Replace hard-coded SEI type constants with symbolic names
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file libavcodec/h264.c
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "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 "mathops.h"
36 #include "rectangle.h"
37 #include "vdpau_internal.h"
38
39 #include "cabac.h"
40 #if ARCH_X86
41 #include "x86/h264_i386.h"
42 #endif
43
44 //#undef NDEBUG
45 #include <assert.h>
46
47 /**
48  * Value of Picture.reference when Picture is not a reference picture, but
49  * is held for delayed output.
50  */
51 #define DELAYED_PIC_REF 4
52
53 static VLC coeff_token_vlc[4];
54 static VLC_TYPE coeff_token_vlc_tables[520+332+280+256][2];
55 static const int coeff_token_vlc_tables_size[4]={520,332,280,256};
56
57 static VLC chroma_dc_coeff_token_vlc;
58 static VLC_TYPE chroma_dc_coeff_token_vlc_table[256][2];
59 static const int chroma_dc_coeff_token_vlc_table_size = 256;
60
61 static VLC total_zeros_vlc[15];
62 static VLC_TYPE total_zeros_vlc_tables[15][512][2];
63 static const int total_zeros_vlc_tables_size = 512;
64
65 static VLC chroma_dc_total_zeros_vlc[3];
66 static VLC_TYPE chroma_dc_total_zeros_vlc_tables[3][8][2];
67 static const int chroma_dc_total_zeros_vlc_tables_size = 8;
68
69 static VLC run_vlc[6];
70 static VLC_TYPE run_vlc_tables[6][8][2];
71 static const int run_vlc_tables_size = 8;
72
73 static VLC run7_vlc;
74 static VLC_TYPE run7_vlc_table[96][2];
75 static const int run7_vlc_table_size = 96;
76
77 static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
78 static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
79 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);
80 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);
81 static Picture * remove_long(H264Context *h, int i, int ref_mask);
82
83 static av_always_inline uint32_t pack16to32(int a, int b){
84 #ifdef WORDS_BIGENDIAN
85    return (b&0xFFFF) + (a<<16);
86 #else
87    return (a&0xFFFF) + (b<<16);
88 #endif
89 }
90
91 static const uint8_t rem6[52]={
92 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,
93 };
94
95 static const uint8_t div6[52]={
96 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,
97 };
98
99 static const int left_block_options[4][8]={
100     {0,1,2,3,7,10,8,11},
101     {2,2,3,3,8,11,8,11},
102     {0,0,1,1,7,10,7,10},
103     {0,2,0,2,7,10,7,10}
104 };
105
106 #define LEVEL_TAB_BITS 8
107 static int8_t cavlc_level_tab[7][1<<LEVEL_TAB_BITS][2];
108
109 static void fill_caches(H264Context *h, int mb_type, int for_deblock){
110     MpegEncContext * const s = &h->s;
111     const int mb_xy= h->mb_xy;
112     int topleft_xy, top_xy, topright_xy, left_xy[2];
113     int topleft_type, top_type, topright_type, left_type[2];
114     const int * left_block;
115     int topleft_partition= -1;
116     int i;
117
118     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
119
120     //FIXME deblocking could skip the intra and nnz parts.
121     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
122         return;
123
124     /* Wow, what a mess, why didn't they simplify the interlacing & intra
125      * stuff, I can't imagine that these complex rules are worth it. */
126
127     topleft_xy = top_xy - 1;
128     topright_xy= top_xy + 1;
129     left_xy[1] = left_xy[0] = mb_xy-1;
130     left_block = left_block_options[0];
131     if(FRAME_MBAFF){
132         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
133         const int top_pair_xy      = pair_xy     - s->mb_stride;
134         const int topleft_pair_xy  = top_pair_xy - 1;
135         const int topright_pair_xy = top_pair_xy + 1;
136         const int topleft_mb_field_flag  = IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
137         const int top_mb_field_flag      = IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
138         const int topright_mb_field_flag = IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
139         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
140         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
141         const int bottom = (s->mb_y & 1);
142         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);
143
144         if (curr_mb_field_flag && (bottom || top_mb_field_flag)){
145             top_xy -= s->mb_stride;
146         }
147         if (curr_mb_field_flag && (bottom || topleft_mb_field_flag)){
148             topleft_xy -= s->mb_stride;
149         } else if(bottom && !curr_mb_field_flag && left_mb_field_flag) {
150             topleft_xy += s->mb_stride;
151             // take top left mv from the middle of the mb, as opposed to all other modes which use the bottom right partition
152             topleft_partition = 0;
153         }
154         if (curr_mb_field_flag && (bottom || topright_mb_field_flag)){
155             topright_xy -= s->mb_stride;
156         }
157         if (left_mb_field_flag != curr_mb_field_flag) {
158             left_xy[1] = left_xy[0] = pair_xy - 1;
159             if (curr_mb_field_flag) {
160                 left_xy[1] += s->mb_stride;
161                 left_block = left_block_options[3];
162             } else {
163                 left_block= left_block_options[2 - bottom];
164             }
165         }
166     }
167
168     h->top_mb_xy = top_xy;
169     h->left_mb_xy[0] = left_xy[0];
170     h->left_mb_xy[1] = left_xy[1];
171     if(for_deblock){
172         topleft_type = 0;
173         topright_type = 0;
174         top_type     = h->slice_table[top_xy     ] < 0xFFFF ? s->current_picture.mb_type[top_xy]     : 0;
175         left_type[0] = h->slice_table[left_xy[0] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[0]] : 0;
176         left_type[1] = h->slice_table[left_xy[1] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[1]] : 0;
177
178         if(MB_MBAFF && !IS_INTRA(mb_type)){
179             int list;
180             for(list=0; list<h->list_count; list++){
181                 //These values where changed for ease of performing MC, we need to change them back
182                 //FIXME maybe we can make MC and loop filter use the same values or prevent
183                 //the MC code from changing ref_cache and rather use a temporary array.
184                 if(USES_LIST(mb_type,list)){
185                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
186                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
187                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
188                     ref += h->b8_stride;
189                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
190                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
191                 }
192             }
193         }
194     }else{
195         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
196         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
197         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
198         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
199         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
200
201     if(IS_INTRA(mb_type)){
202         int type_mask= h->pps.constrained_intra_pred ? IS_INTRA(-1) : -1;
203         h->topleft_samples_available=
204         h->top_samples_available=
205         h->left_samples_available= 0xFFFF;
206         h->topright_samples_available= 0xEEEA;
207
208         if(!(top_type & type_mask)){
209             h->topleft_samples_available= 0xB3FF;
210             h->top_samples_available= 0x33FF;
211             h->topright_samples_available= 0x26EA;
212         }
213         if(IS_INTERLACED(mb_type) != IS_INTERLACED(left_type[0])){
214             if(IS_INTERLACED(mb_type)){
215                 if(!(left_type[0] & type_mask)){
216                     h->topleft_samples_available&= 0xDFFF;
217                     h->left_samples_available&= 0x5FFF;
218                 }
219                 if(!(left_type[1] & type_mask)){
220                     h->topleft_samples_available&= 0xFF5F;
221                     h->left_samples_available&= 0xFF5F;
222                 }
223             }else{
224                 int left_typei = h->slice_table[left_xy[0] + s->mb_stride ] == h->slice_num
225                                 ? s->current_picture.mb_type[left_xy[0] + s->mb_stride] : 0;
226                 assert(left_xy[0] == left_xy[1]);
227                 if(!((left_typei & type_mask) && (left_type[0] & type_mask))){
228                     h->topleft_samples_available&= 0xDF5F;
229                     h->left_samples_available&= 0x5F5F;
230                 }
231             }
232         }else{
233             if(!(left_type[0] & type_mask)){
234                 h->topleft_samples_available&= 0xDF5F;
235                 h->left_samples_available&= 0x5F5F;
236             }
237         }
238
239         if(!(topleft_type & type_mask))
240             h->topleft_samples_available&= 0x7FFF;
241
242         if(!(topright_type & type_mask))
243             h->topright_samples_available&= 0xFBFF;
244
245         if(IS_INTRA4x4(mb_type)){
246             if(IS_INTRA4x4(top_type)){
247                 h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
248                 h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
249                 h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
250                 h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
251             }else{
252                 int pred;
253                 if(!(top_type & type_mask))
254                     pred= -1;
255                 else{
256                     pred= 2;
257                 }
258                 h->intra4x4_pred_mode_cache[4+8*0]=
259                 h->intra4x4_pred_mode_cache[5+8*0]=
260                 h->intra4x4_pred_mode_cache[6+8*0]=
261                 h->intra4x4_pred_mode_cache[7+8*0]= pred;
262             }
263             for(i=0; i<2; i++){
264                 if(IS_INTRA4x4(left_type[i])){
265                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
266                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
267                 }else{
268                     int pred;
269                     if(!(left_type[i] & type_mask))
270                         pred= -1;
271                     else{
272                         pred= 2;
273                     }
274                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
275                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
276                 }
277             }
278         }
279     }
280     }
281
282
283 /*
284 0 . T T. T T T T
285 1 L . .L . . . .
286 2 L . .L . . . .
287 3 . T TL . . . .
288 4 L . .L . . . .
289 5 L . .. . . . .
290 */
291 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
292     if(top_type){
293         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
294         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
295         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
296         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
297
298         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
299         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
300
301         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
302         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
303
304     }else{
305         h->non_zero_count_cache[4+8*0]=
306         h->non_zero_count_cache[5+8*0]=
307         h->non_zero_count_cache[6+8*0]=
308         h->non_zero_count_cache[7+8*0]=
309
310         h->non_zero_count_cache[1+8*0]=
311         h->non_zero_count_cache[2+8*0]=
312
313         h->non_zero_count_cache[1+8*3]=
314         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
315
316     }
317
318     for (i=0; i<2; i++) {
319         if(left_type[i]){
320             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
321             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
322             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
323             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
324         }else{
325             h->non_zero_count_cache[3+8*1 + 2*8*i]=
326             h->non_zero_count_cache[3+8*2 + 2*8*i]=
327             h->non_zero_count_cache[0+8*1 +   8*i]=
328             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
329         }
330     }
331
332     if( h->pps.cabac ) {
333         // top_cbp
334         if(top_type) {
335             h->top_cbp = h->cbp_table[top_xy];
336         } else if(IS_INTRA(mb_type)) {
337             h->top_cbp = 0x1C0;
338         } else {
339             h->top_cbp = 0;
340         }
341         // left_cbp
342         if (left_type[0]) {
343             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
344         } else if(IS_INTRA(mb_type)) {
345             h->left_cbp = 0x1C0;
346         } else {
347             h->left_cbp = 0;
348         }
349         if (left_type[0]) {
350             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
351         }
352         if (left_type[1]) {
353             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
354         }
355     }
356
357 #if 1
358     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
359         int list;
360         for(list=0; list<h->list_count; list++){
361             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
362                 /*if(!h->mv_cache_clean[list]){
363                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
364                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
365                     h->mv_cache_clean[list]= 1;
366                 }*/
367                 continue;
368             }
369             h->mv_cache_clean[list]= 0;
370
371             if(USES_LIST(top_type, list)){
372                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
373                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
374                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
375                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
376                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
377                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
378                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
379                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
380                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
381                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
382             }else{
383                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
384                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
385                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
386                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
387                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
388             }
389
390             for(i=0; i<2; i++){
391                 int cache_idx = scan8[0] - 1 + i*2*8;
392                 if(USES_LIST(left_type[i], list)){
393                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
394                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
395                     *(uint32_t*)h->mv_cache[list][cache_idx  ]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[0+i*2]];
396                     *(uint32_t*)h->mv_cache[list][cache_idx+8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[1+i*2]];
397                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
398                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
399                 }else{
400                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
401                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
402                     h->ref_cache[list][cache_idx  ]=
403                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
404                 }
405             }
406
407             if(for_deblock || ((IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred) && !FRAME_MBAFF))
408                 continue;
409
410             if(USES_LIST(topleft_type, list)){
411                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
412                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
413                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
414                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
415             }else{
416                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
417                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
418             }
419
420             if(USES_LIST(topright_type, list)){
421                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
422                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
423                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
424                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
425             }else{
426                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
427                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
428             }
429
430             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
431                 continue;
432
433             h->ref_cache[list][scan8[5 ]+1] =
434             h->ref_cache[list][scan8[7 ]+1] =
435             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
436             h->ref_cache[list][scan8[4 ]] =
437             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
438             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
439             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
440             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
441             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
442             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
443
444             if( h->pps.cabac ) {
445                 /* XXX beurk, Load mvd */
446                 if(USES_LIST(top_type, list)){
447                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
448                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
449                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
450                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
451                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
452                 }else{
453                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
454                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
455                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
456                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
457                 }
458                 if(USES_LIST(left_type[0], list)){
459                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
460                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 0*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[0]];
461                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[1]];
462                 }else{
463                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
464                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
465                 }
466                 if(USES_LIST(left_type[1], list)){
467                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
468                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 2*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[2]];
469                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 3*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[3]];
470                 }else{
471                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
472                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
473                 }
474                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
475                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
476                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
477                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
478                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
479
480                 if(h->slice_type_nos == FF_B_TYPE){
481                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
482
483                     if(IS_DIRECT(top_type)){
484                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
485                     }else if(IS_8X8(top_type)){
486                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
487                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
488                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
489                     }else{
490                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
491                     }
492
493                     if(IS_DIRECT(left_type[0]))
494                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
495                     else if(IS_8X8(left_type[0]))
496                         h->direct_cache[scan8[0] - 1 + 0*8]= h->direct_table[h->mb2b8_xy[left_xy[0]] + 1 + h->b8_stride*(left_block[0]>>1)];
497                     else
498                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
499
500                     if(IS_DIRECT(left_type[1]))
501                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
502                     else if(IS_8X8(left_type[1]))
503                         h->direct_cache[scan8[0] - 1 + 2*8]= h->direct_table[h->mb2b8_xy[left_xy[1]] + 1 + h->b8_stride*(left_block[2]>>1)];
504                     else
505                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
506                 }
507             }
508
509             if(FRAME_MBAFF){
510 #define MAP_MVS\
511                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
512                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
513                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
514                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
515                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
516                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
517                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
518                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
519                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
520                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
521                 if(MB_FIELD){
522 #define MAP_F2F(idx, mb_type)\
523                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
524                         h->ref_cache[list][idx] <<= 1;\
525                         h->mv_cache[list][idx][1] /= 2;\
526                         h->mvd_cache[list][idx][1] /= 2;\
527                     }
528                     MAP_MVS
529 #undef MAP_F2F
530                 }else{
531 #define MAP_F2F(idx, mb_type)\
532                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
533                         h->ref_cache[list][idx] >>= 1;\
534                         h->mv_cache[list][idx][1] <<= 1;\
535                         h->mvd_cache[list][idx][1] <<= 1;\
536                     }
537                     MAP_MVS
538 #undef MAP_F2F
539                 }
540             }
541         }
542     }
543 #endif
544
545     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
546 }
547
548 static inline void write_back_intra_pred_mode(H264Context *h){
549     const int mb_xy= h->mb_xy;
550
551     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
552     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
553     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
554     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
555     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
556     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
557     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
558 }
559
560 /**
561  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
562  */
563 static inline int check_intra4x4_pred_mode(H264Context *h){
564     MpegEncContext * const s = &h->s;
565     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
566     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
567     int i;
568
569     if(!(h->top_samples_available&0x8000)){
570         for(i=0; i<4; i++){
571             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
572             if(status<0){
573                 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);
574                 return -1;
575             } else if(status){
576                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
577             }
578         }
579     }
580
581     if((h->left_samples_available&0x8888)!=0x8888){
582         static const int mask[4]={0x8000,0x2000,0x80,0x20};
583         for(i=0; i<4; i++){
584             if(!(h->left_samples_available&mask[i])){
585                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
586                 if(status<0){
587                     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);
588                     return -1;
589                 } else if(status){
590                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
591                 }
592             }
593         }
594     }
595
596     return 0;
597 } //FIXME cleanup like next
598
599 /**
600  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
601  */
602 static inline int check_intra_pred_mode(H264Context *h, int mode){
603     MpegEncContext * const s = &h->s;
604     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
605     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
606
607     if(mode > 6U) {
608         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);
609         return -1;
610     }
611
612     if(!(h->top_samples_available&0x8000)){
613         mode= top[ mode ];
614         if(mode<0){
615             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);
616             return -1;
617         }
618     }
619
620     if((h->left_samples_available&0x8080) != 0x8080){
621         mode= left[ mode ];
622         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
623             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
624         }
625         if(mode<0){
626             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);
627             return -1;
628         }
629     }
630
631     return mode;
632 }
633
634 /**
635  * gets the predicted intra4x4 prediction mode.
