Support for lossless and inter H264 4:2:2.
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of Libav.
6  *
7  * Libav 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  * Libav 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 Libav; 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
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "libavutil/imgutils.h"
29 #include "internal.h"
30 #include "dsputil.h"
31 #include "avcodec.h"
32 #include "mpegvideo.h"
33 #include "h264.h"
34 #include "h264data.h"
35 #include "h264_mvpred.h"
36 #include "golomb.h"
37 #include "mathops.h"
38 #include "rectangle.h"
39 #include "thread.h"
40 #include "vdpau_internal.h"
41 #include "libavutil/avassert.h"
42
43 #include "cabac.h"
44
45 //#undef NDEBUG
46 #include <assert.h>
47
48 static const uint8_t rem6[QP_MAX_NUM+1]={
49 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, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
50 };
51
52 static const uint8_t div6[QP_MAX_NUM+1]={
53 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, 8, 8, 9, 9, 9, 9, 9, 9,10,10,10,10,
54 };
55
56 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
57     PIX_FMT_DXVA2_VLD,
58     PIX_FMT_VAAPI_VLD,
59     PIX_FMT_YUVJ420P,
60     PIX_FMT_NONE
61 };
62
63 /**
64  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
65  */
66 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
67     MpegEncContext * const s = &h->s;
68     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
69     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
70     int i;
71
72     if(!(h->top_samples_available&0x8000)){
73         for(i=0; i<4; i++){
74             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
75             if(status<0){
76                 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);
77                 return -1;
78             } else if(status){
79                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
80             }
81         }
82     }
83
84     if((h->left_samples_available&0x8888)!=0x8888){
85         static const int mask[4]={0x8000,0x2000,0x80,0x20};
86         for(i=0; i<4; i++){
87             if(!(h->left_samples_available&mask[i])){
88                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
89                 if(status<0){
90                     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);
91                     return -1;
92                 } else if(status){
93                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
94                 }
95             }
96         }
97     }
98
99     return 0;
100 } //FIXME cleanup like ff_h264_check_intra_pred_mode
101
102 /**
103  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
104  */
105 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
106     MpegEncContext * const s = &h->s;
107     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
108     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
109
110     if(mode > 6U) {
111         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);
112         return -1;
113     }
114
115     if(!(h->top_samples_available&0x8000)){
116         mode= top[ mode ];
117         if(mode<0){
118             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);
119             return -1;
120         }
121     }
122
123     if((h->left_samples_available&0x8080) != 0x8080){
124         mode= left[ mode ];
125         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
126             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
127         }
128         if(mode<0){
129             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);
130             return -1;
131         }
132     }
133
134     return mode;
135 }
136
137 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
138     int i, si, di;
139     uint8_t *dst;
140     int bufidx;
141
142 //    src[0]&0x80;                //forbidden bit
143     h->nal_ref_idc= src[0]>>5;
144     h->nal_unit_type= src[0]&0x1F;
145
146     src++; length--;
147
148 #if HAVE_FAST_UNALIGNED
149 # if HAVE_FAST_64BIT
150 #   define RS 7
151     for(i=0; i+1<length; i+=9){
152         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
153 # else
154 #   define RS 3
155     for(i=0; i+1<length; i+=5){
156         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
157 # endif
158             continue;
159         if(i>0 && !src[i]) i--;
160         while(src[i]) i++;
161 #else
162 #   define RS 0
163     for(i=0; i+1<length; i+=2){
164         if(src[i]) continue;
165         if(i>0 && src[i-1]==0) i--;
166 #endif
167         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
168             if(src[i+2]!=3){
169                 /* startcode, so we must be past the end */
170                 length=i;
171             }
172             break;
173         }
174         i-= RS;
175     }
176
177     if(i>=length-1){ //no escaped 0
178         *dst_length= length;
179         *consumed= length+1; //+1 for the header
180         return src;
181     }
182
183     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
184     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
185     dst= h->rbsp_buffer[bufidx];
186
187     if (dst == NULL){
188         return NULL;
189     }
190
191 //printf("decoding esc\n");
192     memcpy(dst, src, i);
193     si=di=i;
194     while(si+2<length){
195         //remove escapes (very rare 1:2^22)
196         if(src[si+2]>3){
197             dst[di++]= src[si++];
198             dst[di++]= src[si++];
199         }else if(src[si]==0 && src[si+1]==0){
200             if(src[si+2]==3){ //escape
201                 dst[di++]= 0;
202                 dst[di++]= 0;
203                 si+=3;
204                 continue;
205             }else //next start code
206                 goto nsc;
207         }
208
209         dst[di++]= src[si++];
210     }
211     while(si<length)
212         dst[di++]= src[si++];
213 nsc:
214
215     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
216
217     *dst_length= di;
218     *consumed= si + 1;//+1 for the header
219 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
220     return dst;
221 }
222
223 /**
224  * Identify the exact end of the bitstream
225  * @return the length of the trailing, or 0 if damaged
226  */
227 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
228     int v= *src;
229     int r;
230
231     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
232
233     for(r=1; r<9; r++){
234         if(v&1) return r;
235         v>>=1;
236     }
237     return 0;
238 }
239
240 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
241                                  int y_offset, int list){
242     int raw_my= h->mv_cache[list][ scan8[n] ][1];
243     int filter_height= (raw_my&3) ? 2 : 0;
244     int full_my= (raw_my>>2) + y_offset;
245     int top = full_my - filter_height, bottom = full_my + height + filter_height;
246
247     return FFMAX(abs(top), bottom);
248 }
249
250 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
251                                int y_offset, int list0, int list1, int *nrefs){
252     MpegEncContext * const s = &h->s;
253     int my;
254
255     y_offset += 16*(s->mb_y >> MB_FIELD);
256
257     if(list0){
258         int ref_n = h->ref_cache[0][ scan8[n] ];
259         Picture *ref= &h->ref_list[0][ref_n];
260
261         // Error resilience puts the current picture in the ref list.
262         // Don't try to wait on these as it will cause a deadlock.
263         // Fields can wait on each other, though.
264         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
265            (ref->f.reference & 3) != s->picture_structure) {
266             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
267             if (refs[0][ref_n] < 0) nrefs[0] += 1;
268             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
269         }
270     }
271
272     if(list1){
273         int ref_n = h->ref_cache[1][ scan8[n] ];
274         Picture *ref= &h->ref_list[1][ref_n];
275
276         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
277            (ref->f.reference & 3) != s->picture_structure) {
278             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
279             if (refs[1][ref_n] < 0) nrefs[1] += 1;
280             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
281         }
282     }
283 }
284
285 /**
286  * Wait until all reference frames are available for MC operations.
287  *
288  * @param h the H264 context
289  */
290 static void await_references(H264Context *h){
291     MpegEncContext * const s = &h->s;
292     const int mb_xy= h->mb_xy;
293     const int mb_type = s->current_picture.f.mb_type[mb_xy];
294     int refs[2][48];
295     int nrefs[2] = {0};
296     int ref, list;
297
298     memset(refs, -1, sizeof(refs));
299
300     if(IS_16X16(mb_type)){
301         get_lowest_part_y(h, refs, 0, 16, 0,
302                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
303     }else if(IS_16X8(mb_type)){
304         get_lowest_part_y(h, refs, 0, 8, 0,
305                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
306         get_lowest_part_y(h, refs, 8, 8, 8,
307                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
308     }else if(IS_8X16(mb_type)){
309         get_lowest_part_y(h, refs, 0, 16, 0,
310                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
311         get_lowest_part_y(h, refs, 4, 16, 0,
312                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
313     }else{
314         int i;
315
316         assert(IS_8X8(mb_type));
317
318         for(i=0; i<4; i++){
319             const int sub_mb_type= h->sub_mb_type[i];
320             const int n= 4*i;
321             int y_offset= (i&2)<<2;
322
323             if(IS_SUB_8X8(sub_mb_type)){
324                 get_lowest_part_y(h, refs, n  , 8, y_offset,
325                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
326             }else if(IS_SUB_8X4(sub_mb_type)){
327                 get_lowest_part_y(h, refs, n  , 4, y_offset,
328                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
329                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
330                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
331             }else if(IS_SUB_4X8(sub_mb_type)){
332                 get_lowest_part_y(h, refs, n  , 8, y_offset,
333                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
334                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
335                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
336             }else{
337                 int j;
338                 assert(IS_SUB_4X4(sub_mb_type));
339                 for(j=0; j<4; j++){
340                     int sub_y_offset= y_offset + 2*(j&2);
341                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
342                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
343                 }
344             }
345         }
346     }
347
348     for(list=h->list_count-1; list>=0; list--){
349         for(ref=0; ref<48 && nrefs[list]; ref++){
350             int row = refs[list][ref];
351             if(row >= 0){
352                 Picture *ref_pic = &h->ref_list[list][ref];
353                 int ref_field = ref_pic->f.reference - 1;
354                 int ref_field_picture = ref_pic->field_picture;
355                 int pic_height = 16*s->mb_height >> ref_field_picture;
356
357                 row <<= MB_MBAFF;
358                 nrefs[list]--;
359
360                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
361                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
362                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
363                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
364                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
365                 }else if(FIELD_PICTURE){
366                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
367                 }else{
368                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
369                 }
370             }
371         }
372     }
373 }
374
375 #if 0
376 /**
377  * DCT transforms the 16 dc values.
378  * @param qp quantization parameter ??? FIXME
379  */
380 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
381 //    const int qmul= dequant_coeff[qp][0];
382     int i;
383     int temp[16]; //FIXME check if this is a good idea
384     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
385     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
386
387     for(i=0; i<4; i++){
388         const int offset= y_offset[i];
389         const int z0= block[offset+stride*0] + block[offset+stride*4];
390         const int z1= block[offset+stride*0] - block[offset+stride*4];
391         const int z2= block[offset+stride*1] - block[offset+stride*5];
392         const int z3= block[offset+stride*1] + block[offset+stride*5];
393
394         temp[4*i+0]= z0+z3;
395         temp[4*i+1]= z1+z2;
396         temp[4*i+2]= z1-z2;
397         temp[4*i+3]= z0-z3;
398     }
399
400     for(i=0; i<4; i++){
401         const int offset= x_offset[i];
402         const int z0= temp[4*0+i] + temp[4*2+i];
403         const int z1= temp[4*0+i] - temp[4*2+i];
404         const int z2= temp[4*1+i] - temp[4*3+i];
405         const int z3= temp[4*1+i] + temp[4*3+i];
406
407         block[stride*0 +offset]= (z0 + z3)>>1;
408         block[stride*2 +offset]= (z1 + z2)>>1;
409         block[stride*8 +offset]= (z1 - z2)>>1;
410         block[stride*10+offset]= (z0 - z3)>>1;
411     }
412 }
413 #endif
414
415 #undef xStride
416 #undef stride
417
418 #if 0
419 static void chroma_dc_dct_c(DCTELEM *block){
420     const int stride= 16*2;
421     const int xStride= 16;
422     int a,b,c,d,e;
423
424     a= block[stride*0 + xStride*0];
425     b= block[stride*0 + xStride*1];
426     c= block[stride*1 + xStride*0];
427     d= block[stride*1 + xStride*1];
428
429     e= a-b;
430     a= a+b;
431     b= c-d;
432     c= c+d;
433
434     block[stride*0 + xStride*0]= (a+c);
435     block[stride*0 + xStride*1]= (e+b);
436     block[stride*1 + xStride*0]= (a-c);
437     block[stride*1 + xStride*1]= (e-b);
438 }
439 #endif
440
441 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
442                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
443                            int src_x_offset, int src_y_offset,
444                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
445                            int pixel_shift, int chroma444){
446     MpegEncContext * const s = &h->s;
447     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
448     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
449     const int luma_xy= (mx&3) + ((my&3)<<2);
450     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
451     uint8_t * src_y = pic->f.data[0] + offset;
452     uint8_t * src_cb, * src_cr;
453     int extra_width= h->emu_edge_width;
454     int extra_height= h->emu_edge_height;
455     int emu=0;
456     const int full_mx= mx>>2;
457     const int full_my= my>>2;
458     const int pic_width  = 16*s->mb_width;
459     const int pic_height = 16*s->mb_height >> MB_FIELD;
460     int ysh;
461
462     if(mx&7) extra_width -= 3;
463     if(my&7) extra_height -= 3;
464
465     if(   full_mx < 0-extra_width
466        || full_my < 0-extra_height
467        || full_mx + 16/*FIXME*/ > pic_width + extra_width
468        || full_my + 16/*FIXME*/ > pic_height + extra_height){
469         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_y - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
470                                 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
471             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
472         emu=1;
473     }
474
475     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
476     if(!square){
477         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
478     }
479
480     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
481
482     if(chroma444){
483         src_cb = pic->f.data[1] + offset;
484         if(emu){
485             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
486                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
487             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
488         }
489         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
490         if(!square){
491             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
492         }
493
494         src_cr = pic->f.data[2] + offset;
495         if(emu){
496             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
497                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
498             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
499         }
500         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
501         if(!square){
502             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
503         }
504         return;
505     }
506
507     ysh = 3 - !!(CHROMA422);
508     if(!CHROMA422 && MB_FIELD){
509         // chroma offset when predicting from a field of opposite parity
510         my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
511         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
512     }
513
514     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) + (my >> ysh) * h->mb_uvlinesize;
515     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) + (my >> ysh) * h->mb_uvlinesize;
516
517     if(emu){
518         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize, 9, (16>>!(CHROMA422)) + 1, (mx>>3), (my>>ysh), pic_width>>1, pic_height>>!(CHROMA422));
519             src_cb= s->edge_emu_buffer;
520     }
521     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height << !!(CHROMA422), mx&7, (my << !!(CHROMA422)) &7);
522
523     if(emu){
524         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize, 9, (16>>!(CHROMA422)) + 1, (mx>>3), (my>>ysh), pic_width>>1, pic_height>>!(CHROMA422));
525             src_cr= s->edge_emu_buffer;
526     }
527     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height << !!(CHROMA422), mx&7, (my << !!(CHROMA422)) &7);
528 }
529
530 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
531                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
532                            int x_offset, int y_offset,
533                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
534                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
535                            int list0, int list1, int pixel_shift, int chroma444){
536     MpegEncContext * const s = &h->s;
537     qpel_mc_func *qpix_op=  qpix_put;
538     h264_chroma_mc_func chroma_op= chroma_put;
539
540     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
541     if(chroma444){
542         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
543         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
544     } else if (CHROMA422) {
545         dest_cb += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
546         dest_cr += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
547     }else{
548         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
549         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
550     }
551     x_offset += 8*s->mb_x;
552     y_offset += 8*(s->mb_y >> MB_FIELD);
553
554     if(list0){
555         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
556         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
557                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
558                            qpix_op, chroma_op, pixel_shift, chroma444);
559
