Add VDPAU hardware accelerated decoding for H264 which can be used by
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file h264.c
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "dsputil.h"
29 #include "avcodec.h"
30 #include "mpegvideo.h"
31 #include "h264.h"
32 #include "h264data.h"
33 #include "h264_parser.h"
34 #include "golomb.h"
35 #include "rectangle.h"
36 #include "vdpau_internal.h"
37
38 #include "cabac.h"
39 #ifdef ARCH_X86
40 #include "x86/h264_i386.h"
41 #endif
42
43 //#undef NDEBUG
44 #include <assert.h>
45
46 /**
47  * Value of Picture.reference when Picture is not a reference picture, but
48  * is held for delayed output.
49  */
50 #define DELAYED_PIC_REF 4
51
52 static VLC coeff_token_vlc[4];
53 static VLC_TYPE coeff_token_vlc_tables[520+332+280+256][2];
54 static const int coeff_token_vlc_tables_size[4]={520,332,280,256};
55
56 static VLC chroma_dc_coeff_token_vlc;
57 static VLC_TYPE chroma_dc_coeff_token_vlc_table[256][2];
58 static const int chroma_dc_coeff_token_vlc_table_size = 256;
59
60 static VLC total_zeros_vlc[15];
61 static VLC_TYPE total_zeros_vlc_tables[15][512][2];
62 static const int total_zeros_vlc_tables_size = 512;
63
64 static VLC chroma_dc_total_zeros_vlc[3];
65 static VLC_TYPE chroma_dc_total_zeros_vlc_tables[3][8][2];
66 static const int chroma_dc_total_zeros_vlc_tables_size = 8;
67
68 static VLC run_vlc[6];
69 static VLC_TYPE run_vlc_tables[6][8][2];
70 static const int run_vlc_tables_size = 8;
71
72 static VLC run7_vlc;
73 static VLC_TYPE run7_vlc_table[96][2];
74 static const int run7_vlc_table_size = 96;
75
76 static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
77 static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
78 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);
79 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);
80 static Picture * remove_long(H264Context *h, int i, int ref_mask);
81
82 static av_always_inline uint32_t pack16to32(int a, int b){
83 #ifdef WORDS_BIGENDIAN
84    return (b&0xFFFF) + (a<<16);
85 #else
86    return (a&0xFFFF) + (b<<16);
87 #endif
88 }
89
90 static const uint8_t rem6[52]={
91 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,
92 };
93
94 static const uint8_t div6[52]={
95 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,
96 };
97
98 static const int left_block_options[4][8]={
99     {0,1,2,3,7,10,8,11},
100     {2,2,3,3,8,11,8,11},
101     {0,0,1,1,7,10,7,10},
102     {0,2,0,2,7,10,7,10}
103 };
104
105 #define LEVEL_TAB_BITS 8
106 static int8_t cavlc_level_tab[7][1<<LEVEL_TAB_BITS][2];
107
108 static void fill_caches(H264Context *h, int mb_type, int for_deblock){
109     MpegEncContext * const s = &h->s;
110     const int mb_xy= h->mb_xy;
111     int topleft_xy, top_xy, topright_xy, left_xy[2];
112     int topleft_type, top_type, topright_type, left_type[2];
113     const int * left_block;
114     int topleft_partition= -1;
115     int i;
116
117     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
118
119     //FIXME deblocking could skip the intra and nnz parts.
120     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
121         return;
122
123     /* Wow, what a mess, why didn't they simplify the interlacing & intra
124      * stuff, I can't imagine that these complex rules are worth it. */
125
126     topleft_xy = top_xy - 1;
127     topright_xy= top_xy + 1;
128     left_xy[1] = left_xy[0] = mb_xy-1;
129     left_block = left_block_options[0];
130     if(FRAME_MBAFF){
131         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
132         const int top_pair_xy      = pair_xy     - s->mb_stride;
133         const int topleft_pair_xy  = top_pair_xy - 1;
134         const int topright_pair_xy = top_pair_xy + 1;
135         const int topleft_mb_field_flag  = IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
136         const int top_mb_field_flag      = IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
137         const int topright_mb_field_flag = IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
138         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
139         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
140         const int bottom = (s->mb_y & 1);
141         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);
142
143         if (curr_mb_field_flag && (bottom || top_mb_field_flag)){
144             top_xy -= s->mb_stride;
145         }
146         if (curr_mb_field_flag && (bottom || topleft_mb_field_flag)){
147             topleft_xy -= s->mb_stride;
148         } else if(bottom && !curr_mb_field_flag && left_mb_field_flag) {
149             topleft_xy += s->mb_stride;
150             // take top left mv from the middle of the mb, as opposed to all other modes which use the bottom right partition
151             topleft_partition = 0;
152         }
153         if (curr_mb_field_flag && (bottom || topright_mb_field_flag)){
154             topright_xy -= s->mb_stride;
155         }
156         if (left_mb_field_flag != curr_mb_field_flag) {
157             left_xy[1] = left_xy[0] = pair_xy - 1;
158             if (curr_mb_field_flag) {
159                 left_xy[1] += s->mb_stride;
160                 left_block = left_block_options[3];
161             } else {
162                 left_block= left_block_options[2 - bottom];
163             }
164         }
165     }
166
167     h->top_mb_xy = top_xy;
168     h->left_mb_xy[0] = left_xy[0];
169     h->left_mb_xy[1] = left_xy[1];
170     if(for_deblock){
171         topleft_type = 0;
172         topright_type = 0;
173         top_type     = h->slice_table[top_xy     ] < 0xFFFF ? s->current_picture.mb_type[top_xy]     : 0;
174         left_type[0] = h->slice_table[left_xy[0] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[0]] : 0;
175         left_type[1] = h->slice_table[left_xy[1] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[1]] : 0;
176
177         if(MB_MBAFF && !IS_INTRA(mb_type)){
178             int list;
179             for(list=0; list<h->list_count; list++){
180                 //These values where changed for ease of performing MC, we need to change them back
181                 //FIXME maybe we can make MC and loop filter use the same values or prevent
182                 //the MC code from changing ref_cache and rather use a temporary array.
183                 if(USES_LIST(mb_type,list)){
184                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
185                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
186                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
187                     ref += h->b8_stride;
188                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
189                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
190                 }
191             }
192         }
193     }else{
194         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
195         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
196         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
197         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
198         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
199
200     if(IS_INTRA(mb_type)){
201         int type_mask= h->pps.constrained_intra_pred ? IS_INTRA(-1) : -1;
202         h->topleft_samples_available=
203         h->top_samples_available=
204         h->left_samples_available= 0xFFFF;
205         h->topright_samples_available= 0xEEEA;
206
207         if(!(top_type & type_mask)){
208             h->topleft_samples_available= 0xB3FF;
209             h->top_samples_available= 0x33FF;
210             h->topright_samples_available= 0x26EA;
211         }
212         if(IS_INTERLACED(mb_type) != IS_INTERLACED(left_type[0])){
213             if(IS_INTERLACED(mb_type)){
214                 if(!(left_type[0] & type_mask)){
215                     h->topleft_samples_available&= 0xDFFF;
216                     h->left_samples_available&= 0x5FFF;
217                 }
218                 if(!(left_type[1] & type_mask)){
219                     h->topleft_samples_available&= 0xFF5F;
220                     h->left_samples_available&= 0xFF5F;
221                 }
222             }else{
223                 int left_typei = h->slice_table[left_xy[0] + s->mb_stride ] == h->slice_num
224                                 ? s->current_picture.mb_type[left_xy[0] + s->mb_stride] : 0;
225                 assert(left_xy[0] == left_xy[1]);
226                 if(!((left_typei & type_mask) && (left_type[0] & type_mask))){
227                     h->topleft_samples_available&= 0xDF5F;
228                     h->left_samples_available&= 0x5F5F;
229                 }
230             }
231         }else{
232             if(!(left_type[0] & type_mask)){
233                 h->topleft_samples_available&= 0xDF5F;
234                 h->left_samples_available&= 0x5F5F;
235             }
236         }
237
238         if(!(topleft_type & type_mask))
239             h->topleft_samples_available&= 0x7FFF;
240
241         if(!(topright_type & type_mask))
242             h->topright_samples_available&= 0xFBFF;
243
244         if(IS_INTRA4x4(mb_type)){
245             if(IS_INTRA4x4(top_type)){
246                 h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
247                 h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
248                 h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
249                 h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
250             }else{
251                 int pred;
252                 if(!(top_type & type_mask))
253                     pred= -1;
254                 else{
255                     pred= 2;
256                 }
257                 h->intra4x4_pred_mode_cache[4+8*0]=
258                 h->intra4x4_pred_mode_cache[5+8*0]=
259                 h->intra4x4_pred_mode_cache[6+8*0]=
260                 h->intra4x4_pred_mode_cache[7+8*0]= pred;
261             }
262             for(i=0; i<2; i++){
263                 if(IS_INTRA4x4(left_type[i])){
264                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
265                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
266                 }else{
267                     int pred;
268                     if(!(left_type[i] & type_mask))
269                         pred= -1;
270                     else{
271                         pred= 2;
272                     }
273                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
274                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
275                 }
276             }
277         }
278     }
279     }
280
281
282 /*
283 0 . T T. T T T T
284 1 L . .L . . . .
285 2 L . .L . . . .
286 3 . T TL . . . .
287 4 L . .L . . . .
288 5 L . .. . . . .
289 */
290 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
291     if(top_type){
292         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
293         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
294         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
295         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
296
297         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
298         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
299
300         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
301         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
302
303     }else{
304         h->non_zero_count_cache[4+8*0]=
305         h->non_zero_count_cache[5+8*0]=
306         h->non_zero_count_cache[6+8*0]=
307         h->non_zero_count_cache[7+8*0]=
308
309         h->non_zero_count_cache[1+8*0]=
310         h->non_zero_count_cache[2+8*0]=
311
312         h->non_zero_count_cache[1+8*3]=
313         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
314
315     }
316
317     for (i=0; i<2; i++) {
318         if(left_type[i]){
319             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
320             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
321             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
322             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
323         }else{
324             h->non_zero_count_cache[3+8*1 + 2*8*i]=
325             h->non_zero_count_cache[3+8*2 + 2*8*i]=
326             h->non_zero_count_cache[0+8*1 +   8*i]=
327             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
328         }
329     }
330
331     if( h->pps.cabac ) {
332         // top_cbp
333         if(top_type) {
334             h->top_cbp = h->cbp_table[top_xy];
335         } else if(IS_INTRA(mb_type)) {
336             h->top_cbp = 0x1C0;
337         } else {
338             h->top_cbp = 0;
339         }
340         // left_cbp
341         if (left_type[0]) {
342             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
343         } else if(IS_INTRA(mb_type)) {
344             h->left_cbp = 0x1C0;
345         } else {
346             h->left_cbp = 0;
347         }
348         if (left_type[0]) {
349             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
350         }
351         if (left_type[1]) {
352             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
353         }
354     }
355
356 #if 1
357     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
358         int list;
359         for(list=0; list<h->list_count; list++){
360             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
361                 /*if(!h->mv_cache_clean[list]){
362                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
363                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
364                     h->mv_cache_clean[list]= 1;
365                 }*/
366                 continue;
367             }
368             h->mv_cache_clean[list]= 0;
369
370             if(USES_LIST(top_type, list)){
371                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
372                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
373                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
374                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
375                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
376                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
377                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
378                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
379                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
380                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
381             }else{
382                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
383                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
384                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
385                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
386                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
387             }
388
389             for(i=0; i<2; i++){
390                 int cache_idx = scan8[0] - 1 + i*2*8;
391                 if(USES_LIST(left_type[i], list)){
392                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
393                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
394                     *(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]];
395                     *(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]];
396                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
397                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
398                 }else{
399                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
400                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
401                     h->ref_cache[list][cache_idx  ]=
402                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
403                 }
404             }
405
406             if(for_deblock || ((IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred) && !FRAME_MBAFF))
407                 continue;
408
409             if(USES_LIST(topleft_type, list)){
410                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
411                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
412                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
413                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
414             }else{
415                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
416                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
417             }
418
419             if(USES_LIST(topright_type, list)){
420                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
421                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
422                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
423                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
424             }else{
425                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
426                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
427             }
428
429             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
430                 continue;
431
432             h->ref_cache[list][scan8[5 ]+1] =
433             h->ref_cache[list][scan8[7 ]+1] =
434             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
435             h->ref_cache[list][scan8[4 ]] =
436             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
437             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
438             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
439             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
440             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
441             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
442
443             if( h->pps.cabac ) {
444                 /* XXX beurk, Load mvd */
445                 if(USES_LIST(top_type, list)){
446                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
447                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
448                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
449                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
450                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
451                 }else{
452                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
453                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
454                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
455                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
456                 }
457                 if(USES_LIST(left_type[0], list)){
458                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
459                     *(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]];
460                     *(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]];
461                 }else{
462                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
463                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
464                 }
465                 if(USES_LIST(left_type[1], list)){
466                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
467                     *(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]];
468                     *(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]];
469                 }else{
470                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
471                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
472                 }
473                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
474                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
475                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
476                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
477                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
478
479                 if(h->slice_type_nos == FF_B_TYPE){
480                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
481
482                     if(IS_DIRECT(top_type)){
483                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
484                     }else if(IS_8X8(top_type)){
485                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
486                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
487                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
488                     }else{
489                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
490                     }
491
492                     if(IS_DIRECT(left_type[0]))
493                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
494                     else if(IS_8X8(left_type[0]))
495                         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)];
496                     else
497                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
498
499                     if(IS_DIRECT(left_type[1]))
500                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
501                     else if(IS_8X8(left_type[1]))
502                         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)];
503                     else
504                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
505                 }
506             }
507
508             if(FRAME_MBAFF){
509 #define MAP_MVS\
510                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
511                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
512                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
513                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
514                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
515                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
516                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
517                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
518                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
519                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
520                 if(MB_FIELD){
521 #define MAP_F2F(idx, mb_type)\
522                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
523                         h->ref_cache[list][idx] <<= 1;\
524                         h->mv_cache[list][idx][1] /= 2;\
525                         h->mvd_cache[list][idx][1] /= 2;\
526                     }
527                     MAP_MVS
528 #undef MAP_F2F
529                 }else{
530 #define MAP_F2F(idx, mb_type)\
531                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
532                         h->ref_cache[list][idx] >>= 1;\
533                         h->mv_cache[list][idx][1] <<= 1;\
534                         h->mvd_cache[list][idx][1] <<= 1;\
535                     }
536                     MAP_MVS
537 #undef MAP_F2F
538                 }
539             }
540         }
541     }
542 #endif
543
544     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
545 }
546
547 static inline void write_back_intra_pred_mode(H264Context *h){
548     const int mb_xy= h->mb_xy;
549
550     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
551     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
552     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
553     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
554     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
555     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
556     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
557 }
558
559 /**
560  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
561  */
562 static inline int check_intra4x4_pred_mode(H264Context *h){
563     MpegEncContext * const s = &h->s;
564     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
565     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
566     int i;
567
568     if(!(h->top_samples_available&0x8000)){
569         for(i=0; i<4; i++){
570             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
571             if(status<0){
572                 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);
573                 return -1;
574             } else if(status){
575                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
576             }
577         }
578     }
579
580     if((h->left_samples_available&0x8888)!=0x8888){
581         static const int mask[4]={0x8000,0x2000,0x80,0x20};
582         for(i=0; i<4; i++){
583             if(!(h->left_samples_available&mask[i])){
584                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
585                 if(status<0){
586                     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);
587                     return -1;
588                 } else if(status){
589                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
590                 }
591             }
592         }
593     }
594
595     return 0;
596 } //FIXME cleanup like next
597
598 /**
599  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
600  */
601 static inline int check_intra_pred_mode(H264Context *h, int mode){
602     MpegEncContext * const s = &h->s;
603     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
604     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
605
606     if(mode > 6U) {
607         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);
608         return -1;
609     }
610
611     if(!(h->top_samples_available&0x8000)){
612         mode= top[ mode ];
613         if(mode<0){
614             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);
615             return -1;
616         }
617     }
618
619     if((h->left_samples_available&0x8080) != 0x8080){
620         mode= left[ mode ];
621         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
622             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
623         }
624         if(mode<0){
625             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);
626             return -1;
627         }
628     }
629
630     return mode;
631 }
632
633 /**
634  * gets the predicted intra4x4 prediction mode.
