ec827c0e02a29a3224285aabca33f29fdd922f40
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file h264.c
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "dsputil.h"
29 #include "avcodec.h"
30 #include "mpegvideo.h"
31 #include "h264.h"
32 #include "h264data.h"
33 #include "h264_parser.h"
34 #include "golomb.h"
35 #include "rectangle.h"
36
37 #include "cabac.h"
38 #ifdef ARCH_X86
39 #include "i386/h264_i386.h"
40 #endif
41
42 //#undef NDEBUG
43 #include <assert.h>
44
45 /**
46  * Value of Picture.reference when Picture is not a reference picture, but
47  * is held for delayed output.
48  */
49 #define DELAYED_PIC_REF 4
50
51 static VLC coeff_token_vlc[4];
52 static VLC_TYPE coeff_token_vlc_tables[520+332+280+256][2];
53 static const int coeff_token_vlc_tables_size[4]={520,332,280,256};
54
55 static VLC chroma_dc_coeff_token_vlc;
56 static VLC_TYPE chroma_dc_coeff_token_vlc_table[256][2];
57 static const int chroma_dc_coeff_token_vlc_table_size = 256;
58
59 static VLC total_zeros_vlc[15];
60 static VLC_TYPE total_zeros_vlc_tables[15][512][2];
61 static const int total_zeros_vlc_tables_size = 512;
62
63 static VLC chroma_dc_total_zeros_vlc[3];
64 static VLC_TYPE chroma_dc_total_zeros_vlc_tables[3][8][2];
65 static const int chroma_dc_total_zeros_vlc_tables_size = 8;
66
67 static VLC run_vlc[6];
68 static VLC_TYPE run_vlc_tables[6][8][2];
69 static const int run_vlc_tables_size = 8;
70
71 static VLC run7_vlc;
72 static VLC_TYPE run7_vlc_table[96][2];
73 static const int run7_vlc_table_size = 96;
74
75 static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
76 static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
77 static void filter_mb( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
78 static void filter_mb_fast( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
79 static Picture * remove_long(H264Context *h, int i, int ref_mask);
80
81 static av_always_inline uint32_t pack16to32(int a, int b){
82 #ifdef WORDS_BIGENDIAN
83    return (b&0xFFFF) + (a<<16);
84 #else
85    return (a&0xFFFF) + (b<<16);
86 #endif
87 }
88
89 const uint8_t ff_rem6[52]={
90 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
91 };
92
93 const uint8_t ff_div6[52]={
94 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8,
95 };
96
97 static const int left_block_options[4][8]={
98     {0,1,2,3,7,10,8,11},
99     {2,2,3,3,8,11,8,11},
100     {0,0,1,1,7,10,7,10},
101     {0,2,0,2,7,10,7,10}
102 };
103
104 static void fill_caches(H264Context *h, int mb_type, int for_deblock){
105     MpegEncContext * const s = &h->s;
106     const int mb_xy= h->mb_xy;
107     int topleft_xy, top_xy, topright_xy, left_xy[2];
108     int topleft_type, top_type, topright_type, left_type[2];
109     int * left_block;
110     int topleft_partition= -1;
111     int i;
112
113     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
114
115     //FIXME deblocking could skip the intra and nnz parts.
116     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
117         return;
118
119     /* Wow, what a mess, why didn't they simplify the interlacing & intra
120      * stuff, I can't imagine that these complex rules are worth it. */
121
122     topleft_xy = top_xy - 1;
123     topright_xy= top_xy + 1;
124     left_xy[1] = left_xy[0] = mb_xy-1;
125     left_block = left_block_options[0];
126     if(FRAME_MBAFF){
127         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
128         const int top_pair_xy      = pair_xy     - s->mb_stride;
129         const int topleft_pair_xy  = top_pair_xy - 1;
130         const int topright_pair_xy = top_pair_xy + 1;
131         const int topleft_mb_frame_flag  = !IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
132         const int top_mb_frame_flag      = !IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
133         const int topright_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
134         const int left_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
135         const int curr_mb_frame_flag = !IS_INTERLACED(mb_type);
136         const int bottom = (s->mb_y & 1);
137         tprintf(s->avctx, "fill_caches: curr_mb_frame_flag:%d, left_mb_frame_flag:%d, topleft_mb_frame_flag:%d, top_mb_frame_flag:%d, topright_mb_frame_flag:%d\n", curr_mb_frame_flag, left_mb_frame_flag, topleft_mb_frame_flag, top_mb_frame_flag, topright_mb_frame_flag);
138         if (bottom
139                 ? !curr_mb_frame_flag // bottom macroblock
140                 : (!curr_mb_frame_flag && !top_mb_frame_flag) // top macroblock
141                 ) {
142             top_xy -= s->mb_stride;
143         }
144         if (bottom
145                 ? !curr_mb_frame_flag // bottom macroblock
146                 : (!curr_mb_frame_flag && !topleft_mb_frame_flag) // top macroblock
147                 ) {
148             topleft_xy -= s->mb_stride;
149         } else if(bottom && curr_mb_frame_flag && !left_mb_frame_flag) {
150             topleft_xy += s->mb_stride;
151             // take top left mv from the middle of the mb, as opposed to all other modes which use the bottom right partition
152             topleft_partition = 0;
153         }
154         if (bottom
155                 ? !curr_mb_frame_flag // bottom macroblock
156                 : (!curr_mb_frame_flag && !topright_mb_frame_flag) // top macroblock
157                 ) {
158             topright_xy -= s->mb_stride;
159         }
160         if (left_mb_frame_flag != curr_mb_frame_flag) {
161             left_xy[1] = left_xy[0] = pair_xy - 1;
162             if (curr_mb_frame_flag) {
163                 if (bottom) {
164                     left_block = left_block_options[1];
165                 } else {
166                     left_block= left_block_options[2];
167                 }
168             } else {
169                 left_xy[1] += s->mb_stride;
170                 left_block = left_block_options[3];
171             }
172         }
173     }
174
175     h->top_mb_xy = top_xy;
176     h->left_mb_xy[0] = left_xy[0];
177     h->left_mb_xy[1] = left_xy[1];
178     if(for_deblock){
179         topleft_type = 0;
180         topright_type = 0;
181         top_type     = h->slice_table[top_xy     ] < 255 ? s->current_picture.mb_type[top_xy]     : 0;
182         left_type[0] = h->slice_table[left_xy[0] ] < 255 ? s->current_picture.mb_type[left_xy[0]] : 0;
183         left_type[1] = h->slice_table[left_xy[1] ] < 255 ? s->current_picture.mb_type[left_xy[1]] : 0;
184
185         if(FRAME_MBAFF && !IS_INTRA(mb_type)){
186             int list;
187             int v = *(uint16_t*)&h->non_zero_count[mb_xy][14];
188             for(i=0; i<16; i++)
189                 h->non_zero_count_cache[scan8[i]] = (v>>i)&1;
190             for(list=0; list<h->list_count; list++){
191                 if(USES_LIST(mb_type,list)){
192                     uint32_t *src = (uint32_t*)s->current_picture.motion_val[list][h->mb2b_xy[mb_xy]];
193                     uint32_t *dst = (uint32_t*)h->mv_cache[list][scan8[0]];
194                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
195                     for(i=0; i<4; i++, dst+=8, src+=h->b_stride){
196                         dst[0] = src[0];
197                         dst[1] = src[1];
198                         dst[2] = src[2];
199                         dst[3] = src[3];
200                     }
201                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
202                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = pack16to32(ref[0],ref[1])*0x0101;
203                     ref += h->b8_stride;
204                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
205                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = pack16to32(ref[0],ref[1])*0x0101;
206                 }else{
207                     fill_rectangle(&h-> mv_cache[list][scan8[ 0]], 4, 4, 8, 0, 4);
208                     fill_rectangle(&h->ref_cache[list][scan8[ 0]], 4, 4, 8, (uint8_t)LIST_NOT_USED, 1);
209                 }
210             }
211         }
212     }else{
213         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
214         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
215         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
216         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
217         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
218     }
219
220     if(IS_INTRA(mb_type)){
221         h->topleft_samples_available=
222         h->top_samples_available=
223         h->left_samples_available= 0xFFFF;
224         h->topright_samples_available= 0xEEEA;
225
226         if(!IS_INTRA(top_type) && (top_type==0 || h->pps.constrained_intra_pred)){
227             h->topleft_samples_available= 0xB3FF;
228             h->top_samples_available= 0x33FF;
229             h->topright_samples_available= 0x26EA;
230         }
231         for(i=0; i<2; i++){
232             if(!IS_INTRA(left_type[i]) && (left_type[i]==0 || h->pps.constrained_intra_pred)){
233                 h->topleft_samples_available&= 0xDF5F;
234                 h->left_samples_available&= 0x5F5F;
235             }
236         }
237
238         if(!IS_INTRA(topleft_type) && (topleft_type==0 || h->pps.constrained_intra_pred))
239             h->topleft_samples_available&= 0x7FFF;
240
241         if(!IS_INTRA(topright_type) && (topright_type==0 || h->pps.constrained_intra_pred))
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 || (IS_INTER(top_type) && h->pps.constrained_intra_pred))
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] || (IS_INTER(left_type[i]) && h->pps.constrained_intra_pred))
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 0 . T T. T T T T
283 1 L . .L . . . .
284 2 L . .L . . . .
285 3 . T TL . . . .
286 4 L . .L . . . .
287 5 L . .. . . . .
288 */
289 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
290     if(top_type){
291         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
292         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
293         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
294         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
295
296         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
297         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
298
299         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
300         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
301
302     }else{
303         h->non_zero_count_cache[4+8*0]=
304         h->non_zero_count_cache[5+8*0]=
305         h->non_zero_count_cache[6+8*0]=
306         h->non_zero_count_cache[7+8*0]=
307
308         h->non_zero_count_cache[1+8*0]=
309         h->non_zero_count_cache[2+8*0]=
310
311         h->non_zero_count_cache[1+8*3]=
312         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
313
314     }
315
316     for (i=0; i<2; i++) {
317         if(left_type[i]){
318             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
319             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
320             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
321             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
322         }else{
323             h->non_zero_count_cache[3+8*1 + 2*8*i]=
324             h->non_zero_count_cache[3+8*2 + 2*8*i]=
325             h->non_zero_count_cache[0+8*1 +   8*i]=
326             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
327         }
328     }
329
330     if( h->pps.cabac ) {
331         // top_cbp
332         if(top_type) {
333             h->top_cbp = h->cbp_table[top_xy];
334         } else if(IS_INTRA(mb_type)) {
335             h->top_cbp = 0x1C0;
336         } else {
337             h->top_cbp = 0;
338         }
339         // left_cbp
340         if (left_type[0]) {
341             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
342         } else if(IS_INTRA(mb_type)) {
343             h->left_cbp = 0x1C0;
344         } else {
345             h->left_cbp = 0;
346         }
347         if (left_type[0]) {
348             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
349         }
350         if (left_type[1]) {
351             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
352         }
353     }
354
355 #if 1
356     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
357         int list;
358         for(list=0; list<h->list_count; list++){
359             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
360                 /*if(!h->mv_cache_clean[list]){
361                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
362                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
363                     h->mv_cache_clean[list]= 1;
364                 }*/
365                 continue;
366             }
367             h->mv_cache_clean[list]= 0;
368
369             if(USES_LIST(top_type, list)){
370                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
371                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
372                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
373                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
374                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
375                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
376                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
377                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
378                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
379                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
380             }else{
381                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
382                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
383                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
384                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
385                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
386             }
387
388             for(i=0; i<2; i++){
389                 int cache_idx = scan8[0] - 1 + i*2*8;
390                 if(USES_LIST(left_type[i], list)){
391                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
392                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
393                     *(uint32_t*)h->mv_cache[list][cache_idx  ]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[0+i*2]];
394                     *(uint32_t*)h->mv_cache[list][cache_idx+8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[1+i*2]];
395                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
396                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
397                 }else{
398                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
399                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
400                     h->ref_cache[list][cache_idx  ]=
401                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
402                 }
403             }
404
405             if((for_deblock || (IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred)) && !FRAME_MBAFF)
406                 continue;
407
408             if(USES_LIST(topleft_type, list)){
409                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
410                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
411                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
412                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
413             }else{
414                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
415                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
416             }
417
418             if(USES_LIST(topright_type, list)){
419                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
420                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
421                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
422                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
423             }else{
424                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
425                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
426             }
427
428             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
429                 continue;
430
431             h->ref_cache[list][scan8[5 ]+1] =
432             h->ref_cache[list][scan8[7 ]+1] =
433             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
434             h->ref_cache[list][scan8[4 ]] =
435             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
436             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
437             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
438             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
439             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
440             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
441
442             if( h->pps.cabac ) {
443                 /* XXX beurk, Load mvd */
444                 if(USES_LIST(top_type, list)){
445                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
446                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
447                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
448                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
449                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
450                 }else{
451                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
452                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
453                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
454                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
455                 }
456                 if(USES_LIST(left_type[0], list)){
457                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
458                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 0*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[0]];
459                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[1]];
460                 }else{
461                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
462                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
463                 }
464                 if(USES_LIST(left_type[1], list)){
465                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
466                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 2*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[2]];
467                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 3*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[3]];
468                 }else{
469                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
470                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
471                 }
472                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
473                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
474                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
475                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
476                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
477
478                 if(h->slice_type_nos == FF_B_TYPE){
479                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
480
481                     if(IS_DIRECT(top_type)){
482                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
483                     }else if(IS_8X8(top_type)){
484                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
485                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
486                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
487                     }else{
488                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
489                     }
490
491                     if(IS_DIRECT(left_type[0]))
492                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
493                     else if(IS_8X8(left_type[0]))
494                         h->direct_cache[scan8[0] - 1 + 0*8]= h->direct_table[h->mb2b8_xy[left_xy[0]] + 1 + h->b8_stride*(left_block[0]>>1)];
495                     else
496                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
497
498                     if(IS_DIRECT(left_type[1]))
499                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
500                     else if(IS_8X8(left_type[1]))
501                         h->direct_cache[scan8[0] - 1 + 2*8]= h->direct_table[h->mb2b8_xy[left_xy[1]] + 1 + h->b8_stride*(left_block[2]>>1)];
502                     else
503                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
504                 }
505             }
506
507             if(FRAME_MBAFF){
508 #define MAP_MVS\
509                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
510                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
511                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
512                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
513                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
514                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
515                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
516                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
517                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
518                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
519                 if(MB_FIELD){
520 #define MAP_F2F(idx, mb_type)\
521                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
522                         h->ref_cache[list][idx] <<= 1;\
523                         h->mv_cache[list][idx][1] /= 2;\
524                         h->mvd_cache[list][idx][1] /= 2;\
525                     }
526                     MAP_MVS
527 #undef MAP_F2F
528                 }else{
529 #define MAP_F2F(idx, mb_type)\
530                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
531                         h->ref_cache[list][idx] >>= 1;\
532                         h->mv_cache[list][idx][1] <<= 1;\
533                         h->mvd_cache[list][idx][1] <<= 1;\
534                     }
535                     MAP_MVS
536 #undef MAP_F2F
537                 }
538             }
539         }
540     }
541 #endif
542
543     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
544 }
545
546 static inline void write_back_intra_pred_mode(H264Context *h){
547     const int mb_xy= h->mb_xy;
548
549     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
550     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
551     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
552     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
553     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
554     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
555     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
556 }
557
558 /**
559  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
560  */
561 static inline int check_intra4x4_pred_mode(H264Context *h){
562     MpegEncContext * const s = &h->s;
563     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
564     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
565     int i;
566
567     if(!(h->top_samples_available&0x8000)){
568         for(i=0; i<4; i++){
569             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
570             if(status<0){
571                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
572                 return -1;
573             } else if(status){
574                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
575             }
576         }
577     }
578
579     if(!(h->left_samples_available&0x8000)){
580         for(i=0; i<4; i++){
581             int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
582             if(status<0){
583                 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);
584                 return -1;
585             } else if(status){
586                 h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
587             }
588         }
589     }
590
591     return 0;
592 } //FIXME cleanup like next
593
594 /**
595  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
596  */
597 static inline int check_intra_pred_mode(H264Context *h, int mode){
598     MpegEncContext * const s = &h->s;
599     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
600     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
601
602     if(mode > 6U) {
603         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);
604         return -1;
605     }
606
607     if(!(h->top_samples_available&0x8000)){
608         mode= top[ mode ];
609         if(mode<0){
610             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);
611             return -1;
612         }
613     }
614
615     if(!(h->left_samples_available&0x8000)){
616         mode= left[ mode ];
617         if(mode<0){
618             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);
619             return -1;
620         }
621     }
622
623     return mode;
624 }
625
626 /**
627  * gets the predicted intra4x4 prediction mode.
