619b618765ba652d8dc6c8625872ebc795687a68
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file h264.c
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "dsputil.h"
29 #include "avcodec.h"
30 #include "mpegvideo.h"
31 #include "h264.h"
32 #include "h264data.h"
33 #include "h264_parser.h"
34 #include "golomb.h"
35 #include "rectangle.h"
36
37 #include "cabac.h"
38 #ifdef ARCH_X86
39 #include "i386/h264_i386.h"
40 #endif
41
42 //#undef NDEBUG
43 #include <assert.h>
44
45 /**
46  * Value of Picture.reference when Picture is not a reference picture, but
47  * is held for delayed output.
48  */
49 #define DELAYED_PIC_REF 4
50
51 static VLC coeff_token_vlc[4];
52 static VLC_TYPE coeff_token_vlc_tables[520+332+280+256][2];
53 static const int coeff_token_vlc_tables_size[4]={520,332,280,256};
54
55 static VLC chroma_dc_coeff_token_vlc;
56 static VLC_TYPE chroma_dc_coeff_token_vlc_table[256][2];
57 static const int chroma_dc_coeff_token_vlc_table_size = 256;
58
59 static VLC total_zeros_vlc[15];
60 static VLC_TYPE total_zeros_vlc_tables[15][512][2];
61 static const int total_zeros_vlc_tables_size = 512;
62
63 static VLC chroma_dc_total_zeros_vlc[3];
64 static VLC_TYPE chroma_dc_total_zeros_vlc_tables[3][8][2];
65 static const int chroma_dc_total_zeros_vlc_tables_size = 8;
66
67 static VLC run_vlc[6];
68 static VLC_TYPE run_vlc_tables[6][8][2];
69 static const int run_vlc_tables_size = 8;
70
71 static VLC run7_vlc;
72 static VLC_TYPE run7_vlc_table[96][2];
73 static const int run7_vlc_table_size = 96;
74
75 static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
76 static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
77 static void filter_mb( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
78 static void filter_mb_fast( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
79 static Picture * remove_long(H264Context *h, int i, int ref_mask);
80
81 static av_always_inline uint32_t pack16to32(int a, int b){
82 #ifdef WORDS_BIGENDIAN
83    return (b&0xFFFF) + (a<<16);
84 #else
85    return (a&0xFFFF) + (b<<16);
86 #endif
87 }
88
89 const uint8_t ff_rem6[52]={
90 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
91 };
92
93 const uint8_t ff_div6[52]={
94 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8,
95 };
96
97 static const int left_block_options[4][8]={
98     {0,1,2,3,7,10,8,11},
99     {2,2,3,3,8,11,8,11},
100     {0,0,1,1,7,10,7,10},
101     {0,2,0,2,7,10,7,10}
102 };
103
104 static void fill_caches(H264Context *h, int mb_type, int for_deblock){
105     MpegEncContext * const s = &h->s;
106     const int mb_xy= h->mb_xy;
107     int topleft_xy, top_xy, topright_xy, left_xy[2];
108     int topleft_type, top_type, topright_type, left_type[2];
109     int * left_block;
110     int topleft_partition= -1;
111     int i;
112
113     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
114
115     //FIXME deblocking could skip the intra and nnz parts.
116     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
117         return;
118
119     /* Wow, what a mess, why didn't they simplify the interlacing & intra
120      * stuff, I can't imagine that these complex rules are worth it. */
121
122     topleft_xy = top_xy - 1;
123     topright_xy= top_xy + 1;
124     left_xy[1] = left_xy[0] = mb_xy-1;
125     left_block = left_block_options[0];
126     if(FRAME_MBAFF){
127         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
128         const int top_pair_xy      = pair_xy     - s->mb_stride;
129         const int topleft_pair_xy  = top_pair_xy - 1;
130         const int topright_pair_xy = top_pair_xy + 1;
131         const int topleft_mb_frame_flag  = !IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
132         const int top_mb_frame_flag      = !IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
133         const int topright_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
134         const int left_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
135         const int curr_mb_frame_flag = !IS_INTERLACED(mb_type);
136         const int bottom = (s->mb_y & 1);
137         tprintf(s->avctx, "fill_caches: curr_mb_frame_flag:%d, left_mb_frame_flag:%d, topleft_mb_frame_flag:%d, top_mb_frame_flag:%d, topright_mb_frame_flag:%d\n", curr_mb_frame_flag, left_mb_frame_flag, topleft_mb_frame_flag, top_mb_frame_flag, topright_mb_frame_flag);
138         if (bottom
139                 ? !curr_mb_frame_flag // bottom macroblock
140                 : (!curr_mb_frame_flag && !top_mb_frame_flag) // top macroblock
141                 ) {
142             top_xy -= s->mb_stride;
143         }
144         if (bottom
145                 ? !curr_mb_frame_flag // bottom macroblock
146                 : (!curr_mb_frame_flag && !topleft_mb_frame_flag) // top macroblock
147                 ) {
148             topleft_xy -= s->mb_stride;
149         } else if(bottom && curr_mb_frame_flag && !left_mb_frame_flag) {
150             topleft_xy += s->mb_stride;
151             // take top left mv from the middle of the mb, as opposed to all other modes which use the bottom right partition
152             topleft_partition = 0;
153         }
154         if (bottom
155                 ? !curr_mb_frame_flag // bottom macroblock
156                 : (!curr_mb_frame_flag && !topright_mb_frame_flag) // top macroblock
157                 ) {
158             topright_xy -= s->mb_stride;
159         }
160         if (left_mb_frame_flag != curr_mb_frame_flag) {
161             left_xy[1] = left_xy[0] = pair_xy - 1;
162             if (curr_mb_frame_flag) {
163                 if (bottom) {
164                     left_block = left_block_options[1];
165                 } else {
166                     left_block= left_block_options[2];
167                 }
168             } else {
169                 left_xy[1] += s->mb_stride;
170                 left_block = left_block_options[3];
171             }
172         }
173     }
174
175     h->top_mb_xy = top_xy;
176     h->left_mb_xy[0] = left_xy[0];
177     h->left_mb_xy[1] = left_xy[1];
178     if(for_deblock){
179         topleft_type = 0;
180         topright_type = 0;
181         top_type     = h->slice_table[top_xy     ] < 255 ? s->current_picture.mb_type[top_xy]     : 0;
182         left_type[0] = h->slice_table[left_xy[0] ] < 255 ? s->current_picture.mb_type[left_xy[0]] : 0;
183         left_type[1] = h->slice_table[left_xy[1] ] < 255 ? s->current_picture.mb_type[left_xy[1]] : 0;
184
185         if(MB_MBAFF && !IS_INTRA(mb_type)){
186             int list;
187             for(list=0; list<h->list_count; list++){
188                 //These values where changed for ease of performing MC, we need to change them back
189                 //FIXME maybe we can make MC and loop filter use the same values or prevent
190                 //the MC code from changing ref_cache and rather use a temporary array.
191                 if(USES_LIST(mb_type,list)){
192                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
193                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
194                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
195                     ref += h->b8_stride;
196                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
197                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
198                 }
199             }
200         }
201     }else{
202         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
203         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
204         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
205         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
206         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
207     }
208
209     if(IS_INTRA(mb_type)){
210         h->topleft_samples_available=
211         h->top_samples_available=
212         h->left_samples_available= 0xFFFF;
213         h->topright_samples_available= 0xEEEA;
214
215         if(!IS_INTRA(top_type) && (top_type==0 || h->pps.constrained_intra_pred)){
216             h->topleft_samples_available= 0xB3FF;
217             h->top_samples_available= 0x33FF;
218             h->topright_samples_available= 0x26EA;
219         }
220         for(i=0; i<2; i++){
221             if(!IS_INTRA(left_type[i]) && (left_type[i]==0 || h->pps.constrained_intra_pred)){
222                 h->topleft_samples_available&= 0xDF5F;
223                 h->left_samples_available&= 0x5F5F;
224             }
225         }
226
227         if(!IS_INTRA(topleft_type) && (topleft_type==0 || h->pps.constrained_intra_pred))
228             h->topleft_samples_available&= 0x7FFF;
229
230         if(!IS_INTRA(topright_type) && (topright_type==0 || h->pps.constrained_intra_pred))
231             h->topright_samples_available&= 0xFBFF;
232
233         if(IS_INTRA4x4(mb_type)){
234             if(IS_INTRA4x4(top_type)){
235                 h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
236                 h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
237                 h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
238                 h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
239             }else{
240                 int pred;
241                 if(!top_type || (IS_INTER(top_type) && h->pps.constrained_intra_pred))
242                     pred= -1;
243                 else{
244                     pred= 2;
245                 }
246                 h->intra4x4_pred_mode_cache[4+8*0]=
247                 h->intra4x4_pred_mode_cache[5+8*0]=
248                 h->intra4x4_pred_mode_cache[6+8*0]=
249                 h->intra4x4_pred_mode_cache[7+8*0]= pred;
250             }
251             for(i=0; i<2; i++){
252                 if(IS_INTRA4x4(left_type[i])){
253                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
254                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
255                 }else{
256                     int pred;
257                     if(!left_type[i] || (IS_INTER(left_type[i]) && h->pps.constrained_intra_pred))
258                         pred= -1;
259                     else{
260                         pred= 2;
261                     }
262                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
263                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
264                 }
265             }
266         }
267     }
268
269
270 /*
271 0 . T T. T T T T
272 1 L . .L . . . .
273 2 L . .L . . . .
274 3 . T TL . . . .
275 4 L . .L . . . .
276 5 L . .. . . . .
277 */
278 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
279     if(top_type){
280         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
281         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
282         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
283         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
284
285         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
286         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
287
288         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
289         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
290
291     }else{
292         h->non_zero_count_cache[4+8*0]=
293         h->non_zero_count_cache[5+8*0]=
294         h->non_zero_count_cache[6+8*0]=
295         h->non_zero_count_cache[7+8*0]=
296
297         h->non_zero_count_cache[1+8*0]=
298         h->non_zero_count_cache[2+8*0]=
299
300         h->non_zero_count_cache[1+8*3]=
301         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
302
303     }
304
305     for (i=0; i<2; i++) {
306         if(left_type[i]){
307             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
308             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
309             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
310             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
311         }else{
312             h->non_zero_count_cache[3+8*1 + 2*8*i]=
313             h->non_zero_count_cache[3+8*2 + 2*8*i]=
314             h->non_zero_count_cache[0+8*1 +   8*i]=
315             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
316         }
317     }
318
319     if( h->pps.cabac ) {
320         // top_cbp
321         if(top_type) {
322             h->top_cbp = h->cbp_table[top_xy];
323         } else if(IS_INTRA(mb_type)) {
324             h->top_cbp = 0x1C0;
325         } else {
326             h->top_cbp = 0;
327         }
328         // left_cbp
329         if (left_type[0]) {
330             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
331         } else if(IS_INTRA(mb_type)) {
332             h->left_cbp = 0x1C0;
333         } else {
334             h->left_cbp = 0;
335         }
336         if (left_type[0]) {
337             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
338         }
339         if (left_type[1]) {
340             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
341         }
342     }
343
344 #if 1
345     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
346         int list;
347         for(list=0; list<h->list_count; list++){
348             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
349                 /*if(!h->mv_cache_clean[list]){
350                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
351                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
352                     h->mv_cache_clean[list]= 1;
353                 }*/
354                 continue;
355             }
356             h->mv_cache_clean[list]= 0;
357
358             if(USES_LIST(top_type, list)){
359                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
360                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
361                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
362                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
363                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
364                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
365                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
366                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
367                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
368                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
369             }else{
370                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
371                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
372                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
373                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
374                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
375             }
376
377             for(i=0; i<2; i++){
378                 int cache_idx = scan8[0] - 1 + i*2*8;
379                 if(USES_LIST(left_type[i], list)){
380                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
381                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
382                     *(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]];
383                     *(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]];
384                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
385                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
386                 }else{
387                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
388                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
389                     h->ref_cache[list][cache_idx  ]=
390                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
391                 }
392             }
393
394             if(for_deblock || ((IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred) && !FRAME_MBAFF))
395                 continue;
396
397             if(USES_LIST(topleft_type, list)){
398                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
399                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
400                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
401                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
402             }else{
403                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
404                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
405             }
406
407             if(USES_LIST(topright_type, list)){
408                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
409                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
410                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
411                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
412             }else{
413                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
414                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
415             }
416
417             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
418                 continue;
419
420             h->ref_cache[list][scan8[5 ]+1] =
421             h->ref_cache[list][scan8[7 ]+1] =
422             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
423             h->ref_cache[list][scan8[4 ]] =
424             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
425             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
426             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
427             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
428             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
429             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
430
431             if( h->pps.cabac ) {
432                 /* XXX beurk, Load mvd */
433                 if(USES_LIST(top_type, list)){
434                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
435                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
436                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
437                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
438                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
439                 }else{
440                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
441                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
442                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
443                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
444                 }
445                 if(USES_LIST(left_type[0], list)){
446                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
447                     *(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]];
448                     *(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]];
449                 }else{
450                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
451                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
452                 }
453                 if(USES_LIST(left_type[1], list)){
454                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
455                     *(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]];
456                     *(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]];
457                 }else{
458                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
459                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
460                 }
461                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
462                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
463                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
464                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
465                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
466
467                 if(h->slice_type_nos == FF_B_TYPE){
468                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
469
470                     if(IS_DIRECT(top_type)){
471                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
472                     }else if(IS_8X8(top_type)){
473                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
474                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
475                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
476                     }else{
477                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
478                     }
479
480                     if(IS_DIRECT(left_type[0]))
481                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
482                     else if(IS_8X8(left_type[0]))
483                         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)];
484                     else
485                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
486
487                     if(IS_DIRECT(left_type[1]))
488                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
489                     else if(IS_8X8(left_type[1]))
490                         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)];
491                     else
492                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
493                 }
494             }
495
496             if(FRAME_MBAFF){
497 #define MAP_MVS\
498                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
499                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
500                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
501                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
502                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
503                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
504                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
505                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
506                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
507                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
508                 if(MB_FIELD){
509 #define MAP_F2F(idx, mb_type)\
510                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
511                         h->ref_cache[list][idx] <<= 1;\
512                         h->mv_cache[list][idx][1] /= 2;\
513                         h->mvd_cache[list][idx][1] /= 2;\
514                     }
515                     MAP_MVS
516 #undef MAP_F2F
517                 }else{
518 #define MAP_F2F(idx, mb_type)\
519                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
520                         h->ref_cache[list][idx] >>= 1;\
521                         h->mv_cache[list][idx][1] <<= 1;\
522                         h->mvd_cache[list][idx][1] <<= 1;\
523                     }
524                     MAP_MVS
525 #undef MAP_F2F
526                 }
527             }
528         }
529     }
530 #endif
531
532     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
533 }
534
535 static inline void write_back_intra_pred_mode(H264Context *h){
536     const int mb_xy= h->mb_xy;
537
538     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
539     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
540     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
541     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
542     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
543     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
544     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
545 }
546
547 /**
548  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
549  */
550 static inline int check_intra4x4_pred_mode(H264Context *h){
551     MpegEncContext * const s = &h->s;
552     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
553     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
554     int i;
555
556     if(!(h->top_samples_available&0x8000)){
557         for(i=0; i<4; i++){
558             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
559             if(status<0){
560                 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);
561                 return -1;
562             } else if(status){
563                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
564             }
565         }
566     }
567
568     if(!(h->left_samples_available&0x8000)){
569         for(i=0; i<4; i++){
570             int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
571             if(status<0){
572                 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);
573                 return -1;
574             } else if(status){
575                 h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
576             }
577         }
578     }
579
580     return 0;
581 } //FIXME cleanup like next
582
583 /**
584  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
585  */
586 static inline int check_intra_pred_mode(H264Context *h, int mode){
587     MpegEncContext * const s = &h->s;
588     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
589     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
590
591     if(mode > 6U) {
592         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);
593         return -1;
594     }
595
596     if(!(h->top_samples_available&0x8000)){
597         mode= top[ mode ];
598         if(mode<0){
599             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);
600             return -1;
601         }
602     }
603
604     if(!(h->left_samples_available&0x8000)){
605         mode= left[ mode ];
606         if(mode<0){
607             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);
608             return -1;
609         }
610     }
611
612     return mode;
613 }
614
615 /**
616  * gets the predicted intra4x4 prediction mode.