635  */
636 static inline int pred_intra_mode(H264Context *h, int n){
637     const int index8= scan8[n];
638     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
639     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
640     const int min= FFMIN(left, top);
641
642     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
643
644     if(min<0) return DC_PRED;
645     else      return min;
646 }
647
648 static inline void write_back_non_zero_count(H264Context *h){
649     const int mb_xy= h->mb_xy;
650
651     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
652     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
653     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
654     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
655     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
656     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
657     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
658
659     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
660     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
661     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
662
663     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
664     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
665     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
666 }
667
668 /**
669  * gets the predicted number of non-zero coefficients.
670  * @param n block index
671  */
672 static inline int pred_non_zero_count(H264Context *h, int n){
673     const int index8= scan8[n];
674     const int left= h->non_zero_count_cache[index8 - 1];
675     const int top = h->non_zero_count_cache[index8 - 8];
676     int i= left + top;
677
678     if(i<64) i= (i+1)>>1;
679
680     tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);
681
682     return i&31;
683 }
684
685 static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
686     const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
687     MpegEncContext *s = &h->s;
688
689     /* there is no consistent mapping of mvs to neighboring locations that will
690      * make mbaff happy, so we can't move all this logic to fill_caches */
691     if(FRAME_MBAFF){
692         const uint32_t *mb_types = s->current_picture_ptr->mb_type;
693         const int16_t *mv;
694         *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
695         *C = h->mv_cache[list][scan8[0]-2];
696
697         if(!MB_FIELD
698            && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
699             int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
700             if(IS_INTERLACED(mb_types[topright_xy])){
701 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
702                 const int x4 = X4, y4 = Y4;\
703                 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
704                 if(!USES_LIST(mb_type,list))\
705                     return LIST_NOT_USED;\
706                 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
707                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
708                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
709                 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
710
711                 SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
712             }
713         }
714         if(topright_ref == PART_NOT_AVAILABLE
715            && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
716            && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
717             if(!MB_FIELD
718                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
719                 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
720             }
721             if(MB_FIELD
722                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
723                && i >= scan8[0]+8){
724                 // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
725                 SET_DIAG_MV(/2, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
726             }
727         }
728 #undef SET_DIAG_MV
729     }
730
731     if(topright_ref != PART_NOT_AVAILABLE){
732         *C= h->mv_cache[list][ i - 8 + part_width ];
733         return topright_ref;
734     }else{
735         tprintf(s->avctx, "topright MV not available\n");
736
737         *C= h->mv_cache[list][ i - 8 - 1 ];
738         return h->ref_cache[list][ i - 8 - 1 ];
739     }
740 }
741
742 /**
743  * gets the predicted MV.
744  * @param n the block index
745  * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
746  * @param mx the x component of the predicted motion vector
747  * @param my the y component of the predicted motion vector
748  */
749 static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
750     const int index8= scan8[n];
751     const int top_ref=      h->ref_cache[list][ index8 - 8 ];
752     const int left_ref=     h->ref_cache[list][ index8 - 1 ];
753     const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
754     const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
755     const int16_t * C;
756     int diagonal_ref, match_count;
757
758     assert(part_width==1 || part_width==2 || part_width==4);
759
760 /* mv_cache
761   B . . A T T T T
762   U . . L . . , .
763   U . . L . . . .
764   U . . L . . , .
765   . . . L . . . .
766 */
767
768     diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
769     match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
770     tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
771     if(match_count > 1){ //most common
772         *mx= mid_pred(A[0], B[0], C[0]);
773         *my= mid_pred(A[1], B[1], C[1]);
774     }else if(match_count==1){
775         if(left_ref==ref){
776             *mx= A[0];
777             *my= A[1];
778         }else if(top_ref==ref){
779             *mx= B[0];
780             *my= B[1];
781         }else{
782             *mx= C[0];
783             *my= C[1];
784         }
785     }else{
786         if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
787             *mx= A[0];
788             *my= A[1];
789         }else{
790             *mx= mid_pred(A[0], B[0], C[0]);
791             *my= mid_pred(A[1], B[1], C[1]);
792         }
793     }
794
795     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);
796 }
797
798 /**
799  * gets the directionally predicted 16x8 MV.
800  * @param n the block index
801  * @param mx the x component of the predicted motion vector
802  * @param my the y component of the predicted motion vector
803  */
804 static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
805     if(n==0){
806         const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
807         const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
808
809         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);
810
811         if(top_ref == ref){
812             *mx= B[0];
813             *my= B[1];
814             return;
815         }
816     }else{
817         const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
818         const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];
819
820         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);
821
822         if(left_ref == ref){
823             *mx= A[0];
824             *my= A[1];
825             return;
826         }
827     }
828
829     //RARE
830     pred_motion(h, n, 4, list, ref, mx, my);
831 }
832
833 /**
834  * gets the directionally predicted 8x16 MV.
835  * @param n the block index
836  * @param mx the x component of the predicted motion vector
837  * @param my the y component of the predicted motion vector
838  */
839 static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
840     if(n==0){
841         const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
842         const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];
843
844         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);
845
846         if(left_ref == ref){
847             *mx= A[0];
848             *my= A[1];
849             return;
850         }
851     }else{
852         const int16_t * C;
853         int diagonal_ref;
854
855         diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);
856
857         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);
858
859         if(diagonal_ref == ref){
860             *mx= C[0];
861             *my= C[1];
862             return;
863         }
864     }
865
866     //RARE
867     pred_motion(h, n, 2, list, ref, mx, my);
868 }
869
870 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
871     const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
872     const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];
873
874     tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);
875
876     if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
877        || !( top_ref | *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ])
878        || !(left_ref | *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ])){
879
880         *mx = *my = 0;
881         return;
882     }
883
884     pred_motion(h, 0, 4, 0, 0, mx, my);
885
886     return;
887 }
888
889 static int get_scale_factor(H264Context * const h, int poc, int poc1, int i){
890     int poc0 = h->ref_list[0][i].poc;
891     int td = av_clip(poc1 - poc0, -128, 127);
892     if(td == 0 || h->ref_list[0][i].long_ref){
893         return 256;
894     }else{
895         int tb = av_clip(poc - poc0, -128, 127);
896         int tx = (16384 + (FFABS(td) >> 1)) / td;
897         return av_clip((tb*tx + 32) >> 6, -1024, 1023);
898     }
899 }
900
901 static inline void direct_dist_scale_factor(H264Context * const h){
902     MpegEncContext * const s = &h->s;
903     const int poc = h->s.current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
904     const int poc1 = h->ref_list[1][0].poc;
905     int i, field;
906     for(field=0; field<2; field++){
907         const int poc  = h->s.current_picture_ptr->field_poc[field];
908         const int poc1 = h->ref_list[1][0].field_poc[field];
909         for(i=0; i < 2*h->ref_count[0]; i++)
910             h->dist_scale_factor_field[field][i^field] = get_scale_factor(h, poc, poc1, i+16);
911     }
912
913     for(i=0; i<h->ref_count[0]; i++){
914         h->dist_scale_factor[i] = get_scale_factor(h, poc, poc1, i);
915     }
916 }
917
918 static void fill_colmap(H264Context *h, int map[2][16+32], int list, int field, int colfield, int mbafi){
919     MpegEncContext * const s = &h->s;
920     Picture * const ref1 = &h->ref_list[1][0];
921     int j, old_ref, rfield;
922     int start= mbafi ? 16                      : 0;
923     int end  = mbafi ? 16+2*h->ref_count[list] : h->ref_count[list];
924     int interl= mbafi || s->picture_structure != PICT_FRAME;
925
926     /* bogus; fills in for missing frames */
927     memset(map[list], 0, sizeof(map[list]));
928
929     for(rfield=0; rfield<2; rfield++){
930         for(old_ref=0; old_ref<ref1->ref_count[colfield][list]; old_ref++){
931             int poc = ref1->ref_poc[colfield][list][old_ref];
932
933             if     (!interl)
934                 poc |= 3;
935             else if( interl && (poc&3) == 3) //FIXME store all MBAFF references so this isnt needed
936                 poc= (poc&~3) + rfield + 1;
937
938             for(j=start; j<end; j++){
939                 if(4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3) == poc){
940                     int cur_ref= mbafi ? (j-16)^field : j;
941                     map[list][2*old_ref + (rfield^field) + 16] = cur_ref;
942                     if(rfield == field)
943                         map[list][old_ref] = cur_ref;
944                     break;
945                 }
946             }
947         }
948     }
949 }
950
951 static inline void direct_ref_list_init(H264Context * const h){
952     MpegEncContext * const s = &h->s;
953     Picture * const ref1 = &h->ref_list[1][0];
954     Picture * const cur = s->current_picture_ptr;
955     int list, j, field;
956     int sidx= (s->picture_structure&1)^1;
957     int ref1sidx= (ref1->reference&1)^1;
958
959     for(list=0; list<2; list++){
960         cur->ref_count[sidx][list] = h->ref_count[list];
961         for(j=0; j<h->ref_count[list]; j++)
962             cur->ref_poc[sidx][list][j] = 4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3);
963     }
964
965     if(s->picture_structure == PICT_FRAME){
966         memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0]));
967         memcpy(cur->ref_poc  [1], cur->ref_poc  [0], sizeof(cur->ref_poc  [0]));
968     }
969
970     cur->mbaff= FRAME_MBAFF;
971
972     if(cur->pict_type != FF_B_TYPE || h->direct_spatial_mv_pred)
973         return;
974
975     for(list=0; list<2; list++){
976         fill_colmap(h, h->map_col_to_list0, list, sidx, ref1sidx, 0);
977         for(field=0; field<2; field++)
978             fill_colmap(h, h->map_col_to_list0_field[field], list, field, field, 1);
979     }
980 }
981
982 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
983     MpegEncContext * const s = &h->s;
984     int b8_stride = h->b8_stride;
985     int b4_stride = h->b_stride;
986     int mb_xy = h->mb_xy;
987     int mb_type_col[2];
988     const int16_t (*l1mv0)[2], (*l1mv1)[2];
989     const int8_t *l1ref0, *l1ref1;
990     const int is_b8x8 = IS_8X8(*mb_type);
991     unsigned int sub_mb_type;
992     int i8, i4;
993
994 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
995
996     if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL
997         if(!IS_INTERLACED(*mb_type)){                    //     AFR/FR    -> AFL/FL
998             int cur_poc = s->current_picture_ptr->poc;
999             int *col_poc = h->ref_list[1]->field_poc;
1000             int col_parity = FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc);
1001             mb_xy= s->mb_x + ((s->mb_y&~1) + col_parity)*s->mb_stride;
1002             b8_stride = 0;
1003         }else if(!(s->picture_structure & h->ref_list[1][0].reference) && !h->ref_list[1][0].mbaff){// FL -> FL & differ parity
1004             int fieldoff= 2*(h->ref_list[1][0].reference)-3;
1005             mb_xy += s->mb_stride*fieldoff;
1006         }
1007         goto single_col;
1008     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
1009         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
1010             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
1011             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
1012             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
1013             b8_stride *= 3;
1014             b4_stride *= 6;
1015             //FIXME IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag
1016             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
1017                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
1018                 && !is_b8x8){
1019                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1020                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
1021             }else{
1022                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1023                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1024             }
1025         }else{                                           //     AFR/FR    -> AFR/FR
1026 single_col:
1027             mb_type_col[0] =
1028             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
1029             if(IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag){
1030                 /* FIXME save sub mb types from previous frames (or derive from MVs)
1031                 * so we know exactly what block size to use */
1032                 sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
1033                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1034             }else if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
1035                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1036                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
1037             }else{
1038                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1039                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1040             }
1041         }
1042     }
1043
1044     l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
1045     l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
1046     l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
1047     l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
1048     if(!b8_stride){
1049         if(s->mb_y&1){
1050             l1ref0 += h->b8_stride;
1051             l1ref1 += h->b8_stride;
1052             l1mv0  +=  2*b4_stride;
1053             l1mv1  +=  2*b4_stride;
1054         }
1055     }
1056
1057     if(h->direct_spatial_mv_pred){
1058         int ref[2];
1059         int mv[2][2];
1060         int list;
1061
1062         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
1063
1064         /* ref = min(neighbors) */
1065         for(list=0; list<2; list++){
1066             int refa = h->ref_cache[list][scan8[0] - 1];
1067             int refb = h->ref_cache[list][scan8[0] - 8];
