Implement avcodec_decode_video2(), _audio3() and _subtitle2() which takes an
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file libavcodec/h264.c
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "internal.h"
29 #include "dsputil.h"
30 #include "avcodec.h"
31 #include "mpegvideo.h"
32 #include "h264.h"
33 #include "h264data.h"
34 #include "h264_parser.h"
35 #include "golomb.h"
36 #include "mathops.h"
37 #include "rectangle.h"
38 #include "vdpau_internal.h"
39
40 #include "cabac.h"
41 #if ARCH_X86
42 #include "x86/h264_i386.h"
43 #endif
44
45 //#undef NDEBUG
46 #include <assert.h>
47
48 /**
49  * Value of Picture.reference when Picture is not a reference picture, but
50  * is held for delayed output.
51  */
52 #define DELAYED_PIC_REF 4
53
54 static VLC coeff_token_vlc[4];
55 static VLC_TYPE coeff_token_vlc_tables[520+332+280+256][2];
56 static const int coeff_token_vlc_tables_size[4]={520,332,280,256};
57
58 static VLC chroma_dc_coeff_token_vlc;
59 static VLC_TYPE chroma_dc_coeff_token_vlc_table[256][2];
60 static const int chroma_dc_coeff_token_vlc_table_size = 256;
61
62 static VLC total_zeros_vlc[15];
63 static VLC_TYPE total_zeros_vlc_tables[15][512][2];
64 static const int total_zeros_vlc_tables_size = 512;
65
66 static VLC chroma_dc_total_zeros_vlc[3];
67 static VLC_TYPE chroma_dc_total_zeros_vlc_tables[3][8][2];
68 static const int chroma_dc_total_zeros_vlc_tables_size = 8;
69
70 static VLC run_vlc[6];
71 static VLC_TYPE run_vlc_tables[6][8][2];
72 static const int run_vlc_tables_size = 8;
73
74 static VLC run7_vlc;
75 static VLC_TYPE run7_vlc_table[96][2];
76 static const int run7_vlc_table_size = 96;
77
78 static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
79 static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
80 static void filter_mb( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
81 static void filter_mb_fast( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
82 static Picture * remove_long(H264Context *h, int i, int ref_mask);
83
84 static av_always_inline uint32_t pack16to32(int a, int b){
85 #ifdef WORDS_BIGENDIAN
86    return (b&0xFFFF) + (a<<16);
87 #else
88    return (a&0xFFFF) + (b<<16);
89 #endif
90 }
91
92 static const uint8_t rem6[52]={
93 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
94 };
95
96 static const uint8_t div6[52]={
97 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8,
98 };
99
100 static const uint8_t left_block_options[4][8]={
101     {0,1,2,3,7,10,8,11},
102     {2,2,3,3,8,11,8,11},
103     {0,0,1,1,7,10,7,10},
104     {0,2,0,2,7,10,7,10}
105 };
106
107 #define LEVEL_TAB_BITS 8
108 static int8_t cavlc_level_tab[7][1<<LEVEL_TAB_BITS][2];
109
110 static void fill_caches(H264Context *h, int mb_type, int for_deblock){
111     MpegEncContext * const s = &h->s;
112     const int mb_xy= h->mb_xy;
113     int topleft_xy, top_xy, topright_xy, left_xy[2];
114     int topleft_type, top_type, topright_type, left_type[2];
115     const uint8_t * left_block;
116     int topleft_partition= -1;
117     int i;
118
119     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
120
121     //FIXME deblocking could skip the intra and nnz parts.
122     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
123         return;
124
125     /* Wow, what a mess, why didn't they simplify the interlacing & intra
126      * stuff, I can't imagine that these complex rules are worth it. */
127
128     topleft_xy = top_xy - 1;
129     topright_xy= top_xy + 1;
130     left_xy[1] = left_xy[0] = mb_xy-1;
131     left_block = left_block_options[0];
132     if(FRAME_MBAFF){
133         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
134         const int top_pair_xy      = pair_xy     - s->mb_stride;
135         const int topleft_pair_xy  = top_pair_xy - 1;
136         const int topright_pair_xy = top_pair_xy + 1;
137         const int topleft_mb_field_flag  = IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
138         const int top_mb_field_flag      = IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
139         const int topright_mb_field_flag = IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
140         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
141         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
142         const int bottom = (s->mb_y & 1);
143         tprintf(s->avctx, "fill_caches: curr_mb_field_flag:%d, left_mb_field_flag:%d, topleft_mb_field_flag:%d, top_mb_field_flag:%d, topright_mb_field_flag:%d\n", curr_mb_field_flag, left_mb_field_flag, topleft_mb_field_flag, top_mb_field_flag, topright_mb_field_flag);
144
145         if (curr_mb_field_flag && (bottom || top_mb_field_flag)){
146             top_xy -= s->mb_stride;
147         }
148         if (curr_mb_field_flag && (bottom || topleft_mb_field_flag)){
149             topleft_xy -= s->mb_stride;
150         } else if(bottom && !curr_mb_field_flag && left_mb_field_flag) {
151             topleft_xy += s->mb_stride;
152             // take top left mv from the middle of the mb, as opposed to all other modes which use the bottom right partition
153             topleft_partition = 0;
154         }
155         if (curr_mb_field_flag && (bottom || topright_mb_field_flag)){
156             topright_xy -= s->mb_stride;
157         }
158         if (left_mb_field_flag != curr_mb_field_flag) {
159             left_xy[1] = left_xy[0] = pair_xy - 1;
160             if (curr_mb_field_flag) {
161                 left_xy[1] += s->mb_stride;
162                 left_block = left_block_options[3];
163             } else {
164                 left_block= left_block_options[2 - bottom];
165             }
166         }
167     }
168
169     h->top_mb_xy = top_xy;
170     h->left_mb_xy[0] = left_xy[0];
171     h->left_mb_xy[1] = left_xy[1];
172     if(for_deblock){
173         topleft_type = 0;
174         topright_type = 0;
175         top_type     = h->slice_table[top_xy     ] < 0xFFFF ? s->current_picture.mb_type[top_xy]     : 0;
176         left_type[0] = h->slice_table[left_xy[0] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[0]] : 0;
177         left_type[1] = h->slice_table[left_xy[1] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[1]] : 0;
178
179         if(MB_MBAFF && !IS_INTRA(mb_type)){
180             int list;
181             for(list=0; list<h->list_count; list++){
182                 //These values where changed for ease of performing MC, we need to change them back
183                 //FIXME maybe we can make MC and loop filter use the same values or prevent
184                 //the MC code from changing ref_cache and rather use a temporary array.
185                 if(USES_LIST(mb_type,list)){
186                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
187                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
188                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
189                     ref += h->b8_stride;
190                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
191                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
192                 }
193             }
194         }
195     }else{
196         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
197         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
198         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
199         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
200         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
201
202     if(IS_INTRA(mb_type)){
203         int type_mask= h->pps.constrained_intra_pred ? IS_INTRA(-1) : -1;
204         h->topleft_samples_available=
205         h->top_samples_available=
206         h->left_samples_available= 0xFFFF;
207         h->topright_samples_available= 0xEEEA;
208
209         if(!(top_type & type_mask)){
210             h->topleft_samples_available= 0xB3FF;
211             h->top_samples_available= 0x33FF;
212             h->topright_samples_available= 0x26EA;
213         }
214         if(IS_INTERLACED(mb_type) != IS_INTERLACED(left_type[0])){
215             if(IS_INTERLACED(mb_type)){
216                 if(!(left_type[0] & type_mask)){
217                     h->topleft_samples_available&= 0xDFFF;
218                     h->left_samples_available&= 0x5FFF;
219                 }
220                 if(!(left_type[1] & type_mask)){
221                     h->topleft_samples_available&= 0xFF5F;
222                     h->left_samples_available&= 0xFF5F;
223                 }
224             }else{
225                 int left_typei = h->slice_table[left_xy[0] + s->mb_stride ] == h->slice_num
226                                 ? s->current_picture.mb_type[left_xy[0] + s->mb_stride] : 0;
227                 assert(left_xy[0] == left_xy[1]);
228                 if(!((left_typei & type_mask) && (left_type[0] & type_mask))){
229                     h->topleft_samples_available&= 0xDF5F;
230                     h->left_samples_available&= 0x5F5F;
231                 }
232             }
233         }else{
234             if(!(left_type[0] & type_mask)){
235                 h->topleft_samples_available&= 0xDF5F;
236                 h->left_samples_available&= 0x5F5F;
237             }
238         }
239
240         if(!(topleft_type & type_mask))
241             h->topleft_samples_available&= 0x7FFF;
242
243         if(!(topright_type & type_mask))
244             h->topright_samples_available&= 0xFBFF;
245
246         if(IS_INTRA4x4(mb_type)){
247             if(IS_INTRA4x4(top_type)){
248                 h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
249                 h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
250                 h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
251                 h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
252             }else{
253                 int pred;
254                 if(!(top_type & type_mask))
255                     pred= -1;
256                 else{
257                     pred= 2;
258                 }
259                 h->intra4x4_pred_mode_cache[4+8*0]=
260                 h->intra4x4_pred_mode_cache[5+8*0]=
261                 h->intra4x4_pred_mode_cache[6+8*0]=
262                 h->intra4x4_pred_mode_cache[7+8*0]= pred;
263             }
264             for(i=0; i<2; i++){
265                 if(IS_INTRA4x4(left_type[i])){
266                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
267                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
268                 }else{
269                     int pred;
270                     if(!(left_type[i] & type_mask))
271                         pred= -1;
272                     else{
273                         pred= 2;
274                     }
275                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
276                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
277                 }
278             }
279         }
280     }
281     }
282
283
284 /*
285 0 . T T. T T T T
286 1 L . .L . . . .
287 2 L . .L . . . .
288 3 . T TL . . . .
289 4 L . .L . . . .
290 5 L . .. . . . .