560         qpix_op=  qpix_avg;
561         chroma_op= chroma_avg;
562     }
563
564     if(list1){
565         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
566         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
567                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
568                            qpix_op, chroma_op, pixel_shift, chroma444);
569     }
570 }
571
572 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
573                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
574                            int x_offset, int y_offset,
575                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
576                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
577                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
578                            int list0, int list1, int pixel_shift, int chroma444){
579     MpegEncContext * const s = &h->s;
580
581     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
582     if(chroma444){
583         chroma_weight_avg = luma_weight_avg;
584         chroma_weight_op = luma_weight_op;
585         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
586         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
587     } else if (CHROMA422) {
588         dest_cb += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
589         dest_cr += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
590     }else{
591         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
592         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
593     }
594     x_offset += 8*s->mb_x;
595     y_offset += 8*(s->mb_y >> MB_FIELD);
596
597     if(list0 && list1){
598         /* don't optimize for luma-only case, since B-frames usually
599          * use implicit weights => chroma too. */
600         uint8_t *tmp_cb = s->obmc_scratchpad;
601         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
602         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
603         int refn0 = h->ref_cache[0][ scan8[n] ];
604         int refn1 = h->ref_cache[1][ scan8[n] ];
605
606         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
607                     dest_y, dest_cb, dest_cr,
608                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
609         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
610                     tmp_y, tmp_cb, tmp_cr,
611                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
612
613         if(h->use_weight == 2){
614             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
615             int weight1 = 64 - weight0;
616             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
617             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
618             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
619             if (CHROMA422) {
620                 chroma_weight_avg(dest_cb + chroma_height * h->mb_uvlinesize,
621                                   tmp_cb + chroma_height * h->mb_uvlinesize,
622                                   h->mb_uvlinesize, 5, weight0, weight1, 0);
623                 chroma_weight_avg(dest_cr + chroma_height * h->mb_uvlinesize,
624                                   tmp_cr + chroma_height * h->mb_uvlinesize,
625                                   h->mb_uvlinesize, 5, weight0, weight1, 0);
626             }
627         }else{
628             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
629                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
630                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
631             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
632                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
633                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
634             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
635                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
636                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
637             if (CHROMA422) {
638                 chroma_weight_avg(dest_cb + chroma_height * h->mb_uvlinesize,
639                                   tmp_cb + chroma_height * h->mb_uvlinesize,
640                                   h->mb_uvlinesize, h->chroma_log2_weight_denom,
641                                   h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
642                                   h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
643                 chroma_weight_avg(dest_cr + chroma_height * h->mb_uvlinesize,
644                                   tmp_cr + chroma_height * h->mb_uvlinesize,
645                                   h->mb_uvlinesize, h->chroma_log2_weight_denom,
646                                   h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
647                                   h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
648             }
649         }
650     }else{
651         int list = list1 ? 1 : 0;
652         int refn = h->ref_cache[list][ scan8[n] ];
653         Picture *ref= &h->ref_list[list][refn];
654         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
655                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
656                     qpix_put, chroma_put, pixel_shift, chroma444);
657
658         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
659                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
660         if(h->use_weight_chroma){
661             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
662                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
663             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
664                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
665             if (CHROMA422) {
666                 chroma_weight_op(dest_cb + chroma_height * h->mb_uvlinesize,
667                                  h->mb_uvlinesize, h->chroma_log2_weight_denom,
668                                  h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
669                 chroma_weight_op(dest_cr + chroma_height * h->mb_uvlinesize,
670                                  h->mb_uvlinesize, h->chroma_log2_weight_denom,
671                                  h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
672             }
673         }
674     }
675 }
676
677 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
678                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
679                            int x_offset, int y_offset,
680                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
681                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
682                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
683                            int list0, int list1, int pixel_shift, int chroma444){
684     if((h->use_weight==2 && list0 && list1
685         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
686        || h->use_weight==1)
687         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
688                          x_offset, y_offset, qpix_put, chroma_put,
689                          weight_op[0], weight_op[3], weight_avg[0],
690                          weight_avg[3], list0, list1, pixel_shift, chroma444);
691     else
692         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
693                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
694                     chroma_avg, list0, list1, pixel_shift, chroma444);
695 }
696
697 static inline void prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma444){
698     /* fetch pixels for estimated mv 4 macroblocks ahead
699      * optimized for 64byte cache lines */
700     MpegEncContext * const s = &h->s;
701     const int refn = h->ref_cache[list][scan8[0]];
702     if(refn >= 0){
703         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
704         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
705         uint8_t **src = h->ref_list[list][refn].f.data;
706         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
707         s->dsp.prefetch(src[0]+off, s->linesize, 4);
708         if(chroma444){
709             s->dsp.prefetch(src[1]+off, s->linesize, 4);
710             s->dsp.prefetch(src[2]+off, s->linesize, 4);
711         }else{
712             off= ((mx>>1) << pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + (64 << pixel_shift);
713             s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
714         }
715     }
716 }
717
718 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
719                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
720                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
721                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
722                       int pixel_shift, int chroma444){
723     MpegEncContext * const s = &h->s;
724     const int mb_xy= h->mb_xy;
725     const int mb_type = s->current_picture.f.mb_type[mb_xy];
726
727     assert(IS_INTER(mb_type));
728
729     if(HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
730         await_references(h);
731     prefetch_motion(h, 0, pixel_shift, chroma444);
732
733     if(IS_16X16(mb_type)){
734         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
735                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
736                 weight_op, weight_avg,
737                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
738                 pixel_shift, chroma444);
739     }else if(IS_16X8(mb_type)){
740         mc_part(h, 0, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
741                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
742                 &weight_op[1], &weight_avg[1],
743                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
744                 pixel_shift, chroma444);
745         mc_part(h, 8, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
746                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
747                 &weight_op[1], &weight_avg[1],
748                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
749                 pixel_shift, chroma444);
750     }else if(IS_8X16(mb_type)){
751         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
752                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
753                 &weight_op[2], &weight_avg[2],
754                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
755                 pixel_shift, chroma444);
756         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
757                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
758                 &weight_op[2], &weight_avg[2],
759                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
760                 pixel_shift, chroma444);
761     }else{
762         int i;
763
764         assert(IS_8X8(mb_type));
765
766         for(i=0; i<4; i++){
767             const int sub_mb_type= h->sub_mb_type[i];
768             const int n= 4*i;
769             int x_offset= (i&1)<<2;
770             int y_offset= (i&2)<<1;
771
772             if(IS_SUB_8X8(sub_mb_type)){
773                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
774                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
775                     &weight_op[3], &weight_avg[3],
776                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
777                     pixel_shift, chroma444);
778             }else if(IS_SUB_8X4(sub_mb_type)){
779                 mc_part(h, n  , 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
780                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
781                     &weight_op[4], &weight_avg[4],
782                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
783                     pixel_shift, chroma444);
784                 mc_part(h, n+2, 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
785                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
786                     &weight_op[4], &weight_avg[4],
787                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
788                     pixel_shift, chroma444);
789             }else if(IS_SUB_4X8(sub_mb_type)){
790                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
791                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
792                     &weight_op[5], &weight_avg[5],
793                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
794                     pixel_shift, chroma444);
795                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
796                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
797                     &weight_op[5], &weight_avg[5],
798                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
799                     pixel_shift, chroma444);
800             }else{
801                 int j;
802                 assert(IS_SUB_4X4(sub_mb_type));
803                 for(j=0; j<4; j++){
804                     int sub_x_offset= x_offset + 2*(j&1);
805                     int sub_y_offset= y_offset +   (j&2);
806                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
807                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
808                         &weight_op[6], &weight_avg[6],
809                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
810                         pixel_shift, chroma444);
811                 }
812             }
813         }
814     }
815
816     prefetch_motion(h, 1, pixel_shift, chroma444);
817 }
818
819 static void free_tables(H264Context *h, int free_rbsp){
820     int i;
821     H264Context *hx;
822
823     av_freep(&h->intra4x4_pred_mode);
824     av_freep(&h->chroma_pred_mode_table);
825     av_freep(&h->cbp_table);
826     av_freep(&h->mvd_table[0]);
827     av_freep(&h->mvd_table[1]);
828     av_freep(&h->direct_table);
829     av_freep(&h->non_zero_count);
830     av_freep(&h->slice_table_base);
831     h->slice_table= NULL;
832     av_freep(&h->list_counts);
833
834     av_freep(&h->mb2b_xy);
835     av_freep(&h->mb2br_xy);
836
837     for(i = 0; i < MAX_THREADS; i++) {
838         hx = h->thread_context[i];
839         if(!hx) continue;
840         av_freep(&hx->top_borders[1]);
841         av_freep(&hx->top_borders[0]);
842         av_freep(&hx->s.obmc_scratchpad);
843         if (free_rbsp){
844             av_freep(&hx->rbsp_buffer[1]);
845             av_freep(&hx->rbsp_buffer[0]);
846             hx->rbsp_buffer_size[0] = 0;
847             hx->rbsp_buffer_size[1] = 0;
848         }
849         if (i) av_freep(&h->thread_context[i]);
850     }
851 }
852
853 static void init_dequant8_coeff_table(H264Context *h){
854     int i,j,q,x;
855     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
856
857     for(i=0; i<6; i++ ){
858         h->dequant8_coeff[i] = h->dequant8_buffer[i];
859         for(j=0; j<i; j++){
860             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
861                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
862                 break;
863             }
864         }
865         if(j<i)
866             continue;
867
868         for(q=0; q<max_qp+1; q++){
869             int shift = div6[q];
870             int idx = rem6[q];
871             for(x=0; x<64; x++)
872                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
873                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
874                     h->pps.scaling_matrix8[i][x]) << shift;
875         }
876     }
877 }
878
879 static void init_dequant4_coeff_table(H264Context *h){
880     int i,j,q,x;
881     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
882     for(i=0; i<6; i++ ){
883         h->dequant4_coeff[i] = h->dequant4_buffer[i];
884         for(j=0; j<i; j++){
885             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
886                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
887                 break;
888             }
889         }
890         if(j<i)
891             continue;
892
893         for(q=0; q<max_qp+1; q++){
894             int shift = div6[q] + 2;
895             int idx = rem6[q];
896             for(x=0; x<16; x++)
897                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
898                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
899                     h->pps.scaling_matrix4[i][x]) << shift;
900         }
901     }
902 }
903
904 static void init_dequant_tables(H264Context *h){
905     int i,x;
906     init_dequant4_coeff_table(h);
907     if(h->pps.transform_8x8_mode)
908         init_dequant8_coeff_table(h);
909     if(h->sps.transform_bypass){
910         for(i=0; i<6; i++)
911             for(x=0; x<16; x++)
912                 h->dequant4_coeff[i][0][x] = 1<<6;
913         if(h->pps.transform_8x8_mode)
914             for(i=0; i<6; i++)
915                 for(x=0; x<64; x++)
916                     h->dequant8_coeff[i][0][x] = 1<<6;
917     }
918 }
919
920
921 int ff_h264_alloc_tables(H264Context *h){
922     MpegEncContext * const s = &h->s;
923     const int big_mb_num= s->mb_stride * (s->mb_height+1);
924     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
925     int x,y;
926
927     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
928
929     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
930     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
931     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
932
933     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
934     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
935     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
936     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
937     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
938
939     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
940     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
941
942     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
943     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
944     for(y=0; y<s->mb_height; y++){
945         for(x=0; x<s->mb_width; x++){
946             const int mb_xy= x + y*s->mb_stride;
947             const int b_xy = 4*x + 4*y*h->b_stride;
948
949             h->mb2b_xy [mb_xy]= b_xy;
950             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
951         }
952     }
953
954     s->obmc_scratchpad = NULL;
955
956     if(!h->dequant4_coeff[0])
957         init_dequant_tables(h);
958
959     return 0;
960 fail:
961     free_tables(h, 1);
962     return -1;
963 }
964
965 /**
966  * Mimic alloc_tables(), but for every context thread.
967  */
968 static void clone_tables(H264Context *dst, H264Context *src, int i){
969     MpegEncContext * const s = &src->s;
970     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
971     dst->non_zero_count           = src->non_zero_count;
972     dst->slice_table              = src->slice_table;
973     dst->cbp_table                = src->cbp_table;
974     dst->mb2b_xy                  = src->mb2b_xy;
975     dst->mb2br_xy                 = src->mb2br_xy;
976     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
977     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
978     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
979     dst->direct_table             = src->direct_table;
980     dst->list_counts              = src->list_counts;
981
982     dst->s.obmc_scratchpad = NULL;
983     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma, src->sps.chroma_format_idc);
984 }
985
986 /**
987  * Init context
988  * Allocate buffers which are not shared amongst multiple threads.