1068             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1069             if(refc == PART_NOT_AVAILABLE)
1070                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1071             ref[list] = FFMIN3((unsigned)refa, (unsigned)refb, (unsigned)refc);
1072             if(ref[list] < 0)
1073                 ref[list] = -1;
1074         }
1075
1076         if(ref[0] < 0 && ref[1] < 0){
1077             ref[0] = ref[1] = 0;
1078             mv[0][0] = mv[0][1] =
1079             mv[1][0] = mv[1][1] = 0;
1080         }else{
1081             for(list=0; list<2; list++){
1082                 if(ref[list] >= 0)
1083                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1084                 else
1085                     mv[list][0] = mv[list][1] = 0;
1086             }
1087         }
1088
1089         if(ref[1] < 0){
1090             if(!is_b8x8)
1091                 *mb_type &= ~MB_TYPE_L1;
1092             sub_mb_type &= ~MB_TYPE_L1;
1093         }else if(ref[0] < 0){
1094             if(!is_b8x8)
1095                 *mb_type &= ~MB_TYPE_L0;
1096             sub_mb_type &= ~MB_TYPE_L0;
1097         }
1098
1099         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1100             for(i8=0; i8<4; i8++){
1101                 int x8 = i8&1;
1102                 int y8 = i8>>1;
1103                 int xy8 = x8+y8*b8_stride;
1104                 int xy4 = 3*x8+y8*b4_stride;
1105                 int a=0, b=0;
1106
1107                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1108                     continue;
1109                 h->sub_mb_type[i8] = sub_mb_type;
1110
1111                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1112                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1113                 if(!IS_INTRA(mb_type_col[y8])
1114                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
1115                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
1116                     if(ref[0] > 0)
1117                         a= pack16to32(mv[0][0],mv[0][1]);
1118                     if(ref[1] > 0)
1119                         b= pack16to32(mv[1][0],mv[1][1]);
1120                 }else{
1121                     a= pack16to32(mv[0][0],mv[0][1]);
1122                     b= pack16to32(mv[1][0],mv[1][1]);
1123                 }
1124                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
1125                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
1126             }
1127         }else if(IS_16X16(*mb_type)){
1128             int a=0, b=0;
1129
1130             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1131             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1132             if(!IS_INTRA(mb_type_col[0])
1133                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1134                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1135                        && (h->x264_build>33 || !h->x264_build)))){
1136                 if(ref[0] > 0)
1137                     a= pack16to32(mv[0][0],mv[0][1]);
1138                 if(ref[1] > 0)
1139                     b= pack16to32(mv[1][0],mv[1][1]);
1140             }else{
1141                 a= pack16to32(mv[0][0],mv[0][1]);
1142                 b= pack16to32(mv[1][0],mv[1][1]);
1143             }
1144             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1145             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1146         }else{
1147             for(i8=0; i8<4; i8++){
1148                 const int x8 = i8&1;
1149                 const int y8 = i8>>1;
1150
1151                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1152                     continue;
1153                 h->sub_mb_type[i8] = sub_mb_type;
1154
1155                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1156                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1157                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1158                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1159
1160                 /* col_zero_flag */
1161                 if(!IS_INTRA(mb_type_col[0]) && (   l1ref0[x8 + y8*b8_stride] == 0
1162                                               || (l1ref0[x8 + y8*b8_stride] < 0 && l1ref1[x8 + y8*b8_stride] == 0
1163                                                   && (h->x264_build>33 || !h->x264_build)))){
1164                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*b8_stride] == 0 ? l1mv0 : l1mv1;
1165                     if(IS_SUB_8X8(sub_mb_type)){
1166                         const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1167                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1168                             if(ref[0] == 0)
1169                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1170                             if(ref[1] == 0)
1171                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1172                         }
1173                     }else
1174                     for(i4=0; i4<4; i4++){
1175                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1176                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1177                             if(ref[0] == 0)
1178                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1179                             if(ref[1] == 0)
1180                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1181                         }
1182                     }
1183                 }
1184             }
1185         }
1186     }else{ /* direct temporal mv pred */
1187         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1188         const int *dist_scale_factor = h->dist_scale_factor;
1189         int ref_offset= 0;
1190
1191         if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
1192             map_col_to_list0[0] = h->map_col_to_list0_field[s->mb_y&1][0];
1193             map_col_to_list0[1] = h->map_col_to_list0_field[s->mb_y&1][1];
1194             dist_scale_factor   =h->dist_scale_factor_field[s->mb_y&1];
1195         }
1196         if(h->ref_list[1][0].mbaff && IS_INTERLACED(mb_type_col[0]))
1197             ref_offset += 16;
1198
1199         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1200             /* FIXME assumes direct_8x8_inference == 1 */
1201             int y_shift  = 2*!IS_INTERLACED(*mb_type);
1202
1203             for(i8=0; i8<4; i8++){
1204                 const int x8 = i8&1;
1205                 const int y8 = i8>>1;
1206                 int ref0, scale;
1207                 const int16_t (*l1mv)[2]= l1mv0;
1208
1209                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1210                     continue;
1211                 h->sub_mb_type[i8] = sub_mb_type;
1212
1213                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1214                 if(IS_INTRA(mb_type_col[y8])){
1215                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1216                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1217                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1218                     continue;
1219                 }
1220
1221                 ref0 = l1ref0[x8 + y8*b8_stride];
1222                 if(ref0 >= 0)
1223                     ref0 = map_col_to_list0[0][ref0 + ref_offset];
1224                 else{
1225                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
1226                     l1mv= l1mv1;
1227                 }
1228                 scale = dist_scale_factor[ref0];
1229                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1230
1231                 {
1232                     const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride];
1233                     int my_col = (mv_col[1]<<y_shift)/2;
1234                     int mx = (scale * mv_col[0] + 128) >> 8;
1235                     int my = (scale * my_col + 128) >> 8;
1236                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1237                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1238                 }
1239             }
1240             return;
1241         }
1242
1243         /* one-to-one mv scaling */
1244
1245         if(IS_16X16(*mb_type)){
1246             int ref, mv0, mv1;
1247
1248             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1249             if(IS_INTRA(mb_type_col[0])){
1250                 ref=mv0=mv1=0;
1251             }else{
1252                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset]
1253                                                 : map_col_to_list0[1][l1ref1[0] + ref_offset];
1254                 const int scale = dist_scale_factor[ref0];
1255                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1256                 int mv_l0[2];
1257                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1258                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1259                 ref= ref0;
1260                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1261                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1262             }
1263             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1264             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1265             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1266         }else{
1267             for(i8=0; i8<4; i8++){
1268                 const int x8 = i8&1;
1269                 const int y8 = i8>>1;
1270                 int ref0, scale;
1271                 const int16_t (*l1mv)[2]= l1mv0;
1272
1273                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1274                     continue;
1275                 h->sub_mb_type[i8] = sub_mb_type;
1276                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1277                 if(IS_INTRA(mb_type_col[0])){
1278                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1279                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1280                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1281                     continue;
1282                 }
1283
1284                 ref0 = l1ref0[x8 + y8*b8_stride] + ref_offset;
1285                 if(ref0 >= 0)
1286                     ref0 = map_col_to_list0[0][ref0];
1287                 else{
1288                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
1289                     l1mv= l1mv1;
1290                 }
1291                 scale = dist_scale_factor[ref0];
1292
1293                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1294                 if(IS_SUB_8X8(sub_mb_type)){
1295                     const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1296                     int mx = (scale * mv_col[0] + 128) >> 8;
1297                     int my = (scale * mv_col[1] + 128) >> 8;
1298                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1299                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1300                 }else
1301                 for(i4=0; i4<4; i4++){
1302                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1303                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1304                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1305                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1306                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1307                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1308                 }
1309             }
1310         }
1311     }
1312 }
1313
1314 static inline void write_back_motion(H264Context *h, int mb_type){
1315     MpegEncContext * const s = &h->s;
1316     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1317     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1318     int list;
1319
1320     if(!USES_LIST(mb_type, 0))
1321         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1322
1323     for(list=0; list<h->list_count; list++){
1324         int y;
1325         if(!USES_LIST(mb_type, list))
1326             continue;
1327
1328         for(y=0; y<4; y++){
1329             *(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];
1330             *(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];
1331         }
1332         if( h->pps.cabac ) {
1333             if(IS_SKIP(mb_type))
1334                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1335             else
1336             for(y=0; y<4; y++){
1337                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1338                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1339             }
1340         }
1341
1342         {
1343             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1344             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1345             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1346             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1347             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1348         }
1349     }
1350
1351     if(h->slice_type_nos == FF_B_TYPE && h->pps.cabac){
1352         if(IS_8X8(mb_type)){
1353             uint8_t *direct_table = &h->direct_table[b8_xy];
1354             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1355             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1356             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1357         }
1358     }
1359 }
1360
1361 /**
1362  * Decodes a network abstraction layer unit.
1363  * @param consumed is the number of bytes used as input
1364  * @param length is the length of the array
1365  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1366  * @returns decoded bytes, might be src+1 if no escapes
1367  */
1368 static const uint8_t *decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
1369     int i, si, di;
1370     uint8_t *dst;
1371     int bufidx;
1372
1373 //    src[0]&0x80;                //forbidden bit
1374     h->nal_ref_idc= src[0]>>5;
1375     h->nal_unit_type= src[0]&0x1F;
1376
1377     src++; length--;
1378 #if 0
1379     for(i=0; i<length; i++)
1380         printf("%2X ", src[i]);
1381 #endif
1382
1383 #ifdef HAVE_FAST_UNALIGNED
1384 # ifdef HAVE_FAST_64BIT
1385 #   define RS 7
1386     for(i=0; i+1<length; i+=9){
1387         if(!((~*(uint64_t*)(src+i) & (*(uint64_t*)(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
1388 # else
1389 #   define RS 3
1390     for(i=0; i+1<length; i+=5){
1391         if(!((~*(uint32_t*)(src+i) & (*(uint32_t*)(src+i) - 0x01000101U)) & 0x80008080U))
1392 # endif
1393             continue;
1394         if(i>0 && !src[i]) i--;
1395         while(src[i]) i++;
1396 #else
1397 #   define RS 0
1398     for(i=0; i+1<length; i+=2){
1399         if(src[i]) continue;
1400         if(i>0 && src[i-1]==0) i--;
1401 #endif
1402         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1403             if(src[i+2]!=3){
1404                 /* startcode, so we must be past the end */
1405                 length=i;
1406             }
1407             break;
1408         }
1409         i-= RS;
1410     }
1411
1412     if(i>=length-1){ //no escaped 0
1413         *dst_length= length;
1414         *consumed= length+1; //+1 for the header
1415         return src;
1416     }
1417
1418     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1419     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
1420     dst= h->rbsp_buffer[bufidx];
1421
1422     if (dst == NULL){
1423         return NULL;
1424     }
1425
1426 //printf("decoding esc\n");
1427     memcpy(dst, src, i);
1428     si=di=i;
1429     while(si+2<length){
1430         //remove escapes (very rare 1:2^22)
1431         if(src[si+2]>3){
1432             dst[di++]= src[si++];
1433             dst[di++]= src[si++];
1434         }else if(src[si]==0 && src[si+1]==0){
1435             if(src[si+2]==3){ //escape
1436                 dst[di++]= 0;
1437                 dst[di++]= 0;
1438                 si+=3;
1439                 continue;
1440             }else //next start code
1441                 goto nsc;
1442         }
1443
1444         dst[di++]= src[si++];
1445     }
1446     while(si<length)
1447         dst[di++]= src[si++];
1448 nsc:
1449
1450     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
1451
1452     *dst_length= di;
1453     *consumed= si + 1;//+1 for the header
1454 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1455     return dst;
1456 }
1457
1458 /**
1459  * identifies the exact end of the bitstream
1460  * @return the length of the trailing, or 0 if damaged
1461  */
1462 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src){
1463     int v= *src;
1464     int r;
1465
1466     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1467
1468     for(r=1; r<9; r++){
1469         if(v&1) return r;
1470         v>>=1;
1471     }
1472     return 0;
1473 }
1474
1475 /**
1476  * IDCT transforms the 16 dc values and dequantizes them.