628  */
629 static inline int pred_intra_mode(H264Context *h, int n){
630     const int index8= scan8[n];
631     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
632     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
633     const int min= FFMIN(left, top);
634
635     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
636
637     if(min<0) return DC_PRED;
638     else      return min;
639 }
640
641 static inline void write_back_non_zero_count(H264Context *h){
642     const int mb_xy= h->mb_xy;
643
644     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
645     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
646     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
647     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
648     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
649     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
650     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
651
652     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
653     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
654     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
655
656     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
657     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
658     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
659
660     if(FRAME_MBAFF){
661         // store all luma nnzs, for deblocking
662         int v = 0, i;
663         for(i=0; i<16; i++)
664             v += (!!h->non_zero_count_cache[scan8[i]]) << i;
665         *(uint16_t*)&h->non_zero_count[mb_xy][14] = v;
666     }
667 }
668
669 /**
670  * gets the predicted number of non-zero coefficients.
671  * @param n block index
672  */
673 static inline int pred_non_zero_count(H264Context *h, int n){
674     const int index8= scan8[n];
675     const int left= h->non_zero_count_cache[index8 - 1];
676     const int top = h->non_zero_count_cache[index8 - 8];
677     int i= left + top;
678
679     if(i<64) i= (i+1)>>1;
680
681     tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);
682
683     return i&31;
684 }
685
686 static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
687     const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
688     MpegEncContext *s = &h->s;
689
690     /* there is no consistent mapping of mvs to neighboring locations that will
691      * make mbaff happy, so we can't move all this logic to fill_caches */
692     if(FRAME_MBAFF){
693         const uint32_t *mb_types = s->current_picture_ptr->mb_type;
694         const int16_t *mv;
695         *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
696         *C = h->mv_cache[list][scan8[0]-2];
697
698         if(!MB_FIELD
699            && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
700             int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
701             if(IS_INTERLACED(mb_types[topright_xy])){
702 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
703                 const int x4 = X4, y4 = Y4;\
704                 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
705                 if(!USES_LIST(mb_type,list))\
706                     return LIST_NOT_USED;\
707                 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
708                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
709                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
710                 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
711
712                 SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
713             }
714         }
715         if(topright_ref == PART_NOT_AVAILABLE
716            && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
717            && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
718             if(!MB_FIELD
719                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
720                 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
721             }
722             if(MB_FIELD
723                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
724                && i >= scan8[0]+8){
725                 // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
726                 SET_DIAG_MV(/2, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
727             }
728         }
729 #undef SET_DIAG_MV
730     }
731
732     if(topright_ref != PART_NOT_AVAILABLE){
733         *C= h->mv_cache[list][ i - 8 + part_width ];
734         return topright_ref;
735     }else{
736         tprintf(s->avctx, "topright MV not available\n");
737
738         *C= h->mv_cache[list][ i - 8 - 1 ];
739         return h->ref_cache[list][ i - 8 - 1 ];
740     }
741 }
742
743 /**
744  * gets the predicted MV.
745  * @param n the block index
746  * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
747  * @param mx the x component of the predicted motion vector
748  * @param my the y component of the predicted motion vector
749  */
750 static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
751     const int index8= scan8[n];
752     const int top_ref=      h->ref_cache[list][ index8 - 8 ];
753     const int left_ref=     h->ref_cache[list][ index8 - 1 ];
754     const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
755     const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
756     const int16_t * C;
757     int diagonal_ref, match_count;
758
759     assert(part_width==1 || part_width==2 || part_width==4);
760
761 /* mv_cache
762   B . . A T T T T
763   U . . L . . , .
764   U . . L . . . .
765   U . . L . . , .
766   . . . L . . . .
767 */
768
769     diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
770     match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
771     tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
772     if(match_count > 1){ //most common
773         *mx= mid_pred(A[0], B[0], C[0]);
774         *my= mid_pred(A[1], B[1], C[1]);
775     }else if(match_count==1){
776         if(left_ref==ref){
777             *mx= A[0];
778             *my= A[1];
779         }else if(top_ref==ref){
780             *mx= B[0];
781             *my= B[1];
782         }else{
783             *mx= C[0];
784             *my= C[1];
785         }
786     }else{
787         if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
788             *mx= A[0];
789             *my= A[1];
790         }else{
791             *mx= mid_pred(A[0], B[0], C[0]);
792             *my= mid_pred(A[1], B[1], C[1]);
793         }
794     }
795
796     tprintf(h->s.avctx, "pred_motion (%2d %2d %2d) (%2d %2d %2d) (%2d %2d %2d) -> (%2d %2d %2d) at %2d %2d %d list %d\n", top_ref, B[0], B[1],                    diagonal_ref, C[0], C[1], left_ref, A[0], A[1], ref, *mx, *my, h->s.mb_x, h->s.mb_y, n, list);
797 }
798
799 /**
800  * gets the directionally predicted 16x8 MV.
801  * @param n the block index
802  * @param mx the x component of the predicted motion vector
803  * @param my the y component of the predicted motion vector
804  */
805 static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
806     if(n==0){
807         const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
808         const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
809
810         tprintf(h->s.avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n", top_ref, B[0], B[1], h->s.mb_x, h->s.mb_y, n, list);
811
812         if(top_ref == ref){
813             *mx= B[0];
814             *my= B[1];
815             return;
816         }
817     }else{
818         const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
819         const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];
820
821         tprintf(h->s.avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n", left_ref, A[0], A[1], h->s.mb_x, h->s.mb_y, n, list);
822
823         if(left_ref == ref){
824             *mx= A[0];
825             *my= A[1];
826             return;
827         }
828     }
829
830     //RARE
831     pred_motion(h, n, 4, list, ref, mx, my);
832 }
833
834 /**
835  * gets the directionally predicted 8x16 MV.
836  * @param n the block index
837  * @param mx the x component of the predicted motion vector
838  * @param my the y component of the predicted motion vector
839  */
840 static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
841     if(n==0){
842         const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
843         const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];
844
845         tprintf(h->s.avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n", left_ref, A[0], A[1], h->s.mb_x, h->s.mb_y, n, list);
846
847         if(left_ref == ref){
848             *mx= A[0];
849             *my= A[1];
850             return;
851         }
852     }else{
853         const int16_t * C;
854         int diagonal_ref;
855
856         diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);
857
858         tprintf(h->s.avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n", diagonal_ref, C[0], C[1], h->s.mb_x, h->s.mb_y, n, list);
859
860         if(diagonal_ref == ref){
861             *mx= C[0];
862             *my= C[1];
863             return;
864         }
865     }
866
867     //RARE
868     pred_motion(h, n, 2, list, ref, mx, my);
869 }
870
871 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
872     const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
873     const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];
874
875     tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);
876
877     if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
878        || (top_ref == 0  && *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ] == 0)
879        || (left_ref == 0 && *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ] == 0)){
880
881         *mx = *my = 0;
882         return;
883     }
884
885     pred_motion(h, 0, 4, 0, 0, mx, my);
886
887     return;
888 }
889
890 static inline void direct_dist_scale_factor(H264Context * const h){
891     MpegEncContext * const s = &h->s;
892     const int poc = h->s.current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
893     const int poc1 = h->ref_list[1][0].poc;
894     int i;
895     for(i=0; i<h->ref_count[0]; i++){
896         int poc0 = h->ref_list[0][i].poc;
897         int td = av_clip(poc1 - poc0, -128, 127);
898         if(td == 0 || h->ref_list[0][i].long_ref){
899             h->dist_scale_factor[i] = 256;
900         }else{
901             int tb = av_clip(poc - poc0, -128, 127);
902             int tx = (16384 + (FFABS(td) >> 1)) / td;
903             h->dist_scale_factor[i] = av_clip((tb*tx + 32) >> 6, -1024, 1023);
904         }
905     }
906     if(FRAME_MBAFF){
907         for(i=0; i<h->ref_count[0]; i++){
908             h->dist_scale_factor_field[2*i] =
909             h->dist_scale_factor_field[2*i+1] = h->dist_scale_factor[i];
910         }
911     }
912 }
913 static inline void direct_ref_list_init(H264Context * const h){
914     MpegEncContext * const s = &h->s;
915     Picture * const ref1 = &h->ref_list[1][0];
916     Picture * const cur = s->current_picture_ptr;
917     int list, i, j;
918     int sidx= s->picture_structure&1;
919     int ref1sidx= ref1->reference&1;
920     for(list=0; list<2; list++){
921         cur->ref_count[sidx][list] = h->ref_count[list];
922         for(j=0; j<h->ref_count[list]; j++)
923             cur->ref_poc[sidx][list][j] = 4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3);
924     }
925     if(s->picture_structure == PICT_FRAME){
926         memcpy(cur->ref_count[0], cur->ref_count[1], sizeof(cur->ref_count[0]));
927         memcpy(cur->ref_poc  [0], cur->ref_poc  [1], sizeof(cur->ref_poc  [0]));
928     }
929     if(cur->pict_type != FF_B_TYPE || h->direct_spatial_mv_pred)
930         return;
931     for(list=0; list<2; list++){
932         for(i=0; i<ref1->ref_count[ref1sidx][list]; i++){
933             int poc = ref1->ref_poc[ref1sidx][list][i];
934             if(((poc&3) == 3) != (s->picture_structure == PICT_FRAME))
935                 poc= (poc&~3) + s->picture_structure;
936             h->map_col_to_list0[list][i] = 0; /* bogus; fills in for missing frames */
937             for(j=0; j<h->ref_count[list]; j++)
938                 if(4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3) == poc){
939                     h->map_col_to_list0[list][i] = j;
940                     break;
941                 }
942         }
943     }
944     if(FRAME_MBAFF){
945         for(list=0; list<2; list++){
946             for(i=0; i<ref1->ref_count[ref1sidx][list]; i++){
947                 j = h->map_col_to_list0[list][i];
948                 h->map_col_to_list0_field[list][2*i] = 2*j;
949                 h->map_col_to_list0_field[list][2*i+1] = 2*j+1;
950             }
951         }
952     }
953 }
954
955 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
956     MpegEncContext * const s = &h->s;
957     int b8_stride = h->b8_stride;
958     int b4_stride = h->b_stride;
959     int mb_xy = h->mb_xy;
960     int mb_type_col[2];
961     const int16_t (*l1mv0)[2], (*l1mv1)[2];
962     const int8_t *l1ref0, *l1ref1;
963     const int is_b8x8 = IS_8X8(*mb_type);
964     unsigned int sub_mb_type;
965     int i8, i4;
966
967 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
968
969     if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL
970         if(h->ref_list[1][0].reference == PICT_FRAME){   // AFL/AFR/FR/FL -> AFL
971             if(!IS_INTERLACED(*mb_type)){                //     AFR/FR    -> AFL
972                 int cur_poc = s->current_picture_ptr->poc;
973                 int *col_poc = h->ref_list[1]->field_poc;
974                 int col_parity = FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc);
975                 mb_xy= s->mb_x + ((s->mb_y&~1) + col_parity)*s->mb_stride;
976                 l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
977                 l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
978                 l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
979                 l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
980                 if(s->mb_y&1){
981                     l1ref0 +=   b8_stride;
982                     l1ref1 +=   b8_stride;
983                     l1mv0  += 2*b4_stride;
984                     l1mv1  += 2*b4_stride;
985                 }
986                 b8_stride = 0;
987             }
988         }else if(!(s->picture_structure & h->ref_list[1][0].reference)){// FL -> FL & differ parity
989             int fieldoff= 2*(h->ref_list[1][0].reference)-3;
990             mb_xy += s->mb_stride*fieldoff;
991         }
992         goto single_col;
993     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
994         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
995             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
996             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
997             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
998             b8_stride *= 3;
999             b4_stride *= 6;
1000             //FIXME IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag
1001             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
1002                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
1003                 && !is_b8x8){
1004                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1005                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
1006             }else{
1007                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1008                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1009             }
1010         }else{                                           //     AFR/FR    -> AFR/FR
1011 single_col:
1012             mb_type_col[0] =
1013             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
1014     if(IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag){
1015         /* FIXME save sub mb types from previous frames (or derive from MVs)
1016          * so we know exactly what block size to use */
1017         sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
1018         *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1019     }else if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
1020         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1021         *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
1022     }else{
1023         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1024         *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1025     }
1026         }
1027     }
1028
1029     if(b8_stride){
1030         l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
1031         l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
1032         l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
1033         l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
1034     }
1035
1036     if(h->direct_spatial_mv_pred){
1037         int ref[2];
1038         int mv[2][2];
1039         int list;
1040
1041         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
1042
1043         /* ref = min(neighbors) */
1044         for(list=0; list<2; list++){
1045             int refa = h->ref_cache[list][scan8[0] - 1];
1046             int refb = h->ref_cache[list][scan8[0] - 8];
1047             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1048             if(refc == PART_NOT_AVAILABLE)
1049                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1050             ref[list] = FFMIN3((unsigned)refa, (unsigned)refb, (unsigned)refc);
1051             if(ref[list] < 0)
1052                 ref[list] = -1;
1053         }
1054
1055         if(ref[0] < 0 && ref[1] < 0){
1056             ref[0] = ref[1] = 0;
1057             mv[0][0] = mv[0][1] =
1058             mv[1][0] = mv[1][1] = 0;
1059         }else{
1060             for(list=0; list<2; list++){
1061                 if(ref[list] >= 0)
1062                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1063                 else
1064                     mv[list][0] = mv[list][1] = 0;
1065             }
1066         }
1067
1068         if(ref[1] < 0){
1069             if(!is_b8x8)
1070                 *mb_type &= ~MB_TYPE_L1;
1071             sub_mb_type &= ~MB_TYPE_L1;
1072         }else if(ref[0] < 0){
1073             if(!is_b8x8)
1074                 *mb_type &= ~MB_TYPE_L0;
1075             sub_mb_type &= ~MB_TYPE_L0;
1076         }
1077
1078         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1079             for(i8=0; i8<4; i8++){
1080                 int x8 = i8&1;
1081                 int y8 = i8>>1;
1082                 int xy8 = x8+y8*b8_stride;
1083                 int xy4 = 3*x8+y8*b4_stride;
1084                 int a=0, b=0;
1085
1086                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1087                     continue;
1088                 h->sub_mb_type[i8] = sub_mb_type;
1089
1090                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1091                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1092                 if(!IS_INTRA(mb_type_col[y8])
