Use local variabes for *stride, where local variables exist.
[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                 b8_stride = 0;
977             }
978         }else if(!(s->picture_structure & h->ref_list[1][0].reference)){// FL -> FL & differ parity
979             int fieldoff= 2*(h->ref_list[1][0].reference)-3;
980             mb_xy += s->mb_stride*fieldoff;
981         }
982         goto single_col;
983     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
984         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
985             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
986             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
987             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
988             b8_stride *= 3;
989             b4_stride *= 6;
990             //FIXME IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag
991             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
992                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
993                 && !is_b8x8){
994                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
995                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
996             }else{
997                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
998                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
999             }
1000         }else{                                           //     AFR/FR    -> AFR/FR
1001 single_col:
1002             mb_type_col[0] =
1003             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
1004             if(IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag){
1005                 /* FIXME save sub mb types from previous frames (or derive from MVs)
1006                 * so we know exactly what block size to use */
1007                 sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
1008                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1009             }else if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
1010                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1011                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
1012             }else{
1013                 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1014                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
1015             }
1016         }
1017     }
1018
1019     l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
1020     l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
1021     l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
1022     l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
1023     if(!b8_stride){
1024         if(s->mb_y&1){
1025             l1ref0 += h->b8_stride;
1026             l1ref1 += h->b8_stride;
1027             l1mv0  +=  2*b4_stride;
1028             l1mv1  +=  2*b4_stride;
1029         }
1030     }
1031
1032     if(h->direct_spatial_mv_pred){
1033         int ref[2];
1034         int mv[2][2];
1035         int list;
1036
1037         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
1038
1039         /* ref = min(neighbors) */
1040         for(list=0; list<2; list++){
1041             int refa = h->ref_cache[list][scan8[0] - 1];
1042             int refb = h->ref_cache[list][scan8[0] - 8];
1043             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1044             if(refc == PART_NOT_AVAILABLE)
1045                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1046             ref[list] = FFMIN3((unsigned)refa, (unsigned)refb, (unsigned)refc);
1047             if(ref[list] < 0)
1048                 ref[list] = -1;
1049         }
1050
1051         if(ref[0] < 0 && ref[1] < 0){
1052             ref[0] = ref[1] = 0;
1053             mv[0][0] = mv[0][1] =
1054             mv[1][0] = mv[1][1] = 0;
1055         }else{
1056             for(list=0; list<2; list++){
1057                 if(ref[list] >= 0)
1058                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1059                 else
1060                     mv[list][0] = mv[list][1] = 0;
1061             }
1062         }
1063
1064         if(ref[1] < 0){
1065             if(!is_b8x8)
1066                 *mb_type &= ~MB_TYPE_L1;
1067             sub_mb_type &= ~MB_TYPE_L1;
1068         }else if(ref[0] < 0){
1069             if(!is_b8x8)
1070                 *mb_type &= ~MB_TYPE_L0;
1071             sub_mb_type &= ~MB_TYPE_L0;
1072         }
1073
1074         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1075             for(i8=0; i8<4; i8++){
1076                 int x8 = i8&1;
1077                 int y8 = i8>>1;
1078                 int xy8 = x8+y8*b8_stride;
1079                 int xy4 = 3*x8+y8*b4_stride;
1080                 int a=0, b=0;
1081
1082                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1083                     continue;
1084                 h->sub_mb_type[i8] = sub_mb_type;
1085
1086                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1087                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1088                 if(!IS_INTRA(mb_type_col[y8])
1089                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
1090                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
1091                     if(ref[0] > 0)
1092                         a= pack16to32(mv[0][0],mv[0][1]);
1093                     if(ref[1] > 0)
1094                         b= pack16to32(mv[1][0],mv[1][1]);
1095                 }else{
1096                     a= pack16to32(mv[0][0],mv[0][1]);
1097                     b= pack16to32(mv[1][0],mv[1][1]);
1098                 }
1099                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
1100                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
1101             }
1102         }else if(IS_16X16(*mb_type)){
1103             int a=0, b=0;
1104
1105             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1106             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1107             if(!IS_INTRA(mb_type_col[0])
1108                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1109                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1110                        && (h->x264_build>33 || !h->x264_build)))){
1111                 if(ref[0] > 0)
1112                     a= pack16to32(mv[0][0],mv[0][1]);
1113                 if(ref[1] > 0)
1114                     b= pack16to32(mv[1][0],mv[1][1]);
1115             }else{
1116                 a= pack16to32(mv[0][0],mv[0][1]);
1117                 b= pack16to32(mv[1][0],mv[1][1]);
1118             }
1119             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1120             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1121         }else{
1122             for(i8=0; i8<4; i8++){
1123                 const int x8 = i8&1;
1124                 const int y8 = i8>>1;
1125
1126                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1127                     continue;
1128                 h->sub_mb_type[i8] = sub_mb_type;
1129
1130                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1131                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1132                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1133                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1134
1135                 /* col_zero_flag */
1136                 if(!IS_INTRA(mb_type_col[0]) && (   l1ref0[x8 + y8*b8_stride] == 0
1137                                               || (l1ref0[x8 + y8*b8_stride] < 0 && l1ref1[x8 + y8*b8_stride] == 0
1138                                                   && (h->x264_build>33 || !h->x264_build)))){
1139                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*b8_stride] == 0 ? l1mv0 : l1mv1;
1140                     if(IS_SUB_8X8(sub_mb_type)){
1141                         const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1142                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1143                             if(ref[0] == 0)
1144                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1145                             if(ref[1] == 0)
1146                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1147                         }
1148                     }else
1149                     for(i4=0; i4<4; i4++){
1150                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1151                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1152                             if(ref[0] == 0)
1153                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1154                             if(ref[1] == 0)
1155                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1156                         }
1157                     }
1158                 }
1159             }
1160         }
1161     }else{ /* direct temporal mv pred */
1162         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1163         const int *dist_scale_factor = h->dist_scale_factor;
1164
1165         if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
1166             map_col_to_list0[0] = h->map_col_to_list0_field[0];
1167             map_col_to_list0[1] = h->map_col_to_list0_field[1];
1168             dist_scale_factor = h->dist_scale_factor_field;
1169         }
1170         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
1171             /* FIXME assumes direct_8x8_inference == 1 */
1172             int y_shift  = 2*!IS_INTERLACED(*mb_type);
1173             int ref_shift= FRAME_MBAFF ? y_shift : 1;
1174
1175             for(i8=0; i8<4; i8++){
1176                 const int x8 = i8&1;
1177                 const int y8 = i8>>1;
1178                 int ref0, scale;
1179                 const int16_t (*l1mv)[2]= l1mv0;
1180
1181                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1182                     continue;
1183                 h->sub_mb_type[i8] = sub_mb_type;
1184
1185                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1186                 if(IS_INTRA(mb_type_col[y8])){
1187                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1188                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1189                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1190                     continue;
1191                 }
1192
1193                 ref0 = l1ref0[x8 + y8*b8_stride];
1194                 if(ref0 >= 0)
1195                     ref0 = map_col_to_list0[0][ref0*2>>ref_shift];
1196                 else{
1197                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride]*2>>ref_shift];
1198                     l1mv= l1mv1;
1199                 }
1200                 scale = dist_scale_factor[ref0];
1201                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1202
1203                 {
1204                     const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride];
1205                     int my_col = (mv_col[1]<<y_shift)/2;
1206                     int mx = (scale * mv_col[0] + 128) >> 8;
1207                     int my = (scale * my_col + 128) >> 8;
1208                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1209                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1210                 }
1211             }
1212             return;
1213         }
1214
1215         /* one-to-one mv scaling */
1216
1217         if(IS_16X16(*mb_type)){
1218             int ref, mv0, mv1;
1219
1220             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1221             if(IS_INTRA(mb_type_col[0])){
1222                 ref=mv0=mv1=0;
1223             }else{
1224                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0]]
1225                                                 : map_col_to_list0[1][l1ref1[0]];
1226                 const int scale = dist_scale_factor[ref0];
1227                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1228                 int mv_l0[2];
1229                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1230                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1231                 ref= ref0;
1232                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1233                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1234             }
1235             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1236             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1237             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1238         }else{
1239             for(i8=0; i8<4; i8++){
1240                 const int x8 = i8&1;
1241                 const int y8 = i8>>1;
1242                 int ref0, scale;
1243                 const int16_t (*l1mv)[2]= l1mv0;
1244
1245                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1246                     continue;
1247                 h->sub_mb_type[i8] = sub_mb_type;
1248                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1249                 if(IS_INTRA(mb_type_col[0])){
1250                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1251                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1252                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1253                     continue;
1254                 }
1255
1256                 ref0 = l1ref0[x8 + y8*b8_stride];
1257                 if(ref0 >= 0)
1258                     ref0 = map_col_to_list0[0][ref0];
1259                 else{
1260                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride]];
1261                     l1mv= l1mv1;
1262                 }
1263                 scale = dist_scale_factor[ref0];
1264
1265                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1266                 if(IS_SUB_8X8(sub_mb_type)){
1267                     const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
1268                     int mx = (scale * mv_col[0] + 128) >> 8;
1269                     int my = (scale * mv_col[1] + 128) >> 8;
1270                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1271                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1272                 }else
1273                 for(i4=0; i4<4; i4++){
1274                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
1275                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1276                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1277                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1278                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1279                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1280                 }
1281             }
1282         }
1283     }
1284 }
1285
1286 static inline void write_back_motion(H264Context *h, int mb_type){
1287     MpegEncContext * const s = &h->s;
1288     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1289     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1290     int list;
1291
1292     if(!USES_LIST(mb_type, 0))
1293         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1294
1295     for(list=0; list<h->list_count; list++){
1296         int y;
1297         if(!USES_LIST(mb_type, list))
1298             continue;
1299
1300         for(y=0; y<4; y++){
1301             *(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];
1302             *(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];
1303         }
1304         if( h->pps.cabac ) {
1305             if(IS_SKIP(mb_type))
1306                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1307             else
1308             for(y=0; y<4; y++){
1309                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1310                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1311             }
1312         }
1313
1314         {
1315             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1316             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1317             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1318             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1319             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1320         }
1321     }
1322
1323     if(h->slice_type_nos == FF_B_TYPE && h->pps.cabac){
1324         if(IS_8X8(mb_type)){
1325             uint8_t *direct_table = &h->direct_table[b8_xy];
1326             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1327             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1328             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1329         }
1330     }
1331 }
1332
1333 /**
1334  * Decodes a network abstraction layer unit.