1477  * @param qp quantization parameter
1478  */
1479 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1480 #define stride 16
1481     int i;
1482     int temp[16]; //FIXME check if this is a good idea
1483     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1484     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1485
1486 //memset(block, 64, 2*256);
1487 //return;
1488     for(i=0; i<4; i++){
1489         const int offset= y_offset[i];
1490         const int z0= block[offset+stride*0] + block[offset+stride*4];
1491         const int z1= block[offset+stride*0] - block[offset+stride*4];
1492         const int z2= block[offset+stride*1] - block[offset+stride*5];
1493         const int z3= block[offset+stride*1] + block[offset+stride*5];
1494
1495         temp[4*i+0]= z0+z3;
1496         temp[4*i+1]= z1+z2;
1497         temp[4*i+2]= z1-z2;
1498         temp[4*i+3]= z0-z3;
1499     }
1500
1501     for(i=0; i<4; i++){
1502         const int offset= x_offset[i];
1503         const int z0= temp[4*0+i] + temp[4*2+i];
1504         const int z1= temp[4*0+i] - temp[4*2+i];
1505         const int z2= temp[4*1+i] - temp[4*3+i];
1506         const int z3= temp[4*1+i] + temp[4*3+i];
1507
1508         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
1509         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1510         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1511         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1512     }
1513 }
1514
1515 #if 0
1516 /**
1517  * DCT transforms the 16 dc values.
1518  * @param qp quantization parameter ??? FIXME
1519  */
1520 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1521 //    const int qmul= dequant_coeff[qp][0];
1522     int i;
1523     int temp[16]; //FIXME check if this is a good idea
1524     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1525     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1526
1527     for(i=0; i<4; i++){
1528         const int offset= y_offset[i];
1529         const int z0= block[offset+stride*0] + block[offset+stride*4];
1530         const int z1= block[offset+stride*0] - block[offset+stride*4];
1531         const int z2= block[offset+stride*1] - block[offset+stride*5];
1532         const int z3= block[offset+stride*1] + block[offset+stride*5];
1533
1534         temp[4*i+0]= z0+z3;
1535         temp[4*i+1]= z1+z2;
1536         temp[4*i+2]= z1-z2;
1537         temp[4*i+3]= z0-z3;
1538     }
1539
1540     for(i=0; i<4; i++){
1541         const int offset= x_offset[i];
1542         const int z0= temp[4*0+i] + temp[4*2+i];
1543         const int z1= temp[4*0+i] - temp[4*2+i];
1544         const int z2= temp[4*1+i] - temp[4*3+i];
1545         const int z3= temp[4*1+i] + temp[4*3+i];
1546
1547         block[stride*0 +offset]= (z0 + z3)>>1;
1548         block[stride*2 +offset]= (z1 + z2)>>1;
1549         block[stride*8 +offset]= (z1 - z2)>>1;
1550         block[stride*10+offset]= (z0 - z3)>>1;
1551     }
1552 }
1553 #endif
1554
1555 #undef xStride
1556 #undef stride
1557
1558 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1559     const int stride= 16*2;
1560     const int xStride= 16;
1561     int a,b,c,d,e;
1562
1563     a= block[stride*0 + xStride*0];
1564     b= block[stride*0 + xStride*1];
1565     c= block[stride*1 + xStride*0];
1566     d= block[stride*1 + xStride*1];
1567
1568     e= a-b;
1569     a= a+b;
1570     b= c-d;
1571     c= c+d;
1572
1573     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1574     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1575     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1576     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1577 }
1578
1579 #if 0
1580 static void chroma_dc_dct_c(DCTELEM *block){
1581     const int stride= 16*2;
1582     const int xStride= 16;
1583     int a,b,c,d,e;
1584
1585     a= block[stride*0 + xStride*0];
1586     b= block[stride*0 + xStride*1];
1587     c= block[stride*1 + xStride*0];
1588     d= block[stride*1 + xStride*1];
1589
1590     e= a-b;
1591     a= a+b;
1592     b= c-d;
1593     c= c+d;
1594
1595     block[stride*0 + xStride*0]= (a+c);
1596     block[stride*0 + xStride*1]= (e+b);
1597     block[stride*1 + xStride*0]= (a-c);
1598     block[stride*1 + xStride*1]= (e-b);
1599 }
1600 #endif
1601
1602 /**
1603  * gets the chroma qp.
1604  */
1605 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1606     return h->pps.chroma_qp_table[t][qscale];
1607 }
1608
1609 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1610                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1611                            int src_x_offset, int src_y_offset,
1612                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1613     MpegEncContext * const s = &h->s;
1614     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1615     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1616     const int luma_xy= (mx&3) + ((my&3)<<2);
1617     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1618     uint8_t * src_cb, * src_cr;
1619     int extra_width= h->emu_edge_width;
1620     int extra_height= h->emu_edge_height;
1621     int emu=0;
1622     const int full_mx= mx>>2;
1623     const int full_my= my>>2;
1624     const int pic_width  = 16*s->mb_width;
1625     const int pic_height = 16*s->mb_height >> MB_FIELD;
1626
1627     if(mx&7) extra_width -= 3;
1628     if(my&7) extra_height -= 3;
1629
1630     if(   full_mx < 0-extra_width
1631        || full_my < 0-extra_height
1632        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1633        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1634         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);
1635             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1636         emu=1;
1637     }
1638
1639     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1640     if(!square){
1641         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1642     }
1643
1644     if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1645
1646     if(MB_FIELD){
1647         // chroma offset when predicting from a field of opposite parity
1648         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1649         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1650     }
1651     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1652     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1653
1654     if(emu){
1655         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);
1656             src_cb= s->edge_emu_buffer;
1657     }
1658     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1659
1660     if(emu){
1661         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);
1662             src_cr= s->edge_emu_buffer;
1663     }
1664     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1665 }
1666
1667 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1668                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1669                            int x_offset, int y_offset,
1670                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1671                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1672                            int list0, int list1){
1673     MpegEncContext * const s = &h->s;
1674     qpel_mc_func *qpix_op=  qpix_put;
1675     h264_chroma_mc_func chroma_op= chroma_put;
1676
1677     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1678     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1679     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1680     x_offset += 8*s->mb_x;
1681     y_offset += 8*(s->mb_y >> MB_FIELD);
1682
1683     if(list0){
1684         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1685         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1686                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1687                            qpix_op, chroma_op);
1688
1689         qpix_op=  qpix_avg;
1690         chroma_op= chroma_avg;
1691     }
1692
1693     if(list1){
1694         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1695         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1696                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1697                            qpix_op, chroma_op);
1698     }
1699 }
1700
1701 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1702                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1703                            int x_offset, int y_offset,
1704                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1705                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1706                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1707                            int list0, int list1){
1708     MpegEncContext * const s = &h->s;
1709
1710     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1711     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1712     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1713     x_offset += 8*s->mb_x;
1714     y_offset += 8*(s->mb_y >> MB_FIELD);
1715
1716     if(list0 && list1){
1717         /* don't optimize for luma-only case, since B-frames usually
1718          * use implicit weights => chroma too. */
1719         uint8_t *tmp_cb = s->obmc_scratchpad;
1720         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1721         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1722         int refn0 = h->ref_cache[0][ scan8[n] ];
1723         int refn1 = h->ref_cache[1][ scan8[n] ];
1724
1725         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1726                     dest_y, dest_cb, dest_cr,
1727                     x_offset, y_offset, qpix_put, chroma_put);
1728         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1729                     tmp_y, tmp_cb, tmp_cr,
1730                     x_offset, y_offset, qpix_put, chroma_put);
1731
1732         if(h->use_weight == 2){
1733             int weight0 = h->implicit_weight[refn0][refn1];
1734             int weight1 = 64 - weight0;
1735             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1736             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1737             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1738         }else{
1739             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1740                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1741                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1742             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1743                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1744                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1745             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1746                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1747                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1748         }
1749     }else{
1750         int list = list1 ? 1 : 0;
1751         int refn = h->ref_cache[list][ scan8[n] ];
1752         Picture *ref= &h->ref_list[list][refn];
1753         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1754                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1755                     qpix_put, chroma_put);
1756
1757         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1758                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1759         if(h->use_weight_chroma){
1760             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1761                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1762             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1763                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1764         }
1765     }
1766 }
1767
1768 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1769                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1770                            int x_offset, int y_offset,
1771                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1772                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1773                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1774                            int list0, int list1){
1775     if((h->use_weight==2 && list0 && list1
1776         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1777        || h->use_weight==1)
1778         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1779                          x_offset, y_offset, qpix_put, chroma_put,
1780                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1781     else
1782         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1783                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1784 }
1785
1786 static inline void prefetch_motion(H264Context *h, int list){
1787     /* fetch pixels for estimated mv 4 macroblocks ahead
1788      * optimized for 64byte cache lines */
1789     MpegEncContext * const s = &h->s;
1790     const int refn = h->ref_cache[list][scan8[0]];
1791     if(refn >= 0){
1792         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1793         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1794         uint8_t **src= h->ref_list[list][refn].data;
1795         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1796         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1797         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1798         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1799     }
1800 }
1801
1802 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1803                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1804                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1805                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1806     MpegEncContext * const s = &h->s;
1807     const int mb_xy= h->mb_xy;
1808     const int mb_type= s->current_picture.mb_type[mb_xy];
1809
1810     assert(IS_INTER(mb_type));
1811
1812     prefetch_motion(h, 0);
1813
1814     if(IS_16X16(mb_type)){
1815         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1816                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1817                 &weight_op[0], &weight_avg[0],
1818                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1819     }else if(IS_16X8(mb_type)){
1820         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1821                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1822                 &weight_op[1], &weight_avg[1],
1823                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1824         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1825                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1826                 &weight_op[1], &weight_avg[1],
1827                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1828     }else if(IS_8X16(mb_type)){
1829         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1830                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1831                 &weight_op[2], &weight_avg[2],
1832                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1833         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1834                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1835                 &weight_op[2], &weight_avg[2],
1836                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1837     }else{
1838         int i;
1839
1840         assert(IS_8X8(mb_type));
1841
1842         for(i=0; i<4; i++){
1843             const int sub_mb_type= h->sub_mb_type[i];
1844             const int n= 4*i;
1845             int x_offset= (i&1)<<2;
1846             int y_offset= (i&2)<<1;
1847
1848             if(IS_SUB_8X8(sub_mb_type)){
1849                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1850                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1851                     &weight_op[3], &weight_avg[3],
1852                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1853             }else if(IS_SUB_8X4(sub_mb_type)){
1854                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1855                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1856                     &weight_op[4], &weight_avg[4],
1857                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1858                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1859                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1860                     &weight_op[4], &weight_avg[4],
1861                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1862             }else if(IS_SUB_4X8(sub_mb_type)){
1863                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1864                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1865                     &weight_op[5], &weight_avg[5],
1866                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1867                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1868                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1869                     &weight_op[5], &weight_avg[5],
1870                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1871             }else{
1872                 int j;
1873                 assert(IS_SUB_4X4(sub_mb_type));
1874                 for(j=0; j<4; j++){
1875                     int sub_x_offset= x_offset + 2*(j&1);
1876                     int sub_y_offset= y_offset +   (j&2);
1877                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1878                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1879                         &weight_op[6], &weight_avg[6],
1880                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1881                 }
1882             }
1883         }
1884     }
1885
1886     prefetch_motion(h, 1);
1887 }
1888
1889 static av_cold void init_cavlc_level_tab(void){
1890     int suffix_length, mask;
1891     unsigned int i;
1892
1893     for(suffix_length=0; suffix_length<7; suffix_length++){
1894         for(i=0; i<(1<<LEVEL_TAB_BITS); i++){
1895             int prefix= LEVEL_TAB_BITS - av_log2(2*i);
1896             int level_code= (prefix<<suffix_length) + (i>>(LEVEL_TAB_BITS-prefix-1-suffix_length)) - (1<<suffix_length);
1897
1898             mask= -(level_code&1);
1899             level_code= (((2+level_code)>>1) ^ mask) - mask;
1900             if(prefix + 1 + suffix_length <= LEVEL_TAB_BITS){
1901                 cavlc_level_tab[suffix_length][i][0]= level_code;
1902                 cavlc_level_tab[suffix_length][i][1]= prefix + 1 + suffix_length;
1903             }else if(prefix + 1 <= LEVEL_TAB_BITS){
1904                 cavlc_level_tab[suffix_length][i][0]= prefix+100;
1905                 cavlc_level_tab[suffix_length][i][1]= prefix + 1;
1906             }else{
1907                 cavlc_level_tab[suffix_length][i][0]= LEVEL_TAB_BITS+100;
1908                 cavlc_level_tab[suffix_length][i][1]= LEVEL_TAB_BITS;
1909             }
1910         }
1911     }
1912 }
1913
1914 static av_cold void decode_init_vlc(void){
1915     static int done = 0;
1916
1917     if (!done) {
1918         int i;
1919         int offset;
1920         done = 1;
1921
1922         chroma_dc_coeff_token_vlc.table = chroma_dc_coeff_token_vlc_table;
1923         chroma_dc_coeff_token_vlc.table_allocated = chroma_dc_coeff_token_vlc_table_size;
1924         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1925                  &chroma_dc_coeff_token_len [0], 1, 1,
1926                  &chroma_dc_coeff_token_bits[0], 1, 1,
1927                  INIT_VLC_USE_NEW_STATIC);
1928
1929         offset = 0;
1930         for(i=0; i<4; i++){
1931             coeff_token_vlc[i].table = coeff_token_vlc_tables+offset;
1932             coeff_token_vlc[i].table_allocated = coeff_token_vlc_tables_size[i];
1933             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1934                      &coeff_token_len [i][0], 1, 1,
1935                      &coeff_token_bits[i][0], 1, 1,
1936                      INIT_VLC_USE_NEW_STATIC);
1937             offset += coeff_token_vlc_tables_size[i];
1938         }
1939         /*
1940          * This is a one time safety check to make sure that
1941          * the packed static coeff_token_vlc table sizes
1942          * were initialized correctly.
1943          */
1944         assert(offset == FF_ARRAY_ELEMS(coeff_token_vlc_tables));
1945
1946         for(i=0; i<3; i++){
1947             chroma_dc_total_zeros_vlc[i].table = chroma_dc_total_zeros_vlc_tables[i];
1948             chroma_dc_total_zeros_vlc[i].table_allocated = chroma_dc_total_zeros_vlc_tables_size;
1949             init_vlc(&chroma_dc_total_zeros_vlc[i],
1950                      CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1951                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1952                      &chroma_dc_total_zeros_bits[i][0], 1, 1,
1953                      INIT_VLC_USE_NEW_STATIC);
1954         }
1955         for(i=0; i<15; i++){
1956             total_zeros_vlc[i].table = total_zeros_vlc_tables[i];
1957             total_zeros_vlc[i].table_allocated = total_zeros_vlc_tables_size;
1958             init_vlc(&total_zeros_vlc[i],
1959                      TOTAL_ZEROS_VLC_BITS, 16,
1960                      &total_zeros_len [i][0], 1, 1,
1961                      &total_zeros_bits[i][0], 1, 1,
1962                      INIT_VLC_USE_NEW_STATIC);
1963         }
1964
1965         for(i=0; i<6; i++){
1966             run_vlc[i].table = run_vlc_tables[i];
1967             run_vlc[i].table_allocated = run_vlc_tables_size;
1968             init_vlc(&run_vlc[i],
1969                      RUN_VLC_BITS, 7,
1970                      &run_len [i][0], 1, 1,
1971                      &run_bits[i][0], 1, 1,
1972                      INIT_VLC_USE_NEW_STATIC);
1973         }
1974         run7_vlc.table = run7_vlc_table,
1975         run7_vlc.table_allocated = run7_vlc_table_size;
1976         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
1977                  &run_len [6][0], 1, 1,
1978                  &run_bits[6][0], 1, 1,
1979                  INIT_VLC_USE_NEW_STATIC);
1980
1981         init_cavlc_level_tab();
1982     }
1983 }
1984
1985 static void free_tables(H264Context *h){
1986     int i;
1987     H264Context *hx;
1988     av_freep(&h->intra4x4_pred_mode);
1989     av_freep(&h->chroma_pred_mode_table);
1990     av_freep(&h->cbp_table);
1991     av_freep(&h->mvd_table[0]);
1992     av_freep(&h->mvd_table[1]);
1993     av_freep(&h->direct_table);
1994     av_freep(&h->non_zero_count);
1995     av_freep(&h->slice_table_base);
1996     h->slice_table= NULL;
1997
1998     av_freep(&h->mb2b_xy);
1999     av_freep(&h->mb2b8_xy);
2000
2001     for(i = 0; i < h->s.avctx->thread_count; i++) {
2002         hx = h->thread_context[i];
2003         if(!hx) continue;
2004         av_freep(&hx->top_borders[1]);
2005         av_freep(&hx->top_borders[0]);
2006         av_freep(&hx->s.obmc_scratchpad);
2007     }
2008 }
2009
2010 static void init_dequant8_coeff_table(H264Context *h){
2011     int i,q,x;
2012     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2013     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2014     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2015
2016     for(i=0; i<2; i++ ){
2017         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2018             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2019             break;
2020         }
2021
2022         for(q=0; q<52; q++){
2023             int shift = div6[q];
2024             int idx = rem6[q];
2025             for(x=0; x<64; x++)
2026                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2027                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2028                     h->pps.scaling_matrix8[i][x]) << shift;
2029         }
2030     }
2031 }
2032
2033 static void init_dequant4_coeff_table(H264Context *h){
2034     int i,j,q,x;
2035     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2036     for(i=0; i<6; i++ ){
2037         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2038         for(j=0; j<i; j++){
2039             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2040                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2041                 break;
2042             }
2043         }
2044         if(j<i)
2045             continue;
2046
2047         for(q=0; q<52; q++){
2048             int shift = div6[q] + 2;
2049             int idx = rem6[q];
2050             for(x=0; x<16; x++)
2051                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2052                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2053                     h->pps.scaling_matrix4[i][x]) << shift;
2054         }
2055     }
2056 }
2057
2058 static void init_dequant_tables(H264Context *h){
2059     int i,x;
2060     init_dequant4_coeff_table(h);
2061     if(h->pps.transform_8x8_mode)
2062         init_dequant8_coeff_table(h);
2063     if(h->sps.transform_bypass){
2064         for(i=0; i<6; i++)
2065             for(x=0; x<16; x++)
2066                 h->dequant4_coeff[i][0][x] = 1<<6;
2067         if(h->pps.transform_8x8_mode)
2068             for(i=0; i<2; i++)
2069                 for(x=0; x<64; x++)
2070                     h->dequant8_coeff[i][0][x] = 1<<6;
2071     }
2072 }
2073
2074
2075 /**
2076  * allocates tables.
2077  * needs width/height
2078  */
2079 static int alloc_tables(H264Context *h){
2080     MpegEncContext * const s = &h->s;
2081     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2082     int x,y;
2083
2084     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2085
2086     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2087     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base))
2088     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2089
2090     CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2091     CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2092     CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2093     CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2094
2095     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
2096     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2097
2098     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2099     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2100     for(y=0; y<s->mb_height; y++){
2101         for(x=0; x<s->mb_width; x++){
2102             const int mb_xy= x + y*s->mb_stride;
2103             const int b_xy = 4*x + 4*y*h->b_stride;
2104             const int b8_xy= 2*x + 2*y*h->b8_stride;
2105
2106             h->mb2b_xy [mb_xy]= b_xy;
2107             h->mb2b8_xy[mb_xy]= b8_xy;
2108         }
2109     }
2110
2111     s->obmc_scratchpad = NULL;
2112
2113     if(!h->dequant4_coeff[0])
2114         init_dequant_tables(h);
2115
2116     return 0;
2117 fail:
2118     free_tables(h);
2119     return -1;
2120 }
2121
2122 /**
2123  * Mimic alloc_tables(), but for every context thread.
2124  */
2125 static void clone_tables(H264Context *dst, H264Context *src){
2126     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2127     dst->non_zero_count           = src->non_zero_count;
2128     dst->slice_table              = src->slice_table;
2129     dst->cbp_table                = src->cbp_table;
2130     dst->mb2b_xy                  = src->mb2b_xy;
2131     dst->mb2b8_xy                 = src->mb2b8_xy;
2132     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2133     dst->mvd_table[0]             = src->mvd_table[0];
2134     dst->mvd_table[1]             = src->mvd_table[1];
2135     dst->direct_table             = src->direct_table;
2136
2137     dst->s.obmc_scratchpad = NULL;
2138     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2139 }
2140
2141 /**
2142  * Init context
2143  * Allocate buffers which are not shared amongst multiple threads.
2144  */
2145 static int context_init(H264Context *h){
2146     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2147     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2148
2149     return 0;
2150 fail:
2151     return -1; // free_tables will clean up for us
2152 }
2153
2154 static av_cold void common_init(H264Context *h){
2155     MpegEncContext * const s = &h->s;
2156
2157     s->width = s->avctx->width;
2158     s->height = s->avctx->height;
2159     s->codec_id= s->avctx->codec->id;
2160
2161     ff_h264_pred_init(&h->hpc, s->codec_id);
2162
2163     h->dequant_coeff_pps= -1;
2164     s->unrestricted_mv=1;
2165     s->decode=1; //FIXME
2166
2167     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
2168
2169     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2170     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2171 }
2172
2173 static av_cold int decode_init(AVCodecContext *avctx){
2174     H264Context *h= avctx->priv_data;
2175     MpegEncContext * const s = &h->s;
2176
2177     MPV_decode_defaults(s);
2178
2179     s->avctx = avctx;
2180     common_init(h);
2181
2182     s->out_format = FMT_H264;
2183     s->workaround_bugs= avctx->workaround_bugs;
2184
2185     // set defaults
2186 //    s->decode_mb= ff_h263_decode_mb;
2187     s->quarter_sample = 1;
2188     s->low_delay= 1;
2189
2190     if(avctx->codec_id == CODEC_ID_SVQ3)
2191         avctx->pix_fmt= PIX_FMT_YUVJ420P;
2192     else if(avctx->codec_id == CODEC_ID_H264_VDPAU)
2193         avctx->pix_fmt= PIX_FMT_VDPAU_H264;
2194     else
2195         avctx->pix_fmt= PIX_FMT_YUV420P;
2196
2197     decode_init_vlc();
2198
2199     if(avctx->extradata_size > 0 && avctx->extradata &&
2200        *(char *)avctx->extradata == 1){
2201         h->is_avc = 1;
2202         h->got_avcC = 0;
2203     } else {
2204         h->is_avc = 0;
2205     }
2206
2207     h->thread_context[0] = h;
2208     h->outputed_poc = INT_MIN;
2209     h->prev_poc_msb= 1<<16;
2210     return 0;
2211 }
2212
2213 static int frame_start(H264Context *h){
2214     MpegEncContext * const s = &h->s;
2215     int i;
2216
2217     if(MPV_frame_start(s, s->avctx) < 0)
2218         return -1;
2219     ff_er_frame_start(s);
2220     /*
2221      * MPV_frame_start uses pict_type to derive key_frame.
2222      * This is incorrect for H.264; IDR markings must be used.
2223      * Zero here; IDR markings per slice in frame or fields are ORed in later.
2224      * See decode_nal_units().
2225      */
2226     s->current_picture_ptr->key_frame= 0;
2227
2228     assert(s->linesize && s->uvlinesize);
2229
2230     for(i=0; i<16; i++){
2231         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2232         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2233     }
2234     for(i=0; i<4; i++){
2235         h->block_offset[16+i]=
2236         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2237         h->block_offset[24+16+i]=
2238         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2239     }
2240
2241     /* can't be in alloc_tables because linesize isn't known there.
2242      * FIXME: redo bipred weight to not require extra buffer? */
2243     for(i = 0; i < s->avctx->thread_count; i++)
2244         if(!h->thread_context[i]->s.obmc_scratchpad)
2245             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2246
2247     /* some macroblocks will be accessed before they're available */
2248     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2249         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
2250
2251 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2252
2253     // We mark the current picture as non-reference after allocating it, so
2254     // that if we break out due to an error it can be released automatically
2255     // in the next MPV_frame_start().
2256     // SVQ3 as well as most other codecs have only last/next/current and thus
2257     // get released even with set reference, besides SVQ3 and others do not
2258     // mark frames as reference later "naturally".
2259     if(s->codec_id != CODEC_ID_SVQ3)
2260         s->current_picture_ptr->reference= 0;
2261
2262     s->current_picture_ptr->field_poc[0]=
2263     s->current_picture_ptr->field_poc[1]= INT_MAX;
2264     assert(s->current_picture_ptr->long_ref==0);
2265
2266     return 0;
2267 }
2268
2269 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){
2270     MpegEncContext * const s = &h->s;
2271     int i;
2272     int step    = 1;
2273     int offset  = 1;
2274     int uvoffset= 1;
2275     int top_idx = 1;
2276     int skiplast= 0;
2277
2278     src_y  -=   linesize;
2279     src_cb -= uvlinesize;
2280     src_cr -= uvlinesize;
2281
2282     if(!simple && FRAME_MBAFF){
2283         if(s->mb_y&1){
2284             offset  = MB_MBAFF ? 1 : 17;
2285             uvoffset= MB_MBAFF ? 1 : 9;
2286             if(!MB_MBAFF){
2287                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 0)= *(uint64_t*)(src_y +  15*linesize);
2288                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 8)= *(uint64_t*)(src_y +8+15*linesize);
2289                 if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2290                     *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+7*uvlinesize);
2291                     *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+7*uvlinesize);
2292                 }
2293             }
2294         }else{
2295             if(!MB_MBAFF){
2296                 h->left_border[0]= h->top_borders[0][s->mb_x][15];
2297                 if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2298                     h->left_border[34   ]= h->top_borders[0][s->mb_x][16+7  ];
2299                     h->left_border[34+18]= h->top_borders[0][s->mb_x][16+8+7];
2300                 }
2301                 skiplast= 1;
2302             }
2303             offset  =
2304             uvoffset=
2305             top_idx = MB_MBAFF ? 0 : 1;
2306         }
2307         step= MB_MBAFF ? 2 : 1;
2308     }
2309
2310     // There are two lines saved, the line above the the top macroblock of a pair,
2311     // and the line above the bottom macroblock
2312     h->left_border[offset]= h->top_borders[top_idx][s->mb_x][15];
2313     for(i=1; i<17 - skiplast; i++){
2314         h->left_border[offset+i*step]= src_y[15+i*  linesize];
2315     }
2316
2317     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2318     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2319
2320     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2321         h->left_border[uvoffset+34   ]= h->top_borders[top_idx][s->mb_x][16+7];
2322         h->left_border[uvoffset+34+18]= h->top_borders[top_idx][s->mb_x][24+7];
2323         for(i=1; i<9 - skiplast; i++){
2324             h->left_border[uvoffset+34   +i*step]= src_cb[7+i*uvlinesize];
2325             h->left_border[uvoffset+34+18+i*step]= src_cr[7+i*uvlinesize];
2326         }
2327         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2328         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2329     }
2330 }
2331
2332 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){
2333     MpegEncContext * const s = &h->s;
2334     int temp8, i;
2335     uint64_t temp64;
2336     int deblock_left;
2337     int deblock_top;
2338     int mb_xy;
2339     int step    = 1;
2340     int offset  = 1;
2341     int uvoffset= 1;
2342     int top_idx = 1;
2343
2344     if(!simple && FRAME_MBAFF){
2345         if(s->mb_y&1){
2346             offset  = MB_MBAFF ? 1 : 17;
2347             uvoffset= MB_MBAFF ? 1 : 9;
2348         }else{
2349             offset  =
2350             uvoffset=
2351             top_idx = MB_MBAFF ? 0 : 1;
2352         }
2353         step= MB_MBAFF ? 2 : 1;
2354     }
2355
2356     if(h->deblocking_filter == 2) {
2357         mb_xy = h->mb_xy;
2358         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2359         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2360     } else {
2361         deblock_left = (s->mb_x > 0);
2362         deblock_top =  (s->mb_y > !!MB_FIELD);
2363     }
2364
2365     src_y  -=   linesize + 1;
2366     src_cb -= uvlinesize + 1;
2367     src_cr -= uvlinesize + 1;
2368
2369 #define XCHG(a,b,t,xchg)\
2370 t= a;\
2371 if(xchg)\
2372     a= b;\
2373 b= t;
2374
2375     if(deblock_left){
2376         for(i = !deblock_top; i<16; i++){
2377             XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, xchg);
2378         }
2379         XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, 1);
2380     }
2381
2382     if(deblock_top){
2383         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2384         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2385         if(s->mb_x+1 < s->mb_width){
2386             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2387         }
2388     }
2389
2390     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2391         if(deblock_left){
2392             for(i = !deblock_top; i<8; i++){
2393                 XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, xchg);
2394                 XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, xchg);
2395             }
2396             XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, 1);
2397             XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, 1);
2398         }
2399         if(deblock_top){
2400             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2401             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2402         }
2403     }
2404 }
2405
2406 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2407     MpegEncContext * const s = &h->s;
2408     const int mb_x= s->mb_x;
2409     const int mb_y= s->mb_y;
2410     const int mb_xy= h->mb_xy;
2411     const int mb_type= s->current_picture.mb_type[mb_xy];
2412     uint8_t  *dest_y, *dest_cb, *dest_cr;
2413     int linesize, uvlinesize /*dct_offset*/;
2414     int i;
2415     int *block_offset = &h->block_offset[0];
2416     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2417     const int is_h264 = simple || s->codec_id == CODEC_ID_H264;
2418     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2419     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2420
2421     dest_y  = s->current_picture.data[0] + (mb_x + mb_y * s->linesize  ) * 16;
2422     dest_cb = s->current_picture.data[1] + (mb_x + mb_y * s->uvlinesize) * 8;
2423     dest_cr = s->current_picture.data[2] + (mb_x + mb_y * s->uvlinesize) * 8;
2424
2425     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2426     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2427
2428     if (!simple && MB_FIELD) {
2429         linesize   = h->mb_linesize   = s->linesize * 2;
2430         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2431         block_offset = &h->block_offset[24];
2432         if(mb_y&1){ //FIXME move out of this function?
2433             dest_y -= s->linesize*15;
2434             dest_cb-= s->uvlinesize*7;
2435             dest_cr-= s->uvlinesize*7;
2436         }
2437         if(FRAME_MBAFF) {
2438             int list;
2439             for(list=0; list<h->list_count; list++){
2440                 if(!USES_LIST(mb_type, list))
2441                     continue;
2442                 if(IS_16X16(mb_type)){
2443                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2444                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2445                 }else{
2446                     for(i=0; i<16; i+=4){
2447                         int ref = h->ref_cache[list][scan8[i]];
2448                         if(ref >= 0)
2449                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2450                     }
2451                 }
2452             }
2453         }
2454     } else {
2455         linesize   = h->mb_linesize   = s->linesize;
2456         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2457 //        dct_offset = s->linesize * 16;
2458     }
2459
2460     if (!simple && IS_INTRA_PCM(mb_type)) {
2461         for (i=0; i<16; i++) {
2462             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
2463         }
2464         for (i=0; i<8; i++) {
2465             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
2466             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
2467         }
2468     } else {
2469         if(IS_INTRA(mb_type)){
2470             if(h->deblocking_filter)
2471                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2472
2473             if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2474                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2475                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2476             }
2477
2478             if(IS_INTRA4x4(mb_type)){
2479                 if(simple || !s->encoding){
2480                     if(IS_8x8DCT(mb_type)){
2481                         if(transform_bypass){
2482                             idct_dc_add =
2483                             idct_add    = s->dsp.add_pixels8;
2484                         }else{
2485                             idct_dc_add = s->dsp.h264_idct8_dc_add;
2486                             idct_add    = s->dsp.h264_idct8_add;
2487                         }
2488                         for(i=0; i<16; i+=4){
2489                             uint8_t * const ptr= dest_y + block_offset[i];
2490                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2491                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
2492                                 h->hpc.pred8x8l_add[dir](ptr, h->mb + i*16, linesize);
2493                             }else{
2494                                 const int nnz = h->non_zero_count_cache[ scan8[i] ];
2495                                 h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2496                                                             (h->topright_samples_available<<i)&0x4000, linesize);
2497                                 if(nnz){
2498                                     if(nnz == 1 && h->mb[i*16])
2499                                         idct_dc_add(ptr, h->mb + i*16, linesize);
2500                                     else
2501                                         idct_add   (ptr, h->mb + i*16, linesize);
2502                                 }
2503                             }
2504                         }
2505                     }else{
2506                         if(transform_bypass){
2507                             idct_dc_add =
2508                             idct_add    = s->dsp.add_pixels4;
2509                         }else{
2510                             idct_dc_add = s->dsp.h264_idct_dc_add;
2511                             idct_add    = s->dsp.h264_idct_add;
2512                         }
2513                         for(i=0; i<16; i++){
2514                             uint8_t * const ptr= dest_y + block_offset[i];
2515                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2516
2517                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
2518                                 h->hpc.pred4x4_add[dir](ptr, h->mb + i*16, linesize);
2519                             }else{
2520                                 uint8_t *topright;
2521                                 int nnz, tr;
2522                                 if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2523                                     const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2524                                     assert(mb_y || linesize <= block_offset[i]);
2525                                     if(!topright_avail){
2526                                         tr= ptr[3 - linesize]*0x01010101;
2527                                         topright= (uint8_t*) &tr;
2528                                     }else
2529                                         topright= ptr + 4 - linesize;
2530                                 }else
2531                                     topright= NULL;
2532
2533                                 h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2534                                 nnz = h->non_zero_count_cache[ scan8[i] ];
2535                                 if(nnz){
2536                                     if(is_h264){
2537                                         if(nnz == 1 && h->mb[i*16])
2538                                             idct_dc_add(ptr, h->mb + i*16, linesize);
2539                                         else
2540                                             idct_add   (ptr, h->mb + i*16, linesize);
2541                                     }else
2542                                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2543                                 }
2544                             }
2545                         }
2546                     }
2547                 }
2548             }else{
2549                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2550                 if(is_h264){
2551                     if(!transform_bypass)
2552                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2553                 }else
2554                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2555             }
2556             if(h->deblocking_filter)
2557                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2558         }else if(is_h264){
2559             hl_motion(h, dest_y, dest_cb, dest_cr,
2560                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2561                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2562                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2563         }
2564
2565
2566         if(!IS_INTRA4x4(mb_type)){
2567             if(is_h264){
2568                 if(IS_INTRA16x16(mb_type)){
2569                     if(transform_bypass){
2570                         if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
2571                             h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb, linesize);
2572                         }else{
2573                             for(i=0; i<16; i++){
2574                                 if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
2575                                     s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + i*16, linesize);
2576                             }
2577                         }
2578                     }else{
2579                          s->dsp.h264_idct_add16intra(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2580                     }
2581                 }else if(h->cbp&15){
2582                     if(transform_bypass){
2583                         const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2584                         idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2585                         for(i=0; i<16; i+=di){
2586                             if(h->non_zero_count_cache[ scan8[i] ]){
2587                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2588                             }
2589                         }
2590                     }else{
2591                         if(IS_8x8DCT(mb_type)){
2592                             s->dsp.h264_idct8_add4(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2593                         }else{
2594                             s->dsp.h264_idct_add16(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
2595                         }
2596                     }
2597                 }
2598             }else{
2599                 for(i=0; i<16; i++){
2600                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2601                         uint8_t * const ptr= dest_y + block_offset[i];
2602                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2603                     }
2604                 }
2605             }
2606         }
2607
2608         if((simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
2609             uint8_t *dest[2] = {dest_cb, dest_cr};
2610             if(transform_bypass){
2611                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
2612                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + 16*16, uvlinesize);
2613                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 20, h->mb + 20*16, uvlinesize);
2614                 }else{
2615                     idct_add = s->dsp.add_pixels4;
2616                     for(i=16; i<16+8; i++){
2617                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
2618                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2619                     }
2620                 }
2621             }else{
2622                 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]);
2623                 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]);
2624                 if(is_h264){
2625                     idct_add = s->dsp.h264_idct_add;
2626                     idct_dc_add = s->dsp.h264_idct_dc_add;
2627                     for(i=16; i<16+8; i++){
2628                         if(h->non_zero_count_cache[ scan8[i] ])
2629                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2630                         else if(h->mb[i*16])
2631                             idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2632                     }
2633                 }else{
2634                     for(i=16; i<16+8; i++){
2635                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2636                             uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2637                             svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2638                         }
2639                     }
2640                 }
2641             }
2642         }
2643     }
2644     if(h->cbp || IS_INTRA(mb_type))
2645         s->dsp.clear_blocks(h->mb);
2646
2647     if(h->deblocking_filter) {
2648         backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2649         fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2650         h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2651         h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2652         if (!simple && FRAME_MBAFF) {
2653             filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2654         } else {
2655             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2656         }
2657     }
2658 }
2659
2660 /**
2661  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2662  */
2663 static void hl_decode_mb_simple(H264Context *h){
2664     hl_decode_mb_internal(h, 1);
2665 }
2666
2667 /**
2668  * Process a macroblock; this handles edge cases, such as interlacing.
2669  */
2670 static void av_noinline hl_decode_mb_complex(H264Context *h){
2671     hl_decode_mb_internal(h, 0);
2672 }
2673
2674 static void hl_decode_mb(H264Context *h){
2675     MpegEncContext * const s = &h->s;
2676     const int mb_xy= h->mb_xy;
2677     const int mb_type= s->current_picture.mb_type[mb_xy];
2678     int is_complex = ENABLE_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2679
2680     if(ENABLE_H264_ENCODER && !s->decode)
2681         return;
2682
2683     if (is_complex)
2684         hl_decode_mb_complex(h);
2685     else hl_decode_mb_simple(h);
2686 }
2687
2688 static void pic_as_field(Picture *pic, const int parity){
2689     int i;
2690     for (i = 0; i < 4; ++i) {
2691         if (parity == PICT_BOTTOM_FIELD)
2692             pic->data[i] += pic->linesize[i];
2693         pic->reference = parity;
2694         pic->linesize[i] *= 2;
2695     }
2696     pic->poc= pic->field_poc[parity == PICT_BOTTOM_FIELD];
2697 }
2698
2699 static int split_field_copy(Picture *dest, Picture *src,
2700                             int parity, int id_add){
2701     int match = !!(src->reference & parity);
2702
2703     if (match) {
2704         *dest = *src;
2705         if(parity != PICT_FRAME){
2706             pic_as_field(dest, parity);
2707             dest->pic_id *= 2;
2708             dest->pic_id += id_add;
2709         }
2710     }
2711
2712     return match;
2713 }
2714
2715 static int build_def_list(Picture *def, Picture **in, int len, int is_long, int sel){
2716     int i[2]={0};
2717     int index=0;
2718
2719     while(i[0]<len || i[1]<len){
2720         while(i[0]<len && !(in[ i[0] ] && (in[ i[0] ]->reference & sel)))
2721             i[0]++;
2722         while(i[1]<len && !(in[ i[1] ] && (in[ i[1] ]->reference & (sel^3))))
2723             i[1]++;
2724         if(i[0] < len){
2725             in[ i[0] ]->pic_id= is_long ? i[0] : in[ i[0] ]->frame_num;
2726             split_field_copy(&def[index++], in[ i[0]++ ], sel  , 1);
2727         }
2728         if(i[1] < len){
2729             in[ i[1] ]->pic_id= is_long ? i[1] : in[ i[1] ]->frame_num;
2730             split_field_copy(&def[index++], in[ i[1]++ ], sel^3, 0);
2731         }
2732     }
2733
2734     return index;
2735 }
2736
2737 static int add_sorted(Picture **sorted, Picture **src, int len, int limit, int dir){
2738     int i, best_poc;
2739     int out_i= 0;
2740
2741     for(;;){
2742         best_poc= dir ? INT_MIN : INT_MAX;
2743
2744         for(i=0; i<len; i++){
2745             const int poc= src[i]->poc;
2746             if(((poc > limit) ^ dir) && ((poc < best_poc) ^ dir)){
2747                 best_poc= poc;
2748                 sorted[out_i]= src[i];
2749             }
2750         }
2751         if(best_poc == (dir ? INT_MIN : INT_MAX))
2752             break;
2753         limit= sorted[out_i++]->poc - dir;
2754     }
2755     return out_i;
2756 }
2757
2758 /**
2759  * fills the default_ref_list.
2760  */
2761 static int fill_default_ref_list(H264Context *h){
2762     MpegEncContext * const s = &h->s;
2763     int i, len;
2764
2765     if(h->slice_type_nos==FF_B_TYPE){
2766         Picture *sorted[32];
2767         int cur_poc, list;
2768         int lens[2];
2769
2770         if(FIELD_PICTURE)
2771             cur_poc= s->current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
2772         else
2773             cur_poc= s->current_picture_ptr->poc;
2774
2775         for(list= 0; list<2; list++){
2776             len= add_sorted(sorted    , h->short_ref, h->short_ref_count, cur_poc, 1^list);
2777             len+=add_sorted(sorted+len, h->short_ref, h->short_ref_count, cur_poc, 0^list);
2778             assert(len<=32);
2779             len= build_def_list(h->default_ref_list[list]    , sorted     , len, 0, s->picture_structure);
2780             len+=build_def_list(h->default_ref_list[list]+len, h->long_ref, 16 , 1, s->picture_structure);
2781             assert(len<=32);
2782
2783             if(len < h->ref_count[list])
2784                 memset(&h->default_ref_list[list][len], 0, sizeof(Picture)*(h->ref_count[list] - len));
2785             lens[list]= len;
2786         }
2787
2788         if(lens[0] == lens[1] && lens[1] > 1){
2789             for(i=0; h->default_ref_list[0][i].data[0] == h->default_ref_list[1][i].data[0] && i<lens[0]; i++);
2790             if(i == lens[0])
2791                 FFSWAP(Picture, h->default_ref_list[1][0], h->default_ref_list[1][1]);
2792         }
2793     }else{
2794         len = build_def_list(h->default_ref_list[0]    , h->short_ref, h->short_ref_count, 0, s->picture_structure);
2795         len+= build_def_list(h->default_ref_list[0]+len, h-> long_ref, 16                , 1, s->picture_structure);
2796         assert(len <= 32);
2797         if(len < h->ref_count[0])
2798             memset(&h->default_ref_list[0][len], 0, sizeof(Picture)*(h->ref_count[0] - len));
2799     }
2800 #ifdef TRACE
2801     for (i=0; i<h->ref_count[0]; i++) {
2802         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]);
2803     }
2804     if(h->slice_type_nos==FF_B_TYPE){
2805         for (i=0; i<h->ref_count[1]; i++) {
2806             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]);
2807         }
2808     }
2809 #endif
2810     return 0;
2811 }
2812
2813 static void print_short_term(H264Context *h);
2814 static void print_long_term(H264Context *h);
2815
2816 /**
2817  * Extract structure information about the picture described by pic_num in
2818  * the current decoding context (frame or field). Note that pic_num is
2819  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
2820  * @param pic_num picture number for which to extract structure information
2821  * @param structure one of PICT_XXX describing structure of picture
2822  *                      with pic_num
2823  * @return frame number (short term) or long term index of picture
2824  *         described by pic_num
2825  */
2826 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
2827     MpegEncContext * const s = &h->s;
2828
2829     *structure = s->picture_structure;
2830     if(FIELD_PICTURE){
2831         if (!(pic_num & 1))
2832             /* opposite field */
2833             *structure ^= PICT_FRAME;
2834         pic_num >>= 1;
2835     }
2836
2837     return pic_num;
2838 }
2839
2840 static int decode_ref_pic_list_reordering(H264Context *h){
2841     MpegEncContext * const s = &h->s;
2842     int list, index, pic_structure;
2843
2844     print_short_term(h);
2845     print_long_term(h);
2846
2847     for(list=0; list<h->list_count; list++){
2848         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
2849
2850         if(get_bits1(&s->gb)){
2851             int pred= h->curr_pic_num;
2852
2853             for(index=0; ; index++){
2854                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb_31(&s->gb);
2855                 unsigned int pic_id;
2856                 int i;
2857                 Picture *ref = NULL;
2858
2859                 if(reordering_of_pic_nums_idc==3)
2860                     break;
2861
2862                 if(index >= h->ref_count[list]){
2863                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
2864                     return -1;
2865                 }
2866
2867                 if(reordering_of_pic_nums_idc<3){
2868                     if(reordering_of_pic_nums_idc<2){
2869                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
2870                         int frame_num;
2871
2872                         if(abs_diff_pic_num > h->max_pic_num){
2873                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
2874                             return -1;
2875                         }
2876
2877                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
2878                         else                                pred+= abs_diff_pic_num;
2879                         pred &= h->max_pic_num - 1;
2880
2881                         frame_num = pic_num_extract(h, pred, &pic_structure);
2882
2883                         for(i= h->short_ref_count-1; i>=0; i--){
2884                             ref = h->short_ref[i];
2885                             assert(ref->reference);
2886                             assert(!ref->long_ref);
2887                             if(
2888                                    ref->frame_num == frame_num &&
2889                                    (ref->reference & pic_structure)
2890                               )
2891                                 break;
2892                         }
2893                         if(i>=0)
2894                             ref->pic_id= pred;
2895                     }else{
2896                         int long_idx;
2897                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
2898
2899                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
2900
2901                         if(long_idx>31){
2902                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
2903                             return -1;
2904                         }
2905                         ref = h->long_ref[long_idx];
2906                         assert(!(ref && !ref->reference));
2907                         if(ref && (ref->reference & pic_structure)){
2908                             ref->pic_id= pic_id;
2909                             assert(ref->long_ref);
2910                             i=0;
2911                         }else{
2912                             i=-1;
2913                         }
2914                     }
2915
2916                     if (i < 0) {
2917                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
2918                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
2919                     } else {
2920                         for(i=index; i+1<h->ref_count[list]; i++){
2921                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
2922                                 break;
2923                         }
2924                         for(; i > index; i--){
2925                             h->ref_list[list][i]= h->ref_list[list][i-1];
2926                         }
2927                         h->ref_list[list][index]= *ref;
2928                         if (FIELD_PICTURE){
2929                             pic_as_field(&h->ref_list[list][index], pic_structure);
2930                         }
2931                     }
2932                 }else{
2933                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
2934                     return -1;
2935                 }
2936             }
2937         }
2938     }
2939     for(list=0; list<h->list_count; list++){
2940         for(index= 0; index < h->ref_count[list]; index++){
2941             if(!h->ref_list[list][index].data[0]){
2942                 av_log(h->s.avctx, AV_LOG_ERROR, "Missing reference picture\n");
2943                 h->ref_list[list][index]= s->current_picture; //FIXME this is not a sensible solution
2944             }
2945         }
2946     }
2947
2948     return 0;
2949 }
2950
2951 static void fill_mbaff_ref_list(H264Context *h){
2952     int list, i, j;
2953     for(list=0; list<2; list++){ //FIXME try list_count
2954         for(i=0; i<h->ref_count[list]; i++){
2955             Picture *frame = &h->ref_list[list][i];
2956             Picture *field = &h->ref_list[list][16+2*i];
2957             field[0] = *frame;
2958             for(j=0; j<3; j++)
2959                 field[0].linesize[j] <<= 1;
2960             field[0].reference = PICT_TOP_FIELD;
2961             field[0].poc= field[0].field_poc[0];
2962             field[1] = field[0];
2963             for(j=0; j<3; j++)
2964                 field[1].data[j] += frame->linesize[j];
2965             field[1].reference = PICT_BOTTOM_FIELD;
2966             field[1].poc= field[1].field_poc[1];
2967
2968             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
2969             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
2970             for(j=0; j<2; j++){
2971                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
2972                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
2973             }
2974         }
2975     }
2976     for(j=0; j<h->ref_count[1]; j++){
2977         for(i=0; i<h->ref_count[0]; i++)
2978             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
2979         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
2980         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
2981     }
2982 }
2983
2984 static int pred_weight_table(H264Context *h){
2985     MpegEncContext * const s = &h->s;
2986     int list, i;
2987     int luma_def, chroma_def;
2988
2989     h->use_weight= 0;
2990     h->use_weight_chroma= 0;
2991     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2992     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2993     luma_def = 1<<h->luma_log2_weight_denom;
2994     chroma_def = 1<<h->chroma_log2_weight_denom;
2995
2996     for(list=0; list<2; list++){
2997         for(i=0; i<h->ref_count[list]; i++){
2998             int luma_weight_flag, chroma_weight_flag;
2999
3000             luma_weight_flag= get_bits1(&s->gb);
3001             if(luma_weight_flag){
3002                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3003                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3004                 if(   h->luma_weight[list][i] != luma_def
3005                    || h->luma_offset[list][i] != 0)
3006                     h->use_weight= 1;
3007             }else{
3008                 h->luma_weight[list][i]= luma_def;
3009                 h->luma_offset[list][i]= 0;
3010             }
3011
3012             if(CHROMA){
3013                 chroma_weight_flag= get_bits1(&s->gb);
3014                 if(chroma_weight_flag){
3015                     int j;
3016                     for(j=0; j<2; j++){
3017                         h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3018                         h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3019                         if(   h->chroma_weight[list][i][j] != chroma_def
3020                         || h->chroma_offset[list][i][j] != 0)
3021                             h->use_weight_chroma= 1;
3022                     }
3023                 }else{
3024                     int j;
3025                     for(j=0; j<2; j++){
3026                         h->chroma_weight[list][i][j]= chroma_def;
3027                         h->chroma_offset[list][i][j]= 0;
3028                     }
3029                 }
3030             }
3031         }
3032         if(h->slice_type_nos != FF_B_TYPE) break;
3033     }
3034     h->use_weight= h->use_weight || h->use_weight_chroma;
3035     return 0;
3036 }
3037
3038 static void implicit_weight_table(H264Context *h){
3039     MpegEncContext * const s = &h->s;
3040     int ref0, ref1;
3041     int cur_poc = s->current_picture_ptr->poc;
3042
3043     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3044        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3045         h->use_weight= 0;
3046         h->use_weight_chroma= 0;
3047         return;
3048     }
3049
3050     h->use_weight= 2;
3051     h->use_weight_chroma= 2;
3052     h->luma_log2_weight_denom= 5;
3053     h->chroma_log2_weight_denom= 5;
3054
3055     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3056         int poc0 = h->ref_list[0][ref0].poc;
3057         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3058             int poc1 = h->ref_list[1][ref1].poc;
3059             int td = av_clip(poc1 - poc0, -128, 127);
3060             if(td){
3061                 int tb = av_clip(cur_poc - poc0, -128, 127);
3062                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3063                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3064                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3065                     h->implicit_weight[ref0][ref1] = 32;
3066                 else
3067                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3068             }else
3069                 h->implicit_weight[ref0][ref1] = 32;
3070         }
3071     }
3072 }
3073
3074 /**
3075  * Mark a picture as no longer needed for reference. The refmask
3076  * argument allows unreferencing of individual fields or the whole frame.
3077  * If the picture becomes entirely unreferenced, but is being held for
3078  * display purposes, it is marked as such.
3079  * @param refmask mask of fields to unreference; the mask is bitwise
3080  *                anded with the reference marking of pic
3081  * @return non-zero if pic becomes entirely unreferenced (except possibly
3082  *         for display purposes) zero if one of the fields remains in
3083  *         reference
3084  */
3085 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3086     int i;
3087     if (pic->reference &= refmask) {
3088         return 0;
3089     } else {
3090         for(i = 0; h->delayed_pic[i]; i++)
3091             if(pic == h->delayed_pic[i]){
3092                 pic->reference=DELAYED_PIC_REF;
3093                 break;
3094             }
3095         return 1;
3096     }
3097 }
3098
3099 /**
3100  * instantaneous decoder refresh.
3101  */
3102 static void idr(H264Context *h){
3103     int i;
3104
3105     for(i=0; i<16; i++){
3106         remove_long(h, i, 0);
3107     }
3108     assert(h->long_ref_count==0);
3109
3110     for(i=0; i<h->short_ref_count; i++){
3111         unreference_pic(h, h->short_ref[i], 0);
3112         h->short_ref[i]= NULL;
3113     }
3114     h->short_ref_count=0;
3115     h->prev_frame_num= 0;
3116     h->prev_frame_num_offset= 0;
3117     h->prev_poc_msb=
3118     h->prev_poc_lsb= 0;
3119 }
3120
3121 /* forget old pics after a seek */
3122 static void flush_dpb(AVCodecContext *avctx){
3123     H264Context *h= avctx->priv_data;
3124     int i;
3125     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
3126         if(h->delayed_pic[i])
3127             h->delayed_pic[i]->reference= 0;
3128         h->delayed_pic[i]= NULL;
3129     }
3130     h->outputed_poc= INT_MIN;
3131     idr(h);
3132     if(h->s.current_picture_ptr)
3133         h->s.current_picture_ptr->reference= 0;
3134     h->s.first_field= 0;
3135     ff_mpeg_flush(avctx);
3136 }
3137
3138 /**
3139  * Find a Picture in the short term reference list by frame number.
3140  * @param frame_num frame number to search for
3141  * @param idx the index into h->short_ref where returned picture is found
3142  *            undefined if no picture found.
3143  * @return pointer to the found picture, or NULL if no pic with the provided
3144  *                 frame number is found
3145  */
3146 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3147     MpegEncContext * const s = &h->s;
3148     int i;
3149
3150     for(i=0; i<h->short_ref_count; i++){
3151         Picture *pic= h->short_ref[i];
3152         if(s->avctx->debug&FF_DEBUG_MMCO)
3153             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3154         if(pic->frame_num == frame_num) {
3155             *idx = i;
3156             return pic;
3157         }
3158     }
3159     return NULL;
3160 }
3161
3162 /**
3163  * Remove a picture from the short term reference list by its index in
3164  * that list.  This does no checking on the provided index; it is assumed
3165  * to be valid. Other list entries are shifted down.
3166  * @param i index into h->short_ref of picture to remove.
3167  */
3168 static void remove_short_at_index(H264Context *h, int i){
3169     assert(i >= 0 && i < h->short_ref_count);
3170     h->short_ref[i]= NULL;
3171     if (--h->short_ref_count)
3172         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3173 }
3174
3175 /**
3176  *
3177  * @return the removed picture or NULL if an error occurs
3178  */
3179 static Picture * remove_short(H264Context *h, int frame_num, int ref_mask){
3180     MpegEncContext * const s = &h->s;
3181     Picture *pic;
3182     int i;
3183
3184     if(s->avctx->debug&FF_DEBUG_MMCO)
3185         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3186
3187     pic = find_short(h, frame_num, &i);
3188     if (pic){
3189         if(unreference_pic(h, pic, ref_mask))
3190         remove_short_at_index(h, i);
3191     }
3192
3193     return pic;
3194 }
3195
3196 /**
3197  * Remove a picture from the long term reference list by its index in
3198  * that list.
3199  * @return the removed picture or NULL if an error occurs
3200  */
3201 static Picture * remove_long(H264Context *h, int i, int ref_mask){
3202     Picture *pic;
3203
3204     pic= h->long_ref[i];
3205     if (pic){
3206         if(unreference_pic(h, pic, ref_mask)){
3207             assert(h->long_ref[i]->long_ref == 1);
3208             h->long_ref[i]->long_ref= 0;
3209             h->long_ref[i]= NULL;
3210             h->long_ref_count--;
3211         }
3212     }
3213
3214     return pic;
3215 }
3216
3217 /**
3218  * print short term list
3219  */
3220 static void print_short_term(H264Context *h) {
3221     uint32_t i;
3222     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3223         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3224         for(i=0; i<h->short_ref_count; i++){
3225             Picture *pic= h->short_ref[i];
3226             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3227         }
3228     }
3229 }
3230
3231 /**
3232  * print long term list
3233  */
3234 static void print_long_term(H264Context *h) {
3235     uint32_t i;
3236     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3237         av_log(h->s.avctx, AV_LOG_DEBUG, "long term list:\n");
3238         for(i = 0; i < 16; i++){
3239             Picture *pic= h->long_ref[i];
3240             if (pic) {
3241                 av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3242             }
3243         }
3244     }
3245 }
3246
3247 /**
3248  * Executes the reference picture marking (memory management control operations).
3249  */
3250 static int execute_ref_pic_marking(H264Context *h, MMCO *mmco, int mmco_count){
3251     MpegEncContext * const s = &h->s;
3252     int i, j;
3253     int current_ref_assigned=0;
3254     Picture *pic;
3255
3256     if((s->avctx->debug&FF_DEBUG_MMCO) && mmco_count==0)
3257         av_log(h->s.avctx, AV_LOG_DEBUG, "no mmco here\n");
3258
3259     for(i=0; i<mmco_count; i++){
3260         int structure, frame_num;
3261         if(s->avctx->debug&FF_DEBUG_MMCO)
3262             av_log(h->s.avctx, AV_LOG_DEBUG, "mmco:%d %d %d\n", h->mmco[i].opcode, h->mmco[i].short_pic_num, h->mmco[i].long_arg);
3263
3264         if(   mmco[i].opcode == MMCO_SHORT2UNUSED
3265            || mmco[i].opcode == MMCO_SHORT2LONG){
3266             frame_num = pic_num_extract(h, mmco[i].short_pic_num, &structure);
3267             pic = find_short(h, frame_num, &j);
3268             if(!pic){
3269                 if(mmco[i].opcode != MMCO_SHORT2LONG || !h->long_ref[mmco[i].long_arg]
3270                    || h->long_ref[mmco[i].long_arg]->frame_num != frame_num)
3271                 av_log(h->s.avctx, AV_LOG_ERROR, "mmco: unref short failure\n");
3272                 continue;
3273             }
3274         }
3275
3276         switch(mmco[i].opcode){
3277         case MMCO_SHORT2UNUSED:
3278             if(s->avctx->debug&FF_DEBUG_MMCO)
3279                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref short %d count %d\n", h->mmco[i].short_pic_num, h->short_ref_count);
3280             remove_short(h, frame_num, structure ^ PICT_FRAME);
3281             break;
3282         case MMCO_SHORT2LONG:
3283                 if (h->long_ref[mmco[i].long_arg] != pic)
3284                     remove_long(h, mmco[i].long_arg, 0);
3285
3286                 remove_short_at_index(h, j);
3287                 h->long_ref[ mmco[i].long_arg ]= pic;
3288                 if (h->long_ref[ mmco[i].long_arg ]){
3289                     h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3290                     h->long_ref_count++;
3291                 }
3292             break;
3293         case MMCO_LONG2UNUSED:
3294             j = pic_num_extract(h, mmco[i].long_arg, &structure);
3295             pic = h->long_ref[j];
3296             if (pic) {
3297                 remove_long(h, j, structure ^ PICT_FRAME);
3298             } else if(s->avctx->debug&FF_DEBUG_MMCO)
3299                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref long failure\n");
3300             break;
3301         case MMCO_LONG:
3302                     // Comment below left from previous code as it is an interresting note.
3303                     /* First field in pair is in short term list or
3304                      * at a different long term index.
3305                      * This is not allowed; see 7.4.3.3, notes 2 and 3.
3306                      * Report the problem and keep the pair where it is,
3307                      * and mark this field valid.
3308                      */
3309
3310             if (h->long_ref[mmco[i].long_arg] != s->current_picture_ptr) {
3311                 remove_long(h, mmco[i].long_arg, 0);
3312
3313                 h->long_ref[ mmco[i].long_arg ]= s->current_picture_ptr;
3314                 h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3315                 h->long_ref_count++;
3316             }
3317
3318             s->current_picture_ptr->reference |= s->picture_structure;
3319             current_ref_assigned=1;
3320             break;
3321         case MMCO_SET_MAX_LONG:
3322             assert(mmco[i].long_arg <= 16);
3323             // just remove the long term which index is greater than new max
3324             for(j = mmco[i].long_arg; j<16; j++){
3325                 remove_long(h, j, 0);
3326             }
3327             break;
3328         case MMCO_RESET:
3329             while(h->short_ref_count){
3330                 remove_short(h, h->short_ref[0]->frame_num, 0);
3331             }
3332             for(j = 0; j < 16; j++) {
3333                 remove_long(h, j, 0);
3334             }
3335             s->current_picture_ptr->poc=
3336             s->current_picture_ptr->field_poc[0]=
3337             s->current_picture_ptr->field_poc[1]=
3338             h->poc_lsb=
3339             h->poc_msb=
3340             h->frame_num=
3341             s->current_picture_ptr->frame_num= 0;
3342             break;
3343         default: assert(0);
3344         }
3345     }
3346
3347     if (!current_ref_assigned) {
3348         /* Second field of complementary field pair; the first field of
3349          * which is already referenced. If short referenced, it
3350          * should be first entry in short_ref. If not, it must exist
3351          * in long_ref; trying to put it on the short list here is an
3352          * error in the encoded bit stream (ref: 7.4.3.3, NOTE 2 and 3).
3353          */
3354         if (h->short_ref_count && h->short_ref[0] == s->current_picture_ptr) {
3355             /* Just mark the second field valid */
3356             s->current_picture_ptr->reference = PICT_FRAME;
3357         } else if (s->current_picture_ptr->long_ref) {
3358             av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term reference "
3359                                              "assignment for second field "
3360                                              "in complementary field pair "
3361                                              "(first field is long term)\n");
3362         } else {
3363             pic= remove_short(h, s->current_picture_ptr->frame_num, 0);
3364             if(pic){
3365                 av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term buffer state detected\n");
3366             }
3367
3368             if(h->short_ref_count)
3369                 memmove(&h->short_ref[1], &h->short_ref[0], h->short_ref_count*sizeof(Picture*));
3370
3371             h->short_ref[0]= s->current_picture_ptr;
3372             h->short_ref_count++;
3373             s->current_picture_ptr->reference |= s->picture_structure;
3374         }
3375     }
3376
3377     if (h->long_ref_count + h->short_ref_count > h->sps.ref_frame_count){
3378
3379         /* We have too many reference frames, probably due to corrupted
3380          * stream. Need to discard one frame. Prevents overrun of the
3381          * short_ref and long_ref buffers.
3382          */
3383         av_log(h->s.avctx, AV_LOG_ERROR,
3384                "number of reference frames exceeds max (probably "
3385                "corrupt input), discarding one\n");
3386
3387         if (h->long_ref_count && !h->short_ref_count) {
3388             for (i = 0; i < 16; ++i)
3389                 if (h->long_ref[i])
3390                     break;
3391
3392             assert(i < 16);
3393             remove_long(h, i, 0);
3394         } else {
3395             pic = h->short_ref[h->short_ref_count - 1];
3396             remove_short(h, pic->frame_num, 0);
3397         }
3398     }
3399
3400     print_short_term(h);
3401     print_long_term(h);
3402     return 0;
3403 }
3404
3405 static int decode_ref_pic_marking(H264Context *h, GetBitContext *gb){
3406     MpegEncContext * const s = &h->s;
3407     int i;
3408
3409     h->mmco_index= 0;
3410     if(h->nal_unit_type == NAL_IDR_SLICE){ //FIXME fields
3411         s->broken_link= get_bits1(gb) -1;
3412         if(get_bits1(gb)){
3413             h->mmco[0].opcode= MMCO_LONG;
3414             h->mmco[0].long_arg= 0;
3415             h->mmco_index= 1;
3416         }
3417     }else{
3418         if(get_bits1(gb)){ // adaptive_ref_pic_marking_mode_flag
3419             for(i= 0; i<MAX_MMCO_COUNT; i++) {
3420                 MMCOOpcode opcode= get_ue_golomb_31(gb);
3421
3422                 h->mmco[i].opcode= opcode;
3423                 if(opcode==MMCO_SHORT2UNUSED || opcode==MMCO_SHORT2LONG){
3424                     h->mmco[i].short_pic_num= (h->curr_pic_num - get_ue_golomb(gb) - 1) & (h->max_pic_num - 1);
3425 /*                    if(h->mmco[i].short_pic_num >= h->short_ref_count || h->short_ref[ h->mmco[i].short_pic_num ] == NULL){
3426                         av_log(s->avctx, AV_LOG_ERROR,&