636  */
637 static inline int pred_intra_mode(H264Context *h, int n){
638     const int index8= scan8[n];
639     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
640     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
641     const int min= FFMIN(left, top);
642
643     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
644
645     if(min<0) return DC_PRED;
646     else      return min;
647 }
648
649 static inline void write_back_non_zero_count(H264Context *h){
650     const int mb_xy= h->mb_xy;
651
652     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
653     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
654     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
655     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
656     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
657     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
658     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
659
660     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
661     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
662     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
663
664     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
665     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
666     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
667 }
668
669 /**
670  * gets the predicted number of non-zero coefficients.
671  * @param n block index
672  */
673 static inline int pred_non_zero_count(H264Context *h, int n){
674     const int index8= scan8[n];
675     const int left= h->non_zero_count_cache[index8 - 1];
676     const int top = h->non_zero_count_cache[index8 - 8];
677     int i= left + top;
678
679     if(i<64) i= (i+1)>>1;
680
681     tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);
682
683     return i&31;
684 }
685
686 static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
687     const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
688     MpegEncContext *s = &h->s;
689
690     /* there is no consistent mapping of mvs to neighboring locations that will
691      * make mbaff happy, so we can't move all this logic to fill_caches */
692     if(FRAME_MBAFF){
693         const uint32_t *mb_types = s->current_picture_ptr->mb_type;
694         const int16_t *mv;
695         *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
696         *C = h->mv_cache[list][scan8[0]-2];
697
698         if(!MB_FIELD
699            && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
700             int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
701             if(IS_INTERLACED(mb_types[topright_xy])){
702 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
703                 const int x4 = X4, y4 = Y4;\
704                 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
705                 if(!USES_LIST(mb_type,list))\
706                     return LIST_NOT_USED;\
707                 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
708                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
709                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
710                 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
711
712                 SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
713             }
714         }
715         if(topright_ref == PART_NOT_AVAILABLE
716            && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
717            && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
718             if(!MB_FIELD
719                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
720                 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
721             }
722             if(MB_FIELD
723                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
724                && i >= scan8[0]+8){
725                 // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
726                 SET_DIAG_MV(/2, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
727             }
728         }
729 #undef SET_DIAG_MV
730     }
731
732     if(topright_ref != PART_NOT_AVAILABLE){
733         *C= h->mv_cache[list][ i - 8 + part_width ];
734         return topright_ref;
735     }else{
736         tprintf(s->avctx, "topright MV not available\n");
737
738         *C= h->mv_cache[list][ i - 8 - 1 ];
739         return h->ref_cache[list][ i - 8 - 1 ];
740     }
741 }
742
743 /**
744  * gets the predicted MV.
745  * @param n the block index
746  * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
747  * @param mx the x component of the predicted motion vector
748  * @param my the y component of the predicted motion vector
749  */
750 static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
751     const int index8= scan8[n];
752     const int top_ref=      h->ref_cache[list][ index8 - 8 ];
753     const int left_ref=     h->ref_cache[list][ index8 - 1 ];
754     const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
755     const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
756     const int16_t * C;
757     int diagonal_ref, match_count;
758
759     assert(part_width==1 || part_width==2 || part_width==4);
760
761 /* mv_cache
762   B . . A T T T T
763   U . . L . . , .
764   U . . L . . . .
765   U . . L . . , .
766   . . . L . . . .
767 */
768
769     diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
770     match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
771     tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
772     if(match_count > 1){ //most common
773         *mx= mid_pred(A[0], B[0], C[0]);
774         *my= mid_pred(A[1], B[1], C[1]);
775     }else if(match_count==1){
776         if(left_ref==ref){
777             *mx= A[0];
778             *my= A[1];
779         }else if(top_ref==ref){
780             *mx= B[0];
781             *my= B[1];
782         }else{
783             *mx= C[0];
784             *my= C[1];
785         }
786     }else{
787         if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
788             *mx= A[0];
789             *my= A[1];
790         }else{
791             *mx= mid_pred(A[0], B[0], C[0]);
792             *my= mid_pred(A[1], B[1], C[1]);
793         }
794     }
795
796     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);
797 }
798
799 /**
800  * gets the directionally predicted 16x8 MV.
801  * @param n the block index
802  * @param mx the x component of the predicted motion vector
803  * @param my the y component of the predicted motion vector
804  */
805 static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
806     if(n==0){
807         const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
808         const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
809
810         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);
811
812         if(top_ref == ref){
813             *mx= B[0];
814             *my= B[1];
815             return;
816         }
817     }else{
818         const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
819         const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];
820
821         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);
822
823         if(left_ref == ref){
824             *mx= A[0];
825             *my= A[1];
826             return;
827         }
828     }
829
830     //RARE
831     pred_motion(h, n, 4, list, ref, mx, my);
832 }
833
834 /**
835  * gets the directionally predicted 8x16 MV.
836  * @param n the block index
837  * @param mx the x component of the predicted motion vector
838  * @param my the y component of the predicted motion vector
839  */
840 static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
841     if(n==0){
842         const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
843         const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];
844
845         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);
846
847         if(left_ref == ref){
848             *mx= A[0];
849             *my= A[1];
850             return;
851         }
852     }else{
853         const int16_t * C;
854         int diagonal_ref;
855
856         diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);
857
858         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);
859
860         if(diagonal_ref == ref){
861             *mx= C[0];
862             *my= C[1];
863             return;
864         }
865     }
866
867     //RARE
868     pred_motion(h, n, 2, list, ref, mx, my);
869 }
870
871 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
872     const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
873     const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];
874
875     tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);
876
877     if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
878        || !( top_ref | *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ])
879        || !(left_ref | *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ])){
880
881         *mx = *my = 0;
882         return;
883     }
884
885     pred_motion(h, 0, 4, 0, 0, mx, my);
886
887     return;
888 }
889
890 static int get_scale_factor(H264Context * const h, int poc, int poc1, int i){
891     int poc0 = h->ref_list[0][i].poc;
892     int td = av_clip(poc1 - poc0, -128, 127);
893     if(td == 0 || h->ref_list[0][i].long_ref){
894         return 256;
895     }else{
896         int tb = av_clip(poc - poc0, -128, 127);
897         int tx = (16384 + (FFABS(td) >> 1)) / td;
898         return av_clip((tb*tx + 32) >> 6, -1024, 1023);
899     }
900 }
901
902 static inline void direct_dist_scale_factor(H264Context * const h){
903     MpegEncContext * const s = &h->s;
904     const int poc = h->s.current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
905     const int poc1 = h->ref_list[1][0].poc;
906     int i, field;
907     for(field=0; field<2; field++){
908         const int poc  = h->s.current_picture_ptr->field_poc[field];
909         const int poc1 = h->ref_list[1][0].field_poc[field];
910         for(i=0; i < 2*h->ref_count[0]; i++)
911             h->dist_scale_factor_field[field][i^field] = get_scale_factor(h, poc, poc1, i+16);
912     }
913
914     for(i=0; i<h->ref_count[0]; i++){
915         h->dist_scale_factor[i] = get_scale_factor(h, poc, poc1, i);
916     }
917 }
918
919 static void fill_colmap(H264Context *h, int map[2][16+32], int list, int field, int colfield, int mbafi){
920     MpegEncContext * const s = &h->s;
921     Picture * const ref1 = &h->ref_list[1][0];
922     int j, old_ref, rfield;
923     int start= mbafi ? 16                      : 0;
924     int end  = mbafi ? 16+2*h->ref_count[list] : h->ref_count[list];
925     int interl= mbafi || s->picture_structure != PICT_FRAME;
926
927     /* bogus; fills in for missing frames */
928     memset(map[list], 0, sizeof(map[list]));
929
930     for(rfield=0; rfield<2; rfield++){
931         for(old_ref=0; old_ref<ref1->ref_count[colfield][list]; old_ref++){
932             int poc = ref1->ref_poc[colfield][list][old_ref];
933
934             if     (!interl)
935                 poc |= 3;
936             else if( interl && (poc&3) == 3) //FIXME store all MBAFF references so this isnt needed
937                 poc= (poc&~3) + rfield + 1;
938
939             for(j=start; j<end; j++){
940                 if(4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3) == poc){
941                     int cur_ref= mbafi ? (j-16)^field : j;
942                     map[list][2*old_ref + (rfield^field) + 16] = cur_ref;
943                     if(rfield == field)
944                         map[list][old_ref] = cur_ref;
945                     break;
946                 }
947             }
948         }
949     }
950 }
951
952 static inline void direct_ref_list_init(H264Context * const h){
953     MpegEncContext * const s = &h->s;
954     Picture * const ref1 = &h->ref_list[1][0];
955     Picture * const cur = s->current_picture_ptr;
956     int list, j, field;
957     int sidx= (s->picture_structure&1)^1;
958     int ref1sidx= (ref1->reference&1)^1;
959
960     for(list=0; list<2; list++){
961         cur->ref_count[sidx][list] = h->ref_count[list];
962         for(j=0; j<h->ref_count[list]; j++)
963             cur->ref_poc[sidx][list][j] = 4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3);
964     }
965
966     if(s->picture_structure == PICT_FRAME){
967         memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0]));
968         memcpy(cur->ref_poc  [1], cur->ref_poc  [0], sizeof(cur->ref_poc  [0]));
969     }
970
971     cur->mbaff= FRAME_MBAFF;
972
973     if(cur->pict_type != FF_B_TYPE || h->direct_spatial_mv_pred)
974         return;
975
976     for(list=0; list<2; list++){
977         fill_colmap(h, h->map_col_to_list0, list, sidx, ref1sidx, 0);
978         for(field=0; field<2; field++)
979             fill_colmap(h, h->map_col_to_list0_field[field], list, field, field, 1);
980     }
981 }
982
983 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
984     MpegEncContext * const s = &h->s;
985     int b8_stride = h->b8_stride;
986     int b4_stride = h->b_stride;
987     int mb_xy = h->mb_xy;
988     int mb_type_col[2];
989     const int16_t (*l1mv0)[2], (*l1mv1)[2];
990     const int8_t *l1ref0, *l1ref1;
991     const int is_b8x8 = IS_8X8(*mb_type);
992     unsigned int sub_mb_type;
993     int i8, i4;
994
995 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
996
997     if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL
998         if(!IS_INTERLACED(*mb_type)){                    //     AFR/FR    -> AFL/FL
999             int cur_poc = s->current_picture_ptr->poc;
1000             int *col_poc = h->ref_list[1]->field_poc;
1001             int col_parity = FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc);
1002             mb_xy= s->mb_x + ((s->mb_y&~1) + col_parity)*s->mb_stride;
1003             b8_stride = 0;
1004         }else if(!(s->picture_structure & h->ref_list[1][0].reference) && !h->ref_list[1][0].mbaff){// FL -> FL & differ parity
1005             int fieldoff= 2*(h->ref_list[1][0].reference)-3;
1006             mb_xy += s->mb_stride*fieldoff;
1007         }
1008         goto single_col;
1009     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
1010         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
1011             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
1012             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
1013             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
1014             b8_stride *= 3;
1015             b4_stride *= 6;
1016             //FIXME IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag
1017             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
1018                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
1019                 && !is_b8x8){
1020                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1021                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
1022             }else{
1023                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1024                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1025             }
1026         }else{                                           //     AFR/FR    -> AFR/FR
1027 single_col:
1028             mb_type_col[0] =
1029             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
1030             if(IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag){
1031                 /* FIXME save sub mb types from previous frames (or derive from MVs)
1032                 * so we know exactly what block size to use */
1033                 sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
1034                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1035             }else if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
1036                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1037                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
1038             }else{
1039                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1040                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1041             }
1042         }
1043     }
1044
1045     l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
1046     l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
1047     l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
1048     l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
1049     if(!b8_stride){
1050         if(s->mb_y&1){
1051             l1ref0 += h->b8_stride;
1052             l1ref1 += h->b8_stride;
1053             l1mv0  +=  2*b4_stride;
1054             l1mv1  +=  2*b4_stride;
1055         }
1056     }
1057
1058     if(h->direct_spatial_mv_pred){
1059         int ref[2];
1060         int mv[2][2];
1061         int list;
1062
1063         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
1064
1065         /* ref = min(neighbors) */
1066         for(list=0; list<2; list++){
1067             int refa = h->ref_cache[list][scan8[0] - 1];
1068             int refb = h->ref_cache[list][scan8[0] - 8];
1069             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1070             if(refc == PART_NOT_AVAILABLE)
1071                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1072             ref[list] = FFMIN3((unsigned)refa, (unsigned)refb, (unsigned)refc);
1073             if(ref[list] < 0)
1074                 ref[list] = -1;
1075         }
1076
1077         if(ref[0] < 0 && ref[1] < 0){
1078             ref[0] = ref[1] = 0;
1079             mv[0][0] = mv[0][1] =
1080             mv[1][0] = mv[1][1] = 0;
1081         }else{
1082             for(list=0; list<2; list++){
1083                 if(ref[list] >= 0)
1084                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1085                 else
1086                     mv[list][0] = mv[list][1] = 0;
1087             }
1088         }
1089
1090         if(ref[1] < 0){
1091             if(!is_b8x8)
1092                 *mb_type &= ~MB_TYPE_L1;
1093             sub_mb_type &= ~MB_TYPE_L1;
1094         }else if(ref[0] < 0){
1095             if(!is_b8x8)
1096                 *mb_type &= ~MB_TYPE_L0;
1097             sub_mb_type &= ~MB_TYPE_L0;
1098         }
1099
1100         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1101             for(i8=0; i8<4; i8++){
1102                 int x8 = i8&1;
1103                 int y8 = i8>>1;
1104                 int xy8 = x8+y8*b8_stride;
1105                 int xy4 = 3*x8+y8*b4_stride;
1106                 int a=0, b=0;
1107
1108                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1109                     continue;
1110                 h->sub_mb_type[i8] = sub_mb_type;
1111
1112                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1113                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1114                 if(!IS_INTRA(mb_type_col[y8])
1115                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
1116                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
1117                     if(ref[0] > 0)
1118                         a= pack16to32(mv[0][0],mv[0][1]);
1119                     if(ref[1] > 0)