989  */
990 static int context_init(H264Context *h){
991     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
992     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
993
994     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
995     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
996
997     return 0;
998 fail:
999     return -1; // free_tables will clean up for us
1000 }
1001
1002 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
1003
1004 static av_cold void common_init(H264Context *h){
1005     MpegEncContext * const s = &h->s;
1006
1007     s->width = s->avctx->width;
1008     s->height = s->avctx->height;
1009     s->codec_id= s->avctx->codec->id;
1010
1011     ff_h264dsp_init(&h->h264dsp, 8, 1);
1012     ff_h264_pred_init(&h->hpc, s->codec_id, 8, 1);
1013
1014     h->dequant_coeff_pps= -1;
1015     s->unrestricted_mv=1;
1016     s->decode=1; //FIXME
1017
1018     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
1019
1020     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
1021     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
1022 }
1023
1024 int ff_h264_decode_extradata(H264Context *h)
1025 {
1026     AVCodecContext *avctx = h->s.avctx;
1027
1028     if(avctx->extradata[0] == 1){
1029         int i, cnt, nalsize;
1030         unsigned char *p = avctx->extradata;
1031
1032         h->is_avc = 1;
1033
1034         if(avctx->extradata_size < 7) {
1035             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1036             return -1;
1037         }
1038         /* sps and pps in the avcC always have length coded with 2 bytes,
1039            so put a fake nal_length_size = 2 while parsing them */
1040         h->nal_length_size = 2;
1041         // Decode sps from avcC
1042         cnt = *(p+5) & 0x1f; // Number of sps
1043         p += 6;
1044         for (i = 0; i < cnt; i++) {
1045             nalsize = AV_RB16(p) + 2;
1046             if (p - avctx->extradata + nalsize > avctx->extradata_size)
1047                 return -1;
1048             if(decode_nal_units(h, p, nalsize) < 0) {
1049                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1050                 return -1;
1051             }
1052             p += nalsize;
1053         }
1054         // Decode pps from avcC
1055         cnt = *(p++); // Number of pps
1056         for (i = 0; i < cnt; i++) {
1057             nalsize = AV_RB16(p) + 2;
1058             if (p - avctx->extradata + nalsize > avctx->extradata_size)
1059                 return -1;
1060             if (decode_nal_units(h, p, nalsize) < 0) {
1061                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1062                 return -1;
1063             }
1064             p += nalsize;
1065         }
1066         // Now store right nal length size, that will be use to parse all other nals
1067         h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1068     } else {
1069         h->is_avc = 0;
1070         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1071             return -1;
1072     }
1073     return 0;
1074 }
1075
1076 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1077     H264Context *h= avctx->priv_data;
1078     MpegEncContext * const s = &h->s;
1079
1080     MPV_decode_defaults(s);
1081
1082     s->avctx = avctx;
1083     common_init(h);
1084
1085     s->out_format = FMT_H264;
1086     s->workaround_bugs= avctx->workaround_bugs;
1087
1088     // set defaults
1089 //    s->decode_mb= ff_h263_decode_mb;
1090     s->quarter_sample = 1;
1091     if(!avctx->has_b_frames)
1092     s->low_delay= 1;
1093
1094     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1095
1096     ff_h264_decode_init_vlc();
1097
1098     h->pixel_shift = 0;
1099     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1100
1101     h->thread_context[0] = h;
1102     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1103     h->prev_poc_msb= 1<<16;
1104     h->x264_build = -1;
1105     ff_h264_reset_sei(h);
1106     if(avctx->codec_id == CODEC_ID_H264){
1107         if(avctx->ticks_per_frame == 1){
1108             s->avctx->time_base.den *=2;
1109         }
1110         avctx->ticks_per_frame = 2;
1111     }
1112
1113     if(avctx->extradata_size > 0 && avctx->extradata &&
1114         ff_h264_decode_extradata(h))
1115         return -1;
1116
1117     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1118         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1119         s->low_delay = 0;
1120     }
1121
1122     return 0;
1123 }
1124
1125 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1126 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1127 {
1128     int i;
1129
1130     for (i=0; i<count; i++){
1131         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1132                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1133                 !from[i]));
1134         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1135     }
1136 }
1137
1138 static void copy_parameter_set(void **to, void **from, int count, int size)
1139 {
1140     int i;
1141
1142     for (i=0; i<count; i++){
1143         if (to[i] && !from[i]) av_freep(&to[i]);
1144         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1145
1146         if (from[i]) memcpy(to[i], from[i], size);
1147     }
1148 }
1149
1150 static int decode_init_thread_copy(AVCodecContext *avctx){
1151     H264Context *h= avctx->priv_data;
1152
1153     if (!avctx->is_copy) return 0;
1154     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1155     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1156
1157     return 0;
1158 }
1159
1160 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1161 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1162     H264Context *h= dst->priv_data, *h1= src->priv_data;
1163     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1164     int inited = s->context_initialized, err;
1165     int i;
1166
1167     if(dst == src || !s1->context_initialized) return 0;
1168
1169     err = ff_mpeg_update_thread_context(dst, src);
1170     if(err) return err;
1171
1172     //FIXME handle width/height changing
1173     if(!inited){
1174         for(i = 0; i < MAX_SPS_COUNT; i++)
1175             av_freep(h->sps_buffers + i);
1176
1177         for(i = 0; i < MAX_PPS_COUNT; i++)
1178             av_freep(h->pps_buffers + i);
1179
1180         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1181         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1182         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1183         if (ff_h264_alloc_tables(h) < 0) {
1184             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1185             return AVERROR(ENOMEM);
1186         }
1187         context_init(h);
1188
1189         for(i=0; i<2; i++){
1190             h->rbsp_buffer[i] = NULL;
1191             h->rbsp_buffer_size[i] = 0;
1192         }
1193
1194         h->thread_context[0] = h;
1195
1196         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1197         // so this has to be allocated here
1198         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1199
1200         s->dsp.clear_blocks(h->mb);
1201         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1202     }
1203
1204     //extradata/NAL handling
1205     h->is_avc          = h1->is_avc;
1206
1207     //SPS/PPS
1208     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1209     h->sps             = h1->sps;
1210     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1211     h->pps             = h1->pps;
1212
1213     //Dequantization matrices
1214     //FIXME these are big - can they be only copied when PPS changes?
1215     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1216
1217     for(i=0; i<6; i++)
1218         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1219
1220     for(i=0; i<6; i++)
1221         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1222
1223     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1224
1225     //POC timing
1226     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1227
1228     //reference lists
1229     copy_fields(h, h1, ref_count, list_count);
1230     copy_fields(h, h1, ref_list,  intra_gb);
1231     copy_fields(h, h1, short_ref, cabac_init_idc);
1232
1233     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1234     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1235     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1236
1237     h->last_slice_type = h1->last_slice_type;
1238
1239     if(!s->current_picture_ptr) return 0;
1240
1241     if(!s->dropable) {
1242         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1243         h->prev_poc_msb     = h->poc_msb;
1244         h->prev_poc_lsb     = h->poc_lsb;
1245     }
1246     h->prev_frame_num_offset= h->frame_num_offset;
1247     h->prev_frame_num       = h->frame_num;
1248     h->outputed_poc         = h->next_outputed_poc;
1249
1250     return err;
1251 }
1252
1253 int ff_h264_frame_start(H264Context *h){
1254     MpegEncContext * const s = &h->s;
1255     int i;
1256     const int pixel_shift = h->pixel_shift;
1257     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1258
1259     if(MPV_frame_start(s, s->avctx) < 0)
1260         return -1;
1261     ff_er_frame_start(s);
1262     /*
1263      * MPV_frame_start uses pict_type to derive key_frame.
1264      * This is incorrect for H.264; IDR markings must be used.
1265      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1266      * See decode_nal_units().
1267      */
1268     s->current_picture_ptr->f.key_frame = 0;
1269     s->current_picture_ptr->mmco_reset= 0;
1270
1271     assert(s->linesize && s->uvlinesize);
1272
1273     for(i=0; i<16; i++){
1274         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1275         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1276     }
1277     for(i=0; i<16; i++){
1278         h->block_offset[16+i]=
1279         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1280         h->block_offset[48+16+i]=
1281         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1282     }
1283
1284     /* can't be in alloc_tables because linesize isn't known there.
1285      * FIXME: redo bipred weight to not require extra buffer? */
1286     for(i = 0; i < thread_count; i++)
1287         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1288             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1289
1290     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1291     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1292
1293 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1294
1295     // We mark the current picture as non-reference after allocating it, so
1296     // that if we break out due to an error it can be released automatically
1297     // in the next MPV_frame_start().
1298     // SVQ3 as well as most other codecs have only last/next/current and thus
1299     // get released even with set reference, besides SVQ3 and others do not
1300     // mark frames as reference later "naturally".
1301     if(s->codec_id != CODEC_ID_SVQ3)
1302         s->current_picture_ptr->f.reference = 0;
1303
1304     s->current_picture_ptr->field_poc[0]=
1305     s->current_picture_ptr->field_poc[1]= INT_MAX;
1306
1307     h->next_output_pic = NULL;
1308
1309     assert(s->current_picture_ptr->long_ref==0);
1310
1311     return 0;
1312 }
1313
1314 /**
1315   * Run setup operations that must be run after slice header decoding.
1316   * This includes finding the next displayed frame.
1317   *
1318   * @param h h264 master context
1319   * @param setup_finished enough NALs have been read that we can call
1320   * ff_thread_finish_setup()
1321   */
1322 static void decode_postinit(H264Context *h, int setup_finished){
1323     MpegEncContext * const s = &h->s;
1324     Picture *out = s->current_picture_ptr;
1325     Picture *cur = s->current_picture_ptr;
1326     int i, pics, out_of_order, out_idx;
1327
1328     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1329     s->current_picture_ptr->f.pict_type   = s->pict_type;
1330
1331     if (h->next_output_pic) return;
1332
1333     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1334         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1335         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1336         //to find this yet, so we assume the worst for now.
1337         //if (setup_finished)
1338         //    ff_thread_finish_setup(s->avctx);
1339         return;
1340     }
1341
1342     cur->f.interlaced_frame = 0;
1343     cur->f.repeat_pict      = 0;
1344
1345     /* Signal interlacing information externally. */
1346     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1347
1348     if(h->sps.pic_struct_present_flag){
1349         switch (h->sei_pic_struct)
1350         {
1351         case SEI_PIC_STRUCT_FRAME:
1352             break;
1353         case SEI_PIC_STRUCT_TOP_FIELD:
1354         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1355             cur->f.interlaced_frame = 1;
1356             break;
1357         case SEI_PIC_STRUCT_TOP_BOTTOM:
1358         case SEI_PIC_STRUCT_BOTTOM_TOP:
1359             if (FIELD_OR_MBAFF_PICTURE)
1360                 cur->f.interlaced_frame = 1;
1361             else
1362                 // try to flag soft telecine progressive
1363                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1364             break;
1365         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1366         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1367             // Signal the possibility of telecined film externally (pic_struct 5,6)
1368             // From these hints, let the applications decide if they apply deinterlacing.
1369             cur->f.repeat_pict = 1;
1370             break;
1371         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1372             // Force progressive here, as doubling interlaced frame is a bad idea.
1373             cur->f.repeat_pict = 2;
1374             break;
1375         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1376             cur->f.repeat_pict = 4;
1377             break;
1378         }
1379
1380         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1381             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1382     }else{
1383         /* Derive interlacing flag from used decoding process. */
1384         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1385     }
1386     h->prev_interlaced_frame = cur->f.interlaced_frame;
1387
1388     if (cur->field_poc[0] != cur->field_poc[1]){
1389         /* Derive top_field_first from field pocs. */
1390         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1391     }else{
1392         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1393             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1394             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1395               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1396                 cur->f.top_field_first = 1;
1397             else
1398                 cur->f.top_field_first = 0;
1399         }else{
1400             /* Most likely progressive */
1401             cur->f.top_field_first = 0;
1402         }
1403     }
1404
1405     //FIXME do something with unavailable reference frames
1406
1407     /* Sort B-frames into display order */
1408
1409     if(h->sps.bitstream_restriction_flag
1410        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1411         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1412         s->low_delay = 0;
1413     }
1414
1415     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1416        && !h->sps.bitstream_restriction_flag){
1417         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1418         s->low_delay= 0;
1419     }
1420
1421     pics = 0;
1422     while(h->delayed_pic[pics]) pics++;
1423
1424     assert(pics <= MAX_DELAYED_PIC_COUNT);
1425
1426     h->delayed_pic[pics++] = cur;
1427     if (cur->f.reference == 0)
1428         cur->f.reference = DELAYED_PIC_REF;
1429
1430     out = h->delayed_pic[0];
1431     out_idx = 0;
1432     for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1433         if(h->delayed_pic[i]->poc < out->poc){
1434             out = h->delayed_pic[i];
1435             out_idx = i;
1436         }
1437     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1438         h->next_outputed_poc= INT_MIN;
1439     out_of_order = out->poc < h->next_outputed_poc;
1440
1441     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1442         { }
1443     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1444        || (s->low_delay &&
1445         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1446          || cur->f.pict_type == AV_PICTURE_TYPE_B)))
1447     {
1448         s->low_delay = 0;
1449         s->avctx->has_b_frames++;
1450     }
1451
1452     if(out_of_order || pics > s->avctx->has_b_frames){
1453         out->f.reference &= ~DELAYED_PIC_REF;
1454         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1455                          // or else the first thread can release the picture and reuse it unsafely
1456         for(i=out_idx; h->delayed_pic[i]; i++)
1457             h->delayed_pic[i] = h->delayed_pic[i+1];
1458     }
1459     if(!out_of_order && pics > s->avctx->has_b_frames){
1460         h->next_output_pic = out;
1461         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1462             h->next_outputed_poc = INT_MIN;
1463         } else
1464             h->next_outputed_poc = out->poc;
1465     }else{
1466         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1467     }
1468
1469     if (setup_finished)
1470         ff_thread_finish_setup(s->avctx);
1471 }
1472
1473 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1474                                               uint8_t *src_cb, uint8_t *src_cr,
1475                                               int linesize, int uvlinesize, int simple)
1476 {
1477     MpegEncContext * const s = &h->s;
1478     uint8_t *top_border;
1479     int top_idx = 1;
1480     const int pixel_shift = h->pixel_shift;
1481     int chroma444 = CHROMA444;
1482     int chroma422 = CHROMA422;
1483
1484     src_y  -=   linesize;
1485     src_cb -= uvlinesize;
1486     src_cr -= uvlinesize;
1487
1488     if(!simple && FRAME_MBAFF){
1489         if(s->mb_y&1){
1490             if(!MB_MBAFF){
1491                 top_border = h->top_borders[0][s->mb_x];
1492                 AV_COPY128(top_border, src_y + 15*linesize);
1493                 if (pixel_shift)
1494                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1495                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1496                     if(chroma444){
1497                         if (pixel_shift){
1498                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1499                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1500                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1501                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1502                         } else {
1503                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1504                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1505                         }
1506                     } else if(chroma422) {
1507                         if (pixel_shift) {
1508                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1509                             AV_COPY128(top_border+48, src_cr + 15*uvlinesize);
1510                         } else {
1511                             AV_COPY64(top_border+16, src_cb +  15*uvlinesize);
1512                             AV_COPY64(top_border+24, src_cr +  15*uvlinesize);
1513                         }
1514                     } else {
1515                         if (pixel_shift) {
1516                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1517                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1518                         } else {
1519                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1520                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1521                         }
1522                     }
1523                 }
1524             }
1525         }else if(MB_MBAFF){
1526             top_idx = 0;
1527         }else
1528             return;
1529     }
1530
1531     top_border = h->top_borders[top_idx][s->mb_x];
1532     // There are two lines saved, the line above the the top macroblock of a pair,
1533     // and the line above the bottom macroblock
1534     AV_COPY128(top_border, src_y + 16*linesize);
1535     if (pixel_shift)
1536         AV_COPY128(top_border+16, src_y+16*linesize+16);
1537
1538     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1539         if(chroma444){
1540             if (pixel_shift){
1541                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1542                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1543                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1544                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1545             } else {
1546                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1547                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1548             }
1549         } else if(chroma422) {
1550             if (pixel_shift) {
1551                 AV_COPY128(top_border+32, src_cb+16*uvlinesize);
1552                 AV_COPY128(top_border+48, src_cr+16*uvlinesize);
1553             } else {
1554                 AV_COPY64(top_border+16, src_cb+16*uvlinesize);
1555                 AV_COPY64(top_border+24, src_cr+16*uvlinesize);
1556             }
1557         } else {
1558             if (pixel_shift) {
1559                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1560                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1561             } else {
1562                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1563                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1564             }
1565         }
1566     }
1567 }
1568
1569 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1570                                   uint8_t *src_cb, uint8_t *src_cr,
1571                                   int linesize, int uvlinesize,
1572                                   int xchg, int chroma444,
1573                                   int simple, int pixel_shift){
1574     MpegEncContext * const s = &h->s;
1575     int deblock_topleft;
1576     int deblock_top;
1577     int top_idx = 1;
1578     uint8_t *top_border_m1;
1579     uint8_t *top_border;
1580
1581     if(!simple && FRAME_MBAFF){
1582         if(s->mb_y&1){
1583             if(!MB_MBAFF)
1584                 return;
1585         }else{
1586             top_idx = MB_MBAFF ? 0 : 1;
1587         }
1588     }
1589
1590     if(h->deblocking_filter == 2) {
1591         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1592         deblock_top     = h->top_type;
1593     } else {
1594         deblock_topleft = (s->mb_x > 0);
1595         deblock_top     = (s->mb_y > !!MB_FIELD);
1596     }
1597
1598     src_y  -=   linesize + 1 + pixel_shift;
1599     src_cb -= uvlinesize + 1 + pixel_shift;
1600     src_cr -= uvlinesize + 1 + pixel_shift;
1601
