cabac: split cabac.h into declarations and function definitions
[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 "cabac.h"
31 #include "cabac_functions.h"
32 #include "dsputil.h"
33 #include "avcodec.h"
34 #include "mpegvideo.h"
35 #include "h264.h"
36 #include "h264data.h"
37 #include "h264_mvpred.h"
38 #include "golomb.h"
39 #include "mathops.h"
40 #include "rectangle.h"
41 #include "thread.h"
42 #include "vdpau_internal.h"
43 #include "libavutil/avassert.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_VDA_VLD,
60     PIX_FMT_YUVJ420P,
61     PIX_FMT_NONE
62 };
63
64 /**
65  * Check if the top & left blocks are available if needed and
66  * change the dc mode so it only uses the available blocks.
67  */
68 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
69     MpegEncContext * const s = &h->s;
70     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
71     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
72     int i;
73
74     if(!(h->top_samples_available&0x8000)){
75         for(i=0; i<4; i++){
76             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
77             if(status<0){
78                 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);
79                 return -1;
80             } else if(status){
81                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
82             }
83         }
84     }
85
86     if((h->left_samples_available&0x8888)!=0x8888){
87         static const int mask[4]={0x8000,0x2000,0x80,0x20};
88         for(i=0; i<4; i++){
89             if(!(h->left_samples_available&mask[i])){
90                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
91                 if(status<0){
92                     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);
93                     return -1;
94                 } else if(status){
95                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
96                 }
97             }
98         }
99     }
100
101     return 0;
102 } //FIXME cleanup like ff_h264_check_intra_pred_mode
103
104 /**
105  * Check if the top & left blocks are available if needed and
106  * change the dc mode so it only uses the available blocks.
107  */
108 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
109     MpegEncContext * const s = &h->s;
110     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
111     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
112
113     if(mode > 6U) {
114         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);
115         return -1;
116     }
117
118     if(!(h->top_samples_available&0x8000)){
119         mode= top[ mode ];
120         if(mode<0){
121             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);
122             return -1;
123         }
124     }
125
126     if((h->left_samples_available&0x8080) != 0x8080){
127         mode= left[ mode ];
128         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
129             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
130         }
131         if(mode<0){
132             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);
133             return -1;
134         }
135     }
136
137     return mode;
138 }
139
140 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
141     int i, si, di;
142     uint8_t *dst;
143     int bufidx;
144
145 //    src[0]&0x80;                //forbidden bit
146     h->nal_ref_idc= src[0]>>5;
147     h->nal_unit_type= src[0]&0x1F;
148
149     src++; length--;
150
151 #if HAVE_FAST_UNALIGNED
152 # if HAVE_FAST_64BIT
153 #   define RS 7
154     for(i=0; i+1<length; i+=9){
155         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
156 # else
157 #   define RS 3
158     for(i=0; i+1<length; i+=5){
159         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
160 # endif
161             continue;
162         if(i>0 && !src[i]) i--;
163         while(src[i]) i++;
164 #else
165 #   define RS 0
166     for(i=0; i+1<length; i+=2){
167         if(src[i]) continue;
168         if(i>0 && src[i-1]==0) i--;
169 #endif
170         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
171             if(src[i+2]!=3){
172                 /* startcode, so we must be past the end */
173                 length=i;
174             }
175             break;
176         }
177         i-= RS;
178     }
179
180     if(i>=length-1){ //no escaped 0
181         *dst_length= length;
182         *consumed= length+1; //+1 for the header
183         return src;
184     }
185
186     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
187     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
188     dst= h->rbsp_buffer[bufidx];
189
190     if (dst == NULL){
191         return NULL;
192     }
193
194 //printf("decoding esc\n");
195     memcpy(dst, src, i);
196     si=di=i;
197     while(si+2<length){
198         //remove escapes (very rare 1:2^22)
199         if(src[si+2]>3){
200             dst[di++]= src[si++];
201             dst[di++]= src[si++];
202         }else if(src[si]==0 && src[si+1]==0){
203             if(src[si+2]==3){ //escape
204                 dst[di++]= 0;
205                 dst[di++]= 0;
206                 si+=3;
207                 continue;
208             }else //next start code
209                 goto nsc;
210         }
211
212         dst[di++]= src[si++];
213     }
214     while(si<length)
215         dst[di++]= src[si++];
216 nsc:
217
218     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
219
220     *dst_length= di;
221     *consumed= si + 1;//+1 for the header
222 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
223     return dst;
224 }
225
226 /**
227  * Identify the exact end of the bitstream
228  * @return the length of the trailing, or 0 if damaged
229  */
230 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
231     int v= *src;
232     int r;
233
