9cdac246058a6efa7e67c740975abe72311af7a1
[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 libavcodec/h264.c
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "internal.h"
29 #include "dsputil.h"
30 #include "avcodec.h"
31 #include "mpegvideo.h"
32 #include "h264.h"
33 #include "h264data.h"
34 #include "h264_parser.h"
35 #include "golomb.h"
36 #include "mathops.h"
37 #include "rectangle.h"
38 #include "vdpau_internal.h"
39
40 #include "cabac.h"
41 #if ARCH_X86
42 #include "x86/h264_i386.h"
43 #endif
44
45 //#undef NDEBUG
46 #include <assert.h>
47
48 /**
49  * Value of Picture.reference when Picture is not a reference picture, but
50  * is held for delayed output.
51  */
52 #define DELAYED_PIC_REF 4
53
54 static VLC coeff_token_vlc[4];
55 static VLC_TYPE coeff_token_vlc_tables[520+332+280+256][2];
56 static const int coeff_token_vlc_tables_size[4]={520,332,280,256};
57
58 static VLC chroma_dc_coeff_token_vlc;
59 static VLC_TYPE chroma_dc_coeff_token_vlc_table[256][2];
60 static const int chroma_dc_coeff_token_vlc_table_size = 256;
61
62 static VLC total_zeros_vlc[15];
63 static VLC_TYPE total_zeros_vlc_tables[15][512][2];
64 static const int total_zeros_vlc_tables_size = 512;
65
66 static VLC chroma_dc_total_zeros_vlc[3];
67 static VLC_TYPE chroma_dc_total_zeros_vlc_tables[3][8][2];
68 static const int chroma_dc_total_zeros_vlc_tables_size = 8;
69
70 static VLC run_vlc[6];
71 static VLC_TYPE run_vlc_tables[6][8][2];
72 static const int run_vlc_tables_size = 8;
73
74 static VLC run7_vlc;
75 static VLC_TYPE run7_vlc_table[96][2];
76 static const int run7_vlc_table_size = 96;
77
78 static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
79 static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
80 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);
81 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);
82 static Picture * remove_long(H264Context *h, int i, int ref_mask);
83
84 static av_always_inline uint32_t pack16to32(int a, int b){
85 #ifdef WORDS_BIGENDIAN
86    return (b&0xFFFF) + (a<<16);
87 #else
88    return (a&0xFFFF) + (b<<16);
89 #endif
90 }
91
92 static const uint8_t rem6[52]={
93 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,
94 };
95
96 static const uint8_t div6[52]={
97 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,
98 };
99
100 static const uint8_t left_block_options[4][8]={
101     {0,1,2,3,7,10,8,11},
102     {2,2,3,3,8,11,8,11},
103     {0,0,1,1,7,10,7,10},
104     {0,2,0,2,7,10,7,10}
105 };
106
107 #define LEVEL_TAB_BITS 8
108 static int8_t cavlc_level_tab[7][1<<LEVEL_TAB_BITS][2];
109
110 static void fill_caches(H264Context *h, int mb_type, int for_deblock){
111     MpegEncContext * const s = &h->s;
112     const int mb_xy= h->mb_xy;
113     int topleft_xy, top_xy, topright_xy, left_xy[2];
114     int topleft_type, top_type, topright_type, left_type[2];
115     const uint8_t * left_block;
116     int topleft_partition= -1;
117     int i;
118
119     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
120
121     //FIXME deblocking could skip the intra and nnz parts.
122     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
123         return;
124
125     /* Wow, what a mess, why didn't they simplify the interlacing & intra
126      * stuff, I can't imagine that these complex rules are worth it. */
127
128     topleft_xy = top_xy - 1;
129     topright_xy= top_xy + 1;
130     left_xy[1] = left_xy[0] = mb_xy-1;
131     left_block = left_block_options[0];
132     if(FRAME_MBAFF){
133         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
134         const int top_pair_xy      = pair_xy     - s->mb_stride;
135         const int topleft_pair_xy  = top_pair_xy - 1;
136         const int topright_pair_xy = top_pair_xy + 1;
137         const int topleft_mb_field_flag  = IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
138         const int top_mb_field_flag      = IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
139         const int topright_mb_field_flag = IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
140         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
141         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
142         const int bottom = (s->mb_y & 1);
143         tprintf(s->avctx, "fill_caches: curr_mb_field_flag:%d, left_mb_field_flag:%d, topleft_mb_field_flag:%d, top_mb_field_flag:%d, topright_mb_field_flag:%d\n", curr_mb_field_flag, left_mb_field_flag, topleft_mb_field_flag, top_mb_field_flag, topright_mb_field_flag);
144
145         if (curr_mb_field_flag && (bottom || top_mb_field_flag)){
146             top_xy -= s->mb_stride;
147         }
148         if (curr_mb_field_flag && (bottom || topleft_mb_field_flag)){
149             topleft_xy -= s->mb_stride;
150         } else if(bottom && !curr_mb_field_flag && left_mb_field_flag) {
151             topleft_xy += s->mb_stride;
152             // take top left mv from the middle of the mb, as opposed to all other modes which use the bottom right partition
153             topleft_partition = 0;
154         }
155         if (curr_mb_field_flag && (bottom || topright_mb_field_flag)){
156             topright_xy -= s->mb_stride;
157         }
158         if (left_mb_field_flag != curr_mb_field_flag) {
159             left_xy[1] = left_xy[0] = pair_xy - 1;
160             if (curr_mb_field_flag) {
161                 left_xy[1] += s->mb_stride;
162                 left_block = left_block_options[3];
163             } else {
164                 left_block= left_block_options[2 - bottom];
165             }
166         }
167     }
168
169     h->top_mb_xy = top_xy;
170     h->left_mb_xy[0] = left_xy[0];
171     h->left_mb_xy[1] = left_xy[1];
172     if(for_deblock){
173         topleft_type = 0;
174         topright_type = 0;
175         top_type     = h->slice_table[top_xy     ] < 0xFFFF ? s->current_picture.mb_type[top_xy]     : 0;
176         left_type[0] = h->slice_table[left_xy[0] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[0]] : 0;
177         left_type[1] = h->slice_table[left_xy[1] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[1]] : 0;
178
179         if(MB_MBAFF && !IS_INTRA(mb_type)){
180             int list;
181             for(list=0; list<h->list_count; list++){
182                 //These values where changed for ease of performing MC, we need to change them back
183                 //FIXME maybe we can make MC and loop filter use the same values or prevent
184                 //the MC code from changing ref_cache and rather use a temporary array.
185                 if(USES_LIST(mb_type,list)){
186                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
187                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
188                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
189                     ref += h->b8_stride;
190                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
191                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
192                 }
193             }
194         }
195     }else{
196         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
197         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
198         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
199         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
200         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
201
202     if(IS_INTRA(mb_type)){
203         int type_mask= h->pps.constrained_intra_pred ? IS_INTRA(-1) : -1;
204         h->topleft_samples_available=
205         h->top_samples_available=
206         h->left_samples_available= 0xFFFF;
207         h->topright_samples_available= 0xEEEA;
208
209         if(!(top_type & type_mask)){
210             h->topleft_samples_available= 0xB3FF;
211             h->top_samples_available= 0x33FF;
212             h->topright_samples_available= 0x26EA;
213         }
214         if(IS_INTERLACED(mb_type) != IS_INTERLACED(left_type[0])){
215             if(IS_INTERLACED(mb_type)){
216                 if(!(left_type[0] & type_mask)){
217                     h->topleft_samples_available&= 0xDFFF;
218                     h->left_samples_available&= 0x5FFF;
219                 }
220                 if(!(left_type[1] & type_mask)){
221                     h->topleft_samples_available&= 0xFF5F;
222                     h->left_samples_available&= 0xFF5F;
223                 }
224             }else{
225                 int left_typei = h->slice_table[left_xy[0] + s->mb_stride ] == h->slice_num
226                                 ? s->current_picture.mb_type[left_xy[0] + s->mb_stride] : 0;
227                 assert(left_xy[0] == left_xy[1]);
228                 if(!((left_typei & type_mask) && (left_type[0] & type_mask))){
229                     h->topleft_samples_available&= 0xDF5F;
230                     h->left_samples_available&= 0x5F5F;
231                 }
232             }
233         }else{
234             if(!(left_type[0] & type_mask)){
235                 h->topleft_samples_available&= 0xDF5F;
236                 h->left_samples_available&= 0x5F5F;
237             }
238         }
239
240         if(!(topleft_type & type_mask))
241             h->topleft_samples_available&= 0x7FFF;
242
243         if(!(topright_type & type_mask))
244             h->topright_samples_available&= 0xFBFF;
245
246         if(IS_INTRA4x4(mb_type)){
247             if(IS_INTRA4x4(top_type)){
248                 h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
249                 h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
250                 h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
251                 h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
252             }else{
253                 int pred;
254                 if(!(top_type & type_mask))
255                     pred= -1;
256                 else{
257                     pred= 2;
258                 }
259                 h->intra4x4_pred_mode_cache[4+8*0]=
260                 h->intra4x4_pred_mode_cache[5+8*0]=
261                 h->intra4x4_pred_mode_cache[6+8*0]=
262                 h->intra4x4_pred_mode_cache[7+8*0]= pred;
263             }
264             for(i=0; i<2; i++){
265                 if(IS_INTRA4x4(left_type[i])){
266                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
267                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
268                 }else{
269                     int pred;
270                     if(!(left_type[i] & type_mask))
271                         pred= -1;
272                     else{
273                         pred= 2;
274                     }
275                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
276                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
277                 }
278             }
279         }
280     }
281     }
282
283
284 /*
285 0 . T T. T T T T
286 1 L . .L . . . .
287 2 L . .L . . . .
288 3 . T TL . . . .
289 4 L . .L . . . .
290 5 L . .. . . . .
291 */
292 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
293     if(top_type){
294         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
295         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
296         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
297         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
298
299         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
300         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
301
302         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
303         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
304
305     }else{
306         h->non_zero_count_cache[4+8*0]=
307         h->non_zero_count_cache[5+8*0]=
308         h->non_zero_count_cache[6+8*0]=
309         h->non_zero_count_cache[7+8*0]=
310
311         h->non_zero_count_cache[1+8*0]=
312         h->non_zero_count_cache[2+8*0]=
313
314         h->non_zero_count_cache[1+8*3]=
315         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
316
317     }
318
319     for (i=0; i<2; i++) {
320         if(left_type[i]){
321             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
322             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
323             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
324             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
325         }else{
326             h->non_zero_count_cache[3+8*1 + 2*8*i]=
327             h->non_zero_count_cache[3+8*2 + 2*8*i]=
328             h->non_zero_count_cache[0+8*1 +   8*i]=
329             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
330         }
331     }
332
333     if( h->pps.cabac ) {
334         // top_cbp
335         if(top_type) {
336             h->top_cbp = h->cbp_table[top_xy];
337         } else if(IS_INTRA(mb_type)) {
338             h->top_cbp = 0x1C0;
339         } else {
340             h->top_cbp = 0;
341         }
342         // left_cbp
343         if (left_type[0]) {
344             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
345         } else if(IS_INTRA(mb_type)) {
346             h->left_cbp = 0x1C0;
347         } else {
348             h->left_cbp = 0;
349         }
350         if (left_type[0]) {
351             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
352         }
353         if (left_type[1]) {
354             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
355         }
356     }
357
358 #if 1
359     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
360         int list;
361         for(list=0; list<h->list_count; list++){
362             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
363                 /*if(!h->mv_cache_clean[list]){
364                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
365                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
366                     h->mv_cache_clean[list]= 1;
367                 }*/
368                 continue;
369             }
370             h->mv_cache_clean[list]= 0;
371
372             if(USES_LIST(top_type, list)){
373                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
374                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
375                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
376                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
377                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
378                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
379                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
380                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
381                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
382                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
383             }else{
384                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
385                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
386                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
387                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
388                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
389             }
390
391             for(i=0; i<2; i++){
392                 int cache_idx = scan8[0] - 1 + i*2*8;
393                 if(USES_LIST(left_type[i], list)){
394                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
395                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
396                     *(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]];
397                     *(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]];
398                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
399                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
400                 }else{
401                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
402                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
403                     h->ref_cache[list][cache_idx  ]=
404                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
405                 }
406             }
407
408             if(for_deblock || ((IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred) && !FRAME_MBAFF))
409                 continue;
410
411             if(USES_LIST(topleft_type, list)){
412                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
413                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
414                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
415                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
416             }else{
417                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
418                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
419             }
420
421             if(USES_LIST(topright_type, list)){
422                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
423                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
424                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
425                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
426             }else{
427                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
428                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
429             }
430
431             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
432                 continue;
433
434             h->ref_cache[list][scan8[5 ]+1] =
435             h->ref_cache[list][scan8[7 ]+1] =
436             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
437             h->ref_cache[list][scan8[4 ]] =
438             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
439             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
440             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
441             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
442             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
443             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
444
445             if( h->pps.cabac ) {
446                 /* XXX beurk, Load mvd */
447                 if(USES_LIST(top_type, list)){
448                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
449                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
450                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
451                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
452                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
453                 }else{
454                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
455                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
456                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
457                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
458                 }
459                 if(USES_LIST(left_type[0], list)){
460                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
461                     *(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]];
462                     *(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]];
463                 }else{
464                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
465                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
466                 }
467                 if(USES_LIST(left_type[1], list)){
468                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
469                     *(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]];
470                     *(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]];
471                 }else{
472                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
473                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
474                 }
475                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
476                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
477                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
478                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
479                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
480
481                 if(h->slice_type_nos == FF_B_TYPE){
482                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
483
484                     if(IS_DIRECT(top_type)){
485                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
486                     }else if(IS_8X8(top_type)){
487                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
488                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
489                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
490                     }else{
491                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
492                     }
493
494                     if(IS_DIRECT(left_type[0]))
495                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
496                     else if(IS_8X8(left_type[0]))
497                         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)];
498                     else
499                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
500
501                     if(IS_DIRECT(left_type[1]))
502                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
503                     else if(IS_8X8(left_type[1]))
504                         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)];
505                     else
506                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
507                 }
508             }
509
510             if(FRAME_MBAFF){
511 #define MAP_MVS\
512                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
513                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
514                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
515                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
516                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
517                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
518                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
519                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
520                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
521                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
522                 if(MB_FIELD){
523 #define MAP_F2F(idx, mb_type)\
524                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
525                         h->ref_cache[list][idx] <<= 1;\
526                         h->mv_cache[list][idx][1] /= 2;\
527                         h->mvd_cache[list][idx][1] /= 2;\
528                     }
529                     MAP_MVS
530 #undef MAP_F2F
531                 }else{
532 #define MAP_F2F(idx, mb_type)\
533                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
534                         h->ref_cache[list][idx] >>= 1;\
535                         h->mv_cache[list][idx][1] <<= 1;\
536                         h->mvd_cache[list][idx][1] <<= 1;\
537                     }
538                     MAP_MVS
539 #undef MAP_F2F
540                 }
541             }
542         }
543     }
544 #endif
545
546     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
547 }
548
549 static inline void write_back_intra_pred_mode(H264Context *h){
550     const int mb_xy= h->mb_xy;
551
552     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
553     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
554     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
555     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
556     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
557     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
558     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
559 }
560
561 /**
562  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
563  */
564 static inline int check_intra4x4_pred_mode(H264Context *h){
565     MpegEncContext * const s = &h->s;
566     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
567     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
568     int i;
569
570     if(!(h->top_samples_available&0x8000)){
571         for(i=0; i<4; i++){
572             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
573             if(status<0){
574                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
575                 return -1;
576             } else if(status){
577                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
578             }
579         }
580     }
581
582     if((h->left_samples_available&0x8888)!=0x8888){
583         static const int mask[4]={0x8000,0x2000,0x80,0x20};
584         for(i=0; i<4; i++){
585             if(!(h->left_samples_available&mask[i])){
586                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
587                 if(status<0){
588                     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);
589                     return -1;
590                 } else if(status){
591                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
592                 }
593             }
594         }
595     }
596
597     return 0;
598 } //FIXME cleanup like next
599
600 /**
601  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
602  */
603 static inline int check_intra_pred_mode(H264Context *h, int mode){
604     MpegEncContext * const s = &h->s;
605     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
606     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
607
608     if(mode > 6U) {
609         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);
610         return -1;
611     }
612
613     if(!(h->top_samples_available&0x8000)){
614         mode= top[ mode ];
615         if(mode<0){
616             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);
617             return -1;
618         }
619     }
620
621     if((h->left_samples_available&0x8080) != 0x8080){
622         mode= left[ mode ];
623         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
624             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
625         }
626         if(mode<0){
627             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);
628             return -1;
629         }
630     }
631
632     return mode;
633 }
634
635 /**
636  * gets the predicted intra4x4 prediction mode.