1602     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1603     top_border    = h->top_borders[top_idx][s->mb_x];
1604
1605 #define XCHG(a,b,xchg)\
1606     if (pixel_shift) {\
1607         if (xchg) {\
1608             AV_SWAP64(b+0,a+0);\
1609             AV_SWAP64(b+8,a+8);\
1610         } else {\
1611             AV_COPY128(b,a); \
1612         }\
1613     } else \
1614 if (xchg) AV_SWAP64(b,a);\
1615 else      AV_COPY64(b,a);
1616
1617     if(deblock_top){
1618         if(deblock_topleft){
1619             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1620         }
1621         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1622         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1623         if(s->mb_x+1 < s->mb_width){
1624             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1625         }
1626     }
1627     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1628         if(chroma444){
1629             if(deblock_topleft){
1630                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1631                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1632             }
1633             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1634             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1635             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1636             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1637             if(s->mb_x+1 < s->mb_width){
1638                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1639                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1640             }
1641         } else {
1642             if(deblock_top){
1643                 if(deblock_topleft){
1644                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1645                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1646                 }
1647                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1648                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1649             }
1650         }
1651     }
1652 }
1653
1654 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1655     if (high_bit_depth) {
1656         return AV_RN32A(((int32_t*)mb) + index);
1657     } else
1658         return AV_RN16A(mb + index);
1659 }
1660
1661 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1662     if (high_bit_depth) {
1663         AV_WN32A(((int32_t*)mb) + index, value);
1664     } else
1665         AV_WN16A(mb + index, value);
1666 }
1667
1668 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1669                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1670 {
1671     MpegEncContext * const s = &h->s;
1672     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1673     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1674     int i;
1675     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1676     block_offset += 16*p;
1677     if(IS_INTRA4x4(mb_type)){
1678         if(simple || !s->encoding){
1679             if(IS_8x8DCT(mb_type)){
1680                 if(transform_bypass){
1681                     idct_dc_add =
1682                     idct_add    = s->dsp.add_pixels8;
1683                 }else{
1684                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1685                     idct_add    = h->h264dsp.h264_idct8_add;
1686                 }
1687                 for(i=0; i<16; i+=4){
1688                     uint8_t * const ptr= dest_y + block_offset[i];
1689                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1690                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1691                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1692                     }else{
1693                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1694                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1695                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1696                         if(nnz){
1697                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1698                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1699                             else
1700                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1701                         }
1702                     }
1703                 }
1704             }else{
1705                 if(transform_bypass){
1706                     idct_dc_add =
1707                     idct_add    = s->dsp.add_pixels4;
1708                 }else{
1709                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1710                     idct_add    = h->h264dsp.h264_idct_add;
1711                 }
1712                 for(i=0; i<16; i++){
1713                     uint8_t * const ptr= dest_y + block_offset[i];
1714                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1715
1716                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1717                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1718                     }else{
1719                         uint8_t *topright;
1720                         int nnz, tr;
1721                         uint64_t tr_high;
1722                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1723                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1724                             assert(s->mb_y || linesize <= block_offset[i]);
1725                             if(!topright_avail){
1726                                 if (pixel_shift) {
1727                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1728                                     topright= (uint8_t*) &tr_high;
1729                                 } else {
1730                                     tr= ptr[3 - linesize]*0x01010101u;
1731                                     topright= (uint8_t*) &tr;
1732                                 }
1733                             }else
1734                                 topright= ptr + (4 << pixel_shift) - linesize;
1735                         }else
1736                             topright= NULL;
1737
1738                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1739                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1740                         if(nnz){
1741                             if(is_h264){
1742                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1743                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1744                                 else
1745                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1746                             }else
1747                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1748                         }
1749                     }
1750                 }
1751             }
1752         }
1753     }else{
1754         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1755         if(is_h264){
1756             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1757                 if(!transform_bypass)
1758                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1759                 else{
1760                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1761                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1762                     for(i = 0; i < 16; i++)
1763                         dctcoef_set(h->mb+p*256, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1764                 }
1765             }
1766         }else
1767             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1768     }
1769 }
1770
1771 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1772                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1773 {
1774     MpegEncContext * const s = &h->s;
1775     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1776     int i;
1777     block_offset += 16*p;
1778     if(!IS_INTRA4x4(mb_type)){
1779         if(is_h264){
1780             if(IS_INTRA16x16(mb_type)){
1781                 if(transform_bypass){
1782                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1783                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1784                     }else{
1785                         for(i=0; i<16; i++){
1786                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1787                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1788                         }
1789                     }
1790                 }else{
1791                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1792                 }
1793             }else if(h->cbp&15){
1794                 if(transform_bypass){
1795                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1796                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1797                     for(i=0; i<16; i+=di){
1798                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1799                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1800                         }
1801                     }
1802                 }else{
1803                     if(IS_8x8DCT(mb_type)){
1804                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1805                     }else{
1806                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1807                     }
1808                 }
1809             }
1810         }else{
1811             for(i=0; i<16; i++){
1812                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1813                     uint8_t * const ptr= dest_y + block_offset[i];
1814                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1815                 }
1816             }
1817         }
1818     }
1819 }
1820
1821 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1822     MpegEncContext * const s = &h->s;
1823     const int mb_x= s->mb_x;
1824     const int mb_y= s->mb_y;
1825     const int mb_xy= h->mb_xy;
1826     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1827     uint8_t  *dest_y, *dest_cb, *dest_cr;
1828     int linesize, uvlinesize /*dct_offset*/;
1829     int i, j;
1830     int *block_offset = &h->block_offset[0];
1831     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1832     /* is_h264 should always be true if SVQ3 is disabled. */
1833     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1834     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1835     const int block_h = 16 >> s->chroma_y_shift;
1836
1837     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1838     dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1839     dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1840
1841     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1842     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1843
1844     h->list_counts[mb_xy]= h->list_count;
1845
1846     if (!simple && MB_FIELD) {
1847         linesize   = h->mb_linesize   = s->linesize * 2;
1848         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1849         block_offset = &h->block_offset[48];
1850         if(mb_y&1){ //FIXME move out of this function?
1851             dest_y -= s->linesize*15;
1852             dest_cb-= s->uvlinesize * (block_h - 1);
1853             dest_cr-= s->uvlinesize * (block_h - 1);
1854         }
1855         if(FRAME_MBAFF) {
1856             int list;
1857             for(list=0; list<h->list_count; list++){
1858                 if(!USES_LIST(mb_type, list))
1859                     continue;
1860                 if(IS_16X16(mb_type)){
1861                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1862                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1863                 }else{
1864                     for(i=0; i<16; i+=4){
1865                         int ref = h->ref_cache[list][scan8[i]];
1866                         if(ref >= 0)
1867                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1868                     }
1869                 }
1870             }
1871         }
1872     } else {
1873         linesize   = h->mb_linesize   = s->linesize;
1874         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1875 //        dct_offset = s->linesize * 16;
1876     }
1877
1878     if (!simple && IS_INTRA_PCM(mb_type)) {
1879         if (pixel_shift) {
1880             const int bit_depth = h->sps.bit_depth_luma;
1881             int j;
1882             GetBitContext gb;
1883             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1884
1885             for (i = 0; i < 16; i++) {
1886                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1887                 for (j = 0; j < 16; j++)
1888                     tmp_y[j] = get_bits(&gb, bit_depth);
1889             }
1890             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1891                 if (!h->sps.chroma_format_idc) {
1892                     for (i = 0; i < block_h; i++) {
1893                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1894                         for (j = 0; j < 8; j++) {
1895                             tmp_cb[j] = 1 << (bit_depth - 1);
1896                         }
1897                     }
1898                     for (i = 0; i < block_h; i++) {
1899                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1900                         for (j = 0; j < 8; j++) {
1901                             tmp_cr[j] = 1 << (bit_depth - 1);
1902                         }
1903                     }
1904                 } else {
1905                     for (i = 0; i < block_h; i++) {
1906                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1907                         for (j = 0; j < 8; j++)
1908                             tmp_cb[j] = get_bits(&gb, bit_depth);
1909                     }
1910                     for (i = 0; i < block_h; i++) {
1911                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1912                         for (j = 0; j < 8; j++)
1913                             tmp_cr[j] = get_bits(&gb, bit_depth);
1914                     }
1915                 }
1916             }
1917         } else {
1918             for (i=0; i<16; i++) {
1919                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1920             }
1921             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1922                 if (!h->sps.chroma_format_idc) {
1923                     for (i = 0; i < block_h; i++) {
1924                         memset(dest_cb + i*uvlinesize, 128, 8);
1925                         memset(dest_cr + i*uvlinesize, 128, 8);
1926                     }
1927                 } else {
1928                     for (i = 0; i < block_h; i++) {
1929                         memcpy(dest_cb + i*uvlinesize, h->mb + 128 + i*4,  8);
1930                         memcpy(dest_cr + i*uvlinesize, h->mb + 160 + i*4,  8);
1931                     }
1932                 }
1933             }
1934         }
1935     } else {
1936         if(IS_INTRA(mb_type)){
1937             if(h->deblocking_filter)
1938                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1939
1940             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1941                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1942                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1943             }
1944
1945             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1946
1947             if(h->deblocking_filter)
1948                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1949         }else if(is_h264){
1950             hl_motion(h, dest_y, dest_cb, dest_cr,
1951                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1952                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1953                       h->h264dsp.weight_h264_pixels_tab,
1954                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 0);
1955         }
1956
1957         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1958
1959         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1960             uint8_t *dest[2] = {dest_cb, dest_cr};
1961             if(transform_bypass){
1962                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1963                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
1964                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
1965                 }else{
1966                     idct_add = s->dsp.add_pixels4;
1967                     for(j=1; j<3; j++){
1968                         for(i=j*16; i<j*16+4; i++){
1969                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1970                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1971                         }
1972                         if (CHROMA422) {
1973                             for(i=j*16+4; i<j*16+8; i++){
1974                                 if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1975                                     idct_add   (dest[j-1] + block_offset[i+4], h->mb + (i*16 << pixel_shift), uvlinesize);
1976                             }
1977                         }
1978                     }
1979                 }
1980             }else{
1981                 if(is_h264){
1982                     int qp[2];
1983                     if (CHROMA422) {
1984                         qp[0] = h->chroma_qp[0] + 3;
1985                         qp[1] = h->chroma_qp[1] + 3;
1986                     } else {
1987                         qp[0] = h->chroma_qp[0];
1988                         qp[1] = h->chroma_qp[1];
1989                     }
1990                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1991                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*1 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][qp[0]][0]);
1992                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1993                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*2 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][qp[1]][0]);
1994                     h->h264dsp.h264_idct_add8(dest, block_offset,
1995                                               h->mb, uvlinesize,
1996                                               h->non_zero_count_cache);
1997                 }else{
1998                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16*16*1, h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1999                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16*16*2, h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
2000                     for(j=1; j<3; j++){
2001                         for(i=j*16; i<j*16+4; i++){
2002                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2003                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
2004                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
2005                             }
2006                         }
2007                     }
2008                 }
2009             }
2010         }
2011     }
2012     if(h->cbp || IS_INTRA(mb_type))
2013     {
2014         s->dsp.clear_blocks(h->mb);
2015         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2016     }
2017 }
2018
2019 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
2020     MpegEncContext * const s = &h->s;
2021     const int mb_x= s->mb_x;
2022     const int mb_y= s->mb_y;
2023     const int mb_xy= h->mb_xy;
2024     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2025     uint8_t  *dest[3];
2026     int linesize;
2027     int i, j, p;
2028     int *block_offset = &h->block_offset[0];
2029     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2030     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
2031
2032     for (p = 0; p < plane_count; p++)
2033     {
2034         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
2035         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
2036     }
2037
2038     h->list_counts[mb_xy]= h->list_count;
2039
2040     if (!simple && MB_FIELD) {
2041         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
2042         block_offset = &h->block_offset[48];
2043         if(mb_y&1) //FIXME move out of this function?
2044             for (p = 0; p < 3; p++)
2045                 dest[p] -= s->linesize*15;
2046         if(FRAME_MBAFF) {
2047             int list;
2048             for(list=0; list<h->list_count; list++){
2049                 if(!USES_LIST(mb_type, list))
2050                     continue;
2051                 if(IS_16X16(mb_type)){
2052                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2053                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2054                 }else{
2055                     for(i=0; i<16; i+=4){
2056                         int ref = h->ref_cache[list][scan8[i]];
2057                         if(ref >= 0)
2058                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2059                     }
2060                 }
2061             }
2062         }
2063     } else {
2064         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2065     }
2066
2067     if (!simple && IS_INTRA_PCM(mb_type)) {
2068         if (pixel_shift) {
2069             const int bit_depth = h->sps.bit_depth_luma;
2070             GetBitContext gb;
2071             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
2072
2073             for (p = 0; p < plane_count; p++) {
2074                 for (i = 0; i < 16; i++) {
2075                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2076                     for (j = 0; j < 16; j++)
2077                         tmp[j] = get_bits(&gb, bit_depth);
2078                 }
2079             }
2080         } else {
2081             for (p = 0; p < plane_count; p++) {
2082                 for (i = 0; i < 16; i++) {
2083                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2084                 }
2085             }
2086         }
2087     } else {
2088         if(IS_INTRA(mb_type)){
2089             if(h->deblocking_filter)
2090                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2091
2092             for (p = 0; p < plane_count; p++)
2093                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2094
2095             if(h->deblocking_filter)
2096                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2097         }else{
2098             hl_motion(h, dest[0], dest[1], dest[2],
2099                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2100                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2101                       h->h264dsp.weight_h264_pixels_tab,
2102                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 1);
2103         }
2104
2105         for (p = 0; p < plane_count; p++)
2106             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2107     }
2108     if(h->cbp || IS_INTRA(mb_type))
2109     {
2110         s->dsp.clear_blocks(h->mb);
2111         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2112     }
2113 }
2114
2115 /**
2116  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2117  */
2118 #define hl_decode_mb_simple(sh, bits) \
2119 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2120     hl_decode_mb_internal(h, 1, sh); \
2121 }
2122 hl_decode_mb_simple(0, 8);
2123 hl_decode_mb_simple(1, 16);
2124
2125 /**
2126  * Process a macroblock; this handles edge cases, such as interlacing.