1335  * @param consumed is the number of bytes used as input
1336  * @param length is the length of the array
1337  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1338  * @returns decoded bytes, might be src+1 if no escapes
1339  */
1340 static const uint8_t *decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
1341     int i, si, di;
1342     uint8_t *dst;
1343     int bufidx;
1344
1345 //    src[0]&0x80;                //forbidden bit
1346     h->nal_ref_idc= src[0]>>5;
1347     h->nal_unit_type= src[0]&0x1F;
1348
1349     src++; length--;
1350 #if 0
1351     for(i=0; i<length; i++)
1352         printf("%2X ", src[i]);
1353 #endif
1354     for(i=0; i+1<length; i+=2){
1355         if(src[i]) continue;
1356         if(i>0 && src[i-1]==0) i--;
1357         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1358             if(src[i+2]!=3){
1359                 /* startcode, so we must be past the end */
1360                 length=i;
1361             }
1362             break;
1363         }
1364     }
1365
1366     if(i>=length-1){ //no escaped 0
1367         *dst_length= length;
1368         *consumed= length+1; //+1 for the header
1369         return src;
1370     }
1371
1372     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1373     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length);
1374     dst= h->rbsp_buffer[bufidx];
1375
1376     if (dst == NULL){
1377         return NULL;
1378     }
1379
1380 //printf("decoding esc\n");
1381     si=di=0;
1382     while(si<length){
1383         //remove escapes (very rare 1:2^22)
1384         if(si+2<length && src[si]==0 && src[si+1]==0 && src[si+2]<=3){
1385             if(src[si+2]==3){ //escape
1386                 dst[di++]= 0;
1387                 dst[di++]= 0;
1388                 si+=3;
1389                 continue;
1390             }else //next start code
1391                 break;
1392         }
1393
1394         dst[di++]= src[si++];
1395     }
1396
1397     *dst_length= di;
1398     *consumed= si + 1;//+1 for the header
1399 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1400     return dst;
1401 }
1402
1403 /**
1404  * identifies the exact end of the bitstream
1405  * @return the length of the trailing, or 0 if damaged
1406  */
1407 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src){
1408     int v= *src;
1409     int r;
1410
1411     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1412
1413     for(r=1; r<9; r++){
1414         if(v&1) return r;
1415         v>>=1;
1416     }
1417     return 0;
1418 }
1419
1420 /**
1421  * IDCT transforms the 16 dc values and dequantizes them.
1422  * @param qp quantization parameter
1423  */
1424 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1425 #define stride 16
1426     int i;
1427     int temp[16]; //FIXME check if this is a good idea
1428     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1429     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1430
1431 //memset(block, 64, 2*256);
1432 //return;
1433     for(i=0; i<4; i++){
1434         const int offset= y_offset[i];
1435         const int z0= block[offset+stride*0] + block[offset+stride*4];
1436         const int z1= block[offset+stride*0] - block[offset+stride*4];
1437         const int z2= block[offset+stride*1] - block[offset+stride*5];
1438         const int z3= block[offset+stride*1] + block[offset+stride*5];
1439
1440         temp[4*i+0]= z0+z3;
1441         temp[4*i+1]= z1+z2;
1442         temp[4*i+2]= z1-z2;
1443         temp[4*i+3]= z0-z3;
1444     }
1445
1446     for(i=0; i<4; i++){
1447         const int offset= x_offset[i];
1448         const int z0= temp[4*0+i] + temp[4*2+i];
1449         const int z1= temp[4*0+i] - temp[4*2+i];
1450         const int z2= temp[4*1+i] - temp[4*3+i];
1451         const int z3= temp[4*1+i] + temp[4*3+i];
1452
1453         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
1454         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1455         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1456         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1457     }
1458 }
1459
1460 #if 0
1461 /**
1462  * DCT transforms the 16 dc values.
1463  * @param qp quantization parameter ??? FIXME
1464  */
1465 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1466 //    const int qmul= dequant_coeff[qp][0];
1467     int i;
1468     int temp[16]; //FIXME check if this is a good idea
1469     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1470     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1471
1472     for(i=0; i<4; i++){
1473         const int offset= y_offset[i];
1474         const int z0= block[offset+stride*0] + block[offset+stride*4];
1475         const int z1= block[offset+stride*0] - block[offset+stride*4];
1476         const int z2= block[offset+stride*1] - block[offset+stride*5];
1477         const int z3= block[offset+stride*1] + block[offset+stride*5];
1478
1479         temp[4*i+0]= z0+z3;
1480         temp[4*i+1]= z1+z2;
1481         temp[4*i+2]= z1-z2;
1482         temp[4*i+3]= z0-z3;
1483     }
1484
1485     for(i=0; i<4; i++){
1486         const int offset= x_offset[i];
1487         const int z0= temp[4*0+i] + temp[4*2+i];
1488         const int z1= temp[4*0+i] - temp[4*2+i];
1489         const int z2= temp[4*1+i] - temp[4*3+i];
1490         const int z3= temp[4*1+i] + temp[4*3+i];
1491
1492         block[stride*0 +offset]= (z0 + z3)>>1;
1493         block[stride*2 +offset]= (z1 + z2)>>1;
1494         block[stride*8 +offset]= (z1 - z2)>>1;
1495         block[stride*10+offset]= (z0 - z3)>>1;
1496     }
1497 }
1498 #endif
1499
1500 #undef xStride
1501 #undef stride
1502
1503 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1504     const int stride= 16*2;
1505     const int xStride= 16;
1506     int a,b,c,d,e;
1507
1508     a= block[stride*0 + xStride*0];
1509     b= block[stride*0 + xStride*1];
1510     c= block[stride*1 + xStride*0];
1511     d= block[stride*1 + xStride*1];
1512
1513     e= a-b;
1514     a= a+b;
1515     b= c-d;
1516     c= c+d;
1517
1518     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1519     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1520     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1521     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1522 }
1523
1524 #if 0
1525 static void chroma_dc_dct_c(DCTELEM *block){
1526     const int stride= 16*2;
1527     const int xStride= 16;
1528     int a,b,c,d,e;
1529
1530     a= block[stride*0 + xStride*0];
1531     b= block[stride*0 + xStride*1];
1532     c= block[stride*1 + xStride*0];
1533     d= block[stride*1 + xStride*1];
1534
1535     e= a-b;
1536     a= a+b;
1537     b= c-d;
1538     c= c+d;
1539
1540     block[stride*0 + xStride*0]= (a+c);
1541     block[stride*0 + xStride*1]= (e+b);
1542     block[stride*1 + xStride*0]= (a-c);
1543     block[stride*1 + xStride*1]= (e-b);
1544 }
1545 #endif
1546
1547 /**
1548  * gets the chroma qp.