1093                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
1094                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
1095                     if(ref[0] > 0)
1096                         a= pack16to32(mv[0][0],mv[0][1]);
1097                     if(ref[1] > 0)
1098                         b= pack16to32(mv[1][0],mv[1][1]);
1099                 }else{
1100                     a= pack16to32(mv[0][0],mv[0][1]);
1101                     b= pack16to32(mv[1][0],mv[1][1]);
1102                 }
1103                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
1104                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
1105             }
1106         }else if(IS_16X16(*mb_type)){
1107             int a=0, b=0;
1108
1109             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1110             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1111             if(!IS_INTRA(mb_type_col[0])
1112                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1113                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1114                        && (h->x264_build>33 || !h->x264_build)))){
1115                 if(ref[0] > 0)
1116                     a= pack16to32(mv[0][0],mv[0][1]);
1117                 if(ref[1] > 0)
1118                     b= pack16to32(mv[1][0],mv[1][1]);
1119             }else{
1120                 a= pack16to32(mv[0][0],mv[0][1]);
1121                 b= pack16to32(mv[1][0],mv[1][1]);
1122             }
1123             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1124             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1125         }else{
1126             for(i8=0; i8<4; i8++){
1127                 const int x8 = i8&1;
1128                 const int y8 = i8>>1;
1129
1130                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1131                     continue;
1132                 h->sub_mb_type[i8] = sub_mb_type;
1133
1134                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1135                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1136                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1137                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1138
1139                 /* col_zero_flag */
1140                 if(!IS_INTRA(mb_type_col[0]) && (   l1ref0[x8 + y8*h->b8_stride] == 0
1141                                               || (l1ref0[x8 + y8*h->b8_stride] < 0 && l1ref1[x8 + y8*h->b8_stride] == 0
1142                                                   && (h->x264_build>33 || !h->x264_build)))){
1143                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*h->b8_stride] == 0 ? l1mv0 : l1mv1;
1144                     if(IS_SUB_8X8(sub_mb_type)){
1145                         const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1146                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1147                             if(ref[0] == 0)
1148                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1149                             if(ref[1] == 0)
1150                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1151                         }
1152                     }else
1153                     for(i4=0; i4<4; i4++){
1154                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1155                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1156                             if(ref[0] == 0)
1157                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1158                             if(ref[1] == 0)
1159                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1160                         }
1161                     }
1162                 }
1163             }
1164         }
1165     }else{ /* direct temporal mv pred */
1166         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1167         const int *dist_scale_factor = h->dist_scale_factor;
1168
1169             if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
1170                 map_col_to_list0[0] = h->map_col_to_list0_field[0];
1171                 map_col_to_list0[1] = h->map_col_to_list0_field[1];
1172                 dist_scale_factor = h->dist_scale_factor_field;
1173             }
1174             if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1175                 /* FIXME assumes direct_8x8_inference == 1 */
1176                 int y_shift;
1177                 int ref_shift;
1178
1179                 if(IS_INTERLACED(*mb_type)){
1180                     /* frame to field scaling */
1181                     y_shift = 0;
1182                     ref_shift= FRAME_MBAFF ? 0 : 1;
1183                 }else{
1184                     y_shift = 2;
1185                     ref_shift= FRAME_MBAFF ? 2 : 1;
1186                 }
1187
1188                 for(i8=0; i8<4; i8++){
1189                     const int x8 = i8&1;
1190                     const int y8 = i8>>1;
1191                     int ref0, scale;
1192                     const int16_t (*l1mv)[2]= l1mv0;
1193
1194                     if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1195                         continue;
1196                     h->sub_mb_type[i8] = sub_mb_type;
1197
1198                     fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1199                     if(IS_INTRA(mb_type_col[y8])){
1200                         fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1201                         fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1202                         fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1203                         continue;
1204                     }
1205
1206                     ref0 = l1ref0[x8 + y8*b8_stride];
1207                     if(ref0 >= 0)
1208                         ref0 = map_col_to_list0[0][ref0*2>>ref_shift];
1209                     else{
1210                         ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride]*2>>ref_shift];
1211                         l1mv= l1mv1;
1212                     }
1213                     scale = dist_scale_factor[ref0];
1214                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1215
1216                     {
1217                         const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride];
1218                         int my_col = (mv_col[1]<<y_shift)/2;
1219                         int mx = (scale * mv_col[0] + 128) >> 8;
1220                         int my = (scale * my_col + 128) >> 8;
1221                         fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1222                         fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1223                     }
1224                 }
1225                 return;
1226             }
1227
1228         /* one-to-one mv scaling */
1229
1230         if(IS_16X16(*mb_type)){
1231             int ref, mv0, mv1;
1232
1233             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1234             if(IS_INTRA(mb_type_col[0])){
1235                 ref=mv0=mv1=0;
1236             }else{
1237                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0]]
1238                                                 : map_col_to_list0[1][l1ref1[0]];
1239                 const int scale = dist_scale_factor[ref0];
1240                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1241                 int mv_l0[2];
1242                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1243                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1244                 ref= ref0;
1245                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1246                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1247             }
1248             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1249             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1250             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1251         }else{
1252             for(i8=0; i8<4; i8++){
1253                 const int x8 = i8&1;
1254                 const int y8 = i8>>1;
1255                 int ref0, scale;
1256                 const int16_t (*l1mv)[2]= l1mv0;
1257
1258                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1259                     continue;
1260                 h->sub_mb_type[i8] = sub_mb_type;
1261                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1262                 if(IS_INTRA(mb_type_col[0])){
1263                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1264                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1265                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1266                     continue;
1267                 }
1268
1269                 ref0 = l1ref0[x8 + y8*h->b8_stride];
1270                 if(ref0 >= 0)
1271                     ref0 = map_col_to_list0[0][ref0];
1272                 else{
1273                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*h->b8_stride]];
1274                     l1mv= l1mv1;
1275                 }
1276                 scale = dist_scale_factor[ref0];
1277
1278                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1279                 if(IS_SUB_8X8(sub_mb_type)){
1280                     const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1281                     int mx = (scale * mv_col[0] + 128) >> 8;
1282                     int my = (scale * mv_col[1] + 128) >> 8;
1283                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1284                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1285                 }else
1286                 for(i4=0; i4<4; i4++){
1287                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1288                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1289                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1290                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1291                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1292                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1293                 }
1294             }
1295         }
1296     }
1297 }
1298
1299 static inline void write_back_motion(H264Context *h, int mb_type){
1300     MpegEncContext * const s = &h->s;
1301     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1302     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1303     int list;
1304
1305     if(!USES_LIST(mb_type, 0))
1306         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1307
1308     for(list=0; list<h->list_count; list++){
1309         int y;
1310         if(!USES_LIST(mb_type, list))
1311             continue;
1312
1313         for(y=0; y<4; y++){
1314             *(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];
1315             *(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];
1316         }
1317         if( h->pps.cabac ) {
1318             if(IS_SKIP(mb_type))
1319                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1320             else
1321             for(y=0; y<4; y++){
1322                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1323                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1324             }
1325         }
1326
1327         {
1328             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1329             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1330             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1331             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1332             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1333         }
1334     }
1335
1336     if(h->slice_type_nos == FF_B_TYPE && h->pps.cabac){
1337         if(IS_8X8(mb_type)){
1338             uint8_t *direct_table = &h->direct_table[b8_xy];
1339             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1340             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1341             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1342         }
1343     }
1344 }
1345
1346 /**
1347  * Decodes a network abstraction layer unit.
1348  * @param consumed is the number of bytes used as input
1349  * @param length is the length of the array
1350  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1351  * @returns decoded bytes, might be src+1 if no escapes
1352  */
1353 static const uint8_t *decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
1354     int i, si, di;
1355     uint8_t *dst;
1356     int bufidx;
1357
1358 //    src[0]&0x80;                //forbidden bit
1359     h->nal_ref_idc= src[0]>>5;
1360     h->nal_unit_type= src[0]&0x1F;
1361
1362     src++; length--;
1363 #if 0
1364     for(i=0; i<length; i++)
1365         printf("%2X ", src[i]);
1366 #endif
1367     for(i=0; i+1<length; i+=2){
1368         if(src[i]) continue;
1369         if(i>0 && src[i-1]==0) i--;
1370         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1371             if(src[i+2]!=3){
1372                 /* startcode, so we must be past the end */
1373                 length=i;
1374             }
1375             break;
1376         }
1377     }
1378
1379     if(i>=length-1){ //no escaped 0
1380         *dst_length= length;
1381         *consumed= length+1; //+1 for the header
1382         return src;
1383     }
1384
1385     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1386     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length);
1387     dst= h->rbsp_buffer[bufidx];
1388
1389     if (dst == NULL){
1390         return NULL;
1391     }
1392
1393 //printf("decoding esc\n");
1394     si=di=0;
1395     while(si<length){
1396         //remove escapes (very rare 1:2^22)
1397         if(si+2<length && src[si]==0 && src[si+1]==0 && src[si+2]<=3){
1398             if(src[si+2]==3){ //escape
1399                 dst[di++]= 0;
1400                 dst[di++]= 0;
1401                 si+=3;
1402                 continue;
1403             }else //next start code
1404                 break;
1405         }
1406
1407         dst[di++]= src[si++];
1408     }
1409
1410     *dst_length= di;
1411     *consumed= si + 1;//+1 for the header
1412 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1413     return dst;
1414 }
1415
1416 /**
1417  * identifies the exact end of the bitstream
1418  * @return the length of the trailing, or 0 if damaged
1419  */
1420 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src){
1421     int v= *src;
1422     int r;
1423
1424     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1425
1426     for(r=1; r<9; r++){
1427         if(v&1) return r;
1428         v>>=1;
1429     }
1430     return 0;
1431 }
1432
1433 /**
1434  * IDCT transforms the 16 dc values and dequantizes them.
1435  * @param qp quantization parameter
1436  */
1437 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1438 #define stride 16
1439     int i;
1440     int temp[16]; //FIXME check if this is a good idea
1441     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1442     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1443
1444 //memset(block, 64, 2*256);
1445 //return;
1446     for(i=0; i<4; i++){
1447         const int offset= y_offset[i];
1448         const int z0= block[offset+stride*0] + block[offset+stride*4];
1449         const int z1= block[offset+stride*0] - block[offset+stride*4];
1450         const int z2= block[offset+stride*1] - block[offset+stride*5];
1451         const int z3= block[offset+stride*1] + block[offset+stride*5];
1452
1453         temp[4*i+0]= z0+z3;
1454         temp[4*i+1]= z1+z2;
1455         temp[4*i+2]= z1-z2;
1456         temp[4*i+3]= z0-z3;
1457     }
1458
1459     for(i=0; i<4; i++){
1460         const int offset= x_offset[i];
1461         const int z0= temp[4*0+i] + temp[4*2+i];
1462         const int z1= temp[4*0+i] - temp[4*2+i];
1463         const int z2= temp[4*1+i] - temp[4*3+i];
1464         const int z3= temp[4*1+i] + temp[4*3+i];
1465
1466         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
1467         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1468         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1469         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1470     }
1471 }
1472
1473 #if 0
1474 /**
1475  * DCT transforms the 16 dc values.
1476  * @param qp quantization parameter ??? FIXME
1477  */
1478 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1479 //    const int qmul= dequant_coeff[qp][0];
1480     int i;
1481     int temp[16]; //FIXME check if this is a good idea
1482     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1483     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1484
1485     for(i=0; i<4; i++){
1486         const int offset= y_offset[i];
1487         const int z0= block[offset+stride*0] + block[offset+stride*4];
1488         const int z1= block[offset+stride*0] - block[offset+stride*4];
1489         const int z2= block[offset+stride*1] - block[offset+stride*5];
1490         const int z3= block[offset+stride*1] + block[offset+stride*5];
1491
1492         temp[4*i+0]= z0+z3;
1493         temp[4*i+1]= z1+z2;
1494         temp[4*i+2]= z1-z2;
1495         temp[4*i+3]= z0-z3;
1496     }
1497
1498     for(i=0; i<4; i++){
1499         const int offset= x_offset[i];
1500         const int z0= temp[4*0+i] + temp[4*2+i];
1501         const int z1= temp[4*0+i] - temp[4*2+i];
1502         const int z2= temp[4*1+i] - temp[4*3+i];
1503         const int z3= temp[4*1+i] + temp[4*3+i];
1504
1505         block[stride*0 +offset]= (z0 + z3)>>1;
1506         block[stride*2 +offset]= (z1 + z2)>>1;
1507         block[stride*8 +offset]= (z1 - z2)>>1;
1508         block[stride*10+offset]= (z0 - z3)>>1;
1509     }
1510 }
1511 #endif
1512
1513 #undef xStride
1514 #undef stride
1515
1516 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1517     const int stride= 16*2;
1518     const int xStride= 16;
1519     int a,b,c,d,e;
1520
1521     a= block[stride*0 + xStride*0];
1522     b= block[stride*0 + xStride*1];
1523     c= block[stride*1 + xStride*0];
1524     d= block[stride*1 + xStride*1];
1525
1526     e= a-b;
1527     a= a+b;
1528     b= c-d;
1529     c= c+d;
1530
1531     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1532     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1533     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1534     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1535 }
1536
1537 #if 0
1538 static void chroma_dc_dct_c(DCTELEM *block){
1539     const int stride= 16*2;
1540     const int xStride= 16;
1541     int a,b,c,d,e;
1542
1543     a= block[stride*0 + xStride*0];
1544     b= block[stride*0 + xStride*1];
1545     c= block[stride*1 + xStride*0];
1546     d= block[stride*1 + xStride*1];
1547
1548     e= a-b;
1549     a= a+b;
1550     b= c-d;
1551     c= c+d;
1552
1553     block[stride*0 + xStride*0]= (a+c);
1554     block[stride*0 + xStride*1]= (e+b);
1555     block[stride*1 + xStride*0]= (a-c);
1556     block[stride*1 + xStride*1]= (e-b);
1557 }
1558 #endif
1559
1560 /**
1561  * gets the chroma qp.