617  */
618 static inline int pred_intra_mode(H264Context *h, int n){
619     const int index8= scan8[n];
620     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
621     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
622     const int min= FFMIN(left, top);
623
624     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
625
626     if(min<0) return DC_PRED;
627     else      return min;
628 }
629
630 static inline void write_back_non_zero_count(H264Context *h){
631     const int mb_xy= h->mb_xy;
632
633     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
634     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
635     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
636     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
637     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
638     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
639     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
640
641     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
642     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
643     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
644
645     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
646     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
647     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
648 }
649
650 /**
651  * gets the predicted number of non-zero coefficients.
652  * @param n block index
653  */
654 static inline int pred_non_zero_count(H264Context *h, int n){
655     const int index8= scan8[n];
656     const int left= h->non_zero_count_cache[index8 - 1];
657     const int top = h->non_zero_count_cache[index8 - 8];
658     int i= left + top;
659
660     if(i<64) i= (i+1)>>1;
661
662     tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);
663
664     return i&31;
665 }
666
667 static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
668     const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
669     MpegEncContext *s = &h->s;
670
671     /* there is no consistent mapping of mvs to neighboring locations that will
672      * make mbaff happy, so we can't move all this logic to fill_caches */
673     if(FRAME_MBAFF){
674         const uint32_t *mb_types = s->current_picture_ptr->mb_type;
675         const int16_t *mv;
676         *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
677         *C = h->mv_cache[list][scan8[0]-2];
678
679         if(!MB_FIELD
680            && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
681             int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
682             if(IS_INTERLACED(mb_types[topright_xy])){
683 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
684                 const int x4 = X4, y4 = Y4;\
685                 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
686                 if(!USES_LIST(mb_type,list))\
687                     return LIST_NOT_USED;\
688                 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
689                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
690                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
691                 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
692
693                 SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
694             }
695         }
696         if(topright_ref == PART_NOT_AVAILABLE
697            && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
698            && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
699             if(!MB_FIELD
700                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
701                 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
702             }
703             if(MB_FIELD
704                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
705                && i >= scan8[0]+8){
706                 // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
707                 SET_DIAG_MV(/2, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
708             }
709         }
710 #undef SET_DIAG_MV
711     }
712
713     if(topright_ref != PART_NOT_AVAILABLE){
714         *C= h->mv_cache[list][ i - 8 + part_width ];
715         return topright_ref;
716     }else{
717         tprintf(s->avctx, "topright MV not available\n");
718
719         *C= h->mv_cache[list][ i - 8 - 1 ];
720         return h->ref_cache[list][ i - 8 - 1 ];
721     }
722 }
723
724 /**
725  * gets the predicted MV.
726  * @param n the block index
727  * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
728  * @param mx the x component of the predicted motion vector
729  * @param my the y component of the predicted motion vector
730  */
731 static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
732     const int index8= scan8[n];
733     const int top_ref=      h->ref_cache[list][ index8 - 8 ];
734     const int left_ref=     h->ref_cache[list][ index8 - 1 ];
735     const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
736     const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
737     const int16_t * C;
738     int diagonal_ref, match_count;
739
740     assert(part_width==1 || part_width==2 || part_width==4);
741
742 /* mv_cache
743   B . . A T T T T
744   U . . L . . , .
745   U . . L . . . .
746   U . . L . . , .
747   . . . L . . . .
748 */
749
750     diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
751     match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
752     tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
753     if(match_count > 1){ //most common
754         *mx= mid_pred(A[0], B[0], C[0]);
755         *my= mid_pred(A[1], B[1], C[1]);
756     }else if(match_count==1){
757         if(left_ref==ref){
758             *mx= A[0];
759             *my= A[1];
760         }else if(top_ref==ref){
761             *mx= B[0];
762             *my= B[1];
763         }else{
764             *mx= C[0];
765             *my= C[1];
766         }
767     }else{
768         if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
769             *mx= A[0];
770             *my= A[1];
771         }else{
772             *mx= mid_pred(A[0], B[0], C[0]);
773             *my= mid_pred(A[1], B[1], C[1]);
774         }
775     }
776
777     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);
778 }
779
780 /**
781  * gets the directionally predicted 16x8 MV.
782  * @param n the block index
783  * @param mx the x component of the predicted motion vector
784  * @param my the y component of the predicted motion vector
785  */
786 static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
787     if(n==0){
788         const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
789         const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
790
791         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);
792
793         if(top_ref == ref){
794             *mx= B[0];
795             *my= B[1];
796             return;
797         }
798     }else{
799         const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
800         const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];
801
802         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);
803
804         if(left_ref == ref){
805             *mx= A[0];
806             *my= A[1];
807             return;
808         }
809     }
810
811     //RARE
812     pred_motion(h, n, 4, list, ref, mx, my);
813 }
814
815 /**
816  * gets the directionally predicted 8x16 MV.
817  * @param n the block index
818  * @param mx the x component of the predicted motion vector
819  * @param my the y component of the predicted motion vector
820  */
821 static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
822     if(n==0){
823         const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
824         const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];
825
826         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);
827
828         if(left_ref == ref){
829             *mx= A[0];
830             *my= A[1];
831             return;
832         }
833     }else{
834         const int16_t * C;
835         int diagonal_ref;
836
837         diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);
838
839         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);
840
841         if(diagonal_ref == ref){
842             *mx= C[0];
843             *my= C[1];
844             return;
845         }
846     }
847
848     //RARE
849     pred_motion(h, n, 2, list, ref, mx, my);
850 }
851
852 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
853     const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
854     const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];
855
856     tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);
857
858     if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
859        || (top_ref == 0  && *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ] == 0)
860        || (left_ref == 0 && *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ] == 0)){
861
862         *mx = *my = 0;
863         return;
864     }
865
866     pred_motion(h, 0, 4, 0, 0, mx, my);
867
868     return;
869 }
870
871 static inline void direct_dist_scale_factor(H264Context * const h){
872     MpegEncContext * const s = &h->s;
873     const int poc = h->s.current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
874     const int poc1 = h->ref_list[1][0].poc;
875     int i;
876     for(i=0; i<h->ref_count[0]; i++){
877         int poc0 = h->ref_list[0][i].poc;
878         int td = av_clip(poc1 - poc0, -128, 127);
879         if(td == 0 || h->ref_list[0][i].long_ref){
880             h->dist_scale_factor[i] = 256;
881         }else{
882             int tb = av_clip(poc - poc0, -128, 127);
883             int tx = (16384 + (FFABS(td) >> 1)) / td;
884             h->dist_scale_factor[i] = av_clip((tb*tx + 32) >> 6, -1024, 1023);
885         }
886     }
887     if(FRAME_MBAFF){
888         for(i=0; i<h->ref_count[0]; i++){
889             h->dist_scale_factor_field[2*i] =
890             h->dist_scale_factor_field[2*i+1] = h->dist_scale_factor[i];
891         }
892     }
893 }
894 static inline void direct_ref_list_init(H264Context * const h){
895     MpegEncContext * const s = &h->s;
896     Picture * const ref1 = &h->ref_list[1][0];
897     Picture * const cur = s->current_picture_ptr;
898     int list, i, j;
899     int sidx= s->picture_structure&1;
900     int ref1sidx= ref1->reference&1;
901     for(list=0; list<2; list++){
902         cur->ref_count[sidx][list] = h->ref_count[list];
903         for(j=0; j<h->ref_count[list]; j++)
904             cur->ref_poc[sidx][list][j] = 4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3);
905     }
906     if(s->picture_structure == PICT_FRAME){
907         memcpy(cur->ref_count[0], cur->ref_count[1], sizeof(cur->ref_count[0]));
908         memcpy(cur->ref_poc  [0], cur->ref_poc  [1], sizeof(cur->ref_poc  [0]));
909     }
910     if(cur->pict_type != FF_B_TYPE || h->direct_spatial_mv_pred)
911         return;
912     for(list=0; list<2; list++){
913         for(i=0; i<ref1->ref_count[ref1sidx][list]; i++){
914             int poc = ref1->ref_poc[ref1sidx][list][i];
915             if(((poc&3) == 3) != (s->picture_structure == PICT_FRAME))
916                 poc= (poc&~3) + s->picture_structure;
917             h->map_col_to_list0[list][i] = 0; /* bogus; fills in for missing frames */
918             for(j=0; j<h->ref_count[list]; j++)
919                 if(4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3) == poc){
920                     h->map_col_to_list0[list][i] = j;
921                     break;
922                 }
923         }
924     }
925     if(FRAME_MBAFF){
926         for(list=0; list<2; list++){
927             for(i=0; i<ref1->ref_count[ref1sidx][list]; i++){
928                 j = h->map_col_to_list0[list][i];
929                 h->map_col_to_list0_field[list][2*i] = 2*j;
930                 h->map_col_to_list0_field[list][2*i+1] = 2*j+1;
931             }
932         }
933     }
934 }
935
936 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
937     MpegEncContext * const s = &h->s;
938     int b8_stride = h->b8_stride;
939     int b4_stride = h->b_stride;
940     int mb_xy = h->mb_xy;
941     int mb_type_col[2];
942     const int16_t (*l1mv0)[2], (*l1mv1)[2];
943     const int8_t *l1ref0, *l1ref1;
944     const int is_b8x8 = IS_8X8(*mb_type);
945     unsigned int sub_mb_type;
946     int i8, i4;
947
948 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
949
950     if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL
951         if(h->ref_list[1][0].reference == PICT_FRAME){   // AFL/AFR/FR/FL -> AFL
952             if(!IS_INTERLACED(*mb_type)){                //     AFR/FR    -> AFL
953                 int cur_poc = s->current_picture_ptr->poc;
954                 int *col_poc = h->ref_list[1]->field_poc;
955                 int col_parity = FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc);
956                 mb_xy= s->mb_x + ((s->mb_y&~1) + col_parity)*s->mb_stride;
957                 b8_stride = 0;
958             }
959         }else if(!(s->picture_structure & h->ref_list[1][0].reference)){// FL -> FL & differ parity
960             int fieldoff= 2*(h->ref_list[1][0].reference)-3;
961             mb_xy += s->mb_stride*fieldoff;
962         }
963         goto single_col;
964     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
965         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
966             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
967             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
968             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
969             b8_stride *= 3;
970             b4_stride *= 6;
971             //FIXME IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag
972             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
973                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
974                 && !is_b8x8){
975                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
976                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
977             }else{
978                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
979                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
980             }
981         }else{                                           //     AFR/FR    -> AFR/FR
982 single_col:
983             mb_type_col[0] =
984             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
985             if(IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag){
986                 /* FIXME save sub mb types from previous frames (or derive from MVs)
987                 * so we know exactly what block size to use */
988                 sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
989                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
990             }else if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
991                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
992                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
993             }else{
994                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
995                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
996             }
997         }
998     }
999
1000     l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
1001     l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
1002     l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
1003     l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
1004     if(!b8_stride){
1005         if(s->mb_y&1){
1006             l1ref0 += h->b8_stride;
1007             l1ref1 += h->b8_stride;
1008             l1mv0  +=  2*b4_stride;
1009             l1mv1  +=  2*b4_stride;
1010         }
1011     }
1012
1013     if(h->direct_spatial_mv_pred){
1014         int ref[2];
1015         int mv[2][2];
1016         int list;
1017
1018         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
1019
1020         /* ref = min(neighbors) */
1021         for(list=0; list<2; list++){
1022             int refa = h->ref_cache[list][scan8[0] - 1];
1023             int refb = h->ref_cache[list][scan8[0] - 8];
1024             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1025             if(refc == PART_NOT_AVAILABLE)
1026                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1027             ref[list] = FFMIN3((unsigned)refa, (unsigned)refb, (unsigned)refc);
1028             if(ref[list] < 0)
1029                 ref[list] = -1;
1030         }
1031
1032         if(ref[0] < 0 && ref[1] < 0){
1033             ref[0] = ref[1] = 0;
1034             mv[0][0] = mv[0][1] =
1035             mv[1][0] = mv[1][1] = 0;
1036         }else{
1037             for(list=0; list<2; list++){
1038                 if(ref[list] >= 0)
1039                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1040                 else
1041                     mv[list][0] = mv[list][1] = 0;
1042             }
1043         }
1044
1045         if(ref[1] < 0){
1046             if(!is_b8x8)
1047                 *mb_type &= ~MB_TYPE_L1;
1048             sub_mb_type &= ~MB_TYPE_L1;
1049         }else if(ref[0] < 0){
1050             if(!is_b8x8)
1051                 *mb_type &= ~MB_TYPE_L0;
1052             sub_mb_type &= ~MB_TYPE_L0;
1053         }
1054
1055         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1056             for(i8=0; i8<4; i8++){
1057                 int x8 = i8&1;
1058                 int y8 = i8>>1;
1059                 int xy8 = x8+y8*b8_stride;
1060                 int xy4 = 3*x8+y8*b4_stride;
1061                 int a=0, b=0;
1062
1063                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1064                     continue;
1065                 h->sub_mb_type[i8] = sub_mb_type;
1066
1067                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1068                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1069                 if(!IS_INTRA(mb_type_col[y8])
1070                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
1071                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
1072                     if(ref[0] > 0)
1073                         a= pack16to32(mv[0][0],mv[0][1]);
1074                     if(ref[1] > 0)
1075                         b= pack16to32(mv[1][0],mv[1][1]);
1076                 }else{
1077                     a= pack16to32(mv[0][0],mv[0][1]);
1078                     b= pack16to32(mv[1][0],mv[1][1]);
1079                 }
1080                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
1081                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
1082             }
1083         }else if(IS_16X16(*mb_type)){
1084             int a=0, b=0;
1085
1086             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1087             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1088             if(!IS_INTRA(mb_type_col[0])
1089                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1090                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1091                        && (h->x264_build>33 || !h->x264_build)))){
1092                 if(ref[0] > 0)
1093                     a= pack16to32(mv[0][0],mv[0][1]);
1094                 if(ref[1] > 0)
1095                     b= pack16to32(mv[1][0],mv[1][1]);
1096             }else{
1097                 a= pack16to32(mv[0][0],mv[0][1]);
1098                 b= pack16to32(mv[1][0],mv[1][1]);
1099             }
1100             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1101             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1102         }else{
1103             for(i8=0; i8<4; i8++){
1104                 const int x8 = i8&1;
1105                 const int y8 = i8>>1;
1106
1107                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1108                     continue;
1109                 h->sub_mb_type[i8] = sub_mb_type;
1110
1111                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1112                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1113                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1114                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1115
1116                 /* col_zero_flag */
1117                 if(!IS_INTRA(mb_type_col[0]) && (   l1ref0[x8 + y8*b8_stride] == 0
1118                                               || (l1ref0[x8 + y8*b8_stride] < 0 && l1ref1[x8 + y8*b8_stride] == 0
1119                                                   && (h->x264_build>33 || !h->x264_build)))){
1120                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*b8_stride] == 0 ? l1mv0 : l1mv1;
1121                     if(IS_SUB_8X8(sub_mb_type)){
1122                         const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1123                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1124                             if(ref[0] == 0)
1125                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1126                             if(ref[1] == 0)
1127                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1128                         }
1129                     }else
1130                     for(i4=0; i4<4; i4++){
1131                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1132                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1133                             if(ref[0] == 0)
1134                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1135                             if(ref[1] == 0)
1136                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1137                         }
1138                     }
1139                 }
1140             }
1141         }
1142     }else{ /* direct temporal mv pred */
1143         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1144         const int *dist_scale_factor = h->dist_scale_factor;
1145
1146         if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
1147             map_col_to_list0[0] = h->map_col_to_list0_field[0];
1148             map_col_to_list0[1] = h->map_col_to_list0_field[1];
1149             dist_scale_factor = h->dist_scale_factor_field;
1150         }
1151         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1152             /* FIXME assumes direct_8x8_inference == 1 */
1153             int y_shift  = 2*!IS_INTERLACED(*mb_type);
1154             int ref_shift= FRAME_MBAFF ? y_shift : 1;
1155
1156             for(i8=0; i8<4; i8++){
1157                 const int x8 = i8&1;
1158                 const int y8 = i8>>1;
1159                 int ref0, scale;
1160                 const int16_t (*l1mv)[2]= l1mv0;
1161
1162                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1163                     continue;
1164                 h->sub_mb_type[i8] = sub_mb_type;
1165
1166                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1167                 if(IS_INTRA(mb_type_col[y8])){
1168                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1169                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1170                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1171                     continue;
1172                 }
1173
1174                 ref0 = l1ref0[x8 + y8*b8_stride];
1175                 if(ref0 >= 0)
1176                     ref0 = map_col_to_list0[0][ref0*2>>ref_shift];
1177                 else{
1178                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride]*2>>ref_shift];
1179                     l1mv= l1mv1;
1180                 }
1181                 scale = dist_scale_factor[ref0];
1182                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1183
1184                 {
1185                     const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride];
1186                     int my_col = (mv_col[1]<<y_shift)/2;
1187                     int mx = (scale * mv_col[0] + 128) >> 8;
1188                     int my = (scale * my_col + 128) >> 8;
1189                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1190                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1191                 }
1192             }
1193             return;
1194         }
1195
1196         /* one-to-one mv scaling */
1197
1198         if(IS_16X16(*mb_type)){
1199             int ref, mv0, mv1;
1200
1201             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1202             if(IS_INTRA(mb_type_col[0])){
1203                 ref=mv0=mv1=0;
1204             }else{
1205                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0]]
1206                                                 : map_col_to_list0[1][l1ref1[0]];
1207                 const int scale = dist_scale_factor[ref0];
1208                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1209                 int mv_l0[2];
1210                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1211                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1212                 ref= ref0;
1213                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1214                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1215             }
1216             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1217             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1218             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1219         }else{
1220             for(i8=0; i8<4; i8++){
1221                 const int x8 = i8&1;
1222                 const int y8 = i8>>1;
1223                 int ref0, scale;
1224                 const int16_t (*l1mv)[2]= l1mv0;
1225
1226                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1227                     continue;
1228                 h->sub_mb_type[i8] = sub_mb_type;
1229                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1230                 if(IS_INTRA(mb_type_col[0])){
1231                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1232                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1233                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1234                     continue;
1235                 }
1236
1237                 ref0 = l1ref0[x8 + y8*b8_stride];
1238                 if(ref0 >= 0)
1239                     ref0 = map_col_to_list0[0][ref0];
1240                 else{
1241                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride]];
1242                     l1mv= l1mv1;
1243                 }
1244                 scale = dist_scale_factor[ref0];
1245
1246                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1247                 if(IS_SUB_8X8(sub_mb_type)){
1248                     const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1249                     int mx = (scale * mv_col[0] + 128) >> 8;
1250                     int my = (scale * mv_col[1] + 128) >> 8;
1251                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1252                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1253                 }else
1254                 for(i4=0; i4<4; i4++){
1255                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1256                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1257                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1258                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1259                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1260                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1261                 }
1262             }
1263         }
1264     }
1265 }
1266
1267 static inline void write_back_motion(H264Context *h, int mb_type){
1268     MpegEncContext * const s = &h->s;
1269     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1270     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1271     int list;
1272
1273     if(!USES_LIST(mb_type, 0))
1274         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1275
1276     for(list=0; list<h->list_count; list++){
1277         int y;
1278         if(!USES_LIST(mb_type, list))
1279             continue;
1280
1281         for(y=0; y<4; y++){
1282             *(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];
1283             *(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];
1284         }
1285         if( h->pps.cabac ) {
1286             if(IS_SKIP(mb_type))
1287                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1288             else
1289             for(y=0; y<4; y++){
1290                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1291                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1292             }
1293         }
1294
1295         {
1296             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1297             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1298             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1299             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1300             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1301         }
1302     }
1303
1304     if(h->slice_type_nos == FF_B_TYPE && h->pps.cabac){
1305         if(IS_8X8(mb_type)){
1306             uint8_t *direct_table = &h->direct_table[b8_xy];
1307             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1308             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1309             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1310         }
1311     }
1312 }
1313
1314 /**
1315  * Decodes a network abstraction layer unit.