1549  */
1550 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1551     return h->pps.chroma_qp_table[t][qscale];
1552 }
1553
1554 //FIXME need to check that this does not overflow signed 32 bit for low qp, I am not sure, it's very close
1555 //FIXME check that gcc inlines this (and optimizes intra & separate_dc stuff away)
1556 static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int separate_dc){
1557     int i;
1558     const int * const quant_table= quant_coeff[qscale];
1559     const int bias= intra ? (1<<QUANT_SHIFT)/3 : (1<<QUANT_SHIFT)/6;
1560     const unsigned int threshold1= (1<<QUANT_SHIFT) - bias - 1;
1561     const unsigned int threshold2= (threshold1<<1);
1562     int last_non_zero;
1563
1564     if(separate_dc){
1565         if(qscale<=18){
1566             //avoid overflows
1567             const int dc_bias= intra ? (1<<(QUANT_SHIFT-2))/3 : (1<<(QUANT_SHIFT-2))/6;
1568             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT-2)) - dc_bias - 1;
1569             const unsigned int dc_threshold2= (dc_threshold1<<1);
1570
1571             int level= block[0]*quant_coeff[qscale+18][0];
1572             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1573                 if(level>0){
1574                     level= (dc_bias + level)>>(QUANT_SHIFT-2);
1575                     block[0]= level;
1576                 }else{
1577                     level= (dc_bias - level)>>(QUANT_SHIFT-2);
1578                     block[0]= -level;
1579                 }
1580 //                last_non_zero = i;
1581             }else{
1582                 block[0]=0;
1583             }
1584         }else{
1585             const int dc_bias= intra ? (1<<(QUANT_SHIFT+1))/3 : (1<<(QUANT_SHIFT+1))/6;
1586             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT+1)) - dc_bias - 1;
1587             const unsigned int dc_threshold2= (dc_threshold1<<1);
1588
1589             int level= block[0]*quant_table[0];
1590             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1591                 if(level>0){
1592                     level= (dc_bias + level)>>(QUANT_SHIFT+1);
1593                     block[0]= level;
1594                 }else{
1595                     level= (dc_bias - level)>>(QUANT_SHIFT+1);
1596                     block[0]= -level;
1597                 }
1598 //                last_non_zero = i;
1599             }else{
1600                 block[0]=0;
1601             }
1602         }
1603         last_non_zero= 0;
1604         i=1;
1605     }else{
1606         last_non_zero= -1;
1607         i=0;
1608     }
1609
1610     for(; i<16; i++){
1611         const int j= scantable[i];
1612         int level= block[j]*quant_table[j];
1613
1614 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
1615 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
1616         if(((unsigned)(level+threshold1))>threshold2){
1617             if(level>0){
1618                 level= (bias + level)>>QUANT_SHIFT;
1619                 block[j]= level;
1620             }else{
1621                 level= (bias - level)>>QUANT_SHIFT;
1622                 block[j]= -level;
1623             }
1624             last_non_zero = i;
1625         }else{
1626             block[j]=0;
1627         }
1628     }
1629
1630     return last_non_zero;
1631 }
1632
1633 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1634                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1635                            int src_x_offset, int src_y_offset,
1636                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1637     MpegEncContext * const s = &h->s;
1638     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1639     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1640     const int luma_xy= (mx&3) + ((my&3)<<2);
1641     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1642     uint8_t * src_cb, * src_cr;
1643     int extra_width= h->emu_edge_width;
1644     int extra_height= h->emu_edge_height;
1645     int emu=0;
1646     const int full_mx= mx>>2;
1647     const int full_my= my>>2;
1648     const int pic_width  = 16*s->mb_width;
1649     const int pic_height = 16*s->mb_height >> MB_FIELD;
1650
1651     if(!pic->data[0]) //FIXME this is unacceptable, some sensible error concealment must be done for missing reference frames
1652         return;
1653
1654     if(mx&7) extra_width -= 3;
1655     if(my&7) extra_height -= 3;
1656
1657     if(   full_mx < 0-extra_width
1658        || full_my < 0-extra_height
1659        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1660        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1661         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);
1662             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1663         emu=1;
1664     }
1665
1666     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1667     if(!square){
1668         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1669     }
1670
1671     if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1672
1673     if(MB_FIELD){
1674         // chroma offset when predicting from a field of opposite parity
1675         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1676         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1677     }
1678     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1679     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1680
1681     if(emu){
1682         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);
1683             src_cb= s->edge_emu_buffer;
1684     }
1685     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1686
1687     if(emu){
1688         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);
1689             src_cr= s->edge_emu_buffer;
1690     }
1691     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1692 }
1693
1694 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1695                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1696                            int x_offset, int y_offset,
1697                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1698                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1699                            int list0, int list1){
1700     MpegEncContext * const s = &h->s;
1701     qpel_mc_func *qpix_op=  qpix_put;
1702     h264_chroma_mc_func chroma_op= chroma_put;
1703
1704     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1705     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1706     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1707     x_offset += 8*s->mb_x;
1708     y_offset += 8*(s->mb_y >> MB_FIELD);
1709
1710     if(list0){
1711         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1712         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1713                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1714                            qpix_op, chroma_op);
1715
1716         qpix_op=  qpix_avg;
1717         chroma_op= chroma_avg;
1718     }
1719
1720     if(list1){
1721         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1722         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1723                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1724                            qpix_op, chroma_op);
1725     }
1726 }
1727
1728 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1729                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1730                            int x_offset, int y_offset,
1731                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1732                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1733                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1734                            int list0, int list1){
1735     MpegEncContext * const s = &h->s;
1736
1737     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1738     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1739     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1740     x_offset += 8*s->mb_x;
1741     y_offset += 8*(s->mb_y >> MB_FIELD);
1742
1743     if(list0 && list1){
1744         /* don't optimize for luma-only case, since B-frames usually
1745          * use implicit weights => chroma too. */
1746         uint8_t *tmp_cb = s->obmc_scratchpad;
1747         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1748         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1749         int refn0 = h->ref_cache[0][ scan8[n] ];
1750         int refn1 = h->ref_cache[1][ scan8[n] ];
1751
1752         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1753                     dest_y, dest_cb, dest_cr,
1754                     x_offset, y_offset, qpix_put, chroma_put);
1755         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1756                     tmp_y, tmp_cb, tmp_cr,
1757                     x_offset, y_offset, qpix_put, chroma_put);
1758
1759         if(h->use_weight == 2){
1760             int weight0 = h->implicit_weight[refn0][refn1];
1761             int weight1 = 64 - weight0;
1762             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1763             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1764             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1765         }else{
1766             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1767                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1768                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1769             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1770                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1771                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1772             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1773                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1774                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1775         }
1776     }else{
1777         int list = list1 ? 1 : 0;
1778         int refn = h->ref_cache[list][ scan8[n] ];
1779         Picture *ref= &h->ref_list[list][refn];
1780         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1781                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1782                     qpix_put, chroma_put);
1783
1784         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1785                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1786         if(h->use_weight_chroma){
1787             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1788                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1789             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1790                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1791         }
1792     }
1793 }
1794
1795 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1796                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1797                            int x_offset, int y_offset,
1798                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1799                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1800                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1801                            int list0, int list1){
1802     if((h->use_weight==2 && list0 && list1
1803         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1804        || h->use_weight==1)
1805         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1806                          x_offset, y_offset, qpix_put, chroma_put,
1807                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1808     else
1809         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1810                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1811 }
1812
1813 static inline void prefetch_motion(H264Context *h, int list){
1814     /* fetch pixels for estimated mv 4 macroblocks ahead
1815      * optimized for 64byte cache lines */
1816     MpegEncContext * const s = &h->s;
1817     const int refn = h->ref_cache[list][scan8[0]];
1818     if(refn >= 0){
1819         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1820         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1821         uint8_t **src= h->ref_list[list][refn].data;
1822         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1823         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1824         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1825         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1826     }
1827 }
1828
1829 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1830                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1831                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1832                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1833     MpegEncContext * const s = &h->s;
1834     const int mb_xy= h->mb_xy;
1835     const int mb_type= s->current_picture.mb_type[mb_xy];
1836
1837     assert(IS_INTER(mb_type));
1838
1839     prefetch_motion(h, 0);
1840
1841     if(IS_16X16(mb_type)){
1842         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1843                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1844                 &weight_op[0], &weight_avg[0],
1845                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1846     }else if(IS_16X8(mb_type)){
1847         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1848                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1849                 &weight_op[1], &weight_avg[1],
1850                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1851         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1852                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1853                 &weight_op[1], &weight_avg[1],
1854                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1855     }else if(IS_8X16(mb_type)){
1856         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1857                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1858                 &weight_op[2], &weight_avg[2],
1859                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1860         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1861                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1862                 &weight_op[2], &weight_avg[2],
1863                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1864     }else{
1865         int i;
1866
1867         assert(IS_8X8(mb_type));
1868
1869         for(i=0; i<4; i++){
1870             const int sub_mb_type= h->sub_mb_type[i];
1871             const int n= 4*i;
1872             int x_offset= (i&1)<<2;
1873             int y_offset= (i&2)<<1;
1874
1875             if(IS_SUB_8X8(sub_mb_type)){
1876                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1877                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1878                     &weight_op[3], &weight_avg[3],
1879                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1880             }else if(IS_SUB_8X4(sub_mb_type)){
1881                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1882                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1883                     &weight_op[4], &weight_avg[4],
1884                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1885                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1886                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1887                     &weight_op[4], &weight_avg[4],
1888                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1889             }else if(IS_SUB_4X8(sub_mb_type)){
1890                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1891                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1892                     &weight_op[5], &weight_avg[5],
1893                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1894                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1895                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1896                     &weight_op[5], &weight_avg[5],
1897                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1898             }else{
1899                 int j;
1900                 assert(IS_SUB_4X4(sub_mb_type));
1901                 for(j=0; j<4; j++){
1902                     int sub_x_offset= x_offset + 2*(j&1);
1903                     int sub_y_offset= y_offset +   (j&2);
1904                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1905                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1906                         &weight_op[6], &weight_avg[6],
1907                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1908                 }
1909             }
1910         }
1911     }
1912
1913     prefetch_motion(h, 1);
1914 }
1915
1916 static av_cold void decode_init_vlc(void){
1917     static int done = 0;
1918
1919     if (!done) {
1920         int i;
1921         int offset;
1922         done = 1;
1923
1924         chroma_dc_coeff_token_vlc.table = chroma_dc_coeff_token_vlc_table;
1925         chroma_dc_coeff_token_vlc.table_allocated = chroma_dc_coeff_token_vlc_table_size;
1926         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1927                  &chroma_dc_coeff_token_len [0], 1, 1,
1928                  &chroma_dc_coeff_token_bits[0], 1, 1,
1929                  INIT_VLC_USE_NEW_STATIC);
1930
1931         offset = 0;
1932         for(i=0; i<4; i++){
1933             coeff_token_vlc[i].table = coeff_token_vlc_tables+offset;
1934             coeff_token_vlc[i].table_allocated = coeff_token_vlc_tables_size[i];
1935             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1936                      &coeff_token_len [i][0], 1, 1,
1937                      &coeff_token_bits[i][0], 1, 1,
1938                      INIT_VLC_USE_NEW_STATIC);
1939             offset += coeff_token_vlc_tables_size[i];
1940         }
1941         /*
1942          * This is a one time safety check to make sure that
1943          * the packed static coeff_token_vlc table sizes
1944          * were initialized correctly.