1562  */
1563 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1564     return h->pps.chroma_qp_table[t][qscale];
1565 }
1566
1567 //FIXME need to check that this does not overflow signed 32 bit for low qp, I am not sure, it's very close
1568 //FIXME check that gcc inlines this (and optimizes intra & separate_dc stuff away)
1569 static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int separate_dc){
1570     int i;
1571     const int * const quant_table= quant_coeff[qscale];
1572     const int bias= intra ? (1<<QUANT_SHIFT)/3 : (1<<QUANT_SHIFT)/6;
1573     const unsigned int threshold1= (1<<QUANT_SHIFT) - bias - 1;
1574     const unsigned int threshold2= (threshold1<<1);
1575     int last_non_zero;
1576
1577     if(separate_dc){
1578         if(qscale<=18){
1579             //avoid overflows
1580             const int dc_bias= intra ? (1<<(QUANT_SHIFT-2))/3 : (1<<(QUANT_SHIFT-2))/6;
1581             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT-2)) - dc_bias - 1;
1582             const unsigned int dc_threshold2= (dc_threshold1<<1);
1583
1584             int level= block[0]*quant_coeff[qscale+18][0];
1585             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1586                 if(level>0){
1587                     level= (dc_bias + level)>>(QUANT_SHIFT-2);
1588                     block[0]= level;
1589                 }else{
1590                     level= (dc_bias - level)>>(QUANT_SHIFT-2);
1591                     block[0]= -level;
1592                 }
1593 //                last_non_zero = i;
1594             }else{
1595                 block[0]=0;
1596             }
1597         }else{
1598             const int dc_bias= intra ? (1<<(QUANT_SHIFT+1))/3 : (1<<(QUANT_SHIFT+1))/6;
1599             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT+1)) - dc_bias - 1;
1600             const unsigned int dc_threshold2= (dc_threshold1<<1);
1601
1602             int level= block[0]*quant_table[0];
1603             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1604                 if(level>0){
1605                     level= (dc_bias + level)>>(QUANT_SHIFT+1);
1606                     block[0]= level;
1607                 }else{
1608                     level= (dc_bias - level)>>(QUANT_SHIFT+1);
1609                     block[0]= -level;
1610                 }
1611 //                last_non_zero = i;
1612             }else{
1613                 block[0]=0;
1614             }
1615         }
1616         last_non_zero= 0;
1617         i=1;
1618     }else{
1619         last_non_zero= -1;
1620         i=0;
1621     }
1622
1623     for(; i<16; i++){
1624         const int j= scantable[i];
1625         int level= block[j]*quant_table[j];
1626
1627 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
1628 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
1629         if(((unsigned)(level+threshold1))>threshold2){
1630             if(level>0){
1631                 level= (bias + level)>>QUANT_SHIFT;
1632                 block[j]= level;
1633             }else{
1634                 level= (bias - level)>>QUANT_SHIFT;
1635                 block[j]= -level;
1636             }
1637             last_non_zero = i;
1638         }else{
1639             block[j]=0;
1640         }
1641     }
1642
1643     return last_non_zero;
1644 }
1645
1646 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1647                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1648                            int src_x_offset, int src_y_offset,
1649                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1650     MpegEncContext * const s = &h->s;
1651     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1652     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1653     const int luma_xy= (mx&3) + ((my&3)<<2);
1654     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1655     uint8_t * src_cb, * src_cr;
1656     int extra_width= h->emu_edge_width;
1657     int extra_height= h->emu_edge_height;
1658     int emu=0;
1659     const int full_mx= mx>>2;
1660     const int full_my= my>>2;
1661     const int pic_width  = 16*s->mb_width;
1662     const int pic_height = 16*s->mb_height >> MB_FIELD;
1663
1664     if(!pic->data[0]) //FIXME this is unacceptable, some sensible error concealment must be done for missing reference frames
1665         return;
1666
1667     if(mx&7) extra_width -= 3;
1668     if(my&7) extra_height -= 3;
1669
1670     if(   full_mx < 0-extra_width
1671        || full_my < 0-extra_height
1672        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1673        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1674         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);
1675             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1676         emu=1;
1677     }
1678
1679     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1680     if(!square){
1681         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1682     }
1683
1684     if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1685
1686     if(MB_FIELD){
1687         // chroma offset when predicting from a field of opposite parity
1688         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1689         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1690     }
1691     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1692     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1693
1694     if(emu){
1695         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);
1696             src_cb= s->edge_emu_buffer;
1697     }
1698     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1699
1700     if(emu){
1701         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);
1702             src_cr= s->edge_emu_buffer;
1703     }
1704     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1705 }
1706
1707 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1708                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1709                            int x_offset, int y_offset,
1710                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1711                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1712                            int list0, int list1){
1713     MpegEncContext * const s = &h->s;
1714     qpel_mc_func *qpix_op=  qpix_put;
1715     h264_chroma_mc_func chroma_op= chroma_put;
1716
1717     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1718     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1719     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1720     x_offset += 8*s->mb_x;
1721     y_offset += 8*(s->mb_y >> MB_FIELD);
1722
1723     if(list0){
1724         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1725         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1726                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1727                            qpix_op, chroma_op);
1728
1729         qpix_op=  qpix_avg;
1730         chroma_op= chroma_avg;
1731     }
1732
1733     if(list1){
1734         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1735         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1736                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1737                            qpix_op, chroma_op);
1738     }
1739 }
1740
1741 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1742                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1743                            int x_offset, int y_offset,
1744                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1745                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1746                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1747                            int list0, int list1){
1748     MpegEncContext * const s = &h->s;
1749
1750     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1751     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1752     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1753     x_offset += 8*s->mb_x;
1754     y_offset += 8*(s->mb_y >> MB_FIELD);
1755
1756     if(list0 && list1){
1757         /* don't optimize for luma-only case, since B-frames usually
1758          * use implicit weights => chroma too. */
1759         uint8_t *tmp_cb = s->obmc_scratchpad;
1760         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1761         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1762         int refn0 = h->ref_cache[0][ scan8[n] ];
1763         int refn1 = h->ref_cache[1][ scan8[n] ];
1764
1765         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1766                     dest_y, dest_cb, dest_cr,
1767                     x_offset, y_offset, qpix_put, chroma_put);
1768         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1769                     tmp_y, tmp_cb, tmp_cr,
1770                     x_offset, y_offset, qpix_put, chroma_put);
1771
1772         if(h->use_weight == 2){
1773             int weight0 = h->implicit_weight[refn0][refn1];
1774             int weight1 = 64 - weight0;
1775             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1776             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1777             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1778         }else{
1779             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1780                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1781                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1782             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1783                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1784                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1785             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1786                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1787                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1788         }
1789     }else{
1790         int list = list1 ? 1 : 0;
1791         int refn = h->ref_cache[list][ scan8[n] ];
1792         Picture *ref= &h->ref_list[list][refn];
1793         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1794                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1795                     qpix_put, chroma_put);
1796
1797         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1798                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1799         if(h->use_weight_chroma){
1800             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1801                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1802             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1803                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1804         }
1805     }
1806 }
1807
1808 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1809                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1810                            int x_offset, int y_offset,
1811                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1812                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1813                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1814                            int list0, int list1){
1815     if((h->use_weight==2 && list0 && list1
1816         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1817        || h->use_weight==1)
1818         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1819                          x_offset, y_offset, qpix_put, chroma_put,
1820                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1821     else
1822         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1823                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1824 }
1825
1826 static inline void prefetch_motion(H264Context *h, int list){
1827     /* fetch pixels for estimated mv 4 macroblocks ahead
1828      * optimized for 64byte cache lines */
1829     MpegEncContext * const s = &h->s;
1830     const int refn = h->ref_cache[list][scan8[0]];
1831     if(refn >= 0){
1832         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1833         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1834         uint8_t **src= h->ref_list[list][refn].data;
1835         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1836         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1837         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1838         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1839     }
1840 }
1841
1842 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1843                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1844                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1845                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1846     MpegEncContext * const s = &h->s;
1847     const int mb_xy= h->mb_xy;
1848     const int mb_type= s->current_picture.mb_type[mb_xy];
1849
1850     assert(IS_INTER(mb_type));
1851
1852     prefetch_motion(h, 0);
1853
1854     if(IS_16X16(mb_type)){
1855         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1856                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1857                 &weight_op[0], &weight_avg[0],
1858                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1859     }else if(IS_16X8(mb_type)){
1860         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1861                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1862                 &weight_op[1], &weight_avg[1],
1863                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1864         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1865                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1866                 &weight_op[1], &weight_avg[1],
1867                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1868     }else if(IS_8X16(mb_type)){
1869         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1870                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1871                 &weight_op[2], &weight_avg[2],
1872                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1873         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1874                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1875                 &weight_op[2], &weight_avg[2],
1876                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1877     }else{
1878         int i;
1879
1880         assert(IS_8X8(mb_type));
1881
1882         for(i=0; i<4; i++){
1883             const int sub_mb_type= h->sub_mb_type[i];
1884             const int n= 4*i;
1885             int x_offset= (i&1)<<2;
1886             int y_offset= (i&2)<<1;
1887
1888             if(IS_SUB_8X8(sub_mb_type)){
1889                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1890                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1891                     &weight_op[3], &weight_avg[3],
1892                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1893             }else if(IS_SUB_8X4(sub_mb_type)){
1894                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1895                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1896                     &weight_op[4], &weight_avg[4],
1897                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1898                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1899                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1900                     &weight_op[4], &weight_avg[4],
1901                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1902             }else if(IS_SUB_4X8(sub_mb_type)){
1903                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1904                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1905                     &weight_op[5], &weight_avg[5],
1906                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1907                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1908                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1909                     &weight_op[5], &weight_avg[5],
1910                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1911             }else{
1912                 int j;
1913                 assert(IS_SUB_4X4(sub_mb_type));
1914                 for(j=0; j<4; j++){
1915                     int sub_x_offset= x_offset + 2*(j&1);
1916                     int sub_y_offset= y_offset +   (j&2);
1917                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1918                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1919                         &weight_op[6], &weight_avg[6],
1920                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1921                 }
1922             }
1923         }
1924     }
1925
1926     prefetch_motion(h, 1);
1927 }
1928
1929 static av_cold void decode_init_vlc(void){
1930     static int done = 0;
1931
1932     if (!done) {
1933         int i;
1934         int offset;
1935         done = 1;
1936
1937         chroma_dc_coeff_token_vlc.table = chroma_dc_coeff_token_vlc_table;
1938         chroma_dc_coeff_token_vlc.table_allocated = chroma_dc_coeff_token_vlc_table_size;
1939         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1940                  &chroma_dc_coeff_token_len [0], 1, 1,
1941                  &chroma_dc_coeff_token_bits[0], 1, 1,
1942                  INIT_VLC_USE_NEW_STATIC);
1943
1944         offset = 0;
1945         for(i=0; i<4; i++){
1946             coeff_token_vlc[i].table = coeff_token_vlc_tables+offset;
1947             coeff_token_vlc[i].table_allocated = coeff_token_vlc_tables_size[i];
1948             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1949                      &coeff_token_len [i][0], 1, 1,
1950                      &coeff_token_bits[i][0], 1, 1,
1951                      INIT_VLC_USE_NEW_STATIC);
1952             offset += coeff_token_vlc_tables_size[i];
1953         }
1954         /*
1955          * This is a one time safety check to make sure that
1956          * the packed static coeff_token_vlc table sizes
1957          * were initialized correctly.
1958          */
1959         assert(offset == sizeof(coeff_token_vlc_tables)/(sizeof(VLC_TYPE)*2));
1960
1961         for(i=0; i<3; i++){
1962             chroma_dc_total_zeros_vlc[i].table = chroma_dc_total_zeros_vlc_tables[i];
1963             chroma_dc_total_zeros_vlc[i].table_allocated = chroma_dc_total_zeros_vlc_tables_size;
1964             init_vlc(&chroma_dc_total_zeros_vlc[i],
1965                      CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1966                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1967                      &chroma_dc_total_zeros_bits[i][0], 1, 1,
1968                      INIT_VLC_USE_NEW_STATIC);
1969         }
1970         for(i=0; i<15; i++){
1971             total_zeros_vlc[i].table = total_zeros_vlc_tables[i];
1972             total_zeros_vlc[i].table_allocated = total_zeros_vlc_tables_size;
1973             init_vlc(&total_zeros_vlc[i],
1974                      TOTAL_ZEROS_VLC_BITS, 16,
1975                      &total_zeros_len [i][0], 1, 1,
1976                      &total_zeros_bits[i][0], 1, 1,
1977                      INIT_VLC_USE_NEW_STATIC);
1978         }
1979
1980         for(i=0; i<6; i++){
1981             run_vlc[i].table = run_vlc_tables[i];
1982             run_vlc[i].table_allocated = run_vlc_tables_size;
1983             init_vlc(&run_vlc[i],
1984                      RUN_VLC_BITS, 7,
1985                      &run_len [i][0], 1, 1,
1986                      &run_bits[i][0], 1, 1,
1987                      INIT_VLC_USE_NEW_STATIC);
1988         }
1989         run7_vlc.table = run7_vlc_table,
1990         run7_vlc.table_allocated = run7_vlc_table_size;
1991         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
1992                  &run_len [6][0], 1, 1,
1993                  &run_bits[6][0], 1, 1,
1994                  INIT_VLC_USE_NEW_STATIC);
1995     }
1996 }
1997
1998 static void free_tables(H264Context *h){
1999     int i;
2000     H264Context *hx;
2001     av_freep(&h->intra4x4_pred_mode);
2002     av_freep(&h->chroma_pred_mode_table);
2003     av_freep(&h->cbp_table);
2004     av_freep(&h->mvd_table[0]);
2005     av_freep(&h->mvd_table[1]);
2006     av_freep(&h->direct_table);
2007     av_freep(&h->non_zero_count);
2008     av_freep(&h->slice_table_base);
2009     h->slice_table= NULL;
2010
2011     av_freep(&h->mb2b_xy);
2012     av_freep(&h->mb2b8_xy);
2013
2014     for(i = 0; i < MAX_SPS_COUNT; i++)
2015         av_freep(h->sps_buffers + i);
2016
2017     for(i = 0; i < MAX_PPS_COUNT; i++)
2018         av_freep(h->pps_buffers + i);
2019
2020     for(i = 0; i < h->s.avctx->thread_count; i++) {
2021         hx = h->thread_context[i];
2022         if(!hx) continue;
2023         av_freep(&hx->top_borders[1]);
2024         av_freep(&hx->top_borders[0]);
2025         av_freep(&hx->s.obmc_scratchpad);
2026     }
2027 }
2028
2029 static void init_dequant8_coeff_table(H264Context *h){
2030     int i,q,x;
2031     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2032     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2033     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2034
2035     for(i=0; i<2; i++ ){
2036         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2037             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2038             break;
2039         }
2040
2041         for(q=0; q<52; q++){
2042             int shift = ff_div6[q];
2043             int idx = ff_rem6[q];
2044             for(x=0; x<64; x++)
2045                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2046                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2047                     h->pps.scaling_matrix8[i][x]) << shift;
2048         }
2049     }
2050 }
2051
2052 static void init_dequant4_coeff_table(H264Context *h){
2053     int i,j,q,x;
2054     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2055     for(i=0; i<6; i++ ){
2056         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2057         for(j=0; j<i; j++){
2058             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2059                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2060                 break;
2061             }
2062         }
2063         if(j<i)
2064             continue;
2065
2066         for(q=0; q<52; q++){
2067             int shift = ff_div6[q] + 2;
2068             int idx = ff_rem6[q];
2069             for(x=0; x<16; x++)
2070                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2071                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2072                     h->pps.scaling_matrix4[i][x]) << shift;
2073         }
2074     }
2075 }
2076
2077 static void init_dequant_tables(H264Context *h){
2078     int i,x;
2079     init_dequant4_coeff_table(h);
2080     if(h->pps.transform_8x8_mode)
2081         init_dequant8_coeff_table(h);
2082     if(h->sps.transform_bypass){
2083         for(i=0; i<6; i++)
2084             for(x=0; x<16; x++)
2085                 h->dequant4_coeff[i][0][x] = 1<<6;
2086         if(h->pps.transform_8x8_mode)
2087             for(i=0; i<2; i++)
2088                 for(x=0; x<64; x++)
2089                     h->dequant8_coeff[i][0][x] = 1<<6;
2090     }
2091 }
2092
2093
2094 /**
2095  * allocates tables.
2096  * needs width/height
2097  */
2098 static int alloc_tables(H264Context *h){
2099     MpegEncContext * const s = &h->s;
2100     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2101     int x,y;
2102
2103     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2104
2105     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2106     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(uint8_t))
2107     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2108
2109     CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2110     CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2111     CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2112     CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2113
2114     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(uint8_t));
2115     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2116
2117     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2118     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2119     for(y=0; y<s->mb_height; y++){
2120         for(x=0; x<s->mb_width; x++){
2121             const int mb_xy= x + y*s->mb_stride;
2122             const int b_xy = 4*x + 4*y*h->b_stride;
2123             const int b8_xy= 2*x + 2*y*h->b8_stride;
2124
2125             h->mb2b_xy [mb_xy]= b_xy;
2126             h->mb2b8_xy[mb_xy]= b8_xy;
2127         }
2128     }
2129
2130     s->obmc_scratchpad = NULL;
2131
2132     if(!h->dequant4_coeff[0])
2133         init_dequant_tables(h);
2134
2135     return 0;
2136 fail:
2137     free_tables(h);
2138     return -1;
2139 }
2140
2141 /**
2142  * Mimic alloc_tables(), but for every context thread.
2143  */
2144 static void clone_tables(H264Context *dst, H264Context *src){
2145     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2146     dst->non_zero_count           = src->non_zero_count;
2147     dst->slice_table              = src->slice_table;
2148     dst->cbp_table                = src->cbp_table;
2149     dst->mb2b_xy                  = src->mb2b_xy;
2150     dst->mb2b8_xy                 = src->mb2b8_xy;
2151     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2152     dst->mvd_table[0]             = src->mvd_table[0];
2153     dst->mvd_table[1]             = src->mvd_table[1];
2154     dst->direct_table             = src->direct_table;
2155
2156     dst->s.obmc_scratchpad = NULL;
2157     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2158 }
2159
2160 /**
2161  * Init context
2162  * Allocate buffers which are not shared amongst multiple threads.
2163  */
2164 static int context_init(H264Context *h){
2165     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2166     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2167
2168     return 0;
2169 fail:
2170     return -1; // free_tables will clean up for us
2171 }
2172
2173 static av_cold void common_init(H264Context *h){
2174     MpegEncContext * const s = &h->s;
2175
2176     s->width = s->avctx->width;
2177     s->height = s->avctx->height;
2178     s->codec_id= s->avctx->codec->id;
2179
2180     ff_h264_pred_init(&h->hpc, s->codec_id);
2181
2182     h->dequant_coeff_pps= -1;
2183     s->unrestricted_mv=1;
2184     s->decode=1; //FIXME
2185
2186     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2187     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2188 }
2189
2190 static av_cold int decode_init(AVCodecContext *avctx){
2191     H264Context *h= avctx->priv_data;
2192     MpegEncContext * const s = &h->s;
2193
2194     MPV_decode_defaults(s);
2195
2196     s->avctx = avctx;
2197     common_init(h);
2198
2199     s->out_format = FMT_H264;
2200     s->workaround_bugs= avctx->workaround_bugs;
2201
2202     // set defaults
2203 //    s->decode_mb= ff_h263_decode_mb;
2204     s->quarter_sample = 1;
2205     s->low_delay= 1;
2206
2207     if(avctx->codec_id == CODEC_ID_SVQ3)
2208         avctx->pix_fmt= PIX_FMT_YUVJ420P;
2209     else
2210         avctx->pix_fmt= PIX_FMT_YUV420P;
2211
2212     decode_init_vlc();
2213
2214     if(avctx->extradata_size > 0 && avctx->extradata &&
2215        *(char *)avctx->extradata == 1){
2216         h->is_avc = 1;
2217         h->got_avcC = 0;
2218     } else {
2219         h->is_avc = 0;
2220     }
2221
2222     h->thread_context[0] = h;
2223     h->outputed_poc = INT_MIN;
2224     return 0;
2225 }
2226
2227 static int frame_start(H264Context *h){
2228     MpegEncContext * const s = &h->s;
2229     int i;
2230
2231     if(MPV_frame_start(s, s->avctx) < 0)
2232         return -1;
2233     ff_er_frame_start(s);
2234     /*
2235      * MPV_frame_start uses pict_type to derive key_frame.