1316  * @param consumed is the number of bytes used as input
1317  * @param length is the length of the array
1318  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1319  * @returns decoded bytes, might be src+1 if no escapes
1320  */
1321 static const uint8_t *decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
1322     int i, si, di;
1323     uint8_t *dst;
1324     int bufidx;
1325
1326 //    src[0]&0x80;                //forbidden bit
1327     h->nal_ref_idc= src[0]>>5;
1328     h->nal_unit_type= src[0]&0x1F;
1329
1330     src++; length--;
1331 #if 0
1332     for(i=0; i<length; i++)
1333         printf("%2X ", src[i]);
1334 #endif
1335     for(i=0; i+1<length; i+=2){
1336         if(src[i]) continue;
1337         if(i>0 && src[i-1]==0) i--;
1338         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1339             if(src[i+2]!=3){
1340                 /* startcode, so we must be past the end */
1341                 length=i;
1342             }
1343             break;
1344         }
1345     }
1346
1347     if(i>=length-1){ //no escaped 0
1348         *dst_length= length;
1349         *consumed= length+1; //+1 for the header
1350         return src;
1351     }
1352
1353     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1354     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length);
1355     dst= h->rbsp_buffer[bufidx];
1356
1357     if (dst == NULL){
1358         return NULL;
1359     }
1360
1361 //printf("decoding esc\n");
1362     si=di=0;
1363     while(si<length){
1364         //remove escapes (very rare 1:2^22)
1365         if(si+2<length && src[si]==0 && src[si+1]==0 && src[si+2]<=3){
1366             if(src[si+2]==3){ //escape
1367                 dst[di++]= 0;
1368                 dst[di++]= 0;
1369                 si+=3;
1370                 continue;
1371             }else //next start code
1372                 break;
1373         }
1374
1375         dst[di++]= src[si++];
1376     }
1377
1378     *dst_length= di;
1379     *consumed= si + 1;//+1 for the header
1380 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1381     return dst;
1382 }
1383
1384 /**
1385  * identifies the exact end of the bitstream
1386  * @return the length of the trailing, or 0 if damaged
1387  */
1388 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src){
1389     int v= *src;
1390     int r;
1391
1392     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1393
1394     for(r=1; r<9; r++){
1395         if(v&1) return r;
1396         v>>=1;
1397     }
1398     return 0;
1399 }
1400
1401 /**
1402  * IDCT transforms the 16 dc values and dequantizes them.
1403  * @param qp quantization parameter
1404  */
1405 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1406 #define stride 16
1407     int i;
1408     int temp[16]; //FIXME check if this is a good idea
1409     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1410     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1411
1412 //memset(block, 64, 2*256);
1413 //return;
1414     for(i=0; i<4; i++){
1415         const int offset= y_offset[i];
1416         const int z0= block[offset+stride*0] + block[offset+stride*4];
1417         const int z1= block[offset+stride*0] - block[offset+stride*4];
1418         const int z2= block[offset+stride*1] - block[offset+stride*5];
1419         const int z3= block[offset+stride*1] + block[offset+stride*5];
1420
1421         temp[4*i+0]= z0+z3;
1422         temp[4*i+1]= z1+z2;
1423         temp[4*i+2]= z1-z2;
1424         temp[4*i+3]= z0-z3;
1425     }
1426
1427     for(i=0; i<4; i++){
1428         const int offset= x_offset[i];
1429         const int z0= temp[4*0+i] + temp[4*2+i];
1430         const int z1= temp[4*0+i] - temp[4*2+i];
1431         const int z2= temp[4*1+i] - temp[4*3+i];
1432         const int z3= temp[4*1+i] + temp[4*3+i];
1433
1434         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
1435         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1436         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1437         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1438     }
1439 }
1440
1441 #if 0
1442 /**
1443  * DCT transforms the 16 dc values.
1444  * @param qp quantization parameter ??? FIXME
1445  */
1446 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1447 //    const int qmul= dequant_coeff[qp][0];
1448     int i;
1449     int temp[16]; //FIXME check if this is a good idea
1450     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1451     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1452
1453     for(i=0; i<4; i++){
1454         const int offset= y_offset[i];
1455         const int z0= block[offset+stride*0] + block[offset+stride*4];
1456         const int z1= block[offset+stride*0] - block[offset+stride*4];
1457         const int z2= block[offset+stride*1] - block[offset+stride*5];
1458         const int z3= block[offset+stride*1] + block[offset+stride*5];
1459
1460         temp[4*i+0]= z0+z3;
1461         temp[4*i+1]= z1+z2;
1462         temp[4*i+2]= z1-z2;
1463         temp[4*i+3]= z0-z3;
1464     }
1465
1466     for(i=0; i<4; i++){
1467         const int offset= x_offset[i];
1468         const int z0= temp[4*0+i] + temp[4*2+i];
1469         const int z1= temp[4*0+i] - temp[4*2+i];
1470         const int z2= temp[4*1+i] - temp[4*3+i];
1471         const int z3= temp[4*1+i] + temp[4*3+i];
1472
1473         block[stride*0 +offset]= (z0 + z3)>>1;
1474         block[stride*2 +offset]= (z1 + z2)>>1;
1475         block[stride*8 +offset]= (z1 - z2)>>1;
1476         block[stride*10+offset]= (z0 - z3)>>1;
1477     }
1478 }
1479 #endif
1480
1481 #undef xStride
1482 #undef stride
1483
1484 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1485     const int stride= 16*2;
1486     const int xStride= 16;
1487     int a,b,c,d,e;
1488
1489     a= block[stride*0 + xStride*0];
1490     b= block[stride*0 + xStride*1];
1491     c= block[stride*1 + xStride*0];
1492     d= block[stride*1 + xStride*1];
1493
1494     e= a-b;
1495     a= a+b;
1496     b= c-d;
1497     c= c+d;
1498
1499     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1500     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1501     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1502     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1503 }
1504
1505 #if 0
1506 static void chroma_dc_dct_c(DCTELEM *block){
1507     const int stride= 16*2;
1508     const int xStride= 16;
1509     int a,b,c,d,e;
1510
1511     a= block[stride*0 + xStride*0];
1512     b= block[stride*0 + xStride*1];
1513     c= block[stride*1 + xStride*0];
1514     d= block[stride*1 + xStride*1];
1515
1516     e= a-b;
1517     a= a+b;
1518     b= c-d;
1519     c= c+d;
1520
1521     block[stride*0 + xStride*0]= (a+c);
1522     block[stride*0 + xStride*1]= (e+b);
1523     block[stride*1 + xStride*0]= (a-c);
1524     block[stride*1 + xStride*1]= (e-b);
1525 }
1526 #endif
1527
1528 /**
1529  * gets the chroma qp.
1530  */
1531 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1532     return h->pps.chroma_qp_table[t][qscale];
1533 }
1534
1535 //FIXME need to check that this does not overflow signed 32 bit for low qp, I am not sure, it's very close
1536 //FIXME check that gcc inlines this (and optimizes intra & separate_dc stuff away)
1537 static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int separate_dc){
1538     int i;
1539     const int * const quant_table= quant_coeff[qscale];
1540     const int bias= intra ? (1<<QUANT_SHIFT)/3 : (1<<QUANT_SHIFT)/6;
1541     const unsigned int threshold1= (1<<QUANT_SHIFT) - bias - 1;
1542     const unsigned int threshold2= (threshold1<<1);
1543     int last_non_zero;
1544
1545     if(separate_dc){
1546         if(qscale<=18){
1547             //avoid overflows
1548             const int dc_bias= intra ? (1<<(QUANT_SHIFT-2))/3 : (1<<(QUANT_SHIFT-2))/6;
1549             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT-2)) - dc_bias - 1;
1550             const unsigned int dc_threshold2= (dc_threshold1<<1);
1551
1552             int level= block[0]*quant_coeff[qscale+18][0];
1553             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1554                 if(level>0){
1555                     level= (dc_bias + level)>>(QUANT_SHIFT-2);
1556                     block[0]= level;
1557                 }else{
1558                     level= (dc_bias - level)>>(QUANT_SHIFT-2);
1559                     block[0]= -level;
1560                 }
1561 //                last_non_zero = i;
1562             }else{
1563                 block[0]=0;
1564             }
1565         }else{
1566             const int dc_bias= intra ? (1<<(QUANT_SHIFT+1))/3 : (1<<(QUANT_SHIFT+1))/6;
1567             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT+1)) - dc_bias - 1;
1568             const unsigned int dc_threshold2= (dc_threshold1<<1);
1569
1570             int level= block[0]*quant_table[0];
1571             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1572                 if(level>0){
1573                     level= (dc_bias + level)>>(QUANT_SHIFT+1);
1574                     block[0]= level;
1575                 }else{
1576                     level= (dc_bias - level)>>(QUANT_SHIFT+1);
1577                     block[0]= -level;
1578                 }
1579 //                last_non_zero = i;
1580             }else{
1581                 block[0]=0;
1582             }
1583         }
1584         last_non_zero= 0;
1585         i=1;
1586     }else{
1587         last_non_zero= -1;
1588         i=0;
1589     }
1590
1591     for(; i<16; i++){
1592         const int j= scantable[i];
1593         int level= block[j]*quant_table[j];
1594
1595 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
1596 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
1597         if(((unsigned)(level+threshold1))>threshold2){
1598             if(level>0){
1599                 level= (bias + level)>>QUANT_SHIFT;
1600                 block[j]= level;
1601             }else{
1602                 level= (bias - level)>>QUANT_SHIFT;
1603                 block[j]= -level;
1604             }
1605             last_non_zero = i;
1606         }else{
1607             block[j]=0;
1608         }
1609     }
1610
1611     return last_non_zero;
1612 }
1613
1614 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1615                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1616                            int src_x_offset, int src_y_offset,
1617                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1618     MpegEncContext * const s = &h->s;
1619     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1620     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1621     const int luma_xy= (mx&3) + ((my&3)<<2);
1622     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1623     uint8_t * src_cb, * src_cr;
1624     int extra_width= h->emu_edge_width;
1625     int extra_height= h->emu_edge_height;
1626     int emu=0;
1627     const int full_mx= mx>>2;
1628     const int full_my= my>>2;
1629     const int pic_width  = 16*s->mb_width;
1630     const int pic_height = 16*s->mb_height >> MB_FIELD;
1631
1632     if(!pic->data[0]) //FIXME this is unacceptable, some sensible error concealment must be done for missing reference frames
1633         return;
1634
1635     if(mx&7) extra_width -= 3;
1636     if(my&7) extra_height -= 3;
1637
1638     if(   full_mx < 0-extra_width
1639        || full_my < 0-extra_height
1640        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1641        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1642         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);
1643             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1644         emu=1;
1645     }
1646
1647     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1648     if(!square){
1649         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1650     }
1651
1652     if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1653
1654     if(MB_FIELD){
1655         // chroma offset when predicting from a field of opposite parity
1656         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1657         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1658     }
1659     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1660     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1661
1662     if(emu){
1663         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);
1664             src_cb= s->edge_emu_buffer;
1665     }
1666     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1667
1668     if(emu){
1669         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);
1670             src_cr= s->edge_emu_buffer;
1671     }
1672     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1673 }
1674
1675 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1676                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1677                            int x_offset, int y_offset,
1678                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1679                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1680                            int list0, int list1){
1681     MpegEncContext * const s = &h->s;
1682     qpel_mc_func *qpix_op=  qpix_put;
1683     h264_chroma_mc_func chroma_op= chroma_put;
1684
1685     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1686     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1687     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1688     x_offset += 8*s->mb_x;
1689     y_offset += 8*(s->mb_y >> MB_FIELD);
1690
1691     if(list0){
1692         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1693         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1694                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1695                            qpix_op, chroma_op);
1696
1697         qpix_op=  qpix_avg;
1698         chroma_op= chroma_avg;
1699     }
1700
1701     if(list1){
1702         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1703         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1704                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1705                            qpix_op, chroma_op);
1706     }
1707 }
1708
1709 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1710                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1711                            int x_offset, int y_offset,
1712                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1713                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1714                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1715                            int list0, int list1){
1716     MpegEncContext * const s = &h->s;
1717
1718     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1719     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1720     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1721     x_offset += 8*s->mb_x;
1722     y_offset += 8*(s->mb_y >> MB_FIELD);
1723
1724     if(list0 && list1){
1725         /* don't optimize for luma-only case, since B-frames usually
1726          * use implicit weights => chroma too. */
1727         uint8_t *tmp_cb = s->obmc_scratchpad;
1728         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1729         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1730         int refn0 = h->ref_cache[0][ scan8[n] ];
1731         int refn1 = h->ref_cache[1][ scan8[n] ];
1732
1733         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1734                     dest_y, dest_cb, dest_cr,
1735                     x_offset, y_offset, qpix_put, chroma_put);
1736         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1737                     tmp_y, tmp_cb, tmp_cr,
1738                     x_offset, y_offset, qpix_put, chroma_put);
1739
1740         if(h->use_weight == 2){
1741             int weight0 = h->implicit_weight[refn0][refn1];
1742             int weight1 = 64 - weight0;
1743             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1744             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1745             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1746         }else{
1747             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1748                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1749                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1750             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1751                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1752                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1753             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1754                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1755                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1756         }
1757     }else{
1758         int list = list1 ? 1 : 0;
1759         int refn = h->ref_cache[list][ scan8[n] ];
1760         Picture *ref= &h->ref_list[list][refn];
1761         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1762                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1763                     qpix_put, chroma_put);
1764
1765         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1766                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1767         if(h->use_weight_chroma){
1768             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1769                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1770             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1771                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1772         }
1773     }
1774 }
1775
1776 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1777                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1778                            int x_offset, int y_offset,