1120                         b= pack16to32(mv[1][0],mv[1][1]);
1121                 }else{
1122                     a= pack16to32(mv[0][0],mv[0][1]);
1123                     b= pack16to32(mv[1][0],mv[1][1]);
1124                 }
1125                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
1126                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
1127             }
1128         }else if(IS_16X16(*mb_type)){
1129             int a=0, b=0;
1130
1131             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1132             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1133             if(!IS_INTRA(mb_type_col[0])
1134                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1135                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1136                        && (h->x264_build>33 || !h->x264_build)))){
1137                 if(ref[0] > 0)
1138                     a= pack16to32(mv[0][0],mv[0][1]);
1139                 if(ref[1] > 0)
1140                     b= pack16to32(mv[1][0],mv[1][1]);
1141             }else{
1142                 a= pack16to32(mv[0][0],mv[0][1]);
1143                 b= pack16to32(mv[1][0],mv[1][1]);
1144             }
1145             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1146             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1147         }else{
1148             for(i8=0; i8<4; i8++){
1149                 const int x8 = i8&1;
1150                 const int y8 = i8>>1;
1151
1152                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1153                     continue;
1154                 h->sub_mb_type[i8] = sub_mb_type;
1155
1156                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1157                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1158                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1159                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1160
1161                 /* col_zero_flag */
1162                 if(!IS_INTRA(mb_type_col[0]) && (   l1ref0[x8 + y8*b8_stride] == 0
1163                                               || (l1ref0[x8 + y8*b8_stride] < 0 && l1ref1[x8 + y8*b8_stride] == 0
1164                                                   && (h->x264_build>33 || !h->x264_build)))){
1165                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*b8_stride] == 0 ? l1mv0 : l1mv1;
1166                     if(IS_SUB_8X8(sub_mb_type)){
1167                         const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1168                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1169                             if(ref[0] == 0)
1170                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1171                             if(ref[1] == 0)
1172                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1173                         }
1174                     }else
1175                     for(i4=0; i4<4; i4++){
1176                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1177                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1178                             if(ref[0] == 0)
1179                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1180                             if(ref[1] == 0)
1181                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1182                         }
1183                     }
1184                 }
1185             }
1186         }
1187     }else{ /* direct temporal mv pred */
1188         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1189         const int *dist_scale_factor = h->dist_scale_factor;
1190         int ref_offset= 0;
1191
1192         if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
1193             map_col_to_list0[0] = h->map_col_to_list0_field[s->mb_y&1][0];
1194             map_col_to_list0[1] = h->map_col_to_list0_field[s->mb_y&1][1];
1195             dist_scale_factor   =h->dist_scale_factor_field[s->mb_y&1];
1196         }
1197         if(h->ref_list[1][0].mbaff && IS_INTERLACED(mb_type_col[0]))
1198             ref_offset += 16;
1199
1200         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1201             /* FIXME assumes direct_8x8_inference == 1 */
1202             int y_shift  = 2*!IS_INTERLACED(*mb_type);
1203
1204             for(i8=0; i8<4; i8++){
1205                 const int x8 = i8&1;
1206                 const int y8 = i8>>1;
1207                 int ref0, scale;
1208                 const int16_t (*l1mv)[2]= l1mv0;
1209
1210                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1211                     continue;
1212                 h->sub_mb_type[i8] = sub_mb_type;
1213
1214                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1215                 if(IS_INTRA(mb_type_col[y8])){
1216                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1217                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1218                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1219                     continue;
1220                 }
1221
1222                 ref0 = l1ref0[x8 + y8*b8_stride];
1223                 if(ref0 >= 0)
1224                     ref0 = map_col_to_list0[0][ref0 + ref_offset];
1225                 else{
1226                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
1227                     l1mv= l1mv1;
1228                 }
1229                 scale = dist_scale_factor[ref0];
1230                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1231
1232                 {
1233                     const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride];
1234                     int my_col = (mv_col[1]<<y_shift)/2;
1235                     int mx = (scale * mv_col[0] + 128) >> 8;
1236                     int my = (scale * my_col + 128) >> 8;
1237                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1238                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1239                 }
1240             }
1241             return;
1242         }
1243
1244         /* one-to-one mv scaling */
1245
1246         if(IS_16X16(*mb_type)){
1247             int ref, mv0, mv1;
1248
1249             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1250             if(IS_INTRA(mb_type_col[0])){
1251                 ref=mv0=mv1=0;
1252             }else{
1253                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset]
1254                                                 : map_col_to_list0[1][l1ref1[0] + ref_offset];
1255                 const int scale = dist_scale_factor[ref0];
1256                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1257                 int mv_l0[2];
1258                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1259                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1260                 ref= ref0;
1261                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1262                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1263             }
1264             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1265             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1266             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1267         }else{
1268             for(i8=0; i8<4; i8++){
1269                 const int x8 = i8&1;
1270                 const int y8 = i8>>1;
1271                 int ref0, scale;
1272                 const int16_t (*l1mv)[2]= l1mv0;
1273
1274                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1275                     continue;
1276                 h->sub_mb_type[i8] = sub_mb_type;
1277                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1278                 if(IS_INTRA(mb_type_col[0])){
1279                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1280                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1281                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1282                     continue;
1283                 }
1284
1285                 ref0 = l1ref0[x8 + y8*b8_stride] + ref_offset;
1286                 if(ref0 >= 0)
1287                     ref0 = map_col_to_list0[0][ref0];
1288                 else{
1289                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
1290                     l1mv= l1mv1;
1291                 }
1292                 scale = dist_scale_factor[ref0];
1293
1294                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1295                 if(IS_SUB_8X8(sub_mb_type)){
1296                     const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1297                     int mx = (scale * mv_col[0] + 128) >> 8;
1298                     int my = (scale * mv_col[1] + 128) >> 8;
1299                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1300                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1301                 }else
1302                 for(i4=0; i4<4; i4++){
1303                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1304                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1305                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1306                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1307                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1308                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1309                 }
1310             }
1311         }
1312     }
1313 }
1314
1315 static inline void write_back_motion(H264Context *h, int mb_type){
1316     MpegEncContext * const s = &h->s;
1317     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1318     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1319     int list;
1320
1321     if(!USES_LIST(mb_type, 0))
1322         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1323
1324     for(list=0; list<h->list_count; list++){
1325         int y;
1326         if(!USES_LIST(mb_type, list))
1327             continue;
1328
1329         for(y=0; y<4; y++){
1330             *(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];
1331             *(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];
1332         }
1333         if( h->pps.cabac ) {
1334             if(IS_SKIP(mb_type))
1335                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1336             else
1337             for(y=0; y<4; y++){
1338                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1339                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1340             }
1341         }
1342
1343         {
1344             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1345             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1346             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1347             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1348             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1349         }
1350     }
1351
1352     if(h->slice_type_nos == FF_B_TYPE && h->pps.cabac){
1353         if(IS_8X8(mb_type)){
1354             uint8_t *direct_table = &h->direct_table[b8_xy];
1355             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1356             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1357             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1358         }
1359     }
1360 }
1361
1362 /**
1363  * Decodes a network abstraction layer unit.
1364  * @param consumed is the number of bytes used as input
1365  * @param length is the length of the array
1366  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1367  * @returns decoded bytes, might be src+1 if no escapes
1368  */
1369 static const uint8_t *decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
1370     int i, si, di;
1371     uint8_t *dst;
1372     int bufidx;
1373
1374 //    src[0]&0x80;                //forbidden bit
1375     h->nal_ref_idc= src[0]>>5;
1376     h->nal_unit_type= src[0]&0x1F;
1377
1378     src++; length--;
1379 #if 0
1380     for(i=0; i<length; i++)
1381         printf("%2X ", src[i]);
1382 #endif
1383
1384 #if HAVE_FAST_UNALIGNED
1385 # if HAVE_FAST_64BIT
1386 #   define RS 7
1387     for(i=0; i+1<length; i+=9){
1388         if(!((~*(uint64_t*)(src+i) & (*(uint64_t*)(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
1389 # else
1390 #   define RS 3
1391     for(i=0; i+1<length; i+=5){
1392         if(!((~*(uint32_t*)(src+i) & (*(uint32_t*)(src+i) - 0x01000101U)) & 0x80008080U))
1393 # endif
1394             continue;
1395         if(i>0 && !src[i]) i--;
1396         while(src[i]) i++;
1397 #else
1398 #   define RS 0
1399     for(i=0; i+1<length; i+=2){
1400         if(src[i]) continue;
1401         if(i>0 && src[i-1]==0) i--;
1402 #endif
1403         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1404             if(src[i+2]!=3){
1405                 /* startcode, so we must be past the end */
1406                 length=i;
1407             }
1408             break;
1409         }
1410         i-= RS;
1411     }
1412
1413     if(i>=length-1){ //no escaped 0
1414         *dst_length= length;
1415         *consumed= length+1; //+1 for the header
1416         return src;
1417     }
1418
1419     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1420     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
1421     dst= h->rbsp_buffer[bufidx];
1422
1423     if (dst == NULL){
1424         return NULL;
1425     }
1426
1427 //printf("decoding esc\n");
1428     memcpy(dst, src, i);
1429     si=di=i;
1430     while(si+2<length){
1431         //remove escapes (very rare 1:2^22)
1432         if(src[si+2]>3){
1433             dst[di++]= src[si++];
1434             dst[di++]= src[si++];
1435         }else if(src[si]==0 && src[si+1]==0){
1436             if(src[si+2]==3){ //escape
1437                 dst[di++]= 0;
1438                 dst[di++]= 0;
1439                 si+=3;
1440                 continue;
1441             }else //next start code
1442                 goto nsc;
1443         }
1444
1445         dst[di++]= src[si++];
1446     }
1447     while(si<length)
1448         dst[di++]= src[si++];
1449 nsc:
1450
1451     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
1452
1453     *dst_length= di;
1454     *consumed= si + 1;//+1 for the header
1455 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1456     return dst;
1457 }
1458
1459 /**
1460  * identifies the exact end of the bitstream
1461  * @return the length of the trailing, or 0 if damaged
1462  */
1463 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src){
1464     int v= *src;
1465     int r;
1466
1467     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1468
1469     for(r=1; r<9; r++){
1470         if(v&1) return r;
1471         v>>=1;
1472     }
1473     return 0;
1474 }
1475
1476 /**
1477  * IDCT transforms the 16 dc values and dequantizes them.
1478  * @param qp quantization parameter
1479  */
1480 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1481 #define stride 16
1482     int i;
1483     int temp[16]; //FIXME check if this is a good idea
1484     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1485     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1486
1487 //memset(block, 64, 2*256);
1488 //return;
1489     for(i=0; i<4; i++){
1490         const int offset= y_offset[i];
1491         const int z0= block[offset+stride*0] + block[offset+stride*4];
1492         const int z1= block[offset+stride*0] - block[offset+stride*4];
1493         const int z2= block[offset+stride*1] - block[offset+stride*5];
1494         const int z3= block[offset+stride*1] + block[offset+stride*5];
1495
1496         temp[4*i+0]= z0+z3;
1497         temp[4*i+1]= z1+z2;
1498         temp[4*i+2]= z1-z2;
1499         temp[4*i+3]= z0-z3;
1500     }
1501
1502     for(i=0; i<4; i++){
1503         const int offset= x_offset[i];
1504         const int z0= temp[4*0+i] + temp[4*2+i];
1505         const int z1= temp[4*0+i] - temp[4*2+i];
1506         const int z2= temp[4*1+i] - temp[4*3+i];
1507         const int z3= temp[4*1+i] + temp[4*3+i];
1508
1509         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
1510         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1511         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1512         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1513     }
1514 }
1515
1516 #if 0
1517 /**
1518  * DCT transforms the 16 dc values.
1519  * @param qp quantization parameter ??? FIXME
1520  */
1521 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1522 //    const int qmul= dequant_coeff[qp][0];
1523     int i;
1524     int temp[16]; //FIXME check if this is a good idea
1525     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1526     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1527
1528     for(i=0; i<4; i++){
1529         const int offset= y_offset[i];
1530         const int z0= block[offset+stride*0] + block[offset+stride*4];
1531         const int z1= block[offset+stride*0] - block[offset+stride*4];
1532         const int z2= block[offset+stride*1] - block[offset+stride*5];
1533         const int z3= block[offset+stride*1] + block[offset+stride*5];
1534
1535         temp[4*i+0]= z0+z3;
1536         temp[4*i+1]= z1+z2;
1537         temp[4*i+2]= z1-z2;
1538         temp[4*i+3]= z0-z3;
1539     }
1540
1541     for(i=0; i<4; i++){
1542         const int offset= x_offset[i];
1543         const int z0= temp[4*0+i] + temp[4*2+i];
1544         const int z1= temp[4*0+i] - temp[4*2+i];
1545         const int z2= temp[4*1+i] - temp[4*3+i];
1546         const int z3= temp[4*1+i] + temp[4*3+i];
1547
1548         block[stride*0 +offset]= (z0 + z3)>>1;
1549         block[stride*2 +offset]= (z1 + z2)>>1;
1550         block[stride*8 +offset]= (z1 - z2)>>1;
1551         block[stride*10+offset]= (z0 - z3)>>1;
1552     }
1553 }
1554 #endif
1555
1556 #undef xStride
1557 #undef stride
1558
1559 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1560     const int stride= 16*2;
1561     const int xStride= 16;
1562     int a,b,c,d,e;
1563
1564     a= block[stride*0 + xStride*0];
1565     b= block[stride*0 + xStride*1];
1566     c= block[stride*1 + xStride*0];
1567     d= block[stride*1 + xStride*1];
1568
1569     e= a-b;
1570     a= a+b;
1571     b= c-d;
1572     c= c+d;
1573
1574     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1575     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1576     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1577     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1578 }
1579
1580 #if 0
1581 static void chroma_dc_dct_c(DCTELEM *block){
1582     const int stride= 16*2;
1583     const int xStride= 16;
1584     int a,b,c,d,e;
1585
1586     a= block[stride*0 + xStride*0];
1587     b= block[stride*0 + xStride*1];
1588     c= block[stride*1 + xStride*0];
1589     d= block[stride*1 + xStride*1];
1590
1591     e= a-b;
1592     a= a+b;
1593     b= c-d;
1594     c= c+d;
1595
1596     block[stride*0 + xStride*0]= (a+c);
1597     block[stride*0 + xStride*1]= (e+b);
1598     block[stride*1 + xStride*0]= (a-c);
1599     block[stride*1 + xStride*1]= (e-b);
1600 }
1601 #endif
1602
1603 /**
1604  * gets the chroma qp.