1945          */
1946         assert(offset == sizeof(coeff_token_vlc_tables)/(sizeof(VLC_TYPE)*2));
1947
1948         for(i=0; i<3; i++){
1949             chroma_dc_total_zeros_vlc[i].table = chroma_dc_total_zeros_vlc_tables[i];
1950             chroma_dc_total_zeros_vlc[i].table_allocated = chroma_dc_total_zeros_vlc_tables_size;
1951             init_vlc(&chroma_dc_total_zeros_vlc[i],
1952                      CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1953                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1954                      &chroma_dc_total_zeros_bits[i][0], 1, 1,
1955                      INIT_VLC_USE_NEW_STATIC);
1956         }
1957         for(i=0; i<15; i++){
1958             total_zeros_vlc[i].table = total_zeros_vlc_tables[i];
1959             total_zeros_vlc[i].table_allocated = total_zeros_vlc_tables_size;
1960             init_vlc(&total_zeros_vlc[i],
1961                      TOTAL_ZEROS_VLC_BITS, 16,
1962                      &total_zeros_len [i][0], 1, 1,
1963                      &total_zeros_bits[i][0], 1, 1,
1964                      INIT_VLC_USE_NEW_STATIC);
1965         }
1966
1967         for(i=0; i<6; i++){
1968             run_vlc[i].table = run_vlc_tables[i];
1969             run_vlc[i].table_allocated = run_vlc_tables_size;
1970             init_vlc(&run_vlc[i],
1971                      RUN_VLC_BITS, 7,
1972                      &run_len [i][0], 1, 1,
1973                      &run_bits[i][0], 1, 1,
1974                      INIT_VLC_USE_NEW_STATIC);
1975         }
1976         run7_vlc.table = run7_vlc_table,
1977         run7_vlc.table_allocated = run7_vlc_table_size;
1978         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
1979                  &run_len [6][0], 1, 1,
1980                  &run_bits[6][0], 1, 1,
1981                  INIT_VLC_USE_NEW_STATIC);
1982     }
1983 }
1984
1985 static void free_tables(H264Context *h){
1986     int i;
1987     H264Context *hx;
1988     av_freep(&h->intra4x4_pred_mode);
1989     av_freep(&h->chroma_pred_mode_table);
1990     av_freep(&h->cbp_table);
1991     av_freep(&h->mvd_table[0]);
1992     av_freep(&h->mvd_table[1]);
1993     av_freep(&h->direct_table);
1994     av_freep(&h->non_zero_count);
1995     av_freep(&h->slice_table_base);
1996     h->slice_table= NULL;
1997
1998     av_freep(&h->mb2b_xy);
1999     av_freep(&h->mb2b8_xy);
2000
2001     for(i = 0; i < MAX_SPS_COUNT; i++)
2002         av_freep(h->sps_buffers + i);
2003
2004     for(i = 0; i < MAX_PPS_COUNT; i++)
2005         av_freep(h->pps_buffers + i);
2006
2007     for(i = 0; i < h->s.avctx->thread_count; i++) {
2008         hx = h->thread_context[i];
2009         if(!hx) continue;
2010         av_freep(&hx->top_borders[1]);
2011         av_freep(&hx->top_borders[0]);
2012         av_freep(&hx->s.obmc_scratchpad);
2013     }
2014 }
2015
2016 static void init_dequant8_coeff_table(H264Context *h){
2017     int i,q,x;
2018     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2019     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2020     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2021
2022     for(i=0; i<2; i++ ){
2023         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2024             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2025             break;
2026         }
2027
2028         for(q=0; q<52; q++){
2029             int shift = ff_div6[q];
2030             int idx = ff_rem6[q];
2031             for(x=0; x<64; x++)
2032                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2033                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2034                     h->pps.scaling_matrix8[i][x]) << shift;
2035         }
2036     }
2037 }
2038
2039 static void init_dequant4_coeff_table(H264Context *h){
2040     int i,j,q,x;
2041     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2042     for(i=0; i<6; i++ ){
2043         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2044         for(j=0; j<i; j++){
2045             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2046                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2047                 break;
2048             }
2049         }
2050         if(j<i)
2051             continue;
2052
2053         for(q=0; q<52; q++){
2054             int shift = ff_div6[q] + 2;
2055             int idx = ff_rem6[q];
2056             for(x=0; x<16; x++)
2057                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2058                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2059                     h->pps.scaling_matrix4[i][x]) << shift;
2060         }
2061     }
2062 }
2063
2064 static void init_dequant_tables(H264Context *h){
2065     int i,x;
2066     init_dequant4_coeff_table(h);
2067     if(h->pps.transform_8x8_mode)
2068         init_dequant8_coeff_table(h);
2069     if(h->sps.transform_bypass){
2070         for(i=0; i<6; i++)
2071             for(x=0; x<16; x++)
2072                 h->dequant4_coeff[i][0][x] = 1<<6;
2073         if(h->pps.transform_8x8_mode)
2074             for(i=0; i<2; i++)
2075                 for(x=0; x<64; x++)
2076                     h->dequant8_coeff[i][0][x] = 1<<6;
2077     }
2078 }
2079
2080
2081 /**
2082  * allocates tables.
2083  * needs width/height
2084  */
2085 static int alloc_tables(H264Context *h){
2086     MpegEncContext * const s = &h->s;
2087     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2088     int x,y;
2089
2090     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2091
2092     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2093     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(uint8_t))
2094     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2095
2096     CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2097     CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2098     CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2099     CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2100
2101     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(uint8_t));
2102     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2103
2104     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2105     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2106     for(y=0; y<s->mb_height; y++){
2107         for(x=0; x<s->mb_width; x++){
2108             const int mb_xy= x + y*s->mb_stride;
2109             const int b_xy = 4*x + 4*y*h->b_stride;
2110             const int b8_xy= 2*x + 2*y*h->b8_stride;
2111
2112             h->mb2b_xy [mb_xy]= b_xy;
2113             h->mb2b8_xy[mb_xy]= b8_xy;
2114         }
2115     }
2116
2117     s->obmc_scratchpad = NULL;
2118
2119     if(!h->dequant4_coeff[0])
2120         init_dequant_tables(h);
2121
2122     return 0;
2123 fail:
2124     free_tables(h);
2125     return -1;
2126 }
2127
2128 /**
2129  * Mimic alloc_tables(), but for every context thread.
2130  */
2131 static void clone_tables(H264Context *dst, H264Context *src){
2132     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2133     dst->non_zero_count           = src->non_zero_count;
2134     dst->slice_table              = src->slice_table;
2135     dst->cbp_table                = src->cbp_table;
2136     dst->mb2b_xy                  = src->mb2b_xy;
2137     dst->mb2b8_xy                 = src->mb2b8_xy;
2138     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2139     dst->mvd_table[0]             = src->mvd_table[0];
2140     dst->mvd_table[1]             = src->mvd_table[1];
2141     dst->direct_table             = src->direct_table;
2142
2143     dst->s.obmc_scratchpad = NULL;
2144     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2145 }
2146
2147 /**
2148  * Init context
2149  * Allocate buffers which are not shared amongst multiple threads.
2150  */
2151 static int context_init(H264Context *h){
2152     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2153     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2154
2155     return 0;
2156 fail:
2157     return -1; // free_tables will clean up for us
2158 }
2159
2160 static av_cold void common_init(H264Context *h){
2161     MpegEncContext * const s = &h->s;
2162
2163     s->width = s->avctx->width;
2164     s->height = s->avctx->height;
2165     s->codec_id= s->avctx->codec->id;
2166
2167     ff_h264_pred_init(&h->hpc, s->codec_id);
2168
2169     h->dequant_coeff_pps= -1;
2170     s->unrestricted_mv=1;
2171     s->decode=1; //FIXME
2172
2173     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2174     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2175 }
2176
2177 static av_cold int decode_init(AVCodecContext *avctx){
2178     H264Context *h= avctx->priv_data;
2179     MpegEncContext * const s = &h->s;
2180
2181     MPV_decode_defaults(s);
2182
2183     s->avctx = avctx;
2184     common_init(h);
2185
2186     s->out_format = FMT_H264;
2187     s->workaround_bugs= avctx->workaround_bugs;
2188
2189     // set defaults
2190 //    s->decode_mb= ff_h263_decode_mb;
2191     s->quarter_sample = 1;
2192     s->low_delay= 1;
2193
2194     if(avctx->codec_id == CODEC_ID_SVQ3)
2195         avctx->pix_fmt= PIX_FMT_YUVJ420P;
2196     else
2197         avctx->pix_fmt= PIX_FMT_YUV420P;
2198
2199     decode_init_vlc();
2200
2201     if(avctx->extradata_size > 0 && avctx->extradata &&
2202        *(char *)avctx->extradata == 1){
2203         h->is_avc = 1;
2204         h->got_avcC = 0;
2205     } else {
2206         h->is_avc = 0;
2207     }
2208
2209     h->thread_context[0] = h;
2210     h->outputed_poc = INT_MIN;
2211     return 0;
2212 }
2213
2214 static int frame_start(H264Context *h){
2215     MpegEncContext * const s = &h->s;
2216     int i;
2217
2218     if(MPV_frame_start(s, s->avctx) < 0)
2219         return -1;
2220     ff_er_frame_start(s);
2221     /*
2222      * MPV_frame_start uses pict_type to derive key_frame.
2223      * This is incorrect for H.264; IDR markings must be used.
2224      * Zero here; IDR markings per slice in frame or fields are ORed in later.
2225      * See decode_nal_units().
2226      */
2227     s->current_picture_ptr->key_frame= 0;
2228
2229     assert(s->linesize && s->uvlinesize);
2230
2231     for(i=0; i<16; i++){
2232         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2233         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2234     }
2235     for(i=0; i<4; i++){
2236         h->block_offset[16+i]=
2237         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2238         h->block_offset[24+16+i]=
2239         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2240     }
2241
2242     /* can't be in alloc_tables because linesize isn't known there.
2243      * FIXME: redo bipred weight to not require extra buffer? */
2244     for(i = 0; i < s->avctx->thread_count; i++)
2245         if(!h->thread_context[i]->s.obmc_scratchpad)
2246             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2247
2248     /* some macroblocks will be accessed before they're available */
2249     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2250         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(uint8_t));
2251
2252 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2253
2254     // We mark the current picture as non-reference after allocating it, so
2255     // that if we break out due to an error it can be released automatically
2256     // in the next MPV_frame_start().
2257     // SVQ3 as well as most other codecs have only last/next/current and thus
2258     // get released even with set reference, besides SVQ3 and others do not
2259     // mark frames as reference later "naturally".