2236      * This is incorrect for H.264; IDR markings must be used.
2237      * Zero here; IDR markings per slice in frame or fields are ORed in later.
2238      * See decode_nal_units().
2239      */
2240     s->current_picture_ptr->key_frame= 0;
2241
2242     assert(s->linesize && s->uvlinesize);
2243
2244     for(i=0; i<16; i++){
2245         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2246         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2247     }
2248     for(i=0; i<4; i++){
2249         h->block_offset[16+i]=
2250         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2251         h->block_offset[24+16+i]=
2252         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2253     }
2254
2255     /* can't be in alloc_tables because linesize isn't known there.
2256      * FIXME: redo bipred weight to not require extra buffer? */
2257     for(i = 0; i < s->avctx->thread_count; i++)
2258         if(!h->thread_context[i]->s.obmc_scratchpad)
2259             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2260
2261     /* some macroblocks will be accessed before they're available */
2262     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2263         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(uint8_t));
2264
2265 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2266
2267     // We mark the current picture as non-reference after allocating it, so
2268     // that if we break out due to an error it can be released automatically
2269     // in the next MPV_frame_start().
2270     // SVQ3 as well as most other codecs have only last/next/current and thus
2271     // get released even with set reference, besides SVQ3 and others do not
2272     // mark frames as reference later "naturally".
2273     if(s->codec_id != CODEC_ID_SVQ3)
2274         s->current_picture_ptr->reference= 0;
2275
2276     s->current_picture_ptr->field_poc[0]=
2277     s->current_picture_ptr->field_poc[1]= INT_MAX;
2278     assert(s->current_picture_ptr->long_ref==0);
2279
2280     return 0;
2281 }
2282
2283 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){
2284     MpegEncContext * const s = &h->s;
2285     int i;
2286
2287     src_y  -=   linesize;
2288     src_cb -= uvlinesize;
2289     src_cr -= uvlinesize;
2290
2291     // There are two lines saved, the line above the the top macroblock of a pair,
2292     // and the line above the bottom macroblock
2293     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2294     for(i=1; i<17; i++){
2295         h->left_border[i]= src_y[15+i*  linesize];
2296     }
2297
2298     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2299     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2300
2301     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2302         h->left_border[17  ]= h->top_borders[0][s->mb_x][16+7];
2303         h->left_border[17+9]= h->top_borders[0][s->mb_x][24+7];
2304         for(i=1; i<9; i++){
2305             h->left_border[i+17  ]= src_cb[7+i*uvlinesize];
2306             h->left_border[i+17+9]= src_cr[7+i*uvlinesize];
2307         }
2308         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2309         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2310     }
2311 }
2312
2313 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){
2314     MpegEncContext * const s = &h->s;
2315     int temp8, i;
2316     uint64_t temp64;
2317     int deblock_left;
2318     int deblock_top;
2319     int mb_xy;
2320
2321     if(h->deblocking_filter == 2) {
2322         mb_xy = h->mb_xy;
2323         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2324         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2325     } else {
2326         deblock_left = (s->mb_x > 0);
2327         deblock_top =  (s->mb_y > 0);
2328     }
2329
2330     src_y  -=   linesize + 1;
2331     src_cb -= uvlinesize + 1;
2332     src_cr -= uvlinesize + 1;
2333
2334 #define XCHG(a,b,t,xchg)\
2335 t= a;\
2336 if(xchg)\
2337     a= b;\
2338 b= t;
2339
2340     if(deblock_left){
2341         for(i = !deblock_top; i<17; i++){
2342             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2343         }
2344     }
2345
2346     if(deblock_top){
2347         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2348         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2349         if(s->mb_x+1 < s->mb_width){
2350             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2351         }
2352     }
2353
2354     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2355         if(deblock_left){
2356             for(i = !deblock_top; i<9; i++){
2357                 XCHG(h->left_border[i+17  ], src_cb[i*uvlinesize], temp8, xchg);
2358                 XCHG(h->left_border[i+17+9], src_cr[i*uvlinesize], temp8, xchg);
2359             }
2360         }
2361         if(deblock_top){
2362             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2363             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2364         }
2365     }
2366 }
2367
2368 static inline void backup_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize){
2369     MpegEncContext * const s = &h->s;
2370     int i;
2371
2372     src_y  -= 2 *   linesize;
2373     src_cb -= 2 * uvlinesize;
2374     src_cr -= 2 * uvlinesize;
2375
2376     // There are two lines saved, the line above the the top macroblock of a pair,
2377     // and the line above the bottom macroblock
2378     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2379     h->left_border[1]= h->top_borders[1][s->mb_x][15];
2380     for(i=2; i<34; i++){
2381         h->left_border[i]= src_y[15+i*  linesize];
2382     }
2383
2384     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  32*linesize);
2385     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+32*linesize);
2386     *(uint64_t*)(h->top_borders[1][s->mb_x]+0)= *(uint64_t*)(src_y +  33*linesize);
2387     *(uint64_t*)(h->top_borders[1][s->mb_x]+8)= *(uint64_t*)(src_y +8+33*linesize);
2388
2389     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2390         h->left_border[34     ]= h->top_borders[0][s->mb_x][16+7];
2391         h->left_border[34+   1]= h->top_borders[1][s->mb_x][16+7];
2392         h->left_border[34+18  ]= h->top_borders[0][s->mb_x][24+7];
2393         h->left_border[34+18+1]= h->top_borders[1][s->mb_x][24+7];
2394         for(i=2; i<18; i++){
2395             h->left_border[i+34   ]= src_cb[7+i*uvlinesize];
2396             h->left_border[i+34+18]= src_cr[7+i*uvlinesize];
2397         }
2398         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+16*uvlinesize);
2399         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+16*uvlinesize);
2400         *(uint64_t*)(h->top_borders[1][s->mb_x]+16)= *(uint64_t*)(src_cb+17*uvlinesize);
2401         *(uint64_t*)(h->top_borders[1][s->mb_x]+24)= *(uint64_t*)(src_cr+17*uvlinesize);
2402     }
2403 }
2404
2405 static inline void xchg_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int xchg){
2406     MpegEncContext * const s = &h->s;
2407     int temp8, i;
2408     uint64_t temp64;
2409     int deblock_left = (s->mb_x > 0);
2410     int deblock_top  = (s->mb_y > 1);
2411
2412     tprintf(s->avctx, "xchg_pair_border: src_y:%p src_cb:%p src_cr:%p ls:%d uvls:%d\n", src_y, src_cb, src_cr, linesize, uvlinesize);
2413
2414     src_y  -= 2 *   linesize + 1;
2415     src_cb -= 2 * uvlinesize + 1;
2416     src_cr -= 2 * uvlinesize + 1;
2417
2418 #define XCHG(a,b,t,xchg)\
2419 t= a;\
2420 if(xchg)\
2421     a= b;\
2422 b= t;
2423
2424     if(deblock_left){
2425         for(i = (!deblock_top)<<1; i<34; i++){
2426             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2427         }
2428     }
2429
2430     if(deblock_top){
2431         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2432         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2433         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+0), *(uint64_t*)(src_y +1 +linesize), temp64, xchg);
2434         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+8), *(uint64_t*)(src_y +9 +linesize), temp64, 1);
2435         if(s->mb_x+1 < s->mb_width){
2436             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2437             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x+1]), *(uint64_t*)(src_y +17 +linesize), temp64, 1);
2438         }
2439     }
2440
2441     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2442         if(deblock_left){
2443             for(i = (!deblock_top) << 1; i<18; i++){
2444                 XCHG(h->left_border[i+34   ], src_cb[i*uvlinesize], temp8, xchg);
2445                 XCHG(h->left_border[i+34+18], src_cr[i*uvlinesize], temp8, xchg);
2446             }
2447         }
2448         if(deblock_top){
2449             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2450             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2451             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+16), *(uint64_t*)(src_cb+1 +uvlinesize), temp64, 1);
2452             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+24), *(uint64_t*)(src_cr+1 +uvlinesize), temp64, 1);
2453         }
2454     }
2455 }
2456
2457 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2458     MpegEncContext * const s = &h->s;
2459     const int mb_x= s->mb_x;
2460     const int mb_y= s->mb_y;
2461     const int mb_xy= h->mb_xy;
2462     const int mb_type= s->current_picture.mb_type[mb_xy];
2463     uint8_t  *dest_y, *dest_cb, *dest_cr;
2464     int linesize, uvlinesize /*dct_offset*/;
2465     int i;
2466     int *block_offset = &h->block_offset[0];
2467     const unsigned int bottom = mb_y & 1;
2468     const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass), is_h264 = (simple || s->codec_id == CODEC_ID_H264);
2469     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2470     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2471
2472     dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2473     dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2474     dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2475
2476     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2477     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2478
2479     if (!simple && MB_FIELD) {
2480         linesize   = h->mb_linesize   = s->linesize * 2;
2481         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2482         block_offset = &h->block_offset[24];
2483         if(mb_y&1){ //FIXME move out of this function?
2484             dest_y -= s->linesize*15;
2485             dest_cb-= s->uvlinesize*7;
2486             dest_cr-= s->uvlinesize*7;
2487         }
2488         if(FRAME_MBAFF) {
2489             int list;
2490             for(list=0; list<h->list_count; list++){
2491                 if(!USES_LIST(mb_type, list))
2492                     continue;
2493                 if(IS_16X16(mb_type)){
2494                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2495                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2496                 }else{
2497                     for(i=0; i<16; i+=4){
2498                         //FIXME can refs be smaller than 8x8 when !direct_8x8_inference ?
2499                         int ref = h->ref_cache[list][scan8[i]];
2500                         if(ref >= 0)
2501                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2502                     }
2503                 }
2504             }
2505         }
2506     } else {
2507         linesize   = h->mb_linesize   = s->linesize;
2508         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2509 //        dct_offset = s->linesize * 16;
2510     }
2511
2512     if(transform_bypass){
2513         idct_dc_add =
2514         idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2515     }else if(IS_8x8DCT(mb_type)){
2516         idct_dc_add = s->dsp.h264_idct8_dc_add;
2517         idct_add = s->dsp.h264_idct8_add;
2518     }else{
2519         idct_dc_add = s->dsp.h264_idct_dc_add;
2520         idct_add = s->dsp.h264_idct_add;
2521     }
2522
2523     if(!simple && FRAME_MBAFF && h->deblocking_filter && IS_INTRA(mb_type)
2524        && (!bottom || !IS_INTRA(s->current_picture.mb_type[mb_xy-s->mb_stride]))){
2525         int mbt_y = mb_y&~1;
2526         uint8_t *top_y  = s->current_picture.data[0] + (mbt_y * 16* s->linesize  ) + mb_x * 16;
2527         uint8_t *top_cb = s->current_picture.data[1] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2528         uint8_t *top_cr = s->current_picture.data[2] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2529         xchg_pair_border(h, top_y, top_cb, top_cr, s->linesize, s->uvlinesize, 1);
2530     }
2531
2532     if (!simple && IS_INTRA_PCM(mb_type)) {
2533         for (i=0; i<16; i++) {
2534             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
2535         }
2536         for (i=0; i<8; i++) {
2537             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
2538             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
2539         }
2540     } else {
2541         if(IS_INTRA(mb_type)){
2542             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2543                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2544
2545             if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2546                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2547                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2548             }
2549
2550             if(IS_INTRA4x4(mb_type)){
2551                 if(simple || !s->encoding){
2552                     if(IS_8x8DCT(mb_type)){
2553                         for(i=0; i<16; i+=4){
2554                             uint8_t * const ptr= dest_y + block_offset[i];
2555                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2556                             const int nnz = h->non_zero_count_cache[ scan8[i] ];
2557                             h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2558                                                    (h->topright_samples_available<<i)&0x4000, linesize);
2559                             if(nnz){
2560                                 if(nnz == 1 && h->mb[i*16])
2561                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2562                                 else
2563                                     idct_add(ptr, h->mb + i*16, linesize);
2564                             }
2565                         }
2566                     }else
2567                     for(i=0; i<16; i++){
2568                         uint8_t * const ptr= dest_y + block_offset[i];
2569                         uint8_t *topright;
2570                         const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2571                         int nnz, tr;
2572
2573                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2574                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2575                             assert(mb_y || linesize <= block_offset[i]);
2576                             if(!topright_avail){
2577                                 tr= ptr[3 - linesize]*0x01010101;
2578                                 topright= (uint8_t*) &tr;
2579                             }else
2580                                 topright= ptr + 4 - linesize;
2581                         }else
2582                             topright= NULL;
2583
2584                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2585                         nnz = h->non_zero_count_cache[ scan8[i] ];
2586                         if(nnz){
2587                             if(is_h264){
2588                                 if(nnz == 1 && h->mb[i*16])
2589                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2590                                 else
2591                                     idct_add(ptr, h->mb + i*16, linesize);
2592                             }else
2593                                 svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2594                         }
2595                     }
2596                 }
2597             }else{
2598                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2599                 if(is_h264){
2600                     if(!transform_bypass)
2601                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2602                 }else
2603                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2604             }
2605             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2606                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2607         }else if(is_h264){
2608             hl_motion(h, dest_y, dest_cb, dest_cr,
2609                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2610                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2611                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2612         }
2613
2614
2615         if(!IS_INTRA4x4(mb_type)){
2616             if(is_h264){
2617                 if(IS_INTRA16x16(mb_type)){
2618                     for(i=0; i<16; i++){
2619                         if(h->non_zero_count_cache[ scan8[i] ])
2620                             idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2621                         else if(h->mb[i*16])
2622                             idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2623                     }
2624                 }else{
2625                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2626                     for(i=0; i<16; i+=di){
2627                         int nnz = h->non_zero_count_cache[ scan8[i] ];
2628                         if(nnz){
2629                             if(nnz==1 && h->mb[i*16])
2630                                 idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2631                             else
2632                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2633                         }
2634                     }
2635                 }
2636             }else{
2637                 for(i=0; i<16; i++){
2638                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2639                         uint8_t * const ptr= dest_y + block_offset[i];
2640                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2641                     }
2642                 }
2643             }
2644         }
2645
2646         if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2647             uint8_t *dest[2] = {dest_cb, dest_cr};
2648             if(transform_bypass){
2649                 idct_add = idct_dc_add = s->dsp.add_pixels4;
2650             }else{
2651                 idct_add = s->dsp.h264_idct_add;
2652                 idct_dc_add = s->dsp.h264_idct_dc_add;
2653                 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]);
2654                 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]);
2655             }
2656             if(is_h264){
2657                 for(i=16; i<16+8; i++){
2658                     if(h->non_zero_count_cache[ scan8[i] ])
2659                         idct_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2660                     else if(h->mb[i*16])
2661                         idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2662                 }
2663             }else{
2664                 for(i=16; i<16+8; i++){
2665                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2666                         uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2667                         svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2668                     }
2669                 }
2670             }
2671         }
2672     }
2673     if(h->deblocking_filter) {
2674         if (!simple && FRAME_MBAFF) {
2675             //FIXME try deblocking one mb at a time?