1605  */
1606 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1607     return h->pps.chroma_qp_table[t][qscale];
1608 }
1609
1610 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1611                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1612                            int src_x_offset, int src_y_offset,
1613                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1614     MpegEncContext * const s = &h->s;
1615     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1616     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1617     const int luma_xy= (mx&3) + ((my&3)<<2);
1618     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1619     uint8_t * src_cb, * src_cr;
1620     int extra_width= h->emu_edge_width;
1621     int extra_height= h->emu_edge_height;
1622     int emu=0;
1623     const int full_mx= mx>>2;
1624     const int full_my= my>>2;
1625     const int pic_width  = 16*s->mb_width;
1626     const int pic_height = 16*s->mb_height >> MB_FIELD;
1627
1628     if(mx&7) extra_width -= 3;
1629     if(my&7) extra_height -= 3;
1630
1631     if(   full_mx < 0-extra_width
1632        || full_my < 0-extra_height
1633        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1634        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1635         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);
1636             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1637         emu=1;
1638     }
1639
1640     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1641     if(!square){
1642         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1643     }
1644
1645     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1646
1647     if(MB_FIELD){
1648         // chroma offset when predicting from a field of opposite parity
1649         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1650         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1651     }
1652     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1653     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1654
1655     if(emu){
1656         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);
1657             src_cb= s->edge_emu_buffer;
1658     }
1659     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1660
1661     if(emu){
1662         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);
1663             src_cr= s->edge_emu_buffer;
1664     }
1665     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1666 }
1667
1668 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1669                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1670                            int x_offset, int y_offset,
1671                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1672                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1673                            int list0, int list1){
1674     MpegEncContext * const s = &h->s;
1675     qpel_mc_func *qpix_op=  qpix_put;
1676     h264_chroma_mc_func chroma_op= chroma_put;
1677
1678     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1679     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1680     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1681     x_offset += 8*s->mb_x;
1682     y_offset += 8*(s->mb_y >> MB_FIELD);
1683
1684     if(list0){
1685         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1686         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1687                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1688                            qpix_op, chroma_op);
1689
1690         qpix_op=  qpix_avg;
1691         chroma_op= chroma_avg;
1692     }
1693
1694     if(list1){
1695         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1696         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1697                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1698                            qpix_op, chroma_op);
1699     }
1700 }
1701
1702 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1703                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1704                            int x_offset, int y_offset,
1705                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1706                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1707                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1708                            int list0, int list1){
1709     MpegEncContext * const s = &h->s;
1710
1711     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1712     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1713     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1714     x_offset += 8*s->mb_x;
1715     y_offset += 8*(s->mb_y >> MB_FIELD);
1716
1717     if(list0 && list1){
1718         /* don't optimize for luma-only case, since B-frames usually
1719          * use implicit weights => chroma too. */
1720         uint8_t *tmp_cb = s->obmc_scratchpad;
1721         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1722         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1723         int refn0 = h->ref_cache[0][ scan8[n] ];
1724         int refn1 = h->ref_cache[1][ scan8[n] ];
1725
1726         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1727                     dest_y, dest_cb, dest_cr,
1728                     x_offset, y_offset, qpix_put, chroma_put);
1729         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1730                     tmp_y, tmp_cb, tmp_cr,
1731                     x_offset, y_offset, qpix_put, chroma_put);
1732
1733         if(h->use_weight == 2){
1734             int weight0 = h->implicit_weight[refn0][refn1];
1735             int weight1 = 64 - weight0;
1736             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1737             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1738             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1739         }else{
1740             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1741                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1742                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1743             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1744                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1745                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1746             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1747                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1748                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1749         }
1750     }else{
1751         int list = list1 ? 1 : 0;
1752         int refn = h->ref_cache[list][ scan8[n] ];
1753         Picture *ref= &h->ref_list[list][refn];
1754         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1755                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1756                     qpix_put, chroma_put);
1757
1758         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1759                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1760         if(h->use_weight_chroma){
1761             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1762                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1763             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1764                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1765         }
1766     }
1767 }
1768
1769 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1770                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1771                            int x_offset, int y_offset,
1772                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1773                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1774                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1775                            int list0, int list1){
1776     if((h->use_weight==2 && list0 && list1
1777         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1778        || h->use_weight==1)
1779         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1780                          x_offset, y_offset, qpix_put, chroma_put,
1781                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1782     else
1783         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1784                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1785 }
1786
1787 static inline void prefetch_motion(H264Context *h, int list){
1788     /* fetch pixels for estimated mv 4 macroblocks ahead
1789      * optimized for 64byte cache lines */
1790     MpegEncContext * const s = &h->s;
1791     const int refn = h->ref_cache[list][scan8[0]];
1792     if(refn >= 0){
1793         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1794         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1795         uint8_t **src= h->ref_list[list][refn].data;
1796         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1797         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1798         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1799         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1800     }
1801 }
1802
1803 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1804                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1805                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1806                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1807     MpegEncContext * const s = &h->s;
1808     const int mb_xy= h->mb_xy;
1809     const int mb_type= s->current_picture.mb_type[mb_xy];
1810
1811     assert(IS_INTER(mb_type));
1812
1813     prefetch_motion(h, 0);
1814
1815     if(IS_16X16(mb_type)){
1816         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1817                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1818                 &weight_op[0], &weight_avg[0],
1819                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1820     }else if(IS_16X8(mb_type)){
1821         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1822                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1823                 &weight_op[1], &weight_avg[1],
1824                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1825         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1826                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1827                 &weight_op[1], &weight_avg[1],
1828                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1829     }else if(IS_8X16(mb_type)){
1830         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1831                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1832                 &weight_op[2], &weight_avg[2],
1833                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1834         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1835                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1836                 &weight_op[2], &weight_avg[2],
1837                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1838     }else{
1839         int i;
1840
1841         assert(IS_8X8(mb_type));
1842
1843         for(i=0; i<4; i++){
1844             const int sub_mb_type= h->sub_mb_type[i];
1845             const int n= 4*i;
1846             int x_offset= (i&1)<<2;
1847             int y_offset= (i&2)<<1;
1848
1849             if(IS_SUB_8X8(sub_mb_type)){
1850                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1851                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1852                     &weight_op[3], &weight_avg[3],
1853                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1854             }else if(IS_SUB_8X4(sub_mb_type)){
1855                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1856                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1857                     &weight_op[4], &weight_avg[4],
1858                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1859                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1860                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1861                     &weight_op[4], &weight_avg[4],
1862                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1863             }else if(IS_SUB_4X8(sub_mb_type)){
1864                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1865                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1866                     &weight_op[5], &weight_avg[5],
1867                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1868                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1869                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1870                     &weight_op[5], &weight_avg[5],
1871                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1872             }else{
1873                 int j;
1874                 assert(IS_SUB_4X4(sub_mb_type));
1875                 for(j=0; j<4; j++){
1876                     int sub_x_offset= x_offset + 2*(j&1);
1877                     int sub_y_offset= y_offset +   (j&2);
1878                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1879                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1880                         &weight_op[6], &weight_avg[6],
1881                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1882                 }
1883             }
1884         }
1885     }
1886
1887     prefetch_motion(h, 1);
1888 }
1889
1890 static av_cold void init_cavlc_level_tab(void){
1891     int suffix_length, mask;
1892     unsigned int i;
1893
1894     for(suffix_length=0; suffix_length<7; suffix_length++){
1895         for(i=0; i<(1<<LEVEL_TAB_BITS); i++){
1896             int prefix= LEVEL_TAB_BITS - av_log2(2*i);
1897             int level_code= (prefix<<suffix_length) + (i>>(LEVEL_TAB_BITS-prefix-1-suffix_length)) - (1<<suffix_length);
1898
1899             mask= -(level_code&1);
1900             level_code= (((2+level_code)>>1) ^ mask) - mask;
1901             if(prefix + 1 + suffix_length <= LEVEL_TAB_BITS){
1902                 cavlc_level_tab[suffix_length][i][0]= level_code;
1903                 cavlc_level_tab[suffix_length][i][1]= prefix + 1 + suffix_length;
1904             }else if(prefix + 1 <= LEVEL_TAB_BITS){
1905                 cavlc_level_tab[suffix_length][i][0]= prefix+100;
1906                 cavlc_level_tab[suffix_length][i][1]= prefix + 1;
1907             }else{
1908                 cavlc_level_tab[suffix_length][i][0]= LEVEL_TAB_BITS+100;
1909                 cavlc_level_tab[suffix_length][i][1]= LEVEL_TAB_BITS;
1910             }
1911         }
1912     }
1913 }
1914
1915 static av_cold void decode_init_vlc(void){
1916     static int done = 0;
1917
1918     if (!done) {
1919         int i;
1920         int offset;
1921         done = 1;
1922
1923         chroma_dc_coeff_token_vlc.table = chroma_dc_coeff_token_vlc_table;
1924         chroma_dc_coeff_token_vlc.table_allocated = chroma_dc_coeff_token_vlc_table_size;
1925         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1926                  &chroma_dc_coeff_token_len [0], 1, 1,
1927                  &chroma_dc_coeff_token_bits[0], 1, 1,
1928                  INIT_VLC_USE_NEW_STATIC);
1929
1930         offset = 0;
1931         for(i=0; i<4; i++){
1932             coeff_token_vlc[i].table = coeff_token_vlc_tables+offset;
1933             coeff_token_vlc[i].table_allocated = coeff_token_vlc_tables_size[i];
1934             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1935                      &coeff_token_len [i][0], 1, 1,
1936                      &coeff_token_bits[i][0], 1, 1,
1937                      INIT_VLC_USE_NEW_STATIC);
1938             offset += coeff_token_vlc_tables_size[i];
1939         }
1940         /*
1941          * This is a one time safety check to make sure that
1942          * the packed static coeff_token_vlc table sizes
1943          * were initialized correctly.
1944          */
1945         assert(offset == FF_ARRAY_ELEMS(coeff_token_vlc_tables));
1946
1947         for(i=0; i<3; i++){
1948             chroma_dc_total_zeros_vlc[i].table = chroma_dc_total_zeros_vlc_tables[i];
1949             chroma_dc_total_zeros_vlc[i].table_allocated = chroma_dc_total_zeros_vlc_tables_size;
1950             init_vlc(&chroma_dc_total_zeros_vlc[i],
1951                      CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1952                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1953                      &chroma_dc_total_zeros_bits[i][0], 1, 1,
1954                      INIT_VLC_USE_NEW_STATIC);
1955         }
1956         for(i=0; i<15; i++){
1957             total_zeros_vlc[i].table = total_zeros_vlc_tables[i];
1958             total_zeros_vlc[i].table_allocated = total_zeros_vlc_tables_size;
1959             init_vlc(&total_zeros_vlc[i],
1960                      TOTAL_ZEROS_VLC_BITS, 16,
1961                      &total_zeros_len [i][0], 1, 1,
1962                      &total_zeros_bits[i][0], 1, 1,
1963                      INIT_VLC_USE_NEW_STATIC);
1964         }
1965
1966         for(i=0; i<6; i++){
1967             run_vlc[i].table = run_vlc_tables[i];
1968             run_vlc[i].table_allocated = run_vlc_tables_size;
1969             init_vlc(&run_vlc[i],
1970                      RUN_VLC_BITS, 7,
1971                      &run_len [i][0], 1, 1,
1972                      &run_bits[i][0], 1, 1,
1973                      INIT_VLC_USE_NEW_STATIC);
1974         }
1975         run7_vlc.table = run7_vlc_table,
1976         run7_vlc.table_allocated = run7_vlc_table_size;
1977         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
1978                  &run_len [6][0], 1, 1,
1979                  &run_bits[6][0], 1, 1,
1980                  INIT_VLC_USE_NEW_STATIC);
1981
1982         init_cavlc_level_tab();
1983     }
1984 }
1985
1986 static void free_tables(H264Context *h){
1987     int i;
1988     H264Context *hx;
1989     av_freep(&h->intra4x4_pred_mode);
1990     av_freep(&h->chroma_pred_mode_table);
1991     av_freep(&h->cbp_table);
1992     av_freep(&h->mvd_table[0]);
1993     av_freep(&h->mvd_table[1]);
1994     av_freep(&h->direct_table);
1995     av_freep(&h->non_zero_count);
1996     av_freep(&h->slice_table_base);
1997     h->slice_table= NULL;
1998
1999     av_freep(&h->mb2b_xy);
2000     av_freep(&h->mb2b8_xy);
2001
2002     for(i = 0; i < h->s.avctx->thread_count; i++) {
2003         hx = h->thread_context[i];
2004         if(!hx) continue;
2005         av_freep(&hx->top_borders[1]);
2006         av_freep(&hx->top_borders[0]);
2007         av_freep(&hx->s.obmc_scratchpad);
2008     }
2009 }
2010
2011 static void init_dequant8_coeff_table(H264Context *h){
2012     int i,q,x;
2013     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2014     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2015     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2016
2017     for(i=0; i<2; i++ ){
2018         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2019             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2020             break;
2021         }
2022
2023         for(q=0; q<52; q++){
2024             int shift = div6[q];
2025             int idx = rem6[q];
2026             for(x=0; x<64; x++)
2027                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2028                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2029                     h->pps.scaling_matrix8[i][x]) << shift;
2030         }
2031     }
2032 }
2033
2034 static void init_dequant4_coeff_table(H264Context *h){
2035     int i,j,q,x;
2036     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2037     for(i=0; i<6; i++ ){
2038         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2039         for(j=0; j<i; j++){
2040             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2041                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2042                 break;
2043             }
2044         }
2045         if(j<i)
2046             continue;
2047
2048         for(q=0; q<52; q++){
2049             int shift = div6[q] + 2;
2050             int idx = rem6[q];
2051             for(x=0; x<16; x++)
2052                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2053                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2054                     h->pps.scaling_matrix4[i][x]) << shift;
2055         }
2056     }
2057 }
2058
2059 static void init_dequant_tables(H264Context *h){
2060     int i,x;
2061     init_dequant4_coeff_table(h);
2062     if(h->pps.transform_8x8_mode)
2063         init_dequant8_coeff_table(h);
2064     if(h->sps.transform_bypass){
2065         for(i=0; i<6; i++)
2066             for(x=0; x<16; x++)
2067                 h->dequant4_coeff[i][0][x] = 1<<6;
2068         if(h->pps.transform_8x8_mode)
2069             for(i=0; i<2; i++)
2070                 for(x=0; x<64; x++)
2071                     h->dequant8_coeff[i][0][x] = 1<<6;
2072     }
2073 }
2074
2075
2076 /**
2077  * allocates tables.
2078  * needs width/height
2079  */
2080 static int alloc_tables(H264Context *h){
2081     MpegEncContext * const s = &h->s;
2082     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2083     int x,y;
2084
2085     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2086
2087     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2088     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base))
2089     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2090
2091     CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2092     CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2093     CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2094     CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2095
2096     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
2097     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2098
2099     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2100     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2101     for(y=0; y<s->mb_height; y++){
2102         for(x=0; x<s->mb_width; x++){
2103             const int mb_xy= x + y*s->mb_stride;
2104             const int b_xy = 4*x + 4*y*h->b_stride;
2105             const int b8_xy= 2*x + 2*y*h->b8_stride;
2106
2107             h->mb2b_xy [mb_xy]= b_xy;
2108             h->mb2b8_xy[mb_xy]= b8_xy;
2109         }
2110     }
2111
2112     s->obmc_scratchpad = NULL;
2113
2114     if(!h->dequant4_coeff[0])
2115         init_dequant_tables(h);
2116
2117     return 0;
2118 fail:
2119     free_tables(h);
2120     return -1;
2121 }
2122
2123 /**
2124  * Mimic alloc_tables(), but for every context thread.
2125  */
2126 static void clone_tables(H264Context *dst, H264Context *src){
2127     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2128     dst->non_zero_count           = src->non_zero_count;
2129     dst->slice_table              = src->slice_table;
2130     dst->cbp_table                = src->cbp_table;
2131     dst->mb2b_xy                  = src->mb2b_xy;
2132     dst->mb2b8_xy                 = src->mb2b8_xy;
2133     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2134     dst->mvd_table[0]             = src->mvd_table[0];
2135     dst->mvd_table[1]             = src->mvd_table[1];
2136     dst->direct_table             = src->direct_table;
2137
2138     dst->s.obmc_scratchpad = NULL;
2139     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2140 }
2141
2142 /**
2143  * Init context
2144  * Allocate buffers which are not shared amongst multiple threads.
2145  */
2146 static int context_init(H264Context *h){
2147     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2148     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2149
2150     return 0;
2151 fail:
2152     return -1; // free_tables will clean up for us
2153 }
2154
2155 static av_cold void common_init(H264Context *h){
2156     MpegEncContext * const s = &h->s;
2157
2158     s->width = s->avctx->width;
2159     s->height = s->avctx->height;
2160     s->codec_id= s->avctx->codec->id;
2161
2162     ff_h264_pred_init(&h->hpc, s->codec_id);
2163
2164     h->dequant_coeff_pps= -1;
2165     s->unrestricted_mv=1;
2166     s->decode=1; //FIXME
2167
2168     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
2169
2170     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2171     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2172 }
2173
2174 static av_cold int decode_init(AVCodecContext *avctx){
2175     H264Context *h= avctx->priv_data;
2176     MpegEncContext * const s = &h->s;
2177
2178     MPV_decode_defaults(s);
2179
2180     s->avctx = avctx;
2181     common_init(h);
2182
2183     s->out_format = FMT_H264;
2184     s->workaround_bugs= avctx->workaround_bugs;
2185
2186     // set defaults
2187 //    s->decode_mb= ff_h263_decode_mb;
2188     s->quarter_sample = 1;
2189     s->low_delay= 1;
2190
2191     if(avctx->codec_id == CODEC_ID_SVQ3)
2192         avctx->pix_fmt= PIX_FMT_YUVJ420P;
2193     else if(s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2194         avctx->pix_fmt= PIX_FMT_VDPAU_H264;
2195     else
2196         avctx->pix_fmt= PIX_FMT_YUV420P;
2197
2198     decode_init_vlc();
2199
2200     if(avctx->extradata_size > 0 && avctx->extradata &&
2201        *(char *)avctx->extradata == 1){
2202         h->is_avc = 1;
2203         h->got_avcC = 0;
2204     } else {
2205         h->is_avc = 0;
2206     }
2207
2208     h->thread_context[0] = h;
2209     h->outputed_poc = INT_MIN;
2210     h->prev_poc_msb= 1<<16;
2211     return 0;
2212 }
2213
2214 static int frame_start(H264Context *h){
2215     MpegEncContext * const s = &h->s;
2216     int i;
2217
2218     if(MPV_frame_start(s, s->avctx) < 0)
2219         return -1;
2220     ff_er_frame_start(s);
2221     /*
2222      * MPV_frame_start uses pict_type to derive key_frame.