1779                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1780                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1781                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1782                            int list0, int list1){
1783     if((h->use_weight==2 && list0 && list1
1784         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1785        || h->use_weight==1)
1786         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1787                          x_offset, y_offset, qpix_put, chroma_put,
1788                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1789     else
1790         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1791                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1792 }
1793
1794 static inline void prefetch_motion(H264Context *h, int list){
1795     /* fetch pixels for estimated mv 4 macroblocks ahead
1796      * optimized for 64byte cache lines */
1797     MpegEncContext * const s = &h->s;
1798     const int refn = h->ref_cache[list][scan8[0]];
1799     if(refn >= 0){
1800         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1801         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1802         uint8_t **src= h->ref_list[list][refn].data;
1803         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1804         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1805         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1806         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1807     }
1808 }
1809
1810 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1811                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1812                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1813                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1814     MpegEncContext * const s = &h->s;
1815     const int mb_xy= h->mb_xy;
1816     const int mb_type= s->current_picture.mb_type[mb_xy];
1817
1818     assert(IS_INTER(mb_type));
1819
1820     prefetch_motion(h, 0);
1821
1822     if(IS_16X16(mb_type)){
1823         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1824                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1825                 &weight_op[0], &weight_avg[0],
1826                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1827     }else if(IS_16X8(mb_type)){
1828         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1829                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1830                 &weight_op[1], &weight_avg[1],
1831                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1832         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1833                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1834                 &weight_op[1], &weight_avg[1],
1835                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1836     }else if(IS_8X16(mb_type)){
1837         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1838                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1839                 &weight_op[2], &weight_avg[2],
1840                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1841         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1842                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1843                 &weight_op[2], &weight_avg[2],
1844                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1845     }else{
1846         int i;
1847
1848         assert(IS_8X8(mb_type));
1849
1850         for(i=0; i<4; i++){
1851             const int sub_mb_type= h->sub_mb_type[i];
1852             const int n= 4*i;
1853             int x_offset= (i&1)<<2;
1854             int y_offset= (i&2)<<1;
1855
1856             if(IS_SUB_8X8(sub_mb_type)){
1857                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1858                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1859                     &weight_op[3], &weight_avg[3],
1860                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1861             }else if(IS_SUB_8X4(sub_mb_type)){
1862                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1863                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1864                     &weight_op[4], &weight_avg[4],
1865                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1866                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1867                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1868                     &weight_op[4], &weight_avg[4],
1869                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1870             }else if(IS_SUB_4X8(sub_mb_type)){
1871                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1872                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1873                     &weight_op[5], &weight_avg[5],
1874                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1875                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1876                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1877                     &weight_op[5], &weight_avg[5],
1878                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1879             }else{
1880                 int j;
1881                 assert(IS_SUB_4X4(sub_mb_type));
1882                 for(j=0; j<4; j++){
1883                     int sub_x_offset= x_offset + 2*(j&1);
1884                     int sub_y_offset= y_offset +   (j&2);
1885                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1886                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1887                         &weight_op[6], &weight_avg[6],
1888                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1889                 }
1890             }
1891         }
1892     }
1893
1894     prefetch_motion(h, 1);
1895 }
1896
1897 static av_cold void decode_init_vlc(void){
1898     static int done = 0;
1899
1900     if (!done) {
1901         int i;
1902         int offset;
1903         done = 1;
1904
1905         chroma_dc_coeff_token_vlc.table = chroma_dc_coeff_token_vlc_table;
1906         chroma_dc_coeff_token_vlc.table_allocated = chroma_dc_coeff_token_vlc_table_size;
1907         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1908                  &chroma_dc_coeff_token_len [0], 1, 1,
1909                  &chroma_dc_coeff_token_bits[0], 1, 1,
1910                  INIT_VLC_USE_NEW_STATIC);
1911
1912         offset = 0;
1913         for(i=0; i<4; i++){
1914             coeff_token_vlc[i].table = coeff_token_vlc_tables+offset;
1915             coeff_token_vlc[i].table_allocated = coeff_token_vlc_tables_size[i];
1916             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1917                      &coeff_token_len [i][0], 1, 1,
1918                      &coeff_token_bits[i][0], 1, 1,
1919                      INIT_VLC_USE_NEW_STATIC);
1920             offset += coeff_token_vlc_tables_size[i];
1921         }
1922         /*
1923          * This is a one time safety check to make sure that
1924          * the packed static coeff_token_vlc table sizes
1925          * were initialized correctly.
1926          */
1927         assert(offset == sizeof(coeff_token_vlc_tables)/(sizeof(VLC_TYPE)*2));
1928
1929         for(i=0; i<3; i++){
1930             chroma_dc_total_zeros_vlc[i].table = chroma_dc_total_zeros_vlc_tables[i];
1931             chroma_dc_total_zeros_vlc[i].table_allocated = chroma_dc_total_zeros_vlc_tables_size;
1932             init_vlc(&chroma_dc_total_zeros_vlc[i],
1933                      CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1934                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1935                      &chroma_dc_total_zeros_bits[i][0], 1, 1,
1936                      INIT_VLC_USE_NEW_STATIC);
1937         }
1938         for(i=0; i<15; i++){
1939             total_zeros_vlc[i].table = total_zeros_vlc_tables[i];
1940             total_zeros_vlc[i].table_allocated = total_zeros_vlc_tables_size;
1941             init_vlc(&total_zeros_vlc[i],
1942                      TOTAL_ZEROS_VLC_BITS, 16,
1943                      &total_zeros_len [i][0], 1, 1,
1944                      &total_zeros_bits[i][0], 1, 1,
1945                      INIT_VLC_USE_NEW_STATIC);
1946         }
1947
1948         for(i=0; i<6; i++){
1949             run_vlc[i].table = run_vlc_tables[i];
1950             run_vlc[i].table_allocated = run_vlc_tables_size;
1951             init_vlc(&run_vlc[i],
1952                      RUN_VLC_BITS, 7,
1953                      &run_len [i][0], 1, 1,
1954                      &run_bits[i][0], 1, 1,
1955                      INIT_VLC_USE_NEW_STATIC);
1956         }
1957         run7_vlc.table = run7_vlc_table,
1958         run7_vlc.table_allocated = run7_vlc_table_size;
1959         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
1960                  &run_len [6][0], 1, 1,
1961                  &run_bits[6][0], 1, 1,
1962                  INIT_VLC_USE_NEW_STATIC);
1963     }
1964 }
1965
1966 static void free_tables(H264Context *h){
1967     int i;
1968     H264Context *hx;
1969     av_freep(&h->intra4x4_pred_mode);
1970     av_freep(&h->chroma_pred_mode_table);
1971     av_freep(&h->cbp_table);
1972     av_freep(&h->mvd_table[0]);
1973     av_freep(&h->mvd_table[1]);
1974     av_freep(&h->direct_table);
1975     av_freep(&h->non_zero_count);
1976     av_freep(&h->slice_table_base);
1977     h->slice_table= NULL;
1978
1979     av_freep(&h->mb2b_xy);
1980     av_freep(&h->mb2b8_xy);
1981
1982     for(i = 0; i < MAX_SPS_COUNT; i++)
1983         av_freep(h->sps_buffers + i);
1984
1985     for(i = 0; i < MAX_PPS_COUNT; i++)
1986         av_freep(h->pps_buffers + i);
1987
1988     for(i = 0; i < h->s.avctx->thread_count; i++) {
1989         hx = h->thread_context[i];
1990         if(!hx) continue;
1991         av_freep(&hx->top_borders[1]);
1992         av_freep(&hx->top_borders[0]);
1993         av_freep(&hx->s.obmc_scratchpad);
1994     }
1995 }
1996
1997 static void init_dequant8_coeff_table(H264Context *h){
1998     int i,q,x;
1999     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2000     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2001     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2002
2003     for(i=0; i<2; i++ ){
2004         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2005             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2006             break;
2007         }
2008
2009         for(q=0; q<52; q++){
2010             int shift = ff_div6[q];
2011             int idx = ff_rem6[q];
2012             for(x=0; x<64; x++)
2013                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2014                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2015                     h->pps.scaling_matrix8[i][x]) << shift;
2016         }
2017     }
2018 }
2019
2020 static void init_dequant4_coeff_table(H264Context *h){
2021     int i,j,q,x;
2022     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2023     for(i=0; i<6; i++ ){
2024         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2025         for(j=0; j<i; j++){
2026             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2027                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2028                 break;
2029             }
2030         }
2031         if(j<i)
2032             continue;
2033
2034         for(q=0; q<52; q++){
2035             int shift = ff_div6[q] + 2;
2036             int idx = ff_rem6[q];
2037             for(x=0; x<16; x++)
2038                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2039                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2040                     h->pps.scaling_matrix4[i][x]) << shift;
2041         }
2042     }
2043 }
2044
2045 static void init_dequant_tables(H264Context *h){
2046     int i,x;
2047     init_dequant4_coeff_table(h);
2048     if(h->pps.transform_8x8_mode)
2049         init_dequant8_coeff_table(h);
2050     if(h->sps.transform_bypass){
2051         for(i=0; i<6; i++)
2052             for(x=0; x<16; x++)
2053                 h->dequant4_coeff[i][0][x] = 1<<6;
2054         if(h->pps.transform_8x8_mode)
2055             for(i=0; i<2; i++)
2056                 for(x=0; x<64; x++)
2057                     h->dequant8_coeff[i][0][x] = 1<<6;
2058     }
2059 }
2060
2061
2062 /**
2063  * allocates tables.
2064  * needs width/height
2065  */
2066 static int alloc_tables(H264Context *h){
2067     MpegEncContext * const s = &h->s;
2068     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2069     int x,y;
2070
2071     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2072
2073     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2074     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(uint8_t))
2075     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2076
2077     CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2078     CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2079     CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2080     CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2081
2082     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(uint8_t));
2083     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2084
2085     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2086     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2087     for(y=0; y<s->mb_height; y++){
2088         for(x=0; x<s->mb_width; x++){
2089             const int mb_xy= x + y*s->mb_stride;
2090             const int b_xy = 4*x + 4*y*h->b_stride;
2091             const int b8_xy= 2*x + 2*y*h->b8_stride;
2092
2093             h->mb2b_xy [mb_xy]= b_xy;
2094             h->mb2b8_xy[mb_xy]= b8_xy;
2095         }
2096     }
2097
2098     s->obmc_scratchpad = NULL;
2099
2100     if(!h->dequant4_coeff[0])
2101         init_dequant_tables(h);
2102
2103     return 0;
2104 fail:
2105     free_tables(h);
2106     return -1;
2107 }
2108
2109 /**
2110  * Mimic alloc_tables(), but for every context thread.
2111  */
2112 static void clone_tables(H264Context *dst, H264Context *src){
2113     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2114     dst->non_zero_count           = src->non_zero_count;
2115     dst->slice_table              = src->slice_table;
2116     dst->cbp_table                = src->cbp_table;
2117     dst->mb2b_xy                  = src->mb2b_xy;
2118     dst->mb2b8_xy                 = src->mb2b8_xy;
2119     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2120     dst->mvd_table[0]             = src->mvd_table[0];
2121     dst->mvd_table[1]             = src->mvd_table[1];
2122     dst->direct_table             = src->direct_table;
2123
2124     dst->s.obmc_scratchpad = NULL;
2125     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2126 }
2127
2128 /**
2129  * Init context
2130  * Allocate buffers which are not shared amongst multiple threads.
2131  */
2132 static int context_init(H264Context *h){
2133     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2134     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2135
2136     return 0;
2137 fail:
2138     return -1; // free_tables will clean up for us
2139 }
2140
2141 static av_cold void common_init(H264Context *h){
2142     MpegEncContext * const s = &h->s;
2143
2144     s->width = s->avctx->width;
2145     s->height = s->avctx->height;
2146     s->codec_id= s->avctx->codec->id;
2147
2148     ff_h264_pred_init(&h->hpc, s->codec_id);
2149
2150     h->dequant_coeff_pps= -1;
2151     s->unrestricted_mv=1;
2152     s->decode=1; //FIXME
2153
2154     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2155     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2156 }
2157
2158 static av_cold int decode_init(AVCodecContext *avctx){
2159     H264Context *h= avctx->priv_data;
2160     MpegEncContext * const s = &h->s;
2161
2162     MPV_decode_defaults(s);
2163
2164     s->avctx = avctx;
2165     common_init(h);
2166
2167     s->out_format = FMT_H264;
2168     s->workaround_bugs= avctx->workaround_bugs;
2169
2170     // set defaults
2171 //    s->decode_mb= ff_h263_decode_mb;
2172     s->quarter_sample = 1;
2173     s->low_delay= 1;
2174
2175     if(avctx->codec_id == CODEC_ID_SVQ3)
2176         avctx->pix_fmt= PIX_FMT_YUVJ420P;
2177     else
2178         avctx->pix_fmt= PIX_FMT_YUV420P;
2179
2180     decode_init_vlc();
2181
2182     if(avctx->extradata_size > 0 && avctx->extradata &&
2183        *(char *)avctx->extradata == 1){
2184         h->is_avc = 1;
2185         h->got_avcC = 0;
2186     } else {
2187         h->is_avc = 0;
2188     }
2189
2190     h->thread_context[0] = h;
2191     h->outputed_poc = INT_MIN;
2192     return 0;
2193 }
2194
2195 static int frame_start(H264Context *h){
2196     MpegEncContext * const s = &h->s;
2197     int i;
2198
2199     if(MPV_frame_start(s, s->avctx) < 0)
2200         return -1;
2201     ff_er_frame_start(s);
2202     /*
2203      * MPV_frame_start uses pict_type to derive key_frame.
2204      * This is incorrect for H.264; IDR markings must be used.
2205      * Zero here; IDR markings per slice in frame or fields are ORed in later.
2206      * See decode_nal_units().
2207      */
2208     s->current_picture_ptr->key_frame= 0;
2209
2210     assert(s->linesize && s->uvlinesize);
2211
2212     for(i=0; i<16; i++){
2213         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2214         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2215     }
2216     for(i=0; i<4; i++){
2217         h->block_offset[16+i]=
2218         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2219         h->block_offset[24+16+i]=
2220         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2221     }
2222
2223     /* can't be in alloc_tables because linesize isn't known there.