2127  */
2128 static void av_noinline hl_decode_mb_complex(H264Context *h){
2129     hl_decode_mb_internal(h, 0, h->pixel_shift);
2130 }
2131
2132 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2133     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2134 }
2135
2136 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2137     hl_decode_mb_444_internal(h, 1, 0);
2138 }
2139
2140 void ff_h264_hl_decode_mb(H264Context *h){
2141     MpegEncContext * const s = &h->s;
2142     const int mb_xy= h->mb_xy;
2143     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2144     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2145
2146     if (CHROMA444) {
2147         if(is_complex || h->pixel_shift)
2148             hl_decode_mb_444_complex(h);
2149         else
2150             hl_decode_mb_444_simple(h);
2151     } else if (is_complex) {
2152         hl_decode_mb_complex(h);
2153     } else if (h->pixel_shift) {
2154         hl_decode_mb_simple_16(h);
2155     } else
2156         hl_decode_mb_simple_8(h);
2157 }
2158
2159 static int pred_weight_table(H264Context *h){
2160     MpegEncContext * const s = &h->s;
2161     int list, i;
2162     int luma_def, chroma_def;
2163
2164     h->use_weight= 0;
2165     h->use_weight_chroma= 0;
2166     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2167     if(h->sps.chroma_format_idc)
2168         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2169     luma_def = 1<<h->luma_log2_weight_denom;
2170     chroma_def = 1<<h->chroma_log2_weight_denom;
2171
2172     for(list=0; list<2; list++){
2173         h->luma_weight_flag[list]   = 0;
2174         h->chroma_weight_flag[list] = 0;
2175         for(i=0; i<h->ref_count[list]; i++){
2176             int luma_weight_flag, chroma_weight_flag;
2177
2178             luma_weight_flag= get_bits1(&s->gb);
2179             if(luma_weight_flag){
2180                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2181                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2182                 if(   h->luma_weight[i][list][0] != luma_def
2183                    || h->luma_weight[i][list][1] != 0) {
2184                     h->use_weight= 1;
2185                     h->luma_weight_flag[list]= 1;
2186                 }
2187             }else{
2188                 h->luma_weight[i][list][0]= luma_def;
2189                 h->luma_weight[i][list][1]= 0;
2190             }
2191
2192             if(h->sps.chroma_format_idc){
2193                 chroma_weight_flag= get_bits1(&s->gb);
2194                 if(chroma_weight_flag){
2195                     int j;
2196                     for(j=0; j<2; j++){
2197                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2198                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2199                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2200                            || h->chroma_weight[i][list][j][1] != 0) {
2201                             h->use_weight_chroma= 1;
2202                             h->chroma_weight_flag[list]= 1;
2203                         }
2204                     }
2205                 }else{
2206                     int j;
2207                     for(j=0; j<2; j++){
2208                         h->chroma_weight[i][list][j][0]= chroma_def;
2209                         h->chroma_weight[i][list][j][1]= 0;
2210                     }
2211                 }
2212             }
2213         }
2214         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2215     }
2216     h->use_weight= h->use_weight || h->use_weight_chroma;
2217     return 0;
2218 }
2219
2220 /**
2221  * Initialize implicit_weight table.
2222  * @param field  0/1 initialize the weight for interlaced MBAFF
2223  *                -1 initializes the rest
2224  */
2225 static void implicit_weight_table(H264Context *h, int field){
2226     MpegEncContext * const s = &h->s;
2227     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2228
2229     for (i = 0; i < 2; i++) {
2230         h->luma_weight_flag[i]   = 0;
2231         h->chroma_weight_flag[i] = 0;
2232     }
2233
2234     if(field < 0){
2235         if (s->picture_structure == PICT_FRAME) {
2236             cur_poc = s->current_picture_ptr->poc;
2237         } else {
2238             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2239         }
2240     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2241        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2242         h->use_weight= 0;
2243         h->use_weight_chroma= 0;
2244         return;
2245     }
2246         ref_start= 0;
2247         ref_count0= h->ref_count[0];
2248         ref_count1= h->ref_count[1];
2249     }else{
2250         cur_poc = s->current_picture_ptr->field_poc[field];
2251         ref_start= 16;
2252         ref_count0= 16+2*h->ref_count[0];
2253         ref_count1= 16+2*h->ref_count[1];
2254     }
2255
2256     h->use_weight= 2;
2257     h->use_weight_chroma= 2;
2258     h->luma_log2_weight_denom= 5;
2259     h->chroma_log2_weight_denom= 5;
2260
2261     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2262         int poc0 = h->ref_list[0][ref0].poc;
2263         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2264             int w = 32;
2265             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2266                 int poc1 = h->ref_list[1][ref1].poc;
2267                 int td = av_clip(poc1 - poc0, -128, 127);
2268                 if(td){
2269                     int tb = av_clip(cur_poc - poc0, -128, 127);
2270                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2271                     int dist_scale_factor = (tb*tx + 32) >> 8;
2272                     if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2273                         w = 64 - dist_scale_factor;
2274                 }
2275             }
2276             if(field<0){
2277                 h->implicit_weight[ref0][ref1][0]=
2278                 h->implicit_weight[ref0][ref1][1]= w;
2279             }else{
2280                 h->implicit_weight[ref0][ref1][field]=w;
2281             }
2282         }
2283     }
2284 }
2285
2286 /**
2287  * instantaneous decoder refresh.
2288  */
2289 static void idr(H264Context *h){
2290     ff_h264_remove_all_refs(h);
2291     h->prev_frame_num= 0;
2292     h->prev_frame_num_offset= 0;
2293     h->prev_poc_msb=
2294     h->prev_poc_lsb= 0;
2295 }
2296
2297 /* forget old pics after a seek */
2298 static void flush_dpb(AVCodecContext *avctx){
2299     H264Context *h= avctx->priv_data;
2300     int i;
2301     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
2302         if(h->delayed_pic[i])
2303             h->delayed_pic[i]->f.reference = 0;
2304         h->delayed_pic[i]= NULL;
2305     }
2306     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2307     h->prev_interlaced_frame = 1;
2308     idr(h);
2309     if(h->s.current_picture_ptr)
2310         h->s.current_picture_ptr->f.reference = 0;
2311     h->s.first_field= 0;
2312     ff_h264_reset_sei(h);
2313     ff_mpeg_flush(avctx);
2314 }
2315
2316 static int init_poc(H264Context *h){
2317     MpegEncContext * const s = &h->s;
2318     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2319     int field_poc[2];
2320     Picture *cur = s->current_picture_ptr;
2321
2322     h->frame_num_offset= h->prev_frame_num_offset;
2323     if(h->frame_num < h->prev_frame_num)
2324         h->frame_num_offset += max_frame_num;
2325
2326     if(h->sps.poc_type==0){
2327         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2328
2329         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2330             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2331         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2332             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2333         else
2334             h->poc_msb = h->prev_poc_msb;
2335 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2336         field_poc[0] =
2337         field_poc[1] = h->poc_msb + h->poc_lsb;
2338         if(s->picture_structure == PICT_FRAME)
2339             field_poc[1] += h->delta_poc_bottom;
2340     }else if(h->sps.poc_type==1){
2341         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2342         int i;
2343
2344         if(h->sps.poc_cycle_length != 0)
2345             abs_frame_num = h->frame_num_offset + h->frame_num;
2346         else
2347             abs_frame_num = 0;
2348
2349         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2350             abs_frame_num--;
2351
2352         expected_delta_per_poc_cycle = 0;
2353         for(i=0; i < h->sps.poc_cycle_length; i++)
2354             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2355
2356         if(abs_frame_num > 0){
2357             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2358             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2359
2360             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2361             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2362                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2363         } else
2364             expectedpoc = 0;
2365
2366         if(h->nal_ref_idc == 0)
2367             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2368
2369         field_poc[0] = expectedpoc + h->delta_poc[0];
2370         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2371
2372         if(s->picture_structure == PICT_FRAME)
2373             field_poc[1] += h->delta_poc[1];
2374     }else{
2375         int poc= 2*(h->frame_num_offset + h->frame_num);
2376
2377         if(!h->nal_ref_idc)
2378             poc--;
2379
2380         field_poc[0]= poc;
2381         field_poc[1]= poc;
2382     }
2383
2384     if(s->picture_structure != PICT_BOTTOM_FIELD)
2385         s->current_picture_ptr->field_poc[0]= field_poc[0];
2386     if(s->picture_structure != PICT_TOP_FIELD)
2387         s->current_picture_ptr->field_poc[1]= field_poc[1];
2388     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2389
2390     return 0;
2391 }
2392
2393
2394 /**
2395  * initialize scan tables
2396  */
2397 static void init_scan_tables(H264Context *h){
2398     int i;
2399     for(i=0; i<16; i++){
2400 #define T(x) (x>>2) | ((x<<2) & 0xF)
2401         h->zigzag_scan[i] = T(zigzag_scan[i]);
2402         h-> field_scan[i] = T( field_scan[i]);
2403 #undef T
2404     }
2405     for(i=0; i<64; i++){
2406 #define T(x) (x>>3) | ((x&7)<<3)
2407         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2408         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2409         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2410         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2411 #undef T
2412     }
2413     if(h->sps.transform_bypass){ //FIXME same ugly
2414         h->zigzag_scan_q0          = zigzag_scan;
2415         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2416         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2417         h->field_scan_q0           = field_scan;
2418         h->field_scan8x8_q0        = field_scan8x8;
2419         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2420     }else{
2421         h->zigzag_scan_q0          = h->zigzag_scan;
2422         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2423         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2424         h->field_scan_q0           = h->field_scan;
2425         h->field_scan8x8_q0        = h->field_scan8x8;
2426         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2427     }
2428 }
2429
2430 static int field_end(H264Context *h, int in_setup){
2431     MpegEncContext * const s = &h->s;
2432     AVCodecContext * const avctx= s->avctx;
2433     int err = 0;
2434     s->mb_y= 0;
2435
2436     if (!in_setup && !s->dropable)
2437         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2438                                  s->picture_structure==PICT_BOTTOM_FIELD);
2439
2440     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2441         ff_vdpau_h264_set_reference_frames(s);
2442
2443     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2444         if(!s->dropable) {
2445             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2446             h->prev_poc_msb= h->poc_msb;
2447             h->prev_poc_lsb= h->poc_lsb;
2448         }
2449         h->prev_frame_num_offset= h->frame_num_offset;
2450         h->prev_frame_num= h->frame_num;
2451         h->outputed_poc = h->next_outputed_poc;
2452     }
2453
2454     if (avctx->hwaccel) {
2455         if (avctx->hwaccel->end_frame(avctx) < 0)
2456             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2457     }
2458
2459     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2460         ff_vdpau_h264_picture_complete(s);
2461
2462     /*
2463      * FIXME: Error handling code does not seem to support interlaced
2464      * when slices span multiple rows
2465      * The ff_er_add_slice calls don't work right for bottom
2466      * fields; they cause massive erroneous error concealing
2467      * Error marking covers both fields (top and bottom).
2468      * This causes a mismatched s->error_count
2469      * and a bad error table. Further, the error count goes to
2470      * INT_MAX when called for bottom field, because mb_y is
2471      * past end by one (callers fault) and resync_mb_y != 0
2472      * causes problems for the first MB line, too.
2473      */
2474     if (!FIELD_PICTURE)
2475         ff_er_frame_end(s);
2476
2477     MPV_frame_end(s);
2478
2479     h->current_slice=0;
2480
2481     return err;
2482 }
2483
2484 /**
2485  * Replicate H264 "master" context to thread contexts.
2486  */
2487 static void clone_slice(H264Context *dst, H264Context *src)
2488 {
2489     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2490     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2491     dst->s.current_picture      = src->s.current_picture;
2492     dst->s.linesize             = src->s.linesize;
2493     dst->s.uvlinesize           = src->s.uvlinesize;
2494     dst->s.first_field          = src->s.first_field;
2495
2496     dst->prev_poc_msb           = src->prev_poc_msb;
2497     dst->prev_poc_lsb           = src->prev_poc_lsb;
2498     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2499     dst->prev_frame_num         = src->prev_frame_num;
2500     dst->short_ref_count        = src->short_ref_count;
2501
2502     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2503     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2504     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2505     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2506
2507     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2508     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2509 }
2510
2511 /**
2512  * computes profile from profile_idc and constraint_set?_flags
2513  *
2514  * @param sps SPS
2515  *
2516  * @return profile as defined by FF_PROFILE_H264_*
2517  */
2518 int ff_h264_get_profile(SPS *sps)
2519 {
2520     int profile = sps->profile_idc;
2521
2522     switch(sps->profile_idc) {
2523     case FF_PROFILE_H264_BASELINE:
2524         // constraint_set1_flag set to 1
2525         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2526         break;
2527     case FF_PROFILE_H264_HIGH_10:
2528     case FF_PROFILE_H264_HIGH_422:
2529     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2530         // constraint_set3_flag set to 1
2531         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2532         break;
2533     }
2534
2535     return profile;
2536 }
2537
2538 /**
2539  * decodes a slice header.
2540  * This will also call MPV_common_init() and frame_start() as needed.