2676             // the reduction in load/storing mvs and such might outweigh the extra backup/xchg_border
2677             const int mb_y = s->mb_y - 1;
2678             uint8_t  *pair_dest_y, *pair_dest_cb, *pair_dest_cr;
2679             const int mb_xy= mb_x + mb_y*s->mb_stride;
2680             const int mb_type_top   = s->current_picture.mb_type[mb_xy];
2681             const int mb_type_bottom= s->current_picture.mb_type[mb_xy+s->mb_stride];
2682             if (!bottom) return;
2683             pair_dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2684             pair_dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2685             pair_dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2686
2687             if(IS_INTRA(mb_type_top | mb_type_bottom))
2688                 xchg_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize, 0);
2689
2690             backup_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize);
2691             // deblock a pair
2692             // top
2693             s->mb_y--; h->mb_xy -= s->mb_stride;
2694             tprintf(h->s.avctx, "call mbaff filter_mb mb_x:%d mb_y:%d pair_dest_y = %p, dest_y = %p\n", mb_x, mb_y, pair_dest_y, dest_y);
2695             fill_caches(h, mb_type_top, 1); //FIXME don't fill stuff which isn't used by filter_mb
2696             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2697             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2698             filter_mb(h, mb_x, mb_y, pair_dest_y, pair_dest_cb, pair_dest_cr, linesize, uvlinesize);
2699             // bottom
2700             s->mb_y++; h->mb_xy += s->mb_stride;
2701             tprintf(h->s.avctx, "call mbaff filter_mb\n");
2702             fill_caches(h, mb_type_bottom, 1); //FIXME don't fill stuff which isn't used by filter_mb
2703             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2704             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2705             filter_mb(h, mb_x, mb_y+1, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2706         } else {
2707             tprintf(h->s.avctx, "call filter_mb\n");
2708             backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2709             fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2710             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2711             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2712             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2713         }
2714     }
2715 }
2716
2717 /**
2718  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2719  */
2720 static void hl_decode_mb_simple(H264Context *h){
2721     hl_decode_mb_internal(h, 1);
2722 }
2723
2724 /**
2725  * Process a macroblock; this handles edge cases, such as interlacing.
2726  */
2727 static void av_noinline hl_decode_mb_complex(H264Context *h){
2728     hl_decode_mb_internal(h, 0);
2729 }
2730
2731 static void hl_decode_mb(H264Context *h){
2732     MpegEncContext * const s = &h->s;
2733     const int mb_xy= h->mb_xy;
2734     const int mb_type= s->current_picture.mb_type[mb_xy];
2735     int is_complex = FRAME_MBAFF || MB_FIELD || IS_INTRA_PCM(mb_type) || s->codec_id != CODEC_ID_H264 ||
2736                     (ENABLE_GRAY && (s->flags&CODEC_FLAG_GRAY)) || (ENABLE_H264_ENCODER && s->encoding) || ENABLE_SMALL;
2737
2738     if(ENABLE_H264_ENCODER && !s->decode)
2739         return;
2740
2741     if (is_complex)
2742         hl_decode_mb_complex(h);
2743     else hl_decode_mb_simple(h);
2744 }
2745
2746 static void pic_as_field(Picture *pic, const int parity){
2747     int i;
2748     for (i = 0; i < 4; ++i) {
2749         if (parity == PICT_BOTTOM_FIELD)
2750             pic->data[i] += pic->linesize[i];
2751         pic->reference = parity;
2752         pic->linesize[i] *= 2;
2753     }
2754     pic->poc= pic->field_poc[parity == PICT_BOTTOM_FIELD];
2755 }
2756
2757 static int split_field_copy(Picture *dest, Picture *src,
2758                             int parity, int id_add){
2759     int match = !!(src->reference & parity);
2760
2761     if (match) {
2762         *dest = *src;
2763         if(parity != PICT_FRAME){
2764             pic_as_field(dest, parity);
2765             dest->pic_id *= 2;
2766             dest->pic_id += id_add;
2767         }
2768     }
2769
2770     return match;
2771 }
2772
2773 static int build_def_list(Picture *def, Picture **in, int len, int is_long, int sel){
2774     int i[2]={0};
2775     int index=0;
2776
2777     while(i[0]<len || i[1]<len){
2778         while(i[0]<len && !(in[ i[0] ] && (in[ i[0] ]->reference & sel)))
2779             i[0]++;
2780         while(i[1]<len && !(in[ i[1] ] && (in[ i[1] ]->reference & (sel^3))))
2781             i[1]++;
2782         if(i[0] < len){
2783             in[ i[0] ]->pic_id= is_long ? i[0] : in[ i[0] ]->frame_num;
2784             split_field_copy(&def[index++], in[ i[0]++ ], sel  , 1);
2785         }
2786         if(i[1] < len){
2787             in[ i[1] ]->pic_id= is_long ? i[1] : in[ i[1] ]->frame_num;
2788             split_field_copy(&def[index++], in[ i[1]++ ], sel^3, 0);
2789         }
2790     }
2791
2792     return index;
2793 }
2794
2795 static int add_sorted(Picture **sorted, Picture **src, int len, int limit, int dir){
2796     int i, best_poc;
2797     int out_i= 0;
2798
2799     for(;;){
2800         best_poc= dir ? INT_MIN : INT_MAX;
2801
2802         for(i=0; i<len; i++){
2803             const int poc= src[i]->poc;
2804             if(((poc > limit) ^ dir) && ((poc < best_poc) ^ dir)){
2805                 best_poc= poc;
2806                 sorted[out_i]= src[i];
2807             }
2808         }
2809         if(best_poc == (dir ? INT_MIN : INT_MAX))
2810             break;
2811         limit= sorted[out_i++]->poc - dir;
2812     }
2813     return out_i;
2814 }
2815
2816 /**
2817  * fills the default_ref_list.
2818  */
2819 static int fill_default_ref_list(H264Context *h){
2820     MpegEncContext * const s = &h->s;
2821     int i, len;
2822
2823     if(h->slice_type_nos==FF_B_TYPE){
2824         Picture *sorted[32];
2825         int cur_poc, list;
2826         int lens[2];
2827
2828         if(FIELD_PICTURE)
2829             cur_poc= s->current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
2830         else
2831             cur_poc= s->current_picture_ptr->poc;
2832
2833         for(list= 0; list<2; list++){
2834             len= add_sorted(sorted    , h->short_ref, h->short_ref_count, cur_poc, 1^list);
2835             len+=add_sorted(sorted+len, h->short_ref, h->short_ref_count, cur_poc, 0^list);
2836             assert(len<=32);
2837             len= build_def_list(h->default_ref_list[list]    , sorted     , len, 0, s->picture_structure);
2838             len+=build_def_list(h->default_ref_list[list]+len, h->long_ref, 16 , 1, s->picture_structure);
2839             assert(len<=32);
2840
2841             if(len < h->ref_count[list])
2842                 memset(&h->default_ref_list[list][len], 0, sizeof(Picture)*(h->ref_count[list] - len));
2843             lens[list]= len;
2844         }
2845
2846         if(lens[0] == lens[1] && lens[1] > 1){
2847             for(i=0; h->default_ref_list[0][i].data[0] == h->default_ref_list[1][i].data[0] && i<lens[0]; i++);
2848             if(i == lens[0])
2849                 FFSWAP(Picture, h->default_ref_list[1][0], h->default_ref_list[1][1]);
2850         }
2851     }else{
2852         len = build_def_list(h->default_ref_list[0]    , h->short_ref, h->short_ref_count, 0, s->picture_structure);
2853         len+= build_def_list(h->default_ref_list[0]+len, h-> long_ref, 16                , 1, s->picture_structure);
2854         assert(len <= 32);
2855         if(len < h->ref_count[0])
2856             memset(&h->default_ref_list[0][len], 0, sizeof(Picture)*(h->ref_count[0] - len));
2857     }
2858 #ifdef TRACE
2859     for (i=0; i<h->ref_count[0]; i++) {
2860         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]);
2861     }
2862     if(h->slice_type_nos==FF_B_TYPE){
2863         for (i=0; i<h->ref_count[1]; i++) {
2864             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]);
2865         }
2866     }
2867 #endif
2868     return 0;
2869 }
2870
2871 static void print_short_term(H264Context *h);
2872 static void print_long_term(H264Context *h);
2873
2874 /**
2875  * Extract structure information about the picture described by pic_num in
2876  * the current decoding context (frame or field). Note that pic_num is
2877  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
2878  * @param pic_num picture number for which to extract structure information
2879  * @param structure one of PICT_XXX describing structure of picture
2880  *                      with pic_num
2881  * @return frame number (short term) or long term index of picture
2882  *         described by pic_num
2883  */
2884 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
2885     MpegEncContext * const s = &h->s;
2886
2887     *structure = s->picture_structure;
2888     if(FIELD_PICTURE){
2889         if (!(pic_num & 1))
2890             /* opposite field */
2891             *structure ^= PICT_FRAME;
2892         pic_num >>= 1;
2893     }
2894
2895     return pic_num;
2896 }
2897
2898 static int decode_ref_pic_list_reordering(H264Context *h){
2899     MpegEncContext * const s = &h->s;
2900     int list, index, pic_structure;
2901
2902     print_short_term(h);
2903     print_long_term(h);
2904
2905     for(list=0; list<h->list_count; list++){
2906         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
2907
2908         if(get_bits1(&s->gb)){
2909             int pred= h->curr_pic_num;
2910
2911             for(index=0; ; index++){
2912                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb(&s->gb);
2913                 unsigned int pic_id;
2914                 int i;
2915                 Picture *ref = NULL;
2916
2917                 if(reordering_of_pic_nums_idc==3)
2918                     break;
2919
2920                 if(index >= h->ref_count[list]){
2921                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
2922                     return -1;
2923                 }
2924
2925                 if(reordering_of_pic_nums_idc<3){
2926                     if(reordering_of_pic_nums_idc<2){
2927                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
2928                         int frame_num;
2929
2930                         if(abs_diff_pic_num > h->max_pic_num){
2931                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
2932                             return -1;
2933                         }
2934
2935                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
2936                         else                                pred+= abs_diff_pic_num;
2937                         pred &= h->max_pic_num - 1;
2938
2939                         frame_num = pic_num_extract(h, pred, &pic_structure);
2940
2941                         for(i= h->short_ref_count-1; i>=0; i--){
2942                             ref = h->short_ref[i];
2943                             assert(ref->reference);
2944                             assert(!ref->long_ref);
2945                             if(
2946                                    ref->frame_num == frame_num &&
2947                                    (ref->reference & pic_structure)
2948                               )
2949                                 break;
2950                         }
2951                         if(i>=0)
2952                             ref->pic_id= pred;
2953                     }else{
2954                         int long_idx;
2955                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
2956
2957                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
2958
2959                         if(long_idx>31){
2960                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
2961                             return -1;
2962                         }
2963                         ref = h->long_ref[long_idx];
2964                         assert(!(ref && !ref->reference));
2965                         if(ref && (ref->reference & pic_structure)){
2966                             ref->pic_id= pic_id;
2967                             assert(ref->long_ref);
2968                             i=0;
2969                         }else{
2970                             i=-1;
2971                         }
2972                     }
2973
2974                     if (i < 0) {
2975                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
2976                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
2977                     } else {
2978                         for(i=index; i+1<h->ref_count[list]; i++){
2979                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
2980                                 break;
2981                         }
2982                         for(; i > index; i--){
2983                             h->ref_list[list][i]= h->ref_list[list][i-1];
2984                         }
2985                         h->ref_list[list][index]= *ref;
2986                         if (FIELD_PICTURE){
2987                             pic_as_field(&h->ref_list[list][index], pic_structure);
2988                         }
2989                     }
2990                 }else{
2991                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
2992                     return -1;
2993                 }
2994             }
2995         }
2996     }
2997     for(list=0; list<h->list_count; list++){
2998         for(index= 0; index < h->ref_count[list]; index++){
2999             if(!h->ref_list[list][index].data[0]){
3000                 av_log(h->s.avctx, AV_LOG_ERROR, "Missing reference picture\n");
3001                 h->ref_list[list][index]= s->current_picture; //FIXME this is not a sensible solution
3002             }
3003         }
3004     }
3005
3006     if(h->slice_type_nos==FF_B_TYPE && !h->direct_spatial_mv_pred)
3007         direct_dist_scale_factor(h);
3008     direct_ref_list_init(h);
3009     return 0;
3010 }
3011
3012 static void fill_mbaff_ref_list(H264Context *h){
3013     int list, i, j;
3014     for(list=0; list<2; list++){ //FIXME try list_count
3015         for(i=0; i<h->ref_count[list]; i++){
3016             Picture *frame = &h->ref_list[list][i];
3017             Picture *field = &h->ref_list[list][16+2*i];
3018             field[0] = *frame;
3019             for(j=0; j<3; j++)
3020                 field[0].linesize[j] <<= 1;
3021             field[0].reference = PICT_TOP_FIELD;
3022             field[1] = field[0];
3023             for(j=0; j<3; j++)
3024                 field[1].data[j] += frame->linesize[j];
3025             field[1].reference = PICT_BOTTOM_FIELD;
3026
3027             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
3028             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
3029             for(j=0; j<2; j++){
3030                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
3031                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
3032             }
3033         }
3034     }
3035     for(j=0; j<h->ref_count[1]; j++){
3036         for(i=0; i<h->ref_count[0]; i++)
3037             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
3038         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
3039         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
3040     }
3041 }
3042
3043 static int pred_weight_table(H264Context *h){
3044     MpegEncContext * const s = &h->s;
3045     int list, i;
3046     int luma_def, chroma_def;
3047
3048     h->use_weight= 0;
3049     h->use_weight_chroma= 0;
3050     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
3051     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
3052     luma_def = 1<<h->luma_log2_weight_denom;
3053     chroma_def = 1<<h->chroma_log2_weight_denom;
3054
3055     for(list=0; list<2; list++){
3056         for(i=0; i<h->ref_count[list]; i++){
3057             int luma_weight_flag, chroma_weight_flag;
3058
3059             luma_weight_flag= get_bits1(&s->gb);
3060             if(luma_weight_flag){
3061                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3062                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3063                 if(   h->luma_weight[list][i] != luma_def
3064                    || h->luma_offset[list][i] != 0)
3065                     h->use_weight= 1;
3066             }else{
3067                 h->luma_weight[list][i]= luma_def;
3068                 h->luma_offset[list][i]= 0;
3069             }
3070
3071             if(CHROMA){
3072                 chroma_weight_flag= get_bits1(&s->gb);
3073                 if(chroma_weight_flag){
3074                     int j;
3075                     for(j=0; j<2; j++){
3076                         h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3077                         h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3078                         if(   h->chroma_weight[list][i][j] != chroma_def
3079                         || h->chroma_offset[list][i][j] != 0)
3080                             h->use_weight_chroma= 1;
3081                     }
3082                 }else{
3083                     int j;
3084                     for(j=0; j<2; j++){
3085                         h->chroma_weight[list][i][j]= chroma_def;
3086                         h->chroma_offset[list][i][j]= 0;
3087                     }
3088                 }
3089             }
3090         }
3091         if(h->slice_type_nos != FF_B_TYPE) break;
3092     }
3093     h->use_weight= h->use_weight || h->use_weight_chroma;
3094     return 0;
3095 }
3096
3097 static void implicit_weight_table(H264Context *h){
3098     MpegEncContext * const s = &h->s;
3099     int ref0, ref1;
3100     int cur_poc = s->current_picture_ptr->poc;
3101
3102     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3103        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3104         h->use_weight= 0;
3105         h->use_weight_chroma= 0;
3106         return;
3107     }
3108
3109     h->use_weight= 2;
3110     h->use_weight_chroma= 2;
3111     h->luma_log2_weight_denom= 5;
3112     h->chroma_log2_weight_denom= 5;
3113
3114     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3115         int poc0 = h->ref_list[0][ref0].poc;
3116         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3117             int poc1 = h->ref_list[1][ref1].poc;
3118             int td = av_clip(poc1 - poc0, -128, 127);
3119             if(td){
3120                 int tb = av_clip(cur_poc - poc0, -128, 127);
3121                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3122                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3123                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3124                     h->implicit_weight[ref0][ref1] = 32;
3125                 else
3126                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3127             }else
3128                 h->implicit_weight[ref0][ref1] = 32;
3129         }
3130     }
3131 }
3132
3133 /**
3134  * Mark a picture as no longer needed for reference. The refmask
3135  * argument allows unreferencing of individual fields or the whole frame.
3136  * If the picture becomes entirely unreferenced, but is being held for
3137  * display purposes, it is marked as such.
3138  * @param refmask mask of fields to unreference; the mask is bitwise
3139  *                anded with the reference marking of pic
3140  * @return non-zero if pic becomes entirely unreferenced (except possibly
3141  *         for display purposes) zero if one of the fields remains in
3142  *         reference
3143  */
3144 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3145     int i;
3146     if (pic->reference &= refmask) {
3147         return 0;
3148     } else {
3149         for(i = 0; h->delayed_pic[i]; i++)
3150             if(pic == h->delayed_pic[i]){
3151                 pic->reference=DELAYED_PIC_REF;
3152                 break;
3153             }
3154         return 1;
3155     }
3156 }
3157
3158 /**
3159  * instantaneous decoder refresh.