2224      * FIXME: redo bipred weight to not require extra buffer? */
2225     for(i = 0; i < s->avctx->thread_count; i++)
2226         if(!h->thread_context[i]->s.obmc_scratchpad)
2227             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2228
2229     /* some macroblocks will be accessed before they're available */
2230     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2231         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(uint8_t));
2232
2233 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2234
2235     // We mark the current picture as non-reference after allocating it, so
2236     // that if we break out due to an error it can be released automatically
2237     // in the next MPV_frame_start().
2238     // SVQ3 as well as most other codecs have only last/next/current and thus
2239     // get released even with set reference, besides SVQ3 and others do not
2240     // mark frames as reference later "naturally".
2241     if(s->codec_id != CODEC_ID_SVQ3)
2242         s->current_picture_ptr->reference= 0;
2243
2244     s->current_picture_ptr->field_poc[0]=
2245     s->current_picture_ptr->field_poc[1]= INT_MAX;
2246     assert(s->current_picture_ptr->long_ref==0);
2247
2248     return 0;
2249 }
2250
2251 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){
2252     MpegEncContext * const s = &h->s;
2253     int i;
2254     int step    = 1;
2255     int offset  = 1;
2256     int uvoffset= 1;
2257     int top_idx = 1;
2258     int skiplast= 0;
2259
2260     src_y  -=   linesize;
2261     src_cb -= uvlinesize;
2262     src_cr -= uvlinesize;
2263
2264     if(!simple && FRAME_MBAFF){
2265         if(s->mb_y&1){
2266             offset  = MB_MBAFF ? 1 : 17;
2267             uvoffset= MB_MBAFF ? 1 : 9;
2268             if(!MB_MBAFF){
2269                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 0)= *(uint64_t*)(src_y +  15*linesize);
2270                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 8)= *(uint64_t*)(src_y +8+15*linesize);
2271                 if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2272                     *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+7*uvlinesize);
2273                     *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+7*uvlinesize);
2274                 }
2275             }
2276         }else{
2277             if(!MB_MBAFF){
2278                 h->left_border[0]= h->top_borders[0][s->mb_x][15];
2279                 if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2280                     h->left_border[34   ]= h->top_borders[0][s->mb_x][16+7  ];
2281                     h->left_border[34+18]= h->top_borders[0][s->mb_x][16+8+7];
2282                 }
2283                 skiplast= 1;
2284             }
2285             offset  =
2286             uvoffset=
2287             top_idx = MB_MBAFF ? 0 : 1;
2288         }
2289         step= MB_MBAFF ? 2 : 1;
2290     }
2291
2292     // There are two lines saved, the line above the the top macroblock of a pair,
2293     // and the line above the bottom macroblock
2294     h->left_border[offset]= h->top_borders[top_idx][s->mb_x][15];
2295     for(i=1; i<17 - skiplast; i++){
2296         h->left_border[offset+i*step]= src_y[15+i*  linesize];
2297     }
2298
2299     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2300     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2301
2302     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2303         h->left_border[uvoffset+34   ]= h->top_borders[top_idx][s->mb_x][16+7];
2304         h->left_border[uvoffset+34+18]= h->top_borders[top_idx][s->mb_x][24+7];
2305         for(i=1; i<9 - skiplast; i++){
2306             h->left_border[uvoffset+34   +i*step]= src_cb[7+i*uvlinesize];
2307             h->left_border[uvoffset+34+18+i*step]= src_cr[7+i*uvlinesize];
2308         }
2309         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2310         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2311     }
2312 }
2313
2314 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){
2315     MpegEncContext * const s = &h->s;
2316     int temp8, i;
2317     uint64_t temp64;
2318     int deblock_left;
2319     int deblock_top;
2320     int mb_xy;
2321     int step    = 1;
2322     int offset  = 1;
2323     int uvoffset= 1;
2324     int top_idx = 1;
2325
2326     if(!simple && FRAME_MBAFF){
2327         if(s->mb_y&1){
2328             offset  = MB_MBAFF ? 1 : 17;
2329             uvoffset= MB_MBAFF ? 1 : 9;
2330         }else{
2331             offset  =
2332             uvoffset=
2333             top_idx = MB_MBAFF ? 0 : 1;
2334         }
2335         step= MB_MBAFF ? 2 : 1;
2336     }
2337
2338     if(h->deblocking_filter == 2) {
2339         mb_xy = h->mb_xy;
2340         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2341         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2342     } else {
2343         deblock_left = (s->mb_x > 0);
2344         deblock_top =  (s->mb_y > 0);
2345     }
2346
2347     src_y  -=   linesize + 1;
2348     src_cb -= uvlinesize + 1;
2349     src_cr -= uvlinesize + 1;
2350
2351 #define XCHG(a,b,t,xchg)\
2352 t= a;\
2353 if(xchg)\
2354     a= b;\
2355 b= t;
2356
2357     if(deblock_left){
2358         for(i = !deblock_top; i<16; i++){
2359             XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, xchg);
2360         }
2361         XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, 1);
2362     }
2363
2364     if(deblock_top){
2365         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2366         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2367         if(s->mb_x+1 < s->mb_width){
2368             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2369         }
2370     }
2371
2372     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2373         if(deblock_left){
2374             for(i = !deblock_top; i<8; i++){
2375                 XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, xchg);
2376                 XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, xchg);
2377             }
2378             XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, 1);
2379             XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, 1);
2380         }
2381         if(deblock_top){
2382             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2383             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2384         }
2385     }
2386 }
2387
2388 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2389     MpegEncContext * const s = &h->s;
2390     const int mb_x= s->mb_x;
2391     const int mb_y= s->mb_y;
2392     const int mb_xy= h->mb_xy;
2393     const int mb_type= s->current_picture.mb_type[mb_xy];
2394     uint8_t  *dest_y, *dest_cb, *dest_cr;
2395     int linesize, uvlinesize /*dct_offset*/;
2396     int i;
2397     int *block_offset = &h->block_offset[0];
2398     const unsigned int bottom = mb_y & 1;
2399     const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass), is_h264 = (simple || s->codec_id == CODEC_ID_H264);
2400     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2401     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2402
2403     dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2404     dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2405     dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2406
2407     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2408     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2409
2410     if (!simple && MB_FIELD) {
2411         linesize   = h->mb_linesize   = s->linesize * 2;
2412         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2413         block_offset = &h->block_offset[24];
2414         if(mb_y&1){ //FIXME move out of this function?
2415             dest_y -= s->linesize*15;
2416             dest_cb-= s->uvlinesize*7;
2417             dest_cr-= s->uvlinesize*7;
2418         }
2419         if(FRAME_MBAFF) {
2420             int list;
2421             for(list=0; list<h->list_count; list++){
2422                 if(!USES_LIST(mb_type, list))
2423                     continue;
2424                 if(IS_16X16(mb_type)){
2425                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2426                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2427                 }else{
2428                     for(i=0; i<16; i+=4){
2429                         //FIXME can refs be smaller than 8x8 when !direct_8x8_inference ?
2430                         int ref = h->ref_cache[list][scan8[i]];
2431                         if(ref >= 0)
2432                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2433                     }
2434                 }
2435             }
2436         }
2437     } else {
2438         linesize   = h->mb_linesize   = s->linesize;
2439         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2440 //        dct_offset = s->linesize * 16;
2441     }
2442
2443     if(transform_bypass){
2444         idct_dc_add =
2445         idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2446     }else if(IS_8x8DCT(mb_type)){
2447         idct_dc_add = s->dsp.h264_idct8_dc_add;
2448         idct_add = s->dsp.h264_idct8_add;
2449     }else{
2450         idct_dc_add = s->dsp.h264_idct_dc_add;
2451         idct_add = s->dsp.h264_idct_add;
2452     }
2453
2454     if (!simple && IS_INTRA_PCM(mb_type)) {
2455         for (i=0; i<16; i++) {
2456             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
2457         }
2458         for (i=0; i<8; i++) {
2459             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
2460             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
2461         }
2462     } else {
2463         if(IS_INTRA(mb_type)){
2464             if(h->deblocking_filter)
2465                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2466
2467             if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2468                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2469                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2470             }
2471
2472             if(IS_INTRA4x4(mb_type)){
2473                 if(simple || !s->encoding){
2474                     if(IS_8x8DCT(mb_type)){
2475                         for(i=0; i<16; i+=4){
2476                             uint8_t * const ptr= dest_y + block_offset[i];
2477                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2478                             const int nnz = h->non_zero_count_cache[ scan8[i] ];
2479                             h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2480                                                    (h->topright_samples_available<<i)&0x4000, linesize);
2481                             if(nnz){
2482                                 if(nnz == 1 && h->mb[i*16])
2483                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2484                                 else
2485                                     idct_add(ptr, h->mb + i*16, linesize);
2486                             }
2487                         }
2488                     }else
2489                     for(i=0; i<16; i++){
2490                         uint8_t * const ptr= dest_y + block_offset[i];
2491                         uint8_t *topright;
2492                         const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2493                         int nnz, tr;
2494
2495                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2496                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2497                             assert(mb_y || linesize <= block_offset[i]);
2498                             if(!topright_avail){
2499                                 tr= ptr[3 - linesize]*0x01010101;
2500                                 topright= (uint8_t*) &tr;
2501                             }else
2502                                 topright= ptr + 4 - linesize;
2503                         }else
2504                             topright= NULL;
2505
2506                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2507                         nnz = h->non_zero_count_cache[ scan8[i] ];
2508                         if(nnz){
2509                             if(is_h264){
2510                                 if(nnz == 1 && h->mb[i*16])
2511                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2512                                 else
2513                                     idct_add(ptr, h->mb + i*16, linesize);
2514                             }else
2515                                 svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2516                         }
2517                     }
2518                 }
2519             }else{
2520                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2521                 if(is_h264){
2522                     if(!transform_bypass)
2523                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2524                 }else
2525                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2526             }
2527             if(h->deblocking_filter)
2528                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2529         }else if(is_h264){
2530             hl_motion(h, dest_y, dest_cb, dest_cr,
2531                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2532                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2533                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2534         }
2535
2536
2537         if(!IS_INTRA4x4(mb_type)){
2538             if(is_h264){
2539                 if(IS_INTRA16x16(mb_type)){
2540                     for(i=0; i<16; i++){
2541                         if(h->non_zero_count_cache[ scan8[i] ])
2542                             idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2543                         else if(h->mb[i*16])
2544                             idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2545                     }
2546                 }else{
2547                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2548                     for(i=0; i<16; i+=di){
2549                         int nnz = h->non_zero_count_cache[ scan8[i] ];
2550                         if(nnz){
2551                             if(nnz==1 && h->mb[i*16])
2552                                 idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2553                             else
2554                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2555                         }
2556                     }
2557                 }
2558             }else{
2559                 for(i=0; i<16; i++){
2560                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2561                         uint8_t * const ptr= dest_y + block_offset[i];
2562                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2563                     }
2564                 }
2565             }
2566         }
2567
2568         if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2569             uint8_t *dest[2] = {dest_cb, dest_cr};
2570             if(transform_bypass){
2571                 idct_add = idct_dc_add = s->dsp.add_pixels4;
2572             }else{
2573                 idct_add = s->dsp.h264_idct_add;
2574                 idct_dc_add = s->dsp.h264_idct_dc_add;
2575                 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]);
2576                 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]);
2577             }
2578             if(is_h264){
2579                 for(i=16; i<16+8; i++){
2580                     if(h->non_zero_count_cache[ scan8[i] ])
2581                         idct_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2582                     else if(h->mb[i*16])
2583                         idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2584                 }
2585             }else{
2586                 for(i=16; i<16+8; i++){
2587                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2588                         uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2589                         svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2590                     }
2591                 }
2592             }
2593         }
2594     }
2595     if(h->deblocking_filter) {
2596         backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2597         fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2598         h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2599         h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2600         if (!simple && FRAME_MBAFF) {
2601             filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2602         } else {
2603             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2604         }
2605     }
2606 }
2607
2608 /**
2609  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2610  */
2611 static void hl_decode_mb_simple(H264Context *h){
2612     hl_decode_mb_internal(h, 1);
2613 }
2614
2615 /**
2616  * Process a macroblock; this handles edge cases, such as interlacing.
2617  */
2618 static void av_noinline hl_decode_mb_complex(H264Context *h){
2619     hl_decode_mb_internal(h, 0);
2620 }
2621
2622 static void hl_decode_mb(H264Context *h){
2623     MpegEncContext * const s = &h->s;
2624     const int mb_xy= h->mb_xy;
2625     const int mb_type= s->current_picture.mb_type[mb_xy];
2626     int is_complex = FRAME_MBAFF || MB_FIELD || IS_INTRA_PCM(mb_type) || s->codec_id != CODEC_ID_H264 ||
2627                     (ENABLE_GRAY && (s->flags&CODEC_FLAG_GRAY)) || (ENABLE_H264_ENCODER && s->encoding) || ENABLE_SMALL;
2628
2629     if(ENABLE_H264_ENCODER && !s->decode)
2630         return;
2631
2632     if (is_complex)
2633         hl_decode_mb_complex(h);
2634     else hl_decode_mb_simple(h);
2635 }
2636
2637 static void pic_as_field(Picture *pic, const int parity){
2638     int i;
2639     for (i = 0; i < 4; ++i) {
2640         if (parity == PICT_BOTTOM_FIELD)
2641             pic->data[i] += pic->linesize[i];
2642         pic->reference = parity;
2643         pic->linesize[i] *= 2;
2644     }
2645     pic->poc= pic->field_poc[parity == PICT_BOTTOM_FIELD];
2646 }
2647
2648 static int split_field_copy(Picture *dest, Picture *src,
2649                             int parity, int id_add){
2650     int match = !!(src->reference & parity);
2651
2652     if (match) {
2653         *dest = *src;
2654         if(parity != PICT_FRAME){
2655             pic_as_field(dest, parity);
2656             dest->pic_id *= 2;
2657             dest->pic_id += id_add;
2658         }
2659     }
2660
2661     return match;
2662 }
2663
2664 static int build_def_list(Picture *def, Picture **in, int len, int is_long, int sel){
2665     int i[2]={0};
2666     int index=0;
2667
2668     while(i[0]<len || i[1]<len){
2669         while(i[0]<len && !(in[ i[0] ] && (in[ i[0] ]->reference & sel)))
2670             i[0]++;
2671         while(i[1]<len && !(in[ i[1] ] && (in[ i[1] ]->reference & (sel^3))))
2672             i[1]++;
2673         if(i[0] < len){
2674             in[ i[0] ]->pic_id= is_long ? i[0] : in[ i[0] ]->frame_num;
2675             split_field_copy(&def[index++], in[ i[0]++ ], sel  , 1);
2676         }
2677         if(i[1] < len){
2678             in[ i[1] ]->pic_id= is_long ? i[1] : in[ i[1] ]->frame_num;
2679             split_field_copy(&def[index++], in[ i[1]++ ], sel^3, 0);
2680         }
2681     }
2682
2683     return index;
2684 }
2685
2686 static int add_sorted(Picture **sorted, Picture **src, int len, int limit, int dir){
2687     int i, best_poc;
2688     int out_i= 0;
2689
2690     for(;;){
2691         best_poc= dir ? INT_MIN : INT_MAX;
2692
2693         for(i=0; i<len; i++){
2694             const int poc= src[i]->poc;
2695             if(((poc > limit) ^ dir) && ((poc < best_poc) ^ dir)){
2696                 best_poc= poc;
2697                 sorted[out_i]= src[i];
2698             }
2699         }
2700         if(best_poc == (dir ? INT_MIN : INT_MAX))
2701             break;
2702         limit= sorted[out_i++]->poc - dir;
2703     }
2704     return out_i;
2705 }
2706
2707 /**
2708  * fills the default_ref_list.