291 */
292 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
293     if(top_type){
294         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
295         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
296         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
297         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
298
299         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
300         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
301
302         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
303         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
304
305     }else{
306         h->non_zero_count_cache[4+8*0]=
307         h->non_zero_count_cache[5+8*0]=
308         h->non_zero_count_cache[6+8*0]=
309         h->non_zero_count_cache[7+8*0]=
310
311         h->non_zero_count_cache[1+8*0]=
312         h->non_zero_count_cache[2+8*0]=
313
314         h->non_zero_count_cache[1+8*3]=
315         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
316
317     }
318
319     for (i=0; i<2; i++) {
320         if(left_type[i]){
321             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
322             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
323             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
324             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
325         }else{
326             h->non_zero_count_cache[3+8*1 + 2*8*i]=
327             h->non_zero_count_cache[3+8*2 + 2*8*i]=
328             h->non_zero_count_cache[0+8*1 +   8*i]=
329             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
330         }
331     }
332
333     if( h->pps.cabac ) {
334         // top_cbp
335         if(top_type) {
336             h->top_cbp = h->cbp_table[top_xy];
337         } else if(IS_INTRA(mb_type)) {
338             h->top_cbp = 0x1C0;
339         } else {
340             h->top_cbp = 0;
341         }
342         // left_cbp
343         if (left_type[0]) {
344             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
345         } else if(IS_INTRA(mb_type)) {
346             h->left_cbp = 0x1C0;
347         } else {
348             h->left_cbp = 0;
349         }
350         if (left_type[0]) {
351             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
352         }
353         if (left_type[1]) {
354             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
355         }
356     }
357
358 #if 1
359     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
360         int list;
361         for(list=0; list<h->list_count; list++){
362             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
363                 /*if(!h->mv_cache_clean[list]){
364                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
365                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
366                     h->mv_cache_clean[list]= 1;
367                 }*/
368                 continue;
369             }
370             h->mv_cache_clean[list]= 0;
371
372             if(USES_LIST(top_type, list)){
373                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
374                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
375                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
376                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
377                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
378                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
379                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
380                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
381                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
382                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
383             }else{
384                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
385                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
386                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
387                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
388                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
389             }
390
391             for(i=0; i<2; i++){
392                 int cache_idx = scan8[0] - 1 + i*2*8;
393                 if(USES_LIST(left_type[i], list)){
394                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
395                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
396                     *(uint32_t*)h->mv_cache[list][cache_idx  ]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[0+i*2]];
397                     *(uint32_t*)h->mv_cache[list][cache_idx+8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[1+i*2]];
398                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
399                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
400                 }else{
401                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
402                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
403                     h->ref_cache[list][cache_idx  ]=
404                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
405                 }
406             }
407
408             if(for_deblock || ((IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred) && !FRAME_MBAFF))
409                 continue;
410
411             if(USES_LIST(topleft_type, list)){
412                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
413                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
414                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
415                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
416             }else{
417                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
418                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
419             }
420
421             if(USES_LIST(topright_type, list)){
422                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
423                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
424                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
425                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
426             }else{
427                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
428                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
429             }
430
431             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
432                 continue;
433
434             h->ref_cache[list][scan8[5 ]+1] =
435             h->ref_cache[list][scan8[7 ]+1] =
436             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
437             h->ref_cache[list][scan8[4 ]] =
438             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
439             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
440             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
441             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
442             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
443             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
444
445             if( h->pps.cabac ) {
446                 /* XXX beurk, Load mvd */
447                 if(USES_LIST(top_type, list)){
448                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
449                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
450                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
451                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
452                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
453                 }else{
454                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
455                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
456                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
457                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
458                 }
459                 if(USES_LIST(left_type[0], list)){
460                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
461                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 0*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[0]];
462                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[1]];
463                 }else{
464                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
465                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
466                 }
467                 if(USES_LIST(left_type[1], list)){
468                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
469                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 2*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[2]];
470                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 3*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[3]];
471                 }else{
472                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
473                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
474                 }
475                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
476                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
477                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
478                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
479                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
480
481                 if(h->slice_type_nos == FF_B_TYPE){
482                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
483
484                     if(IS_DIRECT(top_type)){
485                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
486                     }else if(IS_8X8(top_type)){
487                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
488                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
489                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
490                     }else{
491                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
492                     }
493
494                     if(IS_DIRECT(left_type[0]))
495                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
496                     else if(IS_8X8(left_type[0]))
497                         h->direct_cache[scan8[0] - 1 + 0*8]= h->direct_table[h->mb2b8_xy[left_xy[0]] + 1 + h->b8_stride*(left_block[0]>>1)];
498                     else
499                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
500
501                     if(IS_DIRECT(left_type[1]))
502                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
503                     else if(IS_8X8(left_type[1]))
504                         h->direct_cache[scan8[0] - 1 + 2*8]= h->direct_table[h->mb2b8_xy[left_xy[1]] + 1 + h->b8_stride*(left_block[2]>>1)];
505                     else
506                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
507                 }
508             }
509
510             if(FRAME_MBAFF){
511 #define MAP_MVS\
512                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
513                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
514                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
515                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
516                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
517                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
518                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
519                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
520                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
521                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
522                 if(MB_FIELD){
523 #define MAP_F2F(idx, mb_type)\
524                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
525                         h->ref_cache[list][idx] <<= 1;\
526                         h->mv_cache[list][idx][1] /= 2;\
527                         h->mvd_cache[list][idx][1] /= 2;\
528                     }
529                     MAP_MVS
530 #undef MAP_F2F
531                 }else{
532 #define MAP_F2F(idx, mb_type)\
533                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
534                         h->ref_cache[list][idx] >>= 1;\
535                         h->mv_cache[list][idx][1] <<= 1;\
536                         h->mvd_cache[list][idx][1] <<= 1;\
537                     }
538                     MAP_MVS
539 #undef MAP_F2F
540                 }
541             }
542         }
543     }
544 #endif
545
546     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
547 }
548
549 static inline void write_back_intra_pred_mode(H264Context *h){
550     const int mb_xy= h->mb_xy;
551
552     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
553     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
554     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
555     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
556     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
557     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
558     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
559 }
560
561 /**
562  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
563  */
564 static inline int check_intra4x4_pred_mode(H264Context *h){
565     MpegEncContext * const s = &h->s;
566     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
567     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
568     int i;
569
570     if(!(h->top_samples_available&0x8000)){
571         for(i=0; i<4; i++){
572             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
573             if(status<0){
574                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
575                 return -1;
576             } else if(status){
577                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
578             }
579         }
580     }
581
582     if((h->left_samples_available&0x8888)!=0x8888){
583         static const int mask[4]={0x8000,0x2000,0x80,0x20};
584         for(i=0; i<4; i++){
585             if(!(h->left_samples_available&mask[i])){
586                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
587                 if(status<0){
588                     av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
589                     return -1;
590                 } else if(status){
591                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
592                 }
593             }
594         }
595     }
596
597     return 0;
598 } //FIXME cleanup like next
599
600 /**
601  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
602  */
603 static inline int check_intra_pred_mode(H264Context *h, int mode){
604     MpegEncContext * const s = &h->s;
605     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
606     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
607
608     if(mode > 6U) {
609         av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
610         return -1;
611     }
612
613     if(!(h->top_samples_available&0x8000)){
614         mode= top[ mode ];
615         if(mode<0){
616             av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
617             return -1;
618         }
619     }
620
621     if((h->left_samples_available&0x8080) != 0x8080){
622         mode= left[ mode ];
623         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
624             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
625         }
626         if(mode<0){
627             av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
628             return -1;
629         }
630     }
631
632     return mode;
633 }
634
635 /**
636  * gets the predicted intra4x4 prediction mode.
637  */
638 static inline int pred_intra_mode(H264Context *h, int n){
639     const int index8= scan8[n];
640     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
641     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
642     const int min= FFMIN(left, top);
643
644     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
645
646     if(min<0) return DC_PRED;
647     else      return min;
648 }
649
650 static inline void write_back_non_zero_count(H264Context *h){
651     const int mb_xy= h->mb_xy;
652
653     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
654     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
655     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
656     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
657     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
658     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
659     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
660
661     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
662     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
663     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
664
665     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
666     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
667     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
668 }
669
670 /**
671  * gets the predicted number of non-zero coefficients.
672  * @param n block index
673  */
674 static inline int pred_non_zero_count(H264Context *h, int n){
675     const int index8= scan8[n];
676     const int left= h->non_zero_count_cache[index8 - 1];
677     const int top = h->non_zero_count_cache[index8 - 8];
678     int i= left + top;
679
680     if(i<64) i= (i+1)>>1;
681
682     tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);
683
684     return i&31;
685 }
686
687 static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
688     const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
689     MpegEncContext *s = &h->s;
690
691     /* there is no consistent mapping of mvs to neighboring locations that will
692      * make mbaff happy, so we can't move all this logic to fill_caches */
693     if(FRAME_MBAFF){
694         const uint32_t *mb_types = s->current_picture_ptr->mb_type;
695         const int16_t *mv;
696         *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
697         *C = h->mv_cache[list][scan8[0]-2];
698
699         if(!MB_FIELD
700            && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
701             int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
702             if(IS_INTERLACED(mb_types[topright_xy])){
703 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
704                 const int x4 = X4, y4 = Y4;\
705                 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
706                 if(!USES_LIST(mb_type,list))\
707                     return LIST_NOT_USED;\
708                 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
709                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
710                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
711                 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
712
713                 SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
714             }
715         }
716         if(topright_ref == PART_NOT_AVAILABLE
717            && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
718            && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
719             if(!MB_FIELD
720                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
721                 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
722             }
723             if(MB_FIELD
724                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
725                && i >= scan8[0]+8){
726                 // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
727                 SET_DIAG_MV(/2, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
728             }
729         }
730 #undef SET_DIAG_MV
731     }
732
733     if(topright_ref != PART_NOT_AVAILABLE){
734         *C= h->mv_cache[list][ i - 8 + part_width ];
735         return topright_ref;
736     }else{
737         tprintf(s->avctx, "topright MV not available\n");
738
739         *C= h->mv_cache[list][ i - 8 - 1 ];
740         return h->ref_cache[list][ i - 8 - 1 ];
741     }
742 }
743
744 /**
745  * gets the predicted MV.
746  * @param n the block index
747  * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
748  * @param mx the x component of the predicted motion vector
749  * @param my the y component of the predicted motion vector
750  */
751 static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
752     const int index8= scan8[n];
753     const int top_ref=      h->ref_cache[list][ index8 - 8 ];
754     const int left_ref=     h->ref_cache[list][ index8 - 1 ];
755     const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
756     const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
757     const int16_t * C;
758     int diagonal_ref, match_count;
759
760     assert(part_width==1 || part_width==2 || part_width==4);
761
762 /* mv_cache
763   B . . A T T T T
764   U . . L . . , .
765   U . . L . . . .
766   U . . L . . , .
767   . . . L . . . .
768 */
769
770     diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
771     match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
772     tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
773     if(match_count > 1){ //most common
774         *mx= mid_pred(A[0], B[0], C[0]);
775         *my= mid_pred(A[1], B[1], C[1]);
776     }else if(match_count==1){
777         if(left_ref==ref){
778             *mx= A[0];
779             *my= A[1];
780         }else if(top_ref==ref){
781             *mx= B[0];
782             *my= B[1];
783         }else{
784             *mx= C[0];
785             *my= C[1];
786         }
787     }else{
788         if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
789             *mx= A[0];
790             *my= A[1];
791         }else{
792             *mx= mid_pred(A[0], B[0], C[0]);
793             *my= mid_pred(A[1], B[1], C[1]);
794         }
795     }
796
797     tprintf(h->s.avctx, "pred_motion (%2d %2d %2d) (%2d %2d %2d) (%2d %2d %2d) -> (%2d %2d %2d) at %2d %2d %d list %d\n", top_ref, B[0], B[1],                    diagonal_ref, C[0], C[1], left_ref, A[0], A[1], ref, *mx, *my, h->s.mb_x, h->s.mb_y, n, list);
798 }
799
800 /**
801  * gets the directionally predicted 16x8 MV.
802  * @param n the block index
803  * @param mx the x component of the predicted motion vector
804  * @param my the y component of the predicted motion vector
805  */
806 static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
807     if(n==0){
808         const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
809         const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
810
811         tprintf(h->s.avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n", top_ref, B[0], B[1], h->s.mb_x, h->s.mb_y, n, list);
812
813         if(top_ref == ref){
814             *mx= B[0];
815             *my= B[1];
816             return;
817         }
818     }else{
819         const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
820         const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];
821
822         tprintf(h->s.avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n", left_ref, A[0], A[1], h->s.mb_x, h->s.mb_y, n, list);
823
824         if(left_ref == ref){
825             *mx= A[0];
826             *my= A[1];
827             return;
828         }
829     }
830
831     //RARE
832     pred_motion(h, n, 4, list, ref, mx, my);
833 }
834
835 /**
836  * gets the directionally predicted 8x16 MV.