2260     if(s->codec_id != CODEC_ID_SVQ3)
2261         s->current_picture_ptr->reference= 0;
2262
2263     s->current_picture_ptr->field_poc[0]=
2264     s->current_picture_ptr->field_poc[1]= INT_MAX;
2265     assert(s->current_picture_ptr->long_ref==0);
2266
2267     return 0;
2268 }
2269
2270 static inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int simple){
2271     MpegEncContext * const s = &h->s;
2272     int i;
2273
2274     src_y  -=   linesize;
2275     src_cb -= uvlinesize;
2276     src_cr -= uvlinesize;
2277
2278     // There are two lines saved, the line above the the top macroblock of a pair,
2279     // and the line above the bottom macroblock
2280     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2281     for(i=1; i<17; i++){
2282         h->left_border[i]= src_y[15+i*  linesize];
2283     }
2284
2285     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2286     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2287
2288     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2289         h->left_border[17  ]= h->top_borders[0][s->mb_x][16+7];
2290         h->left_border[17+9]= h->top_borders[0][s->mb_x][24+7];
2291         for(i=1; i<9; i++){
2292             h->left_border[i+17  ]= src_cb[7+i*uvlinesize];
2293             h->left_border[i+17+9]= src_cr[7+i*uvlinesize];
2294         }
2295         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2296         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2297     }
2298 }
2299
2300 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){
2301     MpegEncContext * const s = &h->s;
2302     int temp8, i;
2303     uint64_t temp64;
2304     int deblock_left;
2305     int deblock_top;
2306     int mb_xy;
2307
2308     if(h->deblocking_filter == 2) {
2309         mb_xy = h->mb_xy;
2310         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2311         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2312     } else {
2313         deblock_left = (s->mb_x > 0);
2314         deblock_top =  (s->mb_y > 0);
2315     }
2316
2317     src_y  -=   linesize + 1;
2318     src_cb -= uvlinesize + 1;
2319     src_cr -= uvlinesize + 1;
2320
2321 #define XCHG(a,b,t,xchg)\
2322 t= a;\
2323 if(xchg)\
2324     a= b;\
2325 b= t;
2326
2327     if(deblock_left){
2328         for(i = !deblock_top; i<17; i++){
2329             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2330         }
2331     }
2332
2333     if(deblock_top){
2334         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2335         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2336         if(s->mb_x+1 < s->mb_width){
2337             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2338         }
2339     }
2340
2341     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2342         if(deblock_left){
2343             for(i = !deblock_top; i<9; i++){
2344                 XCHG(h->left_border[i+17  ], src_cb[i*uvlinesize], temp8, xchg);
2345                 XCHG(h->left_border[i+17+9], src_cr[i*uvlinesize], temp8, xchg);
2346             }
2347         }
2348         if(deblock_top){
2349             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2350             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2351         }
2352     }
2353 }
2354
2355 static inline void backup_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize){
2356     MpegEncContext * const s = &h->s;
2357     int i;
2358
2359     src_y  -= 2 *   linesize;
2360     src_cb -= 2 * uvlinesize;
2361     src_cr -= 2 * uvlinesize;
2362
2363     // There are two lines saved, the line above the the top macroblock of a pair,
2364     // and the line above the bottom macroblock
2365     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2366     h->left_border[1]= h->top_borders[1][s->mb_x][15];
2367     for(i=2; i<34; i++){
2368         h->left_border[i]= src_y[15+i*  linesize];
2369     }
2370
2371     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  32*linesize);
2372     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+32*linesize);
2373     *(uint64_t*)(h->top_borders[1][s->mb_x]+0)= *(uint64_t*)(src_y +  33*linesize);
2374     *(uint64_t*)(h->top_borders[1][s->mb_x]+8)= *(uint64_t*)(src_y +8+33*linesize);
2375
2376     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2377         h->left_border[34     ]= h->top_borders[0][s->mb_x][16+7];
2378         h->left_border[34+   1]= h->top_borders[1][s->mb_x][16+7];
2379         h->left_border[34+18  ]= h->top_borders[0][s->mb_x][24+7];
2380         h->left_border[34+18+1]= h->top_borders[1][s->mb_x][24+7];
2381         for(i=2; i<18; i++){
2382             h->left_border[i+34   ]= src_cb[7+i*uvlinesize];
2383             h->left_border[i+34+18]= src_cr[7+i*uvlinesize];
2384         }
2385         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+16*uvlinesize);
2386         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+16*uvlinesize);
2387         *(uint64_t*)(h->top_borders[1][s->mb_x]+16)= *(uint64_t*)(src_cb+17*uvlinesize);
2388         *(uint64_t*)(h->top_borders[1][s->mb_x]+24)= *(uint64_t*)(src_cr+17*uvlinesize);
2389     }
2390 }
2391
2392 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){
2393     MpegEncContext * const s = &h->s;
2394     int temp8, i;
2395     uint64_t temp64;
2396     int deblock_left = (s->mb_x > 0);
2397     int deblock_top  = (s->mb_y > 1);
2398
2399     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);
2400
2401     src_y  -= 2 *   linesize + 1;
2402     src_cb -= 2 * uvlinesize + 1;
2403     src_cr -= 2 * uvlinesize + 1;
2404
2405 #define XCHG(a,b,t,xchg)\
2406 t= a;\
2407 if(xchg)\
2408     a= b;\
2409 b= t;
2410
2411     if(deblock_left){
2412         for(i = (!deblock_top)<<1; i<34; i++){
2413             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2414         }
2415     }
2416
2417     if(deblock_top){
2418         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2419         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2420         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+0), *(uint64_t*)(src_y +1 +linesize), temp64, xchg);
2421         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+8), *(uint64_t*)(src_y +9 +linesize), temp64, 1);
2422         if(s->mb_x+1 < s->mb_width){
2423             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2424             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x+1]), *(uint64_t*)(src_y +17 +linesize), temp64, 1);
2425         }
2426     }
2427
2428     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2429         if(deblock_left){
2430             for(i = (!deblock_top) << 1; i<18; i++){
2431                 XCHG(h->left_border[i+34   ], src_cb[i*uvlinesize], temp8, xchg);
2432                 XCHG(h->left_border[i+34+18], src_cr[i*uvlinesize], temp8, xchg);
2433             }
2434         }
2435         if(deblock_top){
2436             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2437             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2438             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+16), *(uint64_t*)(src_cb+1 +uvlinesize), temp64, 1);
2439             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+24), *(uint64_t*)(src_cr+1 +uvlinesize), temp64, 1);
2440         }
2441     }
2442 }
2443
2444 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2445     MpegEncContext * const s = &h->s;
2446     const int mb_x= s->mb_x;
2447     const int mb_y= s->mb_y;
2448     const int mb_xy= h->mb_xy;
2449     const int mb_type= s->current_picture.mb_type[mb_xy];
2450     uint8_t  *dest_y, *dest_cb, *dest_cr;
2451     int linesize, uvlinesize /*dct_offset*/;
2452     int i;
2453     int *block_offset = &h->block_offset[0];
2454     const unsigned int bottom = mb_y & 1;
2455     const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass), is_h264 = (simple || s->codec_id == CODEC_ID_H264);
2456     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2457     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2458
2459     dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2460     dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2461     dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2462
2463     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2464     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2465
2466     if (!simple && MB_FIELD) {
2467         linesize   = h->mb_linesize   = s->linesize * 2;
2468         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2469         block_offset = &h->block_offset[24];
2470         if(mb_y&1){ //FIXME move out of this function?
2471             dest_y -= s->linesize*15;
2472             dest_cb-= s->uvlinesize*7;
2473             dest_cr-= s->uvlinesize*7;
2474         }
2475         if(FRAME_MBAFF) {
2476             int list;
2477             for(list=0; list<h->list_count; list++){
2478                 if(!USES_LIST(mb_type, list))
2479                     continue;
2480                 if(IS_16X16(mb_type)){
2481                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2482                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2483                 }else{
2484                     for(i=0; i<16; i+=4){
2485                         //FIXME can refs be smaller than 8x8 when !direct_8x8_inference ?
2486                         int ref = h->ref_cache[list][scan8[i]];
2487                         if(ref >= 0)
2488                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2489                     }
2490                 }
2491             }
2492         }
2493     } else {
2494         linesize   = h->mb_linesize   = s->linesize;
2495         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2496 //        dct_offset = s->linesize * 16;
2497     }
2498
2499     if(transform_bypass){
2500         idct_dc_add =
2501         idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2502     }else if(IS_8x8DCT(mb_type)){
2503         idct_dc_add = s->dsp.h264_idct8_dc_add;
2504         idct_add = s->dsp.h264_idct8_add;
2505     }else{
2506         idct_dc_add = s->dsp.h264_idct_dc_add;
2507         idct_add = s->dsp.h264_idct_add;
2508     }
2509
2510     if(!simple && FRAME_MBAFF && h->deblocking_filter && IS_INTRA(mb_type)
2511        && (!bottom || !IS_INTRA(s->current_picture.mb_type[mb_xy-s->mb_stride]))){
2512         int mbt_y = mb_y&~1;
2513         uint8_t *top_y  = s->current_picture.data[0] + (mbt_y * 16* s->linesize  ) + mb_x * 16;
2514         uint8_t *top_cb = s->current_picture.data[1] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2515         uint8_t *top_cr = s->current_picture.data[2] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2516         xchg_pair_border(h, top_y, top_cb, top_cr, s->linesize, s->uvlinesize, 1);
2517     }
2518
2519     if (!simple && IS_INTRA_PCM(mb_type)) {
2520         for (i=0; i<16; i++) {
2521             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
2522         }
2523         for (i=0; i<8; i++) {
2524             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
2525             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
2526         }
2527     } else {
2528         if(IS_INTRA(mb_type)){
2529             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2530                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2531
2532             if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2533                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2534                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2535             }
2536
2537             if(IS_INTRA4x4(mb_type)){
2538                 if(simple || !s->encoding){
2539                     if(IS_8x8DCT(mb_type)){
2540                         for(i=0; i<16; i+=4){
2541                             uint8_t * const ptr= dest_y + block_offset[i];
2542                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2543                             const int nnz = h->non_zero_count_cache[ scan8[i] ];
2544                             h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2545                                                    (h->topright_samples_available<<i)&0x4000, linesize);
2546                             if(nnz){
2547                                 if(nnz == 1 && h->mb[i*16])
2548                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2549                                 else
2550                                     idct_add(ptr, h->mb + i*16, linesize);
2551                             }
2552                         }
2553                     }else
2554                     for(i=0; i<16; i++){
2555                         uint8_t * const ptr= dest_y + block_offset[i];
2556                         uint8_t *topright;
2557                         const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2558                         int nnz, tr;
2559
2560                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2561                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2562                             assert(mb_y || linesize <= block_offset[i]);
2563                             if(!topright_avail){
2564                                 tr= ptr[3 - linesize]*0x01010101;
2565                                 topright= (uint8_t*) &tr;
2566                             }else
2567                                 topright= ptr + 4 - linesize;
2568                         }else
2569                             topright= NULL;
2570
2571                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2572                         nnz = h->non_zero_count_cache[ scan8[i] ];
2573                         if(nnz){
2574                             if(is_h264){
2575                                 if(nnz == 1 && h->mb[i*16])
2576                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2577                                 else
2578                                     idct_add(ptr, h->mb + i*16, linesize);
2579                             }else
2580                                 svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2581                         }
2582                     }
2583                 }
2584             }else{
2585                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2586                 if(is_h264){
2587                     if(!transform_bypass)
2588                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2589                 }else
2590                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2591             }
2592             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2593                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2594         }else if(is_h264){
2595             hl_motion(h, dest_y, dest_cb, dest_cr,
2596                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2597                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2598                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2599         }
2600
2601
2602         if(!IS_INTRA4x4(mb_type)){
2603             if(is_h264){
2604                 if(IS_INTRA16x16(mb_type)){
2605                     for(i=0; i<16; i++){
2606                         if(h->non_zero_count_cache[ scan8[i] ])
2607                             idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2608                         else if(h->mb[i*16])
2609                             idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2610                     }
2611                 }else{
2612                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2613                     for(i=0; i<16; i+=di){
2614                         int nnz = h->non_zero_count_cache[ scan8[i] ];
2615                         if(nnz){
2616                             if(nnz==1 && h->mb[i*16])
2617                                 idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2618                             else
2619                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2620                         }
2621                     }
2622                 }
2623             }else{
2624                 for(i=0; i<16; i++){
2625                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2626                         uint8_t * const ptr= dest_y + block_offset[i];
2627                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2628                     }
2629                 }
2630             }
2631         }
2632
2633         if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2634             uint8_t *dest[2] = {dest_cb, dest_cr};
2635             if(transform_bypass){
2636                 idct_add = idct_dc_add = s->dsp.add_pixels4;
2637             }else{
2638                 idct_add = s->dsp.h264_idct_add;
2639                 idct_dc_add = s->dsp.h264_idct_dc_add;
2640                 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]);
2641                 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]);
2642             }
2643             if(is_h264){
2644                 for(i=16; i<16+8; i++){
2645                     if(h->non_zero_count_cache[ scan8[i] ])
2646                         idct_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2647                     else if(h->mb[i*16])
2648                         idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2649                 }
2650             }else{
2651                 for(i=16; i<16+8; i++){
2652                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2653                         uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2654                         svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2655                     }
2656                 }
2657             }
2658         }
2659     }
2660     if(h->deblocking_filter) {
2661         if (!simple && FRAME_MBAFF) {
2662             //FIXME try deblocking one mb at a time?