234     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
235
236     for(r=1; r<9; r++){
237         if(v&1) return r;
238         v>>=1;
239     }
240     return 0;
241 }
242
243 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
244                                  int y_offset, int list){
245     int raw_my= h->mv_cache[list][ scan8[n] ][1];
246     int filter_height= (raw_my&3) ? 2 : 0;
247     int full_my= (raw_my>>2) + y_offset;
248     int top = full_my - filter_height, bottom = full_my + height + filter_height;
249
250     return FFMAX(abs(top), bottom);
251 }
252
253 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
254                                int y_offset, int list0, int list1, int *nrefs){
255     MpegEncContext * const s = &h->s;
256     int my;
257
258     y_offset += 16*(s->mb_y >> MB_FIELD);
259
260     if(list0){
261         int ref_n = h->ref_cache[0][ scan8[n] ];
262         Picture *ref= &h->ref_list[0][ref_n];
263
264         // Error resilience puts the current picture in the ref list.
265         // Don't try to wait on these as it will cause a deadlock.
266         // Fields can wait on each other, though.
267         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
268            (ref->f.reference & 3) != s->picture_structure) {
269             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
270             if (refs[0][ref_n] < 0) nrefs[0] += 1;
271             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
272         }
273     }
274
275     if(list1){
276         int ref_n = h->ref_cache[1][ scan8[n] ];
277         Picture *ref= &h->ref_list[1][ref_n];
278
279         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
280            (ref->f.reference & 3) != s->picture_structure) {
281             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
282             if (refs[1][ref_n] < 0) nrefs[1] += 1;
283             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
284         }
285     }
286 }
287
288 /**
289  * Wait until all reference frames are available for MC operations.
290  *
291  * @param h the H264 context
292  */
293 static void await_references(H264Context *h){
294     MpegEncContext * const s = &h->s;
295     const int mb_xy= h->mb_xy;
296     const int mb_type = s->current_picture.f.mb_type[mb_xy];
297     int refs[2][48];
298     int nrefs[2] = {0};
299     int ref, list;
300
301     memset(refs, -1, sizeof(refs));
302
303     if(IS_16X16(mb_type)){
304         get_lowest_part_y(h, refs, 0, 16, 0,
305                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
306     }else if(IS_16X8(mb_type)){
307         get_lowest_part_y(h, refs, 0, 8, 0,
308                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
309         get_lowest_part_y(h, refs, 8, 8, 8,
310                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
311     }else if(IS_8X16(mb_type)){
312         get_lowest_part_y(h, refs, 0, 16, 0,
313                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
314         get_lowest_part_y(h, refs, 4, 16, 0,
315                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
316     }else{
317         int i;
318
319         assert(IS_8X8(mb_type));
320
321         for(i=0; i<4; i++){
322             const int sub_mb_type= h->sub_mb_type[i];
323             const int n= 4*i;
324             int y_offset= (i&2)<<2;
325
326             if(IS_SUB_8X8(sub_mb_type)){
327                 get_lowest_part_y(h, refs, n  , 8, y_offset,
328                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
329             }else if(IS_SUB_8X4(sub_mb_type)){
330                 get_lowest_part_y(h, refs, n  , 4, y_offset,
331                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
332                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
333                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
334             }else if(IS_SUB_4X8(sub_mb_type)){
335                 get_lowest_part_y(h, refs, n  , 8, y_offset,
336                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
337                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
338                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
339             }else{
340                 int j;
341                 assert(IS_SUB_4X4(sub_mb_type));
342                 for(j=0; j<4; j++){
343                     int sub_y_offset= y_offset + 2*(j&2);
344                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
345                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
346                 }
347             }
348         }
349     }
350
351     for(list=h->list_count-1; list>=0; list--){
352         for(ref=0; ref<48 && nrefs[list]; ref++){
353             int row = refs[list][ref];
354             if(row >= 0){
355                 Picture *ref_pic = &h->ref_list[list][ref];
356                 int ref_field = ref_pic->f.reference - 1;
357                 int ref_field_picture = ref_pic->field_picture;
358                 int pic_height = 16*s->mb_height >> ref_field_picture;
359
360                 row <<= MB_MBAFF;
361                 nrefs[list]--;
362
363                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
364                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
365                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
366                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
367                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
368                 }else if(FIELD_PICTURE){
369                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
370                 }else{
371                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
372                 }
373             }
374         }
375     }
376 }
377
378 #if 0
379 /**
380  * DCT transforms the 16 dc values.