837  * @param n the block index
838  * @param mx the x component of the predicted motion vector
839  * @param my the y component of the predicted motion vector
840  */
841 static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
842     if(n==0){
843         const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
844         const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];
845
846         tprintf(h->s.avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n", left_ref, A[0], A[1], h->s.mb_x, h->s.mb_y, n, list);
847
848         if(left_ref == ref){
849             *mx= A[0];
850             *my= A[1];
851             return;
852         }
853     }else{
854         const int16_t * C;
855         int diagonal_ref;
856
857         diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);
858
859         tprintf(h->s.avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n", diagonal_ref, C[0], C[1], h->s.mb_x, h->s.mb_y, n, list);
860
861         if(diagonal_ref == ref){
862             *mx= C[0];
863             *my= C[1];
864             return;
865         }
866     }
867
868     //RARE
869     pred_motion(h, n, 2, list, ref, mx, my);
870 }
871
872 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
873     const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
874     const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];
875
876     tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);
877
878     if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
879        || !( top_ref | *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ])
880        || !(left_ref | *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ])){
881
882         *mx = *my = 0;
883         return;
884     }
885
886     pred_motion(h, 0, 4, 0, 0, mx, my);
887
888     return;
889 }
890
891 static int get_scale_factor(H264Context * const h, int poc, int poc1, int i){
892     int poc0 = h->ref_list[0][i].poc;
893     int td = av_clip(poc1 - poc0, -128, 127);
894     if(td == 0 || h->ref_list[0][i].long_ref){
895         return 256;
896     }else{
897         int tb = av_clip(poc - poc0, -128, 127);
898         int tx = (16384 + (FFABS(td) >> 1)) / td;
899         return av_clip((tb*tx + 32) >> 6, -1024, 1023);
900     }
901 }
902
903 static inline void direct_dist_scale_factor(H264Context * const h){
904     MpegEncContext * const s = &h->s;
905     const int poc = h->s.current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
906     const int poc1 = h->ref_list[1][0].poc;
907     int i, field;
908     for(field=0; field<2; field++){
909         const int poc  = h->s.current_picture_ptr->field_poc[field];
910         const int poc1 = h->ref_list[1][0].field_poc[field];
911         for(i=0; i < 2*h->ref_count[0]; i++)
912             h->dist_scale_factor_field[field][i^field] = get_scale_factor(h, poc, poc1, i+16);
913     }
914
915     for(i=0; i<h->ref_count[0]; i++){
916         h->dist_scale_factor[i] = get_scale_factor(h, poc, poc1, i);
917     }
918 }
919
920 static void fill_colmap(H264Context *h, int map[2][16+32], int list, int field, int colfield, int mbafi){
921     MpegEncContext * const s = &h->s;
922     Picture * const ref1 = &h->ref_list[1][0];
923     int j, old_ref, rfield;
924     int start= mbafi ? 16                      : 0;
925     int end  = mbafi ? 16+2*h->ref_count[list] : h->ref_count[list];
926     int interl= mbafi || s->picture_structure != PICT_FRAME;
927
928     /* bogus; fills in for missing frames */
929     memset(map[list], 0, sizeof(map[list]));
930
931     for(rfield=0; rfield<2; rfield++){
932         for(old_ref=0; old_ref<ref1->ref_count[colfield][list]; old_ref++){
933             int poc = ref1->ref_poc[colfield][list][old_ref];
934
935             if     (!interl)
936                 poc |= 3;
937             else if( interl && (poc&3) == 3) //FIXME store all MBAFF references so this isnt needed
938                 poc= (poc&~3) + rfield + 1;
939
940             for(j=start; j<end; j++){
941                 if(4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3) == poc){
942                     int cur_ref= mbafi ? (j-16)^field : j;
943                     map[list][2*old_ref + (rfield^field) + 16] = cur_ref;
944                     if(rfield == field)
945                         map[list][old_ref] = cur_ref;
946                     break;
947                 }
948             }
949         }
950     }
951 }
952
953 static inline void direct_ref_list_init(H264Context * const h){
954     MpegEncContext * const s = &h->s;
955     Picture * const ref1 = &h->ref_list[1][0];
956     Picture * const cur = s->current_picture_ptr;
957     int list, j, field;
958     int sidx= (s->picture_structure&1)^1;
959     int ref1sidx= (ref1->reference&1)^1;
960
961     for(list=0; list<2; list++){
962         cur->ref_count[sidx][list] = h->ref_count[list];
963         for(j=0; j<h->ref_count[list]; j++)
964             cur->ref_poc[sidx][list][j] = 4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3);
965     }
966
967     if(s->picture_structure == PICT_FRAME){
968         memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0]));
969         memcpy(cur->ref_poc  [1], cur->ref_poc  [0], sizeof(cur->ref_poc  [0]));
970     }
971
972     cur->mbaff= FRAME_MBAFF;
973
974     if(cur->pict_type != FF_B_TYPE || h->direct_spatial_mv_pred)
975         return;
976
977     for(list=0; list<2; list++){
978         fill_colmap(h, h->map_col_to_list0, list, sidx, ref1sidx, 0);
979         for(field=0; field<2; field++)
980             fill_colmap(h, h->map_col_to_list0_field[field], list, field, field, 1);
981     }
982 }
983
984 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
985     MpegEncContext * const s = &h->s;
986     int b8_stride = h->b8_stride;
987     int b4_stride = h->b_stride;
988     int mb_xy = h->mb_xy;
989     int mb_type_col[2];
990     const int16_t (*l1mv0)[2], (*l1mv1)[2];
991     const int8_t *l1ref0, *l1ref1;
992     const int is_b8x8 = IS_8X8(*mb_type);
993     unsigned int sub_mb_type;
994     int i8, i4;
995
996 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
997
998     if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL
999         if(!IS_INTERLACED(*mb_type)){                    //     AFR/FR    -> AFL/FL
1000             int cur_poc = s->current_picture_ptr->poc;
1001             int *col_poc = h->ref_list[1]->field_poc;
1002             int col_parity = FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc);
1003             mb_xy= s->mb_x + ((s->mb_y&~1) + col_parity)*s->mb_stride;
1004             b8_stride = 0;
1005         }else if(!(s->picture_structure & h->ref_list[1][0].reference) && !h->ref_list[1][0].mbaff){// FL -> FL & differ parity
1006             int fieldoff= 2*(h->ref_list[1][0].reference)-3;
1007             mb_xy += s->mb_stride*fieldoff;
1008         }
1009         goto single_col;
1010     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
1011         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
1012             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
1013             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
1014             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
1015             b8_stride *= 3;
1016             b4_stride *= 6;
1017             //FIXME IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag
1018             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
1019                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
1020                 && !is_b8x8){
1021                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1022                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
1023             }else{
1024                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1025                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1026             }
1027         }else{                                           //     AFR/FR    -> AFR/FR
1028 single_col:
1029             mb_type_col[0] =
1030             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
1031             if(IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag){
1032                 /* FIXME save sub mb types from previous frames (or derive from MVs)
1033                 * so we know exactly what block size to use */
1034                 sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
1035                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1036             }else if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
1037                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1038                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
1039             }else{
1040                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1041                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1042             }
1043         }
1044     }
1045
1046     l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
1047     l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
1048     l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
1049     l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
1050     if(!b8_stride){
1051         if(s->mb_y&1){
1052             l1ref0 += h->b8_stride;
1053             l1ref1 += h->b8_stride;
1054             l1mv0  +=  2*b4_stride;
1055             l1mv1  +=  2*b4_stride;
1056         }
1057     }
1058
1059     if(h->direct_spatial_mv_pred){
1060         int ref[2];
1061         int mv[2][2];
1062         int list;
1063
1064         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
1065
1066         /* ref = min(neighbors) */
1067         for(list=0; list<2; list++){
1068             int refa = h->ref_cache[list][scan8[0] - 1];
1069             int refb = h->ref_cache[list][scan8[0] - 8];
1070             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1071             if(refc == PART_NOT_AVAILABLE)
1072                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1073             ref[list] = FFMIN3((unsigned)refa, (unsigned)refb, (unsigned)refc);
1074             if(ref[list] < 0)
1075                 ref[list] = -1;
1076         }
1077
1078         if(ref[0] < 0 && ref[1] < 0){
1079             ref[0] = ref[1] = 0;
1080             mv[0][0] = mv[0][1] =
1081             mv[1][0] = mv[1][1] = 0;
1082         }else{
1083             for(list=0; list<2; list++){
1084                 if(ref[list] >= 0)
1085                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1086                 else
1087                     mv[list][0] = mv[list][1] = 0;
1088             }
1089         }
1090
1091         if(ref[1] < 0){
1092             if(!is_b8x8)
1093                 *mb_type &= ~MB_TYPE_L1;
1094             sub_mb_type &= ~MB_TYPE_L1;
1095         }else if(ref[0] < 0){
1096             if(!is_b8x8)
1097                 *mb_type &= ~MB_TYPE_L0;
1098             sub_mb_type &= ~MB_TYPE_L0;
1099         }
1100
1101         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1102             for(i8=0; i8<4; i8++){
1103                 int x8 = i8&1;
1104                 int y8 = i8>>1;
1105                 int xy8 = x8+y8*b8_stride;
1106                 int xy4 = 3*x8+y8*b4_stride;
1107                 int a=0, b=0;
1108
1109                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1110                     continue;
1111                 h->sub_mb_type[i8] = sub_mb_type;
1112
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                 if(!IS_INTRA(mb_type_col[y8])
1116                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
1117                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
1118                     if(ref[0] > 0)
1119                         a= pack16to32(mv[0][0],mv[0][1]);
1120                     if(ref[1] > 0)
1121                         b= pack16to32(mv[1][0],mv[1][1]);
1122                 }else{
1123                     a= pack16to32(mv[0][0],mv[0][1]);
1124                     b= pack16to32(mv[1][0],mv[1][1]);
1125                 }
1126                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
1127                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
1128             }
1129         }else if(IS_16X16(*mb_type)){
1130             int a=0, b=0;
1131
1132             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1133             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1134             if(!IS_INTRA(mb_type_col[0])
1135                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1136                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1137                        && (h->x264_build>33 || !h->x264_build)))){
1138                 if(ref[0] > 0)
1139                     a= pack16to32(mv[0][0],mv[0][1]);
1140                 if(ref[1] > 0)
1141                     b= pack16to32(mv[1][0],mv[1][1]);
1142             }else{
1143                 a= pack16to32(mv[0][0],mv[0][1]);
1144                 b= pack16to32(mv[1][0],mv[1][1]);
1145             }
1146             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1147             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1148         }else{
1149             for(i8=0; i8<4; i8++){
1150                 const int x8 = i8&1;
1151                 const int y8 = i8>>1;
1152
1153                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1154                     continue;
1155                 h->sub_mb_type[i8] = sub_mb_type;
1156
1157                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1158                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1159                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1160                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1161
1162                 /* col_zero_flag */
1163                 if(!IS_INTRA(mb_type_col[0]) && (   l1ref0[x8 + y8*b8_stride] == 0
1164                                               || (l1ref0[x8 + y8*b8_stride] < 0 && l1ref1[x8 + y8*b8_stride] == 0
1165                                                   && (h->x264_build>33 || !h->x264_build)))){
1166                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*b8_stride] == 0 ? l1mv0 : l1mv1;
1167                     if(IS_SUB_8X8(sub_mb_type)){
1168                         const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1169                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1170                             if(ref[0] == 0)
1171                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1172                             if(ref[1] == 0)
1173                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1174                         }
1175                     }else
1176                     for(i4=0; i4<4; i4++){
1177                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1178                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1179                             if(ref[0] == 0)
1180                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1181                             if(ref[1] == 0)
1182                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1183                         }
1184                     }
1185                 }
1186             }
1187         }