2663             // the reduction in load/storing mvs and such might outweigh the extra backup/xchg_border
2664             const int mb_y = s->mb_y - 1;
2665             uint8_t  *pair_dest_y, *pair_dest_cb, *pair_dest_cr;
2666             const int mb_xy= mb_x + mb_y*s->mb_stride;
2667             const int mb_type_top   = s->current_picture.mb_type[mb_xy];
2668             const int mb_type_bottom= s->current_picture.mb_type[mb_xy+s->mb_stride];
2669             if (!bottom) return;
2670             pair_dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2671             pair_dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2672             pair_dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2673
2674             if(IS_INTRA(mb_type_top | mb_type_bottom))
2675                 xchg_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize, 0);
2676
2677             backup_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize);
2678             // deblock a pair
2679             // top
2680             s->mb_y--; h->mb_xy -= s->mb_stride;
2681             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);
2682             fill_caches(h, mb_type_top, 1); //FIXME don't fill stuff which isn't used by filter_mb
2683             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2684             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2685             filter_mb(h, mb_x, mb_y, pair_dest_y, pair_dest_cb, pair_dest_cr, linesize, uvlinesize);
2686             // bottom
2687             s->mb_y++; h->mb_xy += s->mb_stride;
2688             tprintf(h->s.avctx, "call mbaff filter_mb\n");
2689             fill_caches(h, mb_type_bottom, 1); //FIXME don't fill stuff which isn't used by filter_mb
2690             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2691             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2692             filter_mb(h, mb_x, mb_y+1, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2693         } else {
2694             tprintf(h->s.avctx, "call filter_mb\n");
2695             backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2696             fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2697             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2698             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2699             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2700         }
2701     }
2702 }
2703
2704 /**
2705  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2706  */
2707 static void hl_decode_mb_simple(H264Context *h){
2708     hl_decode_mb_internal(h, 1);
2709 }
2710
2711 /**
2712  * Process a macroblock; this handles edge cases, such as interlacing.
2713  */
2714 static void av_noinline hl_decode_mb_complex(H264Context *h){
2715     hl_decode_mb_internal(h, 0);
2716 }
2717
2718 static void hl_decode_mb(H264Context *h){
2719     MpegEncContext * const s = &h->s;
2720     const int mb_xy= h->mb_xy;
2721     const int mb_type= s->current_picture.mb_type[mb_xy];
2722     int is_complex = FRAME_MBAFF || MB_FIELD || IS_INTRA_PCM(mb_type) || s->codec_id != CODEC_ID_H264 ||
2723                     (ENABLE_GRAY && (s->flags&CODEC_FLAG_GRAY)) || (ENABLE_H264_ENCODER && s->encoding) || ENABLE_SMALL;
2724
2725     if(ENABLE_H264_ENCODER && !s->decode)
2726         return;
2727
2728     if (is_complex)
2729         hl_decode_mb_complex(h);
2730     else hl_decode_mb_simple(h);
2731 }
2732
2733 static void pic_as_field(Picture *pic, const int parity){
2734     int i;
2735     for (i = 0; i < 4; ++i) {
2736         if (parity == PICT_BOTTOM_FIELD)
2737             pic->data[i] += pic->linesize[i];
2738         pic->reference = parity;
2739         pic->linesize[i] *= 2;
2740     }
2741     pic->poc= pic->field_poc[parity == PICT_BOTTOM_FIELD];
2742 }
2743
2744 static int split_field_copy(Picture *dest, Picture *src,
2745                             int parity, int id_add){
2746     int match = !!(src->reference & parity);
2747
2748     if (match) {
2749         *dest = *src;
2750         if(parity != PICT_FRAME){
2751             pic_as_field(dest, parity);
2752             dest->pic_id *= 2;
2753             dest->pic_id += id_add;
2754         }
2755     }
2756
2757     return match;
2758 }
2759
2760 static int build_def_list(Picture *def, Picture **in, int len, int is_long, int sel){
2761     int i[2]={0};
2762     int index=0;
2763
2764     while(i[0]<len || i[1]<len){
2765         while(i[0]<len && !(in[ i[0] ] && (in[ i[0] ]->reference & sel)))
2766             i[0]++;
2767         while(i[1]<len && !(in[ i[1] ] && (in[ i[1] ]->reference & (sel^3))))
2768             i[1]++;
2769         if(i[0] < len){
2770             in[ i[0] ]->pic_id= is_long ? i[0] : in[ i[0] ]->frame_num;
2771             split_field_copy(&def[index++], in[ i[0]++ ], sel  , 1);
2772         }
2773         if(i[1] < len){
2774             in[ i[1] ]->pic_id= is_long ? i[1] : in[ i[1] ]->frame_num;
2775             split_field_copy(&def[index++], in[ i[1]++ ], sel^3, 0);
2776         }
2777     }
2778
2779     return index;
2780 }
2781
2782 static int add_sorted(Picture **sorted, Picture **src, int len, int limit, int dir){
2783     int i, best_poc;
2784     int out_i= 0;
2785
2786     for(;;){
2787         best_poc= dir ? INT_MIN : INT_MAX;
2788
2789         for(i=0; i<len; i++){
2790             const int poc= src[i]->poc;
2791             if(((poc > limit) ^ dir) && ((poc < best_poc) ^ dir)){
2792                 best_poc= poc;
2793                 sorted[out_i]= src[i];
2794             }
2795         }
2796         if(best_poc == (dir ? INT_MIN : INT_MAX))
2797             break;
2798         limit= sorted[out_i++]->poc - dir;
2799     }
2800     return out_i;
2801 }
2802
2803 /**
2804  * fills the default_ref_list.