2709  */
2710 static int fill_default_ref_list(H264Context *h){
2711     MpegEncContext * const s = &h->s;
2712     int i, len;
2713
2714     if(h->slice_type_nos==FF_B_TYPE){
2715         Picture *sorted[32];
2716         int cur_poc, list;
2717         int lens[2];
2718
2719         if(FIELD_PICTURE)
2720             cur_poc= s->current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
2721         else
2722             cur_poc= s->current_picture_ptr->poc;
2723
2724         for(list= 0; list<2; list++){
2725             len= add_sorted(sorted    , h->short_ref, h->short_ref_count, cur_poc, 1^list);
2726             len+=add_sorted(sorted+len, h->short_ref, h->short_ref_count, cur_poc, 0^list);
2727             assert(len<=32);
2728             len= build_def_list(h->default_ref_list[list]    , sorted     , len, 0, s->picture_structure);
2729             len+=build_def_list(h->default_ref_list[list]+len, h->long_ref, 16 , 1, s->picture_structure);
2730             assert(len<=32);
2731
2732             if(len < h->ref_count[list])
2733                 memset(&h->default_ref_list[list][len], 0, sizeof(Picture)*(h->ref_count[list] - len));
2734             lens[list]= len;
2735         }
2736
2737         if(lens[0] == lens[1] && lens[1] > 1){
2738             for(i=0; h->default_ref_list[0][i].data[0] == h->default_ref_list[1][i].data[0] && i<lens[0]; i++);
2739             if(i == lens[0])
2740                 FFSWAP(Picture, h->default_ref_list[1][0], h->default_ref_list[1][1]);
2741         }
2742     }else{
2743         len = build_def_list(h->default_ref_list[0]    , h->short_ref, h->short_ref_count, 0, s->picture_structure);
2744         len+= build_def_list(h->default_ref_list[0]+len, h-> long_ref, 16                , 1, s->picture_structure);
2745         assert(len <= 32);
2746         if(len < h->ref_count[0])
2747             memset(&h->default_ref_list[0][len], 0, sizeof(Picture)*(h->ref_count[0] - len));
2748     }
2749 #ifdef TRACE
2750     for (i=0; i<h->ref_count[0]; i++) {
2751         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]);
2752     }
2753     if(h->slice_type_nos==FF_B_TYPE){
2754         for (i=0; i<h->ref_count[1]; i++) {
2755             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]);
2756         }
2757     }
2758 #endif
2759     return 0;
2760 }
2761
2762 static void print_short_term(H264Context *h);
2763 static void print_long_term(H264Context *h);
2764
2765 /**
2766  * Extract structure information about the picture described by pic_num in
2767  * the current decoding context (frame or field). Note that pic_num is
2768  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
2769  * @param pic_num picture number for which to extract structure information
2770  * @param structure one of PICT_XXX describing structure of picture
2771  *                      with pic_num
2772  * @return frame number (short term) or long term index of picture
2773  *         described by pic_num
2774  */
2775 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
2776     MpegEncContext * const s = &h->s;
2777
2778     *structure = s->picture_structure;
2779     if(FIELD_PICTURE){
2780         if (!(pic_num & 1))
2781             /* opposite field */
2782             *structure ^= PICT_FRAME;
2783         pic_num >>= 1;
2784     }
2785
2786     return pic_num;
2787 }
2788
2789 static int decode_ref_pic_list_reordering(H264Context *h){
2790     MpegEncContext * const s = &h->s;
2791     int list, index, pic_structure;
2792
2793     print_short_term(h);
2794     print_long_term(h);
2795
2796     for(list=0; list<h->list_count; list++){
2797         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
2798
2799         if(get_bits1(&s->gb)){
2800             int pred= h->curr_pic_num;
2801
2802             for(index=0; ; index++){
2803                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb(&s->gb);
2804                 unsigned int pic_id;
2805                 int i;
2806                 Picture *ref = NULL;
2807
2808                 if(reordering_of_pic_nums_idc==3)
2809                     break;
2810
2811                 if(index >= h->ref_count[list]){
2812                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
2813                     return -1;
2814                 }
2815
2816                 if(reordering_of_pic_nums_idc<3){
2817                     if(reordering_of_pic_nums_idc<2){
2818                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
2819                         int frame_num;
2820
2821                         if(abs_diff_pic_num > h->max_pic_num){
2822                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
2823                             return -1;
2824                         }
2825
2826                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
2827                         else                                pred+= abs_diff_pic_num;
2828                         pred &= h->max_pic_num - 1;
2829
2830                         frame_num = pic_num_extract(h, pred, &pic_structure);
2831
2832                         for(i= h->short_ref_count-1; i>=0; i--){
2833                             ref = h->short_ref[i];
2834                             assert(ref->reference);
2835                             assert(!ref->long_ref);
2836                             if(
2837                                    ref->frame_num == frame_num &&
2838                                    (ref->reference & pic_structure)
2839                               )
2840                                 break;
2841                         }
2842                         if(i>=0)
2843                             ref->pic_id= pred;
2844                     }else{
2845                         int long_idx;
2846                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
2847
2848                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
2849
2850                         if(long_idx>31){
2851                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
2852                             return -1;
2853                         }
2854                         ref = h->long_ref[long_idx];
2855                         assert(!(ref && !ref->reference));
2856                         if(ref && (ref->reference & pic_structure)){
2857                             ref->pic_id= pic_id;
2858                             assert(ref->long_ref);
2859                             i=0;
2860                         }else{
2861                             i=-1;
2862                         }
2863                     }
2864
2865                     if (i < 0) {
2866                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
2867                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
2868                     } else {
2869                         for(i=index; i+1<h->ref_count[list]; i++){
2870                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
2871                                 break;
2872                         }
2873                         for(; i > index; i--){
2874                             h->ref_list[list][i]= h->ref_list[list][i-1];
2875                         }
2876                         h->ref_list[list][index]= *ref;
2877                         if (FIELD_PICTURE){
2878                             pic_as_field(&h->ref_list[list][index], pic_structure);
2879                         }
2880                     }
2881                 }else{
2882                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
2883                     return -1;
2884                 }
2885             }
2886         }
2887     }
2888     for(list=0; list<h->list_count; list++){
2889         for(index= 0; index < h->ref_count[list]; index++){
2890             if(!h->ref_list[list][index].data[0]){
2891                 av_log(h->s.avctx, AV_LOG_ERROR, "Missing reference picture\n");
2892                 h->ref_list[list][index]= s->current_picture; //FIXME this is not a sensible solution
2893             }
2894         }
2895     }
2896
2897     if(h->slice_type_nos==FF_B_TYPE && !h->direct_spatial_mv_pred)
2898         direct_dist_scale_factor(h);
2899     direct_ref_list_init(h);
2900     return 0;
2901 }
2902
2903 static void fill_mbaff_ref_list(H264Context *h){
2904     int list, i, j;
2905     for(list=0; list<2; list++){ //FIXME try list_count
2906         for(i=0; i<h->ref_count[list]; i++){
2907             Picture *frame = &h->ref_list[list][i];
2908             Picture *field = &h->ref_list[list][16+2*i];
2909             field[0] = *frame;
2910             for(j=0; j<3; j++)
2911                 field[0].linesize[j] <<= 1;
2912             field[0].reference = PICT_TOP_FIELD;
2913             field[1] = field[0];
2914             for(j=0; j<3; j++)
2915                 field[1].data[j] += frame->linesize[j];
2916             field[1].reference = PICT_BOTTOM_FIELD;
2917
2918             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
2919             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
2920             for(j=0; j<2; j++){
2921                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
2922                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
2923             }
2924         }
2925     }
2926     for(j=0; j<h->ref_count[1]; j++){
2927         for(i=0; i<h->ref_count[0]; i++)
2928             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
2929         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
2930         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
2931     }
2932 }
2933
2934 static int pred_weight_table(H264Context *h){
2935     MpegEncContext * const s = &h->s;
2936     int list, i;
2937     int luma_def, chroma_def;
2938
2939     h->use_weight= 0;
2940     h->use_weight_chroma= 0;
2941     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2942     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2943     luma_def = 1<<h->luma_log2_weight_denom;
2944     chroma_def = 1<<h->chroma_log2_weight_denom;
2945
2946     for(list=0; list<2; list++){
2947         for(i=0; i<h->ref_count[list]; i++){
2948             int luma_weight_flag, chroma_weight_flag;
2949
2950             luma_weight_flag= get_bits1(&s->gb);
2951             if(luma_weight_flag){
2952                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
2953                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
2954                 if(   h->luma_weight[list][i] != luma_def
2955                    || h->luma_offset[list][i] != 0)
2956                     h->use_weight= 1;
2957             }else{
2958                 h->luma_weight[list][i]= luma_def;
2959                 h->luma_offset[list][i]= 0;
2960             }
2961
2962             if(CHROMA){
2963                 chroma_weight_flag= get_bits1(&s->gb);
2964                 if(chroma_weight_flag){
2965                     int j;
2966                     for(j=0; j<2; j++){
2967                         h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
2968                         h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
2969                         if(   h->chroma_weight[list][i][j] != chroma_def
2970                         || h->chroma_offset[list][i][j] != 0)
2971                             h->use_weight_chroma= 1;
2972                     }
2973                 }else{
2974                     int j;
2975                     for(j=0; j<2; j++){
2976                         h->chroma_weight[list][i][j]= chroma_def;
2977                         h->chroma_offset[list][i][j]= 0;
2978                     }
2979                 }
2980             }
2981         }
2982         if(h->slice_type_nos != FF_B_TYPE) break;
2983     }
2984     h->use_weight= h->use_weight || h->use_weight_chroma;
2985     return 0;
2986 }
2987
2988 static void implicit_weight_table(H264Context *h){
2989     MpegEncContext * const s = &h->s;
2990     int ref0, ref1;
2991     int cur_poc = s->current_picture_ptr->poc;
2992
2993     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
2994        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2995         h->use_weight= 0;
2996         h->use_weight_chroma= 0;
2997         return;
2998     }
2999
3000     h->use_weight= 2;
3001     h->use_weight_chroma= 2;
3002     h->luma_log2_weight_denom= 5;
3003     h->chroma_log2_weight_denom= 5;
3004
3005     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3006         int poc0 = h->ref_list[0][ref0].poc;
3007         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3008             int poc1 = h->ref_list[1][ref1].poc;
3009             int td = av_clip(poc1 - poc0, -128, 127);
3010             if(td){
3011                 int tb = av_clip(cur_poc - poc0, -128, 127);
3012                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3013                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3014                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3015                     h->implicit_weight[ref0][ref1] = 32;
3016                 else
3017                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3018             }else
3019                 h->implicit_weight[ref0][ref1] = 32;
3020         }
3021     }
3022 }
3023
3024 /**
3025  * Mark a picture as no longer needed for reference. The refmask
3026  * argument allows unreferencing of individual fields or the whole frame.
3027  * If the picture becomes entirely unreferenced, but is being held for
3028  * display purposes, it is marked as such.
3029  * @param refmask mask of fields to unreference; the mask is bitwise
3030  *                anded with the reference marking of pic
3031  * @return non-zero if pic becomes entirely unreferenced (except possibly
3032  *         for display purposes) zero if one of the fields remains in
3033  *         reference
3034  */
3035 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3036     int i;
3037     if (pic->reference &= refmask) {
3038         return 0;
3039     } else {
3040         for(i = 0; h->delayed_pic[i]; i++)
3041             if(pic == h->delayed_pic[i]){
3042                 pic->reference=DELAYED_PIC_REF;
3043                 break;
3044             }
3045         return 1;
3046     }
3047 }
3048
3049 /**
3050  * instantaneous decoder refresh.
3051  */
3052 static void idr(H264Context *h){
3053     int i;
3054
3055     for(i=0; i<16; i++){
3056         remove_long(h, i, 0);
3057     }
3058     assert(h->long_ref_count==0);
3059
3060     for(i=0; i<h->short_ref_count; i++){
3061         unreference_pic(h, h->short_ref[i], 0);
3062         h->short_ref[i]= NULL;
3063     }
3064     h->short_ref_count=0;
3065     h->prev_frame_num= 0;
3066     h->prev_frame_num_offset= 0;
3067     h->prev_poc_msb=
3068     h->prev_poc_lsb= 0;
3069 }
3070
3071 /* forget old pics after a seek */
3072 static void flush_dpb(AVCodecContext *avctx){
3073     H264Context *h= avctx->priv_data;
3074     int i;
3075     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
3076         if(h->delayed_pic[i])
3077             h->delayed_pic[i]->reference= 0;
3078         h->delayed_pic[i]= NULL;
3079     }
3080     h->outputed_poc= INT_MIN;
3081     idr(h);
3082     if(h->s.current_picture_ptr)
3083         h->s.current_picture_ptr->reference= 0;
3084     h->s.first_field= 0;
3085     ff_mpeg_flush(avctx);
3086 }
3087
3088 /**
3089  * Find a Picture in the short term reference list by frame number.
3090  * @param frame_num frame number to search for
3091  * @param idx the index into h->short_ref where returned picture is found
3092  *            undefined if no picture found.
3093  * @return pointer to the found picture, or NULL if no pic with the provided
3094  *                 frame number is found
3095  */
3096 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3097     MpegEncContext * const s = &h->s;
3098     int i;
3099
3100     for(i=0; i<h->short_ref_count; i++){
3101         Picture *pic= h->short_ref[i];
3102         if(s->avctx->debug&FF_DEBUG_MMCO)
3103             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3104         if(pic->frame_num == frame_num) {
3105             *idx = i;
3106             return pic;
3107         }
3108     }
3109     return NULL;
3110 }
3111
3112 /**
3113  * Remove a picture from the short term reference list by its index in
3114  * that list.  This does no checking on the provided index; it is assumed
3115  * to be valid. Other list entries are shifted down.
3116  * @param i index into h->short_ref of picture to remove.
3117  */
3118 static void remove_short_at_index(H264Context *h, int i){
3119     assert(i >= 0 && i < h->short_ref_count);
3120     h->short_ref[i]= NULL;
3121     if (--h->short_ref_count)
3122         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3123 }
3124
3125 /**
3126  *
3127  * @return the removed picture or NULL if an error occurs
3128  */
3129 static Picture * remove_short(H264Context *h, int frame_num, int ref_mask){
3130     MpegEncContext * const s = &h->s;
3131     Picture *pic;
3132     int i;
3133
3134     if(s->avctx->debug&FF_DEBUG_MMCO)
3135         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3136
3137     pic = find_short(h, frame_num, &i);
3138     if (pic){
3139         if(unreference_pic(h, pic, ref_mask))
3140         remove_short_at_index(h, i);
3141     }
3142
3143     return pic;
3144 }
3145
3146 /**
3147  * Remove a picture from the long term reference list by its index in
3148  * that list.
3149  * @return the removed picture or NULL if an error occurs
3150  */
3151 static Picture * remove_long(H264Context *h, int i, int ref_mask){
3152     Picture *pic;
3153
3154     pic= h->long_ref[i];
3155     if (pic){
3156         if(unreference_pic(h, pic, ref_mask)){
3157             assert(h->long_ref[i]->long_ref == 1);
3158             h->long_ref[i]->long_ref= 0;
3159             h->long_ref[i]= NULL;
3160             h->long_ref_count--;
3161         }
3162     }
3163
3164     return pic;
3165 }
3166
3167 /**
3168  * print short term list
3169  */
3170 static void print_short_term(H264Context *h) {
3171     uint32_t i;
3172     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3173         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3174         for(i=0; i<h->short_ref_count; i++){
3175             Picture *pic= h->short_ref[i];
3176             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3177         }
3178     }
3179 }
3180
3181 /**
3182  * print long term list
3183  */
3184 static void print_long_term(H264Context *h) {
3185     uint32_t i;
3186     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3187         av_log(h->s.avctx, AV_LOG_DEBUG, "long term list:\n");
3188         for(i = 0; i < 16; i++){
3189             Picture *pic= h->long_ref[i];
3190             if (pic) {
3191                 av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3192             }
3193         }
3194     }
3195 }
3196
3197 /**
3198  * Executes the reference picture marking (memory management control operations).