637  */
638 static inline int pred_intra_mode(H264Context *h, int n){
639     const int index8= scan8[n];
640     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
641     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
642     const int min= FFMIN(left, top);
643
644     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
645
646     if(min<0) return DC_PRED;
647     else      return min;
648 }
649
650 static inline void write_back_non_zero_count(H264Context *h){
651     const int mb_xy= h->mb_xy;
652
653     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
654     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
655     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
656     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
657     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
658     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
659     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
660
661     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
662     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
663     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
664
665     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
666     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
667     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
668 }
669
670 /**
671  * gets the predicted number of non-zero coefficients.
672  * @param n block index
673  */
674 static inline int pred_non_zero_count(H264Context *h, int n){
675     const int index8= scan8[n];
676     const int left= h->non_zero_count_cache[index8 - 1];
677     const int top = h->non_zero_count_cache[index8 - 8];
678     int i= left + top;
679
680     if(i<64) i= (i+1)>>1;
681
682     tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);
683
684     return i&31;
685 }
686
687 static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
688     const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
689     MpegEncContext *s = &h->s;
690
691     /* there is no consistent mapping of mvs to neighboring locations that will
692      * make mbaff happy, so we can't move all this logic to fill_caches */
693     if(FRAME_MBAFF){
694         const uint32_t *mb_types = s->current_picture_ptr->mb_type;
695         const int16_t *mv;
696         *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
697         *C = h->mv_cache[list][scan8[0]-2];
698
699         if(!MB_FIELD
700            && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
701             int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
702             if(IS_INTERLACED(mb_types[topright_xy])){
703 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
704                 const int x4 = X4, y4 = Y4;\
705                 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
706                 if(!USES_LIST(mb_type,list))\
707                     return LIST_NOT_USED;\
708                 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
709                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
710                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
711                 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
712
713                 SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
714             }
715         }
716         if(topright_ref == PART_NOT_AVAILABLE
717            && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
718            && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
719             if(!MB_FIELD
720                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
721                 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
722             }
723             if(MB_FIELD
724                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
725                && i >= scan8[0]+8){
726                 // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
727                 SET_DIAG_MV(/2, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
728             }
729         }
730 #undef SET_DIAG_MV
731     }
732
733     if(topright_ref != PART_NOT_AVAILABLE){
734         *C= h->mv_cache[list][ i - 8 + part_width ];
735         return topright_ref;
736     }else{
737         tprintf(s->avctx, "topright MV not available\n");
738
739         *C= h->mv_cache[list][ i - 8 - 1 ];
740         return h->ref_cache[list][ i - 8 - 1 ];
741     }
742 }
743
744 /**
745  * gets the predicted MV.
746  * @param n the block index
747  * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
748  * @param mx the x component of the predicted motion vector
749  * @param my the y component of the predicted motion vector
750  */
751 static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
752     const int index8= scan8[n];
753     const int top_ref=      h->ref_cache[list][ index8 - 8 ];
754     const int left_ref=     h->ref_cache[list][ index8 - 1 ];
755     const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
756     const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
757     const int16_t * C;
758     int diagonal_ref, match_count;
759
760     assert(part_width==1 || part_width==2 || part_width==4);
761
762 /* mv_cache
763   B . . A T T T T
764   U . . L . . , .
765   U . . L . . . .
766   U . . L . . , .
767   . . . L . . . .
768 */
769
770     diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
771     match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
772     tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
773     if(match_count > 1){ //most common
774         *mx= mid_pred(A[0], B[0], C[0]);
775         *my= mid_pred(A[1], B[1], C[1]);
776     }else if(match_count==1){
777         if(left_ref==ref){
778             *mx= A[0];
779             *my= A[1];
780         }else if(top_ref==ref){
781             *mx= B[0];
782             *my= B[1];
783         }else{
784             *mx= C[0];
785             *my= C[1];
786         }
787     }else{
788         if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
789             *mx= A[0];
790             *my= A[1];
791         }else{
792             *mx= mid_pred(A[0], B[0], C[0]);
793             *my= mid_pred(A[1], B[1], C[1]);
794         }
795     }
796
797     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);
798 }
799
800 /**
801  * gets the directionally predicted 16x8 MV.
802  * @param n the block index
803  * @param mx the x component of the predicted motion vector
804  * @param my the y component of the predicted motion vector
805  */
806 static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
807     if(n==0){
808         const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
809         const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
810
811         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);
812
813         if(top_ref == ref){
814             *mx= B[0];
815             *my= B[1];
816             return;
817         }
818     }else{
819         const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
820         const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];
821
822         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);
823
824         if(left_ref == ref){
825             *mx= A[0];
826             *my= A[1];
827             return;
828         }
829     }
830
831     //RARE
832     pred_motion(h, n, 4, list, ref, mx, my);
833 }
834
835 /**
836  * gets the directionally predicted 8x16 MV.
837  * @param n the block index
838  * @param mx the x component of the predicted motion vector
839  * @param my the y component of the predicted motion vector
840  */
841 static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
842     if(n==0){
843         const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
844         const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];
845
846         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);
847
848         if(left_ref == ref){
849             *mx= A[0];
850             *my= A[1];
851             return;
852         }
853     }else{
854         const int16_t * C;
855         int diagonal_ref;
856
857         diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);
858
859         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);
860
861         if(diagonal_ref == ref){
862             *mx= C[0];
863             *my= C[1];
864             return;
865         }
866     }
867
868     //RARE
869     pred_motion(h, n, 2, list, ref, mx, my);
870 }
871
872 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
873     const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
874     const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];
875
876     tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);
877
878     if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
879        || !( top_ref | *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ])
880        || !(left_ref | *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ])){
881
882         *mx = *my = 0;
883         return;
884     }
885
886     pred_motion(h, 0, 4, 0, 0, mx, my);
887
888     return;
889 }
890
891 static int get_scale_factor(H264Context * const h, int poc, int poc1, int i){
892     int poc0 = h->ref_list[0][i].poc;
893     int td = av_clip(poc1 - poc0, -128, 127);
894     if(td == 0 || h->ref_list[0][i].long_ref){
895         return 256;
896     }else{
897         int tb = av_clip(poc - poc0, -128, 127);
898         int tx = (16384 + (FFABS(td) >> 1)) / td;
899         return av_clip((tb*tx + 32) >> 6, -1024, 1023);
900     }
901 }
902
903 static inline void direct_dist_scale_factor(H264Context * const h){
904     MpegEncContext * const s = &h->s;
905     const int poc = h->s.current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
906     const int poc1 = h->ref_list[1][0].poc;
907     int i, field;
908     for(field=0; field<2; field++){
909         const int poc  = h->s.current_picture_ptr->field_poc[field];
910         const int poc1 = h->ref_list[1][0].field_poc[field];
911         for(i=0; i < 2*h->ref_count[0]; i++)
912             h->dist_scale_factor_field[field][i^field] = get_scale_factor(h, poc, poc1, i+16);
913     }
914
915     for(i=0; i<h->ref_count[0]; i++){
916         h->dist_scale_factor[i] = get_scale_factor(h, poc, poc1, i);
917     }
918 }
919
920 static void fill_colmap(H264Context *h, int map[2][16+32], int list, int field, int colfield, int mbafi){
921     MpegEncContext * const s = &h->s;
922     Picture * const ref1 = &h->ref_list[1][0];
923     int j, old_ref, rfield;
924     int start= mbafi ? 16                      : 0;
925     int end  = mbafi ? 16+2*h->ref_count[list] : h->ref_count[list];
926     int interl= mbafi || s->picture_structure != PICT_FRAME;
927
928     /* bogus; fills in for missing frames */
929     memset(map[list], 0, sizeof(map[list]));
930
931     for(rfield=0; rfield<2; rfield++){
932         for(old_ref=0; old_ref<ref1->ref_count[colfield][list]; old_ref++){
933             int poc = ref1->ref_poc[colfield][list][old_ref];
934
935             if     (!interl)
936                 poc |= 3;
937             else if( interl && (poc&3) == 3) //FIXME store all MBAFF references so this isnt needed
938                 poc= (poc&~3) + rfield + 1;
939
940             for(j=start; j<end; j++){
941                 if(4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3) == poc){
942                     int cur_ref= mbafi ? (j-16)^field : j;
943                     map[list][2*old_ref + (rfield^field) + 16] = cur_ref;
944                     if(rfield == field)
945                         map[list][old_ref] = cur_ref;
946                     break;
947                 }
948             }
949         }
950     }
951 }
952
953 static inline void direct_ref_list_init(H264Context * const h){
954     MpegEncContext * const s = &h->s;
955     Picture * const ref1 = &h->ref_list[1][0];
956     Picture * const cur = s->current_picture_ptr;
957     int list, j, field;
958     int sidx= (s->picture_structure&1)^1;
959     int ref1sidx= (ref1->reference&1)^1;
960
961     for(list=0; list<2; list++){
962         cur->ref_count[sidx][list] = h->ref_count[list];
963         for(j=0; j<h->ref_count[list]; j++)
964             cur->ref_poc[sidx][list][j] = 4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3);
965     }
966
967     if(s->picture_structure == PICT_FRAME){
968         memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0]));
969         memcpy(cur->ref_poc  [1], cur->ref_poc  [0], sizeof(cur->ref_poc  [0]));
970     }
971
972     cur->mbaff= FRAME_MBAFF;
973
974     if(cur->pict_type != FF_B_TYPE || h->direct_spatial_mv_pred)
975         return;
976
977     for(list=0; list<2; list++){
978         fill_colmap(h, h->map_col_to_list0, list, sidx, ref1sidx, 0);
979         for(field=0; field<2; field++)
980             fill_colmap(h, h->map_col_to_list0_field[field], list, field, field, 1);
981     }
982 }
983
984 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
985     MpegEncContext * const s = &h->s;
986     int b8_stride = h->b8_stride;
987     int b4_stride = h->b_stride;
988     int mb_xy = h->mb_xy;
989     int mb_type_col[2];
990     const int16_t (*l1mv0)[2], (*l1mv1)[2];
991     const int8_t *l1ref0, *l1ref1;
992     const int is_b8x8 = IS_8X8(*mb_type);
993     unsigned int sub_mb_type;
994     int i8, i4;
995
996     assert(h->ref_list[1][0].reference&3);
997
998 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
999
1000     if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL
1001         if(!IS_INTERLACED(*mb_type)){                    //     AFR/FR    -> AFL/FL
1002             int cur_poc = s->current_picture_ptr->poc;
1003             int *col_poc = h->ref_list[1]->field_poc;
1004             int col_parity = FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc);
1005             mb_xy= s->mb_x + ((s->mb_y&~1) + col_parity)*s->mb_stride;
1006             b8_stride = 0;
1007         }else if(!(s->picture_structure & h->ref_list[1][0].reference) && !h->ref_list[1][0].mbaff){// FL -> FL & differ parity
1008             int fieldoff= 2*(h->ref_list[1][0].reference)-3;
1009             mb_xy += s->mb_stride*fieldoff;
1010         }
1011         goto single_col;
1012     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
1013         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
1014             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
1015             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
1016             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
1017             b8_stride *= 3;
1018             b4_stride *= 6;
1019             //FIXME IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag
1020             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
1021                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
1022                 && !is_b8x8){
1023                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1024                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
1025             }else{
1026                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1027                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1028             }
1029         }else{                                           //     AFR/FR    -> AFR/FR
1030 single_col:
1031             mb_type_col[0] =
1032             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
1033             if(IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag){
1034                 /* FIXME save sub mb types from previous frames (or derive from MVs)
1035                 * so we know exactly what block size to use */
1036                 sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
1037                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1038             }else if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
1039                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1040                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
1041             }else{
1042                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1043                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1044             }
1045         }
1046     }
1047
1048     l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
1049     l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
1050     l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
1051     l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
1052     if(!b8_stride){
1053         if(s->mb_y&1){
1054             l1ref0 += h->b8_stride;
1055             l1ref1 += h->b8_stride;
1056             l1mv0  +=  2*b4_stride;
1057             l1mv1  +=  2*b4_stride;
1058         }
1059     }
1060
1061     if(h->direct_spatial_mv_pred){
1062         int ref[2];
1063         int mv[2][2];
1064         int list;
1065
1066         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
1067
1068         /* ref = min(neighbors) */
1069         for(list=0; list<2; list++){
1070             int refa = h->ref_cache[list][scan8[0] - 1];
1071             int refb = h->ref_cache[list][scan8[0] - 8];
1072             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1073             if(refc == PART_NOT_AVAILABLE)
1074                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1075             ref[list] = FFMIN3((unsigned)refa, (unsigned)refb, (unsigned)refc);
1076             if(ref[list] < 0)
1077                 ref[list] = -1;
1078         }
1079
1080         if(ref[0] < 0 && ref[1] < 0){
1081             ref[0] = ref[1] = 0;
1082             mv[0][0] = mv[0][1] =
1083             mv[1][0] = mv[1][1] = 0;
1084         }else{
1085             for(list=0; list<2; list++){
1086                 if(ref[list] >= 0)
1087                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1088                 else
1089                     mv[list][0] = mv[list][1] = 0;
1090             }
1091         }
1092
1093         if(ref[1] < 0){
1094             if(!is_b8x8)
1095                 *mb_type &= ~MB_TYPE_L1;
1096             sub_mb_type &= ~MB_TYPE_L1;
1097         }else if(ref[0] < 0){
1098             if(!is_b8x8)
1099                 *mb_type &= ~MB_TYPE_L0;
1100             sub_mb_type &= ~MB_TYPE_L0;
1101         }
1102
1103         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1104             for(i8=0; i8<4; i8++){
1105                 int x8 = i8&1;
1106                 int y8 = i8>>1;
1107                 int xy8 = x8+y8*b8_stride;
1108                 int xy4 = 3*x8+y8*b4_stride;
1109                 int a=0, b=0;
1110
1111                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1112                     continue;
1113                 h->sub_mb_type[i8] = sub_mb_type;
1114
1115                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1116                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1117                 if(!IS_INTRA(mb_type_col[y8])
1118                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