2805  */
2806 static int fill_default_ref_list(H264Context *h){
2807     MpegEncContext * const s = &h->s;
2808     int i, len;
2809
2810     if(h->slice_type_nos==FF_B_TYPE){
2811         Picture *sorted[32];
2812         int cur_poc, list;
2813         int lens[2];
2814
2815         if(FIELD_PICTURE)
2816             cur_poc= s->current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
2817         else
2818             cur_poc= s->current_picture_ptr->poc;
2819
2820         for(list= 0; list<2; list++){
2821             len= add_sorted(sorted    , h->short_ref, h->short_ref_count, cur_poc, 1^list);
2822             len+=add_sorted(sorted+len, h->short_ref, h->short_ref_count, cur_poc, 0^list);
2823             assert(len<=32);
2824             len= build_def_list(h->default_ref_list[list]    , sorted     , len, 0, s->picture_structure);
2825             len+=build_def_list(h->default_ref_list[list]+len, h->long_ref, 16 , 1, s->picture_structure);
2826             assert(len<=32);
2827
2828             if(len < h->ref_count[list])
2829                 memset(&h->default_ref_list[list][len], 0, sizeof(Picture)*(h->ref_count[list] - len));
2830             lens[list]= len;
2831         }
2832
2833         if(lens[0] == lens[1] && lens[1] > 1){
2834             for(i=0; h->default_ref_list[0][i].data[0] == h->default_ref_list[1][i].data[0] && i<lens[0]; i++);
2835             if(i == lens[0])
2836                 FFSWAP(Picture, h->default_ref_list[1][0], h->default_ref_list[1][1]);
2837         }
2838     }else{
2839         len = build_def_list(h->default_ref_list[0]    , h->short_ref, h->short_ref_count, 0, s->picture_structure);
2840         len+= build_def_list(h->default_ref_list[0]+len, h-> long_ref, 16                , 1, s->picture_structure);
2841         assert(len <= 32);
2842         if(len < h->ref_count[0])
2843             memset(&h->default_ref_list[0][len], 0, sizeof(Picture)*(h->ref_count[0] - len));
2844     }
2845 #ifdef TRACE
2846     for (i=0; i<h->ref_count[0]; i++) {
2847         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]);
2848     }
2849     if(h->slice_type_nos==FF_B_TYPE){
2850         for (i=0; i<h->ref_count[1]; i++) {
2851             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]);
2852         }
2853     }
2854 #endif
2855     return 0;
2856 }
2857
2858 static void print_short_term(H264Context *h);
2859 static void print_long_term(H264Context *h);
2860
2861 /**
2862  * Extract structure information about the picture described by pic_num in
2863  * the current decoding context (frame or field). Note that pic_num is
2864  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
2865  * @param pic_num picture number for which to extract structure information
2866  * @param structure one of PICT_XXX describing structure of picture
2867  *                      with pic_num
2868  * @return frame number (short term) or long term index of picture
2869  *         described by pic_num
2870  */
2871 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
2872     MpegEncContext * const s = &h->s;
2873
2874     *structure = s->picture_structure;
2875     if(FIELD_PICTURE){
2876         if (!(pic_num & 1))
2877             /* opposite field */
2878             *structure ^= PICT_FRAME;
2879         pic_num >>= 1;
2880     }
2881
2882     return pic_num;
2883 }
2884
2885 static int decode_ref_pic_list_reordering(H264Context *h){
2886     MpegEncContext * const s = &h->s;
2887     int list, index, pic_structure;
2888
2889     print_short_term(h);
2890     print_long_term(h);
2891
2892     for(list=0; list<h->list_count; list++){
2893         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
2894
2895         if(get_bits1(&s->gb)){
2896             int pred= h->curr_pic_num;
2897
2898             for(index=0; ; index++){
2899                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb(&s->gb);
2900                 unsigned int pic_id;
2901                 int i;
2902                 Picture *ref = NULL;
2903
2904                 if(reordering_of_pic_nums_idc==3)
2905                     break;
2906
2907                 if(index >= h->ref_count[list]){
2908                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
2909                     return -1;
2910                 }
2911
2912                 if(reordering_of_pic_nums_idc<3){
2913                     if(reordering_of_pic_nums_idc<2){
2914                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
2915                         int frame_num;
2916
2917                         if(abs_diff_pic_num > h->max_pic_num){
2918                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
2919                             return -1;
2920                         }
2921
2922                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
2923                         else                                pred+= abs_diff_pic_num;
2924                         pred &= h->max_pic_num - 1;
2925
2926                         frame_num = pic_num_extract(h, pred, &pic_structure);
2927
2928                         for(i= h->short_ref_count-1; i>=0; i--){
2929                             ref = h->short_ref[i];
2930                             assert(ref->reference);
2931                             assert(!ref->long_ref);
2932                             if(
2933                                    ref->frame_num == frame_num &&
2934                                    (ref->reference & pic_structure)
2935                               )
2936                                 break;
2937                         }
2938                         if(i>=0)
2939                             ref->pic_id= pred;
2940                     }else{
2941                         int long_idx;
2942                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
2943
2944                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
2945
2946                         if(long_idx>31){
2947                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
2948                             return -1;
2949                         }
2950                         ref = h->long_ref[long_idx];
2951                         assert(!(ref && !ref->reference));
2952                         if(ref && (ref->reference & pic_structure)){
2953                             ref->pic_id= pic_id;
2954                             assert(ref->long_ref);
2955                             i=0;
2956                         }else{
2957                             i=-1;
2958                         }
2959                     }
2960
2961                     if (i < 0) {
2962                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
2963                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
2964                     } else {
2965                         for(i=index; i+1<h->ref_count[list]; i++){
2966                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
2967                                 break;
2968                         }
2969                         for(; i > index; i--){
2970                             h->ref_list[list][i]= h->ref_list[list][i-1];
2971                         }
2972                         h->ref_list[list][index]= *ref;
2973                         if (FIELD_PICTURE){
2974                             pic_as_field(&h->ref_list[list][index], pic_structure);
2975                         }
2976                     }
2977                 }else{
2978                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
2979                     return -1;
2980                 }
2981             }
2982         }
2983     }
2984     for(list=0; list<h->list_count; list++){
2985         for(index= 0; index < h->ref_count[list]; index++){
2986             if(!h->ref_list[list][index].data[0]){
2987                 av_log(h->s.avctx, AV_LOG_ERROR, "Missing reference picture\n");
2988                 h->ref_list[list][index]= s->current_picture; //FIXME this is not a sensible solution
2989             }
2990         }
2991     }
2992
2993     if(h->slice_type_nos==FF_B_TYPE && !h->direct_spatial_mv_pred)
2994         direct_dist_scale_factor(h);
2995     direct_ref_list_init(h);
2996     return 0;
2997 }
2998
2999 static void fill_mbaff_ref_list(H264Context *h){
3000     int list, i, j;
3001     for(list=0; list<2; list++){ //FIXME try list_count
3002         for(i=0; i<h->ref_count[list]; i++){
3003             Picture *frame = &h->ref_list[list][i];
3004             Picture *field = &h->ref_list[list][16+2*i];
3005             field[0] = *frame;
3006             for(j=0; j<3; j++)
3007                 field[0].linesize[j] <<= 1;
3008             field[0].reference = PICT_TOP_FIELD;
3009             field[1] = field[0];
3010             for(j=0; j<3; j++)
3011                 field[1].data[j] += frame->linesize[j];
3012             field[1].reference = PICT_BOTTOM_FIELD;
3013
3014             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
3015             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
3016             for(j=0; j<2; j++){
3017                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
3018                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
3019             }
3020         }
3021     }
3022     for(j=0; j<h->ref_count[1]; j++){
3023         for(i=0; i<h->ref_count[0]; i++)
3024             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
3025         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
3026         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
3027     }
3028 }
3029
3030 static int pred_weight_table(H264Context *h){
3031     MpegEncContext * const s = &h->s;
3032     int list, i;
3033     int luma_def, chroma_def;
3034
3035     h->use_weight= 0;
3036     h->use_weight_chroma= 0;
3037     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
3038     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
3039     luma_def = 1<<h->luma_log2_weight_denom;
3040     chroma_def = 1<<h->chroma_log2_weight_denom;
3041
3042     for(list=0; list<2; list++){
3043         for(i=0; i<h->ref_count[list]; i++){
3044             int luma_weight_flag, chroma_weight_flag;
3045
3046             luma_weight_flag= get_bits1(&s->gb);
3047             if(luma_weight_flag){
3048                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3049                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3050                 if(   h->luma_weight[list][i] != luma_def
3051                    || h->luma_offset[list][i] != 0)
3052                     h->use_weight= 1;
3053             }else{
3054                 h->luma_weight[list][i]= luma_def;
3055                 h->luma_offset[list][i]= 0;
3056             }
3057
3058             if(CHROMA){
3059                 chroma_weight_flag= get_bits1(&s->gb);
3060                 if(chroma_weight_flag){
3061                     int j;
3062                     for(j=0; j<2; j++){
3063                         h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3064                         h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3065                         if(   h->chroma_weight[list][i][j] != chroma_def
3066                         || h->chroma_offset[list][i][j] != 0)
3067                             h->use_weight_chroma= 1;
3068                     }
3069                 }else{
3070                     int j;
3071                     for(j=0; j<2; j++){
3072                         h->chroma_weight[list][i][j]= chroma_def;
3073                         h->chroma_offset[list][i][j]= 0;
3074                     }
3075                 }
3076             }
3077         }
3078         if(h->slice_type_nos != FF_B_TYPE) break;
3079     }
3080     h->use_weight= h->use_weight || h->use_weight_chroma;
3081     return 0;
3082 }
3083
3084 static void implicit_weight_table(H264Context *h){
3085     MpegEncContext * const s = &h->s;
3086     int ref0, ref1;
3087     int cur_poc = s->current_picture_ptr->poc;
3088
3089     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3090        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3091         h->use_weight= 0;
3092         h->use_weight_chroma= 0;
3093         return;
3094     }
3095
3096     h->use_weight= 2;
3097     h->use_weight_chroma= 2;
3098     h->luma_log2_weight_denom= 5;
3099     h->chroma_log2_weight_denom= 5;
3100
3101     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3102         int poc0 = h->ref_list[0][ref0].poc;
3103         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3104             int poc1 = h->ref_list[1][ref1].poc;
3105             int td = av_clip(poc1 - poc0, -128, 127);
3106             if(td){
3107                 int tb = av_clip(cur_poc - poc0, -128, 127);
3108                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3109                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3110                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3111                     h->implicit_weight[ref0][ref1] = 32;
3112                 else
3113                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3114             }else
3115                 h->implicit_weight[ref0][ref1] = 32;
3116         }
3117     }
3118 }
3119
3120 /**
3121  * Mark a picture as no longer needed for reference. The refmask
3122  * argument allows unreferencing of individual fields or the whole frame.
3123  * If the picture becomes entirely unreferenced, but is being held for
3124  * display purposes, it is marked as such.
3125  * @param refmask mask of fields to unreference; the mask is bitwise
3126  *                anded with the reference marking of pic
3127  * @return non-zero if pic becomes entirely unreferenced (except possibly
3128  *         for display purposes) zero if one of the fields remains in
3129  *         reference
3130  */
3131 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3132     int i;
3133     if (pic->reference &= refmask) {
3134         return 0;
3135     } else {
3136         for(i = 0; h->delayed_pic[i]; i++)
3137             if(pic == h->delayed_pic[i]){
3138                 pic->reference=DELAYED_PIC_REF;
3139                 break;
3140             }
3141         return 1;
3142     }
3143 }
3144
3145 /**
3146  * instantaneous decoder refresh.