2223      * This is incorrect for H.264; IDR markings must be used.
2224      * Zero here; IDR markings per slice in frame or fields are ORed in later.
2225      * See decode_nal_units().
2226      */
2227     s->current_picture_ptr->key_frame= 0;
2228
2229     assert(s->linesize && s->uvlinesize);
2230
2231     for(i=0; i<16; i++){
2232         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2233         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2234     }
2235     for(i=0; i<4; i++){
2236         h->block_offset[16+i]=
2237         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2238         h->block_offset[24+16+i]=
2239         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2240     }
2241
2242     /* can't be in alloc_tables because linesize isn't known there.
2243      * FIXME: redo bipred weight to not require extra buffer? */
2244     for(i = 0; i < s->avctx->thread_count; i++)
2245         if(!h->thread_context[i]->s.obmc_scratchpad)
2246             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2247
2248     /* some macroblocks will be accessed before they're available */
2249     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2250         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
2251
2252 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2253
2254     // We mark the current picture as non-reference after allocating it, so
2255     // that if we break out due to an error it can be released automatically
2256     // in the next MPV_frame_start().
2257     // SVQ3 as well as most other codecs have only last/next/current and thus
2258     // get released even with set reference, besides SVQ3 and others do not
2259     // mark frames as reference later "naturally".
2260     if(s->codec_id != CODEC_ID_SVQ3)
2261         s->current_picture_ptr->reference= 0;
2262
2263     s->current_picture_ptr->field_poc[0]=
2264     s->current_picture_ptr->field_poc[1]= INT_MAX;
2265     assert(s->current_picture_ptr->long_ref==0);
2266
2267     return 0;
2268 }
2269
2270 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){
2271     MpegEncContext * const s = &h->s;
2272     int i;
2273     int step    = 1;
2274     int offset  = 1;
2275     int uvoffset= 1;
2276     int top_idx = 1;
2277     int skiplast= 0;
2278
2279     src_y  -=   linesize;
2280     src_cb -= uvlinesize;
2281     src_cr -= uvlinesize;
2282
2283     if(!simple && FRAME_MBAFF){
2284         if(s->mb_y&1){
2285             offset  = MB_MBAFF ? 1 : 17;
2286             uvoffset= MB_MBAFF ? 1 : 9;
2287             if(!MB_MBAFF){
2288                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 0)= *(uint64_t*)(src_y +  15*linesize);
2289                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 8)= *(uint64_t*)(src_y +8+15*linesize);
2290                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2291                     *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+7*uvlinesize);
2292                     *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+7*uvlinesize);
2293                 }
2294             }
2295         }else{
2296             if(!MB_MBAFF){
2297                 h->left_border[0]= h->top_borders[0][s->mb_x][15];
2298                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2299                     h->left_border[34   ]= h->top_borders[0][s->mb_x][16+7  ];
2300                     h->left_border[34+18]= h->top_borders[0][s->mb_x][16+8+7];
2301                 }
2302                 skiplast= 1;
2303             }
2304             offset  =
2305             uvoffset=
2306             top_idx = MB_MBAFF ? 0 : 1;
2307         }
2308         step= MB_MBAFF ? 2 : 1;
2309     }
2310
2311     // There are two lines saved, the line above the the top macroblock of a pair,
2312     // and the line above the bottom macroblock
2313     h->left_border[offset]= h->top_borders[top_idx][s->mb_x][15];
2314     for(i=1; i<17 - skiplast; i++){
2315         h->left_border[offset+i*step]= src_y[15+i*  linesize];
2316     }
2317
2318     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2319     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2320
2321     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2322         h->left_border[uvoffset+34   ]= h->top_borders[top_idx][s->mb_x][16+7];
2323         h->left_border[uvoffset+34+18]= h->top_borders[top_idx][s->mb_x][24+7];
2324         for(i=1; i<9 - skiplast; i++){
2325             h->left_border[uvoffset+34   +i*step]= src_cb[7+i*uvlinesize];
2326             h->left_border[uvoffset+34+18+i*step]= src_cr[7+i*uvlinesize];
2327         }
2328         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2329         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2330     }
2331 }
2332
2333 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){
2334     MpegEncContext * const s = &h->s;
2335     int temp8, i;
2336     uint64_t temp64;
2337     int deblock_left;
2338     int deblock_top;
2339     int mb_xy;
2340     int step    = 1;
2341     int offset  = 1;
2342     int uvoffset= 1;
2343     int top_idx = 1;
2344
2345     if(!simple && FRAME_MBAFF){
2346         if(s->mb_y&1){
2347             offset  = MB_MBAFF ? 1 : 17;
2348             uvoffset= MB_MBAFF ? 1 : 9;
2349         }else{
2350             offset  =
2351             uvoffset=
2352             top_idx = MB_MBAFF ? 0 : 1;
2353         }
2354         step= MB_MBAFF ? 2 : 1;
2355     }
2356
2357     if(h->deblocking_filter == 2) {
2358         mb_xy = h->mb_xy;
2359         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2360         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2361     } else {
2362         deblock_left = (s->mb_x > 0);
2363         deblock_top =  (s->mb_y > !!MB_FIELD);
2364     }
2365
2366     src_y  -=   linesize + 1;
2367     src_cb -= uvlinesize + 1;
2368     src_cr -= uvlinesize + 1;
2369
2370 #define XCHG(a,b,t,xchg)\
2371 t= a;\
2372 if(xchg)\
2373     a= b;\
2374 b= t;
2375
2376     if(deblock_left){
2377         for(i = !deblock_top; i<16; i++){
2378             XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, xchg);
2379         }
2380         XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, 1);
2381     }
2382
2383     if(deblock_top){
2384         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2385         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2386         if(s->mb_x+1 < s->mb_width){
2387             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2388         }
2389     }
2390
2391     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2392         if(deblock_left){
2393             for(i = !deblock_top; i<8; i++){
2394                 XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, xchg);
2395                 XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, xchg);
2396             }
2397             XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, 1);
2398             XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, 1);
2399         }
2400         if(deblock_top){
2401             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2402             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2403         }
2404     }
2405 }
2406
2407 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2408     MpegEncContext * const s = &h->s;
2409     const int mb_x= s->mb_x;
2410     const int mb_y= s->mb_y;
2411     const int mb_xy= h->mb_xy;
2412     const int mb_type= s->current_picture.mb_type[mb_xy];
2413     uint8_t  *dest_y, *dest_cb, *dest_cr;
2414     int linesize, uvlinesize /*dct_offset*/;
2415     int i;
2416     int *block_offset = &h->block_offset[0];
2417     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2418     /* is_h264 should always be true if SVQ3 is disabled. */
2419     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
2420     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2421     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2422
2423     dest_y  = s->current_picture.data[0] + (mb_x + mb_y * s->linesize  ) * 16;
2424     dest_cb = s->current_picture.data[1] + (mb_x + mb_y * s->uvlinesize) * 8;
2425     dest_cr = s->current_picture.data[2] + (mb_x + mb_y * s->uvlinesize) * 8;
2426
2427     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2428     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2429
2430     if (!simple && MB_FIELD) {
2431         linesize   = h->mb_linesize   = s->linesize * 2;
2432         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2433         block_offset = &h->block_offset[24];
2434         if(mb_y&1){ //FIXME move out of this function?
2435             dest_y -= s->linesize*15;
2436             dest_cb-= s->uvlinesize*7;
2437             dest_cr-= s->uvlinesize*7;
2438         }
2439         if(FRAME_MBAFF) {
2440             int list;
2441             for(list=0; list<h->list_count; list++){
2442                 if(!USES_LIST(mb_type, list))
2443                     continue;
2444                 if(IS_16X16(mb_type)){
2445                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2446                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2447                 }else{
2448                     for(i=0; i<16; i+=4){
2449                         int ref = h->ref_cache[list][scan8[i]];
2450                         if(ref >= 0)
2451                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2452                     }
2453                 }
2454             }
2455         }
2456     } else {
2457         linesize   = h->mb_linesize   = s->linesize;
2458         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2459 //        dct_offset = s->linesize * 16;
2460     }
2461
2462     if (!simple && IS_INTRA_PCM(mb_type)) {
2463         for (i=0; i<16; i++) {
2464             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
2465         }
2466         for (i=0; i<8; i++) {
2467             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
2468             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
2469         }
2470     } else {
2471         if(IS_INTRA(mb_type)){
2472             if(h->deblocking_filter)
2473                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2474
2475             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2476                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2477                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2478             }
2479
2480             if(IS_INTRA4x4(mb_type)){
2481                 if(simple || !s->encoding){
2482                     if(IS_8x8DCT(mb_type)){
2483                         if(transform_bypass){
2484                             idct_dc_add =
2485                             idct_add    = s->dsp.add_pixels8;
2486                         }else{
2487                             idct_dc_add = s->dsp.h264_idct8_dc_add;
2488                             idct_add    = s->dsp.h264_idct8_add;
2489                         }
2490                         for(i=0; i<16; i+=4){
2491                             uint8_t * const ptr= dest_y + block_offset[i];
2492                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2493                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
2494                                 h->hpc.pred8x8l_add[dir](ptr, h->mb + i*16, linesize);
2495                             }else{
2496                                 const int nnz = h->non_zero_count_cache[ scan8[i] ];
2497                                 h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2498                                                             (h->topright_samples_available<<i)&0x4000, linesize);
2499                                 if(nnz){
2500                                     if(nnz == 1 && h->mb[i*16])
2501                                         idct_dc_add(ptr, h->mb + i*16, linesize);
2502                                     else
2503                                         idct_add   (ptr, h->mb + i*16, linesize);
2504                                 }
2505                             }
2506                         }
2507                     }else{
2508                         if(transform_bypass){
2509                             idct_dc_add =
2510                             idct_add    = s->dsp.add_pixels4;
2511                         }else{
2512                             idct_dc_add = s->dsp.h264_idct_dc_add;
2513                             idct_add    = s->dsp.h264_idct_add;
2514                         }
2515                         for(i=0; i<16; i++){
2516                             uint8_t * const ptr= dest_y + block_offset[i];
2517                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2518
2519                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
2520                                 h->hpc.pred4x4_add[dir](ptr, h->mb + i*16, linesize);
2521                             }else{
2522                                 uint8_t *topright;
2523                                 int nnz, tr;
2524                                 if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2525                                     const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2526                                     assert(mb_y || linesize <= block_offset[i]);
2527                                     if(!topright_avail){
2528                                         tr= ptr[3 - linesize]*0x01010101;
2529                                         topright= (uint8_t*) &tr;
2530                                     }else
2531                                         topright= ptr + 4 - linesize;
2532                                 }else
2533                                     topright= NULL;
2534
2535                                 h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2536                                 nnz = h->non_zero_count_cache[ scan8[i] ];
2537                                 if(nnz){
2538                                     if(is_h264){
2539                                         if(nnz == 1 && h->mb[i*16])
2540                                             idct_dc_add(ptr, h->mb + i*16, linesize);
2541                                         else
2542                                             idct_add   (ptr, h->mb + i*16, linesize);
2543                                     }else
2544                                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2545                                 }
2546                             }
2547                         }
2548                     }
2549                 }
2550             }else{
2551                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2552                 if(is_h264){
2553                     if(!transform_bypass)
2554                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2555                 }else
2556                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2557             }
2558             if(h->deblocking_filter)
2559                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2560         }else if(is_h264){
2561             hl_motion(h, dest_y, dest_cb, dest_cr,
2562                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2563                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2564                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2565         }
2566
2567
2568         if(!IS_INTRA4x4(mb_type)){
2569             if(is_h264){
2570                 if(IS_INTRA16x16(mb_type)){
2571                     if(transform_bypass){
2572                         if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
2573                             h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb, linesize);
2574                         }else{
2575                             for(i=0; i<16; i++){
2576                                 if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
2577                                     s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + i*16, linesize);
2578                             }
2579                         }
2580                     }else{
2581                          s->dsp.h264_idct_add16intra(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2582                     }
2583                 }else if(h->cbp&15){
2584                     if(transform_bypass){
2585                         const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2586                         idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2587                         for(i=0; i<16; i+=di){
2588                             if(h->non_zero_count_cache[ scan8[i] ]){
2589                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2590                             }
2591                         }
2592                     }else{
2593                         if(IS_8x8DCT(mb_type)){
2594                             s->dsp.h264_idct8_add4(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2595                         }else{
2596                             s->dsp.h264_idct_add16(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2597                         }
2598                     }
2599                 }
2600             }else{
2601                 for(i=0; i<16; i++){
2602                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2603                         uint8_t * const ptr= dest_y + block_offset[i];
2604                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2605                     }
2606                 }
2607             }
2608         }
2609
2610         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
2611             uint8_t *dest[2] = {dest_cb, dest_cr};
2612             if(transform_bypass){
2613                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
2614                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + 16*16, uvlinesize);
2615                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 20, h->mb + 20*16, uvlinesize);
2616                 }else{
2617                     idct_add = s->dsp.add_pixels4;
2618                     for(i=16; i<16+8; i++){
2619                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
2620                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2621                     }
2622                 }
2623             }else{
2624                 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]);
2625                 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]);
2626                 if(is_h264){
2627                     idct_add = s->dsp.h264_idct_add;
2628                     idct_dc_add = s->dsp.h264_idct_dc_add;
2629                     for(i=16; i<16+8; i++){
2630                         if(h->non_zero_count_cache[ scan8[i] ])
2631                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2632                         else if(h->mb[i*16])
2633                             idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2634                     }
2635                 }else{
2636                     for(i=16; i<16+8; i++){
2637                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2638                             uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2639                             svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2640                         }
2641                     }
2642                 }
2643             }
2644         }
2645     }
2646     if(h->cbp || IS_INTRA(mb_type))
2647         s->dsp.clear_blocks(h->mb);
2648
2649     if(h->deblocking_filter) {
2650         backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2651         fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2652         h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2653         h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2654         if (!simple && FRAME_MBAFF) {
2655             filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2656         } else {
2657             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2658         }
2659     }
2660 }
2661
2662 /**
2663  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2664  */
2665 static void hl_decode_mb_simple(H264Context *h){
2666     hl_decode_mb_internal(h, 1);
2667 }
2668
2669 /**
2670  * Process a macroblock; this handles edge cases, such as interlacing.