1119                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
1120                     if(ref[0] > 0)
1121                         a= pack16to32(mv[0][0],mv[0][1]);
1122                     if(ref[1] > 0)
1123                         b= pack16to32(mv[1][0],mv[1][1]);
1124                 }else{
1125                     a= pack16to32(mv[0][0],mv[0][1]);
1126                     b= pack16to32(mv[1][0],mv[1][1]);
1127                 }
1128                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
1129                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
1130             }
1131         }else if(IS_16X16(*mb_type)){
1132             int a=0, b=0;
1133
1134             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1135             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1136             if(!IS_INTRA(mb_type_col[0])
1137                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1138                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1139                        && (h->x264_build>33 || !h->x264_build)))){
1140                 if(ref[0] > 0)
1141                     a= pack16to32(mv[0][0],mv[0][1]);
1142                 if(ref[1] > 0)
1143                     b= pack16to32(mv[1][0],mv[1][1]);
1144             }else{
1145                 a= pack16to32(mv[0][0],mv[0][1]);
1146                 b= pack16to32(mv[1][0],mv[1][1]);
1147             }
1148             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1149             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1150         }else{
1151             for(i8=0; i8<4; i8++){
1152                 const int x8 = i8&1;
1153                 const int y8 = i8>>1;
1154
1155                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1156                     continue;
1157                 h->sub_mb_type[i8] = sub_mb_type;
1158
1159                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1160                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1161                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1162                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1163
1164                 /* col_zero_flag */
1165                 if(!IS_INTRA(mb_type_col[0]) && (   l1ref0[x8 + y8*b8_stride] == 0
1166                                               || (l1ref0[x8 + y8*b8_stride] < 0 && l1ref1[x8 + y8*b8_stride] == 0
1167                                                   && (h->x264_build>33 || !h->x264_build)))){
1168                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*b8_stride] == 0 ? l1mv0 : l1mv1;
1169                     if(IS_SUB_8X8(sub_mb_type)){
1170                         const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1171                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1172                             if(ref[0] == 0)
1173                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1174                             if(ref[1] == 0)
1175                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1176                         }
1177                     }else
1178                     for(i4=0; i4<4; i4++){
1179                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1180                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1181                             if(ref[0] == 0)
1182                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1183                             if(ref[1] == 0)
1184                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1185                         }
1186                     }
1187                 }
1188             }
1189         }
1190     }else{ /* direct temporal mv pred */
1191         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1192         const int *dist_scale_factor = h->dist_scale_factor;
1193         int ref_offset= 0;
1194
1195         if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
1196             map_col_to_list0[0] = h->map_col_to_list0_field[s->mb_y&1][0];
1197             map_col_to_list0[1] = h->map_col_to_list0_field[s->mb_y&1][1];
1198             dist_scale_factor   =h->dist_scale_factor_field[s->mb_y&1];
1199         }
1200         if(h->ref_list[1][0].mbaff && IS_INTERLACED(mb_type_col[0]))
1201             ref_offset += 16;
1202
1203         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1204             /* FIXME assumes direct_8x8_inference == 1 */
1205             int y_shift  = 2*!IS_INTERLACED(*mb_type);
1206
1207             for(i8=0; i8<4; i8++){
1208                 const int x8 = i8&1;
1209                 const int y8 = i8>>1;
1210                 int ref0, scale;
1211                 const int16_t (*l1mv)[2]= l1mv0;
1212
1213                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1214                     continue;
1215                 h->sub_mb_type[i8] = sub_mb_type;
1216
1217                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1218                 if(IS_INTRA(mb_type_col[y8])){
1219                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1220                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1221                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1222                     continue;
1223                 }
1224
1225                 ref0 = l1ref0[x8 + y8*b8_stride];
1226                 if(ref0 >= 0)
1227                     ref0 = map_col_to_list0[0][ref0 + ref_offset];
1228                 else{
1229                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
1230                     l1mv= l1mv1;
1231                 }
1232                 scale = dist_scale_factor[ref0];
1233                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1234
1235                 {
1236                     const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride];
1237                     int my_col = (mv_col[1]<<y_shift)/2;
1238                     int mx = (scale * mv_col[0] + 128) >> 8;
1239                     int my = (scale * my_col + 128) >> 8;
1240                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1241                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1242                 }
1243             }
1244             return;
1245         }
1246
1247         /* one-to-one mv scaling */
1248
1249         if(IS_16X16(*mb_type)){
1250             int ref, mv0, mv1;
1251
1252             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1253             if(IS_INTRA(mb_type_col[0])){
1254                 ref=mv0=mv1=0;
1255             }else{
1256                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset]
1257                                                 : map_col_to_list0[1][l1ref1[0] + ref_offset];
1258                 const int scale = dist_scale_factor[ref0];
1259                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1260                 int mv_l0[2];
1261                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1262                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1263                 ref= ref0;
1264                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1265                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1266             }
1267             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1268             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1269             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1270         }else{
1271             for(i8=0; i8<4; i8++){
1272                 const int x8 = i8&1;
1273                 const int y8 = i8>>1;
1274                 int ref0, scale;
1275                 const int16_t (*l1mv)[2]= l1mv0;
1276
1277                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1278                     continue;
1279                 h->sub_mb_type[i8] = sub_mb_type;
1280                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1281                 if(IS_INTRA(mb_type_col[0])){
1282                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1283                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1284                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1285                     continue;
1286                 }
1287
1288                 ref0 = l1ref0[x8 + y8*b8_stride] + ref_offset;
1289                 if(ref0 >= 0)
1290                     ref0 = map_col_to_list0[0][ref0];
1291                 else{
1292                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
1293                     l1mv= l1mv1;
1294                 }
1295                 scale = dist_scale_factor[ref0];
1296
1297                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1298                 if(IS_SUB_8X8(sub_mb_type)){
1299                     const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1300                     int mx = (scale * mv_col[0] + 128) >> 8;
1301                     int my = (scale * mv_col[1] + 128) >> 8;
1302                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1303                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1304                 }else
1305                 for(i4=0; i4<4; i4++){
1306                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1307                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1308                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1309                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1310                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1311                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1312                 }
1313             }
1314         }
1315     }
1316 }
1317
1318 static inline void write_back_motion(H264Context *h, int mb_type){
1319     MpegEncContext * const s = &h->s;
1320     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1321     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1322     int list;
1323
1324     if(!USES_LIST(mb_type, 0))
1325         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1326
1327     for(list=0; list<h->list_count; list++){
1328         int y;
1329         if(!USES_LIST(mb_type, list))
1330             continue;
1331
1332         for(y=0; y<4; y++){
1333             *(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];
1334             *(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];
1335         }
1336         if( h->pps.cabac ) {
1337             if(IS_SKIP(mb_type))
1338                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1339             else
1340             for(y=0; y<4; y++){
1341                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1342                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1343             }
1344         }
1345
1346         {
1347             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1348             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1349             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1350             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1351             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1352         }
1353     }
1354
1355     if(h->slice_type_nos == FF_B_TYPE && h->pps.cabac){
1356         if(IS_8X8(mb_type)){
1357             uint8_t *direct_table = &h->direct_table[b8_xy];
1358             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1359             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1360             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1361         }
1362     }
1363 }
1364
1365 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
1366     int i, si, di;
1367     uint8_t *dst;
1368     int bufidx;
1369
1370 //    src[0]&0x80;                //forbidden bit
1371     h->nal_ref_idc= src[0]>>5;
1372     h->nal_unit_type= src[0]&0x1F;
1373
1374     src++; length--;
1375 #if 0
1376     for(i=0; i<length; i++)
1377         printf("%2X ", src[i]);
1378 #endif
1379
1380 #if HAVE_FAST_UNALIGNED
1381 # if HAVE_FAST_64BIT
1382 #   define RS 7
1383     for(i=0; i+1<length; i+=9){
1384         if(!((~*(const uint64_t*)(src+i) & (*(const uint64_t*)(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
1385 # else
1386 #   define RS 3
1387     for(i=0; i+1<length; i+=5){
1388         if(!((~*(const uint32_t*)(src+i) & (*(const uint32_t*)(src+i) - 0x01000101U)) & 0x80008080U))
1389 # endif
1390             continue;
1391         if(i>0 && !src[i]) i--;
1392         while(src[i]) i++;
1393 #else
1394 #   define RS 0
1395     for(i=0; i+1<length; i+=2){
1396         if(src[i]) continue;
1397         if(i>0 && src[i-1]==0) i--;
1398 #endif
1399         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1400             if(src[i+2]!=3){
1401                 /* startcode, so we must be past the end */
1402                 length=i;
1403             }
1404             break;
1405         }
1406         i-= RS;
1407     }
1408
1409     if(i>=length-1){ //no escaped 0
1410         *dst_length= length;
1411         *consumed= length+1; //+1 for the header
1412         return src;
1413     }
1414
1415     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1416     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
1417     dst= h->rbsp_buffer[bufidx];
1418
1419     if (dst == NULL){
1420         return NULL;
1421     }
1422
1423 //printf("decoding esc\n");
1424     memcpy(dst, src, i);
1425     si=di=i;
1426     while(si+2<length){
1427         //remove escapes (very rare 1:2^22)
1428         if(src[si+2]>3){
1429             dst[di++]= src[si++];
1430             dst[di++]= src[si++];
1431         }else if(src[si]==0 && src[si+1]==0){
1432             if(src[si+2]==3){ //escape
1433                 dst[di++]= 0;
1434                 dst[di++]= 0;
1435                 si+=3;
1436                 continue;
1437             }else //next start code
1438                 goto nsc;
1439         }
1440
1441         dst[di++]= src[si++];
1442     }
1443     while(si<length)
1444         dst[di++]= src[si++];
1445 nsc:
1446
1447     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
1448
1449     *dst_length= di;
1450     *consumed= si + 1;//+1 for the header
1451 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1452     return dst;
1453 }
1454
1455 int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
1456     int v= *src;
1457     int r;
1458
1459     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1460
1461     for(r=1; r<9; r++){
1462         if(v&1) return r;
1463         v>>=1;
1464     }
1465     return 0;
1466 }
1467
1468 /**
1469  * IDCT transforms the 16 dc values and dequantizes them.
1470  * @param qp quantization parameter
1471  */
1472 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1473 #define stride 16
1474     int i;
1475     int temp[16]; //FIXME check if this is a good idea
1476     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1477     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1478
1479 //memset(block, 64, 2*256);
1480 //return;
1481     for(i=0; i<4; i++){
1482         const int offset= y_offset[i];
1483         const int z0= block[offset+stride*0] + block[offset+stride*4];
1484         const int z1= block[offset+stride*0] - block[offset+stride*4];
1485         const int z2= block[offset+stride*1] - block[offset+stride*5];
1486         const int z3= block[offset+stride*1] + block[offset+stride*5];
1487
1488         temp[4*i+0]= z0+z3;
1489         temp[4*i+1]= z1+z2;
1490         temp[4*i+2]= z1-z2;
1491         temp[4*i+3]= z0-z3;
1492     }
1493
1494     for(i=0; i<4; i++){
1495         const int offset= x_offset[i];
1496         const int z0= temp[4*0+i] + temp[4*2+i];
1497         const int z1= temp[4*0+i] - temp[4*2+i];
1498         const int z2= temp[4*1+i] - temp[4*3+i];
1499         const int z3= temp[4*1+i] + temp[4*3+i];
1500
1501         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
1502         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1503         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1504         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1505     }
1506 }
1507
1508 #if 0
1509 /**
1510  * DCT transforms the 16 dc values.
1511  * @param qp quantization parameter ??? FIXME
1512  */
1513 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1514 //    const int qmul= dequant_coeff[qp][0];
1515     int i;
1516     int temp[16]; //FIXME check if this is a good idea
1517     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1518     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1519
1520     for(i=0; i<4; i++){
1521         const int offset= y_offset[i];
1522         const int z0= block[offset+stride*0] + block[offset+stride*4];
1523         const int z1= block[offset+stride*0] - block[offset+stride*4];
1524         const int z2= block[offset+stride*1] - block[offset+stride*5];
1525         const int z3= block[offset+stride*1] + block[offset+stride*5];
1526
1527         temp[4*i+0]= z0+z3;
1528         temp[4*i+1]= z1+z2;
1529         temp[4*i+2]= z1-z2;
1530         temp[4*i+3]= z0-z3;
1531     }
1532
1533     for(i=0; i<4; i++){
1534         const int offset= x_offset[i];
1535         const int z0= temp[4*0+i] + temp[4*2+i];
1536         const int z1= temp[4*0+i] - temp[4*2+i];
1537         const int z2= temp[4*1+i] - temp[4*3+i];
1538         const int z3= temp[4*1+i] + temp[4*3+i];
1539
1540         block[stride*0 +offset]= (z0 + z3)>>1;
1541         block[stride*2 +offset]= (z1 + z2)>>1;
1542         block[stride*8 +offset]= (z1 - z2)>>1;
1543         block[stride*10+offset]= (z0 - z3)>>1;
1544     }
1545 }
1546 #endif
1547
1548 #undef xStride
1549 #undef stride
1550
1551 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1552     const int stride= 16*2;
1553     const int xStride= 16;
1554     int a,b,c,d,e;
1555
1556     a= block[stride*0 + xStride*0];
1557     b= block[stride*0 + xStride*1];
1558     c= block[stride*1 + xStride*0];
1559     d= block[stride*1 + xStride*1];
1560
1561     e= a-b;
1562     a= a+b;
1563     b= c-d;
1564     c= c+d;
1565
1566     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1567     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1568     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1569     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1570 }
1571
1572 #if 0
1573 static void chroma_dc_dct_c(DCTELEM *block){
1574     const int stride= 16*2;
1575     const int xStride= 16;
1576     int a,b,c,d,e;
1577
1578     a= block[stride*0 + xStride*0];
1579     b= block[stride*0 + xStride*1];
1580     c= block[stride*1 + xStride*0];
1581     d= block[stride*1 + xStride*1];
1582
1583     e= a-b;
1584     a= a+b;
1585     b= c-d;
1586     c= c+d;
1587
1588     block[stride*0 + xStride*0]= (a+c);
1589     block[stride*0 + xStride*1]= (e+b);
1590     block[stride*1 + xStride*0]= (a-c);
1591     block[stride*1 + xStride*1]= (e-b);
1592 }
1593 #endif
1594
1595 /**
1596  * gets the chroma qp.
1597  */
1598 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1599     return h->pps.chroma_qp_table[t][qscale];
1600 }
1601
1602 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1603                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1604                            int src_x_offset, int src_y_offset,
1605                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1606     MpegEncContext * const s = &h->s;
1607     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1608     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1609     const int luma_xy= (mx&3) + ((my&3)<<2);
1610     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1611     uint8_t * src_cb, * src_cr;
1612     int extra_width= h->emu_edge_width;
1613     int extra_height= h->emu_edge_height;
1614     int emu=0;
1615     const int full_mx= mx>>2;
1616     const int full_my= my>>2;
1617     const int pic_width  = 16*s->mb_width;
1618     const int pic_height = 16*s->mb_height >> MB_FIELD;
1619
1620     if(mx&7) extra_width -= 3;
1621     if(my&7) extra_height -= 3;
1622
1623     if(   full_mx < 0-extra_width
1624        || full_my < 0-extra_height
1625        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1626        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1627         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);
1628             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1629         emu=1;
1630     }
1631
1632     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1633     if(!square){
1634         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1635     }
1636
1637     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1638
1639     if(MB_FIELD){
1640         // chroma offset when predicting from a field of opposite parity
1641         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1642         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1643     }
1644     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1645     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1646
1647     if(emu){
1648         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);
1649             src_cb= s->edge_emu_buffer;
1650     }
1651     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1652
1653     if(emu){
1654         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);
1655             src_cr= s->edge_emu_buffer;
1656     }
1657     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1658 }
1659
1660 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1661                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1662                            int x_offset, int y_offset,
1663                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1664                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1665                            int list0, int list1){
1666     MpegEncContext * const s = &h->s;
1667     qpel_mc_func *qpix_op=  qpix_put;
1668     h264_chroma_mc_func chroma_op= chroma_put;
1669
1670     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1671     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1672     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1673     x_offset += 8*s->mb_x;
1674     y_offset += 8*(s->mb_y >> MB_FIELD);
1675
1676     if(list0){
1677         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1678         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1679                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1680                            qpix_op, chroma_op);
1681
1682         qpix_op=  qpix_avg;
1683         chroma_op= chroma_avg;
1684     }
1685
1686     if(list1){
1687         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1688         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1689                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1690                            qpix_op, chroma_op);
1691     }
1692 }
1693
1694 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1695                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1696                            int x_offset, int y_offset,
1697                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1698                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1699                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1700                            int list0, int list1){
1701     MpegEncContext * const s = &h->s;
1702
1703     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1704     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1705     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1706     x_offset += 8*s->mb_x;
1707     y_offset += 8*(s->mb_y >> MB_FIELD);
1708
1709     if(list0 && list1){
1710         /* don't optimize for luma-only case, since B-frames usually
1711          * use implicit weights => chroma too. */
1712         uint8_t *tmp_cb = s->obmc_scratchpad;
1713         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1714         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1715         int refn0 = h->ref_cache[0][ scan8[n] ];
1716         int refn1 = h->ref_cache[1][ scan8[n] ];
1717
1718         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1719                     dest_y, dest_cb, dest_cr,
1720                     x_offset, y_offset, qpix_put, chroma_put);
1721         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1722                     tmp_y, tmp_cb, tmp_cr,
1723                     x_offset, y_offset, qpix_put, chroma_put);
1724
1725         if(h->use_weight == 2){
1726             int weight0 = h->implicit_weight[refn0][refn1];
1727             int weight1 = 64 - weight0;
1728             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1729             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1730             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1731         }else{
1732             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1733                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1734                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1735             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1736                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1737                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1738             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1739                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1740                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1741         }
1742     }else{
1743         int list = list1 ? 1 : 0;
1744         int refn = h->ref_cache[list][ scan8[n] ];
1745         Picture *ref= &h->ref_list[list][refn];
1746         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1747                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1748                     qpix_put, chroma_put);
1749
1750         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1751                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1752         if(h->use_weight_chroma){
1753             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1754                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1755             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1756                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1757         }
1758     }
1759 }
1760
1761 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1762                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1763                            int x_offset, int y_offset,
1764                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1765                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1766                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1767                            int list0, int list1){
1768     if((h->use_weight==2 && list0 && list1
1769         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1770        || h->use_weight==1)
1771         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1772                          x_offset, y_offset, qpix_put, chroma_put,
1773                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1774     else
1775         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1776                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1777 }
1778
1779 static inline void prefetch_motion(H264Context *h, int list){
1780     /* fetch pixels for estimated mv 4 macroblocks ahead
1781      * optimized for 64byte cache lines */
1782     MpegEncContext * const s = &h->s;
1783     const int refn = h->ref_cache[list][scan8[0]];
1784     if(refn >= 0){
1785         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1786         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1787         uint8_t **src= h->ref_list[list][refn].data;
1788         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1789         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1790         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1791         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1792     }
1793 }
1794
1795 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1796                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1797                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1798                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1799     MpegEncContext * const s = &h->s;
1800     const int mb_xy= h->mb_xy;
1801     const int mb_type= s->current_picture.mb_type[mb_xy];
1802
1803     assert(IS_INTER(mb_type));
1804
1805     prefetch_motion(h, 0);
1806
1807     if(IS_16X16(mb_type)){
1808         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1809                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1810                 &weight_op[0], &weight_avg[0],
1811                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1812     }else if(IS_16X8(mb_type)){
1813         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1814                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1815                 &weight_op[1], &weight_avg[1],
1816                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1817         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1818                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1819                 &weight_op[1], &weight_avg[1],
1820                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1821     }else if(IS_8X16(mb_type)){
1822         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1823                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1824                 &weight_op[2], &weight_avg[2],
1825                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1826         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1827                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1828                 &weight_op[2], &weight_avg[2],
1829                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1830     }else{
1831         int i;
1832
1833         assert(IS_8X8(mb_type));
1834
1835         for(i=0; i<4; i++){
1836             const int sub_mb_type= h->sub_mb_type[i];
1837             const int n= 4*i;
1838             int x_offset= (i&1)<<2;
1839             int y_offset= (i&2)<<1;
1840
1841             if(IS_SUB_8X8(sub_mb_type)){
1842                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1843                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1844                     &weight_op[3], &weight_avg[3],
1845                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1846             }else if(IS_SUB_8X4(sub_mb_type)){
1847                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1848                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1849                     &weight_op[4], &weight_avg[4],
1850                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1851                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1852                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1853                     &weight_op[4], &weight_avg[4],
1854                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1855             }else if(IS_SUB_4X8(sub_mb_type)){
1856                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1857                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1858                     &weight_op[5], &weight_avg[5],
1859                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1860                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1861                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1862                     &weight_op[5], &weight_avg[5],
1863                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1864             }else{
1865                 int j;
1866                 assert(IS_SUB_4X4(sub_mb_type));
1867                 for(j=0; j<4; j++){
1868                     int sub_x_offset= x_offset + 2*(j&1);
1869                     int sub_y_offset= y_offset +   (j&2);
1870                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1871                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1872                         &weight_op[6], &weight_avg[6],
1873                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1874                 }
1875             }
1876         }
1877     }
1878
1879     prefetch_motion(h, 1);
1880 }
1881
1882 static av_cold void init_cavlc_level_tab(void){
1883     int suffix_length, mask;
1884     unsigned int i;
1885
1886     for(suffix_length=0; suffix_length<7; suffix_length++){
1887         for(i=0; i<(1<<LEVEL_TAB_BITS); i++){
1888             int prefix= LEVEL_TAB_BITS - av_log2(2*i);
1889             int level_code= (prefix<<suffix_length) + (i>>(LEVEL_TAB_BITS-prefix-1-suffix_length)) - (1<<suffix_length);
1890
1891             mask= -(level_code&1);
1892             level_code= (((2+level_code)>>1) ^ mask) - mask;
1893             if(prefix + 1 + suffix_length <= LEVEL_TAB_BITS){
1894                 cavlc_level_tab[suffix_length][i][0]= level_code;
1895                 cavlc_level_tab[suffix_length][i][1]= prefix + 1 + suffix_length;
1896             }else if(prefix + 1 <= LEVEL_TAB_BITS){
1897                 cavlc_level_tab[suffix_length][i][0]= prefix+100;
1898                 cavlc_level_tab[suffix_length][i][1]= prefix + 1;
1899             }else{
1900                 cavlc_level_tab[suffix_length][i][0]= LEVEL_TAB_BITS+100;
1901                 cavlc_level_tab[suffix_length][i][1]= LEVEL_TAB_BITS;
1902             }
1903         }
1904     }
1905 }
1906
1907 static av_cold void decode_init_vlc(void){
1908     static int done = 0;
1909
1910     if (!done) {
1911         int i;
1912         int offset;
1913         done = 1;
1914
1915         chroma_dc_coeff_token_vlc.table = chroma_dc_coeff_token_vlc_table;
1916         chroma_dc_coeff_token_vlc.table_allocated = chroma_dc_coeff_token_vlc_table_size;
1917         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1918                  &chroma_dc_coeff_token_len [0], 1, 1,
1919                  &chroma_dc_coeff_token_bits[0], 1, 1,
1920                  INIT_VLC_USE_NEW_STATIC);
1921
1922         offset = 0;
1923         for(i=0; i<4; i++){
1924             coeff_token_vlc[i].table = coeff_token_vlc_tables+offset;
1925             coeff_token_vlc[i].table_allocated = coeff_token_vlc_tables_size[i];
1926             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1927                      &coeff_token_len [i][0], 1, 1,
1928                      &coeff_token_bits[i][0], 1, 1,
1929                      INIT_VLC_USE_NEW_STATIC);
1930             offset += coeff_token_vlc_tables_size[i];
1931         }
1932         /*
1933          * This is a one time safety check to make sure that
1934          * the packed static coeff_token_vlc table sizes
1935          * were initialized correctly.
1936          */
1937         assert(offset == FF_ARRAY_ELEMS(coeff_token_vlc_tables));
1938
1939         for(i=0; i<3; i++){
1940             chroma_dc_total_zeros_vlc[i].table = chroma_dc_total_zeros_vlc_tables[i];
1941             chroma_dc_total_zeros_vlc[i].table_allocated = chroma_dc_total_zeros_vlc_tables_size;
1942             init_vlc(&chroma_dc_total_zeros_vlc[i],
1943                      CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1944                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1945                      &chroma_dc_total_zeros_bits[i][0], 1, 1,
1946                      INIT_VLC_USE_NEW_STATIC);
1947         }
1948         for(i=0; i<15; i++){
1949             total_zeros_vlc[i].table = total_zeros_vlc_tables[i];
1950             total_zeros_vlc[i].table_allocated = total_zeros_vlc_tables_size;
1951             init_vlc(&total_zeros_vlc[i],
1952                      TOTAL_ZEROS_VLC_BITS, 16,
1953                      &total_zeros_len [i][0], 1, 1,
1954                      &total_zeros_bits[i][0], 1, 1,
1955                      INIT_VLC_USE_NEW_STATIC);
1956         }
1957
1958         for(i=0; i<6; i++){
1959             run_vlc[i].table = run_vlc_tables[i];
1960             run_vlc[i].table_allocated = run_vlc_tables_size;
1961             init_vlc(&run_vlc[i],
1962                      RUN_VLC_BITS, 7,
1963                      &run_len [i][0], 1, 1,
1964                      &run_bits[i][0], 1, 1,
1965                      INIT_VLC_USE_NEW_STATIC);
1966         }
1967         run7_vlc.table = run7_vlc_table,
1968         run7_vlc.table_allocated = run7_vlc_table_size;
1969         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
1970                  &run_len [6][0], 1, 1,
1971                  &run_bits[6][0], 1, 1,
1972                  INIT_VLC_USE_NEW_STATIC);
1973
1974         init_cavlc_level_tab();
1975     }
1976 }
1977
1978 static void free_tables(H264Context *h){
1979     int i;
1980     H264Context *hx;
1981     av_freep(&h->intra4x4_pred_mode);
1982     av_freep(&h->chroma_pred_mode_table);
1983     av_freep(&h->cbp_table);
1984     av_freep(&h->mvd_table[0]);
1985     av_freep(&h->mvd_table[1]);
1986     av_freep(&h->direct_table);
1987     av_freep(&h->non_zero_count);
1988     av_freep(&h->slice_table_base);
1989     h->slice_table= NULL;
1990
1991     av_freep(&h->mb2b_xy);
1992     av_freep(&h->mb2b8_xy);
1993
1994     for(i = 0; i < MAX_THREADS; i++) {
1995         hx = h->thread_context[i];
1996         if(!hx) continue;
1997         av_freep(&hx->top_borders[1]);
1998         av_freep(&hx->top_borders[0]);
1999         av_freep(&hx->s.obmc_scratchpad);
2000         av_freep(&hx->rbsp_buffer[1]);
2001         av_freep(&hx->rbsp_buffer[0]);
2002         if (i) av_freep(&h->thread_context[i]);
2003     }
2004 }
2005
2006 static void init_dequant8_coeff_table(H264Context *h){
2007     int i,q,x;
2008     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2009     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2010     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2011
2012     for(i=0; i<2; i++ ){
2013         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2014             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2015             break;
2016         }
2017
2018         for(q=0; q<52; q++){
2019             int shift = div6[q];
2020             int idx = rem6[q];
2021             for(x=0; x<64; x++)
2022                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2023                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2024                     h->pps.scaling_matrix8[i][x]) << shift;
2025         }
2026     }
2027 }
2028
2029 static void init_dequant4_coeff_table(H264Context *h){
2030     int i,j,q,x;
2031     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2032     for(i=0; i<6; i++ ){
2033         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2034         for(j=0; j<i; j++){
2035             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2036                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2037                 break;
2038             }
2039         }
2040         if(j<i)
2041             continue;
2042
2043         for(q=0; q<52; q++){
2044             int shift = div6[q] + 2;
2045             int idx = rem6[q];
2046             for(x=0; x<16; x++)
2047                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2048                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2049                     h->pps.scaling_matrix4[i][x]) << shift;
2050         }
2051     }
2052 }
2053
2054 static void init_dequant_tables(H264Context *h){
2055     int i,x;
2056     init_dequant4_coeff_table(h);
2057     if(h->pps.transform_8x8_mode)
2058         init_dequant8_coeff_table(h);
2059     if(h->sps.transform_bypass){
2060         for(i=0; i<6; i++)
2061             for(x=0; x<16; x++)
2062                 h->dequant4_coeff[i][0][x] = 1<<6;
2063         if(h->pps.transform_8x8_mode)
2064             for(i=0; i<2; i++)
2065                 for(x=0; x<64; x++)
2066                     h->dequant8_coeff[i][0][x] = 1<<6;
2067     }
2068 }
2069
2070
2071 /**
2072  * allocates tables.
2073  * needs width/height
2074  */
2075 static int alloc_tables(H264Context *h){
2076     MpegEncContext * const s = &h->s;
2077     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2078     int x,y;
2079
2080     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2081
2082     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2083     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base))
2084     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2085
2086     CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2087     CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2088     CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2089     CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2090
2091     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
2092     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2093
2094     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2095     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2096     for(y=0; y<s->mb_height; y++){
2097         for(x=0; x<s->mb_width; x++){
2098             const int mb_xy= x + y*s->mb_stride;
2099             const int b_xy = 4*x + 4*y*h->b_stride;
2100             const int b8_xy= 2*x + 2*y*h->b8_stride;
2101
2102             h->mb2b_xy [mb_xy]= b_xy;
2103             h->mb2b8_xy[mb_xy]= b8_xy;
2104         }
2105     }
2106
2107     s->obmc_scratchpad = NULL;
2108
2109     if(!h->dequant4_coeff[0])
2110         init_dequant_tables(h);
2111
2112     return 0;
2113 fail:
2114     free_tables(h);
2115     return -1;
2116 }
2117
2118 /**
2119  * Mimic alloc_tables(), but for every context thread.
2120  */
2121 static void clone_tables(H264Context *dst, H264Context *src){
2122     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2123     dst->non_zero_count           = src->non_zero_count;
2124     dst->slice_table              = src->slice_table;
2125     dst->cbp_table                = src->cbp_table;
2126     dst->mb2b_xy                  = src->mb2b_xy;
2127     dst->mb2b8_xy                 = src->mb2b8_xy;
2128     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2129     dst->mvd_table[0]             = src->mvd_table[0];
2130     dst->mvd_table[1]             = src->mvd_table[1];
2131     dst->direct_table             = src->direct_table;
2132
2133     dst->s.obmc_scratchpad = NULL;
2134     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2135 }
2136
2137 /**
2138  * Init context
2139  * Allocate buffers which are not shared amongst multiple threads.
2140  */
2141 static int context_init(H264Context *h){
2142     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2143     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2144
2145     return 0;
2146 fail:
2147     return -1; // free_tables will clean up for us
2148 }
2149
2150 static av_cold void common_init(H264Context *h){
2151     MpegEncContext * const s = &h->s;
2152
2153     s->width = s->avctx->width;
2154     s->height = s->avctx->height;
2155     s->codec_id= s->avctx->codec->id;
2156
2157     ff_h264_pred_init(&h->hpc, s->codec_id);
2158
2159     h->dequant_coeff_pps= -1;
2160     s->unrestricted_mv=1;
2161     s->decode=1; //FIXME
2162
2163     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
2164
2165     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2166     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2167 }
2168
2169 /**
2170  * Reset SEI values at the beginning of the frame.
2171  *
2172  * @param h H.264 context.
2173  */
2174 static void reset_sei(H264Context *h) {
2175     h->sei_recovery_frame_cnt       = -1;
2176     h->sei_dpb_output_delay         =  0;
2177     h->sei_cpb_removal_delay        = -1;
2178     h->sei_buffering_period_present =  0;
2179 }
2180
2181 static av_cold int decode_init(AVCodecContext *avctx){
2182     H264Context *h= avctx->priv_data;
2183     MpegEncContext * const s = &h->s;
2184
2185     MPV_decode_defaults(s);
2186
2187     s->avctx = avctx;
2188     common_init(h);
2189
2190     s->out_format = FMT_H264;
2191     s->workaround_bugs= avctx->workaround_bugs;
2192
2193     // set defaults
2194 //    s->decode_mb= ff_h263_decode_mb;
2195     s->quarter_sample = 1;
2196     if(!avctx->has_b_frames)
2197     s->low_delay= 1;
2198
2199     if(s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2200         avctx->pix_fmt= PIX_FMT_VDPAU_H264;
2201     else
2202         avctx->pix_fmt= avctx->get_format(avctx, avctx->codec->pix_fmts);
2203     avctx->hwaccel = ff_find_hwaccel(avctx->codec->id, avctx->pix_fmt);
2204     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
2205
2206     decode_init_vlc();
2207
2208     if(avctx->extradata_size > 0 && avctx->extradata &&
2209        *(char *)avctx->extradata == 1){
2210         h->is_avc = 1;
2211         h->got_avcC = 0;
2212     } else {
2213         h->is_avc = 0;
2214     }
2215
2216     h->thread_context[0] = h;
2217     h->outputed_poc = INT_MIN;
2218     h->prev_poc_msb= 1<<16;
2219     reset_sei(h);
2220     if(avctx->codec_id == CODEC_ID_H264){
2221         if(avctx->ticks_per_frame == 1){
2222             s->avctx->time_base.den *=2;
2223         }
2224         avctx->ticks_per_frame = 2;
2225     }
2226     return 0;
2227 }
2228
2229 static int frame_start(H264Context *h){
2230     MpegEncContext * const s = &h->s;
2231     int i;
2232
2233     if(MPV_frame_start(s, s->avctx) < 0)
2234         return -1;
2235     ff_er_frame_start(s);
2236     /*
2237      * MPV_frame_start uses pict_type to derive key_frame.
2238      * This is incorrect for H.264; IDR markings must be used.
2239      * Zero here; IDR markings per slice in frame or fields are ORed in later.
2240      * See decode_nal_units().
2241      */
2242     s->current_picture_ptr->key_frame= 0;
2243     s->current_picture_ptr->mmco_reset= 0;
2244
2245     assert(s->linesize && s->uvlinesize);
2246
2247     for(i=0; i<16; i++){
2248         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2249         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2250     }
2251     for(i=0; i<4; i++){
2252         h->block_offset[16+i]=
2253         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2254         h->block_offset[24+16+i]=
2255         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2256     }
2257
2258     /* can't be in alloc_tables because linesize isn't known there.
2259      * FIXME: redo bipred weight to not require extra buffer? */
2260     for(i = 0; i < s->avctx->thread_count; i++)
2261         if(!h->thread_context[i]->s.obmc_scratchpad)
2262             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2263
2264     /* some macroblocks will be accessed before they're available */
2265     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2266         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
2267
2268 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2269
2270     // We mark the current picture as non-reference after allocating it, so
2271     // that if we break out due to an error it can be released automatically
2272     // in the next MPV_frame_start().
2273     // SVQ3 as well as most other codecs have only last/next/current and thus
2274     // get released even with set reference, besides SVQ3 and others do not
2275     // mark frames as reference later "naturally".
2276     if(s->codec_id != CODEC_ID_SVQ3)
2277         s->current_picture_ptr->reference= 0;
2278
2279     s->current_picture_ptr->field_poc[0]=
2280     s->current_picture_ptr->field_poc[1]= INT_MAX;
2281     assert(s->current_picture_ptr->long_ref==0);
2282
2283     return 0;
2284 }
2285
2286 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){
2287     MpegEncContext * const s = &h->s;
2288     int i;
2289     int step    = 1;
2290     int offset  = 1;
2291     int uvoffset= 1;
2292     int top_idx = 1;
2293     int skiplast= 0;
2294
2295     src_y  -=   linesize;
2296     src_cb -= uvlinesize;
2297     src_cr -= uvlinesize;
2298
2299     if(!simple && FRAME_MBAFF){
2300         if(s->mb_y&1){
2301             offset  = MB_MBAFF ? 1 : 17;
2302             uvoffset= MB_MBAFF ? 1 : 9;
2303             if(!MB_MBAFF){
2304                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 0)= *(uint64_t*)(src_y +  15*linesize);
2305                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 8)= *(uint64_t*)(src_y +8+15*linesize);
2306                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2307                     *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+7*uvlinesize);
2308                     *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+7*uvlinesize);
2309                 }
2310             }
2311         }else{
2312             if(!MB_MBAFF){
2313                 h->left_border[0]= h->top_borders[0][s->mb_x][15];
2314                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2315                     h->left_border[34   ]= h->top_borders[0][s->mb_x][16+7  ];
2316                     h->left_border[34+18]= h->top_borders[0][s->mb_x][16+8+7];
2317                 }
2318                 skiplast= 1;
2319             }
2320             offset  =
2321             uvoffset=
2322             top_idx = MB_MBAFF ? 0 : 1;
2323         }
2324         step= MB_MBAFF ? 2 : 1;
2325     }
2326
2327     // There are two lines saved, the line above the the top macroblock of a pair,
2328     // and the line above the bottom macroblock
2329     h->left_border[offset]= h->top_borders[top_idx][s->mb_x][15];
2330     for(i=1; i<17 - skiplast; i++){
2331         h->left_border[offset+i*step]= src_y[15+i*  linesize];
2332     }
2333
2334     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2335     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2336
2337     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2338         h->left_border[uvoffset+34   ]= h->top_borders[top_idx][s->mb_x][16+7];
2339         h->left_border[uvoffset+34+18]= h->top_borders[top_idx][s->mb_x][24+7];
2340         for(i=1; i<9 - skiplast; i++){
2341             h->left_border[uvoffset+34   +i*step]= src_cb[7+i*uvlinesize];
2342             h->left_border[uvoffset+34+18+i*step]= src_cr[7+i*uvlinesize];
2343         }
2344         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2345         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2346     }
2347 }
2348
2349 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){
2350     MpegEncContext * const s = &h->s;
2351     int temp8, i;
2352     uint64_t temp64;
2353     int deblock_left;
2354     int deblock_top;
2355     int mb_xy;
2356     int step    = 1;
2357     int offset  = 1;
2358     int uvoffset= 1;
2359     int top_idx = 1;
2360
2361     if(!simple && FRAME_MBAFF){
2362         if(s->mb_y&1){
2363             offset  = MB_MBAFF ? 1 : 17;
2364             uvoffset= MB_MBAFF ? 1 : 9;
2365         }else{
2366             offset  =
2367             uvoffset=
2368             top_idx = MB_MBAFF ? 0 : 1;
2369         }
2370         step= MB_MBAFF ? 2 : 1;
2371     }
2372
2373     if(h->deblocking_filter == 2) {
2374         mb_xy = h->mb_xy;
2375         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2376         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2377     } else {
2378         deblock_left = (s->mb_x > 0);
2379         deblock_top =  (s->mb_y > !!MB_FIELD);
2380     }
2381
2382     src_y  -=   linesize + 1;
2383     src_cb -= uvlinesize + 1;
2384     src_cr -= uvlinesize + 1;
2385
2386 #define XCHG(a,b,t,xchg)\
2387 t= a;\
2388 if(xchg)\
2389     a= b;\
2390 b= t;
2391
2392     if(deblock_left){
2393         for(i = !deblock_top; i<16; i++){
2394             XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, xchg);
2395         }
2396         XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, 1);
2397     }
2398
2399     if(deblock_top){
2400         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2401         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2402         if(s->mb_x+1 < s->mb_width){
2403             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2404         }
2405     }
2406
2407     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2408         if(deblock_left){
2409             for(i = !deblock_top; i<8; i++){
2410                 XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, xchg);
2411                 XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, xchg);
2412             }
2413             XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, 1);
2414             XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, 1);
2415         }
2416         if(deblock_top){
2417             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2418             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2419         }
2420     }
2421 }
2422
2423 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2424     MpegEncContext * const s = &h->s;
2425     const int mb_x= s->mb_x;
2426     const int mb_y= s->mb_y;
2427     const int mb_xy= h->mb_xy;
2428     const int mb_type= s->current_picture.mb_type[mb_xy];
2429     uint8_t  *dest_y, *dest_cb, *dest_cr;
2430     int linesize, uvlinesize /*dct_offset*/;
2431     int i;
2432     int *block_offset = &h->block_offset[0];
2433     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2434     /* is_h264 should always be true if SVQ3 is disabled. */
2435     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
2436     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2437     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2438
2439     dest_y  = s->current_picture.data[0] + (mb_x + mb_y * s->linesize  ) * 16;
2440     dest_cb = s->current_picture.data[1] + (mb_x + mb_y * s->uvlinesize) * 8;
2441     dest_cr = s->current_picture.data[2] + (mb_x + mb_y * s->uvlinesize) * 8;
2442
2443     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2444     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2445
2446     if (!simple && MB_FIELD) {
2447         linesize   = h->mb_linesize   = s->linesize * 2;
2448         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2449         block_offset = &h->block_offset[24];
2450         if(mb_y&1){ //FIXME move out of this function?
2451             dest_y -= s->linesize*15;
2452             dest_cb-= s->uvlinesize*7;
2453             dest_cr-= s->uvlinesize*7;
2454         }
2455         if(FRAME_MBAFF) {
2456             int list;
2457             for(list=0; list<h->list_count; list++){
2458                 if(!USES_LIST(mb_type, list))
2459                     continue;
2460                 if(IS_16X16(mb_type)){
2461                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2462                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2463                 }else{
2464                     for(i=0; i<16; i+=4){
2465                         int ref = h->ref_cache[list][scan8[i]];
2466                         if(ref >= 0)
2467                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2468                     }
2469                 }
2470             }
2471         }
2472     } else {
2473         linesize   = h->mb_linesize   = s->linesize;
2474         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2475 //        dct_offset = s->linesize * 16;
2476     }
2477
2478     if (!simple && IS_INTRA_PCM(mb_type)) {
2479         for (i=0; i<16; i++) {
2480             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
2481         }
2482         for (i=0; i<8; i++) {
2483             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
2484             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
2485         }
2486     } else {
2487         if(IS_INTRA(mb_type)){
2488             if(h->deblocking_filter)
2489                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2490
2491             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2492                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2493                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2494             }
2495
2496             if(IS_INTRA4x4(mb_type)){
2497                 if(simple || !s->encoding){
2498                     if(IS_8x8DCT(mb_type)){
2499                         if(transform_bypass){
2500                             idct_dc_add =
2501                             idct_add    = s->dsp.add_pixels8;
2502                         }else{
2503                             idct_dc_add = s->dsp.h264_idct8_dc_add;
2504                             idct_add    = s->dsp.h264_idct8_add;
2505                         }
2506                         for(i=0; i<16; i+=4){
2507                             uint8_t * const ptr= dest_y + block_offset[i];
2508                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2509                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
2510                                 h->hpc.pred8x8l_add[dir](ptr, h->mb + i*16, linesize);
2511                             }else{
2512                                 const int nnz = h->non_zero_count_cache[ scan8[i] ];
2513                                 h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2514                                                             (h->topright_samples_available<<i)&0x4000, linesize);
2515                                 if(nnz){
2516                                     if(nnz == 1 && h->mb[i*16])
2517                                         idct_dc_add(ptr, h->mb + i*16, linesize);
2518                                     else
2519                                         idct_add   (ptr, h->mb + i*16, linesize);
2520                                 }
2521                             }
2522                         }
2523                     }else{
2524                         if(transform_bypass){
2525                             idct_dc_add =
2526                             idct_add    = s->dsp.add_pixels4;
2527                         }else{
2528                             idct_dc_add = s->dsp.h264_idct_dc_add;
2529                             idct_add    = s->dsp.h264_idct_add;
2530                         }
2531                         for(i=0; i<16; i++){
2532                             uint8_t * const ptr= dest_y + block_offset[i];
2533                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2534
2535                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
2536                                 h->hpc.pred4x4_add[dir](ptr, h->mb + i*16, linesize);
2537                             }else{
2538                                 uint8_t *topright;
2539                                 int nnz, tr;
2540                                 if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2541                                     const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2542                                     assert(mb_y || linesize <= block_offset[i]);
2543                                     if(!topright_avail){
2544                                         tr= ptr[3 - linesize]*0x01010101;
2545                                         topright= (uint8_t*) &tr;
2546                                     }else
2547                                         topright= ptr + 4 - linesize;
2548                                 }else
2549                                     topright= NULL;
2550
2551                                 h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2552                                 nnz = h->non_zero_count_cache[ scan8[i] ];
2553                                 if(nnz){
2554                                     if(is_h264){
2555                                         if(nnz == 1 && h->mb[i*16])
2556                                             idct_dc_add(ptr, h->mb + i*16, linesize);
2557                                         else
2558                                             idct_add   (ptr, h->mb + i*16, linesize);
2559                                     }else
2560                                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2561                                 }
2562                             }
2563                         }
2564                     }
2565                 }
2566             }else{
2567                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2568                 if(is_h264){
2569                     if(!transform_bypass)
2570                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2571                 }else
2572                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2573             }
2574             if(h->deblocking_filter)
2575                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2576         }else if(is_h264){
2577             hl_motion(h, dest_y, dest_cb, dest_cr,
2578                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2579                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2580                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2581         }
2582
2583
2584         if(!IS_INTRA4x4(mb_type)){
2585             if(is_h264){
2586                 if(IS_INTRA16x16(mb_type)){
2587                     if(transform_bypass){
2588                         if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
2589                             h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb, linesize);
2590                         }else{
2591                             for(i=0; i<16; i++){
2592                                 if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
2593                                     s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + i*16, linesize);
2594                             }
2595                         }
2596                     }else{
2597                          s->dsp.h264_idct_add16intra(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2598                     }
2599                 }else if(h->cbp&15){
2600                     if(transform_bypass){
2601                         const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2602                         idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2603                         for(i=0; i<16; i+=di){
2604                             if(h->non_zero_count_cache[ scan8[i] ]){
2605                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2606                             }
2607                         }
2608                     }else{
2609                         if(IS_8x8DCT(mb_type)){
2610                             s->dsp.h264_idct8_add4(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2611                         }else{
2612                             s->dsp.h264_idct_add16(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2613                         }
2614                     }
2615                 }
2616             }else{
2617                 for(i=0; i<16; i++){
2618                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2619                         uint8_t * const ptr= dest_y + block_offset[i];
2620                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2621                     }
2622                 }
2623             }
2624         }
2625
2626         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
2627             uint8_t *dest[2] = {dest_cb, dest_cr};
2628             if(transform_bypass){
2629                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
2630                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + 16*16, uvlinesize);
2631                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 20, h->mb + 20*16, uvlinesize);
2632                 }else{
2633                     idct_add = s->dsp.add_pixels4;
2634                     for(i=16; i<16+8; i++){
2635                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
2636                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2637                     }
2638                 }
2639             }else{
2640                 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]);
2641                 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]);
2642                 if(is_h264){
2643                     idct_add = s->dsp.h264_idct_add;
2644                     idct_dc_add = s->dsp.h264_idct_dc_add;
2645                     for(i=16; i<16+8; i++){
2646                         if(h->non_zero_count_cache[ scan8[i] ])
2647                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2648                         else if(h->mb[i*16])
2649                             idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2650                     }
2651                 }else{
2652                     for(i=16; i<16+8; i++){
2653                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2654                             uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2655                             svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2656                         }
2657                     }
2658                 }
2659             }
2660         }
2661     }
2662     if(h->cbp || IS_INTRA(mb_type))
2663         s->dsp.clear_blocks(h->mb);
2664
2665     if(h->deblocking_filter) {
2666         backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2667         fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2668         h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2669         h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2670         if (!simple && FRAME_MBAFF) {
2671             filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2672         } else {
2673             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2674         }
2675     }
2676 }
2677
2678 /**
2679  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2680  */
2681 static void hl_decode_mb_simple(H264Context *h){
2682     hl_decode_mb_internal(h, 1);
2683 }
2684
2685 /**
2686  * Process a macroblock; this handles edge cases, such as interlacing.