3199  */
3200 static int execute_ref_pic_marking(H264Context *h, MMCO *mmco, int mmco_count){
3201     MpegEncContext * const s = &h->s;
3202     int i, j;
3203     int current_ref_assigned=0;
3204     Picture *pic;
3205
3206     if((s->avctx->debug&FF_DEBUG_MMCO) && mmco_count==0)
3207         av_log(h->s.avctx, AV_LOG_DEBUG, "no mmco here\n");
3208
3209     for(i=0; i<mmco_count; i++){
3210         int structure, frame_num;
3211         if(s->avctx->debug&FF_DEBUG_MMCO)
3212             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);
3213
3214         if(   mmco[i].opcode == MMCO_SHORT2UNUSED
3215            || mmco[i].opcode == MMCO_SHORT2LONG){
3216             frame_num = pic_num_extract(h, mmco[i].short_pic_num, &structure);
3217             pic = find_short(h, frame_num, &j);
3218             if(!pic){
3219                 if(mmco[i].opcode != MMCO_SHORT2LONG || !h->long_ref[mmco[i].long_arg]
3220                    || h->long_ref[mmco[i].long_arg]->frame_num != frame_num)
3221                 av_log(h->s.avctx, AV_LOG_ERROR, "mmco: unref short failure\n");
3222                 continue;
3223             }
3224         }
3225
3226         switch(mmco[i].opcode){
3227         case MMCO_SHORT2UNUSED:
3228             if(s->avctx->debug&FF_DEBUG_MMCO)
3229                 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);
3230             remove_short(h, frame_num, structure ^ PICT_FRAME);
3231             break;
3232         case MMCO_SHORT2LONG:
3233                 if (h->long_ref[mmco[i].long_arg] != pic)
3234                     remove_long(h, mmco[i].long_arg, 0);
3235
3236                 remove_short_at_index(h, j);
3237                 h->long_ref[ mmco[i].long_arg ]= pic;
3238                 if (h->long_ref[ mmco[i].long_arg ]){
3239                     h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3240                     h->long_ref_count++;
3241                 }
3242             break;
3243         case MMCO_LONG2UNUSED:
3244             j = pic_num_extract(h, mmco[i].long_arg, &structure);
3245             pic = h->long_ref[j];
3246             if (pic) {
3247                 remove_long(h, j, structure ^ PICT_FRAME);
3248             } else if(s->avctx->debug&FF_DEBUG_MMCO)
3249                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref long failure\n");
3250             break;
3251         case MMCO_LONG:
3252                     // Comment below left from previous code as it is an interresting note.
3253                     /* First field in pair is in short term list or
3254                      * at a different long term index.
3255                      * This is not allowed; see 7.4.3.3, notes 2 and 3.
3256                      * Report the problem and keep the pair where it is,
3257                      * and mark this field valid.
3258                      */
3259
3260             if (h->long_ref[mmco[i].long_arg] != s->current_picture_ptr) {
3261                 remove_long(h, mmco[i].long_arg, 0);
3262
3263                 h->long_ref[ mmco[i].long_arg ]= s->current_picture_ptr;
3264                 h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3265                 h->long_ref_count++;
3266             }
3267
3268             s->current_picture_ptr->reference |= s->picture_structure;
3269             current_ref_assigned=1;
3270             break;
3271         case MMCO_SET_MAX_LONG:
3272             assert(mmco[i].long_arg <= 16);
3273             // just remove the long term which index is greater than new max
3274             for(j = mmco[i].long_arg; j<16; j++){
3275                 remove_long(h, j, 0);
3276             }
3277             break;
3278         case MMCO_RESET:
3279             while(h->short_ref_count){
3280                 remove_short(h, h->short_ref[0]->frame_num, 0);
3281             }
3282             for(j = 0; j < 16; j++) {
3283                 remove_long(h, j, 0);
3284             }
3285             s->current_picture_ptr->poc=
3286             s->current_picture_ptr->field_poc[0]=
3287             s->current_picture_ptr->field_poc[1]=
3288             h->poc_lsb=
3289             h->poc_msb=
3290             h->frame_num=
3291             s->current_picture_ptr->frame_num= 0;
3292             break;
3293         default: assert(0);
3294         }
3295     }
3296
3297     if (!current_ref_assigned) {
3298         /* Second field of complementary field pair; the first field of
3299          * which is already referenced. If short referenced, it
3300          * should be first entry in short_ref. If not, it must exist
3301          * in long_ref; trying to put it on the short list here is an
3302          * error in the encoded bit stream (ref: 7.4.3.3, NOTE 2 and 3).
3303          */
3304         if (h->short_ref_count && h->short_ref[0] == s->current_picture_ptr) {
3305             /* Just mark the second field valid */
3306             s->current_picture_ptr->reference = PICT_FRAME;
3307         } else if (s->current_picture_ptr->long_ref) {
3308             av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term reference "
3309                                              "assignment for second field "
3310                                              "in complementary field pair "
3311                                              "(first field is long term)\n");
3312         } else {
3313             pic= remove_short(h, s->current_picture_ptr->frame_num, 0);
3314             if(pic){
3315                 av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term buffer state detected\n");
3316             }
3317
3318             if(h->short_ref_count)
3319                 memmove(&h->short_ref[1], &h->short_ref[0], h->short_ref_count*sizeof(Picture*));
3320
3321             h->short_ref[0]= s->current_picture_ptr;
3322             h->short_ref_count++;
3323             s->current_picture_ptr->reference |= s->picture_structure;
3324         }
3325     }
3326
3327     if (h->long_ref_count + h->short_ref_count > h->sps.ref_frame_count){
3328
3329         /* We have too many reference frames, probably due to corrupted
3330          * stream. Need to discard one frame. Prevents overrun of the
3331          * short_ref and long_ref buffers.
3332          */
3333         av_log(h->s.avctx, AV_LOG_ERROR,
3334                "number of reference frames exceeds max (probably "
3335                "corrupt input), discarding one\n");
3336
3337         if (h->long_ref_count && !h->short_ref_count) {
3338             for (i = 0; i < 16; ++i)
3339                 if (h->long_ref[i])
3340                     break;
3341
3342             assert(i < 16);
3343             remove_long(h, i, 0);
3344         } else {
3345             pic = h->short_ref[h->short_ref_count - 1];
3346             remove_short(h, pic->frame_num, 0);
3347         }
3348     }
3349
3350     print_short_term(h);
3351     print_long_term(h);
3352     return 0;
3353 }
3354
3355 static int decode_ref_pic_marking(H264Context *h, GetBitContext *gb){
3356     MpegEncContext * const s = &h->s;
3357     int i;
3358
3359     h->mmco_index= 0;
3360     if(h->nal_unit_type == NAL_IDR_SLICE){ //FIXME fields
3361         s->broken_link= get_bits1(gb) -1;
3362         if(get_bits1(gb)){
3363             h->mmco[0].opcode= MMCO_LONG;
3364             h->mmco[0].long_arg= 0;
3365             h->mmco_index= 1;
3366         }
3367     }else{
3368         if(get_bits1(gb)){ // adaptive_ref_pic_marking_mode_flag
3369             for(i= 0; i<MAX_MMCO_COUNT; i++) {
3370                 MMCOOpcode opcode= get_ue_golomb(gb);
3371
3372                 h->mmco[i].opcode= opcode;
3373                 if(opcode==MMCO_SHORT2UNUSED || opcode==MMCO_SHORT2LONG){
3374                     h->mmco[i].short_pic_num= (h->curr_pic_num - get_ue_golomb(gb) - 1) & (h->max_pic_num - 1);
3375 /*                    if(h->mmco[i].short_pic_num >= h->short_ref_count || h->short_ref[ h->mmco[i].short_pic_num ] == NULL){
3376                         av_log(s->avctx, AV_LOG_ERROR, "illegal short ref in memory management control operation %d\n", mmco);
3377                         return -1;
3378                     }*/
3379                 }
3380                 if(opcode==MMCO_SHORT2LONG || opcode==MMCO_LONG2UNUSED || opcode==MMCO_LONG || opcode==MMCO_SET_MAX_LONG){
3381                     unsigned int long_arg= get_ue_golomb(gb);
3382                     if(long_arg >= 32 || (long_arg >= 16 && !(opcode == MMCO_LONG2UNUSED && FIELD_PICTURE))){
3383                         av_log(h->s.avctx, AV_LOG_ERROR, "illegal long ref in memory management control operation %d\n", opcode);
3384                         return -1;
3385                     }
3386                     h->mmco[i].long_arg= long_arg;
3387                 }
3388
3389                 if(opcode > (unsigned)MMCO_LONG){
3390                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal memory management control operation %d\n", opcode);
3391                     return -1;
3392                 }
3393                 if(opcode == MMCO_END)
3394                     break;
3395             }
3396             h->mmco_index= i;
3397         }else{
3398             assert(h->long_ref_count + h->short_ref_count <= h->sps.ref_frame_count);
3399
3400             if(h->short_ref_count && h->long_ref_count + h->short_ref_count == h->sps.ref_frame_count &&
3401                     !(FIELD_PICTURE && !s->first_field && s->current_picture_ptr->reference)) {
3402                 h->mmco[0].opcode= MMCO_SHORT2UNUSED;
3403                 h->mmco[0].short_pic_num= h->short_ref[ h->short_ref_count - 1 ]->frame_num;
3404                 h->mmco_index= 1;
3405                 if (FIELD_PICTURE) {
3406                     h->mmco[0].short_pic_num *= 2;
3407                     h->mmco[1].opcode= MMCO_SHORT2UNUSED;
3408                     h->mmco[1].short_pic_num= h->mmco[0].short_pic_num + 1;
3409                     h->mmco_index= 2;
3410                 }
3411             }
3412         }
3413     }
3414
3415     return 0;
3416 }
3417
3418 static int init_poc(H264Context *h){
3419     MpegEncContext * const s = &h->s;
3420     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
3421     int field_poc[2];
3422     Picture *cur = s->current_picture_ptr;
3423
3424     h->frame_num_offset= h->prev_frame_num_offset;
3425     if(h->frame_num < h->prev_frame_num)
3426         h->frame_num_offset += max_frame_num;
3427
3428     if(h->sps.poc_type==0){
3429         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
3430
3431         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
3432             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
3433         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
3434             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
3435         else
3436             h->poc_msb = h->prev_poc_msb;
3437 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
3438         field_poc[0] =
3439         field_poc[1] = h->poc_msb + h->poc_lsb;
3440         if(s->picture_structure == PICT_FRAME)
3441             field_poc[1] += h->delta_poc_bottom;
3442     }else if(h->sps.poc_type==1){
3443         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
3444         int i;
3445
3446         if(h->sps.poc_cycle_length != 0)
3447             abs_frame_num = h->frame_num_offset + h->frame_num;
3448         else
3449             abs_frame_num = 0;
3450
3451         if(h->nal_ref_idc==0 && abs_frame_num > 0)
3452             abs_frame_num--;
3453
3454         expected_delta_per_poc_cycle = 0;
3455         for(i=0; i < h->sps.poc_cycle_length; i++)
3456             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
3457
3458         if(abs_frame_num > 0){
3459             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
3460             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
3461
3462             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
3463             for(i = 0; i <= frame_num_in_poc_cycle; i++)
3464                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
3465         } else
3466             expectedpoc = 0;
3467
3468         if(h->nal_ref_idc == 0)
3469             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
3470
3471         field_poc[0] = expectedpoc + h->delta_poc[0];
3472         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
3473
3474         if(s->picture_structure == PICT_FRAME)
3475             field_poc[1] += h->delta_poc[1];
3476     }else{
3477         int poc= 2*(h->frame_num_offset + h->frame_num);
3478
3479         if(!h->nal_ref_idc)
3480             poc--;
3481
3482         field_poc[0]= poc;
3483         field_poc[1]= poc;
3484     }
3485
3486     if(s->picture_structure != PICT_BOTTOM_FIELD)
3487         s->current_picture_ptr->field_poc[0]= field_poc[0];
3488     if(s->picture_structure != PICT_TOP_FIELD)
3489         s->current_picture_ptr->field_poc[1]= field_poc[1];
3490     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
3491
3492     return 0;
3493 }
3494
3495
3496 /**
3497  * initialize scan tables
3498  */
3499 static void init_scan_tables(H264Context *h){
3500     MpegEncContext * const s = &h->s;
3501     int i;
3502     if(s->dsp.h264_idct_add == ff_h264_idct_add_c){ //FIXME little ugly
3503         memcpy(h->zigzag_scan, zigzag_scan, 16*sizeof(uint8_t));
3504         memcpy(h-> field_scan,  field_scan, 16*sizeof(uint8_t));
3505     }else{
3506         for(i=0; i<16; i++){
3507 #define T(x) (x>>2) | ((x<<2) & 0xF)
3508             h->zigzag_scan[i] = T(zigzag_scan[i]);
3509             h-> field_scan[i] = T( field_scan[i]);
3510 #undef T
3511         }
3512     }
3513     if(s->dsp.h264_idct8_add == ff_h264_idct8_add_c){
3514         memcpy(h->zigzag_scan8x8,       zigzag_scan8x8,       64*sizeof(uint8_t));
3515         memcpy(h->zigzag_scan8x8_cavlc, zigzag_scan8x8_cavlc, 64*sizeof(uint8_t));
3516         memcpy(h->field_scan8x8,        field_scan8x8,        64*sizeof(uint8_t));
3517         memcpy(h->field_scan8x8_cavlc,  field_scan8x8_cavlc,  64*sizeof(uint8_t));
3518     }else{
3519         for(i=0; i<64; i++){
3520 #define T(x) (x>>3) | ((x&7)<<3)
3521             h->zigzag_scan8x8[i]       = T(zigzag_scan8x8[i]);
3522             h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
3523             h->field_scan8x8[i]        = T(field_scan8x8[i]);
3524             h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
3525 #undef T
3526         }
3527     }
3528     if(h->sps.transform_bypass){ //FIXME same ugly
3529         h->zigzag_scan_q0          = zigzag_scan;
3530         h->zigzag_scan8x8_q0       = zigzag_scan8x8;
3531         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
3532         h->field_scan_q0           = field_scan;
3533         h->field_scan8x8_q0        = field_scan8x8;
3534         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
3535     }else{
3536         h->zigzag_scan_q0          = h->zigzag_scan;
3537         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
3538         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
3539         h->field_scan_q0           = h->field_scan;
3540         h->field_scan8x8_q0        = h->field_scan8x8;
3541         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
3542     }
3543 }
3544
3545 /**
3546  * Replicates H264 "master" context to thread contexts.
3547  */
3548 static void clone_slice(H264Context *dst, H264Context *src)
3549 {
3550     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
3551     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
3552     dst->s.current_picture      = src->s.current_picture;
3553     dst->s.linesize             = src->s.linesize;
3554     dst->s.uvlinesize           = src->s.uvlinesize;
3555     dst->s.first_field          = src->s.first_field;
3556
3557     dst->prev_poc_msb           = src->prev_poc_msb;
3558     dst->prev_poc_lsb           = src->prev_poc_lsb;
3559     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
3560     dst->prev_frame_num         = src->prev_frame_num;
3561     dst->short_ref_count        = src->short_ref_count;
3562
3563     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
3564     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
3565     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
3566     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
3567
3568     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
3569     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
3570 }
3571
3572 /**
3573  * decodes a slice header.
3574  * This will also call MPV_common_init() and frame_start() as needed.