2671  */
2672 static void av_noinline hl_decode_mb_complex(H264Context *h){
2673     hl_decode_mb_internal(h, 0);
2674 }
2675
2676 static void hl_decode_mb(H264Context *h){
2677     MpegEncContext * const s = &h->s;
2678     const int mb_xy= h->mb_xy;
2679     const int mb_type= s->current_picture.mb_type[mb_xy];
2680     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2681
2682     if (is_complex)
2683         hl_decode_mb_complex(h);
2684     else hl_decode_mb_simple(h);
2685 }
2686
2687 static void pic_as_field(Picture *pic, const int parity){
2688     int i;
2689     for (i = 0; i < 4; ++i) {
2690         if (parity == PICT_BOTTOM_FIELD)
2691             pic->data[i] += pic->linesize[i];
2692         pic->reference = parity;
2693         pic->linesize[i] *= 2;
2694     }
2695     pic->poc= pic->field_poc[parity == PICT_BOTTOM_FIELD];
2696 }
2697
2698 static int split_field_copy(Picture *dest, Picture *src,
2699                             int parity, int id_add){
2700     int match = !!(src->reference & parity);
2701
2702     if (match) {
2703         *dest = *src;
2704         if(parity != PICT_FRAME){
2705             pic_as_field(dest, parity);
2706             dest->pic_id *= 2;
2707             dest->pic_id += id_add;
2708         }
2709     }
2710
2711     return match;
2712 }
2713
2714 static int build_def_list(Picture *def, Picture **in, int len, int is_long, int sel){
2715     int i[2]={0};
2716     int index=0;
2717
2718     while(i[0]<len || i[1]<len){
2719         while(i[0]<len && !(in[ i[0] ] && (in[ i[0] ]->reference & sel)))
2720             i[0]++;
2721         while(i[1]<len && !(in[ i[1] ] && (in[ i[1] ]->reference & (sel^3))))
2722             i[1]++;
2723         if(i[0] < len){
2724             in[ i[0] ]->pic_id= is_long ? i[0] : in[ i[0] ]->frame_num;
2725             split_field_copy(&def[index++], in[ i[0]++ ], sel  , 1);
2726         }
2727         if(i[1] < len){
2728             in[ i[1] ]->pic_id= is_long ? i[1] : in[ i[1] ]->frame_num;
2729             split_field_copy(&def[index++], in[ i[1]++ ], sel^3, 0);
2730         }
2731     }
2732
2733     return index;
2734 }
2735
2736 static int add_sorted(Picture **sorted, Picture **src, int len, int limit, int dir){
2737     int i, best_poc;
2738     int out_i= 0;
2739
2740     for(;;){
2741         best_poc= dir ? INT_MIN : INT_MAX;
2742
2743         for(i=0; i<len; i++){
2744             const int poc= src[i]->poc;
2745             if(((poc > limit) ^ dir) && ((poc < best_poc) ^ dir)){
2746                 best_poc= poc;
2747                 sorted[out_i]= src[i];
2748             }
2749         }
2750         if(best_poc == (dir ? INT_MIN : INT_MAX))
2751             break;
2752         limit= sorted[out_i++]->poc - dir;
2753     }
2754     return out_i;
2755 }
2756
2757 /**
2758  * fills the default_ref_list.
2759  */
2760 static int fill_default_ref_list(H264Context *h){
2761     MpegEncContext * const s = &h->s;
2762     int i, len;
2763
2764     if(h->slice_type_nos==FF_B_TYPE){
2765         Picture *sorted[32];
2766         int cur_poc, list;
2767         int lens[2];
2768
2769         if(FIELD_PICTURE)
2770             cur_poc= s->current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
2771         else
2772             cur_poc= s->current_picture_ptr->poc;
2773
2774         for(list= 0; list<2; list++){
2775             len= add_sorted(sorted    , h->short_ref, h->short_ref_count, cur_poc, 1^list);
2776             len+=add_sorted(sorted+len, h->short_ref, h->short_ref_count, cur_poc, 0^list);
2777             assert(len<=32);
2778             len= build_def_list(h->default_ref_list[list]    , sorted     , len, 0, s->picture_structure);
2779             len+=build_def_list(h->default_ref_list[list]+len, h->long_ref, 16 , 1, s->picture_structure);
2780             assert(len<=32);
2781
2782             if(len < h->ref_count[list])
2783                 memset(&h->default_ref_list[list][len], 0, sizeof(Picture)*(h->ref_count[list] - len));
2784             lens[list]= len;
2785         }
2786
2787         if(lens[0] == lens[1] && lens[1] > 1){
2788             for(i=0; h->default_ref_list[0][i].data[0] == h->default_ref_list[1][i].data[0] && i<lens[0]; i++);
2789             if(i == lens[0])
2790                 FFSWAP(Picture, h->default_ref_list[1][0], h->default_ref_list[1][1]);
2791         }
2792     }else{
2793         len = build_def_list(h->default_ref_list[0]    , h->short_ref, h->short_ref_count, 0, s->picture_structure);
2794         len+= build_def_list(h->default_ref_list[0]+len, h-> long_ref, 16                , 1, s->picture_structure);
2795         assert(len <= 32);
2796         if(len < h->ref_count[0])
2797             memset(&h->default_ref_list[0][len], 0, sizeof(Picture)*(h->ref_count[0] - len));
2798     }
2799 #ifdef TRACE
2800     for (i=0; i<h->ref_count[0]; i++) {
2801         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]);
2802     }
2803     if(h->slice_type_nos==FF_B_TYPE){
2804         for (i=0; i<h->ref_count[1]; i++) {
2805             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]);
2806         }
2807     }
2808 #endif
2809     return 0;
2810 }
2811
2812 static void print_short_term(H264Context *h);
2813 static void print_long_term(H264Context *h);
2814
2815 /**
2816  * Extract structure information about the picture described by pic_num in
2817  * the current decoding context (frame or field). Note that pic_num is
2818  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
2819  * @param pic_num picture number for which to extract structure information
2820  * @param structure one of PICT_XXX describing structure of picture
2821  *                      with pic_num
2822  * @return frame number (short term) or long term index of picture
2823  *         described by pic_num
2824  */
2825 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
2826     MpegEncContext * const s = &h->s;
2827
2828     *structure = s->picture_structure;
2829     if(FIELD_PICTURE){
2830         if (!(pic_num & 1))
2831             /* opposite field */
2832             *structure ^= PICT_FRAME;
2833         pic_num >>= 1;
2834     }
2835
2836     return pic_num;
2837 }
2838
2839 static int decode_ref_pic_list_reordering(H264Context *h){
2840     MpegEncContext * const s = &h->s;
2841     int list, index, pic_structure;
2842
2843     print_short_term(h);
2844     print_long_term(h);
2845
2846     for(list=0; list<h->list_count; list++){
2847         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
2848
2849         if(get_bits1(&s->gb)){
2850             int pred= h->curr_pic_num;
2851
2852             for(index=0; ; index++){
2853                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb_31(&s->gb);
2854                 unsigned int pic_id;
2855                 int i;
2856                 Picture *ref = NULL;
2857
2858                 if(reordering_of_pic_nums_idc==3)
2859                     break;
2860
2861                 if(index >= h->ref_count[list]){
2862                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
2863                     return -1;
2864                 }
2865
2866                 if(reordering_of_pic_nums_idc<3){
2867                     if(reordering_of_pic_nums_idc<2){
2868                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
2869                         int frame_num;
2870
2871                         if(abs_diff_pic_num > h->max_pic_num){
2872                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
2873                             return -1;
2874                         }
2875
2876                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
2877                         else                                pred+= abs_diff_pic_num;
2878                         pred &= h->max_pic_num - 1;
2879
2880                         frame_num = pic_num_extract(h, pred, &pic_structure);
2881
2882                         for(i= h->short_ref_count-1; i>=0; i--){
2883                             ref = h->short_ref[i];
2884                             assert(ref->reference);
2885                             assert(!ref->long_ref);
2886                             if(
2887                                    ref->frame_num == frame_num &&
2888                                    (ref->reference & pic_structure)
2889                               )
2890                                 break;
2891                         }
2892                         if(i>=0)
2893                             ref->pic_id= pred;
2894                     }else{
2895                         int long_idx;
2896                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
2897
2898                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
2899
2900                         if(long_idx>31){
2901                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
2902                             return -1;
2903                         }
2904                         ref = h->long_ref[long_idx];
2905                         assert(!(ref && !ref->reference));
2906                         if(ref && (ref->reference & pic_structure)){
2907                             ref->pic_id= pic_id;
2908                             assert(ref->long_ref);
2909                             i=0;
2910                         }else{
2911                             i=-1;
2912                         }
2913                     }
2914
2915                     if (i < 0) {
2916                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
2917                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
2918                     } else {
2919                         for(i=index; i+1<h->ref_count[list]; i++){
2920                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
2921                                 break;
2922                         }
2923                         for(; i > index; i--){
2924                             h->ref_list[list][i]= h->ref_list[list][i-1];
2925                         }
2926                         h->ref_list[list][index]= *ref;
2927                         if (FIELD_PICTURE){
2928                             pic_as_field(&h->ref_list[list][index], pic_structure);
2929                         }
2930                     }
2931                 }else{
2932                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
2933                     return -1;
2934                 }
2935             }
2936         }
2937     }
2938     for(list=0; list<h->list_count; list++){
2939         for(index= 0; index < h->ref_count[list]; index++){
2940             if(!h->ref_list[list][index].data[0]){
2941                 av_log(h->s.avctx, AV_LOG_ERROR, "Missing reference picture\n");
2942                 h->ref_list[list][index]= s->current_picture; //FIXME this is not a sensible solution
2943             }
2944         }
2945     }
2946
2947     return 0;
2948 }
2949
2950 static void fill_mbaff_ref_list(H264Context *h){
2951     int list, i, j;
2952     for(list=0; list<2; list++){ //FIXME try list_count
2953         for(i=0; i<h->ref_count[list]; i++){
2954             Picture *frame = &h->ref_list[list][i];
2955             Picture *field = &h->ref_list[list][16+2*i];
2956             field[0] = *frame;
2957             for(j=0; j<3; j++)
2958                 field[0].linesize[j] <<= 1;
2959             field[0].reference = PICT_TOP_FIELD;
2960             field[0].poc= field[0].field_poc[0];
2961             field[1] = field[0];
2962             for(j=0; j<3; j++)
2963                 field[1].data[j] += frame->linesize[j];
2964             field[1].reference = PICT_BOTTOM_FIELD;
2965             field[1].poc= field[1].field_poc[1];
2966
2967             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
2968             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
2969             for(j=0; j<2; j++){
2970                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
2971                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
2972             }
2973         }
2974     }
2975     for(j=0; j<h->ref_count[1]; j++){
2976         for(i=0; i<h->ref_count[0]; i++)
2977             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
2978         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
2979         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
2980     }
2981 }
2982
2983 static int pred_weight_table(H264Context *h){
2984     MpegEncContext * const s = &h->s;
2985     int list, i;
2986     int luma_def, chroma_def;
2987
2988     h->use_weight= 0;
2989     h->use_weight_chroma= 0;
2990     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2991     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2992     luma_def = 1<<h->luma_log2_weight_denom;
2993     chroma_def = 1<<h->chroma_log2_weight_denom;
2994
2995     for(list=0; list<2; list++){
2996         h->luma_weight_flag[list]   = 0;
2997         h->chroma_weight_flag[list] = 0;
2998         for(i=0; i<h->ref_count[list]; i++){
2999             int luma_weight_flag, chroma_weight_flag;
3000
3001             luma_weight_flag= get_bits1(&s->gb);
3002             if(luma_weight_flag){
3003                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3004                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3005                 if(   h->luma_weight[list][i] != luma_def
3006                    || h->luma_offset[list][i] != 0) {
3007                     h->use_weight= 1;
3008                     h->luma_weight_flag[list]= 1;
3009                 }
3010             }else{
3011                 h->luma_weight[list][i]= luma_def;
3012                 h->luma_offset[list][i]= 0;
3013             }
3014
3015             if(CHROMA){
3016                 chroma_weight_flag= get_bits1(&s->gb);
3017                 if(chroma_weight_flag){
3018                     int j;
3019                     for(j=0; j<2; j++){
3020                         h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3021                         h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3022                         if(   h->chroma_weight[list][i][j] != chroma_def
3023                            || h->chroma_offset[list][i][j] != 0) {
3024                             h->use_weight_chroma= 1;
3025                             h->chroma_weight_flag[list]= 1;
3026                         }
3027                     }
3028                 }else{
3029                     int j;
3030                     for(j=0; j<2; j++){
3031                         h->chroma_weight[list][i][j]= chroma_def;
3032                         h->chroma_offset[list][i][j]= 0;
3033                     }
3034                 }
3035             }
3036         }
3037         if(h->slice_type_nos != FF_B_TYPE) break;
3038     }
3039     h->use_weight= h->use_weight || h->use_weight_chroma;
3040     return 0;
3041 }
3042
3043 static void implicit_weight_table(H264Context *h){
3044     MpegEncContext * const s = &h->s;
3045     int ref0, ref1, i;
3046     int cur_poc = s->current_picture_ptr->poc;
3047
3048     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3049        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3050         h->use_weight= 0;
3051         h->use_weight_chroma= 0;
3052         return;
3053     }
3054
3055     h->use_weight= 2;
3056     h->use_weight_chroma= 2;
3057     h->luma_log2_weight_denom= 5;
3058     h->chroma_log2_weight_denom= 5;
3059     for (i = 0; i < 2; i++) {
3060         h->luma_weight_flag[i]   = 0;
3061         h->chroma_weight_flag[i] = 0;
3062     }
3063
3064     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3065         int poc0 = h->ref_list[0][ref0].poc;
3066         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3067             int poc1 = h->ref_list[1][ref1].poc;
3068             int td = av_clip(poc1 - poc0, -128, 127);
3069             if(td){
3070                 int tb = av_clip(cur_poc - poc0, -128, 127);
3071                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3072                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3073                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3074                     h->implicit_weight[ref0][ref1] = 32;
3075                 else
3076                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3077             }else
3078                 h->implicit_weight[ref0][ref1] = 32;
3079         }
3080     }
3081 }
3082
3083 /**
3084  * Mark a picture as no longer needed for reference. The refmask
3085  * argument allows unreferencing of individual fields or the whole frame.
3086  * If the picture becomes entirely unreferenced, but is being held for
3087  * display purposes, it is marked as such.
3088  * @param refmask mask of fields to unreference; the mask is bitwise
3089  *                anded with the reference marking of pic
3090  * @return non-zero if pic becomes entirely unreferenced (except possibly
3091  *         for display purposes) zero if one of the fields remains in
3092  *         reference
3093  */
3094 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3095     int i;
3096     if (pic->reference &= refmask) {
3097         return 0;
3098     } else {
3099         for(i = 0; h->delayed_pic[i]; i++)
3100             if(pic == h->delayed_pic[i]){
3101                 pic->reference=DELAYED_PIC_REF;
3102                 break;
3103             }
3104         return 1;
3105     }
3106 }
3107
3108 /**
3109  * instantaneous decoder refresh.
3110  */
3111 static void idr(H264Context *h){
3112     int i;
3113
3114     for(i=0; i<16; i++){
3115         remove_long(h, i, 0);
3116     }
3117     assert(h->long_ref_count==0);
3118
3119     for(i=0; i<h->short_ref_count; i++){
3120         unreference_pic(h, h->short_ref[i], 0);
3121         h->short_ref[i]= NULL;
3122     }
3123     h->short_ref_count=0;
3124     h->prev_frame_num= 0;
3125     h->prev_frame_num_offset= 0;
3126     h->prev_poc_msb=
3127     h->prev_poc_lsb= 0;
3128 }
3129
3130 /* forget old pics after a seek */
3131 static void flush_dpb(AVCodecContext *avctx){
3132     H264Context *h= avctx->priv_data;
3133     int i;
3134     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
3135         if(h->delayed_pic[i])
3136             h->delayed_pic[i]->reference= 0;
3137         h->delayed_pic[i]= NULL;
3138     }
3139     h->outputed_poc= INT_MIN;
3140     idr(h);
3141     if(h->s.current_picture_ptr)
3142         h->s.current_picture_ptr->reference= 0;
3143     h->s.first_field= 0;
3144     ff_mpeg_flush(avctx);
3145 }
3146
3147 /**
3148  * Find a Picture in the short term reference list by frame number.