2687  */
2688 static void av_noinline hl_decode_mb_complex(H264Context *h){
2689     hl_decode_mb_internal(h, 0);
2690 }
2691
2692 static void hl_decode_mb(H264Context *h){
2693     MpegEncContext * const s = &h->s;
2694     const int mb_xy= h->mb_xy;
2695     const int mb_type= s->current_picture.mb_type[mb_xy];
2696     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2697
2698     if (is_complex)
2699         hl_decode_mb_complex(h);
2700     else hl_decode_mb_simple(h);
2701 }
2702
2703 static void pic_as_field(Picture *pic, const int parity){
2704     int i;
2705     for (i = 0; i < 4; ++i) {
2706         if (parity == PICT_BOTTOM_FIELD)
2707             pic->data[i] += pic->linesize[i];
2708         pic->reference = parity;
2709         pic->linesize[i] *= 2;
2710     }
2711     pic->poc= pic->field_poc[parity == PICT_BOTTOM_FIELD];
2712 }
2713
2714 static int split_field_copy(Picture *dest, Picture *src,
2715                             int parity, int id_add){
2716     int match = !!(src->reference & parity);
2717
2718     if (match) {
2719         *dest = *src;
2720         if(parity != PICT_FRAME){
2721             pic_as_field(dest, parity);
2722             dest->pic_id *= 2;
2723             dest->pic_id += id_add;
2724         }
2725     }
2726
2727     return match;
2728 }
2729
2730 static int build_def_list(Picture *def, Picture **in, int len, int is_long, int sel){
2731     int i[2]={0};
2732     int index=0;
2733
2734     while(i[0]<len || i[1]<len){
2735         while(i[0]<len && !(in[ i[0] ] && (in[ i[0] ]->reference & sel)))
2736             i[0]++;
2737         while(i[1]<len && !(in[ i[1] ] && (in[ i[1] ]->reference & (sel^3))))
2738             i[1]++;
2739         if(i[0] < len){
2740             in[ i[0] ]->pic_id= is_long ? i[0] : in[ i[0] ]->frame_num;
2741             split_field_copy(&def[index++], in[ i[0]++ ], sel  , 1);
2742         }
2743         if(i[1] < len){
2744             in[ i[1] ]->pic_id= is_long ? i[1] : in[ i[1] ]->frame_num;
2745             split_field_copy(&def[index++], in[ i[1]++ ], sel^3, 0);
2746         }
2747     }
2748
2749     return index;
2750 }
2751
2752 static int add_sorted(Picture **sorted, Picture **src, int len, int limit, int dir){
2753     int i, best_poc;
2754     int out_i= 0;
2755
2756     for(;;){
2757         best_poc= dir ? INT_MIN : INT_MAX;
2758
2759         for(i=0; i<len; i++){
2760             const int poc= src[i]->poc;
2761             if(((poc > limit) ^ dir) && ((poc < best_poc) ^ dir)){
2762                 best_poc= poc;
2763                 sorted[out_i]= src[i];
2764             }
2765         }
2766         if(best_poc == (dir ? INT_MIN : INT_MAX))
2767             break;
2768         limit= sorted[out_i++]->poc - dir;
2769     }
2770     return out_i;
2771 }
2772
2773 /**
2774  * fills the default_ref_list.
2775  */
2776 static int fill_default_ref_list(H264Context *h){
2777     MpegEncContext * const s = &h->s;
2778     int i, len;
2779
2780     if(h->slice_type_nos==FF_B_TYPE){
2781         Picture *sorted[32];
2782         int cur_poc, list;
2783         int lens[2];
2784
2785         if(FIELD_PICTURE)
2786             cur_poc= s->current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
2787         else
2788             cur_poc= s->current_picture_ptr->poc;
2789
2790         for(list= 0; list<2; list++){
2791             len= add_sorted(sorted    , h->short_ref, h->short_ref_count, cur_poc, 1^list);
2792             len+=add_sorted(sorted+len, h->short_ref, h->short_ref_count, cur_poc, 0^list);
2793             assert(len<=32);
2794             len= build_def_list(h->default_ref_list[list]    , sorted     , len, 0, s->picture_structure);
2795             len+=build_def_list(h->default_ref_list[list]+len, h->long_ref, 16 , 1, s->picture_structure);
2796             assert(len<=32);
2797
2798             if(len < h->ref_count[list])
2799                 memset(&h->default_ref_list[list][len], 0, sizeof(Picture)*(h->ref_count[list] - len));
2800             lens[list]= len;
2801         }
2802
2803         if(lens[0] == lens[1] && lens[1] > 1){
2804             for(i=0; h->default_ref_list[0][i].data[0] == h->default_ref_list[1][i].data[0] && i<lens[0]; i++);
2805             if(i == lens[0])
2806                 FFSWAP(Picture, h->default_ref_list[1][0], h->default_ref_list[1][1]);
2807         }
2808     }else{
2809         len = build_def_list(h->default_ref_list[0]    , h->short_ref, h->short_ref_count, 0, s->picture_structure);
2810         len+= build_def_list(h->default_ref_list[0]+len, h-> long_ref, 16                , 1, s->picture_structure);
2811         assert(len <= 32);
2812         if(len < h->ref_count[0])
2813             memset(&h->default_ref_list[0][len], 0, sizeof(Picture)*(h->ref_count[0] - len));
2814     }
2815 #ifdef TRACE
2816     for (i=0; i<h->ref_count[0]; i++) {
2817         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]);
2818     }
2819     if(h->slice_type_nos==FF_B_TYPE){
2820         for (i=0; i<h->ref_count[1]; i++) {
2821             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]);
2822         }
2823     }
2824 #endif
2825     return 0;
2826 }
2827
2828 static void print_short_term(H264Context *h);
2829 static void print_long_term(H264Context *h);
2830
2831 /**
2832  * Extract structure information about the picture described by pic_num in
2833  * the current decoding context (frame or field). Note that pic_num is
2834  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
2835  * @param pic_num picture number for which to extract structure information
2836  * @param structure one of PICT_XXX describing structure of picture
2837  *                      with pic_num
2838  * @return frame number (short term) or long term index of picture
2839  *         described by pic_num
2840  */
2841 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
2842     MpegEncContext * const s = &h->s;
2843
2844     *structure = s->picture_structure;
2845     if(FIELD_PICTURE){
2846         if (!(pic_num & 1))
2847             /* opposite field */
2848             *structure ^= PICT_FRAME;
2849         pic_num >>= 1;
2850     }
2851
2852     return pic_num;
2853 }
2854
2855 static int decode_ref_pic_list_reordering(H264Context *h){
2856     MpegEncContext * const s = &h->s;
2857     int list, index, pic_structure;
2858
2859     print_short_term(h);
2860     print_long_term(h);
2861
2862     for(list=0; list<h->list_count; list++){
2863         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
2864
2865         if(get_bits1(&s->gb)){
2866             int pred= h->curr_pic_num;
2867
2868             for(index=0; ; index++){
2869                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb_31(&s->gb);
2870                 unsigned int pic_id;
2871                 int i;
2872                 Picture *ref = NULL;
2873
2874                 if(reordering_of_pic_nums_idc==3)
2875                     break;
2876
2877                 if(index >= h->ref_count[list]){
2878                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
2879                     return -1;
2880                 }
2881
2882                 if(reordering_of_pic_nums_idc<3){
2883                     if(reordering_of_pic_nums_idc<2){
2884                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
2885                         int frame_num;
2886
2887                         if(abs_diff_pic_num > h->max_pic_num){
2888                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
2889                             return -1;
2890                         }
2891
2892                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
2893                         else                                pred+= abs_diff_pic_num;
2894                         pred &= h->max_pic_num - 1;
2895
2896                         frame_num = pic_num_extract(h, pred, &pic_structure);
2897
2898                         for(i= h->short_ref_count-1; i>=0; i--){
2899                             ref = h->short_ref[i];
2900                             assert(ref->reference);
2901                             assert(!ref->long_ref);
2902                             if(
2903                                    ref->frame_num == frame_num &&
2904                                    (ref->reference & pic_structure)
2905                               )
2906                                 break;
2907                         }
2908                         if(i>=0)
2909                             ref->pic_id= pred;
2910                     }else{
2911                         int long_idx;
2912                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
2913
2914                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
2915
2916                         if(long_idx>31){
2917                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
2918                             return -1;
2919                         }
2920                         ref = h->long_ref[long_idx];
2921                         assert(!(ref && !ref->reference));
2922                         if(ref && (ref->reference & pic_structure)){
2923                             ref->pic_id= pic_id;
2924                             assert(ref->long_ref);
2925                             i=0;
2926                         }else{
2927                             i=-1;
2928                         }
2929                     }
2930
2931                     if (i < 0) {
2932                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
2933                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
2934                     } else {
2935                         for(i=index; i+1<h->ref_count[list]; i++){
2936                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
2937                                 break;
2938                         }
2939                         for(; i > index; i--){
2940                             h->ref_list[list][i]= h->ref_list[list][i-1];
2941                         }
2942                         h->ref_list[list][index]= *ref;
2943                         if (FIELD_PICTURE){
2944                             pic_as_field(&h->ref_list[list][index], pic_structure);
2945                         }
2946                     }
2947                 }else{
2948                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
2949                     return -1;
2950                 }
2951             }
2952         }
2953     }
2954     for(list=0; list<h->list_count; list++){
2955         for(index= 0; index < h->ref_count[list]; index++){
2956             if(!h->ref_list[list][index].data[0]){
2957                 av_log(h->s.avctx, AV_LOG_ERROR, "Missing reference picture\n");
2958                 if(h->default_ref_list[list][0].data[0])
2959                     h->ref_list[list][index]= h->default_ref_list[list][0];
2960                 else
2961                     return -1;
2962             }
2963         }
2964     }
2965
2966     return 0;
2967 }
2968
2969 static void fill_mbaff_ref_list(H264Context *h){
2970     int list, i, j;
2971     for(list=0; list<2; list++){ //FIXME try list_count
2972         for(i=0; i<h->ref_count[list]; i++){
2973             Picture *frame = &h->ref_list[list][i];
2974             Picture *field = &h->ref_list[list][16+2*i];
2975             field[0] = *frame;
2976             for(j=0; j<3; j++)
2977                 field[0].linesize[j] <<= 1;
2978             field[0].reference = PICT_TOP_FIELD;
2979             field[0].poc= field[0].field_poc[0];
2980             field[1] = field[0];
2981             for(j=0; j<3; j++)
2982                 field[1].data[j] += frame->linesize[j];
2983             field[1].reference = PICT_BOTTOM_FIELD;
2984             field[1].poc= field[1].field_poc[1];
2985
2986             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
2987             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
2988             for(j=0; j<2; j++){
2989                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
2990                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
2991             }
2992         }
2993     }
2994     for(j=0; j<h->ref_count[1]; j++){
2995         for(i=0; i<h->ref_count[0]; i++)
2996             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
2997         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
2998         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
2999     }
3000 }
3001
3002 static int pred_weight_table(H264Context *h){
3003     MpegEncContext * const s = &h->s;
3004     int list, i;
3005     int luma_def, chroma_def;
3006
3007     h->use_weight= 0;
3008     h->use_weight_chroma= 0;
3009     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
3010     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
3011     luma_def = 1<<h->luma_log2_weight_denom;
3012     chroma_def = 1<<h->chroma_log2_weight_denom;
3013
3014     for(list=0; list<2; list++){
3015         h->luma_weight_flag[list]   = 0;
3016         h->chroma_weight_flag[list] = 0;
3017         for(i=0; i<h->ref_count[list]; i++){
3018             int luma_weight_flag, chroma_weight_flag;
3019
3020             luma_weight_flag= get_bits1(&s->gb);
3021             if(luma_weight_flag){
3022                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3023                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3024                 if(   h->luma_weight[list][i] != luma_def
3025                    || h->luma_offset[list][i] != 0) {
3026                     h->use_weight= 1;
3027                     h->luma_weight_flag[list]= 1;
3028                 }
3029             }else{
3030                 h->luma_weight[list][i]= luma_def;
3031                 h->luma_offset[list][i]= 0;
3032             }
3033
3034             if(CHROMA){
3035                 chroma_weight_flag= get_bits1(&s->gb);
3036                 if(chroma_weight_flag){
3037                     int j;
3038                     for(j=0; j<2; j++){
3039                         h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3040                         h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3041                         if(   h->chroma_weight[list][i][j] != chroma_def
3042                            || h->chroma_offset[list][i][j] != 0) {
3043                             h->use_weight_chroma= 1;
3044                             h->chroma_weight_flag[list]= 1;
3045                         }
3046                     }
3047                 }else{
3048                     int j;
3049                     for(j=0; j<2; j++){
3050                         h->chroma_weight[list][i][j]= chroma_def;
3051                         h->chroma_offset[list][i][j]= 0;
3052                     }
3053                 }
3054             }
3055         }
3056         if(h->slice_type_nos != FF_B_TYPE) break;
3057     }
3058     h->use_weight= h->use_weight || h->use_weight_chroma;
3059     return 0;
3060 }
3061
3062 static void implicit_weight_table(H264Context *h){
3063     MpegEncContext * const s = &h->s;
3064     int ref0, ref1, i;
3065     int cur_poc = s->current_picture_ptr->poc;
3066
3067     for (i = 0; i < 2; i++) {
3068         h->luma_weight_flag[i]   = 0;
3069         h->chroma_weight_flag[i] = 0;
3070     }
3071
3072     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3073        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3074         h->use_weight= 0;
3075         h->use_weight_chroma= 0;
3076         return;
3077     }
3078
3079     h->use_weight= 2;
3080     h->use_weight_chroma= 2;
3081     h->luma_log2_weight_denom= 5;
3082     h->chroma_log2_weight_denom= 5;
3083
3084     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3085         int poc0 = h->ref_list[0][ref0].poc;
3086         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3087             int poc1 = h->ref_list[1][ref1].poc;
3088             int td = av_clip(poc1 - poc0, -128, 127);
3089             if(td){
3090                 int tb = av_clip(cur_poc - poc0, -128, 127);
3091                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3092                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3093                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3094                     h->implicit_weight[ref0][ref1] = 32;
3095                 else
3096                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3097             }else
3098                 h->implicit_weight[ref0][ref1] = 32;
3099         }
3100     }
3101 }
3102
3103 /**
3104  * Mark a picture as no longer needed for reference. The refmask
3105  * argument allows unreferencing of individual fields or the whole frame.
3106  * If the picture becomes entirely unreferenced, but is being held for
3107  * display purposes, it is marked as such.
3108  * @param refmask mask of fields to unreference; the mask is bitwise
3109  *                anded with the reference marking of pic
3110  * @return non-zero if pic becomes entirely unreferenced (except possibly
3111  *         for display purposes) zero if one of the fields remains in
3112  *         reference
3113  */
3114 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3115     int i;
3116     if (pic->reference &= refmask) {
3117         return 0;
3118     } else {
3119         for(i = 0; h->delayed_pic[i]; i++)
3120             if(pic == h->delayed_pic[i]){
3121                 pic->reference=DELAYED_PIC_REF;
3122                 break;
3123             }
3124         return 1;
3125     }
3126 }
3127
3128 /**
3129  * instantaneous decoder refresh.
3130  */
3131 static void idr(H264Context *h){
3132     int i;
3133
3134     for(i=0; i<16; i++){
3135         remove_long(h, i, 0);
3136     }
3137     assert(h->long_ref_count==0);
3138
3139     for(i=0; i<h->short_ref_count; i++){
3140         unreference_pic(h, h->short_ref[i], 0);
3141         h->short_ref[i]= NULL;
3142     }
3143     h->short_ref_count=0;
3144     h->prev_frame_num= 0;
3145     h->prev_frame_num_offset= 0;
3146     h->prev_poc_msb=
3147     h->prev_poc_lsb= 0;
3148 }
3149
3150 /* forget old pics after a seek */
3151 static void flush_dpb(AVCodecContext *avctx){
3152     H264Context *h= avctx->priv_data;
3153     int i;
3154     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
3155         if(h->delayed_pic[i])
3156             h->delayed_pic[i]->reference= 0;
3157         h->delayed_pic[i]= NULL;
3158     }
3159     h->outputed_poc= INT_MIN;
3160     h->prev_interlaced_frame = 1;
3161     idr(h);
3162     if(h->s.current_picture_ptr)
3163         h->s.current_picture_ptr->reference= 0;
3164     h->s.first_field= 0;
3165     reset_sei(h);
3166     ff_mpeg_flush(avctx);
3167 }
3168
3169 /**
3170  * Find a Picture in the short term reference list by frame number.