2541  *
2542  * @param h h264context
2543  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2544  *
2545  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2546  */
2547 static int decode_slice_header(H264Context *h, H264Context *h0){
2548     MpegEncContext * const s = &h->s;
2549     MpegEncContext * const s0 = &h0->s;
2550     unsigned int first_mb_in_slice;
2551     unsigned int pps_id;
2552     int num_ref_idx_active_override_flag;
2553     unsigned int slice_type, tmp, i, j;
2554     int default_ref_list_done = 0;
2555     int last_pic_structure;
2556
2557     s->dropable= h->nal_ref_idc == 0;
2558
2559     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2560     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2561         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2562         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2563     }else{
2564         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2565         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2566     }
2567
2568     first_mb_in_slice= get_ue_golomb(&s->gb);
2569
2570     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2571         if(h0->current_slice && FIELD_PICTURE){
2572             field_end(h, 1);
2573         }
2574
2575         h0->current_slice = 0;
2576         if (!s0->first_field)
2577             s->current_picture_ptr= NULL;
2578     }
2579
2580     slice_type= get_ue_golomb_31(&s->gb);
2581     if(slice_type > 9){
2582         av_log(h->s.avctx, AV_LOG_ERROR, "slice type too large (%d) at %d %d\n", h->slice_type, s->mb_x, s->mb_y);
2583         return -1;
2584     }
2585     if(slice_type > 4){
2586         slice_type -= 5;
2587         h->slice_type_fixed=1;
2588     }else
2589         h->slice_type_fixed=0;
2590
2591     slice_type= golomb_to_pict_type[ slice_type ];
2592     if (slice_type == AV_PICTURE_TYPE_I
2593         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2594         default_ref_list_done = 1;
2595     }
2596     h->slice_type= slice_type;
2597     h->slice_type_nos= slice_type & 3;
2598
2599     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2600
2601     pps_id= get_ue_golomb(&s->gb);
2602     if(pps_id>=MAX_PPS_COUNT){
2603         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2604         return -1;
2605     }
2606     if(!h0->pps_buffers[pps_id]) {
2607         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2608         return -1;
2609     }
2610     h->pps= *h0->pps_buffers[pps_id];
2611
2612     if(!h0->sps_buffers[h->pps.sps_id]) {
2613         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2614         return -1;
2615     }
2616     h->sps = *h0->sps_buffers[h->pps.sps_id];
2617
2618     s->avctx->profile = ff_h264_get_profile(&h->sps);
2619     s->avctx->level   = h->sps.level_idc;
2620     s->avctx->refs    = h->sps.ref_frame_count;
2621
2622     if(h == h0 && h->dequant_coeff_pps != pps_id){
2623         h->dequant_coeff_pps = pps_id;
2624         init_dequant_tables(h);
2625     }
2626
2627     s->mb_width= h->sps.mb_width;
2628     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2629
2630     h->b_stride=  s->mb_width*4;
2631
2632     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2633
2634     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2635     if(h->sps.frame_mbs_only_flag)
2636         s->height= 16*s->mb_height - (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2637     else
2638         s->height= 16*s->mb_height - (2<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2639
2640     if (s->context_initialized
2641         && (   s->width != s->avctx->width || s->height != s->avctx->height
2642             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2643         if(h != h0) {
2644             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2645             return -1;   // width / height changed during parallelized decoding
2646         }
2647         free_tables(h, 0);
2648         flush_dpb(s->avctx);
2649         MPV_common_end(s);
2650     }
2651     if (!s->context_initialized) {
2652         if (h != h0) {
2653             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2654             return -1;
2655         }
2656
2657         avcodec_set_dimensions(s->avctx, s->width, s->height);
2658         s->avctx->sample_aspect_ratio= h->sps.sar;
2659         av_assert0(s->avctx->sample_aspect_ratio.den);
2660
2661         if(h->sps.video_signal_type_present_flag){
2662             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2663             if(h->sps.colour_description_present_flag){
2664                 s->avctx->color_primaries = h->sps.color_primaries;
2665                 s->avctx->color_trc       = h->sps.color_trc;
2666                 s->avctx->colorspace      = h->sps.colorspace;
2667             }
2668         }
2669
2670         if(h->sps.timing_info_present_flag){
2671             int64_t den= h->sps.time_scale;
2672             if(h->x264_build < 44U)
2673                 den *= 2;
2674             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2675                       h->sps.num_units_in_tick, den, 1<<30);
2676         }
2677
2678         switch (h->sps.bit_depth_luma) {
2679             case 9 :
2680                 if (CHROMA444)
2681                     s->avctx->pix_fmt = PIX_FMT_YUV444P9;
2682                 else if (CHROMA422)
2683                     s->avctx->pix_fmt = PIX_FMT_YUV422P9;
2684                 else
2685                     s->avctx->pix_fmt = PIX_FMT_YUV420P9;
2686                 break;
2687             case 10 :
2688                 if (CHROMA444)
2689                     s->avctx->pix_fmt = PIX_FMT_YUV444P10;
2690                 else if (CHROMA422)
2691                     s->avctx->pix_fmt = PIX_FMT_YUV422P10;
2692                 else
2693                     s->avctx->pix_fmt = PIX_FMT_YUV420P10;
2694                 break;
2695             default:
2696                 if (CHROMA444){
2697                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2698                 } else if (CHROMA422) {
2699                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P : PIX_FMT_YUV422P;
2700                 }else{
2701                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2702                                                              s->avctx->codec->pix_fmts ?
2703                                                              s->avctx->codec->pix_fmts :
2704                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2705                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2706                                                              ff_hwaccel_pixfmt_list_420);
2707                 }
2708         }
2709
2710         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2711
2712         if (MPV_common_init(s) < 0) {
2713             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2714             return -1;
2715         }
2716         s->first_field = 0;
2717         h->prev_interlaced_frame = 1;
2718
2719         init_scan_tables(h);
2720         if (ff_h264_alloc_tables(h) < 0) {
2721             av_log(h->s.avctx, AV_LOG_ERROR, "Could not allocate memory for h264\n");
2722             return AVERROR(ENOMEM);
2723         }
2724
2725         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2726             if (context_init(h) < 0) {
2727                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2728                 return -1;
2729             }
2730         } else {
2731             for(i = 1; i < s->avctx->thread_count; i++) {
2732                 H264Context *c;
2733                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2734                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2735                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2736                 c->h264dsp = h->h264dsp;
2737                 c->sps = h->sps;
2738                 c->pps = h->pps;
2739                 c->pixel_shift = h->pixel_shift;
2740                 init_scan_tables(c);
2741                 clone_tables(c, h, i);
2742             }
2743
2744             for(i = 0; i < s->avctx->thread_count; i++)
2745                 if (context_init(h->thread_context[i]) < 0) {
2746                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2747                     return -1;
2748                 }
2749         }
2750     }
2751
2752     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2753
2754     h->mb_mbaff = 0;
2755     h->mb_aff_frame = 0;
2756     last_pic_structure = s0->picture_structure;
2757     if(h->sps.frame_mbs_only_flag){
2758         s->picture_structure= PICT_FRAME;
2759     }else{
2760         if(get_bits1(&s->gb)) { //field_pic_flag
2761             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2762         } else {
2763             s->picture_structure= PICT_FRAME;
2764             h->mb_aff_frame = h->sps.mb_aff;
2765         }
2766     }
2767     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2768
2769     if(h0->current_slice == 0){
2770         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2771         if(h->frame_num != h->prev_frame_num) {
2772             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2773
2774             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2775
2776             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2777                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2778                 if (unwrap_prev_frame_num < 0)
2779                     unwrap_prev_frame_num += max_frame_num;
2780
2781                 h->prev_frame_num = unwrap_prev_frame_num;
2782             }
2783         }
2784
2785         while(h->frame_num !=  h->prev_frame_num &&
2786               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2787             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2788             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2789             if (ff_h264_frame_start(h) < 0)
2790                 return -1;
2791             h->prev_frame_num++;
2792             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2793             s->current_picture_ptr->frame_num= h->prev_frame_num;
2794             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2795             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2796             ff_generate_sliding_window_mmcos(h);
2797             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2798                 s->avctx->error_recognition >= FF_ER_EXPLODE)
2799                 return AVERROR_INVALIDDATA;
2800             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2801              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2802              * about there being no actual duplicates.
2803              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2804              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2805              * be fixed. */
2806             if (h->short_ref_count) {
2807                 if (prev) {
2808                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2809                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2810                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2811                     h->short_ref[0]->poc = prev->poc+2;
2812                 }
2813                 h->short_ref[0]->frame_num = h->prev_frame_num;
2814             }
2815         }
2816
2817         /* See if we have a decoded first field looking for a pair... */
2818         if (s0->first_field) {
2819             assert(s0->current_picture_ptr);
2820             assert(s0->current_picture_ptr->f.data[0]);
2821             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2822
2823             /* figure out if we have a complementary field pair */
2824             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2825                 /*
2826                  * Previous field is unmatched. Don't display it, but let it
2827                  * remain for reference if marked as such.
2828                  */
2829                 s0->current_picture_ptr = NULL;
2830                 s0->first_field = FIELD_PICTURE;
2831
2832             } else {
2833                 if (h->nal_ref_idc &&
2834                         s0->current_picture_ptr->f.reference &&
2835                         s0->current_picture_ptr->frame_num != h->frame_num) {
2836                     /*
2837                      * This and previous field were reference, but had
2838                      * different frame_nums. Consider this field first in
2839                      * pair. Throw away previous field except for reference
2840                      * purposes.
2841                      */
2842                     s0->first_field = 1;
2843                     s0->current_picture_ptr = NULL;
2844
2845                 } else {
2846                     /* Second field in complementary pair */
2847                     s0->first_field = 0;
2848                 }
2849             }
2850
2851         } else {
2852             /* Frame or first field in a potentially complementary pair */
2853             assert(!s0->current_picture_ptr);
2854             s0->first_field = FIELD_PICTURE;
2855         }
2856
2857         if(!FIELD_PICTURE || s0->first_field) {
2858             if (ff_h264_frame_start(h) < 0) {
2859                 s0->first_field = 0;
2860                 return -1;
2861             }
2862         } else {
2863             ff_release_unused_pictures(s, 0);
2864         }
2865     }
2866     if(h != h0)
2867         clone_slice(h, h0);
2868
2869     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2870
2871     assert(s->mb_num == s->mb_width * s->mb_height);
2872     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2873        first_mb_in_slice                    >= s->mb_num){
2874         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2875         return -1;
2876     }
2877     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2878     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2879     if (s->picture_structure == PICT_BOTTOM_FIELD)
2880         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2881     assert(s->mb_y < s->mb_height);
2882
2883     if(s->picture_structure==PICT_FRAME){
2884         h->curr_pic_num=   h->frame_num;
2885         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2886     }else{
2887         h->curr_pic_num= 2*h->frame_num + 1;
2888         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2889     }
2890
2891     if(h->nal_unit_type == NAL_IDR_SLICE){
2892         get_ue_golomb(&s->gb); /* idr_pic_id */
2893     }
2894
2895     if(h->sps.poc_type==0){
2896         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2897
2898         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2899             h->delta_poc_bottom= get_se_golomb(&s->gb);
2900         }
2901     }
2902
2903     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2904         h->delta_poc[0]= get_se_golomb(&s->gb);
2905
2906         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2907             h->delta_poc[1]= get_se_golomb(&s->gb);
2908     }
2909
2910     init_poc(h);
2911
2912     if(h->pps.redundant_pic_cnt_present){
2913         h->redundant_pic_count= get_ue_golomb(&s->gb);
2914     }
2915
2916     //set defaults, might be overridden a few lines later
2917     h->ref_count[0]= h->pps.ref_count[0];
2918     h->ref_count[1]= h->pps.ref_count[1];
2919
2920     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2921         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2922             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2923         }
2924         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2925
2926         if(num_ref_idx_active_override_flag){
2927             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2928             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2929                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2930
2931             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2932                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2933                 h->ref_count[0]= h->ref_count[1]= 1;
2934                 return -1;
2935             }
2936         }
2937         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2938             h->list_count= 2;
2939         else
2940             h->list_count= 1;
2941     }else
2942         h->list_count= 0;
2943
2944     if(!default_ref_list_done){
2945         ff_h264_fill_default_ref_list(h);
2946     }
2947
2948     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0) {
2949         h->ref_count[1]= h->ref_count[0]= 0;
2950         return -1;
2951     }
2952
2953     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2954         s->last_picture_ptr= &h->ref_list[0][0];
2955         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2956     }
2957     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2958         s->next_picture_ptr= &h->ref_list[1][0];
2959         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2960     }
2961
2962     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2963        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2964         pred_weight_table(h);
2965     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2966         implicit_weight_table(h, -1);
2967     }else {
2968         h->use_weight = 0;
2969         for (i = 0; i < 2; i++) {
2970             h->luma_weight_flag[i]   = 0;
2971             h->chroma_weight_flag[i] = 0;
2972         }
2973     }
2974
2975     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
2976        s->avctx->error_recognition >= FF_ER_EXPLODE)
2977         return AVERROR_INVALIDDATA;
2978
2979     if(FRAME_MBAFF){
2980         ff_h264_fill_mbaff_ref_list(h);
2981
2982         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2983             implicit_weight_table(h, 0);
2984             implicit_weight_table(h, 1);
2985         }
2986     }
2987
2988     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2989         ff_h264_direct_dist_scale_factor(h);
2990     ff_h264_direct_ref_list_init(h);
2991
2992     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2993         tmp = get_ue_golomb_31(&s->gb);
2994         if(tmp > 2){
2995             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2996             return -1;
2997         }
2998         h->cabac_init_idc= tmp;
2999     }
3000
3001     h->last_qscale_diff = 0;
3002     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3003     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
3004         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3005         return -1;
3006     }
3007     s->qscale= tmp;
3008     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3009     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3010     //FIXME qscale / qp ... stuff
3011     if(h->slice_type == AV_PICTURE_TYPE_SP){
3012         get_bits1(&s->gb); /* sp_for_switch_flag */
3013     }
3014     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
3015         get_se_golomb(&s->gb); /* slice_qs_delta */
3016     }
3017
3018     h->deblocking_filter = 1;
3019     h->slice_alpha_c0_offset = 52;
3020     h->slice_beta_offset = 52;
3021     if( h->pps.deblocking_filter_parameters_present ) {
3022         tmp= get_ue_golomb_31(&s->gb);
3023         if(tmp > 2){
3024             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
3025             return -1;
3026         }
3027         h->deblocking_filter= tmp;
3028         if(h->deblocking_filter < 2)
3029             h->deblocking_filter^= 1; // 1<->0
3030
3031         if( h->deblocking_filter ) {
3032             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
3033             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
3034             if(   h->slice_alpha_c0_offset > 104U
3035                || h->slice_beta_offset     > 104U){
3036                 av_log(s->avctx, AV_LOG_ERROR, "deblocking filter parameters %d %d out of range\n", h->slice_alpha_c0_offset, h->slice_beta_offset);
3037                 return -1;
3038             }
3039         }
3040     }
3041
3042     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
3043        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
3044        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
3045        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
3046         h->deblocking_filter= 0;
3047
3048     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
3049         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
3050             /* Cheat slightly for speed:
3051                Do not bother to deblock across slices. */
3052             h->deblocking_filter = 2;
3053         } else {
3054             h0->max_contexts = 1;
3055             if(!h0->single_decode_warning) {
3056                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3057                 h0->single_decode_warning = 1;
3058             }
3059             if (h != h0) {
3060                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
3061                 return 1;
3062             }
3063         }
3064     }
3065     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
3066                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
3067                  + 6 * (h->sps.bit_depth_luma - 8);
3068
3069 #if 0 //FMO
3070     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
3071         slice_group_change_cycle= get_bits(&s->gb, ?);
3072 #endif
3073
3074     h0->last_slice_type = slice_type;
3075     h->slice_num = ++h0->current_slice;
3076     if(h->slice_num >= MAX_SLICES){
3077         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
3078     }
3079
3080     for(j=0; j<2; j++){
3081         int id_list[16];
3082         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
3083         for(i=0; i<16; i++){
3084             id_list[i]= 60;
3085             if (h->ref_list[j][i].f.data[0]) {
3086                 int k;
3087                 uint8_t *base = h->ref_list[j][i].f.base[0];
3088                 for(k=0; k<h->short_ref_count; k++)
3089                     if (h->short_ref[k]->f.base[0] == base) {
3090                         id_list[i]= k;
3091                         break;
3092                     }
3093                 for(k=0; k<h->long_ref_count; k++)
3094                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3095                         id_list[i]= h->short_ref_count + k;
3096                         break;
3097                     }
3098             }
3099         }
3100
3101         ref2frm[0]=
3102         ref2frm[1]= -1;
3103         for(i=0; i<16; i++)
3104             ref2frm[i+2]= 4*id_list[i]
3105                           + (h->ref_list[j][i].f.reference & 3);
3106         ref2frm[18+0]=
3107         ref2frm[18+1]= -1;
3108         for(i=16; i<48; i++)
3109             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3110                           + (h->ref_list[j][i].f.reference & 3);
3111     }
3112
3113     //FIXME: fix draw_edges+PAFF+frame threads
3114     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3115     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3116
3117     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3118         av_log(h->s.avctx, AV_LOG_DEBUG, "slice:%d %s mb:%d %c%s%s pps:%u frame:%d poc:%d/%d ref:%d/%d qp:%d loop:%d:%d:%d weight:%d%s %s\n",
3119                h->slice_num,
3120                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3121                first_mb_in_slice,
3122                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3123                pps_id, h->frame_num,
3124                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3125                h->ref_count[0], h->ref_count[1],
3126                s->qscale,
3127                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3128                h->use_weight,
3129                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3130                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3131                );
3132     }
3133
3134     return 0;
3135 }
3136
3137 int ff_h264_get_slice_type(const H264Context *h)
3138 {
3139     switch (h->slice_type) {
3140     case AV_PICTURE_TYPE_P:  return 0;
3141     case AV_PICTURE_TYPE_B:  return 1;
3142     case AV_PICTURE_TYPE_I:  return 2;
3143     case AV_PICTURE_TYPE_SP: return 3;
3144     case AV_PICTURE_TYPE_SI: return 4;
3145     default:         return -1;
3146     }
3147 }
3148
3149 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3150                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3151 {
3152     int b_stride = h->b_stride;
3153     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3154     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3155     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3156         if(USES_LIST(top_type, list)){
3157             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3158             const int b8_xy= 4*top_xy + 2;
3159             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3160             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3161             ref_cache[0 - 1*8]=
3162             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3163             ref_cache[2 - 1*8]=
3164             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3165         }else{
3166             AV_ZERO128(mv_dst - 1*8);
3167             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3168         }
3169
3170         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3171             if(USES_LIST(left_type[LTOP], list)){
3172                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3173                 const int b8_xy= 4*left_xy[LTOP] + 1;
3174                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3175                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3176                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3177                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3178                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3179                 ref_cache[-1 +  0]=
3180                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3181                 ref_cache[-1 + 16]=
3182                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3183             }else{
3184                 AV_ZERO32(mv_dst - 1 + 0);
3185                 AV_ZERO32(mv_dst - 1 + 8);
3186                 AV_ZERO32(mv_dst - 1 +16);
3187                 AV_ZERO32(mv_dst - 1 +24);
3188                 ref_cache[-1 +  0]=
3189                 ref_cache[-1 +  8]=
3190                 ref_cache[-1 + 16]=
3191                 ref_cache[-1 + 24]= LIST_NOT_USED;
3192             }
3193         }
3194     }
3195
3196     if(!USES_LIST(mb_type, list)){
3197         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3198         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3199         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3200         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3201         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3202         return;
3203     }
3204
3205     {
3206         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3207         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3208         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3209         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3210         AV_WN32A(&ref_cache[0*8], ref01);
3211         AV_WN32A(&ref_cache[1*8], ref01);
3212         AV_WN32A(&ref_cache[2*8], ref23);
3213         AV_WN32A(&ref_cache[3*8], ref23);
3214     }
3215
3216     {
3217         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3218         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3219         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3220         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3221         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3222     }
3223 }
3224
3225 /**
3226  *
3227  * @return non zero if the loop filter can be skiped
3228  */
3229 static int fill_filter_caches(H264Context *h, int mb_type){
3230     MpegEncContext * const s = &h->s;
3231     const int mb_xy= h->mb_xy;
3232     int top_xy, left_xy[LEFT_MBS];
3233     int top_type, left_type[LEFT_MBS];
3234     uint8_t *nnz;
3235     uint8_t *nnz_cache;
3236
3237     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3238
3239     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3240      * stuff, I can't imagine that these complex rules are worth it. */
3241
3242     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3243     if(FRAME_MBAFF){
3244         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3245         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3246         if(s->mb_y&1){
3247             if (left_mb_field_flag != curr_mb_field_flag) {
3248                 left_xy[LTOP] -= s->mb_stride;
3249             }
3250         }else{
3251             if(curr_mb_field_flag){
3252                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3253             }
3254             if (left_mb_field_flag != curr_mb_field_flag) {
3255                 left_xy[LBOT] += s->mb_stride;
3256             }
3257         }
3258     }
3259
3260     h->top_mb_xy = top_xy;
3261     h->left_mb_xy[LTOP] = left_xy[LTOP];
3262     h->left_mb_xy[LBOT] = left_xy[LBOT];
3263     {
3264         //for sufficiently low qp, filtering wouldn't do anything
3265         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3266         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3267         int qp = s->current_picture.f.qscale_table[mb_xy];
3268         if(qp <= qp_thresh
3269            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3270            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3271             if(!FRAME_MBAFF)
3272                 return 1;
3273             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3274                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3275                 return 1;
3276         }
3277     }
3278
3279     top_type        = s->current_picture.f.mb_type[top_xy];
3280     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3281     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3282     if(h->deblocking_filter == 2){
3283         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3284         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3285     }else{
3286         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3287         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3288     }
3289     h->top_type       = top_type;
3290     h->left_type[LTOP]= left_type[LTOP];
3291     h->left_type[LBOT]= left_type[LBOT];
3292
3293     if(IS_INTRA(mb_type))
3294         return 0;
3295
3296     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3297     if(h->list_count == 2)
3298         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3299
3300     nnz = h->non_zero_count[mb_xy];
3301     nnz_cache = h->non_zero_count_cache;
3302     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3303     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3304     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3305     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3306     h->cbp= h->cbp_table[mb_xy];
3307
3308     if(top_type){
3309         nnz = h->non_zero_count[top_xy];
3310         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3311     }
3312
3313     if(left_type[LTOP]){
3314         nnz = h->non_zero_count[left_xy[LTOP]];
3315         nnz_cache[3+8*1]= nnz[3+0*4];
3316         nnz_cache[3+8*2]= nnz[3+1*4];
3317         nnz_cache[3+8*3]= nnz[3+2*4];
3318         nnz_cache[3+8*4]= nnz[3+3*4];
3319     }
3320
3321     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3322     if(!CABAC && h->pps.transform_8x8_mode){
3323         if(IS_8x8DCT(top_type)){
3324             nnz_cache[4+8*0]=
3325             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3326             nnz_cache[6+8*0]=
3327             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3328         }
3329         if(IS_8x8DCT(left_type[LTOP])){
3330             nnz_cache[3+8*1]=
3331             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3332         }
3333         if(IS_8x8DCT(left_type[LBOT])){
3334             nnz_cache[3+8*3]=
3335             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3336         }
3337
3338         if(IS_8x8DCT(mb_type)){
3339             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3340             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3341
3342             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3343             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3344
3345             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3346             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3347
3348             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3349             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3350         }
3351     }
3352
3353     return 0;
3354 }
3355
3356 static void loop_filter(H264Context *h, int start_x, int end_x){
3357     MpegEncContext * const s = &h->s;
3358     uint8_t  *dest_y, *dest_cb, *dest_cr;
3359     int linesize, uvlinesize, mb_x, mb_y;
3360     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3361     const int old_slice_type= h->slice_type;
3362     const int pixel_shift = h->pixel_shift;
3363     const int block_h = 16 >> s->chroma_y_shift;
3364
3365     if(h->deblocking_filter) {
3366         for(mb_x= start_x; mb_x<end_x; mb_x++){
3367             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3368                 int mb_xy, mb_type;
3369                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3370                 h->slice_num= h->slice_table[mb_xy];
3371                 mb_type = s->current_picture.f.mb_type[mb_xy];
3372                 h->list_count= h->list_counts[mb_xy];
3373
3374                 if(FRAME_MBAFF)
3375                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3376
3377                 s->mb_x= mb_x;
3378                 s->mb_y= mb_y;
3379                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3380                 dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3381                 dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3382                     //FIXME simplify above
3383
3384                 if (MB_FIELD) {
3385                     linesize   = h->mb_linesize   = s->linesize * 2;
3386                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3387                     if(mb_y&1){ //FIXME move out of this function?
3388                         dest_y -= s->linesize*15;
3389                         dest_cb-= s->uvlinesize * (block_h - 1);
3390                         dest_cr-= s->uvlinesize * (block_h - 1);
3391                     }
3392                 } else {
3393                     linesize   = h->mb_linesize   = s->linesize;
3394                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3395                 }
3396                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
3397                 if(fill_filter_caches(h, mb_type))
3398                     continue;
3399                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3400                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3401
3402                 if (FRAME_MBAFF) {
3403                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3404                 } else {
3405                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3406                 }
3407             }
3408         }
3409     }
3410     h->slice_type= old_slice_type;
3411     s->mb_x= end_x;
3412     s->mb_y= end_mb_y - FRAME_MBAFF;
3413     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3414     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3415 }
3416
3417 static void predict_field_decoding_flag(H264Context *h){
3418     MpegEncContext * const s = &h->s;
3419     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3420     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3421                 ? s->current_picture.f.mb_type[mb_xy - 1]
3422                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3423                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3424                 : 0;
3425     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3426 }
3427
3428 /**
3429  * Draw edges and report progress for the last MB row.
3430  */
3431 static void decode_finish_row(H264Context *h){
3432     MpegEncContext * const s = &h->s;
3433     int top = 16*(s->mb_y >> FIELD_PICTURE);
3434     int height = 16 << FRAME_MBAFF;
3435     int deblock_border = (16 + 4) << FRAME_MBAFF;
3436     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3437
3438     if (h->deblocking_filter) {
3439         if((top + height) >= pic_height)
3440             height += deblock_border;
3441
3442         top -= deblock_border;
3443     }
3444
3445     if (top >= pic_height || (top + height) < h->emu_edge_height)
3446         return;
3447
3448     height = FFMIN(height, pic_height - top);
3449     if (top < h->emu_edge_height) {
3450         height = top+height;
3451         top = 0;
3452     }
3453
3454     ff_draw_horiz_band(s, top, height);
3455
3456     if (s->dropable) return;
3457
3458     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3459                              s->picture_structure==PICT_BOTTOM_FIELD);
3460 }
3461
3462 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3463     H264Context *h = *(void**)arg;
3464     MpegEncContext * const s = &h->s;
3465     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3466     int lf_x_start = s->mb_x;
3467
3468     s->mb_skip_run= -1;
3469
3470     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3471                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3472
3473     if( h->pps.cabac ) {
3474         /* realign */
3475         align_get_bits( &s->gb );
3476
3477         /* init cabac */
3478         ff_init_cabac_states( &h->cabac);
3479         ff_init_cabac_decoder( &h->cabac,
3480                                s->gb.buffer + get_bits_count(&s->gb)/8,
3481                                (get_bits_left(&s->gb) + 7)/8);
3482
3483         ff_h264_init_cabac_states(h);
3484
3485         for(;;){
3486 //START_TIMER
3487             int ret = ff_h264_decode_mb_cabac(h);
3488             int eos;
3489 //STOP_TIMER("decode_mb_cabac")
3490
3491             if(ret>=0) ff_h264_hl_decode_mb(h);
3492
3493             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3494                 s->mb_y++;
3495
3496                 ret = ff_h264_decode_mb_cabac(h);
3497
3498                 if(ret>=0) ff_h264_hl_decode_mb(h);
3499                 s->mb_y--;
3500             }
3501             eos = get_cabac_terminate( &h->cabac );
3502
3503             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3504                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3505                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3506                 return 0;
3507             }
3508             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3509                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d, bytestream (%td)\n", s->mb_x, s->mb_y, h->cabac.bytestream_end - h->cabac.bytestream);
3510                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3511                 return -1;
3512             }
3513
3514             if( ++s->mb_x >= s->mb_width ) {
3515                 loop_filter(h, lf_x_start, s->mb_x);
3516                 s->mb_x = lf_x_start = 0;
3517                 decode_finish_row(h);
3518                 ++s->mb_y;
3519                 if(FIELD_OR_MBAFF_PICTURE) {
3520                     ++s->mb_y;
3521                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3522                         predict_field_decoding_flag(h);
3523                 }
3524             }
3525
3526             if( eos || s->mb_y >= s->mb_height ) {
3527                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3528                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3529                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3530                 return 0;
3531             }
3532         }
3533
3534     } else {
3535         for(;;){
3536             int ret = ff_h264_decode_mb_cavlc(h);
3537
3538             if(ret>=0) ff_h264_hl_decode_mb(h);
3539
3540             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3541                 s->mb_y++;
3542                 ret = ff_h264_decode_mb_cavlc(h);
3543
3544                 if(ret>=0) ff_h264_hl_decode_mb(h);
3545                 s->mb_y--;
3546             }
3547
3548             if(ret<0){
3549                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3550                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3551                 return -1;
3552             }
3553
3554             if(++s->mb_x >= s->mb_width){
3555                 loop_filter(h, lf_x_start, s->mb_x);
3556                 s->mb_x = lf_x_start = 0;
3557                 decode_finish_row(h);
3558                 ++s->mb_y;
3559                 if(FIELD_OR_MBAFF_PICTURE) {
3560                     ++s->mb_y;
3561                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3562                         predict_field_decoding_flag(h);
3563                 }
3564                 if(s->mb_y >= s->mb_height){
3565                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3566
3567                     if(get_bits_count(&s->gb) == s->gb.size_in_bits ) {
3568                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3569
3570                         return 0;
3571                     }else{
3572                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3573
3574                         return -1;
3575                     }
3576                 }
3577             }
3578
3579             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
3580                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3581                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
3582                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3583                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3584
3585                     return 0;
3586                 }else{
3587                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3588
3589                     return -1;
3590                 }
3591             }
3592         }
3593     }
3594 }
3595
3596 /**
3597  * Call decode_slice() for each context.