3149  * @param frame_num frame number to search for
3150  * @param idx the index into h->short_ref where returned picture is found
3151  *            undefined if no picture found.
3152  * @return pointer to the found picture, or NULL if no pic with the provided
3153  *                 frame number is found
3154  */
3155 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3156     MpegEncContext * const s = &h->s;
3157     int i;
3158
3159     for(i=0; i<h->short_ref_count; i++){
3160         Picture *pic= h->short_ref[i];
3161         if(s->avctx->debug&FF_DEBUG_MMCO)
3162             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3163         if(pic->frame_num == frame_num) {
3164             *idx = i;
3165             return pic;
3166         }
3167     }
3168     return NULL;
3169 }
3170
3171 /**
3172  * Remove a picture from the short term reference list by its index in
3173  * that list.  This does no checking on the provided index; it is assumed
3174  * to be valid. Other list entries are shifted down.
3175  * @param i index into h->short_ref of picture to remove.
3176  */
3177 static void remove_short_at_index(H264Context *h, int i){
3178     assert(i >= 0 && i < h->short_ref_count);
3179     h->short_ref[i]= NULL;
3180     if (--h->short_ref_count)
3181         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3182 }
3183
3184 /**
3185  *
3186  * @return the removed picture or NULL if an error occurs
3187  */
3188 static Picture * remove_short(H264Context *h, int frame_num, int ref_mask){
3189     MpegEncContext * const s = &h->s;
3190     Picture *pic;
3191     int i;
3192
3193     if(s->avctx->debug&FF_DEBUG_MMCO)
3194         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3195
3196     pic = find_short(h, frame_num, &i);
3197     if (pic){
3198         if(unreference_pic(h, pic, ref_mask))
3199         remove_short_at_index(h, i);
3200     }
3201
3202     return pic;
3203 }
3204
3205 /**
3206  * Remove a picture from the long term reference list by its index in
3207  * that list.
3208  * @return the removed picture or NULL if an error occurs
3209  */
3210 static Picture * remove_long(H264Context *h, int i, int ref_mask){
3211     Picture *pic;
3212
3213     pic= h->long_ref[i];
3214     if (pic){
3215         if(unreference_pic(h, pic, ref_mask)){
3216             assert(h->long_ref[i]->long_ref == 1);
3217             h->long_ref[i]->long_ref= 0;
3218             h->long_ref[i]= NULL;
3219             h->long_ref_count--;
3220         }
3221     }
3222
3223     return pic;
3224 }
3225
3226 /**
3227  * print short term list
3228  */
3229 static void print_short_term(H264Context *h) {
3230     uint32_t i;
3231     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3232         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3233         for(i=0; i<h->short_ref_count; i++){
3234             Picture *pic= h->short_ref[i];
3235             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3236         }
3237     }
3238 }
3239
3240 /**
3241  * print long term list
3242  */
3243 static void print_long_term(H264Context *h) {
3244     uint32_t i;
3245     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3246         av_log(h->s.avctx, AV_LOG_DEBUG, "long term list:\n");
3247         for(i = 0; i < 16; i++){
3248             Picture *pic= h->long_ref[i];
3249             if (pic) {
3250                 av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3251             }
3252         }
3253     }
3254 }
3255
3256 /**
3257  * Executes the reference picture marking (memory management control operations).
3258  */
3259 static int execute_ref_pic_marking(H264Context *h, MMCO *mmco, int mmco_count){
3260     MpegEncContext * const s = &h->s;
3261     int i, j;
3262     int current_ref_assigned=0;
3263     Picture *pic;
3264
3265     if((s->avctx->debug&FF_DEBUG_MMCO) && mmco_count==0)
3266         av_log(h->s.avctx, AV_LOG_DEBUG, "no mmco here\n");
3267
3268     for(i=0; i<mmco_count; i++){
3269         int structure, frame_num;
3270         if(s->avctx->debug&FF_DEBUG_MMCO)
3271             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);
3272
3273         if(   mmco[i].opcode == MMCO_SHORT2UNUSED
3274            || mmco[i].opcode == MMCO_SHORT2LONG){
3275             frame_num = pic_num_extract(h, mmco[i].short_pic_num, &structure);
3276             pic = find_short(h, frame_num, &j);
3277             if(!pic){
3278                 if(mmco[i].opcode != MMCO_SHORT2LONG || !h->long_ref[mmco[i].long_arg]
3279                    || h->long_ref[mmco[i].long_arg]->frame_num != frame_num)
3280                 av_log(h->s.avctx, AV_LOG_ERROR, "mmco: unref short failure\n");
3281                 continue;
3282             }
3283         }
3284
3285         switch(mmco[i].opcode){
3286         case MMCO_SHORT2UNUSED:
3287             if(s->avctx->debug&FF_DEBUG_MMCO)
3288                 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);
3289             remove_short(h, frame_num, structure ^ PICT_FRAME);
3290             break;
3291         case MMCO_SHORT2LONG:
3292                 if (h->long_ref[mmco[i].long_arg] != pic)
3293                     remove_long(h, mmco[i].long_arg, 0);
3294
3295                 remove_short_at_index(h, j);
3296                 h->long_ref[ mmco[i].long_arg ]= pic;
3297                 if (h->long_ref[ mmco[i].long_arg ]){
3298                     h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3299                     h->long_ref_count++;
3300                 }
3301             break;
3302         case MMCO_LONG2UNUSED:
3303             j = pic_num_extract(h, mmco[i].long_arg, &structure);
3304             pic = h->long_ref[j];
3305             if (pic) {
3306                 remove_long(h, j, structure ^ PICT_FRAME);
3307             } else if(s->avctx->debug&FF_DEBUG_MMCO)
3308                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref long failure\n");
3309             break;
3310         case MMCO_LONG:
3311                     // Comment below left from previous code as it is an interresting note.
3312                     /* First field in pair is in short term list or
3313                      * at a different long term index.
3314                      * This is not allowed; see 7.4.3.3, notes 2 and 3.
3315                      * Report the problem and keep the pair where it is,
3316                      * and mark this field valid.
3317                      */
3318
3319             if (h->long_ref[mmco[i].long_arg] != s->current_picture_ptr) {
3320                 remove_long(h, mmco[i].long_arg, 0);
3321
3322                 h->long_ref[ mmco[i].long_arg ]= s->current_picture_ptr;
3323                 h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3324                 h->long_ref_count++;
3325             }
3326
3327             s->current_picture_ptr->reference |= s->picture_structure;
3328             current_ref_assigned=1;
3329             break;
3330         case MMCO_SET_MAX_LONG:
3331             assert(mmco[i].long_arg <= 16);
3332             // just remove the long term which index is greater than new max
3333             for(j = mmco[i].long_arg; j<16; j++){
3334                 remove_long(h, j, 0);
3335             }
3336             break;
3337         case MMCO_RESET:
3338             while(h->short_ref_count){
3339                 remove_short(h, h->short_ref[0]->frame_num, 0);
3340             }
3341             for(j = 0; j < 16; j++) {
3342                 remove_long(h, j, 0);
3343             }
3344             s->current_picture_ptr->poc=
3345             s->current_picture_ptr->field_poc[0]=
3346             s->current_picture_ptr->field_poc[1]=
3347             h->poc_lsb=
3348             h->poc_msb=
3349             h->frame_num=
3350             s->current_picture_ptr->frame_num= 0;
3351             break;
3352         default: assert(0);
3353         }
3354     }
3355
3356     if (!current_ref_assigned) {
3357         /* Second field of complementary field pair; the first field of
3358          * which is already referenced. If short referenced, it
3359          * should be first entry in short_ref. If not, it must exist
3360          * in long_ref; trying to put it on the short list here is an
3361          * error in the encoded bit stream (ref: 7.4.3.3, NOTE 2 and 3).
3362          */
3363         if (h->short_ref_count && h->short_ref[0] == s->current_picture_ptr) {
3364             /* Just mark the second field valid */
3365             s->current_picture_ptr->reference = PICT_FRAME;
3366         } else if (s->current_picture_ptr->long_ref) {
3367             av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term reference "
3368                                              "assignment for second field "
3369                                              "in complementary field pair "
3370                                              "(first field is long term)\n");
3371         } else {
3372             pic= remove_short(h, s->current_picture_ptr->frame_num, 0);
3373             if(pic){
3374                 av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term buffer state detected\n");
3375             }
3376
3377             if(h->short_ref_count)
3378                 memmove(&h->short_ref[1], &h->short_ref[0], h->short_ref_count*sizeof(Picture*));
3379
3380             h->short_ref[0]= s->current_picture_ptr;
3381             h->short_ref_count++;
3382             s->current_picture_ptr->reference |= s->picture_structure;
3383         }
3384     }
3385
3386     if (h->long_ref_count + h->short_ref_count > h->sps.ref_frame_count){
3387
3388         /* We have too many reference frames, probably due to corrupted
3389          * stream. Need to discard one frame. Prevents overrun of the
3390          * short_ref and long_ref buffers.
3391          */
3392         av_log(h->s.avctx, AV_LOG_ERROR,
3393                "number of reference frames exceeds max (probably "
3394                "corrupt input), discarding one\n");
3395
3396         if (h->long_ref_count && !h->short_ref_count) {
3397             for (i = 0; i < 16; ++i)
3398                 if (h->long_ref[i])
3399                     break;
3400
3401             assert(i < 16);
3402             remove_long(h, i, 0);
3403         } else {
3404             pic = h->short_ref[h->short_ref_count - 1];
3405             remove_short(h, pic->frame_num, 0);
3406         }
3407     }
3408
3409     print_short_term(h);
3410     print_long_term(h);
3411     return 0;
3412 }
3413
3414 static int decode_ref_pic_marking(H264Context *h, GetBitContext *gb){
3415     MpegEncContext * const s = &h->s;
3416     int i;
3417
3418     h->mmco_index= 0;
3419     if(h->nal_unit_type == NAL_IDR_SLICE){ //FIXME fields
3420         s->broken_link= get_bits1(gb) -1;
3421         if(get_bits1(gb)){
3422             h->mmco[0].opcode= MMCO_LONG;
3423             h->mmco[0].long_arg= 0;
3424             h->mmco_index= 1;
3425         }
3426     }else{
3427         if(get_bits1(gb)){ // adaptive_ref_pic_marking_mode_flag
3428             for(i= 0; i<MAX_MMCO_COUNT; i++) {
3429                 MMCOOpcode opcode= get_ue_golomb_31(gb);
3430
3431                 h->mmco[i].opcode= opcode;
3432                 if(opcode==MMCO_SHORT2UNUSED || opcode==MMCO_SHORT2LONG){
3433                     h->mmco[i].short_pic_num= (h->curr_pic_num - get_ue_golomb(gb) - 1) & (h->max_pic_num - 1);
3434 /*                    if(h->mmco[i].short_pic_num >= h->short_ref_count || h->short_ref[ h->mmco[i].short_pic_num ] == NULL){
3435                         av_log(s->avctx, AV_LOG_ERROR, "illegal short ref in memory management control operation %d\n", mmco);
3436                         return -1;
3437                     }*/
3438                 }
3439                 if(opcode==MMCO_SHORT2LONG || opcode==MMCO_LONG2UNUSED || opcode==MMCO_LONG || opcode==MMCO_SET_MAX_LONG){
3440                     unsigned int long_arg= get_ue_golomb_31(gb);
3441                     if(long_arg >= 32 || (long_arg >= 16 && !(opcode == MMCO_LONG2UNUSED && FIELD_PICTURE))){
3442                         av_log(h->s.avctx, AV_LOG_ERROR, "illegal long ref in memory management control operation %d\n", opcode);
3443                         return -1;
3444                     }
3445                     h->mmco[i].long_arg= long_arg;
3446                 }
3447
3448                 if(opcode > (unsigned)MMCO_LONG){
3449                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal memory management control operation %d\n", opcode);
3450                     return -1;
3451                 }
3452                 if(opcode == MMCO_END)
3453                     break;
3454             }
3455             h->mmco_index= i;
3456         }else{
3457             assert(h->long_ref_count + h->short_ref_count <= h->sps.ref_frame_count);
3458
3459             if(h->short_ref_count && h->long_ref_count + h->short_ref_count == h->sps.ref_frame_count &&
3460                     !(FIELD_PICTURE && !s->first_field && s->current_picture_ptr->reference)) {
3461                 h->mmco[0].opcode= MMCO_SHORT2UNUSED;
3462                 h->mmco[0].short_pic_num= h->short_ref[ h->short_ref_count - 1 ]->frame_num;
3463                 h->mmco_index= 1;
3464                 if (FIELD_PICTURE) {
3465                     h->mmco[0].short_pic_num *= 2;
3466                     h->mmco[1].opcode= MMCO_SHORT2UNUSED;
3467                     h->mmco[1].short_pic_num= h->mmco[0].short_pic_num + 1;
3468                     h->mmco_index= 2;
3469                 }
3470             }
3471         }
3472     }
3473
3474     return 0;
3475 }
3476
3477 static int init_poc(H264Context *h){
3478     MpegEncContext * const s = &h->s;
3479     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
3480     int field_poc[2];
3481     Picture *cur = s->current_picture_ptr;
3482
3483     h->frame_num_offset= h->prev_frame_num_offset;
3484     if(h->frame_num < h->prev_frame_num)
3485         h->frame_num_offset += max_frame_num;
3486
3487     if(h->sps.poc_type==0){
3488         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
3489
3490         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
3491             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
3492         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
3493             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
3494         else
3495             h->poc_msb = h->prev_poc_msb;
3496 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
3497         field_poc[0] =
3498         field_poc[1] = h->poc_msb + h->poc_lsb;
3499         if(s->picture_structure == PICT_FRAME)
3500             field_poc[1] += h->delta_poc_bottom;
3501     }else if(h->sps.poc_type==1){
3502         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
3503         int i;
3504
3505         if(h->sps.poc_cycle_length != 0)
3506             abs_frame_num = h->frame_num_offset + h->frame_num;
3507         else
3508             abs_frame_num = 0;
3509
3510         if(h->nal_ref_idc==0 && abs_frame_num > 0)
3511             abs_frame_num--;
3512
3513         expected_delta_per_poc_cycle = 0;
3514         for(i=0; i < h->sps.poc_cycle_length; i++)
3515             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
3516
3517         if(abs_frame_num > 0){
3518             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
3519             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
3520
3521             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
3522             for(i = 0; i <= frame_num_in_poc_cycle; i++)
3523                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
3524         } else
3525             expectedpoc = 0;
3526
3527         if(h->nal_ref_idc == 0)
3528             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
3529
3530         field_poc[0] = expectedpoc + h->delta_poc[0];
3531         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
3532
3533         if(s->picture_structure == PICT_FRAME)
3534             field_poc[1] += h->delta_poc[1];
3535     }else{
3536         int poc= 2*(h->frame_num_offset + h->frame_num);
3537
3538         if(!h->nal_ref_idc)
3539             poc--;
3540
3541         field_poc[0]= poc;
3542         field_poc[1]= poc;
3543     }
3544
3545     if(s->picture_structure != PICT_BOTTOM_FIELD)
3546         s->current_picture_ptr->field_poc[0]= field_poc[0];
3547     if(s->picture_structure != PICT_TOP_FIELD)
3548         s->current_picture_ptr->field_poc[1]= field_poc[1];
3549     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
3550
3551     return 0;
3552 }
3553
3554
3555 /**
3556  * initialize scan tables
3557  */
3558 static void init_scan_tables(H264Context *h){
3559     MpegEncContext * const s = &h->s;
3560     int i;
3561     if(s->dsp.h264_idct_add == ff_h264_idct_add_c){ //FIXME little ugly
3562         memcpy(h->zigzag_scan, zigzag_scan, 16*sizeof(uint8_t));
3563         memcpy(h-> field_scan,  field_scan, 16*sizeof(uint8_t));
3564     }else{
3565         for(i=0; i<16; i++){
3566 #define T(x) (x>>2) | ((x<<2) & 0xF)
3567             h->zigzag_scan[i] = T(zigzag_scan[i]);
3568             h-> field_scan[i] = T( field_scan[i]);
3569 #undef T
3570         }
3571     }
3572     if(s->dsp.h264_idct8_add == ff_h264_idct8_add_c){
3573         memcpy(h->zigzag_scan8x8,       ff_zigzag_direct,     64*sizeof(uint8_t));
3574         memcpy(h->zigzag_scan8x8_cavlc, zigzag_scan8x8_cavlc, 64*sizeof(uint8_t));
3575         memcpy(h->field_scan8x8,        field_scan8x8,        64*sizeof(uint8_t));
3576         memcpy(h->field_scan8x8_cavlc,  field_scan8x8_cavlc,  64*sizeof(uint8_t));
3577     }else{
3578         for(i=0; i<64; i++){
3579 #define T(x) (x>>3) | ((x&7)<<3)
3580             h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
3581             h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
3582             h->field_scan8x8[i]        = T(field_scan8x8[i]);
3583             h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
3584 #undef T
3585         }
3586     }
3587     if(h->sps.transform_bypass){ //FIXME same ugly
3588         h->zigzag_scan_q0          = zigzag_scan;
3589         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
3590         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
3591         h->field_scan_q0           = field_scan;
3592         h->field_scan8x8_q0        = field_scan8x8;
3593         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
3594     }else{
3595         h->zigzag_scan_q0          = h->zigzag_scan;
3596         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
3597         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
3598         h->field_scan_q0           = h->field_scan;
3599         h->field_scan8x8_q0        = h->field_scan8x8;
3600         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
3601     }
3602 }
3603
3604 /**
3605  * Replicates H264 "master" context to thread contexts.