381  * @param qp quantization parameter ??? FIXME
382  */
383 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
384 //    const int qmul= dequant_coeff[qp][0];
385     int i;
386     int temp[16]; //FIXME check if this is a good idea
387     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
388     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
389
390     for(i=0; i<4; i++){
391         const int offset= y_offset[i];
392         const int z0= block[offset+stride*0] + block[offset+stride*4];
393         const int z1= block[offset+stride*0] - block[offset+stride*4];
394         const int z2= block[offset+stride*1] - block[offset+stride*5];
395         const int z3= block[offset+stride*1] + block[offset+stride*5];
396
397         temp[4*i+0]= z0+z3;
398         temp[4*i+1]= z1+z2;
399         temp[4*i+2]= z1-z2;
400         temp[4*i+3]= z0-z3;
401     }
402
403     for(i=0; i<4; i++){
404         const int offset= x_offset[i];
405         const int z0= temp[4*0+i] + temp[4*2+i];
406         const int z1= temp[4*0+i] - temp[4*2+i];
407         const int z2= temp[4*1+i] - temp[4*3+i];
408         const int z3= temp[4*1+i] + temp[4*3+i];
409
410         block[stride*0 +offset]= (z0 + z3)>>1;
411         block[stride*2 +offset]= (z1 + z2)>>1;
412         block[stride*8 +offset]= (z1 - z2)>>1;
413         block[stride*10+offset]= (z0 - z3)>>1;
414     }
415 }
416 #endif
417
418 #undef xStride
419 #undef stride
420
421 #if 0
422 static void chroma_dc_dct_c(DCTELEM *block){
423     const int stride= 16*2;
424     const int xStride= 16;
425     int a,b,c,d,e;
426
427     a= block[stride*0 + xStride*0];
428     b= block[stride*0 + xStride*1];
429     c= block[stride*1 + xStride*0];
430     d= block[stride*1 + xStride*1];
431
432     e= a-b;
433     a= a+b;
434     b= c-d;
435     c= c+d;
436
437     block[stride*0 + xStride*0]= (a+c);
438     block[stride*0 + xStride*1]= (e+b);
439     block[stride*1 + xStride*0]= (a-c);
440     block[stride*1 + xStride*1]= (e-b);
441 }
442 #endif
443
444 static av_always_inline void
445 mc_dir_part(H264Context *h, Picture *pic, int n, int square,
446             int height, int delta, int list,
447             uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
448             int src_x_offset, int src_y_offset,
449             qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
450             int pixel_shift, int chroma_idc)
451 {
452     MpegEncContext * const s = &h->s;
453     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
454     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
455     const int luma_xy= (mx&3) + ((my&3)<<2);
456     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
457     uint8_t * src_y = pic->f.data[0] + offset;
458     uint8_t * src_cb, * src_cr;
459     int extra_width= h->emu_edge_width;
460     int extra_height= h->emu_edge_height;
461     int emu=0;
462     const int full_mx= mx>>2;
463     const int full_my= my>>2;
464     const int pic_width  = 16*s->mb_width;
465     const int pic_height = 16*s->mb_height >> MB_FIELD;
466     int ysh;
467
468     if(mx&7) extra_width -= 3;
469     if(my&7) extra_height -= 3;
470
471     if(   full_mx < 0-extra_width
472        || full_my < 0-extra_height
473        || full_mx + 16/*FIXME*/ > pic_width + extra_width
474        || full_my + 16/*FIXME*/ > pic_height + extra_height){
475         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_y - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
476                                 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
477             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
478         emu=1;
479     }
480
481     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
482     if(!square){
483         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
484     }
485
486     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
487
488     if(chroma_idc == 3 /* yuv444 */){
489         src_cb = pic->f.data[1] + offset;
490         if(emu){
491             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
492                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
493             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
494         }
495         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
496         if(!square){
497             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
498         }
499
500         src_cr = pic->f.data[2] + offset;
501         if(emu){
502             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
503                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
504             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
505         }
506         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
507         if(!square){
508             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
509         }
510         return;
511     }
512
513     ysh = 3 - (chroma_idc == 2 /* yuv422 */);
514     if(chroma_idc == 1 /* yuv420 */ && MB_FIELD){
515         // chroma offset when predicting from a field of opposite parity
516         my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
517         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
518     }
519
520     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) + (my >> ysh) * h->mb_uvlinesize;
521     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) + (my >> ysh) * h->mb_uvlinesize;
522
523     if(emu){
524         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