1188     }else{ /* direct temporal mv pred */
1189         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1190         const int *dist_scale_factor = h->dist_scale_factor;
1191         int ref_offset= 0;
1192
1193         if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
1194             map_col_to_list0[0] = h->map_col_to_list0_field[s->mb_y&1][0];
1195             map_col_to_list0[1] = h->map_col_to_list0_field[s->mb_y&1][1];
1196             dist_scale_factor   =h->dist_scale_factor_field[s->mb_y&1];
1197         }
1198         if(h->ref_list[1][0].mbaff && IS_INTERLACED(mb_type_col[0]))
1199             ref_offset += 16;
1200
1201         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1202             /* FIXME assumes direct_8x8_inference == 1 */
1203             int y_shift  = 2*!IS_INTERLACED(*mb_type);
1204
1205             for(i8=0; i8<4; i8++){
1206                 const int x8 = i8&1;
1207                 const int y8 = i8>>1;
1208                 int ref0, scale;
1209                 const int16_t (*l1mv)[2]= l1mv0;
1210
1211                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1212                     continue;
1213                 h->sub_mb_type[i8] = sub_mb_type;
1214
1215                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1216                 if(IS_INTRA(mb_type_col[y8])){
1217                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1218                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1219                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1220                     continue;
1221                 }
1222
1223                 ref0 = l1ref0[x8 + y8*b8_stride];
1224                 if(ref0 >= 0)
1225                     ref0 = map_col_to_list0[0][ref0 + ref_offset];
1226                 else{
1227                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
1228                     l1mv= l1mv1;
1229                 }
1230                 scale = dist_scale_factor[ref0];
1231                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1232
1233                 {
1234                     const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride];
1235                     int my_col = (mv_col[1]<<y_shift)/2;
1236                     int mx = (scale * mv_col[0] + 128) >> 8;
1237                     int my = (scale * my_col + 128) >> 8;
1238                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1239                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1240                 }
1241             }
1242             return;
1243         }
1244
1245         /* one-to-one mv scaling */
1246
1247         if(IS_16X16(*mb_type)){
1248             int ref, mv0, mv1;
1249
1250             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1251             if(IS_INTRA(mb_type_col[0])){
1252                 ref=mv0=mv1=0;
1253             }else{
1254                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset]
1255                                                 : map_col_to_list0[1][l1ref1[0] + ref_offset];
1256                 const int scale = dist_scale_factor[ref0];
1257                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1258                 int mv_l0[2];
1259                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1260                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1261                 ref= ref0;
1262                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1263                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1264             }
1265             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1266             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1267             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1268         }else{
1269             for(i8=0; i8<4; i8++){
1270                 const int x8 = i8&1;
1271                 const int y8 = i8>>1;
1272                 int ref0, scale;
1273                 const int16_t (*l1mv)[2]= l1mv0;
1274
1275                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1276                     continue;
1277                 h->sub_mb_type[i8] = sub_mb_type;
1278                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1279                 if(IS_INTRA(mb_type_col[0])){
1280                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1281                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1282                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1283                     continue;
1284                 }
1285
1286                 ref0 = l1ref0[x8 + y8*b8_stride] + ref_offset;
1287                 if(ref0 >= 0)
1288                     ref0 = map_col_to_list0[0][ref0];
1289                 else{
1290                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
1291                     l1mv= l1mv1;
1292                 }
1293                 scale = dist_scale_factor[ref0];
1294
1295                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1296                 if(IS_SUB_8X8(sub_mb_type)){
1297                     const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1298                     int mx = (scale * mv_col[0] + 128) >> 8;
1299                     int my = (scale * mv_col[1] + 128) >> 8;
1300                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1301                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1302                 }else
1303                 for(i4=0; i4<4; i4++){
1304                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1305                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1306                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1307                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1308                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1309                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1310                 }
1311             }
1312         }
1313     }
1314 }
1315
1316 static inline void write_back_motion(H264Context *h, int mb_type){
1317     MpegEncContext * const s = &h->s;
1318     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1319     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1320     int list;
1321
1322     if(!USES_LIST(mb_type, 0))
1323         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1324
1325     for(list=0; list<h->list_count; list++){
1326         int y;
1327         if(!USES_LIST(mb_type, list))
1328             continue;
1329
1330         for(y=0; y<4; y++){
1331             *(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];
1332             *(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];
1333         }
1334         if( h->pps.cabac ) {
1335             if(IS_SKIP(mb_type))
1336                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1337             else
1338             for(y=0; y<4; y++){
1339                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1340                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1341             }
1342         }
1343
1344         {
1345             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1346             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1347             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1348             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1349             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1350         }
1351     }
1352
1353     if(h->slice_type_nos == FF_B_TYPE && h->pps.cabac){
1354         if(IS_8X8(mb_type)){
1355             uint8_t *direct_table = &h->direct_table[b8_xy];
1356             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1357             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1358             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1359         }
1360     }
1361 }
1362
1363 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
1364     int i, si, di;
1365     uint8_t *dst;
1366     int bufidx;
1367
1368 //    src[0]&0x80;                //forbidden bit
1369     h->nal_ref_idc= src[0]>>5;
1370     h->nal_unit_type= src[0]&0x1F;
1371
1372     src++; length--;
1373 #if 0
1374     for(i=0; i<length; i++)
1375         printf("%2X ", src[i]);
1376 #endif
1377
1378 #if HAVE_FAST_UNALIGNED
1379 # if HAVE_FAST_64BIT
1380 #   define RS 7
1381     for(i=0; i+1<length; i+=9){
1382         if(!((~*(const uint64_t*)(src+i) & (*(const uint64_t*)(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
1383 # else
1384 #   define RS 3
1385     for(i=0; i+1<length; i+=5){
1386         if(!((~*(const uint32_t*)(src+i) & (*(const uint32_t*)(src+i) - 0x01000101U)) & 0x80008080U))
1387 # endif
1388             continue;
1389         if(i>0 && !src[i]) i--;
1390         while(src[i]) i++;
1391 #else
1392 #   define RS 0
1393     for(i=0; i+1<length; i+=2){
1394         if(src[i]) continue;
1395         if(i>0 && src[i-1]==0) i--;
1396 #endif
1397         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1398             if(src[i+2]!=3){
1399                 /* startcode, so we must be past the end */
1400                 length=i;
1401             }
1402             break;
1403         }
1404         i-= RS;
1405     }
1406
1407     if(i>=length-1){ //no escaped 0
1408         *dst_length= length;
1409         *consumed= length+1; //+1 for the header
1410         return src;
1411     }
1412
1413     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1414     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
1415     dst= h->rbsp_buffer[bufidx];
1416
1417     if (dst == NULL){
1418         return NULL;
1419     }
1420
1421 //printf("decoding esc\n");
1422     memcpy(dst, src, i);
1423     si=di=i;
1424     while(si+2<length){
1425         //remove escapes (very rare 1:2^22)
1426         if(src[si+2]>3){
1427             dst[di++]= src[si++];
1428             dst[di++]= src[si++];
1429         }else if(src[si]==0 && src[si+1]==0){
1430             if(src[si+2]==3){ //escape
1431                 dst[di++]= 0;
1432                 dst[di++]= 0;
1433                 si+=3;
1434                 continue;
1435             }else //next start code
1436                 goto nsc;
1437         }
1438
1439         dst[di++]= src[si++];
1440     }
1441     while(si<length)
1442         dst[di++]= src[si++];
1443 nsc:
1444
1445     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
1446
1447     *dst_length= di;
1448     *consumed= si + 1;//+1 for the header
1449 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1450     return dst;
1451 }
1452
1453 int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
1454     int v= *src;
1455     int r;
1456
1457     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1458
1459     for(r=1; r<9; r++){
1460         if(v&1) return r;
1461         v>>=1;
1462     }
1463     return 0;
1464 }
1465
1466 /**
1467  * IDCT transforms the 16 dc values and dequantizes them.
1468  * @param qp quantization parameter
1469  */
1470 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1471 #define stride 16
1472     int i;
1473     int temp[16]; //FIXME check if this is a good idea
1474     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1475     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1476
1477 //memset(block, 64, 2*256);
1478 //return;
1479     for(i=0; i<4; i++){
1480         const int offset= y_offset[i];
1481         const int z0= block[offset+stride*0] + block[offset+stride*4];
1482         const int z1= block[offset+stride*0] - block[offset+stride*4];
1483         const int z2= block[offset+stride*1] - block[offset+stride*5];
1484         const int z3= block[offset+stride*1] + block[offset+stride*5];
1485
1486         temp[4*i+0]= z0+z3;
1487         temp[4*i+1]= z1+z2;
1488         temp[4*i+2]= z1-z2;
1489         temp[4*i+3]= z0-z3;
1490     }
1491
1492     for(i=0; i<4; i++){
1493         const int offset= x_offset[i];
1494         const int z0= temp[4*0+i] + temp[4*2+i];
1495         const int z1= temp[4*0+i] - temp[4*2+i];
1496         const int z2= temp[4*1+i] - temp[4*3+i];
1497         const int z3= temp[4*1+i] + temp[4*3+i];
1498
1499         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
1500         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1501         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1502         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1503     }
1504 }
1505
1506 #if 0
1507 /**
1508  * DCT transforms the 16 dc values.
1509  * @param qp quantization parameter ??? FIXME
1510  */
1511 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1512 //    const int qmul= dequant_coeff[qp][0];
1513     int i;
1514     int temp[16]; //FIXME check if this is a good idea
1515     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1516     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1517
1518     for(i=0; i<4; i++){
1519         const int offset= y_offset[i];
1520         const int z0= block[offset+stride*0] + block[offset+stride*4];
1521         const int z1= block[offset+stride*0] - block[offset+stride*4];
1522         const int z2= block[offset+stride*1] - block[offset+stride*5];
1523         const int z3= block[offset+stride*1] + block[offset+stride*5];
1524
1525         temp[4*i+0]= z0+z3;
1526         temp[4*i+1]= z1+z2;
1527         temp[4*i+2]= z1-z2;
1528         temp[4*i+3]= z0-z3;
1529     }
1530
1531     for(i=0; i<4; i++){
1532         const int offset= x_offset[i];
1533         const int z0= temp[4*0+i] + temp[4*2+i];
1534         const int z1= temp[4*0+i] - temp[4*2+i];
1535         const int z2= temp[4*1+i] - temp[4*3+i];
1536         const int z3= temp[4*1+i] + temp[4*3+i];
1537
1538         block[stride*0 +offset]= (z0 + z3)>>1;
1539         block[stride*2 +offset]= (z1 + z2)>>1;
1540         block[stride*8 +offset]= (z1 - z2)>>1;
1541         block[stride*10+offset]= (z0 - z3)>>1;
1542     }
1543 }
1544 #endif
1545
1546 #undef xStride
1547 #undef stride
1548
1549 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1550     const int stride= 16*2;
1551     const int xStride= 16;
1552     int a,b,c,d,e;
1553
1554     a= block[stride*0 + xStride*0];
1555     b= block[stride*0 + xStride*1];
1556     c= block[stride*1 + xStride*0];
1557     d= block[stride*1 + xStride*1];
1558
1559     e= a-b;
1560     a= a+b;
1561     b= c-d;
1562     c= c+d;
1563
1564     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1565     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1566     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1567     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1568 }
1569
1570 #if 0
1571 static void chroma_dc_dct_c(DCTELEM *block){
1572     const int stride= 16*2;
1573     const int xStride= 16;
1574     int a,b,c,d,e;
1575
1576     a= block[stride*0 + xStride*0];
1577     b= block[stride*0 + xStride*1];
1578     c= block[stride*1 + xStride*0];
1579     d= block[stride*1 + xStride*1];
1580
1581     e= a-b;
1582     a= a+b;
1583     b= c-d;
1584     c= c+d;
1585
1586     block[stride*0 + xStride*0]= (a+c);
1587     block[stride*0 + xStride*1]= (e+b);
1588     block[stride*1 + xStride*0]= (a-c);
1589     block[stride*1 + xStride*1]= (e-b);
1590 }
1591 #endif
1592
1593 /**
1594  * gets the chroma qp.