3171  * @param frame_num frame number to search for
3172  * @param idx the index into h->short_ref where returned picture is found
3173  *            undefined if no picture found.
3174  * @return pointer to the found picture, or NULL if no pic with the provided
3175  *                 frame number is found
3176  */
3177 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3178     MpegEncContext * const s = &h->s;
3179     int i;
3180
3181     for(i=0; i<h->short_ref_count; i++){
3182         Picture *pic= h->short_ref[i];
3183         if(s->avctx->debug&FF_DEBUG_MMCO)
3184             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3185         if(pic->frame_num == frame_num) {
3186             *idx = i;
3187             return pic;
3188         }
3189     }
3190     return NULL;
3191 }
3192
3193 /**
3194  * Remove a picture from the short term reference list by its index in
3195  * that list.  This does no checking on the provided index; it is assumed
3196  * to be valid. Other list entries are shifted down.
3197  * @param i index into h->short_ref of picture to remove.
3198  */
3199 static void remove_short_at_index(H264Context *h, int i){
3200     assert(i >= 0 && i < h->short_ref_count);
3201     h->short_ref[i]= NULL;
3202     if (--h->short_ref_count)
3203         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3204 }
3205
3206 /**
3207  *
3208  * @return the removed picture or NULL if an error occurs
3209  */
3210 static Picture * remove_short(H264Context *h, int frame_num, int ref_mask){
3211     MpegEncContext * const s = &h->s;
3212     Picture *pic;
3213     int i;
3214
3215     if(s->avctx->debug&FF_DEBUG_MMCO)
3216         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3217
3218     pic = find_short(h, frame_num, &i);
3219     if (pic){
3220         if(unreference_pic(h, pic, ref_mask))
3221         remove_short_at_index(h, i);
3222     }
3223
3224     return pic;
3225 }
3226
3227 /**
3228  * Remove a picture from the long term reference list by its index in
3229  * that list.
3230  * @return the removed picture or NULL if an error occurs
3231  */
3232 static Picture * remove_long(H264Context *h, int i, int ref_mask){
3233     Picture *pic;
3234
3235     pic= h->long_ref[i];
3236     if (pic){
3237         if(unreference_pic(h, pic, ref_mask)){
3238             assert(h->long_ref[i]->long_ref == 1);
3239             h->long_ref[i]->long_ref= 0;
3240             h->long_ref[i]= NULL;
3241             h->long_ref_count--;
3242         }
3243     }
3244
3245     return pic;
3246 }
3247
3248 /**
3249  * print short term list
3250  */
3251 static void print_short_term(H264Context *h) {
3252     uint32_t i;
3253     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3254         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3255         for(i=0; i<h->short_ref_count; i++){
3256             Picture *pic= h->short_ref[i];
3257             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3258         }
3259     }
3260 }
3261
3262 /**
3263  * print long term list
3264  */
3265 static void print_long_term(H264Context *h) {
3266     uint32_t i;
3267     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3268         av_log(h->s.avctx, AV_LOG_DEBUG, "long term list:\n");
3269         for(i = 0; i < 16; i++){
3270             Picture *pic= h->long_ref[i];
3271             if (pic) {
3272                 av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3273             }
3274         }
3275     }
3276 }
3277
3278 /**
3279  * Executes the reference picture marking (memory management control operations).
3280  */
3281 static int execute_ref_pic_marking(H264Context *h, MMCO *mmco, int mmco_count){
3282     MpegEncContext * const s = &h->s;
3283     int i, av_uninit(j);
3284     int current_ref_assigned=0;
3285     Picture *av_uninit(pic);
3286
3287     if((s->avctx->debug&FF_DEBUG_MMCO) && mmco_count==0)
3288         av_log(h->s.avctx, AV_LOG_DEBUG, "no mmco here\n");
3289
3290     for(i=0; i<mmco_count; i++){
3291         int av_uninit(structure), av_uninit(frame_num);
3292         if(s->avctx->debug&FF_DEBUG_MMCO)
3293             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);
3294
3295         if(   mmco[i].opcode == MMCO_SHORT2UNUSED
3296            || mmco[i].opcode == MMCO_SHORT2LONG){
3297             frame_num = pic_num_extract(h, mmco[i].short_pic_num, &structure);
3298             pic = find_short(h, frame_num, &j);
3299             if(!pic){
3300                 if(mmco[i].opcode != MMCO_SHORT2LONG || !h->long_ref[mmco[i].long_arg]
3301                    || h->long_ref[mmco[i].long_arg]->frame_num != frame_num)
3302                 av_log(h->s.avctx, AV_LOG_ERROR, "mmco: unref short failure\n");
3303                 continue;
3304             }
3305         }
3306
3307         switch(mmco[i].opcode){
3308         case MMCO_SHORT2UNUSED:
3309             if(s->avctx->debug&FF_DEBUG_MMCO)
3310                 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);
3311             remove_short(h, frame_num, structure ^ PICT_FRAME);
3312             break;
3313         case MMCO_SHORT2LONG:
3314                 if (h->long_ref[mmco[i].long_arg] != pic)
3315                     remove_long(h, mmco[i].long_arg, 0);
3316
3317                 remove_short_at_index(h, j);
3318                 h->long_ref[ mmco[i].long_arg ]= pic;
3319                 if (h->long_ref[ mmco[i].long_arg ]){
3320                     h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3321                     h->long_ref_count++;
3322                 }
3323             break;
3324         case MMCO_LONG2UNUSED:
3325             j = pic_num_extract(h, mmco[i].long_arg, &structure);
3326             pic = h->long_ref[j];
3327             if (pic) {
3328                 remove_long(h, j, structure ^ PICT_FRAME);
3329             } else if(s->avctx->debug&FF_DEBUG_MMCO)
3330                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref long failure\n");
3331             break;
3332         case MMCO_LONG:
3333                     // Comment below left from previous code as it is an interresting note.
3334                     /* First field in pair is in short term list or
3335                      * at a different long term index.
3336                      * This is not allowed; see 7.4.3.3, notes 2 and 3.
3337                      * Report the problem and keep the pair where it is,
3338                      * and mark this field valid.
3339                      */
3340
3341             if (h->long_ref[mmco[i].long_arg] != s->current_picture_ptr) {
3342                 remove_long(h, mmco[i].long_arg, 0);
3343
3344                 h->long_ref[ mmco[i].long_arg ]= s->current_picture_ptr;
3345                 h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3346                 h->long_ref_count++;
3347             }
3348
3349             s->current_picture_ptr->reference |= s->picture_structure;
3350             current_ref_assigned=1;
3351             break;
3352         case MMCO_SET_MAX_LONG:
3353             assert(mmco[i].long_arg <= 16);
3354             // just remove the long term which index is greater than new max
3355             for(j = mmco[i].long_arg; j<16; j++){
3356                 remove_long(h, j, 0);
3357             }
3358             break;
3359         case MMCO_RESET:
3360             while(h->short_ref_count){
3361                 remove_short(h, h->short_ref[0]->frame_num, 0);
3362             }
3363             for(j = 0; j < 16; j++) {
3364                 remove_long(h, j, 0);
3365             }
3366             s->current_picture_ptr->poc=
3367             s->current_picture_ptr->field_poc[0]=
3368             s->current_picture_ptr->field_poc[1]=
3369             h->poc_lsb=
3370             h->poc_msb=
3371             h->frame_num=
3372             s->current_picture_ptr->frame_num= 0;
3373             s->current_picture_ptr->mmco_reset=1;
3374             break;
3375         default: assert(0);
3376         }
3377     }
3378
3379     if (!current_ref_assigned) {
3380         /* Second field of complementary field pair; the first field of
3381          * which is already referenced. If short referenced, it
3382          * should be first entry in short_ref. If not, it must exist
3383          * in long_ref; trying to put it on the short list here is an
3384          * error in the encoded bit stream (ref: 7.4.3.3, NOTE 2 and 3).
3385          */
3386         if (h->short_ref_count && h->short_ref[0] == s->current_picture_ptr) {
3387             /* Just mark the second field valid */
3388             s->current_picture_ptr->reference = PICT_FRAME;
3389         } else if (s->current_picture_ptr->long_ref) {
3390             av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term reference "
3391                                              "assignment for second field "
3392                                              "in complementary field pair "
3393                                              "(first field is long term)\n");
3394         } else {
3395             pic= remove_short(h, s->current_picture_ptr->frame_num, 0);
3396             if(pic){
3397                 av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term buffer state detected\n");
3398             }
3399
3400             if(h->short_ref_count)
3401                 memmove(&h->short_ref[1], &h->short_ref[0], h->short_ref_count*sizeof(Picture*));
3402
3403             h->short_ref[0]= s->current_picture_ptr;
3404             h->short_ref_count++;
3405             s->current_picture_ptr->reference |= s->picture_structure;
3406         }
3407     }
3408
3409     if (h->long_ref_count + h->short_ref_count > h->sps.ref_frame_count){
3410
3411         /* We have too many reference frames, probably due to corrupted
3412          * stream. Need to discard one frame. Prevents overrun of the
3413          * short_ref and long_ref buffers.
3414          */
3415         av_log(h->s.avctx, AV_LOG_ERROR,
3416                "number of reference frames exceeds max (probably "
3417                "corrupt input), discarding one\n");
3418
3419         if (h->long_ref_count && !h->short_ref_count) {
3420             for (i = 0; i < 16; ++i)
3421                 if (h->long_ref[i])
3422                     break;
3423
3424             assert(i < 16);
3425             remove_long(h, i, 0);
3426         } else {
3427             pic = h->short_ref[h->short_ref_count - 1];
3428             remove_short(h, pic->frame_num, 0);
3429         }
3430     }
3431
3432     print_short_term(h);
3433     print_long_term(h);
3434     return 0;
3435 }
3436
3437 static int decode_ref_pic_marking(H264Context *h, GetBitContext *gb){
3438     MpegEncContext * const s = &h->s;
3439     int i;
3440
3441     h->mmco_index= 0;
3442     if(h->nal_unit_type == NAL_IDR_SLICE){ //FIXME fields
3443         s->broken_link= get_bits1(gb) -1;
3444         if(get_bits1(gb)){
3445             h->mmco[0].opcode= MMCO_LONG;
3446             h->mmco[0].long_arg= 0;
3447             h->mmco_index= 1;
3448         }
3449     }else{
3450         if(get_bits1(gb)){ // adaptive_ref_pic_marking_mode_flag
3451             for(i= 0; i<MAX_MMCO_COUNT; i++) {
3452                 MMCOOpcode opcode= get_ue_golomb_31(gb);
3453
3454                 h->mmco[i].opcode= opcode;
3455                 if(opcode==MMCO_SHORT2UNUSED || opcode==MMCO_SHORT2LONG){
3456                     h->mmco[i].short_pic_num= (h->curr_pic_num - get_ue_golomb(gb) - 1) & (h->max_pic_num - 1);
3457 /*                    if(h->mmco[i].short_pic_num >= h->short_ref_count || h->short_ref[ h->mmco[i].short_pic_num ] == NULL){
3458                         av_log(s->avctx, AV_LOG_ERROR, "illegal short ref in memory management control operation %d\n", mmco);
3459                         return -1;
3460                     }*/
3461                 }
3462                 if(opcode==MMCO_SHORT2LONG || opcode==MMCO_LONG2UNUSED || opcode==MMCO_LONG || opcode==MMCO_SET_MAX_LONG){
3463                     unsigned int long_arg= get_ue_golomb_31(gb);
3464                     if(long_arg >= 32 || (long_arg >= 16 && !(opcode == MMCO_LONG2UNUSED && FIELD_PICTURE))){
3465                         av_log(h->s.avctx, AV_LOG_ERROR, "illegal long ref in memory management control operation %d\n", opcode);
3466                         return -1;
3467                     }
3468                     h->mmco[i].long_arg= long_arg;
3469                 }
3470
3471                 if(opcode > (unsigned)MMCO_LONG){
3472                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal memory management control operation %d\n", opcode);
3473                     return -1;
3474                 }
3475                 if(opcode == MMCO_END)
3476                     break;
3477             }
3478             h->mmco_index= i;
3479         }else{
3480             assert(h->long_ref_count + h->short_ref_count <= h->sps.ref_frame_count);
3481
3482             if(h->short_ref_count && h->long_ref_count + h->short_ref_count == h->sps.ref_frame_count &&
3483                     !(FIELD_PICTURE && !s->first_field && s->current_picture_ptr->reference)) {
3484                 h->mmco[0].opcode= MMCO_SHORT2UNUSED;
3485                 h->mmco[0].short_pic_num= h->short_ref[ h->short_ref_count - 1 ]->frame_num;
3486                 h->mmco_index= 1;
3487                 if (FIELD_PICTURE) {
3488                     h->mmco[0].short_pic_num *= 2;
3489                     h->mmco[1].opcode= MMCO_SHORT2UNUSED;
3490                     h->mmco[1].short_pic_num= h->mmco[0].short_pic_num + 1;
3491                     h->mmco_index= 2;
3492                 }
3493             }
3494         }
3495     }
3496
3497     return 0;
3498 }
3499
3500 static int init_poc(H264Context *h){
3501     MpegEncContext * const s = &h->s;
3502     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
3503     int field_poc[2];
3504     Picture *cur = s->current_picture_ptr;
3505
3506     h->frame_num_offset= h->prev_frame_num_offset;
3507     if(h->frame_num < h->prev_frame_num)
3508         h->frame_num_offset += max_frame_num;
3509
3510     if(h->sps.poc_type==0){
3511         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
3512
3513         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
3514             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
3515         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
3516             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
3517         else
3518             h->poc_msb = h->prev_poc_msb;
3519 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
3520         field_poc[0] =
3521         field_poc[1] = h->poc_msb + h->poc_lsb;
3522         if(s->picture_structure == PICT_FRAME)
3523             field_poc[1] += h->delta_poc_bottom;
3524     }else if(h->sps.poc_type==1){
3525         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
3526         int i;
3527
3528         if(h->sps.poc_cycle_length != 0)
3529             abs_frame_num = h->frame_num_offset + h->frame_num;
3530         else
3531             abs_frame_num = 0;
3532
3533         if(h->nal_ref_idc==0 && abs_frame_num > 0)
3534             abs_frame_num--;
3535
3536         expected_delta_per_poc_cycle = 0;
3537         for(i=0; i < h->sps.poc_cycle_length; i++)
3538             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
3539
3540         if(abs_frame_num > 0){
3541             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
3542             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
3543
3544             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
3545             for(i = 0; i <= frame_num_in_poc_cycle; i++)
3546                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
3547         } else
3548             expectedpoc = 0;
3549
3550         if(h->nal_ref_idc == 0)
3551             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
3552
3553         field_poc[0] = expectedpoc + h->delta_poc[0];
3554         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
3555
3556         if(s->picture_structure == PICT_FRAME)
3557             field_poc[1] += h->delta_poc[1];
3558     }else{
3559         int poc= 2*(h->frame_num_offset + h->frame_num);
3560
3561         if(!h->nal_ref_idc)
3562             poc--;
3563
3564         field_poc[0]= poc;
3565         field_poc[1]= poc;
3566     }
3567
3568     if(s->picture_structure != PICT_BOTTOM_FIELD)
3569         s->current_picture_ptr->field_poc[0]= field_poc[0];
3570     if(s->picture_structure != PICT_TOP_FIELD)
3571         s->current_picture_ptr->field_poc[1]= field_poc[1];
3572     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
3573
3574     return 0;
3575 }
3576
3577
3578 /**
3579  * initialize scan tables
3580  */
3581 static void init_scan_tables(H264Context *h){
3582     MpegEncContext * const s = &h->s;
3583     int i;
3584     if(s->dsp.h264_idct_add == ff_h264_idct_add_c){ //FIXME little ugly
3585         memcpy(h->zigzag_scan, zigzag_scan, 16*sizeof(uint8_t));
3586         memcpy(h-> field_scan,  field_scan, 16*sizeof(uint8_t));
3587     }else{
3588         for(i=0; i<16; i++){
3589 #define T(x) (x>>2) | ((x<<2) & 0xF)
3590             h->zigzag_scan[i] = T(zigzag_scan[i]);
3591             h-> field_scan[i] = T( field_scan[i]);
3592 #undef T
3593         }
3594     }
3595     if(s->dsp.h264_idct8_add == ff_h264_idct8_add_c){
3596         memcpy(h->zigzag_scan8x8,       ff_zigzag_direct,     64*sizeof(uint8_t));
3597         memcpy(h->zigzag_scan8x8_cavlc, zigzag_scan8x8_cavlc, 64*sizeof(uint8_t));
3598         memcpy(h->field_scan8x8,        field_scan8x8,        64*sizeof(uint8_t));
3599         memcpy(h->field_scan8x8_cavlc,  field_scan8x8_cavlc,  64*sizeof(uint8_t));
3600     }else{
3601         for(i=0; i<64; i++){
3602 #define T(x) (x>>3) | ((x&7)<<3)
3603             h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
3604             h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
3605             h->field_scan8x8[i]        = T(field_scan8x8[i]);
3606             h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
3607 #undef T
3608         }
3609     }
3610     if(h->sps.transform_bypass){ //FIXME same ugly
3611         h->zigzag_scan_q0          = zigzag_scan;
3612         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
3613         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
3614         h->field_scan_q0           = field_scan;
3615         h->field_scan8x8_q0        = field_scan8x8;
3616         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
3617     }else{
3618         h->zigzag_scan_q0          = h->zigzag_scan;
3619         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
3620         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
3621         h->field_scan_q0           = h->field_scan;
3622         h->field_scan8x8_q0        = h->field_scan8x8;
3623         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
3624     }
3625 }
3626
3627 static void field_end(H264Context *h){
3628     MpegEncContext * const s = &h->s;
3629     AVCodecContext * const avctx= s->avctx;
3630     s->mb_y= 0;
3631
3632     s->current_picture_ptr->qscale_type= FF_QSCALE_TYPE_H264;
3633     s->current_picture_ptr->pict_type= s->pict_type;
3634
3635     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3636         ff_vdpau_h264_set_reference_frames(s);
3637
3638     if(!s->dropable) {
3639         execute_ref_pic_marking(h, h->mmco, h->mmco_index);
3640         h->prev_poc_msb= h->poc_msb;
3641         h->prev_poc_lsb= h->poc_lsb;
3642     }
3643     h->prev_frame_num_offset= h->frame_num_offset;
3644     h->prev_frame_num= h->frame_num;
3645
3646     if (avctx->hwaccel) {
3647         if (avctx->hwaccel->end_frame(avctx) < 0)
3648             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
3649     }
3650
3651     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3652         ff_vdpau_h264_picture_complete(s);
3653
3654     /*
3655      * FIXME: Error handling code does not seem to support interlaced
3656      * when slices span multiple rows
3657      * The ff_er_add_slice calls don't work right for bottom
3658      * fields; they cause massive erroneous error concealing
3659      * Error marking covers both fields (top and bottom).