3598  *
3599  * @param h h264 master context
3600  * @param context_count number of contexts to execute
3601  */
3602 static int execute_decode_slices(H264Context *h, int context_count){
3603     MpegEncContext * const s = &h->s;
3604     AVCodecContext * const avctx= s->avctx;
3605     H264Context *hx;
3606     int i;
3607
3608     if (s->avctx->hwaccel || s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3609         return 0;
3610     if(context_count == 1) {
3611         return decode_slice(avctx, &h);
3612     } else {
3613         for(i = 1; i < context_count; i++) {
3614             hx = h->thread_context[i];
3615             hx->s.error_recognition = avctx->error_recognition;
3616             hx->s.error_count = 0;
3617         }
3618
3619         avctx->execute(avctx, (void *)decode_slice,
3620                        h->thread_context, NULL, context_count, sizeof(void*));
3621
3622         /* pull back stuff from slices to master context */
3623         hx = h->thread_context[context_count - 1];
3624         s->mb_x = hx->s.mb_x;
3625         s->mb_y = hx->s.mb_y;
3626         s->dropable = hx->s.dropable;
3627         s->picture_structure = hx->s.picture_structure;
3628         for(i = 1; i < context_count; i++)
3629             h->s.error_count += h->thread_context[i]->s.error_count;
3630     }
3631
3632     return 0;
3633 }
3634
3635
3636 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
3637     MpegEncContext * const s = &h->s;
3638     AVCodecContext * const avctx= s->avctx;
3639     H264Context *hx; ///< thread context
3640     int buf_index;
3641     int context_count;
3642     int next_avc;
3643     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3644     int nals_needed=0; ///< number of NALs that need decoding before the next frame thread starts
3645     int nal_index;
3646
3647     h->max_contexts = (HAVE_THREADS && (s->avctx->active_thread_type&FF_THREAD_SLICE)) ? avctx->thread_count : 1;
3648     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
3649         h->current_slice = 0;
3650         if (!s->first_field)
3651             s->current_picture_ptr= NULL;
3652         ff_h264_reset_sei(h);
3653     }
3654
3655     for(;pass <= 1;pass++){
3656         buf_index = 0;
3657         context_count = 0;
3658         next_avc = h->is_avc ? 0 : buf_size;
3659         nal_index = 0;
3660     for(;;){
3661         int consumed;
3662         int dst_length;
3663         int bit_length;
3664         const uint8_t *ptr;
3665         int i, nalsize = 0;
3666         int err;
3667
3668         if(buf_index >= next_avc) {
3669             if(buf_index >= buf_size) break;
3670             nalsize = 0;
3671             for(i = 0; i < h->nal_length_size; i++)
3672                 nalsize = (nalsize << 8) | buf[buf_index++];
3673             if(nalsize <= 0 || nalsize > buf_size - buf_index){
3674                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
3675                 break;
3676             }
3677             next_avc= buf_index + nalsize;
3678         } else {
3679             // start code prefix search
3680             for(; buf_index + 3 < next_avc; buf_index++){
3681                 // This should always succeed in the first iteration.
3682                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
3683                     break;
3684             }
3685
3686             if(buf_index+3 >= buf_size) break;
3687
3688             buf_index+=3;
3689             if(buf_index >= next_avc) continue;
3690         }
3691
3692         hx = h->thread_context[context_count];
3693
3694         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
3695         if (ptr==NULL || dst_length < 0){
3696             return -1;
3697         }
3698         i= buf_index + consumed;
3699         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
3700            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
3701             s->workaround_bugs |= FF_BUG_TRUNCATED;
3702
3703         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
3704         while(ptr[dst_length - 1] == 0 && dst_length > 0)
3705             dst_length--;
3706         }
3707         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
3708
3709         if(s->avctx->debug&FF_DEBUG_STARTCODE){
3710             av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d at %d/%d length %d\n", hx->nal_unit_type, buf_index, buf_size, dst_length);
3711         }
3712
3713         if (h->is_avc && (nalsize != consumed) && nalsize){
3714             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
3715         }
3716
3717         buf_index += consumed;
3718         nal_index++;
3719
3720         if(pass == 0) {
3721             // packets can sometimes contain multiple PPS/SPS
3722             // e.g. two PAFF field pictures in one packet, or a demuxer which splits NALs strangely
3723             // if so, when frame threading we can't start the next thread until we've read all of them
3724             switch (hx->nal_unit_type) {
3725                 case NAL_SPS:
3726                 case NAL_PPS:
3727                     nals_needed = nal_index;
3728                     break;
3729                 case NAL_IDR_SLICE:
3730                 case NAL_SLICE:
3731                     init_get_bits(&hx->s.gb, ptr, bit_length);
3732                     if (!get_ue_golomb(&hx->s.gb))
3733                         nals_needed = nal_index;
3734             }
3735             continue;
3736         }
3737
3738         //FIXME do not discard SEI id
3739         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
3740             continue;
3741
3742       again:
3743         err = 0;
3744         switch(hx->nal_unit_type){
3745         case NAL_IDR_SLICE:
3746             if (h->nal_unit_type != NAL_IDR_SLICE) {
3747                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
3748                 return -1;
3749             }
3750             idr(h); //FIXME ensure we don't loose some frames if there is reordering
3751         case NAL_SLICE:
3752             init_get_bits(&hx->s.gb, ptr, bit_length);
3753             hx->intra_gb_ptr=
3754             hx->inter_gb_ptr= &hx->s.gb;
3755             hx->s.data_partitioning = 0;
3756
3757             if((err = decode_slice_header(hx, h)))
3758                break;
3759
3760             s->current_picture_ptr->f.key_frame |=
3761                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
3762                     (h->sei_recovery_frame_cnt >= 0);
3763
3764             if (h->current_slice == 1) {
3765                 if(!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3766                     decode_postinit(h, nal_index >= nals_needed);
3767                 }
3768
3769                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3770                     return -1;
3771                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3772                     ff_vdpau_h264_picture_start(s);
3773             }
3774
3775             if(hx->redundant_pic_count==0
3776                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3777                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3778                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3779                && avctx->skip_frame < AVDISCARD_ALL){
3780                 if(avctx->hwaccel) {
3781                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
3782                         return -1;
3783                 }else
3784                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
3785                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
3786                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
3787                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
3788                 }else
3789                     context_count++;
3790             }
3791             break;
3792         case NAL_DPA:
3793             init_get_bits(&hx->s.gb, ptr, bit_length);
3794             hx->intra_gb_ptr=
3795             hx->inter_gb_ptr= NULL;
3796
3797             if ((err = decode_slice_header(hx, h)) < 0)
3798                 break;
3799
3800             hx->s.data_partitioning = 1;
3801
3802             break;
3803         case NAL_DPB:
3804             init_get_bits(&hx->intra_gb, ptr, bit_length);
3805             hx->intra_gb_ptr= &hx->intra_gb;
3806             break;
3807         case NAL_DPC:
3808             init_get_bits(&hx->inter_gb, ptr, bit_length);
3809             hx->inter_gb_ptr= &hx->inter_gb;
3810
3811             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
3812                && s->context_initialized
3813                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3814                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3815                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3816                && avctx->skip_frame < AVDISCARD_ALL)
3817                 context_count++;
3818             break;
3819         case NAL_SEI:
3820             init_get_bits(&s->gb, ptr, bit_length);
3821             ff_h264_decode_sei(h);
3822             break;
3823         case NAL_SPS:
3824             init_get_bits(&s->gb, ptr, bit_length);
3825             ff_h264_decode_seq_parameter_set(h);
3826
3827             if (s->flags& CODEC_FLAG_LOW_DELAY ||
3828                 (h->sps.bitstream_restriction_flag && !h->sps.num_reorder_frames))
3829                 s->low_delay=1;
3830
3831             if(avctx->has_b_frames < 2)
3832                 avctx->has_b_frames= !s->low_delay;
3833
3834             if (avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
3835                 h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
3836                 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
3837                     avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3838                     h->cur_chroma_format_idc = h->sps.chroma_format_idc;
3839                     h->pixel_shift = h->sps.bit_depth_luma > 8;
3840
3841                     ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3842                     ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3843                     s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
3844                     dsputil_init(&s->dsp, s->avctx);
3845                 } else {
3846                     av_log(avctx, AV_LOG_ERROR, "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3847                     return -1;
3848                 }
3849             }
3850             break;
3851         case NAL_PPS:
3852             init_get_bits(&s->gb, ptr, bit_length);
3853
3854             ff_h264_decode_picture_parameter_set(h, bit_length);
3855
3856             break;
3857         case NAL_AUD:
3858         case NAL_END_SEQUENCE:
3859         case NAL_END_STREAM:
3860         case NAL_FILLER_DATA:
3861         case NAL_SPS_EXT:
3862         case NAL_AUXILIARY_SLICE:
3863             break;
3864         default:
3865             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
3866         }
3867
3868         if(context_count == h->max_contexts) {
3869             execute_decode_slices(h, context_count);
3870             context_count = 0;
3871         }
3872
3873         if (err < 0)
3874             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3875         else if(err == 1) {
3876             /* Slice could not be decoded in parallel mode, copy down
3877              * NAL unit stuff to context 0 and restart. Note that
3878              * rbsp_buffer is not transferred, but since we no longer
3879              * run in parallel mode this should not be an issue. */
3880             h->nal_unit_type = hx->nal_unit_type;
3881             h->nal_ref_idc   = hx->nal_ref_idc;
3882             hx = h;
3883             goto again;
3884         }
3885     }
3886     }
3887     if(context_count)
3888         execute_decode_slices(h, context_count);
3889     return buf_index;
3890 }
3891
3892 /**
3893  * returns the number of bytes consumed for building the current frame
3894  */
3895 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
3896         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
3897         if(pos+10>buf_size) pos=buf_size; // oops ;)
3898
3899         return pos;
3900 }
3901
3902 static int decode_frame(AVCodecContext *avctx,
3903                              void *data, int *data_size,
3904                              AVPacket *avpkt)
3905 {
3906     const uint8_t *buf = avpkt->data;
3907     int buf_size = avpkt->size;
3908     H264Context *h = avctx->priv_data;
3909     MpegEncContext *s = &h->s;
3910     AVFrame *pict = data;
3911     int buf_index;
3912
3913     s->flags= avctx->flags;
3914     s->flags2= avctx->flags2;
3915
3916    /* end of stream, output what is still in the buffers */
3917  out:
3918     if (buf_size == 0) {
3919         Picture *out;
3920         int i, out_idx;
3921
3922         s->current_picture_ptr = NULL;
3923
3924 //FIXME factorize this with the output code below
3925         out = h->delayed_pic[0];
3926         out_idx = 0;
3927         for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3928             if(h->delayed_pic[i]->poc < out->poc){
3929                 out = h->delayed_pic[i];
3930                 out_idx = i;
3931             }
3932
3933         for(i=out_idx; h->delayed_pic[i]; i++)
3934             h->delayed_pic[i] = h->delayed_pic[i+1];
3935
3936         if(out){
3937             *data_size = sizeof(AVFrame);
3938             *pict= *(AVFrame*)out;
3939         }
3940
3941         return 0;
3942     }
3943
3944     buf_index=decode_nal_units(h, buf, buf_size);
3945     if(buf_index < 0)
3946         return -1;
3947
3948     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
3949         buf_size = 0;
3950         goto out;
3951     }
3952
3953     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
3954         if (avctx->skip_frame >= AVDISCARD_NONREF)
3955             return 0;
3956         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
3957         return -1;
3958     }
3959
3960     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
3961
3962         if(s->flags2 & CODEC_FLAG2_CHUNKS) decode_postinit(h, 1);
3963
3964         field_end(h, 0);
3965
3966         if (!h->next_output_pic) {
3967             /* Wait for second field. */
3968             *data_size = 0;
3969
3970         } else {
3971             *data_size = sizeof(AVFrame);
3972             *pict = *(AVFrame*)h->next_output_pic;
3973         }
3974     }
3975
3976     assert(pict->data[0] || !*data_size);
3977     ff_print_debug_info(s, pict);
3978 //printf("out %d\n", (int)pict->data[0]);
3979
3980     return get_consumed_bytes(s, buf_index, buf_size);
3981 }
3982 #if 0
3983 static inline void fill_mb_avail(H264Context *h){
3984     MpegEncContext * const s = &h->s;
3985     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3986
3987     if(s->mb_y){
3988         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
3989         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
3990         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
3991     }else{
3992         h->mb_avail[0]=
3993         h->mb_avail[1]=
3994         h->mb_avail[2]= 0;
3995     }
3996     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
3997     h->mb_avail[4]= 1; //FIXME move out
3998     h->mb_avail[5]= 0; //FIXME move out
3999 }
4000 #endif
4001
4002 #ifdef TEST
4003 #undef printf
4004 #undef random
4005 #define COUNT 8000
4006 #define SIZE (COUNT*40)
4007 int main(void){
4008     int i;
4009     uint8_t temp[SIZE];
4010     PutBitContext pb;
4011     GetBitContext gb;
4012 //    int int_temp[10000];
4013     DSPContext dsp;
4014     AVCodecContext avctx;
4015
4016     dsputil_init(&dsp, &avctx);
4017
4018     init_put_bits(&pb, temp, SIZE);
4019     printf("testing unsigned exp golomb\n");
4020     for(i=0; i<COUNT; i++){
4021         START_TIMER
4022         set_ue_golomb(&pb, i);
4023         STOP_TIMER("set_ue_golomb");
4024     }
4025     flush_put_bits(&pb);