3606  */
3607 static void clone_slice(H264Context *dst, H264Context *src)
3608 {
3609     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
3610     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
3611     dst->s.current_picture      = src->s.current_picture;
3612     dst->s.linesize             = src->s.linesize;
3613     dst->s.uvlinesize           = src->s.uvlinesize;
3614     dst->s.first_field          = src->s.first_field;
3615
3616     dst->prev_poc_msb           = src->prev_poc_msb;
3617     dst->prev_poc_lsb           = src->prev_poc_lsb;
3618     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
3619     dst->prev_frame_num         = src->prev_frame_num;
3620     dst->short_ref_count        = src->short_ref_count;
3621
3622     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
3623     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
3624     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
3625     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
3626
3627     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
3628     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
3629 }
3630
3631 /**
3632  * decodes a slice header.
3633  * This will also call MPV_common_init() and frame_start() as needed.
3634  *
3635  * @param h h264context
3636  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
3637  *
3638  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
3639  */
3640 static int decode_slice_header(H264Context *h, H264Context *h0){
3641     MpegEncContext * const s = &h->s;
3642     MpegEncContext * const s0 = &h0->s;
3643     unsigned int first_mb_in_slice;
3644     unsigned int pps_id;
3645     int num_ref_idx_active_override_flag;
3646     unsigned int slice_type, tmp, i, j;
3647     int default_ref_list_done = 0;
3648     int last_pic_structure;
3649
3650     s->dropable= h->nal_ref_idc == 0;
3651
3652     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc){
3653         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
3654         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
3655     }else{
3656         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
3657         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
3658     }
3659
3660     first_mb_in_slice= get_ue_golomb(&s->gb);
3661
3662     if((s->flags2 & CODEC_FLAG2_CHUNKS) && first_mb_in_slice == 0){
3663         h0->current_slice = 0;
3664         if (!s0->first_field)
3665             s->current_picture_ptr= NULL;
3666     }
3667
3668     slice_type= get_ue_golomb_31(&s->gb);
3669     if(slice_type > 9){
3670         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);
3671         return -1;
3672     }
3673     if(slice_type > 4){
3674         slice_type -= 5;
3675         h->slice_type_fixed=1;
3676     }else
3677         h->slice_type_fixed=0;
3678
3679     slice_type= golomb_to_pict_type[ slice_type ];
3680     if (slice_type == FF_I_TYPE
3681         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
3682         default_ref_list_done = 1;
3683     }
3684     h->slice_type= slice_type;
3685     h->slice_type_nos= slice_type & 3;
3686
3687     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
3688     if (s->pict_type == FF_B_TYPE && s0->last_picture_ptr == NULL) {
3689         av_log(h->s.avctx, AV_LOG_ERROR,
3690                "B picture before any references, skipping\n");
3691         return -1;
3692     }
3693
3694     pps_id= get_ue_golomb(&s->gb);
3695     if(pps_id>=MAX_PPS_COUNT){
3696         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
3697         return -1;
3698     }
3699     if(!h0->pps_buffers[pps_id]) {
3700         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS referenced\n");
3701         return -1;
3702     }
3703     h->pps= *h0->pps_buffers[pps_id];
3704
3705     if(!h0->sps_buffers[h->pps.sps_id]) {
3706         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS referenced\n");
3707         return -1;
3708     }
3709     h->sps = *h0->sps_buffers[h->pps.sps_id];
3710
3711     if(h == h0 && h->dequant_coeff_pps != pps_id){
3712         h->dequant_coeff_pps = pps_id;
3713         init_dequant_tables(h);
3714     }
3715
3716     s->mb_width= h->sps.mb_width;
3717     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
3718
3719     h->b_stride=  s->mb_width*4;
3720     h->b8_stride= s->mb_width*2;
3721
3722     s->width = 16*s->mb_width - 2*FFMIN(h->sps.crop_right, 7);
3723     if(h->sps.frame_mbs_only_flag)
3724         s->height= 16*s->mb_height - 2*FFMIN(h->sps.crop_bottom, 7);
3725     else
3726         s->height= 16*s->mb_height - 4*FFMIN(h->sps.crop_bottom, 3);
3727
3728     if (s->context_initialized
3729         && (   s->width != s->avctx->width || s->height != s->avctx->height)) {
3730         if(h != h0)
3731             return -1;   // width / height changed during parallelized decoding
3732         free_tables(h);
3733         flush_dpb(s->avctx);
3734         MPV_common_end(s);
3735     }
3736     if (!s->context_initialized) {
3737         if(h != h0)
3738             return -1;  // we cant (re-)initialize context during parallel decoding
3739         if (MPV_common_init(s) < 0)
3740             return -1;
3741         s->first_field = 0;
3742
3743         init_scan_tables(h);
3744         alloc_tables(h);
3745
3746         for(i = 1; i < s->avctx->thread_count; i++) {
3747             H264Context *c;
3748             c = h->thread_context[i] = av_malloc(sizeof(H264Context));
3749             memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
3750             memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
3751             c->sps = h->sps;
3752             c->pps = h->pps;
3753             init_scan_tables(c);
3754             clone_tables(c, h);
3755         }
3756
3757         for(i = 0; i < s->avctx->thread_count; i++)
3758             if(context_init(h->thread_context[i]) < 0)
3759                 return -1;
3760
3761         s->avctx->width = s->width;
3762         s->avctx->height = s->height;
3763         s->avctx->sample_aspect_ratio= h->sps.sar;
3764         if(!s->avctx->sample_aspect_ratio.den)
3765             s->avctx->sample_aspect_ratio.den = 1;
3766
3767         if(h->sps.timing_info_present_flag){
3768             s->avctx->time_base= (AVRational){h->sps.num_units_in_tick * 2, h->sps.time_scale};
3769             if(h->x264_build > 0 && h->x264_build < 44)
3770                 s->avctx->time_base.den *= 2;
3771             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
3772                       s->avctx->time_base.num, s->avctx->time_base.den, 1<<30);
3773         }
3774     }
3775
3776     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
3777
3778     h->mb_mbaff = 0;
3779     h->mb_aff_frame = 0;
3780     last_pic_structure = s0->picture_structure;
3781     if(h->sps.frame_mbs_only_flag){
3782         s->picture_structure= PICT_FRAME;
3783     }else{
3784         if(get_bits1(&s->gb)) { //field_pic_flag
3785             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
3786         } else {
3787             s->picture_structure= PICT_FRAME;
3788             h->mb_aff_frame = h->sps.mb_aff;
3789         }
3790     }
3791     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
3792
3793     if(h0->current_slice == 0){
3794         while(h->frame_num !=  h->prev_frame_num &&
3795               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
3796             av_log(NULL, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
3797             frame_start(h);
3798             h->prev_frame_num++;
3799             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
3800             s->current_picture_ptr->frame_num= h->prev_frame_num;
3801             execute_ref_pic_marking(h, NULL, 0);
3802         }
3803
3804         /* See if we have a decoded first field looking for a pair... */
3805         if (s0->first_field) {
3806             assert(s0->current_picture_ptr);
3807             assert(s0->current_picture_ptr->data[0]);
3808             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
3809
3810             /* figure out if we have a complementary field pair */
3811             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3812                 /*
3813                  * Previous field is unmatched. Don't display it, but let it
3814                  * remain for reference if marked as such.
3815                  */
3816                 s0->current_picture_ptr = NULL;
3817                 s0->first_field = FIELD_PICTURE;
3818
3819             } else {
3820                 if (h->nal_ref_idc &&
3821                         s0->current_picture_ptr->reference &&
3822                         s0->current_picture_ptr->frame_num != h->frame_num) {
3823                     /*
3824                      * This and previous field were reference, but had
3825                      * different frame_nums. Consider this field first in
3826                      * pair. Throw away previous field except for reference
3827                      * purposes.
3828                      */
3829                     s0->first_field = 1;
3830                     s0->current_picture_ptr = NULL;
3831
3832                 } else {
3833                     /* Second field in complementary pair */
3834                     s0->first_field = 0;
3835                 }
3836             }
3837
3838         } else {
3839             /* Frame or first field in a potentially complementary pair */
3840             assert(!s0->current_picture_ptr);
3841             s0->first_field = FIELD_PICTURE;
3842         }
3843
3844         if((!FIELD_PICTURE || s0->first_field) && frame_start(h) < 0) {
3845             s0->first_field = 0;
3846             return -1;
3847         }
3848     }
3849     if(h != h0)
3850         clone_slice(h, h0);
3851
3852     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
3853
3854     assert(s->mb_num == s->mb_width * s->mb_height);
3855     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
3856        first_mb_in_slice                    >= s->mb_num){
3857         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
3858         return -1;
3859     }
3860     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
3861     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
3862     if (s->picture_structure == PICT_BOTTOM_FIELD)
3863         s->resync_mb_y = s->mb_y = s->mb_y + 1;
3864     assert(s->mb_y < s->mb_height);
3865
3866     if(s->picture_structure==PICT_FRAME){
3867         h->curr_pic_num=   h->frame_num;
3868         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
3869     }else{
3870         h->curr_pic_num= 2*h->frame_num + 1;
3871         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
3872     }
3873
3874     if(h->nal_unit_type == NAL_IDR_SLICE){
3875         get_ue_golomb(&s->gb); /* idr_pic_id */
3876     }
3877
3878     if(h->sps.poc_type==0){
3879         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
3880
3881         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
3882             h->delta_poc_bottom= get_se_golomb(&s->gb);
3883         }
3884     }
3885
3886     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
3887         h->delta_poc[0]= get_se_golomb(&s->gb);
3888
3889         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
3890             h->delta_poc[1]= get_se_golomb(&s->gb);
3891     }
3892
3893     init_poc(h);
3894
3895     if(h->pps.redundant_pic_cnt_present){
3896         h->redundant_pic_count= get_ue_golomb(&s->gb);
3897     }
3898
3899     //set defaults, might be overridden a few lines later
3900     h->ref_count[0]= h->pps.ref_count[0];
3901     h->ref_count[1]= h->pps.ref_count[1];
3902
3903     if(h->slice_type_nos != FF_I_TYPE){
3904         if(h->slice_type_nos == FF_B_TYPE){
3905             h->direct_spatial_mv_pred= get_bits1(&s->gb);
3906         }
3907         num_ref_idx_active_override_flag= get_bits1(&s->gb);
3908
3909         if(num_ref_idx_active_override_flag){
3910             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
3911             if(h->slice_type_nos==FF_B_TYPE)
3912                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
3913
3914             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
3915                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3916                 h->ref_count[0]= h->ref_count[1]= 1;
3917                 return -1;
3918             }
3919         }
3920         if(h->slice_type_nos == FF_B_TYPE)
3921             h->list_count= 2;
3922         else
3923             h->list_count= 1;
3924     }else
3925         h->list_count= 0;
3926
3927     if(!default_ref_list_done){
3928         fill_default_ref_list(h);
3929     }
3930
3931     if(h->slice_type_nos!=FF_I_TYPE && decode_ref_pic_list_reordering(h) < 0)
3932         return -1;
3933
3934     if(h->slice_type_nos!=FF_I_TYPE){
3935         s->last_picture_ptr= &h->ref_list[0][0];
3936         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
3937     }
3938     if(h->slice_type_nos==FF_B_TYPE){
3939         s->next_picture_ptr= &h->ref_list[1][0];
3940         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
3941     }
3942
3943     if(   (h->pps.weighted_pred          && h->slice_type_nos == FF_P_TYPE )
3944        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== FF_B_TYPE ) )
3945         pred_weight_table(h);
3946     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== FF_B_TYPE)
3947         implicit_weight_table(h);
3948     else {
3949         h->use_weight = 0;
3950         for (i = 0; i < 2; i++) {
3951             h->luma_weight_flag[i]   = 0;
3952             h->chroma_weight_flag[i] = 0;
3953         }
3954     }
3955
3956     if(h->nal_ref_idc)
3957         decode_ref_pic_marking(h0, &s->gb);
3958
3959     if(FRAME_MBAFF)
3960         fill_mbaff_ref_list(h);
3961
3962     if(h->slice_type_nos==FF_B_TYPE && !h->direct_spatial_mv_pred)
3963         direct_dist_scale_factor(h);
3964     direct_ref_list_init(h);
3965
3966     if( h->slice_type_nos != FF_I_TYPE && h->pps.cabac ){
3967         tmp = get_ue_golomb_31(&s->gb);
3968         if(tmp > 2){
3969             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3970             return -1;
3971         }
3972         h->cabac_init_idc= tmp;
3973     }
3974
3975     h->last_qscale_diff = 0;
3976     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3977     if(tmp>51){
3978         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3979         return -1;
3980     }
3981     s->qscale= tmp;
3982     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3983     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3984     //FIXME qscale / qp ... stuff
3985     if(h->slice_type == FF_SP_TYPE){
3986         get_bits1(&s->gb); /* sp_for_switch_flag */
3987     }
3988     if(h->slice_type==FF_SP_TYPE || h->slice_type == FF_SI_TYPE){
3989         get_se_golomb(&s->gb); /* slice_qs_delta */
3990     }
3991
3992     h->deblocking_filter = 1;
3993     h->slice_alpha_c0_offset = 0;
3994     h->slice_beta_offset = 0;
3995     if( h->pps.deblocking_filter_parameters_present ) {
3996         tmp= get_ue_golomb_31(&s->gb);
3997         if(tmp > 2){
3998             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
3999             return -1;
4000         }
4001         h->deblocking_filter= tmp;
4002         if(h->deblocking_filter < 2)
4003             h->deblocking_filter^= 1; // 1<->0
4004
4005         if( h->deblocking_filter ) {
4006             h->slice_alpha_c0_offset = get_se_golomb(&s->gb) << 1;
4007             h->slice_beta_offset = get_se_golomb(&s->gb) << 1;
4008         }
4009     }
4010
4011     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
4012        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != FF_I_TYPE)
4013        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == FF_B_TYPE)
4014        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
4015         h->deblocking_filter= 0;
4016
4017     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
4018         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {