8a7550f2a6d996b3386d5be9a0ac4f4b7337c1b0
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file h264.c
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "dsputil.h"
29 #include "avcodec.h"
30 #include "mpegvideo.h"
31 #include "h264.h"
32 #include "h264data.h"
33 #include "h264_parser.h"
34 #include "golomb.h"
35 #include "rectangle.h"
36
37 #include "cabac.h"
38 #ifdef ARCH_X86
39 #include "x86/h264_i386.h"
40 #endif
41
42 //#undef NDEBUG
43 #include <assert.h>
44
45 /**
46  * Value of Picture.reference when Picture is not a reference picture, but
47  * is held for delayed output.
48  */
49 #define DELAYED_PIC_REF 4
50
51 static VLC coeff_token_vlc[4];
52 static VLC_TYPE coeff_token_vlc_tables[520+332+280+256][2];
53 static const int coeff_token_vlc_tables_size[4]={520,332,280,256};
54
55 static VLC chroma_dc_coeff_token_vlc;
56 static VLC_TYPE chroma_dc_coeff_token_vlc_table[256][2];
57 static const int chroma_dc_coeff_token_vlc_table_size = 256;
58
59 static VLC total_zeros_vlc[15];
60 static VLC_TYPE total_zeros_vlc_tables[15][512][2];
61 static const int total_zeros_vlc_tables_size = 512;
62
63 static VLC chroma_dc_total_zeros_vlc[3];
64 static VLC_TYPE chroma_dc_total_zeros_vlc_tables[3][8][2];
65 static const int chroma_dc_total_zeros_vlc_tables_size = 8;
66
67 static VLC run_vlc[6];
68 static VLC_TYPE run_vlc_tables[6][8][2];
69 static const int run_vlc_tables_size = 8;
70
71 static VLC run7_vlc;
72 static VLC_TYPE run7_vlc_table[96][2];
73 static const int run7_vlc_table_size = 96;
74
75 static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
76 static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
77 static void filter_mb( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
78 static void filter_mb_fast( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
79 static Picture * remove_long(H264Context *h, int i, int ref_mask);
80
81 static av_always_inline uint32_t pack16to32(int a, int b){
82 #ifdef WORDS_BIGENDIAN
83    return (b&0xFFFF) + (a<<16);
84 #else
85    return (a&0xFFFF) + (b<<16);
86 #endif
87 }
88
89 static const uint8_t rem6[52]={
90 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
91 };
92
93 static const uint8_t div6[52]={
94 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8,
95 };
96
97 static const int left_block_options[4][8]={
98     {0,1,2,3,7,10,8,11},
99     {2,2,3,3,8,11,8,11},
100     {0,0,1,1,7,10,7,10},
101     {0,2,0,2,7,10,7,10}
102 };
103
104 #define LEVEL_TAB_BITS 8
105 static int8_t cavlc_level_tab[7][1<<LEVEL_TAB_BITS][2];
106
107 static void fill_caches(H264Context *h, int mb_type, int for_deblock){
108     MpegEncContext * const s = &h->s;
109     const int mb_xy= h->mb_xy;
110     int topleft_xy, top_xy, topright_xy, left_xy[2];
111     int topleft_type, top_type, topright_type, left_type[2];
112     const int * left_block;
113     int topleft_partition= -1;
114     int i;
115
116     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
117
118     //FIXME deblocking could skip the intra and nnz parts.
119     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
120         return;
121
122     /* Wow, what a mess, why didn't they simplify the interlacing & intra
123      * stuff, I can't imagine that these complex rules are worth it. */
124
125     topleft_xy = top_xy - 1;
126     topright_xy= top_xy + 1;
127     left_xy[1] = left_xy[0] = mb_xy-1;
128     left_block = left_block_options[0];
129     if(FRAME_MBAFF){
130         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
131         const int top_pair_xy      = pair_xy     - s->mb_stride;
132         const int topleft_pair_xy  = top_pair_xy - 1;
133         const int topright_pair_xy = top_pair_xy + 1;
134         const int topleft_mb_field_flag  = IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
135         const int top_mb_field_flag      = IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
136         const int topright_mb_field_flag = IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
137         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
138         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
139         const int bottom = (s->mb_y & 1);
140         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);
141
142         if (curr_mb_field_flag && (bottom || top_mb_field_flag)){
143             top_xy -= s->mb_stride;
144         }
145         if (curr_mb_field_flag && (bottom || topleft_mb_field_flag)){
146             topleft_xy -= s->mb_stride;
147         } else if(bottom && !curr_mb_field_flag && left_mb_field_flag) {
148             topleft_xy += s->mb_stride;
149             // take top left mv from the middle of the mb, as opposed to all other modes which use the bottom right partition
150             topleft_partition = 0;
151         }
152         if (curr_mb_field_flag && (bottom || topright_mb_field_flag)){
153             topright_xy -= s->mb_stride;
154         }
155         if (left_mb_field_flag != curr_mb_field_flag) {
156             left_xy[1] = left_xy[0] = pair_xy - 1;
157             if (curr_mb_field_flag) {
158                 left_xy[1] += s->mb_stride;
159                 left_block = left_block_options[3];
160             } else {
161                 left_block= left_block_options[2 - bottom];
162             }
163         }
164     }
165
166     h->top_mb_xy = top_xy;
167     h->left_mb_xy[0] = left_xy[0];
168     h->left_mb_xy[1] = left_xy[1];
169     if(for_deblock){
170         topleft_type = 0;
171         topright_type = 0;
172         top_type     = h->slice_table[top_xy     ] < 0xFFFF ? s->current_picture.mb_type[top_xy]     : 0;
173         left_type[0] = h->slice_table[left_xy[0] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[0]] : 0;
174         left_type[1] = h->slice_table[left_xy[1] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[1]] : 0;
175
176         if(MB_MBAFF && !IS_INTRA(mb_type)){
177             int list;
178             for(list=0; list<h->list_count; list++){
179                 //These values where changed for ease of performing MC, we need to change them back
180                 //FIXME maybe we can make MC and loop filter use the same values or prevent
181                 //the MC code from changing ref_cache and rather use a temporary array.
182                 if(USES_LIST(mb_type,list)){
183                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
184                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
185                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
186                     ref += h->b8_stride;
187                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
188                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
189                 }
190             }
191         }
192     }else{
193         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
194         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
195         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
196         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
197         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
198
199     if(IS_INTRA(mb_type)){
200         int type_mask= h->pps.constrained_intra_pred ? IS_INTRA(-1) : -1;
201         h->topleft_samples_available=
202         h->top_samples_available=
203         h->left_samples_available= 0xFFFF;
204         h->topright_samples_available= 0xEEEA;
205
206         if(!(top_type & type_mask)){
207             h->topleft_samples_available= 0xB3FF;
208             h->top_samples_available= 0x33FF;
209             h->topright_samples_available= 0x26EA;
210         }
211         if(IS_INTERLACED(mb_type) != IS_INTERLACED(left_type[0])){
212             if(IS_INTERLACED(mb_type)){
213                 if(!(left_type[0] & type_mask)){
214                     h->topleft_samples_available&= 0xDFFF;
215                     h->left_samples_available&= 0x5FFF;
216                 }
217                 if(!(left_type[1] & type_mask)){
218                     h->topleft_samples_available&= 0xFF5F;
219                     h->left_samples_available&= 0xFF5F;
220                 }
221             }else{
222                 int left_typei = h->slice_table[left_xy[0] + s->mb_stride ] == h->slice_num
223                                 ? s->current_picture.mb_type[left_xy[0] + s->mb_stride] : 0;
224                 assert(left_xy[0] == left_xy[1]);
225                 if(!((left_typei & type_mask) && (left_type[0] & type_mask))){
226                     h->topleft_samples_available&= 0xDF5F;
227                     h->left_samples_available&= 0x5F5F;
228                 }
229             }
230         }else{
231             if(!(left_type[0] & type_mask)){
232                 h->topleft_samples_available&= 0xDF5F;
233                 h->left_samples_available&= 0x5F5F;
234             }
235         }
236
237         if(!(topleft_type & type_mask))
238             h->topleft_samples_available&= 0x7FFF;
239
240         if(!(topright_type & type_mask))
241             h->topright_samples_available&= 0xFBFF;
242
243         if(IS_INTRA4x4(mb_type)){
244             if(IS_INTRA4x4(top_type)){
245                 h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
246                 h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
247                 h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
248                 h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
249             }else{
250                 int pred;
251                 if(!(top_type & type_mask))
252                     pred= -1;
253                 else{
254                     pred= 2;
255                 }
256                 h->intra4x4_pred_mode_cache[4+8*0]=
257                 h->intra4x4_pred_mode_cache[5+8*0]=
258                 h->intra4x4_pred_mode_cache[6+8*0]=
259                 h->intra4x4_pred_mode_cache[7+8*0]= pred;
260             }
261             for(i=0; i<2; i++){
262                 if(IS_INTRA4x4(left_type[i])){
263                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
264                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
265                 }else{
266                     int pred;
267                     if(!(left_type[i] & type_mask))
268                         pred= -1;
269                     else{
270                         pred= 2;
271                     }
272                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
273                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
274                 }
275             }
276         }
277     }
278     }
279
280
281 /*
282 0 . T T. T T T T
283 1 L . .L . . . .
284 2 L . .L . . . .
285 3 . T TL . . . .
286 4 L . .L . . . .
287 5 L . .. . . . .
288 */
289 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
290     if(top_type){
291         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
292         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
293         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
294         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
295
296         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
297         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
298
299         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
300         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
301
302     }else{
303         h->non_zero_count_cache[4+8*0]=
304         h->non_zero_count_cache[5+8*0]=
305         h->non_zero_count_cache[6+8*0]=
306         h->non_zero_count_cache[7+8*0]=
307
308         h->non_zero_count_cache[1+8*0]=
309         h->non_zero_count_cache[2+8*0]=
310
311         h->non_zero_count_cache[1+8*3]=
312         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
313
314     }
315
316     for (i=0; i<2; i++) {
317         if(left_type[i]){
318             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
319             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
320             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
321             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
322         }else{
323             h->non_zero_count_cache[3+8*1 + 2*8*i]=
324             h->non_zero_count_cache[3+8*2 + 2*8*i]=
325             h->non_zero_count_cache[0+8*1 +   8*i]=
326             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
327         }
328     }
329
330     if( h->pps.cabac ) {
331         // top_cbp
332         if(top_type) {
333             h->top_cbp = h->cbp_table[top_xy];
334         } else if(IS_INTRA(mb_type)) {
335             h->top_cbp = 0x1C0;
336         } else {
337             h->top_cbp = 0;
338         }
339         // left_cbp
340         if (left_type[0]) {
341             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
342         } else if(IS_INTRA(mb_type)) {
343             h->left_cbp = 0x1C0;
344         } else {
345             h->left_cbp = 0;
346         }
347         if (left_type[0]) {
348             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
349         }
350         if (left_type[1]) {
351             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
352         }
353     }
354
355 #if 1
356     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
357         int list;
358         for(list=0; list<h->list_count; list++){
359             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
360                 /*if(!h->mv_cache_clean[list]){
361                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
362                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
363                     h->mv_cache_clean[list]= 1;
364                 }*/
365                 continue;
366             }
367             h->mv_cache_clean[list]= 0;
368
369             if(USES_LIST(top_type, list)){
370                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
371                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
372                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
373                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
374                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
375                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
376                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
377                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
378                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
379                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
380             }else{
381                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
382                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
383                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
384                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
385                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
386             }
387
388             for(i=0; i<2; i++){
389                 int cache_idx = scan8[0] - 1 + i*2*8;
390                 if(USES_LIST(left_type[i], list)){
391                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
392                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
393                     *(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]];
394                     *(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]];
395                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
396                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
397                 }else{
398                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
399                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
400                     h->ref_cache[list][cache_idx  ]=
401                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
402                 }
403             }
404
405             if(for_deblock || ((IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred) && !FRAME_MBAFF))
406                 continue;
407
408             if(USES_LIST(topleft_type, list)){
409                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
410                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
411                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
412                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
413             }else{
414                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
415                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
416             }
417
418             if(USES_LIST(topright_type, list)){
419                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
420                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
421                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
422                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
423             }else{
424                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
425                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
426             }
427
428             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
429                 continue;
430
431             h->ref_cache[list][scan8[5 ]+1] =
432             h->ref_cache[list][scan8[7 ]+1] =
433             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
434             h->ref_cache[list][scan8[4 ]] =
435             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
436             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
437             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
438             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
439             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
440             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
441
442             if( h->pps.cabac ) {
443                 /* XXX beurk, Load mvd */
444                 if(USES_LIST(top_type, list)){
445                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
446                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
447                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
448                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
449                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
450                 }else{
451                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
452                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
453                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
454                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
455                 }
456                 if(USES_LIST(left_type[0], list)){
457                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
458                     *(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]];
459                     *(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]];
460                 }else{
461                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
462                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
463                 }
464                 if(USES_LIST(left_type[1], list)){
465                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
466                     *(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]];
467                     *(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]];
468                 }else{
469                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
470                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
471                 }
472                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
473                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
474                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
475                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
476                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
477
478                 if(h->slice_type_nos == FF_B_TYPE){
479                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
480
481                     if(IS_DIRECT(top_type)){
482                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
483                     }else if(IS_8X8(top_type)){
484                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
485                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
486                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
487                     }else{
488                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
489                     }
490
491                     if(IS_DIRECT(left_type[0]))
492                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
493                     else if(IS_8X8(left_type[0]))
494                         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)];
495                     else
496                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
497
498                     if(IS_DIRECT(left_type[1]))
499                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
500                     else if(IS_8X8(left_type[1]))
501                         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)];
502                     else
503                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
504                 }
505             }
506
507             if(FRAME_MBAFF){
508 #define MAP_MVS\
509                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
510                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
511                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
512                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
513                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
514                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
515                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
516                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
517                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
518                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
519                 if(MB_FIELD){
520 #define MAP_F2F(idx, mb_type)\
521                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
522                         h->ref_cache[list][idx] <<= 1;\
523                         h->mv_cache[list][idx][1] /= 2;\
524                         h->mvd_cache[list][idx][1] /= 2;\
525                     }
526                     MAP_MVS
527 #undef MAP_F2F
528                 }else{
529 #define MAP_F2F(idx, mb_type)\
530                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
531                         h->ref_cache[list][idx] >>= 1;\
532                         h->mv_cache[list][idx][1] <<= 1;\
533                         h->mvd_cache[list][idx][1] <<= 1;\
534                     }
535                     MAP_MVS
536 #undef MAP_F2F
537                 }
538             }
539         }
540     }
541 #endif
542
543     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
544 }
545
546 static inline void write_back_intra_pred_mode(H264Context *h){
547     const int mb_xy= h->mb_xy;
548
549     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
550     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
551     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
552     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
553     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
554     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
555     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
556 }
557
558 /**
559  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
560  */
561 static inline int check_intra4x4_pred_mode(H264Context *h){
562     MpegEncContext * const s = &h->s;
563     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
564     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
565     int i;
566
567     if(!(h->top_samples_available&0x8000)){
568         for(i=0; i<4; i++){
569             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
570             if(status<0){
571                 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);
572                 return -1;
573             } else if(status){
574                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
575             }
576         }
577     }
578
579     if((h->left_samples_available&0x8888)!=0x8888){
580         static const int mask[4]={0x8000,0x2000,0x80,0x20};
581         for(i=0; i<4; i++){
582             if(!(h->left_samples_available&mask[i])){
583                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
584                 if(status<0){
585                     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);
586                     return -1;
587                 } else if(status){
588                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
589                 }
590             }
591         }
592     }
593
594     return 0;
595 } //FIXME cleanup like next
596
597 /**
598  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
599  */
600 static inline int check_intra_pred_mode(H264Context *h, int mode){
601     MpegEncContext * const s = &h->s;
602     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
603     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
604
605     if(mode > 6U) {
606         av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
607         return -1;
608     }
609
610     if(!(h->top_samples_available&0x8000)){
611         mode= top[ mode ];
612         if(mode<0){
613             av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
614             return -1;
615         }
616     }
617
618     if((h->left_samples_available&0x8080) != 0x8080){
619         mode= left[ mode ];
620         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
621             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
622         }
623         if(mode<0){
624             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);
625             return -1;
626         }
627     }
628
629     return mode;
630 }
631
632 /**
633  * gets the predicted intra4x4 prediction mode.
634  */
635 static inline int pred_intra_mode(H264Context *h, int n){
636     const int index8= scan8[n];
637     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
638     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
639     const int min= FFMIN(left, top);
640
641     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
642
643     if(min<0) return DC_PRED;
644     else      return min;
645 }
646
647 static inline void write_back_non_zero_count(H264Context *h){
648     const int mb_xy= h->mb_xy;
649
650     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
651     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
652     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
653     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
654     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
655     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
656     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
657
658     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
659     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
660     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
661
662     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
663     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
664     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
665 }
666
667 /**
668  * gets the predicted number of non-zero coefficients.
669  * @param n block index
670  */
671 static inline int pred_non_zero_count(H264Context *h, int n){
672     const int index8= scan8[n];
673     const int left= h->non_zero_count_cache[index8 - 1];
674     const int top = h->non_zero_count_cache[index8 - 8];
675     int i= left + top;
676
677     if(i<64) i= (i+1)>>1;
678
679     tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);
680
681     return i&31;
682 }
683
684 static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
685     const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
686     MpegEncContext *s = &h->s;
687
688     /* there is no consistent mapping of mvs to neighboring locations that will
689      * make mbaff happy, so we can't move all this logic to fill_caches */
690     if(FRAME_MBAFF){
691         const uint32_t *mb_types = s->current_picture_ptr->mb_type;
692         const int16_t *mv;
693         *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
694         *C = h->mv_cache[list][scan8[0]-2];
695
696         if(!MB_FIELD
697            && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
698             int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
699             if(IS_INTERLACED(mb_types[topright_xy])){
700 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
701                 const int x4 = X4, y4 = Y4;\
702                 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
703                 if(!USES_LIST(mb_type,list))\
704                     return LIST_NOT_USED;\
705                 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
706                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
707                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
708                 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
709
710                 SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
711             }
712         }
713         if(topright_ref == PART_NOT_AVAILABLE
714            && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
715            && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
716             if(!MB_FIELD
717                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
718                 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
719             }
720             if(MB_FIELD
721                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
722                && i >= scan8[0]+8){
723                 // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
724                 SET_DIAG_MV(/2, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
725             }
726         }
727 #undef SET_DIAG_MV
728     }
729
730     if(topright_ref != PART_NOT_AVAILABLE){
731         *C= h->mv_cache[list][ i - 8 + part_width ];
732         return topright_ref;
733     }else{
734         tprintf(s->avctx, "topright MV not available\n");
735
736         *C= h->mv_cache[list][ i - 8 - 1 ];
737         return h->ref_cache[list][ i - 8 - 1 ];
738     }
739 }
740
741 /**
742  * gets the predicted MV.
743  * @param n the block index
744  * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
745  * @param mx the x component of the predicted motion vector
746  * @param my the y component of the predicted motion vector
747  */
748 static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
749     const int index8= scan8[n];
750     const int top_ref=      h->ref_cache[list][ index8 - 8 ];
751     const int left_ref=     h->ref_cache[list][ index8 - 1 ];
752     const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
753     const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
754     const int16_t * C;
755     int diagonal_ref, match_count;
756
757     assert(part_width==1 || part_width==2 || part_width==4);
758
759 /* mv_cache
760   B . . A T T T T
761   U . . L . . , .
762   U . . L . . . .
763   U . . L . . , .
764   . . . L . . . .
765 */
766
767     diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
768     match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
769     tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
770     if(match_count > 1){ //most common
771         *mx= mid_pred(A[0], B[0], C[0]);
772         *my= mid_pred(A[1], B[1], C[1]);
773     }else if(match_count==1){
774         if(left_ref==ref){
775             *mx= A[0];
776             *my= A[1];
777         }else if(top_ref==ref){
778             *mx= B[0];
779             *my= B[1];
780         }else{
781             *mx= C[0];
782             *my= C[1];
783         }
784     }else{
785         if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
786             *mx= A[0];
787             *my= A[1];
788         }else{
789             *mx= mid_pred(A[0], B[0], C[0]);
790             *my= mid_pred(A[1], B[1], C[1]);
791         }
792     }
793
794     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);
795 }
796
797 /**
798  * gets the directionally predicted 16x8 MV.
799  * @param n the block index
800  * @param mx the x component of the predicted motion vector
801  * @param my the y component of the predicted motion vector
802  */
803 static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
804     if(n==0){
805         const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
806         const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
807
808         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);
809
810         if(top_ref == ref){
811             *mx= B[0];
812             *my= B[1];
813             return;
814         }
815     }else{
816         const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
817         const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];
818
819         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);
820
821         if(left_ref == ref){
822             *mx= A[0];
823             *my= A[1];
824             return;
825         }
826     }
827
828     //RARE
829     pred_motion(h, n, 4, list, ref, mx, my);
830 }
831
832 /**
833  * gets the directionally predicted 8x16 MV.
834  * @param n the block index
835  * @param mx the x component of the predicted motion vector
836  * @param my the y component of the predicted motion vector
837  */
838 static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
839     if(n==0){
840         const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
841         const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];
842
843         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);
844
845         if(left_ref == ref){
846             *mx= A[0];
847             *my= A[1];
848             return;
849         }
850     }else{
851         const int16_t * C;
852         int diagonal_ref;
853
854         diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);
855
856         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);
857
858         if(diagonal_ref == ref){
859             *mx= C[0];
860             *my= C[1];
861             return;
862         }
863     }
864
865     //RARE
866     pred_motion(h, n, 2, list, ref, mx, my);
867 }
868
869 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
870     const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
871     const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];
872
873     tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);
874
875     if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
876        || !( top_ref | *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ])
877        || !(left_ref | *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ])){
878
879         *mx = *my = 0;
880         return;
881     }
882
883     pred_motion(h, 0, 4, 0, 0, mx, my);
884
885     return;
886 }
887
888 static int get_scale_factor(H264Context * const h, int poc, int poc1, int i){
889     int poc0 = h->ref_list[0][i].poc;
890     int td = av_clip(poc1 - poc0, -128, 127);
891     if(td == 0 || h->ref_list[0][i].long_ref){
892         return 256;
893     }else{
894         int tb = av_clip(poc - poc0, -128, 127);
895         int tx = (16384 + (FFABS(td) >> 1)) / td;
896         return av_clip((tb*tx + 32) >> 6, -1024, 1023);
897     }
898 }
899
900 static inline void direct_dist_scale_factor(H264Context * const h){
901     MpegEncContext * const s = &h->s;
902     const int poc = h->s.current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
903     const int poc1 = h->ref_list[1][0].poc;
904     int i, field;
905     for(field=0; field<2; field++){
906         const int poc  = h->s.current_picture_ptr->field_poc[field];
907         const int poc1 = h->ref_list[1][0].field_poc[field];
908         for(i=0; i < 2*h->ref_count[0]; i++)
909             h->dist_scale_factor_field[field][i^field] = get_scale_factor(h, poc, poc1, i+16);
910     }
911
912     for(i=0; i<h->ref_count[0]; i++){
913         h->dist_scale_factor[i] = get_scale_factor(h, poc, poc1, i);
914     }
915 }
916
917 static void fill_colmap(H264Context *h, int map[2][16+32], int list, int field, int colfield, int mbafi){
918     MpegEncContext * const s = &h->s;
919     Picture * const ref1 = &h->ref_list[1][0];
920     int j, old_ref, rfield;
921     int start= mbafi ? 16                      : 0;
922     int end  = mbafi ? 16+2*h->ref_count[list] : h->ref_count[list];
923     int interl= mbafi || s->picture_structure != PICT_FRAME;
924
925     /* bogus; fills in for missing frames */
926     memset(map[list], 0, sizeof(map[list]));
927
928     for(rfield=0; rfield<2; rfield++){
929         for(old_ref=0; old_ref<ref1->ref_count[colfield][list]; old_ref++){
930             int poc = ref1->ref_poc[colfield][list][old_ref];
931
932             if     (!interl)
933                 poc |= 3;
934             else if( interl && (poc&3) == 3) //FIXME store all MBAFF references so this isnt needed
935                 poc= (poc&~3) + rfield + 1;
936
937             for(j=start; j<end; j++){
938                 if(4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3) == poc){
939                     int cur_ref= mbafi ? (j-16)^field : j;
940                     map[list][2*old_ref + (rfield^field) + 16] = cur_ref;
941                     if(rfield == field)
942                         map[list][old_ref] = cur_ref;
943                     break;
944                 }
945             }
946         }
947     }
948 }
949
950 static inline void direct_ref_list_init(H264Context * const h){
951     MpegEncContext * const s = &h->s;
952     Picture * const ref1 = &h->ref_list[1][0];
953     Picture * const cur = s->current_picture_ptr;
954     int list, j, field;
955     int sidx= (s->picture_structure&1)^1;
956     int ref1sidx= (ref1->reference&1)^1;
957
958     for(list=0; list<2; list++){
959         cur->ref_count[sidx][list] = h->ref_count[list];
960         for(j=0; j<h->ref_count[list]; j++)
961             cur->ref_poc[sidx][list][j] = 4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3);
962     }
963
964     if(s->picture_structure == PICT_FRAME){
965         memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0]));
966         memcpy(cur->ref_poc  [1], cur->ref_poc  [0], sizeof(cur->ref_poc  [0]));
967     }
968
969     cur->mbaff= FRAME_MBAFF;
970
971     if(cur->pict_type != FF_B_TYPE || h->direct_spatial_mv_pred)
972         return;
973
974     for(list=0; list<2; list++){
975         fill_colmap(h, h->map_col_to_list0, list, sidx, ref1sidx, 0);
976         for(field=0; field<2; field++)
977             fill_colmap(h, h->map_col_to_list0_field[field], list, field, field, 1);
978     }
979 }
980
981 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
982     MpegEncContext * const s = &h->s;
983     int b8_stride = h->b8_stride;
984     int b4_stride = h->b_stride;
985     int mb_xy = h->mb_xy;
986     int mb_type_col[2];
987     const int16_t (*l1mv0)[2], (*l1mv1)[2];
988     const int8_t *l1ref0, *l1ref1;
989     const int is_b8x8 = IS_8X8(*mb_type);
990     unsigned int sub_mb_type;
991     int i8, i4;
992
993 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
994
995     if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL
996         if(!IS_INTERLACED(*mb_type)){                    //     AFR/FR    -> AFL/FL
997             int cur_poc = s->current_picture_ptr->poc;
998             int *col_poc = h->ref_list[1]->field_poc;
999             int col_parity = FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc);
1000             mb_xy= s->mb_x + ((s->mb_y&~1) + col_parity)*s->mb_stride;
1001             b8_stride = 0;
1002         }else if(!(s->picture_structure & h->ref_list[1][0].reference) && !h->ref_list[1][0].mbaff){// FL -> FL & differ parity
1003             int fieldoff= 2*(h->ref_list[1][0].reference)-3;
1004             mb_xy += s->mb_stride*fieldoff;
1005         }
1006         goto single_col;
1007     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
1008         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
1009             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
1010             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
1011             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
1012             b8_stride *= 3;
1013             b4_stride *= 6;
1014             //FIXME IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag
1015             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
1016                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
1017                 && !is_b8x8){
1018                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1019                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
1020             }else{
1021                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1022                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1023             }
1024         }else{                                           //     AFR/FR    -> AFR/FR
1025 single_col:
1026             mb_type_col[0] =
1027             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
1028             if(IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag){
1029                 /* FIXME save sub mb types from previous frames (or derive from MVs)
1030                 * so we know exactly what block size to use */
1031                 sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
1032                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1033             }else if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
1034                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1035                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
1036             }else{
1037                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1038                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1039             }
1040         }
1041     }
1042
1043     l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
1044     l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
1045     l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
1046     l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
1047     if(!b8_stride){
1048         if(s->mb_y&1){
1049             l1ref0 += h->b8_stride;
1050             l1ref1 += h->b8_stride;
1051             l1mv0  +=  2*b4_stride;
1052             l1mv1  +=  2*b4_stride;
1053         }
1054     }
1055
1056     if(h->direct_spatial_mv_pred){
1057         int ref[2];
1058         int mv[2][2];
1059         int list;
1060
1061         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
1062
1063         /* ref = min(neighbors) */
1064         for(list=0; list<2; list++){
1065             int refa = h->ref_cache[list][scan8[0] - 1];
1066             int refb = h->ref_cache[list][scan8[0] - 8];
1067             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1068             if(refc == PART_NOT_AVAILABLE)
1069                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1070             ref[list] = FFMIN3((unsigned)refa, (unsigned)refb, (unsigned)refc);
1071             if(ref[list] < 0)
1072                 ref[list] = -1;
1073         }
1074
1075         if(ref[0] < 0 && ref[1] < 0){
1076             ref[0] = ref[1] = 0;
1077             mv[0][0] = mv[0][1] =
1078             mv[1][0] = mv[1][1] = 0;
1079         }else{
1080             for(list=0; list<2; list++){
1081                 if(ref[list] >= 0)
1082                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1083                 else
1084                     mv[list][0] = mv[list][1] = 0;
1085             }
1086         }
1087
1088         if(ref[1] < 0){
1089             if(!is_b8x8)
1090                 *mb_type &= ~MB_TYPE_L1;
1091             sub_mb_type &= ~MB_TYPE_L1;
1092         }else if(ref[0] < 0){
1093             if(!is_b8x8)
1094                 *mb_type &= ~MB_TYPE_L0;
1095             sub_mb_type &= ~MB_TYPE_L0;
1096         }
1097
1098         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1099             for(i8=0; i8<4; i8++){
1100                 int x8 = i8&1;
1101                 int y8 = i8>>1;
1102                 int xy8 = x8+y8*b8_stride;
1103                 int xy4 = 3*x8+y8*b4_stride;
1104                 int a=0, b=0;
1105
1106                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1107                     continue;
1108                 h->sub_mb_type[i8] = sub_mb_type;
1109
1110                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1111                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1112                 if(!IS_INTRA(mb_type_col[y8])
1113                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
1114                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
1115                     if(ref[0] > 0)
1116                         a= pack16to32(mv[0][0],mv[0][1]);
1117                     if(ref[1] > 0)
1118                         b= pack16to32(mv[1][0],mv[1][1]);
1119                 }else{
1120                     a= pack16to32(mv[0][0],mv[0][1]);
1121                     b= pack16to32(mv[1][0],mv[1][1]);
1122                 }
1123                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
1124                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
1125             }
1126         }else if(IS_16X16(*mb_type)){
1127             int a=0, b=0;
1128
1129             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1130             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1131             if(!IS_INTRA(mb_type_col[0])
1132                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1133                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1134                        && (h->x264_build>33 || !h->x264_build)))){
1135                 if(ref[0] > 0)
1136                     a= pack16to32(mv[0][0],mv[0][1]);
1137                 if(ref[1] > 0)
1138                     b= pack16to32(mv[1][0],mv[1][1]);
1139             }else{
1140                 a= pack16to32(mv[0][0],mv[0][1]);
1141                 b= pack16to32(mv[1][0],mv[1][1]);
1142             }
1143             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1144             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1145         }else{
1146             for(i8=0; i8<4; i8++){
1147                 const int x8 = i8&1;
1148                 const int y8 = i8>>1;
1149
1150                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1151                     continue;
1152                 h->sub_mb_type[i8] = sub_mb_type;
1153
1154                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1155                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1156                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1157                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1158
1159                 /* col_zero_flag */
1160                 if(!IS_INTRA(mb_type_col[0]) && (   l1ref0[x8 + y8*b8_stride] == 0
1161                                               || (l1ref0[x8 + y8*b8_stride] < 0 && l1ref1[x8 + y8*b8_stride] == 0
1162                                                   && (h->x264_build>33 || !h->x264_build)))){
1163                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*b8_stride] == 0 ? l1mv0 : l1mv1;
1164                     if(IS_SUB_8X8(sub_mb_type)){
1165                         const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1166                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1167                             if(ref[0] == 0)
1168                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1169                             if(ref[1] == 0)
1170                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1171                         }
1172                     }else
1173                     for(i4=0; i4<4; i4++){
1174                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1175                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1176                             if(ref[0] == 0)
1177                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1178                             if(ref[1] == 0)
1179                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1180                         }
1181                     }
1182                 }
1183             }
1184         }
1185     }else{ /* direct temporal mv pred */
1186         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1187         const int *dist_scale_factor = h->dist_scale_factor;
1188         int ref_offset= 0;
1189
1190         if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
1191             map_col_to_list0[0] = h->map_col_to_list0_field[s->mb_y&1][0];
1192             map_col_to_list0[1] = h->map_col_to_list0_field[s->mb_y&1][1];
1193             dist_scale_factor   =h->dist_scale_factor_field[s->mb_y&1];
1194         }
1195         if(h->ref_list[1][0].mbaff && IS_INTERLACED(mb_type_col[0]))
1196             ref_offset += 16;
1197
1198         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1199             /* FIXME assumes direct_8x8_inference == 1 */
1200             int y_shift  = 2*!IS_INTERLACED(*mb_type);
1201
1202             for(i8=0; i8<4; i8++){
1203                 const int x8 = i8&1;
1204                 const int y8 = i8>>1;
1205                 int ref0, scale;
1206                 const int16_t (*l1mv)[2]= l1mv0;
1207
1208                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1209                     continue;
1210                 h->sub_mb_type[i8] = sub_mb_type;
1211
1212                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1213                 if(IS_INTRA(mb_type_col[y8])){
1214                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1215                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1216                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1217                     continue;
1218                 }
1219
1220                 ref0 = l1ref0[x8 + y8*b8_stride];
1221                 if(ref0 >= 0)
1222                     ref0 = map_col_to_list0[0][ref0 + ref_offset];
1223                 else{
1224                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
1225                     l1mv= l1mv1;
1226                 }
1227                 scale = dist_scale_factor[ref0];
1228                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1229
1230                 {
1231                     const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride];
1232                     int my_col = (mv_col[1]<<y_shift)/2;
1233                     int mx = (scale * mv_col[0] + 128) >> 8;
1234                     int my = (scale * my_col + 128) >> 8;
1235                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1236                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1237                 }
1238             }
1239             return;
1240         }
1241
1242         /* one-to-one mv scaling */
1243
1244         if(IS_16X16(*mb_type)){
1245             int ref, mv0, mv1;
1246
1247             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1248             if(IS_INTRA(mb_type_col[0])){
1249                 ref=mv0=mv1=0;
1250             }else{
1251                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset]
1252                                                 : map_col_to_list0[1][l1ref1[0] + ref_offset];
1253                 const int scale = dist_scale_factor[ref0];
1254                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1255                 int mv_l0[2];
1256                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1257                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1258                 ref= ref0;
1259                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1260                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1261             }
1262             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1263             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1264             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1265         }else{
1266             for(i8=0; i8<4; i8++){
1267                 const int x8 = i8&1;
1268                 const int y8 = i8>>1;
1269                 int ref0, scale;
1270                 const int16_t (*l1mv)[2]= l1mv0;
1271
1272                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1273                     continue;
1274                 h->sub_mb_type[i8] = sub_mb_type;
1275                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1276                 if(IS_INTRA(mb_type_col[0])){
1277                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1278                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1279                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1280                     continue;
1281                 }
1282
1283                 ref0 = l1ref0[x8 + y8*b8_stride] + ref_offset;
1284                 if(ref0 >= 0)
1285                     ref0 = map_col_to_list0[0][ref0];
1286                 else{
1287                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
1288                     l1mv= l1mv1;
1289                 }
1290                 scale = dist_scale_factor[ref0];
1291
1292                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1293                 if(IS_SUB_8X8(sub_mb_type)){
1294                     const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1295                     int mx = (scale * mv_col[0] + 128) >> 8;
1296                     int my = (scale * mv_col[1] + 128) >> 8;
1297                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1298                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1299                 }else
1300                 for(i4=0; i4<4; i4++){
1301                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1302                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1303                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1304                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1305                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1306                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1307                 }
1308             }
1309         }
1310     }
1311 }
1312
1313 static inline void write_back_motion(H264Context *h, int mb_type){
1314     MpegEncContext * const s = &h->s;
1315     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1316     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1317     int list;
1318
1319     if(!USES_LIST(mb_type, 0))
1320         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1321
1322     for(list=0; list<h->list_count; list++){
1323         int y;
1324         if(!USES_LIST(mb_type, list))
1325             continue;
1326
1327         for(y=0; y<4; y++){
1328             *(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];
1329             *(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];
1330         }
1331         if( h->pps.cabac ) {
1332             if(IS_SKIP(mb_type))
1333                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1334             else
1335             for(y=0; y<4; y++){
1336                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1337                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1338             }
1339         }
1340
1341         {
1342             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1343             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1344             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1345             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1346             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1347         }
1348     }
1349
1350     if(h->slice_type_nos == FF_B_TYPE && h->pps.cabac){
1351         if(IS_8X8(mb_type)){
1352             uint8_t *direct_table = &h->direct_table[b8_xy];
1353             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1354             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1355             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1356         }
1357     }
1358 }
1359
1360 /**
1361  * Decodes a network abstraction layer unit.
1362  * @param consumed is the number of bytes used as input
1363  * @param length is the length of the array
1364  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1365  * @returns decoded bytes, might be src+1 if no escapes
1366  */
1367 static const uint8_t *decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
1368     int i, si, di;
1369     uint8_t *dst;
1370     int bufidx;
1371
1372 //    src[0]&0x80;                //forbidden bit
1373     h->nal_ref_idc= src[0]>>5;
1374     h->nal_unit_type= src[0]&0x1F;
1375
1376     src++; length--;
1377 #if 0
1378     for(i=0; i<length; i++)
1379         printf("%2X ", src[i]);
1380 #endif
1381
1382 #ifdef HAVE_FAST_UNALIGNED
1383 # ifdef HAVE_FAST_64BIT
1384 #   define RS 7
1385     for(i=0; i+1<length; i+=9){
1386         if(!((~*(uint64_t*)(src+i) & (*(uint64_t*)(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
1387 # else
1388 #   define RS 3
1389     for(i=0; i+1<length; i+=5){
1390         if(!((~*(uint32_t*)(src+i) & (*(uint32_t*)(src+i) - 0x01000101U)) & 0x80008080U))
1391 # endif
1392             continue;
1393         if(i>0 && !src[i]) i--;
1394         while(src[i]) i++;
1395 #else
1396 #   define RS 0
1397     for(i=0; i+1<length; i+=2){
1398         if(src[i]) continue;
1399         if(i>0 && src[i-1]==0) i--;
1400 #endif
1401         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1402             if(src[i+2]!=3){
1403                 /* startcode, so we must be past the end */
1404                 length=i;
1405             }
1406             break;
1407         }
1408         i-= RS;
1409     }
1410
1411     if(i>=length-1){ //no escaped 0
1412         *dst_length= length;
1413         *consumed= length+1; //+1 for the header
1414         return src;
1415     }
1416
1417     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1418     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
1419     dst= h->rbsp_buffer[bufidx];
1420
1421     if (dst == NULL){
1422         return NULL;
1423     }
1424
1425 //printf("decoding esc\n");
1426     memcpy(dst, src, i);
1427     si=di=i;
1428     while(si+2<length){
1429         //remove escapes (very rare 1:2^22)
1430         if(src[si+2]>3){
1431             dst[di++]= src[si++];
1432             dst[di++]= src[si++];
1433         }else if(src[si]==0 && src[si+1]==0){
1434             if(src[si+2]==3){ //escape
1435                 dst[di++]= 0;
1436                 dst[di++]= 0;
1437                 si+=3;
1438                 continue;
1439             }else //next start code
1440                 goto nsc;
1441         }
1442
1443         dst[di++]= src[si++];
1444     }
1445     while(si<length)
1446         dst[di++]= src[si++];
1447 nsc:
1448
1449     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
1450
1451     *dst_length= di;
1452     *consumed= si + 1;//+1 for the header
1453 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1454     return dst;
1455 }
1456
1457 /**
1458  * identifies the exact end of the bitstream
1459  * @return the length of the trailing, or 0 if damaged
1460  */
1461 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src){
1462     int v= *src;
1463     int r;
1464
1465     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1466
1467     for(r=1; r<9; r++){
1468         if(v&1) return r;
1469         v>>=1;
1470     }
1471     return 0;
1472 }
1473
1474 /**
1475  * IDCT transforms the 16 dc values and dequantizes them.
1476  * @param qp quantization parameter
1477  */
1478 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1479 #define stride 16
1480     int i;
1481     int temp[16]; //FIXME check if this is a good idea
1482     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1483     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1484
1485 //memset(block, 64, 2*256);
1486 //return;
1487     for(i=0; i<4; i++){
1488         const int offset= y_offset[i];
1489         const int z0= block[offset+stride*0] + block[offset+stride*4];
1490         const int z1= block[offset+stride*0] - block[offset+stride*4];
1491         const int z2= block[offset+stride*1] - block[offset+stride*5];
1492         const int z3= block[offset+stride*1] + block[offset+stride*5];
1493
1494         temp[4*i+0]= z0+z3;
1495         temp[4*i+1]= z1+z2;
1496         temp[4*i+2]= z1-z2;
1497         temp[4*i+3]= z0-z3;
1498     }
1499
1500     for(i=0; i<4; i++){
1501         const int offset= x_offset[i];
1502         const int z0= temp[4*0+i] + temp[4*2+i];
1503         const int z1= temp[4*0+i] - temp[4*2+i];
1504         const int z2= temp[4*1+i] - temp[4*3+i];
1505         const int z3= temp[4*1+i] + temp[4*3+i];
1506
1507         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
1508         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1509         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1510         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1511     }
1512 }
1513
1514 #if 0
1515 /**
1516  * DCT transforms the 16 dc values.
1517  * @param qp quantization parameter ??? FIXME
1518  */
1519 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1520 //    const int qmul= dequant_coeff[qp][0];
1521     int i;
1522     int temp[16]; //FIXME check if this is a good idea
1523     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1524     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1525
1526     for(i=0; i<4; i++){
1527         const int offset= y_offset[i];
1528         const int z0= block[offset+stride*0] + block[offset+stride*4];
1529         const int z1= block[offset+stride*0] - block[offset+stride*4];
1530         const int z2= block[offset+stride*1] - block[offset+stride*5];
1531         const int z3= block[offset+stride*1] + block[offset+stride*5];
1532
1533         temp[4*i+0]= z0+z3;
1534         temp[4*i+1]= z1+z2;
1535         temp[4*i+2]= z1-z2;
1536         temp[4*i+3]= z0-z3;
1537     }
1538
1539     for(i=0; i<4; i++){
1540         const int offset= x_offset[i];
1541         const int z0= temp[4*0+i] + temp[4*2+i];
1542         const int z1= temp[4*0+i] - temp[4*2+i];
1543         const int z2= temp[4*1+i] - temp[4*3+i];
1544         const int z3= temp[4*1+i] + temp[4*3+i];
1545
1546         block[stride*0 +offset]= (z0 + z3)>>1;
1547         block[stride*2 +offset]= (z1 + z2)>>1;
1548         block[stride*8 +offset]= (z1 - z2)>>1;
1549         block[stride*10+offset]= (z0 - z3)>>1;
1550     }
1551 }
1552 #endif
1553
1554 #undef xStride
1555 #undef stride
1556
1557 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1558     const int stride= 16*2;
1559     const int xStride= 16;
1560     int a,b,c,d,e;
1561
1562     a= block[stride*0 + xStride*0];
1563     b= block[stride*0 + xStride*1];
1564     c= block[stride*1 + xStride*0];
1565     d= block[stride*1 + xStride*1];
1566
1567     e= a-b;
1568     a= a+b;
1569     b= c-d;
1570     c= c+d;
1571
1572     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1573     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1574     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1575     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1576 }
1577
1578 #if 0
1579 static void chroma_dc_dct_c(DCTELEM *block){
1580     const int stride= 16*2;
1581     const int xStride= 16;
1582     int a,b,c,d,e;
1583
1584     a= block[stride*0 + xStride*0];
1585     b= block[stride*0 + xStride*1];
1586     c= block[stride*1 + xStride*0];
1587     d= block[stride*1 + xStride*1];
1588
1589     e= a-b;
1590     a= a+b;
1591     b= c-d;
1592     c= c+d;
1593
1594     block[stride*0 + xStride*0]= (a+c);
1595     block[stride*0 + xStride*1]= (e+b);
1596     block[stride*1 + xStride*0]= (a-c);
1597     block[stride*1 + xStride*1]= (e-b);
1598 }
1599 #endif
1600
1601 /**
1602  * gets the chroma qp.
1603  */
1604 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1605     return h->pps.chroma_qp_table[t][qscale];
1606 }
1607
1608 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1609                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1610                            int src_x_offset, int src_y_offset,
1611                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1612     MpegEncContext * const s = &h->s;
1613     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1614     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1615     const int luma_xy= (mx&3) + ((my&3)<<2);
1616     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1617     uint8_t * src_cb, * src_cr;
1618     int extra_width= h->emu_edge_width;
1619     int extra_height= h->emu_edge_height;
1620     int emu=0;
1621     const int full_mx= mx>>2;
1622     const int full_my= my>>2;
1623     const int pic_width  = 16*s->mb_width;
1624     const int pic_height = 16*s->mb_height >> MB_FIELD;
1625
1626     if(mx&7) extra_width -= 3;
1627     if(my&7) extra_height -= 3;
1628
1629     if(   full_mx < 0-extra_width
1630        || full_my < 0-extra_height
1631        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1632        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1633         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);
1634             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1635         emu=1;
1636     }
1637
1638     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1639     if(!square){
1640         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1641     }
1642
1643     if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1644
1645     if(MB_FIELD){
1646         // chroma offset when predicting from a field of opposite parity
1647         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1648         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1649     }
1650     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1651     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1652
1653     if(emu){
1654         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);
1655             src_cb= s->edge_emu_buffer;
1656     }
1657     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1658
1659     if(emu){
1660         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);
1661             src_cr= s->edge_emu_buffer;
1662     }
1663     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1664 }
1665
1666 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1667                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1668                            int x_offset, int y_offset,
1669                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1670                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1671                            int list0, int list1){
1672     MpegEncContext * const s = &h->s;
1673     qpel_mc_func *qpix_op=  qpix_put;
1674     h264_chroma_mc_func chroma_op= chroma_put;
1675
1676     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1677     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1678     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1679     x_offset += 8*s->mb_x;
1680     y_offset += 8*(s->mb_y >> MB_FIELD);
1681
1682     if(list0){
1683         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1684         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1685                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1686                            qpix_op, chroma_op);
1687
1688         qpix_op=  qpix_avg;
1689         chroma_op= chroma_avg;
1690     }
1691
1692     if(list1){
1693         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1694         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1695                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1696                            qpix_op, chroma_op);
1697     }
1698 }
1699
1700 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1701                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1702                            int x_offset, int y_offset,
1703                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1704                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1705                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1706                            int list0, int list1){
1707     MpegEncContext * const s = &h->s;
1708
1709     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1710     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1711     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1712     x_offset += 8*s->mb_x;
1713     y_offset += 8*(s->mb_y >> MB_FIELD);
1714
1715     if(list0 && list1){
1716         /* don't optimize for luma-only case, since B-frames usually
1717          * use implicit weights => chroma too. */
1718         uint8_t *tmp_cb = s->obmc_scratchpad;
1719         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1720         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1721         int refn0 = h->ref_cache[0][ scan8[n] ];
1722         int refn1 = h->ref_cache[1][ scan8[n] ];
1723
1724         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1725                     dest_y, dest_cb, dest_cr,
1726                     x_offset, y_offset, qpix_put, chroma_put);
1727         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1728                     tmp_y, tmp_cb, tmp_cr,
1729                     x_offset, y_offset, qpix_put, chroma_put);
1730
1731         if(h->use_weight == 2){
1732             int weight0 = h->implicit_weight[refn0][refn1];
1733             int weight1 = 64 - weight0;
1734             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1735             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1736             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1737         }else{
1738             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1739                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1740                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1741             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1742                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1743                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1744             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1745                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1746                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1747         }
1748     }else{
1749         int list = list1 ? 1 : 0;
1750         int refn = h->ref_cache[list][ scan8[n] ];
1751         Picture *ref= &h->ref_list[list][refn];
1752         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1753                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1754                     qpix_put, chroma_put);
1755
1756         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1757                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1758         if(h->use_weight_chroma){
1759             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1760                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1761             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1762                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1763         }
1764     }
1765 }
1766
1767 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1768                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1769                            int x_offset, int y_offset,
1770                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1771                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1772                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1773                            int list0, int list1){
1774     if((h->use_weight==2 && list0 && list1
1775         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1776        || h->use_weight==1)
1777         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1778                          x_offset, y_offset, qpix_put, chroma_put,
1779                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1780     else
1781         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1782                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1783 }
1784
1785 static inline void prefetch_motion(H264Context *h, int list){
1786     /* fetch pixels for estimated mv 4 macroblocks ahead
1787      * optimized for 64byte cache lines */
1788     MpegEncContext * const s = &h->s;
1789     const int refn = h->ref_cache[list][scan8[0]];
1790     if(refn >= 0){
1791         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1792         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1793         uint8_t **src= h->ref_list[list][refn].data;
1794         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1795         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1796         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1797         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1798     }
1799 }
1800
1801 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1802                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1803                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1804                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1805     MpegEncContext * const s = &h->s;
1806     const int mb_xy= h->mb_xy;
1807     const int mb_type= s->current_picture.mb_type[mb_xy];
1808
1809     assert(IS_INTER(mb_type));
1810
1811     prefetch_motion(h, 0);
1812
1813     if(IS_16X16(mb_type)){
1814         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1815                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1816                 &weight_op[0], &weight_avg[0],
1817                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1818     }else if(IS_16X8(mb_type)){
1819         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1820                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1821                 &weight_op[1], &weight_avg[1],
1822                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1823         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1824                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1825                 &weight_op[1], &weight_avg[1],
1826                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1827     }else if(IS_8X16(mb_type)){
1828         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1829                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1830                 &weight_op[2], &weight_avg[2],
1831                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1832         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1833                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1834                 &weight_op[2], &weight_avg[2],
1835                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1836     }else{
1837         int i;
1838
1839         assert(IS_8X8(mb_type));
1840
1841         for(i=0; i<4; i++){
1842             const int sub_mb_type= h->sub_mb_type[i];
1843             const int n= 4*i;
1844             int x_offset= (i&1)<<2;
1845             int y_offset= (i&2)<<1;
1846
1847             if(IS_SUB_8X8(sub_mb_type)){
1848                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1849                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1850                     &weight_op[3], &weight_avg[3],
1851                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1852             }else if(IS_SUB_8X4(sub_mb_type)){
1853                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1854                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1855                     &weight_op[4], &weight_avg[4],
1856                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1857                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1858                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1859                     &weight_op[4], &weight_avg[4],
1860                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1861             }else if(IS_SUB_4X8(sub_mb_type)){
1862                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1863                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1864                     &weight_op[5], &weight_avg[5],
1865                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1866                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1867                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1868                     &weight_op[5], &weight_avg[5],
1869                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1870             }else{
1871                 int j;
1872                 assert(IS_SUB_4X4(sub_mb_type));
1873                 for(j=0; j<4; j++){
1874                     int sub_x_offset= x_offset + 2*(j&1);
1875                     int sub_y_offset= y_offset +   (j&2);
1876                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1877                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1878                         &weight_op[6], &weight_avg[6],
1879                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1880                 }
1881             }
1882         }
1883     }
1884
1885     prefetch_motion(h, 1);
1886 }
1887
1888 static av_cold void init_cavlc_level_tab(void){
1889     int suffix_length, mask;
1890     unsigned int i;
1891
1892     for(suffix_length=0; suffix_length<7; suffix_length++){
1893         for(i=0; i<(1<<LEVEL_TAB_BITS); i++){
1894             int prefix= LEVEL_TAB_BITS - av_log2(2*i);
1895             int level_code= (prefix<<suffix_length) + (i>>(LEVEL_TAB_BITS-prefix-1-suffix_length)) - (1<<suffix_length);
1896
1897             mask= -(level_code&1);
1898             level_code= (((2+level_code)>>1) ^ mask) - mask;
1899             if(prefix + 1 + suffix_length <= LEVEL_TAB_BITS){
1900                 cavlc_level_tab[suffix_length][i][0]= level_code;
1901                 cavlc_level_tab[suffix_length][i][1]= prefix + 1 + suffix_length;
1902             }else if(prefix + 1 <= LEVEL_TAB_BITS){
1903                 cavlc_level_tab[suffix_length][i][0]= prefix+100;
1904                 cavlc_level_tab[suffix_length][i][1]= prefix + 1;
1905             }else{
1906                 cavlc_level_tab[suffix_length][i][0]= LEVEL_TAB_BITS+100;
1907                 cavlc_level_tab[suffix_length][i][1]= LEVEL_TAB_BITS;
1908             }
1909         }
1910     }
1911 }
1912
1913 static av_cold void decode_init_vlc(void){
1914     static int done = 0;
1915
1916     if (!done) {
1917         int i;
1918         int offset;
1919         done = 1;
1920
1921         chroma_dc_coeff_token_vlc.table = chroma_dc_coeff_token_vlc_table;
1922         chroma_dc_coeff_token_vlc.table_allocated = chroma_dc_coeff_token_vlc_table_size;
1923         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1924                  &chroma_dc_coeff_token_len [0], 1, 1,
1925                  &chroma_dc_coeff_token_bits[0], 1, 1,
1926                  INIT_VLC_USE_NEW_STATIC);
1927
1928         offset = 0;
1929         for(i=0; i<4; i++){
1930             coeff_token_vlc[i].table = coeff_token_vlc_tables+offset;
1931             coeff_token_vlc[i].table_allocated = coeff_token_vlc_tables_size[i];
1932             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1933                      &coeff_token_len [i][0], 1, 1,
1934                      &coeff_token_bits[i][0], 1, 1,
1935                      INIT_VLC_USE_NEW_STATIC);
1936             offset += coeff_token_vlc_tables_size[i];
1937         }
1938         /*
1939          * This is a one time safety check to make sure that
1940          * the packed static coeff_token_vlc table sizes
1941          * were initialized correctly.
1942          */
1943         assert(offset == FF_ARRAY_ELEMS(coeff_token_vlc_tables));
1944
1945         for(i=0; i<3; i++){
1946             chroma_dc_total_zeros_vlc[i].table = chroma_dc_total_zeros_vlc_tables[i];
1947             chroma_dc_total_zeros_vlc[i].table_allocated = chroma_dc_total_zeros_vlc_tables_size;
1948             init_vlc(&chroma_dc_total_zeros_vlc[i],
1949                      CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1950                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1951                      &chroma_dc_total_zeros_bits[i][0], 1, 1,
1952                      INIT_VLC_USE_NEW_STATIC);
1953         }
1954         for(i=0; i<15; i++){
1955             total_zeros_vlc[i].table = total_zeros_vlc_tables[i];
1956             total_zeros_vlc[i].table_allocated = total_zeros_vlc_tables_size;
1957             init_vlc(&total_zeros_vlc[i],
1958                      TOTAL_ZEROS_VLC_BITS, 16,
1959                      &total_zeros_len [i][0], 1, 1,
1960                      &total_zeros_bits[i][0], 1, 1,
1961                      INIT_VLC_USE_NEW_STATIC);
1962         }
1963
1964         for(i=0; i<6; i++){
1965             run_vlc[i].table = run_vlc_tables[i];
1966             run_vlc[i].table_allocated = run_vlc_tables_size;
1967             init_vlc(&run_vlc[i],
1968                      RUN_VLC_BITS, 7,
1969                      &run_len [i][0], 1, 1,
1970                      &run_bits[i][0], 1, 1,
1971                      INIT_VLC_USE_NEW_STATIC);
1972         }
1973         run7_vlc.table = run7_vlc_table,
1974         run7_vlc.table_allocated = run7_vlc_table_size;
1975         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
1976                  &run_len [6][0], 1, 1,
1977                  &run_bits[6][0], 1, 1,
1978                  INIT_VLC_USE_NEW_STATIC);
1979
1980         init_cavlc_level_tab();
1981     }
1982 }
1983
1984 static void free_tables(H264Context *h){
1985     int i;
1986     H264Context *hx;
1987     av_freep(&h->intra4x4_pred_mode);
1988     av_freep(&h->chroma_pred_mode_table);
1989     av_freep(&h->cbp_table);
1990     av_freep(&h->mvd_table[0]);
1991     av_freep(&h->mvd_table[1]);
1992     av_freep(&h->direct_table);
1993     av_freep(&h->non_zero_count);
1994     av_freep(&h->slice_table_base);
1995     h->slice_table= NULL;
1996
1997     av_freep(&h->mb2b_xy);
1998     av_freep(&h->mb2b8_xy);
1999
2000     for(i = 0; i < h->s.avctx->thread_count; i++) {
2001         hx = h->thread_context[i];
2002         if(!hx) continue;
2003         av_freep(&hx->top_borders[1]);
2004         av_freep(&hx->top_borders[0]);
2005         av_freep(&hx->s.obmc_scratchpad);
2006     }
2007 }
2008
2009 static void init_dequant8_coeff_table(H264Context *h){
2010     int i,q,x;
2011     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2012     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2013     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2014
2015     for(i=0; i<2; i++ ){
2016         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2017             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2018             break;
2019         }
2020
2021         for(q=0; q<52; q++){
2022             int shift = div6[q];
2023             int idx = rem6[q];
2024             for(x=0; x<64; x++)
2025                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2026                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2027                     h->pps.scaling_matrix8[i][x]) << shift;
2028         }
2029     }
2030 }
2031
2032 static void init_dequant4_coeff_table(H264Context *h){
2033     int i,j,q,x;
2034     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2035     for(i=0; i<6; i++ ){
2036         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2037         for(j=0; j<i; j++){
2038             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2039                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2040                 break;
2041             }
2042         }
2043         if(j<i)
2044             continue;
2045
2046         for(q=0; q<52; q++){
2047             int shift = div6[q] + 2;
2048             int idx = rem6[q];
2049             for(x=0; x<16; x++)
2050                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2051                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2052                     h->pps.scaling_matrix4[i][x]) << shift;
2053         }
2054     }
2055 }
2056
2057 static void init_dequant_tables(H264Context *h){
2058     int i,x;
2059     init_dequant4_coeff_table(h);
2060     if(h->pps.transform_8x8_mode)
2061         init_dequant8_coeff_table(h);
2062     if(h->sps.transform_bypass){
2063         for(i=0; i<6; i++)
2064             for(x=0; x<16; x++)
2065                 h->dequant4_coeff[i][0][x] = 1<<6;
2066         if(h->pps.transform_8x8_mode)
2067             for(i=0; i<2; i++)
2068                 for(x=0; x<64; x++)
2069                     h->dequant8_coeff[i][0][x] = 1<<6;
2070     }
2071 }
2072
2073
2074 /**
2075  * allocates tables.
2076  * needs width/height
2077  */
2078 static int alloc_tables(H264Context *h){
2079     MpegEncContext * const s = &h->s;
2080     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2081     int x,y;
2082
2083     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2084
2085     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2086     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base))
2087     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2088
2089     CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2090     CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2091     CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2092     CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2093
2094     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
2095     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2096
2097     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2098     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2099     for(y=0; y<s->mb_height; y++){
2100         for(x=0; x<s->mb_width; x++){
2101             const int mb_xy= x + y*s->mb_stride;
2102             const int b_xy = 4*x + 4*y*h->b_stride;
2103             const int b8_xy= 2*x + 2*y*h->b8_stride;
2104
2105             h->mb2b_xy [mb_xy]= b_xy;
2106             h->mb2b8_xy[mb_xy]= b8_xy;
2107         }
2108     }
2109
2110     s->obmc_scratchpad = NULL;
2111
2112     if(!h->dequant4_coeff[0])
2113         init_dequant_tables(h);
2114
2115     return 0;
2116 fail:
2117     free_tables(h);
2118     return -1;
2119 }
2120
2121 /**
2122  * Mimic alloc_tables(), but for every context thread.
2123  */
2124 static void clone_tables(H264Context *dst, H264Context *src){
2125     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2126     dst->non_zero_count           = src->non_zero_count;
2127     dst->slice_table              = src->slice_table;
2128     dst->cbp_table                = src->cbp_table;
2129     dst->mb2b_xy                  = src->mb2b_xy;
2130     dst->mb2b8_xy                 = src->mb2b8_xy;
2131     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2132     dst->mvd_table[0]             = src->mvd_table[0];
2133     dst->mvd_table[1]             = src->mvd_table[1];
2134     dst->direct_table             = src->direct_table;
2135
2136     dst->s.obmc_scratchpad = NULL;
2137     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2138 }
2139
2140 /**
2141  * Init context
2142  * Allocate buffers which are not shared amongst multiple threads.
2143  */
2144 static int context_init(H264Context *h){
2145     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2146     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2147
2148     return 0;
2149 fail:
2150     return -1; // free_tables will clean up for us
2151 }
2152
2153 static av_cold void common_init(H264Context *h){
2154     MpegEncContext * const s = &h->s;
2155
2156     s->width = s->avctx->width;
2157     s->height = s->avctx->height;
2158     s->codec_id= s->avctx->codec->id;
2159
2160     ff_h264_pred_init(&h->hpc, s->codec_id);
2161
2162     h->dequant_coeff_pps= -1;
2163     s->unrestricted_mv=1;
2164     s->decode=1; //FIXME
2165
2166     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
2167
2168     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2169     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2170 }
2171
2172 static av_cold int decode_init(AVCodecContext *avctx){
2173     H264Context *h= avctx->priv_data;
2174     MpegEncContext * const s = &h->s;
2175
2176     MPV_decode_defaults(s);
2177
2178     s->avctx = avctx;
2179     common_init(h);
2180
2181     s->out_format = FMT_H264;
2182     s->workaround_bugs= avctx->workaround_bugs;
2183
2184     // set defaults
2185 //    s->decode_mb= ff_h263_decode_mb;
2186     s->quarter_sample = 1;
2187     s->low_delay= 1;
2188
2189     if(avctx->codec_id == CODEC_ID_SVQ3)
2190         avctx->pix_fmt= PIX_FMT_YUVJ420P;
2191     else
2192         avctx->pix_fmt= PIX_FMT_YUV420P;
2193
2194     decode_init_vlc();
2195
2196     if(avctx->extradata_size > 0 && avctx->extradata &&
2197        *(char *)avctx->extradata == 1){
2198         h->is_avc = 1;
2199         h->got_avcC = 0;
2200     } else {
2201         h->is_avc = 0;
2202     }
2203
2204     h->thread_context[0] = h;
2205     h->outputed_poc = INT_MIN;
2206     h->prev_poc_msb= 1<<16;
2207     return 0;
2208 }
2209
2210 static int frame_start(H264Context *h){
2211     MpegEncContext * const s = &h->s;
2212     int i;
2213
2214     if(MPV_frame_start(s, s->avctx) < 0)
2215         return -1;
2216     ff_er_frame_start(s);
2217     /*
2218      * MPV_frame_start uses pict_type to derive key_frame.
2219      * This is incorrect for H.264; IDR markings must be used.
2220      * Zero here; IDR markings per slice in frame or fields are ORed in later.
2221      * See decode_nal_units().
2222      */
2223     s->current_picture_ptr->key_frame= 0;
2224
2225     assert(s->linesize && s->uvlinesize);
2226
2227     for(i=0; i<16; i++){
2228         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2229         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2230     }
2231     for(i=0; i<4; i++){
2232         h->block_offset[16+i]=
2233         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2234         h->block_offset[24+16+i]=
2235         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2236     }
2237
2238     /* can't be in alloc_tables because linesize isn't known there.
2239      * FIXME: redo bipred weight to not require extra buffer? */
2240     for(i = 0; i < s->avctx->thread_count; i++)
2241         if(!h->thread_context[i]->s.obmc_scratchpad)
2242             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2243
2244     /* some macroblocks will be accessed before they're available */
2245     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2246         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
2247
2248 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2249
2250     // We mark the current picture as non-reference after allocating it, so
2251     // that if we break out due to an error it can be released automatically
2252     // in the next MPV_frame_start().
2253     // SVQ3 as well as most other codecs have only last/next/current and thus
2254     // get released even with set reference, besides SVQ3 and others do not
2255     // mark frames as reference later "naturally".
2256     if(s->codec_id != CODEC_ID_SVQ3)
2257         s->current_picture_ptr->reference= 0;
2258
2259     s->current_picture_ptr->field_poc[0]=
2260     s->current_picture_ptr->field_poc[1]= INT_MAX;
2261     assert(s->current_picture_ptr->long_ref==0);
2262
2263     return 0;
2264 }
2265
2266 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){
2267     MpegEncContext * const s = &h->s;
2268     int i;
2269     int step    = 1;
2270     int offset  = 1;
2271     int uvoffset= 1;
2272     int top_idx = 1;
2273     int skiplast= 0;
2274
2275     src_y  -=   linesize;
2276     src_cb -= uvlinesize;
2277     src_cr -= uvlinesize;
2278
2279     if(!simple && FRAME_MBAFF){
2280         if(s->mb_y&1){
2281             offset  = MB_MBAFF ? 1 : 17;
2282             uvoffset= MB_MBAFF ? 1 : 9;
2283             if(!MB_MBAFF){
2284                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 0)= *(uint64_t*)(src_y +  15*linesize);
2285                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 8)= *(uint64_t*)(src_y +8+15*linesize);
2286                 if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2287                     *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+7*uvlinesize);
2288                     *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+7*uvlinesize);
2289                 }
2290             }
2291         }else{
2292             if(!MB_MBAFF){
2293                 h->left_border[0]= h->top_borders[0][s->mb_x][15];
2294                 if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2295                     h->left_border[34   ]= h->top_borders[0][s->mb_x][16+7  ];
2296                     h->left_border[34+18]= h->top_borders[0][s->mb_x][16+8+7];
2297                 }
2298                 skiplast= 1;
2299             }
2300             offset  =
2301             uvoffset=
2302             top_idx = MB_MBAFF ? 0 : 1;
2303         }
2304         step= MB_MBAFF ? 2 : 1;
2305     }
2306
2307     // There are two lines saved, the line above the the top macroblock of a pair,
2308     // and the line above the bottom macroblock
2309     h->left_border[offset]= h->top_borders[top_idx][s->mb_x][15];
2310     for(i=1; i<17 - skiplast; i++){
2311         h->left_border[offset+i*step]= src_y[15+i*  linesize];
2312     }
2313
2314     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2315     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2316
2317     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2318         h->left_border[uvoffset+34   ]= h->top_borders[top_idx][s->mb_x][16+7];
2319         h->left_border[uvoffset+34+18]= h->top_borders[top_idx][s->mb_x][24+7];
2320         for(i=1; i<9 - skiplast; i++){
2321             h->left_border[uvoffset+34   +i*step]= src_cb[7+i*uvlinesize];
2322             h->left_border[uvoffset+34+18+i*step]= src_cr[7+i*uvlinesize];
2323         }
2324         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2325         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2326     }
2327 }
2328
2329 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){
2330     MpegEncContext * const s = &h->s;
2331     int temp8, i;
2332     uint64_t temp64;
2333     int deblock_left;
2334     int deblock_top;
2335     int mb_xy;
2336     int step    = 1;
2337     int offset  = 1;
2338     int uvoffset= 1;
2339     int top_idx = 1;
2340
2341     if(!simple && FRAME_MBAFF){
2342         if(s->mb_y&1){
2343             offset  = MB_MBAFF ? 1 : 17;
2344             uvoffset= MB_MBAFF ? 1 : 9;
2345         }else{
2346             offset  =
2347             uvoffset=
2348             top_idx = MB_MBAFF ? 0 : 1;
2349         }
2350         step= MB_MBAFF ? 2 : 1;
2351     }
2352
2353     if(h->deblocking_filter == 2) {
2354         mb_xy = h->mb_xy;
2355         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2356         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2357     } else {
2358         deblock_left = (s->mb_x > 0);
2359         deblock_top =  (s->mb_y > !!MB_FIELD);
2360     }
2361
2362     src_y  -=   linesize + 1;
2363     src_cb -= uvlinesize + 1;
2364     src_cr -= uvlinesize + 1;
2365
2366 #define XCHG(a,b,t,xchg)\
2367 t= a;\
2368 if(xchg)\
2369     a= b;\
2370 b= t;
2371
2372     if(deblock_left){
2373         for(i = !deblock_top; i<16; i++){
2374             XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, xchg);
2375         }
2376         XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, 1);
2377     }
2378
2379     if(deblock_top){
2380         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2381         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2382         if(s->mb_x+1 < s->mb_width){
2383             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2384         }
2385     }
2386
2387     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2388         if(deblock_left){
2389             for(i = !deblock_top; i<8; i++){
2390                 XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, xchg);
2391                 XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, xchg);
2392             }
2393             XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, 1);
2394             XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, 1);
2395         }
2396         if(deblock_top){
2397             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2398             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2399         }
2400     }
2401 }
2402
2403 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2404     MpegEncContext * const s = &h->s;
2405     const int mb_x= s->mb_x;
2406     const int mb_y= s->mb_y;
2407     const int mb_xy= h->mb_xy;
2408     const int mb_type= s->current_picture.mb_type[mb_xy];
2409     uint8_t  *dest_y, *dest_cb, *dest_cr;
2410     int linesize, uvlinesize /*dct_offset*/;
2411     int i;
2412     int *block_offset = &h->block_offset[0];
2413     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2414     const int is_h264 = simple || s->codec_id == CODEC_ID_H264;
2415     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2416     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2417
2418     dest_y  = s->current_picture.data[0] + (mb_x + mb_y * s->linesize  ) * 16;
2419     dest_cb = s->current_picture.data[1] + (mb_x + mb_y * s->uvlinesize) * 8;
2420     dest_cr = s->current_picture.data[2] + (mb_x + mb_y * s->uvlinesize) * 8;
2421
2422     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2423     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2424
2425     if (!simple && MB_FIELD) {
2426         linesize   = h->mb_linesize   = s->linesize * 2;
2427         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2428         block_offset = &h->block_offset[24];
2429         if(mb_y&1){ //FIXME move out of this function?
2430             dest_y -= s->linesize*15;
2431             dest_cb-= s->uvlinesize*7;
2432             dest_cr-= s->uvlinesize*7;
2433         }
2434         if(FRAME_MBAFF) {
2435             int list;
2436             for(list=0; list<h->list_count; list++){
2437                 if(!USES_LIST(mb_type, list))
2438                     continue;
2439                 if(IS_16X16(mb_type)){
2440                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2441                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2442                 }else{
2443                     for(i=0; i<16; i+=4){
2444                         int ref = h->ref_cache[list][scan8[i]];
2445                         if(ref >= 0)
2446                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2447                     }
2448                 }
2449             }
2450         }
2451     } else {
2452         linesize   = h->mb_linesize   = s->linesize;
2453         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2454 //        dct_offset = s->linesize * 16;
2455     }
2456
2457     if (!simple && IS_INTRA_PCM(mb_type)) {
2458         for (i=0; i<16; i++) {
2459             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
2460         }
2461         for (i=0; i<8; i++) {
2462             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
2463             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
2464         }
2465     } else {
2466         if(IS_INTRA(mb_type)){
2467             if(h->deblocking_filter)
2468                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2469
2470             if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2471                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2472                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2473             }
2474
2475             if(IS_INTRA4x4(mb_type)){
2476                 if(simple || !s->encoding){
2477                     if(IS_8x8DCT(mb_type)){
2478                         if(transform_bypass){
2479                             idct_dc_add =
2480                             idct_add    = s->dsp.add_pixels8;
2481                         }else{
2482                             idct_dc_add = s->dsp.h264_idct8_dc_add;
2483                             idct_add    = s->dsp.h264_idct8_add;
2484                         }
2485                         for(i=0; i<16; i+=4){
2486                             uint8_t * const ptr= dest_y + block_offset[i];
2487                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2488                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
2489                                 h->hpc.pred8x8l_add[dir](ptr, h->mb + i*16, linesize);
2490                             }else{
2491                                 const int nnz = h->non_zero_count_cache[ scan8[i] ];
2492                                 h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2493                                                             (h->topright_samples_available<<i)&0x4000, linesize);
2494                                 if(nnz){
2495                                     if(nnz == 1 && h->mb[i*16])
2496                                         idct_dc_add(ptr, h->mb + i*16, linesize);
2497                                     else
2498                                         idct_add   (ptr, h->mb + i*16, linesize);
2499                                 }
2500                             }
2501                         }
2502                     }else{
2503                         if(transform_bypass){
2504                             idct_dc_add =
2505                             idct_add    = s->dsp.add_pixels4;
2506                         }else{
2507                             idct_dc_add = s->dsp.h264_idct_dc_add;
2508                             idct_add    = s->dsp.h264_idct_add;
2509                         }
2510                         for(i=0; i<16; i++){
2511                             uint8_t * const ptr= dest_y + block_offset[i];
2512                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2513
2514                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
2515                                 h->hpc.pred4x4_add[dir](ptr, h->mb + i*16, linesize);
2516                             }else{
2517                                 uint8_t *topright;
2518                                 int nnz, tr;
2519                                 if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2520                                     const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2521                                     assert(mb_y || linesize <= block_offset[i]);
2522                                     if(!topright_avail){
2523                                         tr= ptr[3 - linesize]*0x01010101;
2524                                         topright= (uint8_t*) &tr;
2525                                     }else
2526                                         topright= ptr + 4 - linesize;
2527                                 }else
2528                                     topright= NULL;
2529
2530                                 h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2531                                 nnz = h->non_zero_count_cache[ scan8[i] ];
2532                                 if(nnz){
2533                                     if(is_h264){
2534                                         if(nnz == 1 && h->mb[i*16])
2535                                             idct_dc_add(ptr, h->mb + i*16, linesize);
2536                                         else
2537                                             idct_add   (ptr, h->mb + i*16, linesize);
2538                                     }else
2539                                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2540                                 }
2541                             }
2542                         }
2543                     }
2544                 }
2545             }else{
2546                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2547                 if(is_h264){
2548                     if(!transform_bypass)
2549                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2550                 }else
2551                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2552             }
2553             if(h->deblocking_filter)
2554                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2555         }else if(is_h264){
2556             hl_motion(h, dest_y, dest_cb, dest_cr,
2557                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2558                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2559                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2560         }
2561
2562
2563         if(!IS_INTRA4x4(mb_type)){
2564             if(is_h264){
2565                 if(IS_INTRA16x16(mb_type)){
2566                     if(transform_bypass){
2567                         if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
2568                             h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb, linesize);
2569                         }else{
2570                             for(i=0; i<16; i++){
2571                                 if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
2572                                     s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + i*16, linesize);
2573                             }
2574                         }
2575                     }else{
2576                          s->dsp.h264_idct_add16intra(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2577                     }
2578                 }else if(h->cbp&15){
2579                     if(transform_bypass){
2580                         const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2581                         idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2582                         for(i=0; i<16; i+=di){
2583                             if(h->non_zero_count_cache[ scan8[i] ]){
2584                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2585                             }
2586                         }
2587                     }else{
2588                         if(IS_8x8DCT(mb_type)){
2589                             s->dsp.h264_idct8_add4(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2590                         }else{
2591                             s->dsp.h264_idct_add16(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2592                         }
2593                     }
2594                 }
2595             }else{
2596                 for(i=0; i<16; i++){
2597                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2598                         uint8_t * const ptr= dest_y + block_offset[i];
2599                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2600                     }
2601                 }
2602             }
2603         }
2604
2605         if((simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
2606             uint8_t *dest[2] = {dest_cb, dest_cr};
2607             if(transform_bypass){
2608                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
2609                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + 16*16, uvlinesize);
2610                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 20, h->mb + 20*16, uvlinesize);
2611                 }else{
2612                     idct_add = s->dsp.add_pixels4;
2613                     for(i=16; i<16+8; i++){
2614                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
2615                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2616                     }
2617                 }
2618             }else{
2619                 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]);
2620                 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]);
2621                 if(is_h264){
2622                     idct_add = s->dsp.h264_idct_add;
2623                     idct_dc_add = s->dsp.h264_idct_dc_add;
2624                     for(i=16; i<16+8; i++){
2625                         if(h->non_zero_count_cache[ scan8[i] ])
2626                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2627                         else if(h->mb[i*16])
2628                             idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2629                     }
2630                 }else{
2631                     for(i=16; i<16+8; i++){
2632                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2633                             uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2634                             svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2635                         }
2636                     }
2637                 }
2638             }
2639         }
2640     }
2641     if(h->cbp || IS_INTRA(mb_type))
2642         s->dsp.clear_blocks(h->mb);
2643
2644     if(h->deblocking_filter) {
2645         backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2646         fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2647         h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2648         h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2649         if (!simple && FRAME_MBAFF) {
2650             filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2651         } else {
2652             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2653         }
2654     }
2655 }
2656
2657 /**
2658  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2659  */
2660 static void hl_decode_mb_simple(H264Context *h){
2661     hl_decode_mb_internal(h, 1);
2662 }
2663
2664 /**
2665  * Process a macroblock; this handles edge cases, such as interlacing.
2666  */
2667 static void av_noinline hl_decode_mb_complex(H264Context *h){
2668     hl_decode_mb_internal(h, 0);
2669 }
2670
2671 static void hl_decode_mb(H264Context *h){
2672     MpegEncContext * const s = &h->s;
2673     const int mb_xy= h->mb_xy;
2674     const int mb_type= s->current_picture.mb_type[mb_xy];
2675     int is_complex = ENABLE_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2676
2677     if(ENABLE_H264_ENCODER && !s->decode)
2678         return;
2679
2680     if (is_complex)
2681         hl_decode_mb_complex(h);
2682     else hl_decode_mb_simple(h);
2683 }
2684
2685 static void pic_as_field(Picture *pic, const int parity){
2686     int i;
2687     for (i = 0; i < 4; ++i) {
2688         if (parity == PICT_BOTTOM_FIELD)
2689             pic->data[i] += pic->linesize[i];
2690         pic->reference = parity;
2691         pic->linesize[i] *= 2;
2692     }
2693     pic->poc= pic->field_poc[parity == PICT_BOTTOM_FIELD];
2694 }
2695
2696 static int split_field_copy(Picture *dest, Picture *src,
2697                             int parity, int id_add){
2698     int match = !!(src->reference & parity);
2699
2700     if (match) {
2701         *dest = *src;
2702         if(parity != PICT_FRAME){
2703             pic_as_field(dest, parity);
2704             dest->pic_id *= 2;
2705             dest->pic_id += id_add;
2706         }
2707     }
2708
2709     return match;
2710 }
2711
2712 static int build_def_list(Picture *def, Picture **in, int len, int is_long, int sel){
2713     int i[2]={0};
2714     int index=0;
2715
2716     while(i[0]<len || i[1]<len){
2717         while(i[0]<len && !(in[ i[0] ] && (in[ i[0] ]->reference & sel)))
2718             i[0]++;
2719         while(i[1]<len && !(in[ i[1] ] && (in[ i[1] ]->reference & (sel^3))))
2720             i[1]++;
2721         if(i[0] < len){
2722             in[ i[0] ]->pic_id= is_long ? i[0] : in[ i[0] ]->frame_num;
2723             split_field_copy(&def[index++], in[ i[0]++ ], sel  , 1);
2724         }
2725         if(i[1] < len){
2726             in[ i[1] ]->pic_id= is_long ? i[1] : in[ i[1] ]->frame_num;
2727             split_field_copy(&def[index++], in[ i[1]++ ], sel^3, 0);
2728         }
2729     }
2730
2731     return index;
2732 }
2733
2734 static int add_sorted(Picture **sorted, Picture **src, int len, int limit, int dir){
2735     int i, best_poc;
2736     int out_i= 0;
2737
2738     for(;;){
2739         best_poc= dir ? INT_MIN : INT_MAX;
2740
2741         for(i=0; i<len; i++){
2742             const int poc= src[i]->poc;
2743             if(((poc > limit) ^ dir) && ((poc < best_poc) ^ dir)){
2744                 best_poc= poc;
2745                 sorted[out_i]= src[i];
2746             }
2747         }
2748         if(best_poc == (dir ? INT_MIN : INT_MAX))
2749             break;
2750         limit= sorted[out_i++]->poc - dir;
2751     }
2752     return out_i;
2753 }
2754
2755 /**
2756  * fills the default_ref_list.
2757  */
2758 static int fill_default_ref_list(H264Context *h){
2759     MpegEncContext * const s = &h->s;
2760     int i, len;
2761
2762     if(h->slice_type_nos==FF_B_TYPE){
2763         Picture *sorted[32];
2764         int cur_poc, list;
2765         int lens[2];
2766
2767         if(FIELD_PICTURE)
2768             cur_poc= s->current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
2769         else
2770             cur_poc= s->current_picture_ptr->poc;
2771
2772         for(list= 0; list<2; list++){
2773             len= add_sorted(sorted    , h->short_ref, h->short_ref_count, cur_poc, 1^list);
2774             len+=add_sorted(sorted+len, h->short_ref, h->short_ref_count, cur_poc, 0^list);
2775             assert(len<=32);
2776             len= build_def_list(h->default_ref_list[list]    , sorted     , len, 0, s->picture_structure);
2777             len+=build_def_list(h->default_ref_list[list]+len, h->long_ref, 16 , 1, s->picture_structure);
2778             assert(len<=32);
2779
2780             if(len < h->ref_count[list])
2781                 memset(&h->default_ref_list[list][len], 0, sizeof(Picture)*(h->ref_count[list] - len));
2782             lens[list]= len;
2783         }
2784
2785         if(lens[0] == lens[1] && lens[1] > 1){
2786             for(i=0; h->default_ref_list[0][i].data[0] == h->default_ref_list[1][i].data[0] && i<lens[0]; i++);
2787             if(i == lens[0])
2788                 FFSWAP(Picture, h->default_ref_list[1][0], h->default_ref_list[1][1]);
2789         }
2790     }else{
2791         len = build_def_list(h->default_ref_list[0]    , h->short_ref, h->short_ref_count, 0, s->picture_structure);
2792         len+= build_def_list(h->default_ref_list[0]+len, h-> long_ref, 16                , 1, s->picture_structure);
2793         assert(len <= 32);
2794         if(len < h->ref_count[0])
2795             memset(&h->default_ref_list[0][len], 0, sizeof(Picture)*(h->ref_count[0] - len));
2796     }
2797 #ifdef TRACE
2798     for (i=0; i<h->ref_count[0]; i++) {
2799         tprintf(h->s.avctx, "List0: %s fn:%d 0x%p\n", (h->default_ref_list[0][i].long_ref ? "LT" : "ST"), h->default_ref_list[0][i].pic_id, h->default_ref_list[0][i].data[0]);
2800     }
2801     if(h->slice_type_nos==FF_B_TYPE){
2802         for (i=0; i<h->ref_count[1]; i++) {
2803             tprintf(h->s.avctx, "List1: %s fn:%d 0x%p\n", (h->default_ref_list[1][i].long_ref ? "LT" : "ST"), h->default_ref_list[1][i].pic_id, h->default_ref_list[1][i].data[0]);
2804         }
2805     }
2806 #endif
2807     return 0;
2808 }
2809
2810 static void print_short_term(H264Context *h);
2811 static void print_long_term(H264Context *h);
2812
2813 /**
2814  * Extract structure information about the picture described by pic_num in
2815  * the current decoding context (frame or field). Note that pic_num is
2816  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
2817  * @param pic_num picture number for which to extract structure information
2818  * @param structure one of PICT_XXX describing structure of picture
2819  *                      with pic_num
2820  * @return frame number (short term) or long term index of picture
2821  *         described by pic_num
2822  */
2823 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
2824     MpegEncContext * const s = &h->s;
2825
2826     *structure = s->picture_structure;
2827     if(FIELD_PICTURE){
2828         if (!(pic_num & 1))
2829             /* opposite field */
2830             *structure ^= PICT_FRAME;
2831         pic_num >>= 1;
2832     }
2833
2834     return pic_num;
2835 }
2836
2837 static int decode_ref_pic_list_reordering(H264Context *h){
2838     MpegEncContext * const s = &h->s;
2839     int list, index, pic_structure;
2840
2841     print_short_term(h);
2842     print_long_term(h);
2843
2844     for(list=0; list<h->list_count; list++){
2845         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
2846
2847         if(get_bits1(&s->gb)){
2848             int pred= h->curr_pic_num;
2849
2850             for(index=0; ; index++){
2851                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb_31(&s->gb);
2852                 unsigned int pic_id;
2853                 int i;
2854                 Picture *ref = NULL;
2855
2856                 if(reordering_of_pic_nums_idc==3)
2857                     break;
2858
2859                 if(index >= h->ref_count[list]){
2860                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
2861                     return -1;
2862                 }
2863
2864                 if(reordering_of_pic_nums_idc<3){
2865                     if(reordering_of_pic_nums_idc<2){
2866                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
2867                         int frame_num;
2868
2869                         if(abs_diff_pic_num > h->max_pic_num){
2870                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
2871                             return -1;
2872                         }
2873
2874                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
2875                         else                                pred+= abs_diff_pic_num;
2876                         pred &= h->max_pic_num - 1;
2877
2878                         frame_num = pic_num_extract(h, pred, &pic_structure);
2879
2880                         for(i= h->short_ref_count-1; i>=0; i--){
2881                             ref = h->short_ref[i];
2882                             assert(ref->reference);
2883                             assert(!ref->long_ref);
2884                             if(
2885                                    ref->frame_num == frame_num &&
2886                                    (ref->reference & pic_structure)
2887                               )
2888                                 break;
2889                         }
2890                         if(i>=0)
2891                             ref->pic_id= pred;
2892                     }else{
2893                         int long_idx;
2894                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
2895
2896                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
2897
2898                         if(long_idx>31){
2899                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
2900                             return -1;
2901                         }
2902                         ref = h->long_ref[long_idx];
2903                         assert(!(ref && !ref->reference));
2904                         if(ref && (ref->reference & pic_structure)){
2905                             ref->pic_id= pic_id;
2906                             assert(ref->long_ref);
2907                             i=0;
2908                         }else{
2909                             i=-1;
2910                         }
2911                     }
2912
2913                     if (i < 0) {
2914                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
2915                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
2916                     } else {
2917                         for(i=index; i+1<h->ref_count[list]; i++){
2918                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
2919                                 break;
2920                         }
2921                         for(; i > index; i--){
2922                             h->ref_list[list][i]= h->ref_list[list][i-1];
2923                         }
2924                         h->ref_list[list][index]= *ref;
2925                         if (FIELD_PICTURE){
2926                             pic_as_field(&h->ref_list[list][index], pic_structure);
2927                         }
2928                     }
2929                 }else{
2930                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
2931                     return -1;
2932                 }
2933             }
2934         }
2935     }
2936     for(list=0; list<h->list_count; list++){
2937         for(index= 0; index < h->ref_count[list]; index++){
2938             if(!h->ref_list[list][index].data[0]){
2939                 av_log(h->s.avctx, AV_LOG_ERROR, "Missing reference picture\n");
2940                 h->ref_list[list][index]= s->current_picture; //FIXME this is not a sensible solution
2941             }
2942         }
2943     }
2944
2945     return 0;
2946 }
2947
2948 static void fill_mbaff_ref_list(H264Context *h){
2949     int list, i, j;
2950     for(list=0; list<2; list++){ //FIXME try list_count
2951         for(i=0; i<h->ref_count[list]; i++){
2952             Picture *frame = &h->ref_list[list][i];
2953             Picture *field = &h->ref_list[list][16+2*i];
2954             field[0] = *frame;
2955             for(j=0; j<3; j++)
2956                 field[0].linesize[j] <<= 1;
2957             field[0].reference = PICT_TOP_FIELD;
2958             field[0].poc= field[0].field_poc[0];
2959             field[1] = field[0];
2960             for(j=0; j<3; j++)
2961                 field[1].data[j] += frame->linesize[j];
2962             field[1].reference = PICT_BOTTOM_FIELD;
2963             field[1].poc= field[1].field_poc[1];
2964
2965             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
2966             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
2967             for(j=0; j<2; j++){
2968                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
2969                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
2970             }
2971         }
2972     }
2973     for(j=0; j<h->ref_count[1]; j++){
2974         for(i=0; i<h->ref_count[0]; i++)
2975             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
2976         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
2977         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
2978     }
2979 }
2980
2981 static int pred_weight_table(H264Context *h){
2982     MpegEncContext * const s = &h->s;
2983     int list, i;
2984     int luma_def, chroma_def;
2985
2986     h->use_weight= 0;
2987     h->use_weight_chroma= 0;
2988     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2989     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2990     luma_def = 1<<h->luma_log2_weight_denom;
2991     chroma_def = 1<<h->chroma_log2_weight_denom;
2992
2993     for(list=0; list<2; list++){
2994         for(i=0; i<h->ref_count[list]; i++){
2995             int luma_weight_flag, chroma_weight_flag;
2996
2997             luma_weight_flag= get_bits1(&s->gb);
2998             if(luma_weight_flag){
2999                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3000                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3001                 if(   h->luma_weight[list][i] != luma_def
3002                    || h->luma_offset[list][i] != 0)
3003                     h->use_weight= 1;
3004             }else{
3005                 h->luma_weight[list][i]= luma_def;
3006                 h->luma_offset[list][i]= 0;
3007             }
3008
3009             if(CHROMA){
3010                 chroma_weight_flag= get_bits1(&s->gb);
3011                 if(chroma_weight_flag){
3012                     int j;
3013                     for(j=0; j<2; j++){
3014                         h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3015                         h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3016                         if(   h->chroma_weight[list][i][j] != chroma_def
3017                         || h->chroma_offset[list][i][j] != 0)
3018                             h->use_weight_chroma= 1;
3019                     }
3020                 }else{
3021                     int j;
3022                     for(j=0; j<2; j++){
3023                         h->chroma_weight[list][i][j]= chroma_def;
3024                         h->chroma_offset[list][i][j]= 0;
3025                     }
3026                 }
3027             }
3028         }
3029         if(h->slice_type_nos != FF_B_TYPE) break;
3030     }
3031     h->use_weight= h->use_weight || h->use_weight_chroma;
3032     return 0;
3033 }
3034
3035 static void implicit_weight_table(H264Context *h){
3036     MpegEncContext * const s = &h->s;
3037     int ref0, ref1;
3038     int cur_poc = s->current_picture_ptr->poc;
3039
3040     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3041        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3042         h->use_weight= 0;
3043         h->use_weight_chroma= 0;
3044         return;
3045     }
3046
3047     h->use_weight= 2;
3048     h->use_weight_chroma= 2;
3049     h->luma_log2_weight_denom= 5;
3050     h->chroma_log2_weight_denom= 5;
3051
3052     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3053         int poc0 = h->ref_list[0][ref0].poc;
3054         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3055             int poc1 = h->ref_list[1][ref1].poc;
3056             int td = av_clip(poc1 - poc0, -128, 127);
3057             if(td){
3058                 int tb = av_clip(cur_poc - poc0, -128, 127);
3059                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3060                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3061                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3062                     h->implicit_weight[ref0][ref1] = 32;
3063                 else
3064                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3065             }else
3066                 h->implicit_weight[ref0][ref1] = 32;
3067         }
3068     }
3069 }
3070
3071 /**
3072  * Mark a picture as no longer needed for reference. The refmask
3073  * argument allows unreferencing of individual fields or the whole frame.
3074  * If the picture becomes entirely unreferenced, but is being held for
3075  * display purposes, it is marked as such.
3076  * @param refmask mask of fields to unreference; the mask is bitwise
3077  *                anded with the reference marking of pic
3078  * @return non-zero if pic becomes entirely unreferenced (except possibly
3079  *         for display purposes) zero if one of the fields remains in
3080  *         reference
3081  */
3082 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3083     int i;
3084     if (pic->reference &= refmask) {
3085         return 0;
3086     } else {
3087         for(i = 0; h->delayed_pic[i]; i++)
3088             if(pic == h->delayed_pic[i]){
3089                 pic->reference=DELAYED_PIC_REF;
3090                 break;
3091             }
3092         return 1;
3093     }
3094 }
3095
3096 /**
3097  * instantaneous decoder refresh.
3098  */
3099 static void idr(H264Context *h){
3100     int i;
3101
3102     for(i=0; i<16; i++){
3103         remove_long(h, i, 0);
3104     }
3105     assert(h->long_ref_count==0);
3106
3107     for(i=0; i<h->short_ref_count; i++){
3108         unreference_pic(h, h->short_ref[i], 0);
3109         h->short_ref[i]= NULL;
3110     }
3111     h->short_ref_count=0;
3112     h->prev_frame_num= 0;
3113     h->prev_frame_num_offset= 0;
3114     h->prev_poc_msb=
3115     h->prev_poc_lsb= 0;
3116 }
3117
3118 /* forget old pics after a seek */
3119 static void flush_dpb(AVCodecContext *avctx){
3120     H264Context *h= avctx->priv_data;
3121     int i;
3122     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
3123         if(h->delayed_pic[i])
3124             h->delayed_pic[i]->reference= 0;
3125         h->delayed_pic[i]= NULL;
3126     }
3127     h->outputed_poc= INT_MIN;
3128     idr(h);
3129     if(h->s.current_picture_ptr)
3130         h->s.current_picture_ptr->reference= 0;
3131     h->s.first_field= 0;
3132     ff_mpeg_flush(avctx);
3133 }
3134
3135 /**
3136  * Find a Picture in the short term reference list by frame number.
3137  * @param frame_num frame number to search for
3138  * @param idx the index into h->short_ref where returned picture is found
3139  *            undefined if no picture found.
3140  * @return pointer to the found picture, or NULL if no pic with the provided
3141  *                 frame number is found
3142  */
3143 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3144     MpegEncContext * const s = &h->s;
3145     int i;
3146
3147     for(i=0; i<h->short_ref_count; i++){
3148         Picture *pic= h->short_ref[i];
3149         if(s->avctx->debug&FF_DEBUG_MMCO)
3150             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3151         if(pic->frame_num == frame_num) {
3152             *idx = i;
3153             return pic;
3154         }
3155     }
3156     return NULL;
3157 }
3158
3159 /**
3160  * Remove a picture from the short term reference list by its index in
3161  * that list.  This does no checking on the provided index; it is assumed
3162  * to be valid. Other list entries are shifted down.
3163  * @param i index into h->short_ref of picture to remove.
3164  */
3165 static void remove_short_at_index(H264Context *h, int i){
3166     assert(i >= 0 && i < h->short_ref_count);
3167     h->short_ref[i]= NULL;
3168     if (--h->short_ref_count)
3169         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3170 }
3171
3172 /**
3173  *
3174  * @return the removed picture or NULL if an error occurs
3175  */
3176 static Picture * remove_short(H264Context *h, int frame_num, int ref_mask){
3177     MpegEncContext * const s = &h->s;
3178     Picture *pic;
3179     int i;
3180
3181     if(s->avctx->debug&FF_DEBUG_MMCO)
3182         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3183
3184     pic = find_short(h, frame_num, &i);
3185     if (pic){
3186         if(unreference_pic(h, pic, ref_mask))
3187         remove_short_at_index(h, i);
3188     }
3189
3190     return pic;
3191 }
3192
3193 /**
3194  * Remove a picture from the long term reference list by its index in
3195  * that list.
3196  * @return the removed picture or NULL if an error occurs
3197  */
3198 static Picture * remove_long(H264Context *h, int i, int ref_mask){
3199     Picture *pic;
3200
3201     pic= h->long_ref[i];
3202     if (pic){
3203         if(unreference_pic(h, pic, ref_mask)){
3204             assert(h->long_ref[i]->long_ref == 1);
3205             h->long_ref[i]->long_ref= 0;
3206             h->long_ref[i]= NULL;
3207             h->long_ref_count--;
3208         }
3209     }
3210
3211     return pic;
3212 }
3213
3214 /**
3215  * print short term list
3216  */
3217 static void print_short_term(H264Context *h) {
3218     uint32_t i;
3219     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3220         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3221         for(i=0; i<h->short_ref_count; i++){
3222             Picture *pic= h->short_ref[i];
3223             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3224         }
3225     }
3226 }
3227
3228 /**
3229  * print long term list
3230  */
3231 static void print_long_term(H264Context *h) {
3232     uint32_t i;
3233     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3234         av_log(h->s.avctx, AV_LOG_DEBUG, "long term list:\n");
3235         for(i = 0; i < 16; i++){
3236             Picture *pic= h->long_ref[i];
3237             if (pic) {
3238                 av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3239             }
3240         }
3241     }
3242 }
3243
3244 /**
3245  * Executes the reference picture marking (memory management control operations).
3246  */
3247 static int execute_ref_pic_marking(H264Context *h, MMCO *mmco, int mmco_count){
3248     MpegEncContext * const s = &h->s;
3249     int i, j;
3250     int current_ref_assigned=0;
3251     Picture *pic;
3252
3253     if((s->avctx->debug&FF_DEBUG_MMCO) && mmco_count==0)
3254         av_log(h->s.avctx, AV_LOG_DEBUG, "no mmco here\n");
3255
3256     for(i=0; i<mmco_count; i++){
3257         int structure, frame_num;
3258         if(s->avctx->debug&FF_DEBUG_MMCO)
3259             av_log(h->s.avctx, AV_LOG_DEBUG, "mmco:%d %d %d\n", h->mmco[i].opcode, h->mmco[i].short_pic_num, h->mmco[i].long_arg);
3260
3261         if(   mmco[i].opcode == MMCO_SHORT2UNUSED
3262            || mmco[i].opcode == MMCO_SHORT2LONG){
3263             frame_num = pic_num_extract(h, mmco[i].short_pic_num, &structure);
3264             pic = find_short(h, frame_num, &j);
3265             if(!pic){
3266                 if(mmco[i].opcode != MMCO_SHORT2LONG || !h->long_ref[mmco[i].long_arg]
3267                    || h->long_ref[mmco[i].long_arg]->frame_num != frame_num)
3268                 av_log(h->s.avctx, AV_LOG_ERROR, "mmco: unref short failure\n");
3269                 continue;
3270             }
3271         }
3272
3273         switch(mmco[i].opcode){
3274         case MMCO_SHORT2UNUSED:
3275             if(s->avctx->debug&FF_DEBUG_MMCO)
3276                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref short %d count %d\n", h->mmco[i].short_pic_num, h->short_ref_count);
3277             remove_short(h, frame_num, structure ^ PICT_FRAME);
3278             break;
3279         case MMCO_SHORT2LONG:
3280                 if (h->long_ref[mmco[i].long_arg] != pic)
3281                     remove_long(h, mmco[i].long_arg, 0);
3282
3283                 remove_short_at_index(h, j);
3284                 h->long_ref[ mmco[i].long_arg ]= pic;
3285                 if (h->long_ref[ mmco[i].long_arg ]){
3286                     h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3287                     h->long_ref_count++;
3288                 }
3289             break;
3290         case MMCO_LONG2UNUSED:
3291             j = pic_num_extract(h, mmco[i].long_arg, &structure);
3292             pic = h->long_ref[j];
3293             if (pic) {
3294                 remove_long(h, j, structure ^ PICT_FRAME);
3295             } else if(s->avctx->debug&FF_DEBUG_MMCO)
3296                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref long failure\n");
3297             break;
3298         case MMCO_LONG:
3299                     // Comment below left from previous code as it is an interresting note.
3300                     /* First field in pair is in short term list or
3301                      * at a different long term index.
3302                      * This is not allowed; see 7.4.3.3, notes 2 and 3.
3303                      * Report the problem and keep the pair where it is,
3304                      * and mark this field valid.
3305                      */
3306
3307             if (h->long_ref[mmco[i].long_arg] != s->current_picture_ptr) {
3308                 remove_long(h, mmco[i].long_arg, 0);
3309
3310                 h->long_ref[ mmco[i].long_arg ]= s->current_picture_ptr;
3311                 h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3312                 h->long_ref_count++;
3313             }
3314
3315             s->current_picture_ptr->reference |= s->picture_structure;
3316             current_ref_assigned=1;
3317             break;
3318         case MMCO_SET_MAX_LONG:
3319             assert(mmco[i].long_arg <= 16);
3320             // just remove the long term which index is greater than new max
3321             for(j = mmco[i].long_arg; j<16; j++){
3322                 remove_long(h, j, 0);
3323             }
3324             break;
3325         case MMCO_RESET:
3326             while(h->short_ref_count){
3327                 remove_short(h, h->short_ref[0]->frame_num, 0);
3328             }
3329             for(j = 0; j < 16; j++) {
3330                 remove_long(h, j, 0);
3331             }
3332             s->current_picture_ptr->poc=
3333             s->current_picture_ptr->field_poc[0]=
3334             s->current_picture_ptr->field_poc[1]=
3335             h->poc_lsb=
3336             h->poc_msb=
3337             h->frame_num=
3338             s->current_picture_ptr->frame_num= 0;
3339             break;
3340         default: assert(0);
3341         }
3342     }
3343
3344     if (!current_ref_assigned) {
3345         /* Second field of complementary field pair; the first field of
3346          * which is already referenced. If short referenced, it
3347          * should be first entry in short_ref. If not, it must exist
3348          * in long_ref; trying to put it on the short list here is an
3349          * error in the encoded bit stream (ref: 7.4.3.3, NOTE 2 and 3).
3350          */
3351         if (h->short_ref_count && h->short_ref[0] == s->current_picture_ptr) {
3352             /* Just mark the second field valid */
3353             s->current_picture_ptr->reference = PICT_FRAME;
3354         } else if (s->current_picture_ptr->long_ref) {
3355             av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term reference "
3356                                              "assignment for second field "
3357                                              "in complementary field pair "
3358                                              "(first field is long term)\n");
3359         } else {
3360             pic= remove_short(h, s->current_picture_ptr->frame_num, 0);
3361             if(pic){
3362                 av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term buffer state detected\n");
3363             }
3364
3365             if(h->short_ref_count)
3366                 memmove(&h->short_ref[1], &h->short_ref[0], h->short_ref_count*sizeof(Picture*));
3367
3368             h->short_ref[0]= s->current_picture_ptr;
3369             h->short_ref_count++;
3370             s->current_picture_ptr->reference |= s->picture_structure;
3371         }
3372     }
3373
3374     if (h->long_ref_count + h->short_ref_count > h->sps.ref_frame_count){
3375
3376         /* We have too many reference frames, probably due to corrupted
3377          * stream. Need to discard one frame. Prevents overrun of the
3378          * short_ref and long_ref buffers.
3379          */
3380         av_log(h->s.avctx, AV_LOG_ERROR,
3381                "number of reference frames exceeds max (probably "
3382                "corrupt input), discarding one\n");
3383
3384         if (h->long_ref_count && !h->short_ref_count) {
3385             for (i = 0; i < 16; ++i)
3386                 if (h->long_ref[i])
3387                     break;
3388
3389             assert(i < 16);
3390             remove_long(h, i, 0);
3391         } else {
3392             pic = h->short_ref[h->short_ref_count - 1];
3393             remove_short(h, pic->frame_num, 0);
3394         }
3395     }
3396
3397     print_short_term(h);
3398     print_long_term(h);
3399     return 0;
3400 }
3401
3402 static int decode_ref_pic_marking(H264Context *h, GetBitContext *gb){
3403     MpegEncContext * const s = &h->s;
3404     int i;
3405
3406     h->mmco_index= 0;
3407     if(h->nal_unit_type == NAL_IDR_SLICE){ //FIXME fields
3408         s->broken_link= get_bits1(gb) -1;
3409         if(get_bits1(gb)){
3410             h->mmco[0].opcode= MMCO_LONG;
3411             h->mmco[0].long_arg= 0;
3412             h->mmco_index= 1;
3413         }
3414     }else{
3415         if(get_bits1(gb)){ // adaptive_ref_pic_marking_mode_flag
3416             for(i= 0; i<MAX_MMCO_COUNT; i++) {
3417                 MMCOOpcode opcode= get_ue_golomb_31(gb);
3418
3419                 h->mmco[i].opcode= opcode;
3420                 if(opcode==MMCO_SHORT2UNUSED || opcode==MMCO_SHORT2LONG){
3421                     h->mmco[i].short_pic_num= (h->curr_pic_num - get_ue_golomb(gb) - 1) & (h->max_pic_num - 1);
3422 /*                    if(h->mmco[i].short_pic_num >= h->short_ref_count || h->short_ref[ h->mmco[i].short_pic_num ] == NULL){
3423                         av_log(s->avctx, AV_LOG_ERROR, "illegal short ref in memory management control operation %d\n", mmco);
3424                         return -1;
3425                     }*/
3426                 }
3427                 if(opcode==MMCO_SHORT2LONG || opcode==MMCO_LONG2UNUSED || opcode==MMCO_LONG || opcode==MMCO_SET_MAX_LONG){
3428                     unsigned int long_arg= get_ue_golomb_31(gb);
3429                     if(long_arg >= 32 || (long_arg >= 16 && !(opcode == MMCO_LONG2UNUSED && FIELD_PICTURE))){
3430                         av_log(h->s.avctx, AV_LOG_ERROR, "illegal long ref in memory management control operation %d\n", opcode);
3431                         return -1;
3432                     }
3433                     h->mmco[i].long_arg= long_arg;
3434                 }
3435
3436                 if(opcode > (unsigned)MMCO_LONG){
3437                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal memory management control operation %d\n", opcode);
3438                     return -1;
3439                 }
3440                 if(opcode == MMCO_END)
3441                     break;
3442             }
3443             h->mmco_index= i;
3444         }else{
3445             assert(h->long_ref_count + h->short_ref_count <= h->sps.ref_frame_count);
3446
3447             if(h->short_ref_count && h->long_ref_count + h->short_ref_count == h->sps.ref_frame_count &&
3448                     !(FIELD_PICTURE && !s->first_field && s->current_picture_ptr->reference)) {
3449                 h->mmco[0].opcode= MMCO_SHORT2UNUSED;
3450                 h->mmco[0].short_pic_num= h->short_ref[ h->short_ref_count - 1 ]->frame_num;
3451                 h->mmco_index= 1;
3452                 if (FIELD_PICTURE) {
3453                     h->mmco[0].short_pic_num *= 2;
3454                     h->mmco[1].opcode= MMCO_SHORT2UNUSED;
3455                     h->mmco[1].short_pic_num= h->mmco[0].short_pic_num + 1;
3456                     h->mmco_index= 2;
3457                 }
3458             }
3459         }
3460     }
3461
3462     return 0;
3463 }
3464
3465 static int init_poc(H264Context *h){
3466     MpegEncContext * const s = &h->s;
3467     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
3468     int field_poc[2];
3469     Picture *cur = s->current_picture_ptr;
3470
3471     h->frame_num_offset= h->prev_frame_num_offset;
3472     if(h->frame_num < h->prev_frame_num)
3473         h->frame_num_offset += max_frame_num;
3474
3475     if(h->sps.poc_type==0){
3476         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
3477
3478         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
3479             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
3480         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
3481             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
3482         else
3483             h->poc_msb = h->prev_poc_msb;
3484 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
3485         field_poc[0] =
3486         field_poc[1] = h->poc_msb + h->poc_lsb;
3487         if(s->picture_structure == PICT_FRAME)
3488             field_poc[1] += h->delta_poc_bottom;
3489     }else if(h->sps.poc_type==1){
3490         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
3491         int i;
3492
3493         if(h->sps.poc_cycle_length != 0)
3494             abs_frame_num = h->frame_num_offset + h->frame_num;
3495         else
3496             abs_frame_num = 0;
3497
3498         if(h->nal_ref_idc==0 && abs_frame_num > 0)
3499             abs_frame_num--;
3500
3501         expected_delta_per_poc_cycle = 0;
3502         for(i=0; i < h->sps.poc_cycle_length; i++)
3503             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
3504
3505         if(abs_frame_num > 0){
3506             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
3507             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
3508
3509             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
3510             for(i = 0; i <= frame_num_in_poc_cycle; i++)
3511                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
3512         } else
3513             expectedpoc = 0;
3514
3515         if(h->nal_ref_idc == 0)
3516             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
3517
3518         field_poc[0] = expectedpoc + h->delta_poc[0];
3519         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
3520
3521         if(s->picture_structure == PICT_FRAME)
3522             field_poc[1] += h->delta_poc[1];
3523     }else{
3524         int poc= 2*(h->frame_num_offset + h->frame_num);
3525
3526         if(!h->nal_ref_idc)
3527             poc--;
3528
3529         field_poc[0]= poc;
3530         field_poc[1]= poc;
3531     }
3532
3533     if(s->picture_structure != PICT_BOTTOM_FIELD)
3534         s->current_picture_ptr->field_poc[0]= field_poc[0];
3535     if(s->picture_structure != PICT_TOP_FIELD)
3536         s->current_picture_ptr->field_poc[1]= field_poc[1];
3537     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
3538
3539     return 0;
3540 }
3541
3542
3543 /**
3544  * initialize scan tables
3545  */
3546 static void init_scan_tables(H264Context *h){
3547     MpegEncContext * const s = &h->s;
3548     int i;
3549     if(s->dsp.h264_idct_add == ff_h264_idct_add_c){ //FIXME little ugly
3550         memcpy(h->zigzag_scan, zigzag_scan, 16*sizeof(uint8_t));
3551         memcpy(h-> field_scan,  field_scan, 16*sizeof(uint8_t));
3552     }else{
3553         for(i=0; i<16; i++){
3554 #define T(x) (x>>2) | ((x<<2) & 0xF)
3555             h->zigzag_scan[i] = T(zigzag_scan[i]);
3556             h-> field_scan[i] = T( field_scan[i]);
3557 #undef T
3558         }
3559     }
3560     if(s->dsp.h264_idct8_add == ff_h264_idct8_add_c){
3561         memcpy(h->zigzag_scan8x8,       zigzag_scan8x8,       64*sizeof(uint8_t));
3562         memcpy(h->zigzag_scan8x8_cavlc, zigzag_scan8x8_cavlc, 64*sizeof(uint8_t));
3563         memcpy(h->field_scan8x8,        field_scan8x8,        64*sizeof(uint8_t));
3564         memcpy(h->field_scan8x8_cavlc,  field_scan8x8_cavlc,  64*sizeof(uint8_t));
3565     }else{
3566         for(i=0; i<64; i++){
3567 #define T(x) (x>>3) | ((x&7)<<3)
3568             h->zigzag_scan8x8[i]       = T(zigzag_scan8x8[i]);
3569             h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
3570             h->field_scan8x8[i]        = T(field_scan8x8[i]);
3571             h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
3572 #undef T
3573         }
3574     }
3575     if(h->sps.transform_bypass){ //FIXME same ugly
3576         h->zigzag_scan_q0          = zigzag_scan;
3577         h->zigzag_scan8x8_q0       = zigzag_scan8x8;
3578         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
3579         h->field_scan_q0           = field_scan;
3580         h->field_scan8x8_q0        = field_scan8x8;
3581         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
3582     }else{
3583         h->zigzag_scan_q0          = h->zigzag_scan;
3584         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
3585         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
3586         h->field_scan_q0           = h->field_scan;
3587         h->field_scan8x8_q0        = h->field_scan8x8;
3588         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
3589     }
3590 }
3591
3592 /**
3593  * Replicates H264 "master" context to thread contexts.
3594  */
3595 static void clone_slice(H264Context *dst, H264Context *src)
3596 {
3597     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
3598     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
3599     dst->s.current_picture      = src->s.current_picture;
3600     dst->s.linesize             = src->s.linesize;
3601     dst->s.uvlinesize           = src->s.uvlinesize;
3602     dst->s.first_field          = src->s.first_field;
3603
3604     dst->prev_poc_msb           = src->prev_poc_msb;
3605     dst->prev_poc_lsb           = src->prev_poc_lsb;
3606     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
3607     dst->prev_frame_num         = src->prev_frame_num;
3608     dst->short_ref_count        = src->short_ref_count;
3609
3610     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
3611     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
3612     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
3613     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
3614
3615     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
3616     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
3617 }
3618
3619 /**
3620  * decodes a slice header.
3621  * This will also call MPV_common_init() and frame_start() as needed.
3622  *
3623  * @param h h264context
3624  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
3625  *
3626  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
3627  */
3628 static int decode_slice_header(H264Context *h, H264Context *h0){
3629     MpegEncContext * const s = &h->s;
3630     MpegEncContext * const s0 = &h0->s;
3631     unsigned int first_mb_in_slice;
3632     unsigned int pps_id;
3633     int num_ref_idx_active_override_flag;
3634     unsigned int slice_type, tmp, i, j;
3635     int default_ref_list_done = 0;
3636     int last_pic_structure;
3637
3638     s->dropable= h->nal_ref_idc == 0;
3639
3640     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc){
3641         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
3642         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
3643     }else{
3644         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
3645         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
3646     }
3647
3648     first_mb_in_slice= get_ue_golomb(&s->gb);
3649
3650     if((s->flags2 & CODEC_FLAG2_CHUNKS) && first_mb_in_slice == 0){
3651         h0->current_slice = 0;
3652         if (!s0->first_field)
3653             s->current_picture_ptr= NULL;
3654     }
3655
3656     slice_type= get_ue_golomb_31(&s->gb);
3657     if(slice_type > 9){
3658         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);
3659         return -1;
3660     }
3661     if(slice_type > 4){
3662         slice_type -= 5;
3663         h->slice_type_fixed=1;
3664     }else
3665         h->slice_type_fixed=0;
3666
3667     slice_type= golomb_to_pict_type[ slice_type ];
3668     if (slice_type == FF_I_TYPE
3669         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
3670         default_ref_list_done = 1;
3671     }
3672     h->slice_type= slice_type;
3673     h->slice_type_nos= slice_type & 3;
3674
3675     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
3676     if (s->pict_type == FF_B_TYPE && s0->last_picture_ptr == NULL) {
3677         av_log(h->s.avctx, AV_LOG_ERROR,
3678                "B picture before any references, skipping\n");
3679         return -1;
3680     }
3681
3682     pps_id= get_ue_golomb(&s->gb);
3683     if(pps_id>=MAX_PPS_COUNT){
3684         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
3685         return -1;
3686     }
3687     if(!h0->pps_buffers[pps_id]) {
3688         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS referenced\n");
3689         return -1;
3690     }
3691     h->pps= *h0->pps_buffers[pps_id];
3692
3693     if(!h0->sps_buffers[h->pps.sps_id]) {
3694         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS referenced\n");
3695         return -1;
3696     }
3697     h->sps = *h0->sps_buffers[h->pps.sps_id];
3698
3699     if(h == h0 && h->dequant_coeff_pps != pps_id){
3700         h->dequant_coeff_pps = pps_id;
3701         init_dequant_tables(h);
3702     }
3703
3704     s->mb_width= h->sps.mb_width;
3705     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
3706
3707     h->b_stride=  s->mb_width*4;
3708     h->b8_stride= s->mb_width*2;
3709
3710     s->width = 16*s->mb_width - 2*FFMIN(h->sps.crop_right, 7);
3711     if(h->sps.frame_mbs_only_flag)
3712         s->height= 16*s->mb_height - 2*FFMIN(h->sps.crop_bottom, 7);
3713     else
3714         s->height= 16*s->mb_height - 4*FFMIN(h->sps.crop_bottom, 3);
3715
3716     if (s->context_initialized
3717         && (   s->width != s->avctx->width || s->height != s->avctx->height)) {
3718         if(h != h0)
3719             return -1;   // width / height changed during parallelized decoding
3720         free_tables(h);
3721         flush_dpb(s->avctx);
3722         MPV_common_end(s);
3723     }
3724     if (!s->context_initialized) {
3725         if(h != h0)
3726             return -1;  // we cant (re-)initialize context during parallel decoding
3727         if (MPV_common_init(s) < 0)
3728             return -1;
3729         s->first_field = 0;
3730
3731         init_scan_tables(h);
3732         alloc_tables(h);
3733
3734         for(i = 1; i < s->avctx->thread_count; i++) {
3735             H264Context *c;
3736             c = h->thread_context[i] = av_malloc(sizeof(H264Context));
3737             memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
3738             memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
3739             c->sps = h->sps;
3740             c->pps = h->pps;
3741             init_scan_tables(c);
3742             clone_tables(c, h);
3743         }
3744
3745         for(i = 0; i < s->avctx->thread_count; i++)
3746             if(context_init(h->thread_context[i]) < 0)
3747                 return -1;
3748
3749         s->avctx->width = s->width;
3750         s->avctx->height = s->height;
3751         s->avctx->sample_aspect_ratio= h->sps.sar;
3752         if(!s->avctx->sample_aspect_ratio.den)
3753             s->avctx->sample_aspect_ratio.den = 1;
3754
3755         if(h->sps.timing_info_present_flag){
3756             s->avctx->time_base= (AVRational){h->sps.num_units_in_tick * 2, h->sps.time_scale};
3757             if(h->x264_build > 0 && h->x264_build < 44)
3758                 s->avctx->time_base.den *= 2;
3759             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
3760                       s->avctx->time_base.num, s->avctx->time_base.den, 1<<30);
3761         }
3762     }
3763
3764     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
3765
3766     h->mb_mbaff = 0;
3767     h->mb_aff_frame = 0;
3768     last_pic_structure = s0->picture_structure;
3769     if(h->sps.frame_mbs_only_flag){
3770         s->picture_structure= PICT_FRAME;
3771     }else{
3772         if(get_bits1(&s->gb)) { //field_pic_flag
3773             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
3774         } else {
3775             s->picture_structure= PICT_FRAME;
3776             h->mb_aff_frame = h->sps.mb_aff;
3777         }
3778     }
3779     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
3780
3781     if(h0->current_slice == 0){
3782         while(h->frame_num !=  h->prev_frame_num &&
3783               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
3784             av_log(NULL, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
3785             frame_start(h);
3786             h->prev_frame_num++;
3787             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
3788             s->current_picture_ptr->frame_num= h->prev_frame_num;
3789             execute_ref_pic_marking(h, NULL, 0);
3790         }
3791
3792         /* See if we have a decoded first field looking for a pair... */
3793         if (s0->first_field) {
3794             assert(s0->current_picture_ptr);
3795             assert(s0->current_picture_ptr->data[0]);
3796             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
3797
3798             /* figure out if we have a complementary field pair */
3799             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3800                 /*
3801                  * Previous field is unmatched. Don't display it, but let it
3802                  * remain for reference if marked as such.
3803                  */
3804                 s0->current_picture_ptr = NULL;
3805                 s0->first_field = FIELD_PICTURE;
3806
3807             } else {
3808                 if (h->nal_ref_idc &&
3809                         s0->current_picture_ptr->reference &&
3810                         s0->current_picture_ptr->frame_num != h->frame_num) {
3811                     /*
3812                      * This and previous field were reference, but had
3813                      * different frame_nums. Consider this field first in
3814                      * pair. Throw away previous field except for reference
3815                      * purposes.
3816                      */
3817                     s0->first_field = 1;
3818                     s0->current_picture_ptr = NULL;
3819
3820                 } else {
3821                     /* Second field in complementary pair */
3822                     s0->first_field = 0;
3823                 }
3824             }
3825
3826         } else {
3827             /* Frame or first field in a potentially complementary pair */
3828             assert(!s0->current_picture_ptr);
3829             s0->first_field = FIELD_PICTURE;
3830         }
3831
3832         if((!FIELD_PICTURE || s0->first_field) && frame_start(h) < 0) {
3833             s0->first_field = 0;
3834             return -1;
3835         }
3836     }
3837     if(h != h0)
3838         clone_slice(h, h0);
3839
3840     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
3841
3842     assert(s->mb_num == s->mb_width * s->mb_height);
3843     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
3844        first_mb_in_slice                    >= s->mb_num){
3845         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
3846         return -1;
3847     }
3848     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
3849     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
3850     if (s->picture_structure == PICT_BOTTOM_FIELD)
3851         s->resync_mb_y = s->mb_y = s->mb_y + 1;
3852     assert(s->mb_y < s->mb_height);
3853
3854     if(s->picture_structure==PICT_FRAME){
3855         h->curr_pic_num=   h->frame_num;
3856         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
3857     }else{
3858         h->curr_pic_num= 2*h->frame_num + 1;
3859         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
3860     }
3861
3862     if(h->nal_unit_type == NAL_IDR_SLICE){
3863         get_ue_golomb(&s->gb); /* idr_pic_id */
3864     }
3865
3866     if(h->sps.poc_type==0){
3867         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
3868
3869         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
3870             h->delta_poc_bottom= get_se_golomb(&s->gb);
3871         }
3872     }
3873
3874     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
3875         h->delta_poc[0]= get_se_golomb(&s->gb);
3876
3877         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
3878             h->delta_poc[1]= get_se_golomb(&s->gb);
3879     }
3880
3881     init_poc(h);
3882
3883     if(h->pps.redundant_pic_cnt_present){
3884         h->redundant_pic_count= get_ue_golomb(&s->gb);
3885     }
3886
3887     //set defaults, might be overridden a few lines later
3888     h->ref_count[0]= h->pps.ref_count[0];
3889     h->ref_count[1]= h->pps.ref_count[1];
3890
3891     if(h->slice_type_nos != FF_I_TYPE){
3892         if(h->slice_type_nos == FF_B_TYPE){
3893             h->direct_spatial_mv_pred= get_bits1(&s->gb);
3894         }
3895         num_ref_idx_active_override_flag= get_bits1(&s->gb);
3896
3897         if(num_ref_idx_active_override_flag){
3898             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
3899             if(h->slice_type_nos==FF_B_TYPE)
3900                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
3901
3902             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
3903                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3904                 h->ref_count[0]= h->ref_count[1]= 1;
3905                 return -1;
3906             }
3907         }
3908         if(h->slice_type_nos == FF_B_TYPE)
3909             h->list_count= 2;
3910         else
3911             h->list_count= 1;
3912     }else
3913         h->list_count= 0;
3914
3915     if(!default_ref_list_done){
3916         fill_default_ref_list(h);
3917     }
3918
3919     if(h->slice_type_nos!=FF_I_TYPE && decode_ref_pic_list_reordering(h) < 0)
3920         return -1;
3921
3922     if(h->slice_type_nos!=FF_I_TYPE){
3923         s->last_picture_ptr= &h->ref_list[0][0];
3924         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
3925     }
3926     if(h->slice_type_nos==FF_B_TYPE){
3927         s->next_picture_ptr= &h->ref_list[1][0];
3928         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
3929     }
3930
3931     if(   (h->pps.weighted_pred          && h->slice_type_nos == FF_P_TYPE )
3932        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== FF_B_TYPE ) )
3933         pred_weight_table(h);
3934     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== FF_B_TYPE)
3935         implicit_weight_table(h);
3936     else
3937         h->use_weight = 0;
3938
3939     if(h->nal_ref_idc)
3940         decode_ref_pic_marking(h0, &s->gb);
3941
3942     if(FRAME_MBAFF)
3943         fill_mbaff_ref_list(h);
3944
3945     if(h->slice_type_nos==FF_B_TYPE && !h->direct_spatial_mv_pred)
3946         direct_dist_scale_factor(h);
3947     direct_ref_list_init(h);
3948
3949     if( h->slice_type_nos != FF_I_TYPE && h->pps.cabac ){
3950         tmp = get_ue_golomb_31(&s->gb);
3951         if(tmp > 2){
3952             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3953             return -1;
3954         }
3955         h->cabac_init_idc= tmp;
3956     }
3957
3958     h->last_qscale_diff = 0;
3959     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3960     if(tmp>51){
3961         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3962         return -1;
3963     }
3964     s->qscale= tmp;
3965     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3966     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3967     //FIXME qscale / qp ... stuff
3968     if(h->slice_type == FF_SP_TYPE){
3969         get_bits1(&s->gb); /* sp_for_switch_flag */
3970     }
3971     if(h->slice_type==FF_SP_TYPE || h->slice_type == FF_SI_TYPE){
3972         get_se_golomb(&s->gb); /* slice_qs_delta */
3973     }
3974
3975     h->deblocking_filter = 1;
3976     h->slice_alpha_c0_offset = 0;
3977     h->slice_beta_offset = 0;
3978     if( h->pps.deblocking_filter_parameters_present ) {
3979         tmp= get_ue_golomb_31(&s->gb);
3980         if(tmp > 2){
3981             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
3982             return -1;
3983         }
3984         h->deblocking_filter= tmp;
3985         if(h->deblocking_filter < 2)
3986             h->deblocking_filter^= 1; // 1<->0
3987
3988         if( h->deblocking_filter ) {
3989             h->slice_alpha_c0_offset = get_se_golomb(&s->gb) << 1;
3990             h->slice_beta_offset = get_se_golomb(&s->gb) << 1;
3991         }
3992     }
3993
3994     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
3995        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != FF_I_TYPE)
3996        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == FF_B_TYPE)
3997        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
3998         h->deblocking_filter= 0;
3999
4000     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
4001         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
4002             /* Cheat slightly for speed:
4003                Do not bother to deblock across slices. */
4004             h->deblocking_filter = 2;
4005         } else {
4006             h0->max_contexts = 1;
4007             if(!h0->single_decode_warning) {
4008                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
4009                 h0->single_decode_warning = 1;
4010             }
4011             if(h != h0)
4012                 return 1; // deblocking switched inside frame
4013         }
4014     }
4015
4016 #if 0 //FMO
4017     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
4018         slice_group_change_cycle= get_bits(&s->gb, ?);
4019 #endif
4020
4021     h0->last_slice_type = slice_type;
4022     h->slice_num = ++h0->current_slice;
4023     if(h->slice_num >= MAX_SLICES){
4024         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
4025     }
4026
4027     for(j=0; j<2; j++){
4028         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
4029         ref2frm[0]=
4030         ref2frm[1]= -1;
4031         for(i=0; i<16; i++)
4032             ref2frm[i+2]= 4*h->ref_list[j][i].frame_num
4033                           +(h->ref_list[j][i].reference&3);
4034         ref2frm[18+0]=
4035         ref2frm[18+1]= -1;
4036         for(i=16; i<48; i++)
4037             ref2frm[i+4]= 4*h->ref_list[j][i].frame_num
4038                           +(h->ref_list[j][i].reference&3);
4039