525                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
526                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
527             src_cb= s->edge_emu_buffer;
528     }
529     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
530               mx&7, (my << (chroma_idc == 2 /* yuv422 */)) &7);
531
532     if(emu){
533         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
534                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
535                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
536             src_cr= s->edge_emu_buffer;
537     }
538     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
539               mx&7, (my << (chroma_idc == 2 /* yuv422 */)) &7);
540 }
541
542 static av_always_inline void
543 mc_part_std(H264Context *h, int n, int square, int height, int delta,
544             uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
545             int x_offset, int y_offset,
546             qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
547             qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
548             int list0, int list1, int pixel_shift, int chroma_idc)
549 {
550     MpegEncContext * const s = &h->s;
551     qpel_mc_func *qpix_op=  qpix_put;
552     h264_chroma_mc_func chroma_op= chroma_put;
553
554     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
555     if (chroma_idc == 3 /* yuv444 */) {
556         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
557         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
558     } else if (chroma_idc == 2 /* yuv422 */) {
559         dest_cb += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
560         dest_cr += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
561     } else /* yuv420 */ {
562         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
563         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
564     }
565     x_offset += 8*s->mb_x;
566     y_offset += 8*(s->mb_y >> MB_FIELD);
567
568     if(list0){
569         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
570         mc_dir_part(h, ref, n, square, height, delta, 0,
571                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
572                            qpix_op, chroma_op, pixel_shift, chroma_idc);
573
574         qpix_op=  qpix_avg;
575         chroma_op= chroma_avg;
576     }
577
578     if(list1){
579         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
580         mc_dir_part(h, ref, n, square, height, delta, 1,
581                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
582                            qpix_op, chroma_op, pixel_shift, chroma_idc);
583     }
584 }
585
586 static av_always_inline void
587 mc_part_weighted(H264Context *h, int n, int square, int height, int delta,
588                  uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
589                  int x_offset, int y_offset,
590                  qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
591                  h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
592                  h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
593                  int list0, int list1, int pixel_shift, int chroma_idc){
594     MpegEncContext * const s = &h->s;
595     int chroma_height;
596
597     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
598     if (chroma_idc == 3 /* yuv444 */) {
599         chroma_height = height;
600         chroma_weight_avg = luma_weight_avg;
601         chroma_weight_op = luma_weight_op;
602         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
603         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
604     } else if (chroma_idc == 2 /* yuv422 */) {
605         chroma_height = height;
606         dest_cb += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
607         dest_cr += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
608     } else /* yuv420 */ {
609         chroma_height = height >> 1;
610         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
611         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
612     }
613     x_offset += 8*s->mb_x;
614     y_offset += 8*(s->mb_y >> MB_FIELD);
615
616     if(list0 && list1){
617         /* don't optimize for luma-only case, since B-frames usually
618          * use implicit weights => chroma too. */
619         uint8_t *tmp_cb = s->obmc_scratchpad;
620         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
621         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
622         int refn0 = h->ref_cache[0][ scan8[n] ];
623         int refn1 = h->ref_cache[1][ scan8[n] ];
624
625         mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
626                     dest_y, dest_cb, dest_cr,
627                     x_offset, y_offset, qpix_put, chroma_put,
628                     pixel_shift, chroma_idc);
629         mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
630                     tmp_y, tmp_cb, tmp_cr,
631                     x_offset, y_offset, qpix_put, chroma_put,
632                     pixel_shift, chroma_idc);
633
634         if(h->use_weight == 2){
635             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
636             int weight1 = 64 - weight0;
637             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize,
638                               height,        5, weight0, weight1, 0);
639             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
640                               chroma_height, 5, weight0, weight1, 0);