3575  *
3576  * @param h h264context
3577  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
3578  *
3579  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
3580  */
3581 static int decode_slice_header(H264Context *h, H264Context *h0){
3582     MpegEncContext * const s = &h->s;
3583     MpegEncContext * const s0 = &h0->s;
3584     unsigned int first_mb_in_slice;
3585     unsigned int pps_id;
3586     int num_ref_idx_active_override_flag;
3587     static const uint8_t slice_type_map[5]= {FF_P_TYPE, FF_B_TYPE, FF_I_TYPE, FF_SP_TYPE, FF_SI_TYPE};
3588     unsigned int slice_type, tmp, i, j;
3589     int default_ref_list_done = 0;
3590     int last_pic_structure;
3591
3592     s->dropable= h->nal_ref_idc == 0;
3593
3594     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc){
3595         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
3596         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
3597     }else{
3598         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
3599         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
3600     }
3601
3602     first_mb_in_slice= get_ue_golomb(&s->gb);
3603
3604     if((s->flags2 & CODEC_FLAG2_CHUNKS) && first_mb_in_slice == 0){
3605         h0->current_slice = 0;
3606         if (!s0->first_field)
3607             s->current_picture_ptr= NULL;
3608     }
3609
3610     slice_type= get_ue_golomb(&s->gb);
3611     if(slice_type > 9){
3612         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);
3613         return -1;
3614     }
3615     if(slice_type > 4){
3616         slice_type -= 5;
3617         h->slice_type_fixed=1;
3618     }else
3619         h->slice_type_fixed=0;
3620
3621     slice_type= slice_type_map[ slice_type ];
3622     if (slice_type == FF_I_TYPE
3623         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
3624         default_ref_list_done = 1;
3625     }
3626     h->slice_type= slice_type;
3627     h->slice_type_nos= slice_type & 3;
3628
3629     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
3630     if (s->pict_type == FF_B_TYPE && s0->last_picture_ptr == NULL) {
3631         av_log(h->s.avctx, AV_LOG_ERROR,
3632                "B picture before any references, skipping\n");
3633         return -1;
3634     }
3635
3636     pps_id= get_ue_golomb(&s->gb);
3637     if(pps_id>=MAX_PPS_COUNT){
3638         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
3639         return -1;
3640     }
3641     if(!h0->pps_buffers[pps_id]) {
3642         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS referenced\n");
3643         return -1;
3644     }
3645     h->pps= *h0->pps_buffers[pps_id];
3646
3647     if(!h0->sps_buffers[h->pps.sps_id]) {
3648         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS referenced\n");
3649         return -1;
3650     }
3651     h->sps = *h0->sps_buffers[h->pps.sps_id];
3652
3653     if(h == h0 && h->dequant_coeff_pps != pps_id){
3654         h->dequant_coeff_pps = pps_id;
3655         init_dequant_tables(h);
3656     }
3657
3658     s->mb_width= h->sps.mb_width;
3659     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
3660
3661     h->b_stride=  s->mb_width*4;
3662     h->b8_stride= s->mb_width*2;
3663
3664     s->width = 16*s->mb_width - 2*FFMIN(h->sps.crop_right, 7);
3665     if(h->sps.frame_mbs_only_flag)
3666         s->height= 16*s->mb_height - 2*FFMIN(h->sps.crop_bottom, 7);
3667     else
3668         s->height= 16*s->mb_height - 4*FFMIN(h->sps.crop_bottom, 3);
3669
3670     if (s->context_initialized
3671         && (   s->width != s->avctx->width || s->height != s->avctx->height)) {
3672         if(h != h0)
3673             return -1;   // width / height changed during parallelized decoding
3674         free_tables(h);
3675         MPV_common_end(s);
3676     }
3677     if (!s->context_initialized) {
3678         if(h != h0)
3679             return -1;  // we cant (re-)initialize context during parallel decoding
3680         if (MPV_common_init(s) < 0)
3681             return -1;
3682         s->first_field = 0;
3683
3684         init_scan_tables(h);
3685         alloc_tables(h);
3686
3687         for(i = 1; i < s->avctx->thread_count; i++) {
3688             H264Context *c;
3689             c = h->thread_context[i] = av_malloc(sizeof(H264Context));
3690             memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
3691             memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
3692             c->sps = h->sps;
3693             c->pps = h->pps;
3694             init_scan_tables(c);
3695             clone_tables(c, h);
3696         }
3697
3698         for(i = 0; i < s->avctx->thread_count; i++)
3699             if(context_init(h->thread_context[i]) < 0)
3700                 return -1;
3701
3702         s->avctx->width = s->width;
3703         s->avctx->height = s->height;
3704         s->avctx->sample_aspect_ratio= h->sps.sar;
3705         if(!s->avctx->sample_aspect_ratio.den)
3706             s->avctx->sample_aspect_ratio.den = 1;
3707
3708         if(h->sps.timing_info_present_flag){
3709             s->avctx->time_base= (AVRational){h->sps.num_units_in_tick * 2, h->sps.time_scale};
3710             if(h->x264_build > 0 && h->x264_build < 44)
3711                 s->avctx->time_base.den *= 2;
3712             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
3713                       s->avctx->time_base.num, s->avctx->time_base.den, 1<<30);
3714         }
3715     }
3716
3717     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
3718
3719     h->mb_mbaff = 0;
3720     h->mb_aff_frame = 0;
3721     last_pic_structure = s0->picture_structure;
3722     if(h->sps.frame_mbs_only_flag){
3723         s->picture_structure= PICT_FRAME;
3724     }else{
3725         if(get_bits1(&s->gb)) { //field_pic_flag
3726             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
3727         } else {
3728             s->picture_structure= PICT_FRAME;
3729             h->mb_aff_frame = h->sps.mb_aff;
3730         }
3731     }
3732     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
3733
3734     if(h0->current_slice == 0){
3735         while(h->frame_num !=  h->prev_frame_num &&
3736               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
3737             av_log(NULL, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
3738             frame_start(h);
3739             h->prev_frame_num++;
3740             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
3741             s->current_picture_ptr->frame_num= h->prev_frame_num;
3742             execute_ref_pic_marking(h, NULL, 0);
3743         }
3744
3745         /* See if we have a decoded first field looking for a pair... */
3746         if (s0->first_field) {
3747             assert(s0->current_picture_ptr);
3748             assert(s0->current_picture_ptr->data[0]);
3749             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
3750
3751             /* figure out if we have a complementary field pair */
3752             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3753                 /*
3754                  * Previous field is unmatched. Don't display it, but let it
3755                  * remain for reference if marked as such.
3756                  */
3757                 s0->current_picture_ptr = NULL;
3758                 s0->first_field = FIELD_PICTURE;
3759
3760             } else {
3761                 if (h->nal_ref_idc &&
3762                         s0->current_picture_ptr->reference &&
3763                         s0->current_picture_ptr->frame_num != h->frame_num) {
3764                     /*
3765                      * This and previous field were reference, but had
3766                      * different frame_nums. Consider this field first in
3767                      * pair. Throw away previous field except for reference
3768                      * purposes.
3769                      */
3770                     s0->first_field = 1;
3771                     s0->current_picture_ptr = NULL;
3772
3773                 } else {
3774                     /* Second field in complementary pair */
3775                     s0->first_field = 0;
3776                 }
3777             }
3778
3779         } else {
3780             /* Frame or first field in a potentially complementary pair */
3781             assert(!s0->current_picture_ptr);
3782             s0->first_field = FIELD_PICTURE;
3783         }
3784
3785         if((!FIELD_PICTURE || s0->first_field) && frame_start(h) < 0) {
3786             s0->first_field = 0;
3787             return -1;
3788         }
3789     }
3790     if(h != h0)
3791         clone_slice(h, h0);
3792
3793     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
3794
3795     assert(s->mb_num == s->mb_width * s->mb_height);
3796     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
3797        first_mb_in_slice                    >= s->mb_num){
3798         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
3799         return -1;
3800     }
3801     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
3802     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
3803     if (s->picture_structure == PICT_BOTTOM_FIELD)
3804         s->resync_mb_y = s->mb_y = s->mb_y + 1;
3805     assert(s->mb_y < s->mb_height);
3806
3807     if(s->picture_structure==PICT_FRAME){
3808         h->curr_pic_num=   h->frame_num;
3809         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
3810     }else{
3811         h->curr_pic_num= 2*h->frame_num + 1;
3812         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
3813     }
3814
3815     if(h->nal_unit_type == NAL_IDR_SLICE){
3816         get_ue_golomb(&s->gb); /* idr_pic_id */
3817     }
3818
3819     if(h->sps.poc_type==0){
3820         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
3821
3822         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
3823             h->delta_poc_bottom= get_se_golomb(&s->gb);
3824         }
3825     }
3826
3827     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
3828         h->delta_poc[0]= get_se_golomb(&s->gb);
3829
3830         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
3831             h->delta_poc[1]= get_se_golomb(&s->gb);
3832     }
3833
3834     init_poc(h);
3835
3836     if(h->pps.redundant_pic_cnt_present){
3837         h->redundant_pic_count= get_ue_golomb(&s->gb);
3838     }
3839
3840     //set defaults, might be overridden a few lines later
3841     h->ref_count[0]= h->pps.ref_count[0];
3842     h->ref_count[1]= h->pps.ref_count[1];
3843
3844     if(h->slice_type_nos != FF_I_TYPE){
3845         if(h->slice_type_nos == FF_B_TYPE){
3846             h->direct_spatial_mv_pred= get_bits1(&s->gb);
3847         }
3848         num_ref_idx_active_override_flag= get_bits1(&s->gb);
3849
3850         if(num_ref_idx_active_override_flag){
3851             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
3852             if(h->slice_type_nos==FF_B_TYPE)
3853                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
3854
3855             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
3856                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3857                 h->ref_count[0]= h->ref_count[1]= 1;
3858                 return -1;
3859             }
3860         }
3861         if(h->slice_type_nos == FF_B_TYPE)
3862             h->list_count= 2;
3863         else
3864             h->list_count= 1;
3865     }else
3866         h->list_count= 0;
3867
3868     if(!default_ref_list_done){
3869         fill_default_ref_list(h);
3870     }
3871
3872     if(h->slice_type_nos!=FF_I_TYPE && decode_ref_pic_list_reordering(h) < 0)
3873         return -1;
3874
3875     if(   (h->pps.weighted_pred          && h->slice_type_nos == FF_P_TYPE )
3876        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== FF_B_TYPE ) )
3877         pred_weight_table(h);
3878     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== FF_B_TYPE)
3879         implicit_weight_table(h);
3880     else
3881         h->use_weight = 0;
3882
3883     if(h->nal_ref_idc)
3884         decode_ref_pic_marking(h0, &s->gb);
3885
3886     if(FRAME_MBAFF)
3887         fill_mbaff_ref_list(h);
3888
3889     if( h->slice_type_nos != FF_I_TYPE && h->pps.cabac ){
3890         tmp = get_ue_golomb(&s->gb);
3891         if(tmp > 2){
3892             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3893             return -1;
3894         }
3895         h->cabac_init_idc= tmp;
3896     }
3897
3898     h->last_qscale_diff = 0;
3899     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3900     if(tmp>51){
3901         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3902         return -1;
3903     }
3904     s->qscale= tmp;
3905     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3906     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3907     //FIXME qscale / qp ... stuff
3908     if(h->slice_type == FF_SP_TYPE){
3909         get_bits1(&s->gb); /* sp_for_switch_flag */
3910     }
3911     if(h->slice_type==FF_SP_TYPE || h->slice_type == FF_SI_TYPE){
3912         get_se_golomb(&s->gb); /* slice_qs_delta */
3913     }
3914
3915     h->deblocking_filter = 1;
3916     h->slice_alpha_c0_offset = 0;
3917     h->slice_beta_offset = 0;
3918     if( h->pps.deblocking_filter_parameters_present ) {
3919         tmp= get_ue_golomb(&s->gb);
3920         if(tmp > 2){
3921             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
3922             return -1;
3923         }
3924         h->deblocking_filter= tmp;
3925         if(h->deblocking_filter < 2)
3926             h->deblocking_filter^= 1; // 1<->0
3927
3928         if( h->deblocking_filter ) {
3929             h->slice_alpha_c0_offset = get_se_golomb(&s->gb) << 1;
3930             h->slice_beta_offset = get_se_golomb(&s->gb) << 1;
3931         }
3932     }
3933
3934     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
3935        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != FF_I_TYPE)
3936        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == FF_B_TYPE)
3937        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
3938         h->deblocking_filter= 0;
3939
3940     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
3941         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
3942             /* Cheat slightly for speed:
3943                Do not bother to deblock across slices. */
3944             h->deblocking_filter = 2;
3945         } else {
3946             h0->max_contexts = 1;
3947             if(!h0->single_decode_warning) {
3948                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3949                 h0->single_decode_warning = 1;
3950             }
3951             if(h != h0)
3952                 return 1; // deblocking switched inside frame
3953         }
3954     }
3955
3956 #if 0 //FMO
3957     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
3958         slice_group_change_cycle= get_bits(&s->gb, ?);
3959 #endif
3960
3961     h0->last_slice_type = slice_type;
3962     h->slice_num = ++h0->current_slice;
3963
3964     for(j=0; j<2; j++){
3965         int *ref2frm= h->ref2frm[h->slice_num&15][j];
3966         ref2frm[0]=
3967         ref2frm[1]= -1;
3968         for(i=0; i<16; i++)
3969             ref2frm[i+2]= 4*h->ref_list[j][i].frame_num
3970                           +(h->ref_list[j][i].reference&3);
3971         ref2frm[18+0]=
3972         ref2frm[18+1]= -1;
3973         for(i=16; i<48; i++)
3974             ref2frm[i+4]= 4*h->ref_list[j][i].frame_num
3975                           +(h->ref_list[j][i].reference&3);
3976     }
3977
3978     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE) ? 0 : 16;
3979     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3980
3981     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3982         av_log(h->s.avctx, AV_LOG_DEBUG, "slice:%d %s mb:%d %c pps:%u frame:%d poc:%d/%d ref:%d/%d qp:%d loop:%d:%d:%d weight:%d%s %s\n",
3983                h->slice_num,
3984                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3985                first_mb_in_slice,
3986                av_get_pict_type_char(h->slice_type),
3987                pps_id, h->frame_num,
3988                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3989                h->ref_count[0], h->ref_count[1],
3990                s->qscale,
3991                h->deblocking_filter, h->slice_alpha_c0_offset/2, h->slice_beta_offset/2,
3992                h->use_weight,
3993                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3994                h->slice_type == FF_B_TYPE ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3995                );
3996     }
3997
3998     return 0;
3999 }
4000
4001 /**
4002  *
4003  */
4004 static inline int get_level_prefix(GetBitContext *gb){
4005     unsigned int buf;
4006     int log;
4007
4008     OPEN_READER(re, gb);
4009     UPDATE_CACHE(re, gb);
4010     buf=GET_CACHE(re, gb);
4011
4012     log= 32 - av_log2(buf);
4013 #ifdef TRACE
4014     print_bin(buf>>(32-log), log);
4015     av_log(NULL, AV_LOG_DEBUG, "%5d %2d %3d lpr @%5d in %s get_level_prefix\n", buf>>(32-log), log, log-1, get_bits_count(gb), __FILE__);
4016 #endif
4017
4018     LAST_SKIP_BITS(re, gb, log);
4019     CLOSE_READER(re, gb);
4020
4021     return log-1;
4022 }
4023
4024 static inline int get_dct8x8_allowed(H264Context *h){
4025     int i;
4026     for(i=0; i<4; i++){
4027         if(!IS_SUB_8X8(h->sub_mb_type[i])
4028            || (!h->sps.direct_8x8_inference_flag && IS_DIRECT(h->sub_mb_type[i])))
4029             return 0;
4030     }
4031     return 1;
4032 }
4033
4034 /**
4035  * decodes a residual block.
4036  * @param n block index
4037  * @param scantable scantable
4038  * @param max_coeff number of coefficients in the block
4039  * @return <0 if an error occurred
4040  */
4041 static int decode_residual(H264Context *h, GetBitContext *gb, DCTELEM *block, int n, const uint8_t *scantable, const uint32_t *qmul, int max_coeff){
4042     MpegEncContext * const s = &h->s;
4043     static const int coeff_token_table_index[17]= {0, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3};
4044     int level[16];
4045     int zeros_left, coeff_num, coeff_token, total_coeff, i, j, trailing_ones, run_before;
4046
4047     //FIXME put trailing_onex into the context
4048
4049     if(n == CHROMA_DC_BLOCK_INDEX){
4050         coeff_token= get_vlc2(gb, chroma_dc_coeff_token_vlc.table, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 1);
4051         total_coeff= coeff_token>>2;
4052     }else{
4053         if(n == LUMA_DC_BLOCK_INDEX){
4054             total_coeff= pred_non_zero_count(h, 0);
4055             coeff_token= get_vlc2(gb, coeff_token_vlc[ coeff_token_table_index[total_coeff] ].table, COEFF_TOKEN_VLC_BITS, 2);
4056             total_coeff= coeff_token>>2;
4057         }else{
4058             total_coeff= pred_non_zero_count(h, n);
4059             coeff_token= get_vlc2(gb, coeff_token_vlc[ coeff_token_table_index[total_coeff] ].table, COEFF_TOKEN_VLC_BITS, 2);
4060             total_coeff= coeff_token>>2;