3660      * This causes a mismatched s->error_count
3661      * and a bad error table. Further, the error count goes to
3662      * INT_MAX when called for bottom field, because mb_y is
3663      * past end by one (callers fault) and resync_mb_y != 0
3664      * causes problems for the first MB line, too.
3665      */
3666     if (!FIELD_PICTURE)
3667         ff_er_frame_end(s);
3668
3669     MPV_frame_end(s);
3670
3671     h->current_slice=0;
3672 }
3673
3674 /**
3675  * Replicates H264 "master" context to thread contexts.
3676  */
3677 static void clone_slice(H264Context *dst, H264Context *src)
3678 {
3679     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
3680     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
3681     dst->s.current_picture      = src->s.current_picture;
3682     dst->s.linesize             = src->s.linesize;
3683     dst->s.uvlinesize           = src->s.uvlinesize;
3684     dst->s.first_field          = src->s.first_field;
3685
3686     dst->prev_poc_msb           = src->prev_poc_msb;
3687     dst->prev_poc_lsb           = src->prev_poc_lsb;
3688     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
3689     dst->prev_frame_num         = src->prev_frame_num;
3690     dst->short_ref_count        = src->short_ref_count;
3691
3692     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
3693     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
3694     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
3695     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
3696
3697     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
3698     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
3699 }
3700
3701 /**
3702  * decodes a slice header.
3703  * This will also call MPV_common_init() and frame_start() as needed.
3704  *
3705  * @param h h264context
3706  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
3707  *
3708  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
3709  */
3710 static int decode_slice_header(H264Context *h, H264Context *h0){
3711     MpegEncContext * const s = &h->s;
3712     MpegEncContext * const s0 = &h0->s;
3713     unsigned int first_mb_in_slice;
3714     unsigned int pps_id;
3715     int num_ref_idx_active_override_flag;
3716     unsigned int slice_type, tmp, i, j;
3717     int default_ref_list_done = 0;
3718     int last_pic_structure;
3719
3720     s->dropable= h->nal_ref_idc == 0;
3721
3722     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc){
3723         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
3724         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
3725     }else{
3726         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
3727         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
3728     }
3729
3730     first_mb_in_slice= get_ue_golomb(&s->gb);
3731
3732     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
3733         if(h0->current_slice && FIELD_PICTURE){
3734             field_end(h);
3735         }
3736
3737         h0->current_slice = 0;
3738         if (!s0->first_field)
3739             s->current_picture_ptr= NULL;
3740     }
3741
3742     slice_type= get_ue_golomb_31(&s->gb);
3743     if(slice_type > 9){
3744         av_log(h->s.avctx, AV_LOG_ERROR, "slice type too large (%d) at %d %d\n", h->slice_type, s->mb_x, s->mb_y);
3745         return -1;
3746     }
3747     if(slice_type > 4){
3748         slice_type -= 5;
3749         h->slice_type_fixed=1;
3750     }else
3751         h->slice_type_fixed=0;
3752
3753     slice_type= golomb_to_pict_type[ slice_type ];
3754     if (slice_type == FF_I_TYPE
3755         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
3756         default_ref_list_done = 1;
3757     }
3758     h->slice_type= slice_type;
3759     h->slice_type_nos= slice_type & 3;
3760
3761     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
3762     if (s->pict_type == FF_B_TYPE && s0->last_picture_ptr == NULL) {
3763         av_log(h->s.avctx, AV_LOG_ERROR,
3764                "B picture before any references, skipping\n");
3765         return -1;
3766     }
3767
3768     pps_id= get_ue_golomb(&s->gb);
3769     if(pps_id>=MAX_PPS_COUNT){
3770         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
3771         return -1;
3772     }
3773     if(!h0->pps_buffers[pps_id]) {
3774         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
3775         return -1;
3776     }
3777     h->pps= *h0->pps_buffers[pps_id];
3778
3779     if(!h0->sps_buffers[h->pps.sps_id]) {
3780         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
3781         return -1;
3782     }
3783     h->sps = *h0->sps_buffers[h->pps.sps_id];
3784
3785     if(h == h0 && h->dequant_coeff_pps != pps_id){
3786         h->dequant_coeff_pps = pps_id;
3787         init_dequant_tables(h);
3788     }
3789
3790     s->mb_width= h->sps.mb_width;
3791     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
3792
3793     h->b_stride=  s->mb_width*4;
3794     h->b8_stride= s->mb_width*2;
3795
3796     s->width = 16*s->mb_width - 2*FFMIN(h->sps.crop_right, 7);
3797     if(h->sps.frame_mbs_only_flag)
3798         s->height= 16*s->mb_height - 2*FFMIN(h->sps.crop_bottom, 7);
3799     else
3800         s->height= 16*s->mb_height - 4*FFMIN(h->sps.crop_bottom, 3);
3801
3802     if (s->context_initialized
3803         && (   s->width != s->avctx->width || s->height != s->avctx->height)) {
3804         if(h != h0)
3805             return -1;   // width / height changed during parallelized decoding
3806         free_tables(h);
3807         flush_dpb(s->avctx);
3808         MPV_common_end(s);
3809     }
3810     if (!s->context_initialized) {
3811         if(h != h0)
3812             return -1;  // we cant (re-)initialize context during parallel decoding
3813         if (MPV_common_init(s) < 0)
3814             return -1;
3815         s->first_field = 0;
3816         h->prev_interlaced_frame = 1;
3817
3818         init_scan_tables(h);
3819         alloc_tables(h);
3820
3821         for(i = 1; i < s->avctx->thread_count; i++) {
3822             H264Context *c;
3823             c = h->thread_context[i] = av_malloc(sizeof(H264Context));
3824             memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
3825             memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
3826             c->sps = h->sps;
3827             c->pps = h->pps;
3828             init_scan_tables(c);
3829             clone_tables(c, h);
3830         }
3831
3832         for(i = 0; i < s->avctx->thread_count; i++)
3833             if(context_init(h->thread_context[i]) < 0)
3834                 return -1;
3835
3836         s->avctx->width = s->width;
3837         s->avctx->height = s->height;
3838         s->avctx->sample_aspect_ratio= h->sps.sar;
3839         if(!s->avctx->sample_aspect_ratio.den)
3840             s->avctx->sample_aspect_ratio.den = 1;
3841
3842         if(h->sps.timing_info_present_flag){
3843             s->avctx->time_base= (AVRational){h->sps.num_units_in_tick, h->sps.time_scale};
3844             if(h->x264_build > 0 && h->x264_build < 44)
3845                 s->avctx->time_base.den *= 2;
3846             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
3847                       s->avctx->time_base.num, s->avctx->time_base.den, 1<<30);
3848         }
3849     }
3850
3851     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
3852
3853     h->mb_mbaff = 0;
3854     h->mb_aff_frame = 0;
3855     last_pic_structure = s0->picture_structure;
3856     if(h->sps.frame_mbs_only_flag){
3857         s->picture_structure= PICT_FRAME;
3858     }else{
3859         if(get_bits1(&s->gb)) { //field_pic_flag
3860             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
3861         } else {
3862             s->picture_structure= PICT_FRAME;
3863             h->mb_aff_frame = h->sps.mb_aff;
3864         }
3865     }
3866     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
3867
3868     if(h0->current_slice == 0){
3869         while(h->frame_num !=  h->prev_frame_num &&
3870               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
3871             av_log(NULL, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
3872             if (frame_start(h) < 0)
3873                 return -1;
3874             h->prev_frame_num++;
3875             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
3876             s->current_picture_ptr->frame_num= h->prev_frame_num;
3877             execute_ref_pic_marking(h, NULL, 0);
3878         }
3879
3880         /* See if we have a decoded first field looking for a pair... */
3881         if (s0->first_field) {
3882             assert(s0->current_picture_ptr);
3883             assert(s0->current_picture_ptr->data[0]);
3884             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
3885
3886             /* figure out if we have a complementary field pair */
3887             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3888                 /*
3889                  * Previous field is unmatched. Don't display it, but let it
3890                  * remain for reference if marked as such.
3891                  */
3892                 s0->current_picture_ptr = NULL;
3893                 s0->first_field = FIELD_PICTURE;
3894
3895             } else {
3896                 if (h->nal_ref_idc &&
3897                         s0->current_picture_ptr->reference &&
3898                         s0->current_picture_ptr->frame_num != h->frame_num) {
3899                     /*
3900                      * This and previous field were reference, but had
3901                      * different frame_nums. Consider this field first in
3902                      * pair. Throw away previous field except for reference
3903                      * purposes.
3904                      */
3905                     s0->first_field = 1;
3906                     s0->current_picture_ptr = NULL;
3907
3908                 } else {
3909                     /* Second field in complementary pair */
3910                     s0->first_field = 0;
3911                 }
3912             }
3913
3914         } else {
3915             /* Frame or first field in a potentially complementary pair */
3916             assert(!s0->current_picture_ptr);
3917             s0->first_field = FIELD_PICTURE;
3918         }
3919
3920         if((!FIELD_PICTURE || s0->first_field) && frame_start(h) < 0) {
3921             s0->first_field = 0;
3922             return -1;
3923         }
3924     }
3925     if(h != h0)
3926         clone_slice(h, h0);
3927
3928     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
3929
3930     assert(s->mb_num == s->mb_width * s->mb_height);
3931     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
3932        first_mb_in_slice                    >= s->mb_num){
3933         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
3934         return -1;
3935     }
3936     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
3937     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
3938     if (s->picture_structure == PICT_BOTTOM_FIELD)
3939         s->resync_mb_y = s->mb_y = s->mb_y + 1;
3940     assert(s->mb_y < s->mb_height);
3941
3942     if(s->picture_structure==PICT_FRAME){
3943         h->curr_pic_num=   h->frame_num;
3944         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
3945     }else{
3946         h->curr_pic_num= 2*h->frame_num + 1;
3947         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
3948     }
3949
3950     if(h->nal_unit_type == NAL_IDR_SLICE){
3951         get_ue_golomb(&s->gb); /* idr_pic_id */
3952     }
3953
3954     if(h->sps.poc_type==0){
3955         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
3956
3957         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
3958             h->delta_poc_bottom= get_se_golomb(&s->gb);
3959         }
3960     }
3961
3962     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
3963         h->delta_poc[0]= get_se_golomb(&s->gb);
3964
3965         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
3966             h->delta_poc[1]= get_se_golomb(&s->gb);
3967     }
3968
3969     init_poc(h);
3970
3971     if(h->pps.redundant_pic_cnt_present){
3972         h->redundant_pic_count= get_ue_golomb(&s->gb);
3973     }
3974
3975     //set defaults, might be overridden a few lines later
3976     h->ref_count[0]= h->pps.ref_count[0];
3977     h->ref_count[1]= h->pps.ref_count[1];
3978
3979     if(h->slice_type_nos != FF_I_TYPE){
3980         if(h->slice_type_nos == FF_B_TYPE){
3981             h->direct_spatial_mv_pred= get_bits1(&s->gb);
3982         }
3983         num_ref_idx_active_override_flag= get_bits1(&s->gb);
3984
3985         if(num_ref_idx_active_override_flag){
3986             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
3987             if(h->slice_type_nos==FF_B_TYPE)
3988                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
3989
3990             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
3991                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3992                 h->ref_count[0]= h->ref_count[1]= 1;
3993                 return -1;
3994             }
3995         }
3996         if(h->slice_type_nos == FF_B_TYPE)
3997             h->list_count= 2;
3998         else
3999             h->list_count= 1;
4000     }else
4001         h->list_count= 0;
4002
4003     if(!default_ref_list_done){
4004         fill_default_ref_list(h);
4005     }
4006
4007     if(h->slice_type_nos!=FF_I_TYPE && decode_ref_pic_list_reordering(h) < 0)
4008         return -1;
4009
4010     if(h->slice_type_nos!=FF_I_TYPE){
4011         s->last_picture_ptr= &h->ref_list[0][0];
4012         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
4013     }
4014     if(h->slice_type_nos==FF_B_TYPE){
4015         s->next_picture_ptr= &h->ref_list[1][0];
4016         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
4017     }
4018
4019     if(   (h->pps.weighted_pred          && h->slice_type_nos == FF_P_TYPE )
4020        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== FF_B_TYPE ) )
4021         pred_weight_table(h);
4022     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== FF_B_TYPE)
4023         implicit_weight_table(h);
4024     else {
4025         h->use_weight = 0;
4026         for (i = 0; i < 2; i++) {
4027             h->luma_weight_flag[i]   = 0;
4028             h->chroma_weight_flag[i] = 0;
4029         }
4030     }
4031
4032     if(h->nal_ref_idc)
4033         decode_ref_pic_marking(h0, &s->gb);
4034
4035     if(FRAME_MBAFF)
4036         fill_mbaff_ref_list(h);
4037
4038     if(h->slice_type_nos==FF_B_TYPE && !h->direct_spatial_mv_pred)
4039         direct_dist_scale_factor(h);
4040     direct_ref_list_init(h);
4041
4042     if( h->slice_type_nos != FF_I_TYPE && h->pps.cabac ){
4043       &nbs