641             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
642                               chroma_height, 5, weight0, weight1, 0);
643         }else{
644             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height, h->luma_log2_weight_denom,
645                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
646                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
647             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
648                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
649                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
650             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
651                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
652                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
653         }
654     }else{
655         int list = list1 ? 1 : 0;
656         int refn = h->ref_cache[list][ scan8[n] ];
657         Picture *ref= &h->ref_list[list][refn];
658         mc_dir_part(h, ref, n, square, height, delta, list,
659                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
660                     qpix_put, chroma_put, pixel_shift, chroma_idc);
661
662         luma_weight_op(dest_y, h->mb_linesize, height, h->luma_log2_weight_denom,
663                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
664         if(h->use_weight_chroma){
665             chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
666                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
667             chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
668                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
669         }
670     }
671 }
672
673 static av_always_inline void
674 mc_part(H264Context *h, int n, int square, int height, int delta,
675         uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
676         int x_offset, int y_offset,
677         qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
678         qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
679         h264_weight_func *weight_op, h264_biweight_func *weight_avg,
680         int list0, int list1, int pixel_shift, int chroma_idc)
681 {
682     if((h->use_weight==2 && list0 && list1
683         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
684        || h->use_weight==1)
685         mc_part_weighted(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
686                          x_offset, y_offset, qpix_put, chroma_put,
687                          weight_op[0], weight_op[1], weight_avg[0],
688                          weight_avg[1], list0, list1, pixel_shift, chroma_idc);
689     else
690         mc_part_std(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
691                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
692                     chroma_avg, list0, list1, pixel_shift, chroma_idc);
693 }
694
695 static av_always_inline void
696 prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma_idc)
697 {
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 (chroma_idc == 3 /* yuv444 */) {
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 chroma_idc)
723 {
724     MpegEncContext * const s = &h->s;
725     const int mb_xy= h->mb_xy;
726     const int mb_type = s->current_picture.f.mb_type[mb_xy];
727
728     assert(IS_INTER(mb_type));
729
730     if(HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
731         await_references(h);
732     prefetch_motion(h, 0, pixel_shift, chroma_idc);
733
734     if(IS_16X16(mb_type)){
735         mc_part(h, 0, 1, 16, 0, dest_y, dest_cb, dest_cr, 0, 0,
736                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
737                 weight_op, weight_avg,
738                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
739                 pixel_shift, chroma_idc);
740     }else if(IS_16X8(mb_type)){
741         mc_part(h, 0, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
742                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
743                 weight_op, weight_avg,
744                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
745                 pixel_shift, chroma_idc);
746         mc_part(h, 8, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
747                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
748                 weight_op, weight_avg,
749                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
750                 pixel_shift, chroma_idc);
751     }else if(IS_8X16(mb_type)){
752         mc_part(h, 0, 0, 16, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
753                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
754                 &weight_op[1], &weight_avg[1],
755                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
756                 pixel_shift, chroma_idc);
757         mc_part(h, 4, 0, 16, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
758                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
759                 &weight_op[1], &weight_avg[1],
760                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
761                 pixel_shift, chroma_idc);
762     }else{
763         int i;
764
765         assert(IS_8X8(mb_type));
766
767         for(i=0; i<4; i++){
768             const int sub_mb_type= h->sub_mb_type[i];
769             const int n= 4*i;
770             int x_offset= (i&1)<<2;
771             int y_offset= (i&2)<<1;
772
773             if(IS_SUB_8X8(sub_mb_type)){
774                 mc_part(h, n, 1, 8, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
775                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
776                     &weight_op[1], &weight_avg[1],
777                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
778                     pixel_shift, chroma_idc);
779             }else if(IS_SUB_8X4(sub_mb_type)){
780                 mc_part(h, n  , 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
781                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
782                     &weight_op[1], &weight_avg[1],
783                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
784                     pixel_shift, chroma_idc);
785                 mc_part(h, n+2, 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
786                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
787                     &weight_op[1], &weight_avg[1],
788                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
789                     pixel_shift, chroma_idc);
790             }else if(IS_SUB_4X8(sub_mb_type)){
791                 mc_part(h, n  , 0, 8, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
792                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
793                     &weight_op[2], &weight_avg[2],
794                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
795                     pixel_shift, chroma_idc);
796                 mc_part(h, n+1, 0, 8, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
797                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
798                     &weight_op[2], &weight_avg[2],
799                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
800                     pixel_shift, chroma_idc);
801             }else{
802                 int j;
803                 assert(IS_SUB_4X4(sub_mb_type));
804                 for(j=0; j<4; j++){
805                     int sub_x_offset= x_offset + 2*(j&1);
806                     int sub_y_offset= y_offset +   (j&2);
807                     mc_part(h, n+j, 1, 4, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
808                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
809                         &weight_op[2], &weight_avg[2],
810                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
811                         pixel_shift, chroma_idc);
812                 }
813             }
814         }
815     }
816
817     prefetch_motion(h, 1, pixel_shift, chroma_idc);
818 }
819
820 static av_always_inline void
821 hl_motion_420(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
822               qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
823               qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
824               h264_weight_func *weight_op, h264_biweight_func *weight_avg,
825               int pixel_shift)
826 {
827     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
828               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 1);
829 }
830
831 static av_always_inline void
832 hl_motion_422(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
833               qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
834               qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
835               h264_weight_func *weight_op, h264_biweight_func *weight_avg,
836               int pixel_shift)
837 {
838     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
839               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 2);
840 }
841
842 static void free_tables(H264Context *h, int free_rbsp){
843     int i;
844     H264Context *hx;
845
846     av_freep(&h->intra4x4_pred_mode);
847     av_freep(&h->chroma_pred_mode_table);
848     av_freep(&h->cbp_table);
849     av_freep(&h->mvd_table[0]);
850     av_freep(&h->mvd_table[1]);
851     av_freep(&h->direct_table);
852     av_freep(&h->non_zero_count);
853     av_freep(&h->slice_table_base);
854     h->slice_table= NULL;
855     av_freep(&h->list_counts);
856
857     av_freep(&h->mb2b_xy);
858     av_freep(&h->mb2br_xy);
859
860     for(i = 0; i < MAX_THREADS; i++) {
861         hx = h->thread_context[i];
862         if(!hx) continue;
863         av_freep(&hx->top_borders[1]);
864         av_freep(&hx->top_borders[0]);
865         av_freep(&hx->s.obmc_scratchpad);
866         if (free_rbsp){
867             av_freep(&hx->rbsp_buffer[1]);
868             av_freep(&hx->rbsp_buffer[0]);
869             hx->rbsp_buffer_size[0] = 0;
870             hx->rbsp_buffer_size[1] = 0;
871         }
872         if (i) av_freep(&h->thread_context[i]);
873     }
874 }
875
876 static void init_dequant8_coeff_table(H264Context *h){
877     int i,j,q,x;
878     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
879
880     for(i=0; i<6; i++ ){
881         h->dequant8_coeff[i] = h->dequant8_buffer[i];
882         for(j=0; j<i; j++){
883             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
884                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
885                 break;
886             }
887         }
888         if(j<i)
889             continue;
890
891         for(q=0; q<max_qp+1; q++){
892             int shift = div6[q];
893             int idx = rem6[q];
894             for(x=0; x<64; x++)
895                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
896                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
897                     h->pps.scaling_matrix8[i][x]) << shift;
898         }
899     }
900 }
901
902 static void init_dequant4_coeff_table(H264Context *h){
903     int i,j,q,x;
904     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
905     for(i=0; i<6; i++ ){
906         h->dequant4_coeff[i] = h->dequant4_buffer[i];
907         for(j=0; j<i; j++){
908             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
909                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
910                 break;
911             }
912         }
913         if(j<i)
914             continue;
915
916         for(q=0; q<max_qp+1; q++){
917             int shift = div6[q] + 2;
918             int idx = rem6[q];
919             for(x=0; x<16; x++)
920                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
921                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
922                     h->pps.scaling_matrix4[i][x]) << shift;
923         }
924     }
925 }
926
927 static void init_dequant_tables(H264Context *h){
928     int i,x;
929     init_dequant4_coeff_table(h);
930     if(h->pps.transform_8x8_mode)
931         init_dequant8_coeff_table(h);
932     if(h->sps.transform_bypass){
933         for(i=0; i<6; i++)
934             for(x=0; x<16; x++)
935                 h->dequant4_coeff[i][0][x] = 1<<6;
936         if(h->pps.transform_8x8_mode)
937             for(i=0; i<6; i++)
938                 for(x=0; x<64; x++)
939                     h->dequant8_coeff[i][0][x] = 1<<6;
940     }
941 }
942
943
944 int ff_h264_alloc_tables(H264Context *h){
945     MpegEncContext * const s = &h->s;
946     const int big_mb_num= s->mb_stride * (s->mb_height+1);
947     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
948     int x,y;
949
950     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
951
952     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
953     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
954     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
955
956     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
957     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
958     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
959     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
960     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
961
962     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
963     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
964
965     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
966     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
967     for(y=0; y<s->mb_height; y++){
968         for(x=0; x<s->mb_width; x++){
969             const int mb_xy= x + y*s->mb_stride;
970             const int b_xy = 4*x + 4*y*h->b_stride;
971
972             h->mb2b_xy [mb_xy]= b_xy;
973             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
974         }
975     }
976
977     s->obmc_scratchpad = NULL;
978
979     if(!h->dequant4_coeff[0])
980         init_dequant_tables(h);
981
982     return 0;
983 fail:
984     free_tables(h, 1);
985     return -1;
986 }
987
988 /**
989  * Mimic alloc_tables(), but for every context thread.
990  */
991 static void clone_tables(H264Context *dst, H264Context *src, int i){
992     MpegEncContext * const s = &src->s;
993     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
994     dst->non_zero_count           = src->non_zero_count;
995     dst->slice_table              = src->slice_table;
996     dst->cbp_table                = src->cbp_table;
997     dst->mb2b_xy                  = src->mb2b_xy;
998     dst->mb2br_xy                 = src->mb2br_xy;
999     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
1000     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
1001     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
1002     dst->direct_table             = src->direct_table;
1003     dst->list_counts              = src->list_counts;
1004
1005     dst->s.obmc_scratchpad = NULL;
1006     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma, src->sps.chroma_format_idc);
1007 }
1008
1009 /**
1010  * Init context
1011  * Allocate buffers which are not shared amongst multiple threads.
1012  */
1013 static int context_init(H264Context *h){
1014     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
1015     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
1016
1017     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
1018     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
1019
1020     return 0;
1021 fail:
1022     return -1; // free_tables will clean up for us
1023 }
1024
1025 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
1026
1027 static av_cold void common_init(H264Context *h){
1028     MpegEncContext * const s = &h->s;
1029
1030     s->width = s->avctx->width;
1031     s->height = s->avctx->height;
1032     s->codec_id= s->avctx->codec->id;
1033
1034     ff_h264dsp_init(&h->h264dsp, 8, 1);
1035     ff_h264_pred_init(&h->hpc, s->codec_id, 8, 1);
1036
1037     h->dequant_coeff_pps= -1;
1038     s->unrestricted_mv=1;
1039
1040     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
1041
1042     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
1043     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
1044 }
1045
1046 int ff_h264_decode_extradata(H264Context *h)
1047 {
1048     AVCodecContext *avctx = h->s.avctx;
1049
1050     if(avctx->extradata[0] == 1){
1051         int i, cnt, nalsize;
1052         unsigned char *p = avctx->extradata;
1053
