s.current_picture_ptr->reference= 0;
3131     h->s.first_field= 0;
3132     ff_mpeg_flush(avctx);
3133 }
3134
3135 /**
3136  * Find a Picture in the short term reference list by frame number.
3137  * @param frame_num frame number to search for
3138  * @param idx the index into h->short_ref where returned picture is found
3139  *            undefined if no picture found.
3140  * @return pointer to the found picture, or NULL if no pic with the provided
3141  *                 frame number is found
3142  */
3143 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3144     MpegEncContext * const s = &h->s;
3145     int i;
3146
3147     for(i=0; i<h->short_ref_count; i++){
3148         Picture *pic= h->short_ref[i];
3149         if(s->avctx->debug&FF_DEBUG_MMCO)
3150             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3151         if(pic->frame_num == frame_num) {
3152             *idx = i;
3153             return pic;
3154         }
3155     }
3156     return NULL;
3157 }
3158
3159 /**
3160  * Remove a picture from the short term reference list by its index in
3161  * that list.  This does no checking on the provided index; it is assumed
3162  * to be valid. Other list entries are shifted down.
3163  * @param i index into h->short_ref of picture to remove.
3164  */
3165 static void remove_short_at_index(H264Context *h, int i){
3166     assert(i >= 0 && i < h->short_ref_count);
3167     h->short_ref[i]= NULL;
3168     if (--h->short_ref_count)
3169         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3170 }
3171
3172 /**
3173  *
3174  * @return the removed picture or NULL if an error occurs
3175  */
3176 static Picture * remove_short(H264Context *h, int frame_num, int ref_mask){
3177     MpegEncContext * const s = &h->s;
3178     Picture *pic;
3179     int i;
3180
3181     if(s->avctx->debug&FF_DEBUG_MMCO)
3182         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3183
3184     pic = find_short(h, frame_num, &i);
3185     if (pic){
3186         if(unreference_pic(h, pic, ref_mask))
3187         remove_short_at_index(h, i);
3188     }
3189
3190     return pic;
3191 }
3192
3193 /**
3194  * Remove a picture from the long term reference list by its index in
3195  * that list.
3196  * @return the removed picture or NULL if an error occurs
3197  */
3198 static Picture * remove_long(H264Context *h, int i, int ref_mask){
3199     Picture *pic;
3200
3201     pic= h->long_ref[i];
3202     if (pic){
3203         if(unreference_pic(h, pic, ref_mask)){
3204             assert(h->long_ref[i]->long_ref == 1);
3205             h->long_ref[i]->long_ref= 0;
3206             h->long_ref[i]= NULL;
3207             h->long_ref_count--;
3208         }
3209     }
3210
3211     return pic;
3212 }
3213
3214 /**
3215  * print short term list
3216  */
3217 static void print_short_term(H264Context *h) {
3218     uint32_t i;
3219     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3220         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3221         for(i=0; i<h->short_ref_count; i++){
3222             Picture *pic= h->short_ref[i];
3223             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3224         }
3225     }
3226 }
3227
3228 /**
3229  * print long term list
3230  */
3231 static void print_long_term(H264Context *h) {
3232     uint32_t i;
3233     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3234         av_log(h->s.avctx, AV_LOG_DEBUG, "long term list:\n");
3235         for(i = 0; i < 16; i++){
3236             Picture *pic= h->long_ref[i];
3237             if (pic) {
3238                 av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3239             }
3240         }
3241     }
3242 }
3243
3244 /**
3245  * Executes the reference picture marking (memory management control operations).
3246  */
3247 static int execute_ref_pic_marking(H264Context *h, MMCO *mmco, int mmco_count){
3248     MpegEncContext * const s = &h->s;
3249     int i, j;
3250     int current_ref_assigned=0;
3251     Picture *pic;
3252
3253     if((s->avctx->debug&FF_DEBUG_MMCO) && mmco_count==0)
3254         av_log(h->s.avctx, AV_LOG_DEBUG, "no mmco here\n");
3255
3256     for(i=0; i<mmco_count; i++){
3257         int structure, frame_num;
3258         if(s->avctx->debug&FF_DEBUG_MMCO)
3259             av_log(h->s.avctx, AV_LOG_DEBUG, "mmco:%d %d %d\n", h->mmco[i].opcode, h->mmco[i].short_pic_num, h->mmco[i].long_arg);
3260
3261         if(   mmco[i].opcode == MMCO_SHORT2UNUSED
3262            || mmco[i].opcode == MMCO_SHORT2LONG){
3263             frame_num = pic_num_extract(h, mmco[i].short_pic_num, &structure);
3264             pic = find_short(h, frame_num, &j);
3265             if(!pic){
3266                 if(mmco[i].opcode != MMCO_SHORT2LONG || !h->long_ref[mmco[i].long_arg]
3267                    || h->long_ref[mmco[i].long_arg]->frame_num != frame_num)
3268                 av_log(h->s.avctx, AV_LOG_ERROR, "mmco: unref short failure\n");
3269                 continue;
3270             }
3271         }
3272
3273         switch(mmco[i].opcode){
3274         case MMCO_SHORT2UNUSED:
3275             if(s->avctx->debug&FF_DEBUG_MMCO)
3276                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref short %d count %d\n", h->mmco[i].short_pic_num, h->short_ref_count);
3277             remove_short(h, frame_num, structure ^ PICT_FRAME);
3278             break;
3279         case MMCO_SHORT2LONG:
3280                 if (h->long_ref[mmco[i].long_arg] != pic)
3281                     remove_long(h, mmco[i].long_arg, 0);
3282
3283                 remove_short_at_index(h, j);
3284                 h->long_ref[ mmco[i].long_arg ]= pic;
3285                 if (h->long_ref[ mmco[i].long_arg ]){
3286                     h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3287                     h->long_ref_count++;
3288                 }
3289             break;
3290         case MMCO_LONG2UNUSED:
3291             j = pic_num_extract(h, mmco[i].long_arg, &structure);
3292             pic = h->long_ref[j];
3293             if (pic) {
3294                 remove_long(h, j, structure ^ PICT_FRAME);
3295             } else if(s->avctx->debug&FF_DEBUG_MMCO)
3296                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref long failure\n");
3297             break;
3298         case MMCO_LONG:
3299                     // Comment below left from previous code as it is an interresting note.
3300                     /* First field in pair is in short term list or
3301                      * at a different long term index.
3302                      * This is not allowed; see 7.4.3.3, notes 2 and 3.
3303                      * Report the problem and keep the pair where it is,
3304                      * and mark this field valid.
3305                      */
3306
3307             if (h->long_ref[mmco[i].long_arg] != s->current_picture_ptr) {
3308                 remove_long(h, mmco[i].long_arg, 0);
3309
3310                 h->long_ref[ mmco[i].long_arg ]= s->current_picture_ptr;
3311                 h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3312                 h->long_ref_count++;
3313             }
3314
3315             s->current_picture_ptr->reference |= s->picture_structure;
3316             current_ref_assigned=1;
3317             break;
3318         case MMCO_SET_MAX_LONG:
3319             assert(mmco[i].long_arg <= 16);
3320             // just remove the long term which index is greater than new max
3321             for(j = mmco[i].long_arg; j<16; j++){
3322                 remove_long(h, j, 0);
3323             }
3324             break;
3325         case MMCO_RESET:
3326             while(h->short_ref_count){
3327                 remove_short(h, h->short_ref[0]->frame_num, 0);
3328             }
3329             for(j = 0; j < 16; j++) {
3330                 remove_long(h, j, 0);
3331             }
3332             s->current_picture_ptr->poc=
3333             s->current_picture_ptr->field_poc[0]=
3334             s->current_picture_ptr->field_poc[1]=
3335             h->poc_lsb=
3336             h->poc_msb=
3337             h->frame_num=
3338             s->current_picture_ptr->frame_num= 0;
3339             break;
3340         default: assert(0);
3341         }
3342     }
3343
3344     if (!current_ref_assigned) {
3345         /* Second field of complementary field pair; the first field of
3346          * which is already referenced. If short referenced, it
3347          * should be first entry in short_ref. If not, it must exist
3348          * in long_ref; trying to put it on the short list here is an
3349          * error in the encoded bit stream (ref: 7.4.3.3, NOTE 2 and 3).
3350          */
3351         if (h->short_ref_count && h->short_ref[0] == s->current_picture_ptr) {
3352             /* Just mark the second field valid */
3353             s->current_picture_ptr->reference = PICT_FRAME;
3354         } else if (s->current_picture_ptr->long_ref) {
3355             av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term reference "
3356                                              "assignment for second field "
3357                                              "in complementary field pair "
3358                                              "(first field is long term)\n");
3359         } else {
3360             pic= remove_short(h, s->current_picture_ptr->frame_num, 0);
3361             if(pic){
3362                 av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term buffer state detected\n");
3363             }
3364
3365             if(h->short_ref_count)
3366                 memmove(&h->short_ref[1], &h->short_ref[0], h->short_ref_count*sizeof(Picture*));
3367
3368             h->short_ref[0]= s->current_picture_ptr;
3369             h->short_ref_count++;
3370             s->current_picture_ptr->reference |= s->picture_structure;
3371         }
3372     }
3373
3374     if (h->long_ref_count + h->short_ref_count > h->sps.ref_frame_count){
3375
3376         /* We have too many reference frames, probably due to corrupted
3377          * stream. Need to discard one frame. Prevents overrun of the
3378          * short_ref and long_ref buffers.
3379          */
3380         av_log(h->s.avctx, AV_LOG_ERROR,
3381                "number of reference frames exceeds max (probably "
3382                "corrupt input), discarding one\n");
3383
3384         if (h->long_ref_count && !h->short_ref_count) {
3385             for (i = 0; i < 16; ++i)
3386                 if (h->long_ref[i])
3387                     break;
3388
3389             assert(i < 16);
3390             remove_long(h, i, 0);
3391         } else {
3392             pic = h->short_ref[h->short_ref_count - 1];
3393             remove_short(h, pic->frame_num, 0);
3394         }
3395     }
3396
3397     print_short_term(h);
3398     print_long_term(h);
3399     return 0;
3400 }
3401
3402 static int decode_ref_pic_marking(H264Context *h, GetBitContext *gb){
3403     MpegEncContext * const s = &h->s;
3404     int i;
3405
3406     h->mmco_index= 0;
3407     if(h->nal_unit_type == NAL_IDR_SLICE){ //FIXME fields
3408         s->broken_link= get_bits1(gb) -1;
3409         if(get_bits1(gb)){
3410             h->mmco[0].opcode= MMCO_LONG;
3411             h->mmco[0].long_arg= 0;
3412             h->mmco_index= 1;
3413         }
3414     }else{
3415         if(get_bits1(gb)){ // adaptive_ref_pic_marking_mode_flag
3416             for(i= 0; i<MAX_MMCO_COUNT; i++) {
3417                 MMCOOpcode opcode= get_ue_golomb(gb);
3418
3419                 h->mmco[i].opcode= opcode;
3420                 if(opcode==MMCO_SHORT2UNUSED || opcode==MMCO_SHORT2LONG){
3421                     h->mmco[i].short_pic_num= (h->curr_pic_num - get_ue_golomb(gb) - 1) & (h->max_pic_num - 1);
3422 /*                    if(h->mmco[i].short_pic_num >= h->short_ref_count || h->short_ref[ h->mmco[i].short_pic_num ] == NULL){
3423                         av_log(s->avctx, AV_LOG_ERROR, "illegal short ref in memory management control operation %d\n", mmco);
3424                         return -1;
3425                     }*/
3426                 }
3427                 if(opcode==MMCO_SHORT2LONG || opcode==MMCO_LONG2UNUSED || opcode==MMCO_LONG || opcode==MMCO_SET_MAX_LONG){
3428                     unsigned int long_arg= get_ue_golomb(gb);
3429                     if(long_arg >= 32 || (long_arg >= 16 && !(opcode == MMCO_LONG2UNUSED && FIELD_PICTURE))){
3430                         av_log(h->s.avctx, AV_LOG_ERROR, "illegal long ref in memory management control operation %d\n", opcode);
3431                         return -1;
3432                     }
3433                     h->mmco[i].long_arg= long_arg;
3434                 }
3435
3436                 if(opcode > (unsigned)MMCO_LONG){
3437                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal memory management control operation %d\n", opcode);
3438                     return -1;
3439                 }
3440                 if(opcode == MMCO_END)