3160  */
3161 static void idr(H264Context *h){
3162     int i;
3163
3164     for(i=0; i<16; i++){
3165         remove_long(h, i, 0);
3166     }
3167     assert(h->long_ref_count==0);
3168
3169     for(i=0; i<h->short_ref_count; i++){
3170         unreference_pic(h, h->short_ref[i], 0);
3171         h->short_ref[i]= NULL;
3172     }
3173     h->short_ref_count=0;
3174     h->prev_frame_num= 0;
3175     h->prev_frame_num_offset= 0;
3176     h->prev_poc_msb=
3177     h->prev_poc_lsb= 0;
3178 }
3179
3180 /* forget old pics after a seek */
3181 static void flush_dpb(AVCodecContext *avctx){
3182     H264Context *h= avctx->priv_data;
3183     int i;
3184     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
3185         if(h->delayed_pic[i])
3186             h->delayed_pic[i]->reference= 0;
3187         h->delayed_pic[i]= NULL;
3188     }
3189     h->outputed_poc= INT_MIN;
3190     idr(h);
3191     if(h->s.current_picture_ptr)
3192         h->s.current_picture_ptr->reference= 0;
3193     h->s.first_field= 0;
3194     ff_mpeg_flush(avctx);
3195 }
3196
3197 /**
3198  * Find a Picture in the short term reference list by frame number.
3199  * @param frame_num frame number to search for
3200  * @param idx the index into h->short_ref where returned picture is found
3201  *            undefined if no picture found.
3202  * @return pointer to the found picture, or NULL if no pic with the provided
3203  *                 frame number is found
3204  */
3205 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3206     MpegEncContext * const s = &h->s;
3207     int i;
3208
3209     for(i=0; i<h->short_ref_count; i++){
3210         Picture *pic= h->short_ref[i];
3211         if(s->avctx->debug&FF_DEBUG_MMCO)
3212             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3213         if(pic->frame_num == frame_num) {
3214             *idx = i;
3215             return pic;
3216         }
3217     }
3218     return NULL;
3219 }
3220
3221 /**
3222  * Remove a picture from the short term reference list by its index in
3223  * that list.  This does no checking on the provided index; it is assumed
3224  * to be valid. Other list entries are shifted down.
3225  * @param i index into h->short_ref of picture to remove.
3226  */
3227 static void remove_short_at_index(H264Context *h, int i){
3228     assert(i >= 0 && i < h->short_ref_count);
3229     h->short_ref[i]= NULL;
3230     if (--h->short_ref_count)
3231         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3232 }
3233
3234 /**
3235  *
3236  * @return the removed picture or NULL if an error occurs
3237  */
3238 static Picture * remove_short(H264Context *h, int frame_num, int ref_mask){
3239     MpegEncContext * const s = &h->s;
3240     Picture *pic;
3241     int i;
3242
3243     if(s->avctx->debug&FF_DEBUG_MMCO)
3244         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3245
3246     pic = find_short(h, frame_num, &i);
3247     if (pic){
3248         if(unreference_pic(h, pic, ref_mask))
3249         remove_short_at_index(h, i);
3250     }
3251
3252     return pic;
3253 }
3254
3255 /**
3256  * Remove a picture from the long term reference list by its index in
3257  * that list.
3258  * @return the removed picture or NULL if an error occurs
3259  */
3260 static Picture * remove_long(H264Context *h, int i, int ref_mask){
3261     Picture *pic;
3262
3263     pic= h->long_ref[i];
3264     if (pic){
3265         if(unreference_pic(h, pic, ref_mask)){
3266             assert(h->long_ref[i]->long_ref == 1);
3267             h->long_ref[i]->long_ref= 0;
3268             h->long_ref[i]= NULL;
3269             h->long_ref_count--;
3270         }
3271     }
3272
3273     return pic;
3274 }
3275
3276 /**
3277  * print short term list
3278  */
3279 static void print_short_term(H264Context *h) {
3280     uint32_t i;
3281     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3282         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3283         for(i=0; i<h->short_ref_count; i++){
3284             Picture *pic= h->short_ref[i];
3285             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3286         }
3287     }
3288 }
3289
3290 /**
3291  * print long term list
3292  */
3293 static void print_long_term(H264Context *h) {
3294     uint32_t i;
3295     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3296         av_log(h->s.avctx, AV_LOG_DEBUG, "long term list:\n");
3297         for(i = 0; i < 16; i++){
3298             Picture *pic= h->long_ref[i];
3299             if (pic) {
3300                 av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3301             }
3302         }
3303     }
3304 }
3305
3306 /**
3307  * Executes the reference picture marking (memory management control operations).
3308  */
3309 static int execute_ref_pic_marking(H264Context *h, MMCO *mmco, int mmco_count){
3310     MpegEncContext * const s = &h->s;
3311     int i, j;
3312     int current_ref_assigned=0;
3313     Picture *pic;
3314
3315     if((s->avctx->debug&FF_DEBUG_MMCO) && mmco_count==0)
3316         av_log(h->s.avctx, AV_LOG_DEBUG, "no mmco here\n");
3317
3318     for(i=0; i<mmco_count; i++){
3319         int structure, frame_num;
3320         if(s->avctx->debug&FF_DEBUG_MMCO)
3321             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);
3322
3323         if(   mmco[i].opcode == MMCO_SHORT2UNUSED
3324            || mmco[i].opcode == MMCO_SHORT2LONG){
3325             frame_num = pic_num_extract(h, mmco[i].short_pic_num, &structure);
3326             pic = find_short(h, frame_num, &j);
3327             if(!pic){
3328                 if(mmco[i].opcode != MMCO_SHORT2LONG || !h->long_ref[mmco[i].long_arg]
3329                    || h->long_ref[mmco[i].long_arg]->frame_num != frame_num)
3330                 av_log(h->s.avctx, AV_LOG_ERROR, "mmco: unref short failure\n");
3331                 continue;
3332             }
3333         }
3334
3335         switch(mmco[i].opcode){
3336         case MMCO_SHORT2UNUSED:
3337             if(s->avctx->debug&FF_DEBUG_MMCO)
3338                 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);
3339             remove_short(h, frame_num, structure ^ PICT_FRAME);
3340             break;
3341         case MMCO_SHORT2LONG:
3342                 if (h->long_ref[mmco[i].long_arg] != pic)
3343                     remove_long(h, mmco[i].long_arg, 0);
3344
3345                 remove_short_at_index(h, j);
3346                 h->long_ref[ mmco[i].long_arg ]= pic;
3347                 if (h->long_ref[ mmco[i].long_arg ]){
3348                     h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3349                     h->long_ref_count++;
3350                 }
3351             break;
3352         case MMCO_LONG2UNUSED:
3353             j = pic_num_extract(h, mmco[i].long_arg, &structure);
3354             pic = h->long_ref[j];
3355             if (pic) {
3356                 remove_long(h, j, structure ^ PICT_FRAME);
3357             } else if(s->avctx->debug&FF_DEBUG_MMCO)
3358                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref long failure\n");
3359             break;
3360         case MMCO_LONG:
3361                     // Comment below left from previous code as it is an interresting note.
3362                     /* First field in pair is in short term list or
3363                      * at a different long term index.
3364                      * This is not allowed; see 7.4.3.3, notes 2 and 3.
3365                      * Report the problem and keep the pair where it is,
3366                      * and mark this field valid.
3367                      */
3368
3369             if (h->long_ref[mmco[i].long_arg] != s->current_picture_ptr) {
3370                 remove_long(h, mmco[i].long_arg, 0);
3371
3372                 h->long_ref[ mmco[i].long_arg ]= s->current_picture_ptr;
3373                 h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3374                 h->long_ref_count++;
3375             }
3376
3377             s->current_picture_ptr->reference |= s->picture_structure;
3378             current_ref_assigned=1;
3379             break;
3380         case MMCO_SET_MAX_LONG:
3381             assert(mmco[i].long_arg <= 16);
3382             // just remove the long term which index is greater than new max
3383             for(j = mmco[i].long_arg; j<16; j++){
3384                 remove_long(h, j, 0);
3385             }
3386             break;
3387         case MMCO_RESET:
3388             while(h->short_ref_count){
3389                 remove_short(h, h->short_ref[0]->frame_num, 0);
3390             }
3391             for(j = 0; j < 16; j++) {
3392                 remove_long(h, j, 0);
3393             }
3394             s->current_picture_ptr->poc=
3395             s->current_picture_ptr->field_poc[0]=
3396             s->current_picture_ptr->field_poc[1]=
3397             h->poc_lsb=
3398             h->poc_msb=
3399             h->frame_num=
3400             s->current_picture_ptr->frame_num= 0;
3401             break;
3402         default: assert(0);
3403         }
3404     }
3405
3406     if (!current_ref_assigned) {
3407         /* Second field of complementary field pair; the first field of
3408          * which is already referenced. If short referenced, it
3409          * should be first entry in short_ref. If not, it must exist
3410          * in long_ref; trying to put it on the short list here is an
3411          * error in the encoded bit stream (ref: 7.4.3.3, NOTE 2 and 3).
3412          */
3413         if (h->short_ref_count && h->short_ref[0] == s->current_picture_ptr) {
3414             /* Just mark the second field valid */
3415             s->current_picture_ptr->reference = PICT_FRAME;
3416         } else if (s->current_picture_ptr->long_ref) {
3417             av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term reference "
3418                                              "assignment for second field "
3419                                              "in complementary field pair "
3420                                              "(first field is long term)\n");
3421         } else {
3422             pic= remove_short(h, s->current_picture_ptr->frame_num, 0);
3423             if(pic){
3424                 av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term buffer state detected\n");
3425             }
3426
3427             if(h->short_ref_count)
3428                 memmove(&h->short_ref[1], &h->short_ref[0], h->short_ref_count*sizeof(Picture*));
3429
3430             h->short_ref[0]= s->current_picture_ptr;
3431             h->short_ref_count++;
3432             s->current_picture_ptr->reference |= s->picture_structure;
3433         }
3434     }
3435
3436     if (h->long_ref_count + h->short_ref_count > h->sps.ref_frame_count){
3437
3438         /* We have too many reference frames, probably due to corrupted
3439          * stream. Need to discard one frame. Prevents overrun of the
3440          * short_ref and long_ref buffers.
3441          */
3442         av_log(h->s.avctx, AV_LOG_ERROR,
3443                "number of reference frames exceeds max (probably "
3444                "corrupt input), discarding one\n");
3445
3446         if (h->long_ref_count && !h->short_ref_count) {
3447             for (i = 0; i < 16; ++i)
3448                 if (h->long_ref[i])
3449                     break;
3450
3451             assert(i < 16);
3452             remove_long(h, i, 0);
3453         } else {
3454             pic = h->short_ref[h->short_ref_count - 1];
3455             remove_short(h, pic->frame_num, 0);
3456         }
3457     }
3458
3459     print_short_term(h);
3460     print_long_term(h);
3461     return 0;
3462 }
3463
3464 static int decode_ref_pic_marking(H264Context *h, GetBitContext *gb){
3465     MpegEncContext * const s = &h->s;
3466     int i;
3467
3468     h->mmco_index= 0;
3469     if(h->nal_unit_type == NAL_IDR_SLICE){ //FIXME fields
3470         s->broken_link= get_bits1(gb) -1;
3471         if(get_bits1(gb)){
3472             h->mmco[0].opcode= MMCO_LONG;
3473             h->mmco[0].long_arg= 0;
3474             h->mmco_index= 1;
3475         }
3476     }else{
3477         if(get_bits1(gb)){ // adaptive_ref_pic_marking_mode_flag
3478             for(i= 0; i<MAX_MMCO_COUNT; i++) {
3479                 MMCOOpcode opcode= get_ue_golomb(gb);
3480
3481                 h->mmco[i].opcode= opcode;
3482                 if(opcode==MMCO_SHORT2UNUSED || opcode==MMCO_SHORT2LONG){
3483                     h->mmco[i].short_pic_num= (h->curr_pic_num - get_ue_golomb(gb) - 1) & (h->max_pic_num - 1);
3484 /*                    if(h->mmco[i].short_pic_num >= h->short_ref_count || h->short_ref[ h->mmco[i].short_pic_num ] == NULL){
3485                         av_log(s->avctx, AV_LOG_ERROR, "illegal short ref in memory management control operation %d\n", mmco);
3486                         return -1;
3487                     }*/
3488                 }
3489                 if(opcode==MMCO_SHORT2LONG || opcode==MMCO_LONG2UNUSED || opcode==MMCO_LONG || opcode==MMCO_SET_MAX_LONG){
3490                     unsigned int long_arg= get_ue_golomb(gb);
3491                     if(long_arg >= 32 || (long_arg >= 16 && !(opcode == MMCO_LONG2UNUSED && FIELD_PICTURE))){
3492                         av_log(h->s.avctx, AV_LOG_ERROR, "illegal long ref in memory management control operation %d\n", opcode);
3493                         return -1;
3494                     }
3495                     h->mmco[i].long_arg= long_arg;
3496                 }
3497
3498                 if(opcode > (unsigned)MMCO_LONG){
3499                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal memory management control operation %d\n", opcode);
3500                     return -1;
3501                 }
3502                 if(opcode == MMCO_END)
3503                     break;
3504             }
3505             h->mmco_index= i;
3506         }else{
3507             assert(h->long_ref_count + h->short_ref_count <= h->sps.ref_frame_count);
3508
3509             if(h->short_ref_count && h->long_ref_count + h->short_ref_count == h->sps.ref_frame_count &&
3510                     !(FIELD_PICTURE && !s->first_field && s->current_picture_ptr->reference)) {
3511                 h->mmco[0].opcode= MMCO_SHORT2UNUSED;
3512                 h->mmco[0].short_pic_num= h->short_ref[ h->short_ref_count - 1 ]->frame_num;
3513                 h->mmco_index= 1;
3514                 if (FIELD_PICTURE) {
3515                     h->mmco[0].short_pic_num *= 2;
3516                     h->mmco[1].opcode= MMCO_SHORT2UNUSED;
3517                     h->mmco[1].short_pic_num= h->mmco[0].short_pic_num + 1;
3518                     h->mmco_index= 2;
3519                 }
3520             }
3521         }
3522     }
3523
3524     return 0;
3525 }
3526
3527 static int init_poc(H264Context *h){
3528     MpegEncContext * const s = &h->s;
3529     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
3530     int field_poc[2];
3531     Picture *cur = s->current_picture_ptr;
3532
3533     h->frame_num_offset= h->prev_frame_num_offset;
3534     if(h->frame_num < h->prev_frame_num)
3535         h->frame_num_offset += max_frame_num;
3536
3537     if(h->sps.poc_type==0){
3538         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
3539
3540         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
3541             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
3542         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
3543             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
3544         else
3545             h->poc_msb = h->prev_poc_msb;
3546 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
3547         field_poc[0] =
3548         field_poc[1] = h->poc_msb + h->poc_lsb;
3549         if(s->picture_structure == PICT_FRAME)
3550             field_poc[1] += h->delta_poc_bottom;
3551     }else if(h->sps.poc_type==1){
3552         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
3553         int i;
3554
3555         if(h->sps.poc_cycle_length != 0)
3556             abs_frame_num = h->frame_num_offset + h->frame_num;
3557         else
3558             abs_frame_num = 0;
3559
3560         if(h->nal_ref_idc==0 && abs_frame_num > 0)
3561             abs_frame_num--;
3562
3563         expected_delta_per_poc_cycle = 0;
3564         for(i=0; i < h->sps.poc_cycle_length; i++)
3565             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
3566
3567         if(abs_frame_num > 0){
3568             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
3569             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
3570
3571             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
3572             for(i = 0; i <= frame_num_in_poc_cycle; i++)
3573                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
3574         } else
3575             expectedpoc = 0;
3576
3577         if(h->nal_ref_idc == 0)
3578             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
3579
3580         field_poc[0] = expectedpoc + h->delta_poc[0];
3581         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
3582
3583         if(s->picture_structure == PICT_FRAME)
3584             field_poc[1] += h->delta_poc[1];
3585     }else{
3586         int poc= 2*(h->frame_num_offset + h->frame_num);
3587
3588         if(!h->nal_ref_idc)
3589             poc--;
3590
3591         field_poc[0]= poc;
3592         field_poc[1]= poc;
3593     }
3594
3595     if(s->picture_structure != PICT_BOTTOM_FIELD)
3596         s->current_picture_ptr->field_poc[0]= field_poc[0];
3597     if(s->picture_structure != PICT_TOP_FIELD)
3598         s->current_picture_ptr->field_poc[1]= field_poc[1];
3599     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
3600
3601     return 0;
3602 }
3603
3604
3605 /**
3606  * initialize scan tables
3607  */
3608 static void init_scan_tables(H264Context *h){
3609     MpegEncContext * const s = &h->s;
3610     int i;
3611     if(s->dsp.h264_idct_add == ff_h264_idct_add_c){ //FIXME little ugly
3612         memcpy(h->zigzag_scan, zigzag_scan, 16*sizeof(uint8_t));
3613         memcpy(h-> field_scan,  field_scan, 16*sizeof(uint8_t));
3614     }else{
3615         for(i=0; i<16; i++){
3616 #define T(x) (x>>2) | ((x<<2) & 0xF)
3617             h->zigzag_scan[i] = T(zigzag_scan[i]);
3618             h-> field_scan[i] = T( field_scan[i]);
3619 #undef T
3620         }
3621     }
3622     if(s->dsp.h264_idct8_add == ff_h264_idct8_add_c){
3623         memcpy(h->zigzag_scan8x8,       zigzag_scan8x8,       64*sizeof(uint8_t));
3624         memcpy(h->zigzag_scan8x8_cavlc, zigzag_scan8x8_cavlc, 64*sizeof(uint8_t));
3625         memcpy(h->field_scan8x8,        field_scan8x8,        64*sizeof(uint8_t));
3626         memcpy(h->field_scan8x8_cavlc,  field_scan8x8_cavlc,  64*sizeof(uint8_t));
3627     }else{
3628         for(i=0; i<64; i++){
3629 #define T(x) (x>>3) | ((x&7)<<3)
3630             h->zigzag_scan8x8[i]       = T(zigzag_scan8x8[i]);
3631             h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
3632             h->field_scan8x8[i]        = T(field_scan8x8[i]);
3633             h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
3634 #undef T
3635         }
3636     }
3637     if(h->sps.transform_bypass){ //FIXME same ugly
3638         h->zigzag_scan_q0          = zigzag_scan;
3639         h->zigzag_scan8x8_q0       = zigzag_scan8x8;
3640         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
3641         h->field_scan_q0           = field_scan;
3642         h->field_scan8x8_q0        = field_scan8x8;
3643         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
3644     }else{
3645         h->zigzag_scan_q0          = h->zigzag_scan;
3646         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
3647         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
3648         h->field_scan_q0           = h->field_scan;
3649         h->field_scan8x8_q0        = h->field_scan8x8;
3650         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
3651     }
3652 }
3653
3654 /**
3655  * Replicates H264 "master" context to thread contexts.