1595  */
1596 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1597     return h->pps.chroma_qp_table[t][qscale];
1598 }
1599
1600 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1601                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1602                            int src_x_offset, int src_y_offset,
1603                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1604     MpegEncContext * const s = &h->s;
1605     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1606     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1607     const int luma_xy= (mx&3) + ((my&3)<<2);
1608     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1609     uint8_t * src_cb, * src_cr;
1610     int extra_width= h->emu_edge_width;
1611     int extra_height= h->emu_edge_height;
1612     int emu=0;
1613     const int full_mx= mx>>2;
1614     const int full_my= my>>2;
1615     const int pic_width  = 16*s->mb_width;
1616     const int pic_height = 16*s->mb_height >> MB_FIELD;
1617
1618     if(mx&7) extra_width -= 3;
1619     if(my&7) extra_height -= 3;
1620
1621     if(   full_mx < 0-extra_width
1622        || full_my < 0-extra_height
1623        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1624        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1625         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);
1626             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1627         emu=1;
1628     }
1629
1630     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1631     if(!square){
1632         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1633     }
1634
1635     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1636
1637     if(MB_FIELD){
1638         // chroma offset when predicting from a field of opposite parity
1639         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1640         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1641     }
1642     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1643     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1644
1645     if(emu){
1646         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);
1647             src_cb= s->edge_emu_buffer;
1648     }
1649     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1650
1651     if(emu){
1652         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);
1653             src_cr= s->edge_emu_buffer;
1654     }
1655     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1656 }
1657
1658 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1659                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1660                            int x_offset, int y_offset,
1661                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1662                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1663                            int list0, int list1){
1664     MpegEncContext * const s = &h->s;
1665     qpel_mc_func *qpix_op=  qpix_put;
1666     h264_chroma_mc_func chroma_op= chroma_put;
1667
1668     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1669     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1670     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1671     x_offset += 8*s->mb_x;
1672     y_offset += 8*(s->mb_y >> MB_FIELD);
1673
1674     if(list0){
1675         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1676         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1677                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1678                            qpix_op, chroma_op);
1679
1680         qpix_op=  qpix_avg;
1681         chroma_op= chroma_avg;
1682     }
1683
1684     if(list1){
1685         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1686         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1687                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1688                            qpix_op, chroma_op);
1689     }
1690 }
1691
1692 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1693                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1694                            int x_offset, int y_offset,
1695                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1696                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1697                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1698                            int list0, int list1){
1699     MpegEncContext * const s = &h->s;
1700
1701     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1702     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1703     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1704     x_offset += 8*s->mb_x;
1705     y_offset += 8*(s->mb_y >> MB_FIELD);
1706
1707     if(list0 && list1){
1708         /* don't optimize for luma-only case, since B-frames usually
1709          * use implicit weights => chroma too. */
1710         uint8_t *tmp_cb = s->obmc_scratchpad;
1711         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1712         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1713         int refn0 = h->ref_cache[0][ scan8[n] ];
1714         int refn1 = h->ref_cache[1][ scan8[n] ];
1715
1716         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1717                     dest_y, dest_cb, dest_cr,
1718                     x_offset, y_offset, qpix_put, chroma_put);
1719         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1720                     tmp_y, tmp_cb, tmp_cr,
1721                     x_offset, y_offset, qpix_put, chroma_put);
1722
1723         if(h->use_weight == 2){
1724             int weight0 = h->implicit_weight[refn0][refn1];
1725             int weight1 = 64 - weight0;
1726             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1727             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1728             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1729         }else{
1730             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1731                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1732                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1733             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1734                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1735                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1736             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1737                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1738                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1739         }
1740     }else{
1741         int list = list1 ? 1 : 0;
1742         int refn = h->ref_cache[list][ scan8[n] ];
1743         Picture *ref= &h->ref_list[list][refn];
1744         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1745                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1746                     qpix_put, chroma_put);
1747
1748         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1749                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1750         if(h->use_weight_chroma){
1751             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1752                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1753             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1754                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1755         }
1756     }
1757 }
1758
1759 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1760                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1761                            int x_offset, int y_offset,
1762                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1763                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1764                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1765                            int list0, int list1){
1766     if((h->use_weight==2 && list0 && list1
1767         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1768        || h->use_weight==1)
1769         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1770                          x_offset, y_offset, qpix_put, chroma_put,
1771                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1772     else
1773         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1774                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1775 }
1776
1777 static inline void prefetch_motion(H264Context *h, int list){
1778     /* fetch pixels for estimated mv 4 macroblocks ahead
1779      * optimized for 64byte cache lines */
1780     MpegEncContext * const s = &h->s;
1781     const int refn = h->ref_cache[list][scan8[0]];
1782     if(refn >= 0){
1783         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1784         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1785         uint8_t **src= h->ref_list[list][refn].data;
1786         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1787         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1788         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1789         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1790     }
1791 }
1792
1793 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1794                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1795                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1796                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1797     MpegEncContext * const s = &h->s;
1798     const int mb_xy= h->mb_xy;
1799     const int mb_type= s->current_picture.mb_type[mb_xy];
1800
1801     assert(IS_INTER(mb_type));
1802
1803     prefetch_motion(h, 0);
1804
1805     if(IS_16X16(mb_type)){
1806         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1807                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1808                 &weight_op[0], &weight_avg[0],
1809                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1810     }else if(IS_16X8(mb_type)){
1811         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1812                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1813                 &weight_op[1], &weight_avg[1],
1814                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1815         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1816                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1817                 &weight_op[1], &weight_avg[1],
1818                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1819     }else if(IS_8X16(mb_type)){
1820         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1821                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1822                 &weight_op[2], &weight_avg[2],
1823                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1824         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1825                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1826                 &weight_op[2], &weight_avg[2],
1827                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1828     }else{
1829         int i;
1830
1831         assert(IS_8X8(mb_type));
1832
1833         for(i=0; i<4; i++){
1834             const int sub_mb_type= h->sub_mb_type[i];
1835             const int n= 4*i;
1836             int x_offset= (i&1)<<2;
1837             int y_offset= (i&2)<<1;
1838
1839             if(IS_SUB_8X8(sub_mb_type)){
1840                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1841                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1842                     &weight_op[3], &weight_avg[3],
1843                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1844             }else if(IS_SUB_8X4(sub_mb_type)){
1845                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1846                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1847                     &weight_op[4], &weight_avg[4],
1848                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1849                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1850                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1851                     &weight_op[4], &weight_avg[4],
1852                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1853             }else if(IS_SUB_4X8(sub_mb_type)){
1854                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1855                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1856                     &weight_op[5], &weight_avg[5],
1857                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1858                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1859                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1860                     &weight_op[5], &weight_avg[5],
1861                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1862             }else{
1863                 int j;
1864                 assert(IS_SUB_4X4(sub_mb_type));
1865                 for(j=0; j<4; j++){
1866                     int sub_x_offset= x_offset + 2*(j&1);
1867                     int sub_y_offset= y_offset +   (j&2);
1868                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1869                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1870                         &weight_op[6], &weight_avg[6],
1871                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1872                 }
1873             }
1874         }
1875     }
1876
1877     prefetch_motion(h, 1);
1878 }
1879
1880 static av_cold void init_cavlc_level_tab(void){
1881     int suffix_length, mask;
1882     unsigned int i;
1883
1884     for(suffix_length=0; suffix_length<7; suffix_length++){
1885         for(i=0; i<(1<<LEVEL_TAB_BITS); i++){
1886             int prefix= LEVEL_TAB_BITS - av_log2(2*i);
1887             int level_code= (prefix<<suffix_length) + (i>>(LEVEL_TAB_BITS-prefix-1-suffix_length)) - (1<<suffix_length);
1888
1889             mask= -(level_code&1);
1890             level_code= (((2+level_code)>>1) ^ mask) - mask;
1891             if(prefix + 1 + suffix_length <= LEVEL_TAB_BITS){
1892                 cavlc_level_tab[suffix_length][i][0]= level_code;
1893                 cavlc_level_tab[suffix_length][i][1]= prefix + 1 + suffix_length;
1894             }else if(prefix + 1 <= LEVEL_TAB_BITS){
1895                 cavlc_level_tab[suffix_length][i][0]= prefix+100;
1896                 cavlc_level_tab[suffix_length][i][1]= prefix + 1;
1897             }else{
1898                 cavlc_level_tab[suffix_length][i][0]= LEVEL_TAB_BITS+100;
1899                 cavlc_level_tab[suffix_length][i][1]= LEVEL_TAB_BITS;
1900             }
1901         }
1902     }
1903 }
1904
1905 static av_cold void decode_init_vlc(void){
1906     static int done = 0;
1907
1908     if (!done) {
1909         int i;
1910         int offset;
1911         done = 1;
1912
1913         chroma_dc_coeff_token_vlc.table = chroma_dc_coeff_token_vlc_table;
1914         chroma_dc_coeff_token_vlc.table_allocated = chroma_dc_coeff_token_vlc_table_size;
1915         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1916                  &chroma_dc_coeff_token_len [0], 1, 1,
1917                  &chroma_dc_coeff_token_bits[0], 1, 1,
1918                  INIT_VLC_USE_NEW_STATIC);
1919
1920         offset = 0;
1921         for(i=0; i<4; i++){
1922             coeff_token_vlc[i].table = coeff_token_vlc_tables+offset;
1923             coeff_token_vlc[i].table_allocated = coeff_token_vlc_tables_size[i];
1924             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1925                      &coeff_token_len [i][0], 1, 1,
1926                      &coeff_token_bits[i][0], 1, 1,
1927                      INIT_VLC_USE_NEW_STATIC);
1928             offset += coeff_token_vlc_tables_size[i];
1929         }
1930         /*
1931          * This is a one time safety check to make sure that
1932          * the packed static coeff_token_vlc table sizes
1933          * were initialized correctly.
1934          */
1935         assert(offset == FF_ARRAY_ELEMS(coeff_token_vlc_tables));
1936
1937         for(i=0; i<3; i++){
1938             chroma_dc_total_zeros_vlc[i].table = chroma_dc_total_zeros_vlc_tables[i];
1939             chroma_dc_total_zeros_vlc[i].table_allocated = chroma_dc_total_zeros_vlc_tables_size;
1940             init_vlc(&chroma_dc_total_zeros_vlc[i],
1941                      CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1942                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1943                      &chroma_dc_total_zeros_bits[i][0], 1, 1,
1944                      INIT_VLC_USE_NEW_STATIC);
1945         }
1946         for(i=0; i<15; i++){
1947             total_zeros_vlc[i].table = total_zeros_vlc_tables[i];
1948             total_zeros_vlc[i].table_allocated = total_zeros_vlc_tables_size;
1949             init_vlc(&total_zeros_vlc[i],
1950                      TOTAL_ZEROS_VLC_BITS, 16,
1951                      &total_zeros_len [i][0], 1, 1,
1952                      &total_zeros_bits[i][0], 1, 1,
1953                      INIT_VLC_USE_NEW_STATIC);
1954         }
1955
1956         for(i=0; i<6; i++){
1957             run_vlc[i].table = run_vlc_tables[i];
1958             run_vlc[i].table_allocated = run_vlc_tables_size;
1959             init_vlc(&run_vlc[i],
1960                      RUN_VLC_BITS, 7,
1961                      &run_len [i][0], 1, 1,
1962                      &run_bits[i][0], 1, 1,
1963                      INIT_VLC_USE_NEW_STATIC);
1964         }
1965         run7_vlc.table = run7_vlc_table,
1966         run7_vlc.table_allocated = run7_vlc_table_size;
1967         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
1968                  &run_len [6][0], 1, 1,
1969                  &run_bits[6][0], 1, 1,
1970                  INIT_VLC_USE_NEW_STATIC);
1971
1972         init_cavlc_level_tab();
1973     }
1974 }
1975
1976 static void free_tables(H264Context *h){
1977     int i;
1978     H264Context *hx;
1979     av_freep(&h->intra4x4_pred_mode);
1980     av_freep(&h->chroma_pred_mode_table);
1981     av_freep(&h->cbp_table);
1982     av_freep(&h->mvd_table[0]);
1983     av_freep(&h->mvd_table[1]);
1984     av_freep(&h->direct_table);
1985     av_freep(&h->non_zero_count);
1986     av_freep(&h->slice_table_base);
1987     h->slice_table= NULL;
1988
1989     av_freep(&h->mb2b_xy);
1990     av_freep(&h->mb2b8_xy);
1991
1992     for(i = 0; i < h->s.avctx->thread_count; i++) {
1993         hx = h->thread_context[i];
1994         if(!hx) continue;
1995         av_freep(&hx->top_borders[1]);
1996         av_freep(&hx->top_borders[0]);
1997         av_freep(&hx->s.obmc_scratchpad);
1998     }
1999 }
2000
2001 static void init_dequant8_coeff_table(H264Context *h){
2002     int i,q,x;
2003     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2004     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2005     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2006
2007     for(i=0; i<2; i++ ){
2008         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2009             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2010             break;
2011         }
2012
2013         for(q=0; q<52; q++){
2014             int shift = div6[q];
2015             int idx = rem6[q];
2016             for(x=0; x<64; x++)
2017                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2018                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2019                     h->pps.scaling_matrix8[i][x]) << shift;
2020         }
2021     }
2022 }
2023
2024 static void init_dequant4_coeff_table(H264Context *h){
2025     int i,j,q,x;
2026     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2027     for(i=0; i<6; i++ ){
2028         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2029         for(j=0; j<i; j++){
2030             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2031                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2032                 break;
2033             }
2034         }
2035         if(j<i)
2036             continue;
2037
2038         for(q=0; q<52; q++){
2039             int shift = div6[q] + 2;
2040             int idx = rem6[q];
2041             for(x=0; x<16; x++)
2042                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2043                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2044                     h->pps.scaling_matrix4[i][x]) << shift;
2045         }
2046     }
2047 }
2048
2049 static void init_dequant_tables(H264Context *h){
2050     int i,x;
2051     init_dequant4_coeff_table(h);
2052     if(h->pps.transform_8x8_mode)
2053         init_dequant8_coeff_table(h);
2054     if(h->sps.transform_bypass){
2055         for(i=0; i<6; i++)
2056             for(x=0; x<16; x++)
2057                 h->dequant4_coeff[i][0][x] = 1<<6;
2058         if(h->pps.transform_8x8_mode)
2059             for(i=0; i<2; i++)
2060                 for(x=0; x<64; x++)
2061                     h->dequant8_coeff[i][0][x] = 1<<6;
2062     }
2063 }
2064
2065
2066 /**
2067  * allocates tables.