3147  */
3148 static void idr(H264Context *h){
3149     int i;
3150
3151     for(i=0; i<16; i++){
3152         remove_long(h, i, 0);
3153     }
3154     assert(h->long_ref_count==0);
3155
3156     for(i=0; i<h->short_ref_count; i++){
3157         unreference_pic(h, h->short_ref[i], 0);
3158         h->short_ref[i]= NULL;
3159     }
3160     h->short_ref_count=0;
3161     h->prev_frame_num= 0;
3162     h->prev_frame_num_offset= 0;
3163     h->prev_poc_msb=
3164     h->prev_poc_lsb= 0;
3165 }
3166
3167 /* forget old pics after a seek */
3168 static void flush_dpb(AVCodecContext *avctx){
3169     H264Context *h= avctx->priv_data;
3170     int i;
3171     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
3172         if(h->delayed_pic[i])
3173             h->delayed_pic[i]->reference= 0;
3174         h->delayed_pic[i]= NULL;
3175     }
3176     h->outputed_poc= INT_MIN;
3177     idr(h);
3178     if(h->s.current_picture_ptr)
3179         h->s.current_picture_ptr->reference= 0;
3180     h->s.first_field= 0;
3181     ff_mpeg_flush(avctx);
3182 }
3183
3184 /**
3185  * Find a Picture in the short term reference list by frame number.
3186  * @param frame_num frame number to search for
3187  * @param idx the index into h->short_ref where returned picture is found
3188  *            undefined if no picture found.
3189  * @return pointer to the found picture, or NULL if no pic with the provided
3190  *                 frame number is found
3191  */
3192 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3193     MpegEncContext * const s = &h->s;
3194     int i;
3195
3196     for(i=0; i<h->short_ref_count; i++){
3197         Picture *pic= h->short_ref[i];
3198         if(s->avctx->debug&FF_DEBUG_MMCO)
3199             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3200         if(pic->frame_num == frame_num) {
3201             *idx = i;
3202             return pic;
3203         }
3204     }
3205     return NULL;
3206 }
3207
3208 /**
3209  * Remove a picture from the short term reference list by its index in
3210  * that list.  This does no checking on the provided index; it is assumed
3211  * to be valid. Other list entries are shifted down.
3212  * @param i index into h->short_ref of picture to remove.
3213  */
3214 static void remove_short_at_index(H264Context *h, int i){
3215     assert(i >= 0 && i < h->short_ref_count);
3216     h->short_ref[i]= NULL;
3217     if (--h->short_ref_count)
3218         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3219 }
3220
3221 /**
3222  *
3223  * @return the removed picture or NULL if an error occurs
3224  */
3225 static Picture * remove_short(H264Context *h, int frame_num, int ref_mask){
3226     MpegEncContext * const s = &h->s;
3227     Picture *pic;
3228     int i;
3229
3230     if(s->avctx->debug&FF_DEBUG_MMCO)
3231         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3232
3233     pic = find_short(h, frame_num, &i);
3234     if (pic){
3235         if(unreference_pic(h, pic, ref_mask))
3236         remove_short_at_index(h, i);
3237     }
3238
3239     return pic;
3240 }
3241
3242 /**
3243  * Remove a picture from the long term reference list by its index in
3244  * that list.
3245  * @return the removed picture or NULL if an error occurs
3246  */
3247 static Picture * remove_long(H264Context *h, int i, int ref_mask){
3248     Picture *pic;
3249
3250     pic= h->long_ref[i];
3251     if (pic){
3252         if(unreference_pic(h, pic, ref_mask)){
3253             assert(h->long_ref[i]->long_ref == 1);
3254             h->long_ref[i]->long_ref= 0;
3255             h->long_ref[i]= NULL;
3256             h->long_ref_count--;
3257         }
3258     }
3259
3260     return pic;
3261 }
3262
3263 /**
3264  * print short term list
3265  */
3266 static void print_short_term(H264Context *h) {
3267     uint32_t i;
3268     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3269         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3270         for(i=0; i<h->short_ref_count; i++){
3271             Picture *pic= h->short_ref[i];
3272             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3273         }
3274     }
3275 }
3276
3277 /**
3278  * print long term list
3279  */
3280 static void print_long_term(H264Context *h) {
3281     uint32_t i;
3282     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3283         av_log(h->s.avctx, AV_LOG_DEBUG, "long term list:\n");
3284         for(i = 0; i < 16; i++){
3285             Picture *pic= h->long_ref[i];
3286             if (pic) {
3287                 av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3288             }
3289         }
3290     }
3291 }
3292
3293 /**
3294  * Executes the reference picture marking (memory management control operations).
3295  */
3296 static int execute_ref_pic_marking(H264Context *h, MMCO *mmco, int mmco_count){
3297     MpegEncContext * const s = &h->s;
3298     int i, j;
3299     int current_ref_assigned=0;
3300     Picture *pic;
3301
3302     if((s->avctx->debug&FF_DEBUG_MMCO) && mmco_count==0)
3303         av_log(h->s.avctx, AV_LOG_DEBUG, "no mmco here\n");
3304
3305     for(i=0; i<mmco_count; i++){
3306         int structure, frame_num;
3307         if(s->avctx->debug&FF_DEBUG_MMCO)
3308             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);
3309
3310         if(   mmco[i].opcode == MMCO_SHORT2UNUSED
3311            || mmco[i].opcode == MMCO_SHORT2LONG){
3312             frame_num = pic_num_extract(h, mmco[i].short_pic_num, &structure);
3313             pic = find_short(h, frame_num, &j);
3314             if(!pic){
3315                 if(mmco[i].opcode != MMCO_SHORT2LONG || !h->long_ref[mmco[i].long_arg]
3316                    || h->long_ref[mmco[i].long_arg]->frame_num != frame_num)
3317                 av_log(h->s.avctx, AV_LOG_ERROR, "mmco: unref short failure\n");
3318                 continue;
3319             }
3320         }
3321
3322         switch(mmco[i].opcode){
3323         case MMCO_SHORT2UNUSED:
3324             if(s->avctx->debug&FF_DEBUG_MMCO)
3325                 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);
3326             remove_short(h, frame_num, structure ^ PICT_FRAME);
3327             break;
3328         case MMCO_SHORT2LONG:
3329                 if (h->long_ref[mmco[i].long_arg] != pic)
3330                     remove_long(h, mmco[i].long_arg, 0);
3331
3332                 remove_short_at_index(h, j);
3333                 h->long_ref[ mmco[i].long_arg ]= pic;
3334                 if (h->long_ref[ mmco[i].long_arg ]){
3335                     h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3336                     h->long_ref_count++;
3337                 }
3338             break;
3339         case MMCO_LONG2UNUSED:
3340             j = pic_num_extract(h, mmco[i].long_arg, &structure);
3341             pic = h->long_ref[j];
3342             if (pic) {
3343                 remove_long(h, j, structure ^ PICT_FRAME);
3344             } else if(s->avctx->debug&FF_DEBUG_MMCO)
3345                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref long failure\n");
3346             break;
3347         case MMCO_LONG:
3348                     // Comment below left from previous code as it is an interresting note.
3349                     /* First field in pair is in short term list or
3350                      * at a different long term index.
3351                      * This is not allowed; see 7.4.3.3, notes 2 and 3.
3352                      * Report the problem and keep the pair where it is,
3353                      * and mark this field valid.
3354                      */
3355
3356             if (h->long_ref[mmco[i].long_arg] != s->current_picture_ptr) {
3357                 remove_long(h, mmco[i].long_arg, 0);
3358
3359                 h->long_ref[ mmco[i].long_arg ]= s->current_picture_ptr;
3360                 h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3361                 h->long_ref_count++;
3362             }
3363
3364             s->current_picture_ptr->reference |= s->picture_structure;
3365             current_ref_assigned=1;
3366             break;
3367         case MMCO_SET_MAX_LONG:
3368             assert(mmco[i].long_arg <= 16);
3369             // just remove the long term which index is greater than new max
3370             for(j = mmco[i].long_arg; j<16; j++){
3371                 remove_long(h, j, 0);
3372             }
3373             break;
3374         case MMCO_RESET:
3375             while(h->short_ref_count){
3376                 remove_short(h, h->short_ref[0]->frame_num, 0);
3377             }
3378             for(j = 0; j < 16; j++) {
3379                 remove_long(h, j, 0);
3380             }
3381             s->current_picture_ptr->poc=
3382             s->current_picture_ptr->field_poc[0]=
3383             s->current_picture_ptr->field_poc[1]=
3384             h->poc_lsb=
3385             h->poc_msb=
3386             h->frame_num=
3387             s->current_picture_ptr->frame_num= 0;
3388             break;
3389         default: assert(0);
3390         }
3391     }
3392
3393     if (!current_ref_assigned) {
3394         /* Second field of complementary field pair; the first field of
3395          * which is already referenced. If short referenced, it
3396          * should be first entry in short_ref. If not, it must exist
3397          * in long_ref; trying to put it on the short list here is an
3398          * error in the encoded bit stream (ref: 7.4.3.3, NOTE 2 and 3).
3399          */
3400         if (h->short_ref_count && h->short_ref[0] == s->current_picture_ptr) {
3401             /* Just mark the second field valid */
3402             s->current_picture_ptr->reference = PICT_FRAME;
3403         } else if (s->current_picture_ptr->long_ref) {
3404             av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term reference "
3405                                              "assignment for second field "
3406                                              "in complementary field pair "
3407                                              "(first field is long term)\n");
3408         } else {
3409             pic= remove_short(h, s->current_picture_ptr->frame_num, 0);
3410             if(pic){
3411                 av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term buffer state detected\n");
3412             }
3413
3414             if(h->short_ref_count)
3415                 memmove(&h->short_ref[1], &h->short_ref[0], h->short_ref_count*sizeof(Picture*));
3416
3417             h->short_ref[0]= s->current_picture_ptr;
3418             h->short_ref_count++;
3419             s->current_picture_ptr->reference |= s->picture_structure;
3420         }
3421     }
3422
3423     if (h->long_ref_count + h->short_ref_count > h->sps.ref_frame_count){
3424
3425         /* We have too many reference frames, probably due to corrupted
3426          * stream. Need to discard one frame. Prevents overrun of the
3427          * short_ref and long_ref buffers.
3428          */
3429         av_log(h->s.avctx, AV_LOG_ERROR,
3430                "number of reference frames exceeds max (probably "
3431                "corrupt input), discarding one\n");
3432
3433         if (h->long_ref_count && !h->short_ref_count) {
3434            &n