3656  */
3657 static void clone_slice(H264Context *dst, H264Context *src)
3658 {
3659     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
3660     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
3661     dst->s.current_picture      = src->s.current_picture;
3662     dst->s.linesize             = src->s.linesize;
3663     dst->s.uvlinesize           = src->s.uvlinesize;
3664     dst->s.first_field          = src->s.first_field;
3665
3666     dst->prev_poc_msb           = src->prev_poc_msb;
3667     dst->prev_poc_lsb           = src->prev_poc_lsb;
3668     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
3669     dst->prev_frame_num         = src->prev_frame_num;
3670     dst->short_ref_count        = src->short_ref_count;
3671
3672     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
3673     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
3674     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
3675     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
3676
3677     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
3678     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
3679 }
3680
3681 /**
3682  * decodes a slice header.
3683  * This will also call MPV_common_init() and frame_start() as needed.
3684  *
3685  * @param h h264context
3686  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
3687  *
3688  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
3689  */
3690 static int decode_slice_header(H264Context *h, H264Context *h0){
3691     MpegEncContext * const s = &h->s;
3692     MpegEncContext * const s0 = &h0->s;
3693     unsigned int first_mb_in_slice;
3694     unsigned int pps_id;
3695     int num_ref_idx_active_override_flag;
3696     static const uint8_t slice_type_map[5]= {FF_P_TYPE, FF_B_TYPE, FF_I_TYPE, FF_SP_TYPE, FF_SI_TYPE};
3697     unsigned int slice_type, tmp, i, j;
3698     int default_ref_list_done = 0;
3699     int last_pic_structure;
3700
3701     s->dropable= h->nal_ref_idc == 0;
3702
3703     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc){
3704         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
3705         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
3706     }else{
3707         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
3708         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
3709     }
3710
3711     first_mb_in_slice= get_ue_golomb(&s->gb);
3712
3713     if((s->flags2 & CODEC_FLAG2_CHUNKS) && first_mb_in_slice == 0){
3714         h0->current_slice = 0;
3715         if (!s0->first_field)
3716             s->current_picture_ptr= NULL;
3717     }
3718
3719     slice_type= get_ue_golomb(&s->gb);
3720     if(slice_type > 9){
3721         av_log(h->s.avctx, AV_LOG_ERROR, "slice type too large (%d) at %d %d\n", h->slice_type, s->mb_x, s->mb_y);
3722         return -1;
3723     }
3724     if(slice_type > 4){
3725         slice_type -= 5;
3726         h->slice_type_fixed=1;
3727     }else
3728         h->slice_type_fixed=0;
3729
3730     slice_type= slice_type_map[ slice_type ];
3731     if (slice_type == FF_I_TYPE
3732         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
3733         default_ref_list_done = 1;
3734     }
3735     h->slice_type= slice_type;
3736     h->slice_type_nos= slice_type & 3;
3737
3738     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
3739     if (s->pict_type == FF_B_TYPE && s0->last_picture_ptr == NULL) {
3740         av_log(h->s.avctx, AV_LOG_ERROR,
3741                "B picture before any references, skipping\n");
3742         return -1;
3743     }
3744
3745     pps_id= get_ue_golomb(&s->gb);
3746     if(pps_id>=MAX_PPS_COUNT){
3747         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
3748         return -1;
3749     }
3750     if(!h0->pps_buffers[pps_id]) {
3751         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS referenced\n");
3752         return -1;
3753     }
3754     h->pps= *h0->pps_buffers[pps_id];
3755
3756     if(!h0->sps_buffers[h->pps.sps_id]) {
3757         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS referenced\n");
3758         return -1;
3759     }
3760     h->sps = *h0->sps_buffers[h->pps.sps_id];
3761
3762     if(h == h0 && h->dequant_coeff_pps != pps_id){
3763         h->dequant_coeff_pps = pps_id;
3764         init_dequant_tables(h);
3765     }
3766
3767     s->mb_width= h->sps.mb_width;
3768     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
3769
3770     h->b_stride=  s->mb_width*4;
3771     h->b8_stride= s->mb_width*2;
3772
3773     s->width = 16*s->mb_width - 2*FFMIN(h->sps.crop_right, 7);
3774     if(h->sps.frame_mbs_only_flag)
3775         s->height= 16*s->mb_height - 2*FFMIN(h->sps.crop_bottom, 7);
3776     else
3777         s->height= 16*s->mb_height - 4*FFMIN(h->sps.crop_bottom, 3);
3778
3779     if (s->context_initialized
3780         && (   s->width != s->avctx->width || s->height != s->avctx->height)) {
3781         if(h != h0)
3782             return -1;   // width / height changed during parallelized decoding
3783         free_tables(h);
3784         MPV_common_end(s);
3785     }
3786     if (!s->context_initialized) {
3787         if(h != h0)
3788             return -1;  // we cant (re-)initialize context during parallel decoding
3789         if (MPV_common_init(s) < 0)
3790             return -1;
3791         s->first_field = 0;
3792
3793         init_scan_tables(h);
3794         alloc_tables(h);
3795
3796         for(i = 1; i < s->avctx->thread_count; i++) {
3797             H264Context *c;
3798             c = h->thread_context[i] = av_malloc(sizeof(H264Context));
3799             memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
3800             memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
3801             c->sps = h->sps;
3802             c->pps = h->pps;
3803             init_scan_tables(c);
3804             clone_tables(c, h);
3805         }
3806
3807         for(i = 0; i < s->avctx->thread_count; i++)
3808             if(context_init(h->thread_context[i]) < 0)
3809                 return -1;
3810
3811         s->avctx->width = s->width;
3812         s->avctx->height = s->height;
3813         s->avctx->sample_aspect_ratio= h->sps.sar;
3814         if(!s->avctx->sample_aspect_ratio.den)
3815             s->avctx->sample_aspect_ratio.den = 1;
3816
3817         if(h->sps.timing_info_present_flag){
3818             s->avctx->time_base= (AVRational){h->sps.num_units_in_tick * 2, h->sps.time_scale};
3819             if(h->x264_build > 0 && h->x264_build < 44)
3820                 s->avctx->time_base.den *= 2;
3821             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
3822                       s->avctx->time_base.num, s->avctx->time_base.den, 1<<30);
3823         }
3824     }
3825
3826     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
3827
3828     h->mb_mbaff = 0;
3829     h->mb_aff_frame = 0;
3830     last_pic_structure = s0->picture_structure;
3831     if(h->sps.frame_mbs_only_flag){
3832         s->picture_structure= PICT_FRAME;
3833     }else{
3834         if(get_bits1(&s->gb)) { //field_pic_flag
3835             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
3836         } else {
3837             s->picture_structure= PICT_FRAME;
3838             h->mb_aff_frame = h->sps.mb_aff;
3839         }
3840     }
3841     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
3842
3843     if(h0->current_slice == 0){
3844         while(h->frame_num !=  h->prev_frame_num &&
3845               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
3846             av_log(NULL, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
3847             frame_start(h);
3848             h->prev_frame_num++;
3849             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
3850             s->current_picture_ptr->frame_num= h->prev_frame_num;
3851             execute_ref_pic_marking(h, NULL, 0);
3852         }
3853
3854         /* See if we have a decoded first field looking for a pair... */
3855         if (s0->first_field) {
3856             assert(s0->current_picture_ptr);
3857             assert(s0->current_picture_ptr->data[0]);
3858             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
3859
3860             /* figure out if we have a complementary field pair */
3861             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3862                 /*
3863                  * Previous field is unmatched. Don't display it, but let it
3864                  * remain for reference if marked as such.
3865                  */
3866                 s0->current_picture_ptr = NULL;
3867                 s0->first_field = FIELD_PICTURE;
3868
3869             } else {
3870                 if (h->nal_ref_idc &&
3871                         s0->current_picture_ptr->reference &&
3872                         s0->current_picture_ptr->frame_num != h->frame_num) {
3873                     /*
3874                      * This and previous field were reference, but had
3875                      * different frame_nums. Consider this field first in
3876                      * pair. Throw away previous field except for reference
3877                      * purposes.
3878                      */
3879                     s0->first_field = 1;
3880                     s0->current_picture_ptr = NULL;
3881
3882                 } else {
3883                     /* Second field in complementary pair */
3884                     s0->first_field = 0;
3885                 }
3886             }
3887
3888         } else {
3889             /* Frame or first field in a potentially complementary pair */
3890             assert(!s0->current_picture_ptr);
3891             s0->first_field = FIELD_PICTURE;
3892         }
3893
3894         if((!FIELD_PICTURE || s0->first_field) && frame_start(h) < 0) {
3895             s0->first_field = 0;
3896             return -1;
3897         }
3898     }
3899     if(h != h0)
3900         clone_slice(h, h0);
3901
3902     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
3903
3904     assert(s->mb_num == s->mb_width * s->mb_height);
3905     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
3906        first_mb_in_slice                    >= s->mb_num){
3907         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
3908         return -1;
3909     }
3910     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
3911     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
3912     if (s->picture_structure == PICT_BOTTOM_FIELD)
3913         s->resync_mb_y = s->mb_y = s->mb_y + 1;
3914     assert(s->mb_y < s->mb_height);
3915
3916     if(s->picture_structure==PICT_FRAME){
3917         h->curr_pic_num=   h->frame_num;
3918         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
3919     }else{
3920         h->curr_pic_num= 2*h->frame_num + 1;
3921         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
3922     }
3923
3924     if(h->nal_unit_type == NAL_IDR_SLICE){
3925         get_ue_golomb(&s->gb); /* idr_pic_id */
3926     }
3927
3928     if(h->sps.poc_type==0){
3929         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
3930
3931         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
3932             h->delta_poc_bottom= get_se_golomb(&s->gb);
3933         }
3934     }
3935
3936     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
3937         h->delta_poc[0]= get_se_golomb(&s->gb);
3938
3939         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
3940             h->delta_poc[1]= get_se_golomb(&s->gb);
3941     }
3942
3943     init_poc(h);
3944
3945     if(h->pps.redundant_pic_cnt_present){
3946         h->redundant_pic_count= get_ue_golomb(&s->gb);
3947     }
3948
3949     //set defaults, might be overridden a few lines later
3950     h->ref_count[0]= h->pps.ref_count[0];
3951     h->ref_count[1]= h->pps.ref_count[1];
3952
3953     if(h->slice_type_nos != FF_I_TYPE){
3954         if(h->slice_type_nos == FF_B_TYPE){
3955             h->direct_spatial_mv_pred= get_bits1(&s->gb);
3956         }
3957         num_ref_idx_active_override_flag= get_bits1(&s->gb);
3958
3959         if(num_ref_idx_active_override_flag){
3960             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
3961             if(h->slice_type_nos==FF_B_TYPE)
3962                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
3963
3964             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
3965                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3966                 h->ref_count[0]= h->ref_count[1]= 1;
3967                 return -1;
3968             }
3969         }
3970         if(h->slice_type_nos == FF_B_TYPE)
3971             h->list_count= 2;
3972         else
3973             h->list_count= 1;
3974     }else
3975         h->list_count= 0;
3976
3977     if(!default_ref_list_done){
3978         fill_default_ref_list(h);
3979     }
3980
3981     if(h->slice_type_nos!=FF_I_TYPE && decode_ref_pic_list_reordering(h) < 0)
3982         return -1;
3983
3984     if(   (h->pps.weighted_pred          && h->slice_type_nos == FF_P_TYPE )
3985        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== FF_B_TYPE ) )
3986         pred_weight_table(h);
3987     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== FF_B_TYPE)
3988         implicit_weight_table(h);
3989     else
3990         h->use_weight = 0;
3991
3992     if(h->nal_ref_idc)
3993         decode_ref_pic_marking(h0, &s->gb);
3994
3995     if(FRAME_MBAFF)
3996         fill_mbaff_ref_list(h);
3997
3998     if( h->slice_type_nos != FF_I_TYPE && h->pps.cabac ){
3999         tmp = get_ue_golomb(&s->gb);
4000         if(tmp > 2){
4001             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
4002             return -1;
4003         }
4004         h->cabac_init_idc= tmp;
4005     }
4006
4007     h->last_qscale_diff = 0;
4008     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
4009     if(tmp>51){
4010         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
4011         return -1;
4012     }
4013     s->qscale= tmp;
4014     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
4015     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
4016     //FIXME qscale / qp ... stuff
4017     if(h->slice_type == FF_SP_TYPE){
4018         get_bits1(&s->gb); /* sp_for_switch_flag */
4019     }
4020     if(h->slice_type==FF_SP_TYPE || h->slice_type == FF_SI_TYPE){
4021         get_se_golomb(&s->gb); /* slice_qs_delta */
4022     }
4023
4024     h->deblocking_filter = 1;
4025     h->slice_alpha_c0_offset = 0;
4026     h->slice_beta_offset = 0;
4027     if( h->pps.deblocking_filter_parameters_present ) {
4028         tmp= get_ue_golomb(&s->gb);
4029         if(tmp > 2){
4030             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
4031             return -1;
4032         }
4033         h->deblocking_filter= tmp;
4034         if(h->deblocking_filter < 2)
4035             h->deblocking_filter^= 1; // 1<->0
4036
4037         if( h->deblocking_filter ) {
4038             h->slice_alpha_c0_offset = get_se_golomb(&s->gb) << 1;
4039             h->slice_beta_offset = get_se_golomb(&s->gb) << 1;
4040         }
4041     }
4042