2068  * needs width/height
2069  */
2070 static int alloc_tables(H264Context *h){
2071     MpegEncContext * const s = &h->s;
2072     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2073     int x,y;
2074
2075     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2076
2077     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2078     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base))
2079     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2080
2081     CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2082     CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2083     CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2084     CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2085
2086     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
2087     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2088
2089     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2090     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2091     for(y=0; y<s->mb_height; y++){
2092         for(x=0; x<s->mb_width; x++){
2093             const int mb_xy= x + y*s->mb_stride;
2094             const int b_xy = 4*x + 4*y*h->b_stride;
2095             const int b8_xy= 2*x + 2*y*h->b8_stride;
2096
2097             h->mb2b_xy [mb_xy]= b_xy;
2098             h->mb2b8_xy[mb_xy]= b8_xy;
2099         }
2100     }
2101
2102     s->obmc_scratchpad = NULL;
2103
2104     if(!h->dequant4_coeff[0])
2105         init_dequant_tables(h);
2106
2107     return 0;
2108 fail:
2109     free_tables(h);
2110     return -1;
2111 }
2112
2113 /**
2114  * Mimic alloc_tables(), but for every context thread.
2115  */
2116 static void clone_tables(H264Context *dst, H264Context *src){
2117     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2118     dst->non_zero_count           = src->non_zero_count;
2119     dst->slice_table              = src->slice_table;
2120     dst->cbp_table                = src->cbp_table;
2121     dst->mb2b_xy                  = src->mb2b_xy;
2122     dst->mb2b8_xy                 = src->mb2b8_xy;
2123     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2124     dst->mvd_table[0]             = src->mvd_table[0];
2125     dst->mvd_table[1]             = src->mvd_table[1];
2126     dst->direct_table             = src->direct_table;
2127
2128     dst->s.obmc_scratchpad = NULL;
2129     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2130 }
2131
2132 /**
2133  * Init context
2134  * Allocate buffers which are not shared amongst multiple threads.
2135  */
2136 static int context_init(H264Context *h){
2137     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2138     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2139
2140     return 0;
2141 fail:
2142     return -1; // free_tables will clean up for us
2143 }
2144
2145 static av_cold void common_init(H264Context *h){
2146     MpegEncContext * const s = &h->s;
2147
2148     s->width = s->avctx->width;
2149     s->height = s->avctx->height;
2150     s->codec_id= s->avctx->codec->id;
2151
2152     ff_h264_pred_init(&h->hpc, s->codec_id);
2153
2154     h->dequant_coeff_pps= -1;
2155     s->unrestricted_mv=1;
2156     s->decode=1; //FIXME
2157
2158     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
2159
2160     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2161     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2162 }
2163
2164 /**
2165  * Reset SEI values at the beginning of the frame.
2166  *
2167  * @param h H.264 context.
2168  */
2169 static void reset_sei(H264Context *h) {
2170     h->sei_recovery_frame_cnt       = -1;
2171     h->sei_dpb_output_delay         =  0;
2172     h->sei_cpb_removal_delay        = -1;
2173     h->sei_buffering_period_present =  0;
2174 }
2175
2176 static av_cold int decode_init(AVCodecContext *avctx){
2177     H264Context *h= avctx->priv_data;
2178     MpegEncContext * const s = &h->s;
2179
2180     MPV_decode_defaults(s);
2181
2182     s->avctx = avctx;
2183     common_init(h);
2184
2185     s->out_format = FMT_H264;
2186     s->workaround_bugs= avctx->workaround_bugs;
2187
2188     // set defaults
2189 //    s->decode_mb= ff_h263_decode_mb;
2190     s->quarter_sample = 1;
2191     if(!avctx->has_b_frames)
2192     s->low_delay= 1;
2193
2194     if(s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2195         avctx->pix_fmt= PIX_FMT_VDPAU_H264;
2196     else
2197         avctx->pix_fmt= avctx->get_format(avctx, avctx->codec->pix_fmts);
2198     avctx->hwaccel = ff_find_hwaccel(avctx->codec->id, avctx->pix_fmt);
2199
2200     decode_init_vlc();
2201
2202     if(avctx->extradata_size > 0 && avctx->extradata &&
2203        *(char *)avctx->extradata == 1){
2204         h->is_avc = 1;
2205         h->got_avcC = 0;
2206     } else {
2207         h->is_avc = 0;
2208     }
2209
2210     h->thread_context[0] = h;
2211     h->outputed_poc = INT_MIN;
2212     h->prev_poc_msb= 1<<16;
2213     reset_sei(h);
2214     if(avctx->codec_id == CODEC_ID_H264){
2215         if(avctx->ticks_per_frame == 1){
2216             s->avctx->time_base.den *=2;
2217         }
2218         avctx->ticks_per_frame = 2;
2219     }
2220     return 0;
2221 }
2222
2223 static int frame_start(H264Context *h){
2224     MpegEncContext * const s = &h->s;
2225     int i;
2226
2227     if(MPV_frame_start(s, s->avctx) < 0)
2228         return -1;
2229     ff_er_frame_start(s);
2230     /*
2231      * MPV_frame_start uses pict_type to derive key_frame.
2232      * This is incorrect for H.264; IDR markings must be used.
2233      * Zero here; IDR markings per slice in frame or fields are ORed in later.
2234      * See decode_nal_units().
2235      */
2236     s->current_picture_ptr->key_frame= 0;
2237
2238     assert(s->linesize && s->uvlinesize);
2239
2240     for(i=0; i<16; i++){
2241         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2242         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2243     }
2244     for(i=0; i<4; i++){
2245         h->block_offset[16+i]=
2246         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2247         h->block_offset[24+16+i]=
2248         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2249     }
2250
2251     /* can't be in alloc_tables because linesize isn't known there.
2252      * FIXME: redo bipred weight to not require extra buffer? */
2253     for(i = 0; i < s->avctx->thread_count; i++)
2254         if(!h->thread_context[i]->s.obmc_scratchpad)
2255             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2256
2257     /* some macroblocks will be accessed before they're available */
2258     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2259         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
2260
2261 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2262
2263     // We mark the current picture as non-reference after allocating it, so
2264     // that if we break out due to an error it can be released automatically
2265     // in the next MPV_frame_start().
2266     // SVQ3 as well as most other codecs have only last/next/current and thus
2267     // get released even with set reference, besides SVQ3 and others do not
2268     // mark frames as reference later "naturally".
2269     if(s->codec_id != CODEC_ID_SVQ3)
2270         s->current_picture_ptr->reference= 0;
2271
2272     s->current_picture_ptr->field_poc[0]=
2273     s->current_picture_ptr->field_poc[1]= INT_MAX;
2274     assert(s->current_picture_ptr->long_ref==0);
2275
2276     return 0;
2277 }
2278
2279 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){
2280     MpegEncContext * const s = &h->s;
2281     int i;
2282     int step    = 1;
2283     int offset  = 1;
2284     int uvoffset= 1;
2285     int top_idx = 1;
2286     int skiplast= 0;
2287
2288     src_y  -=   linesize;
2289     src_cb -= uvlinesize;
2290     src_cr -= uvlinesize;
2291
2292     if(!simple && FRAME_MBAFF){
2293         if(s->mb_y&1){
2294             offset  = MB_MBAFF ? 1 : 17;
2295             uvoffset= MB_MBAFF ? 1 : 9;
2296             if(!MB_MBAFF){
2297                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 0)= *(uint64_t*)(src_y +  15*linesize);
2298                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 8)= *(uint64_t*)(src_y +8+15*linesize);
2299                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2300                     *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+7*uvlinesize);
2301                     *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+7*uvlinesize);
2302                 }
2303             }
2304         }else{
2305             if(!MB_MBAFF){
2306                 h->left_border[0]= h->top_borders[0][s->mb_x][15];
2307                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2308                     h->left_border[34   ]= h->top_borders[0][s->mb_x][16+7  ];
2309                     h->left_border[34+18]= h->top_borders[0][s->mb_x][16+8+7];
2310                 }
2311                 skiplast= 1;
2312             }
2313             offset  =
2314             uvoffset=
2315             top_idx = MB_MBAFF ? 0 : 1;
2316         }
2317         step= MB_MBAFF ? 2 : 1;
2318     }
2319
2320     // There are two lines saved, the line above the the top macroblock of a pair,
2321     // and the line above the bottom macroblock
2322     h->left_border[offset]= h->top_borders[top_idx][s->mb_x][15];
2323     for(i=1; i<17 - skiplast; i++){
2324         h->left_border[offset+i*step]= src_y[15+i*  linesize];
2325     }
2326
2327     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2328     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2329
2330     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2331         h->left_border[uvoffset+34   ]= h->top_borders[top_idx][s->mb_x][16+7];
2332         h->left_border[uvoffset+34+18]= h->top_borders[top_idx][s->mb_x][24+7];
2333         for(i=1; i<9 - skiplast; i++){
2334             h->left_border[uvoffset+34   +i*step]= src_cb[7+i*uvlinesize];
2335             h->left_border[uvoffset+34+18+i*step]= src_cr[7+i*uvlinesize];
2336         }
2337         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2338         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2339     }
2340 }
2341
2342 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){
2343     MpegEncContext * const s = &h->s;
2344     int temp8, i;
2345     uint64_t temp64;
2346     int deblock_left;
2347     int deblock_top;
2348     int mb_xy;
2349     int step    = 1;
2350     int offset  = 1;
2351     int uvoffset= 1;
2352     int top_idx = 1;
2353
2354     if(!simple && FRAME_MBAFF){
2355         if(s->mb_y&1){
2356             offset  = MB_MBAFF ? 1 : 17;
2357             uvoffset= MB_MBAFF ? 1 : 9;
2358         }else{
2359             offset  =
2360             uvoffset=
2361             top_idx = MB_MBAFF ? 0 : 1;
2362         }
2363         step= MB_MBAFF ? 2 : 1;
2364     }
2365
2366     if(h->deblocking_filter == 2) {
2367         mb_xy = h->mb_xy;
2368         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2369         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2370     } else {
2371         deblock_left = (s->mb_x > 0);
2372         deblock_top =  (s->mb_y > !!MB_FIELD);
2373     }
2374
2375     src_y  -=   linesize + 1;
2376     src_cb -= uvlinesize + 1;
2377     src_cr -= uvlinesize + 1;
2378
2379 #define XCHG(a,b,t,xchg)\
2380 t= a;\
2381 if(xchg)\
2382     a= b;\
2383 b= t;
2384
2385     if(deblock_left){
2386         for(i = !deblock_top; i<16; i++){
2387             XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, xchg);
2388         }
2389         XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, 1);
2390     }
2391
2392     if(deblock_top){
2393         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2394         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2395         if(s->mb_x+1 < s->mb_width){
2396             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2397         }
2398     }
2399
2400     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2401         if(deblock_left){
2402             for(i = !deblock_top; i<8; i++){
2403                 XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, xchg);
2404                 XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, xchg);
2405             }
2406             XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, 1);
2407             XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, 1);
2408         }
2409         if(deblock_top){
2410             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2411             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2412         }
2413     }
2414 }
2415
2416 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2417     MpegEncContext * const s = &h->s;
2418     const int mb_x= s->mb_x;
2419     const int mb_y= s->mb_y;
2420     const int mb_xy= h->mb_xy;
2421     const int mb_type= s->current_picture.mb_type[mb_xy];
2422     uint8_t  *dest_y, *dest_cb, *dest_cr;
2423     int linesize, uvlinesize /*dct_offset*/;
2424     int i;
2425     int *block_offset = &h->block_offset[0];
2426     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2427     /* is_h264 should always be true if SVQ3 is disabled. */
2428     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
2429     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2430     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2431
2432     dest_y  = s->current_picture.data[0] + (mb_x + mb_y * s->linesize  ) * 16;
2433     dest_cb = s->current_picture.data[1] + (mb_x + mb_y * s->uvlinesize) * 8;
2434     dest_cr = s->current_picture.data[2] + (mb_x + mb_y * s->uvlinesize) * 8;
2435
2436     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2437     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2438
2439     if (!simple && MB_FIELD) {
2440         linesize   = h->mb_linesize   = s->linesize * 2;
2441         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2442         block_offset = &h->block_offset[24];
2443         if(mb_y&1){ //FIXME move out of this function?
2444             dest_y -= s->linesize*15;
2445             dest_cb-= s->uvlinesize*7;
2446             dest_cr-= s->uvlinesize*7;
2447         }
2448         if(FRAME_MBAFF) {
2449             int list;
2450             for(list=0; list<h->list_count; list++){
2451                 if(!USES_LIST(mb_type, list))
2452                     continue;
2453                 if(IS_16X16(mb_type)){
2454                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2455                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2456                 }else{
2457                     for(i=0; i<16; i+=4){
2458                         int ref = h->ref_cache[list][scan8[i]];
2459                         if(ref >= 0)
2460                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2461                     }
2462                 }
2463             }
2464         }
2465     } else {
2466         linesize   = h->mb_linesize   = s->linesize;
2467         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2468 //        dct_offset = s->linesize * 16;
2469     }
2470
2471     if (!simple && IS_INTRA_PCM(mb_type)) {
2472         for (i=0; i<16; i++) {
2473             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
2474         }
2475         for (i=0; i<8; i++) {
2476             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
2477             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
2478         }
2479     } else {
2480         if(IS_INTRA(mb_type)){
2481             if(h->deblocking_filter)
2482                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2483
2484             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2485                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2486                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2487             }
2488
2489             if(IS_INTRA4x4(mb_type)){
2490                 if(simple || !s->encoding){
2491                     if(IS_8x8DCT(mb_type)){
2492                         if(transform_bypass){
2493                             idct_dc_add =
2494                             idct_add    = s->dsp.add_pixels8;
2495                         }else{
2496                             idct_dc_add = s->dsp.h264_idct8_dc_add;
2497                             idct_add    = s->dsp.h264_idct8_add;
2498                         }
2499                         for(i=0; i<16; i+=4){
2500                             uint8_t * const ptr= dest_y + block_offset[i];
2501                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2502                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
2503                                 h->hpc.pred8x8l_add[dir](ptr, h->mb + i*16, linesize);
2504                             }else{
2505                                 const int nnz = h->non_zero_count_cache[ scan8[i] ];
2506                                 h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2507                                                             (h->topright_samples_available<<i)&0x4000, linesize);
2508                                 if(nnz){
2509                                     if(nnz == 1 && h->mb[i*16])
2510                                         idct_dc_add(ptr, h->mb + i*16, linesize);
2511                                     else
2512                                         idct_add   (ptr, h->mb + i*16, linesize);
2513                                 }
2514                             }
2515                         }
2516                     }else{
2517                         if(transform_bypass){
2518                             idct_dc_add =
2519                             idct_add    = s->dsp.add_pixels4;
2520                         }else{
2521                             idct_dc_add = s->dsp.h264_idct_dc_add;
2522                             idct_add    = s->dsp.h264_idct_add;
2523                         }
2524                         for(i=0; i<16; i++){
2525                             uint8_t * const ptr= dest_y + block_offset[i];
2526                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2527
2528                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
2529                                 h->hpc.pred4x4_add[dir](ptr, h->mb + i*16, linesize);
2530                             }else{
2531                                 uint8_t *topright;
2532                                 int nnz, tr;
2533                                 if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2534                                     const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2535                                     assert(mb_y || linesize <= block_offset[i]);
2536                                     if(!topright_avail){
2537                                         tr= ptr[3 - linesize]*0x01010101;
2538                                         topright= (uint8_t*) &tr;
2539                                     }else
2540                                         topright= ptr + 4 - linesize;
2541                                 }else
2542                                     topright= NULL;
2543
2544                                 h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2545                                 nnz = h->non_zero_count_cache[ scan8[i] ];
2546                                 if(nnz){
2547                                     if(is_h264){
2548                                         if(nnz == 1 && h->mb[i*16])
2549                                             idct_dc_add(ptr, h->mb + i*16, linesize);
2550                                         else
2551                                             idct_add   (ptr, h->mb + i*16, linesize);
2552                                     }else
2553                                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2554                                 }
2555                             }
2556                         }
2557                     }
2558                 }
2559             }else{
2560                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2561                 if(is_h264){
2562                     if(!transform_bypass)
2563                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2564                 }else
2565                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2566             }
2567             if(h->deblocking_filter)
2568                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2569         }else if(is_h264){
2570             hl_motion(h, dest_y, dest_cb, dest_cr,
2571                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2572                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2573                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2574         }
2575
2576
2577         if(!IS_INTRA4x4(mb_type)){
2578             if(is_h264){
2579                 if(IS_INTRA16x16(mb_type)){
2580                     if(transform_bypass){
2581                         if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
2582                             h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb, linesize);
2583                         }else{
2584                             for(i=0; i<16; i++){
2585                                 if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
2586                                     s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + i*16, linesize);
2587                             }
2588                         }
2589                     }else{
2590                          s->dsp.h264_idct_add16intra(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2591                     }
2592                 }else if(h->cbp&15){
2593                     if(transform_bypass){
2594                         const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2595                         idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2596                         for(i=0; i<16; i+=di){
2597                             if(h->non_zero_count_cache[ scan8[i] ]){
2598                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2599                             }
2600                         }
2601                     }else{
2602                         if(IS_8x8DCT(mb_type)){
2603                             s->dsp.h264_idct8_add4(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2604                         }else{
2605                             s->dsp.h264_idct_add16(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2606                         }
2607                     }
2608                 }
2609             }else{
2610                 for(i=0; i<16; i++){
2611                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2612                         uint8_t * const ptr= dest_y + block_offset[i];
2613                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2614                     }
2615                 }
2616             }
2617         }
2618
2619         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
2620             uint8_t *dest[2] = {dest_cb, dest_cr};
2621             if(transform_bypass){
2622                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
2623                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + 16*16, uvlinesize);
2624                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 20, h->mb + 20*16, uvlinesize);
2625                 }else{
2626                     idct_add = s->dsp.add_pixels4;
2627                     for(i=16; i<16+8; i++){
2628                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
2629                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2630                     }
2631                 }
2632             }else{
2633                 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]);
2634                 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]);
2635                 if(is_h264){
2636                     idct_add = s->dsp.h264_idct_add;
2637                     idct_dc_add = s->dsp.h264_idct_dc_add;
2638                     for(i=16; i<16+8; i++){
2639                         if(h->non_zero_count_cache[ scan8[i] ])
2640                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2641                         else if(h->mb[i*16])
2642                             idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2643                     }
2644                 }else{
2645                     for(i=16; i<16+8; i++){
2646                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2647                             uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2648                             svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2649                         }
2650                     }
2651                 }
2652             }
2653         }
2654     }
2655     if(h->cbp || IS_INTRA(mb_type))
2656         s->dsp.clear_blocks(h->mb);
2657
2658     if(h->deblocking_filter) {
2659         backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2660         fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2661         h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2662         h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2663         if (!simple && FRAME_MBAFF) {
2664             filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2665         } else {
2666             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2667         }
2668     }
2669 }
2670
2671 /**
2672  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2673  */
2674 static void hl_decode_mb_simple(H264Context *h){
2675     hl_decode_mb_internal(h, 1);
2676 }
2677
2678 /**
2679  * Process a macroblock; this handles edge cases, such as interlacing.
2680  */
2681 static void av_noinline hl_decode_mb_complex(H264Context *h){
2682     hl_decode_mb_internal(h, 0);
2683 }
2684
2685 static void hl_decode_mb(H264Context *h){
2686     MpegEncContext * const s = &h->s;
2687     const int mb_xy= h->mb_xy;
2688     const int mb_type= s->current_picture.mb_type[mb_xy];
2689     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2690
2691     if (is_complex)
2692         hl_decode_mb_complex(h);
2693     else hl_decode_mb_simple(h);
2694 }
2695
2696 static void pic_as_field(Picture *pic, const int parity){
2697     int i;
2698     for (i = 0; i < 4; ++i) {
2699         if (parity == PICT_BOTTOM_FIELD)
2700             pic->data[i] += pic->linesize[i];
2701         pic->reference = parity;
2702         pic->linesize[i] *= 2;
2703     }
2704     pic->poc= pic->field_poc[parity == PICT_BOTTOM_FIELD];
2705 }
2706
2707 static int split_field_copy(Picture *dest, Picture *src,
2708                             int parity, int id_add){
2709     int match = !!(src->reference & parity);
2710
2711     if (match) {
2712         *dest = *src;
2713         if(parity != PICT_FRAME){
2714             pic_as_field(dest, parity);
2715             dest->pic_id *= 2;
2716             dest->pic_id += id_add;
2717         }
2718     }
2719
2720     return match;
2721 }
2722
2723 static int build_def_list(Picture *def, Picture **in, int len, int is_long, int sel){
2724     int i[2]={0};
2725     int index=0;
2726
2727     while(i[0]<len || i[1]<len){
2728         while(i[0]<len && !(in[ i[0] ] && (in[ i[0] ]->reference & sel)))
2729             i[0]++;
2730         while(i[1]<len && !(in[ i[1] ] && (in[ i[1] ]->reference & (sel^3))))
2731             i[1]++;
2732         if(i[0] < len){
2733             in[ i[0] ]->pic_id= is_long ? i[0] : in[ i[0] ]->frame_num;
2734             split_field_copy(&def[index++], in[ i[0]++ ], sel  , 1);
2735         }
2736         if(i[1] < len){
2737             in[ i[1] ]->pic_id= is_long ? i[1] : in[ i[1] ]->frame_num;
2738             split_field_copy(&def[index++], in[ i[1]++ ], sel^3, 0);
2739         }
2740     }
2741
2742     return index;
2743 }
2744
2745 static int add_sorted(Picture **sorted, Picture **src, int len, int limit, int dir){
2746     int i, best_poc;
2747     int out_i= 0;
2748
2749     for(;;){
2750         best_poc= dir ? INT_MIN : INT_MAX;
2751
2752         for(i=0; i<len; i++){
2753             const int poc= src[i]->poc;
2754             if(((poc > limit) ^ dir) && ((poc < best_poc) ^ dir)){
2755                 best_poc= poc;
2756                 sorted[out_i]= src[i];
2757             }
2758         }
2759         if(best_poc == (dir ? INT_MIN : INT_MAX))
2760             break;
2761         limit= sorted[out_i++]->poc - dir;
2762     }
2763     return out_i;
2764 }
2765
2766 /**
2767  * fills the default_ref_list.
2768  */
2769 static int fill_default_ref_list(H264Context *h){
2770     MpegEncContext * const s = &h->s;
2771     int i, len;
2772
2773     if(h->slice_type_nos==FF_B_TYPE){
2774         Picture *sorted[32];
2775         int cur_poc, list;
2776         int lens[2];
2777
2778         if(FIELD_PICTURE)
2779             cur_poc= s->current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
2780         else
2781             cur_poc= s->current_picture_ptr->poc;
2782
2783         for(list= 0; list<2; list++){
2784             len= add_sorted(sorted    , h->short_ref, h->short_ref_count, cur_poc, 1^list);
2785             len+=add_sorted(sorted+len, h->short_ref, h->short_ref_count, cur_poc, 0^list);
2786             assert(len<=32);
2787             len= build_def_list(h->default_ref_list[list]    , sorted     , len, 0, s->picture_structure);
2788             len+=build_def_list(h->default_ref_list[list]+len, h->long_ref, 16 , 1, s->picture_structure);
2789             assert(len<=32);
2790
2791             if(len < h->ref_count[list])
2792                 memset(&h->default_ref_list[list][len], 0, sizeof(Picture)*(h->ref_count[list] - len));
2793             lens[list]= len;
2794         }
2795
2796         if(lens[0] == lens[1] && lens[1] > 1){
2797             for(i=0; h->default_ref_list[0][i].data[0] == h->default_ref_list[1][i].data[0] && i<lens[0]; i++);
2798             if(i == lens[0])
2799                 FFSWAP(Picture, h->default_ref_list[1][0], h->default_ref_list[1][1]);
2800         }
2801     }else{
2802         len = build_def_list(h->default_ref_list[0]    , h->short_ref, h->short_ref_count, 0, s->picture_structure);
2803         len+= build_def_list(h->default_ref_list[0]+len, h-> long_ref, 16                , 1, s->picture_structure);
2804         assert(len <= 32);
2805         if(len < h->ref_count[0])
2806             memset(&h->default_ref_list[0][len], 0, sizeof(Picture)*(h->ref_count[0] - len));
2807     }
2808 #ifdef TRACE
2809     for (i=0; i<h->ref_count[0]; i++) {
2810         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]);
2811     }
2812     if(h->slice_type_nos==FF_B_TYPE){
2813         for (i=0; i<h->ref_count[1]; i++) {
2814             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]);
2815         }
2816     }
2817 #endif
2818     return 0;
2819 }
2820
2821 static void print_short_term(H264Context *h);
2822 static void print_long_term(H264Context *h);
2823
2824 /**
2825  * Extract structure information about the picture described by pic_num in
2826  * the current decoding context (frame or field). Note that pic_num is
2827  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
2828  * @param pic_num picture number for which to extract structure information
2829  * @param structure one of PICT_XXX describing structure of picture
2830  *                      with pic_num
2831  * @return frame number (short term) or long term index of picture
2832  *         described by pic_num
2833  */
2834 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
2835     MpegEncContext * const s = &h->s;
2836
2837     *structure = s->picture_structure;
2838     if(FIELD_PICTURE){
2839         if (!(pic_num & 1))
2840             /* opposite field */
2841             *structure ^= PICT_FRAME;
2842         pic_num >>= 1;
2843     }
2844
2845     return pic_num;
2846 }
2847
2848 static int decode_ref_pic_list_reordering(H264Context *h){
2849     MpegEncContext * const s = &h->s;
2850     int list, index, pic_structure;
2851
2852     print_short_term(h);
2853     print_long_term(h);
2854
2855     for(list=0; list<h->list_count; list++){
2856         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
2857
2858         if(get_bits1(&s->gb)){
2859             int pred= h->curr_pic_num;
2860
2861             for(index=0; ; index++){
2862                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb_31(&s->gb);
2863                 unsigned int pic_id;
2864                 int i;
2865                 Picture *ref = NULL;
2866
2867                 if(reordering_of_pic_nums_idc==3)
2868                     break;
2869
2870                 if(index >= h->ref_count[list]){
2871                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
2872                     return -1;
2873                 }
2874
2875                 if(reordering_of_pic_nums_idc<3){
2876                     if(reordering_of_pic_nums_idc<2){
2877                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
2878                         int frame_num;
2879
2880                         if(abs_diff_pic_num > h->max_pic_num){
2881                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
2882                             return -1;
2883                         }
2884
2885                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
2886                         else                                pred+= abs_diff_pic_num;
2887                         pred &= h->max_pic_num - 1;
2888
2889                         frame_num = pic_num_extract(h, pred, &pic_structure);
2890
2891                         for(i= h->short_ref_count-1; i>=0; i--){
2892                             ref = h->short_ref[i];
2893                             assert(ref->reference);
2894                             assert(!ref->long_ref);
2895                             if(
2896                                    ref->frame_num == frame_num &&
2897                                    (ref->reference & pic_structure)
2898                               )
2899                                 break;
2900                         }
2901                         if(i>=0)
2902                             ref->pic_id= pred;
2903                     }else{
2904                         int long_idx;
2905                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
2906
2907                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
2908
2909                         if(long_idx>31){
2910                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
2911                             return -1;
2912                         }
2913                         ref = h->long_ref[long_idx];
2914                         assert(!(ref && !ref->reference));
2915                         if(ref && (ref->reference & pic_structure)){
2916                             ref->pic_id= pic_id;
2917                             assert(ref->long_ref);
2918                             i=0;
2919                         }else{
2920                             i=-1;
2921                         }
2922                     }
2923
2924                     if (i < 0) {
2925                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
2926                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
2927                     } else {
2928                         for(i=index; i+1<h->ref_count[list]; i++){
2929                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
2930                                 break;
2931                         }
2932                         for(; i > index; i--){
2933                             h->ref_list[list][i]= h->ref_list[list][i-1];
2934                         }
2935                         h->ref_list[list][index]= *ref;
2936                         if (FIELD_PICTURE){
2937                             pic_as_field(&h->ref_list[list][index], pic_structure);
2938                         }
2939                     }
2940                 }else{
2941                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
2942                     return -1;
2943                 }
2944             }
2945         }
2946     }
2947     for(list=0; list<h->list_count; list++){
2948         for(index= 0; index < h->ref_count[list]; index++){
2949             if(!h->ref_list[list][index].data[0]){
2950                 av_log(h->s.avctx, AV_LOG_ERROR, "Missing reference picture\n");
2951                 h->ref_list[list][index]= s->current_picture; //FIXME this is not a sensible solution
2952             }
2953         }
2954     }
2955
2956     return 0;
2957 }
2958
2959 static void fill_mbaff_ref_list(H264Context *h){
2960     int list, i, j;
2961     for(list=0; list<2; list++){ //FIXME try list_count
2962         for(i=0; i<h->ref_count[list]; i++){
2963             Picture *frame = &h->ref_list[list][i];
2964             Picture *field = &h->ref_list[list][16+2*i];
2965             field[0] = *frame;
2966             for(j=0; j<3; j++)
2967                 field[0].linesize[j] <<= 1;
2968             field[0].reference = PICT_TOP_FIELD;
2969             field[0].poc= field[0].field_poc[0];
2970             field[1] = field[0];
2971             for(j=0; j<3; j++)
2972                 field[1].data[j] += frame->linesize[j];
2973             field[1].reference = PICT_BOTTOM_FIELD;
2974             field[1].poc= field[1].field_poc[1];
2975
2976             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
2977             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
2978             for(j=0; j<2; j++){
2979                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
2980                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
2981             }
2982         }
2983     }
2984     for(j=0; j<h->ref_count[1]; j++){
2985         for(i=0; i<h->ref_count[0]; i++)
2986             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
2987         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
2988         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
2989     }
2990 }
2991
2992 static int pred_weight_table(H264Context *h){
2993     MpegEncContext * const s = &h->s;
2994     int list, i;
2995     int luma_def, chroma_def;
2996
2997     h->use_weight= 0;
2998     h->use_weight_chroma= 0;
2999     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
3000     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
3001     luma_def = 1<<h->luma_log2_weight_denom;
3002     chroma_def = 1<<h->chroma_log2_weight_denom;
3003
3004     for(list=0; list<2; list++){
3005         h->luma_weight_flag[list]   = 0;
3006         h->chroma_weight_flag[list] = 0;
3007         for(i=0; i<h->ref_count[list]; i++){
3008             int luma_weight_flag, chroma_weight_flag;
3009
3010             luma_weight_flag= get_bits1(&s->gb);
3011             if(luma_weight_flag){
3012                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3013                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3014                 if(   h->luma_weight[list][i] != luma_def
3015                    || h->luma_offset[list][i] != 0) {
3016                     h->use_weight= 1;
3017                     h->luma_weight_flag[list]= 1;
3018                 }
3019             }else{
3020                 h->luma_weight[list][i]= luma_def;
3021                 h->luma_offset[list][i]= 0;
3022             }
3023
3024             if(CHROMA){
3025                 chroma_weight_flag= get_bits1(&s->gb);
3026                 if(chroma_weight_flag){
3027                     int j;
3028                     for(j=0; j<2; j++){
3029                         h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3030                         h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3031                         if(   h->chroma_weight[list][i][j] != chroma_def
3032                            || h->chroma_offset[list][i][j] != 0) {
3033                             h->use_weight_chroma= 1;
3034                             h->chroma_weight_flag[list]= 1;
3035                         }
3036                     }
3037                 }else{
3038                     int j;
3039                     for(j=0; j<2; j++){
3040                         h->chroma_weight[list][i][j]= chroma_def;
3041                         h->chroma_offset[list][i][j]= 0;
3042                     }
3043                 }
3044             }
3045         }
3046         if(h->slice_type_nos != FF_B_TYPE) break;
3047     }
3048     h->use_weight= h->use_weight || h->use_weight_chroma;
3049     return 0;
3050 }
3051
3052 static void implicit_weight_table(H264Context *h){
3053     MpegEncContext * const s = &h->s;
3054     int ref0, ref1, i;
3055     int cur_poc = s->current_picture_ptr->poc;
3056
3057     for (i = 0; i < 2; i++) {
3058         h->luma_weight_flag[i]   = 0;
3059         h->chroma_weight_flag[i] = 0;
3060     }
3061
3062     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3063        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3064         h->use_weight= 0;
3065         h->use_weight_chroma= 0;
3066         return;
3067     }
3068
3069     h->use_weight= 2;
3070     h->use_weight_chroma= 2;
3071     h->luma_log2_weight_denom= 5;
3072     h->chroma_log2_weight_denom= 5;
3073
3074     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3075         int poc0 = h->ref_list[0][ref0].poc;
3076         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3077             int poc1 = h->ref_list[1][ref1].poc;
3078             int td = av_clip(poc1 - poc0, -128, 127);
3079             if(td){
3080                 int tb = av_clip(cur_poc - poc0, -128, 127);
3081                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3082                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3083                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3084                     h->implicit_weight[ref0][ref1] = 32;
3085                 else
3086                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3087             }else
3088                 h->implicit_weight[ref0][ref1] = 32;
3089         }
3090     }
3091 }
3092
3093 /**
3094  * Mark a picture as no longer needed for reference. The refmask
3095  * argument allows unreferencing of individual fields or the whole frame.
3096  * If the picture becomes entirely unreferenced, but is being held for
3097  * display purposes, it is marked as such.
3098  * @param refmask mask of fields to unreference; the mask is bitwise
3099  *                anded with the reference marking of pic
3100  * @return non-zero if pic becomes entirely unreferenced (except possibly
3101  *         for display purposes) zero if one of the fields remains in
3102  *         reference
3103  */
3104 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3105     int i;
3106     if (pic->reference &= refmask) {
3107         return 0;
3108     } else {
3109         for(i = 0; h->delayed_pic[i]; i++)
3110             if(pic == h->delayed_pic[i]){
3111                 pic->reference=DELAYED_PIC_REF;
3112                 break;
3113             }
3114         return 1;
3115     }
3116 }
3117
3118 /**
3119  * instantaneous decoder refresh.
3120  */
3121 static void idr(H264Context *h){
3122     int i;
3123
3124     for(i=0; i<16; i++){
3125         remove_long(h, i, 0);
3126     }
3127     assert(h->long_ref_count==0);
3128
3129     for(i=0; i<h->short_ref_count; i++){
3130         unreference_pic(h, h->short_ref[i], 0);
3131         h->short_ref[i]= NULL;
3132     }
3133     h->short_ref_count=0;
3134     h->prev_frame_num= 0;
3135     h->prev_frame_num_offset= 0;
3136     h->prev_poc_msb=
3137     h->prev_poc_lsb= 0;
3138 }
3139
3140 /* forget old pics after a seek */
3141 static void flush_dpb(AVCodecContext *avctx){
3142     H264Context *h= avctx->priv_data;
3143     int i;
3144     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
3145         if(h->delayed_pic[i])
3146             h->delayed_pic[i]->reference= 0;
3147         h->delayed_pic[i]= NULL;
3148     }
3149     h->outputed_poc= INT_MIN;
3150     idr(h);
3151     if(h->s.current_picture_ptr)
3152         h->s.current_picture_ptr->reference= 0;
3153     h->s.first_field= 0;
3154     reset_sei(h);
3155     ff_mpeg_flush(avctx);
3156 }
3157
3158 /**
3159  * Find a Picture in the short term reference list by frame number.
3160  * @param frame_num frame number to search for
3161  * @param idx the index into h->short_ref where returned picture is found
3162  *            undefined if no picture found.
3163  * @return pointer to the found picture, or NULL if no pic with the provided
3164  *                 frame number is found
3165  */
3166 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3167     MpegEncContext * const s = &h->s;
3168     int i;
3169
3170     for(i=0; i<h->short_ref_count; i++){
3171         Picture *pic= h->short_ref[i];
3172         if(s->avctx->debug&FF_DEBUG_MMCO)
3173             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3174         if(pic->frame_num == frame_num) {
3175             *idx = i;
3176             return pic;
3177         }
3178     }
3179     return NULL;
3180 }
3181
3182 /**
3183  * Remove a picture from the short term reference list by its index in
3184  * that list.  This does no checking on the provided index; it is assumed
3185  * to be valid. Other list entries are shifted down.
3186  * @param i index into h->short_ref of picture to remove.
3187  */
3188 static void remove_short_at_index(H264Context *h, int i){
3189     a