1054         h->is_avc = 1;
1055
1056         if(avctx->extradata_size < 7) {
1057             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1058             return -1;
1059         }
1060         /* sps and pps in the avcC always have length coded with 2 bytes,
1061            so put a fake nal_length_size = 2 while parsing them */
1062         h->nal_length_size = 2;
1063         // Decode sps from avcC
1064         cnt = *(p+5) & 0x1f; // Number of sps
1065         p += 6;
1066         for (i = 0; i < cnt; i++) {
1067             nalsize = AV_RB16(p) + 2;
1068             if (p - avctx->extradata + nalsize > avctx->extradata_size)
1069                 return -1;
1070             if(decode_nal_units(h, p, nalsize) < 0) {
1071                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1072                 return -1;
1073             }
1074             p += nalsize;
1075         }
1076         // Decode pps from avcC
1077         cnt = *(p++); // Number of pps
1078         for (i = 0; i < cnt; i++) {
1079             nalsize = AV_RB16(p) + 2;
1080             if (p - avctx->extradata + nalsize > avctx->extradata_size)
1081                 return -1;
1082             if (decode_nal_units(h, p, nalsize) < 0) {
1083                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1084                 return -1;
1085             }
1086             p += nalsize;
1087         }
1088         // Now store right nal length size, that will be use to parse all other nals
1089         h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1090     } else {
1091         h->is_avc = 0;
1092         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1093             return -1;
1094     }
1095     return 0;
1096 }
1097
1098 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1099     H264Context *h= avctx->priv_data;
1100     MpegEncContext * const s = &h->s;
1101     int i;
1102
1103     MPV_decode_defaults(s);
1104
1105     s->avctx = avctx;
1106     common_init(h);
1107
1108     s->out_format = FMT_H264;
1109     s->workaround_bugs= avctx->workaround_bugs;
1110
1111     // set defaults
1112 //    s->decode_mb= ff_h263_decode_mb;
1113     s->quarter_sample = 1;
1114     if(!avctx->has_b_frames)
1115     s->low_delay= 1;
1116
1117     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1118
1119     ff_h264_decode_init_vlc();
1120
1121     h->pixel_shift = 0;
1122     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1123
1124     h->thread_context[0] = h;
1125     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1126     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1127         h->last_pocs[i] = INT_MIN;
1128     h->prev_poc_msb= 1<<16;
1129     h->x264_build = -1;
1130     ff_h264_reset_sei(h);
1131     if(avctx->codec_id == CODEC_ID_H264){
1132         if(avctx->ticks_per_frame == 1){
1133             s->avctx->time_base.den *=2;
1134         }
1135         avctx->ticks_per_frame = 2;
1136     }
1137
1138     if(avctx->extradata_size > 0 && avctx->extradata &&
1139         ff_h264_decode_extradata(h))
1140         return -1;
1141
1142     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1143         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1144         s->low_delay = 0;
1145     }
1146
1147     return 0;
1148 }
1149
1150 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1151 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1152 {
1153     int i;
1154
1155     for (i=0; i<count; i++){
1156         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1157                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1158                 !from[i]));
1159         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1160     }
1161 }
1162
1163 static void copy_parameter_set(void **to, void **from, int count, int size)
1164 {
1165     int i;
1166
1167     for (i=0; i<count; i++){
1168         if (to[i] && !from[i]) av_freep(&to[i]);
1169         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1170
1171         if (from[i]) memcpy(to[i], from[i], size);
1172     }
1173 }
1174
1175 static int decode_init_thread_copy(AVCodecContext *avctx){
1176     H264Context *h= avctx->priv_data;
1177
1178     if (!avctx->internal->is_copy)
1179         return 0;
1180     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1181     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1182
1183     return 0;
1184 }
1185
1186 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1187 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1188     H264Context *h= dst->priv_data, *h1= src->priv_data;
1189     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1190     int inited = s->context_initialized, err;
1191     int i;
1192
1193     if(dst == src || !s1->context_initialized) return 0;
1194
1195     err = ff_mpeg_update_thread_context(dst, src);
1196     if(err) return err;
1197
1198     //FIXME handle width/height changing
1199     if(!inited){
1200         for(i = 0; i < MAX_SPS_COUNT; i++)
1201             av_freep(h->sps_buffers + i);
1202
1203         for(i = 0; i < MAX_PPS_COUNT; i++)
1204             av_freep(h->pps_buffers + i);
1205
1206         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1207         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1208         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1209         if (ff_h264_alloc_tables(h) < 0) {
1210             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1211             return AVERROR(ENOMEM);
1212         }
1213         context_init(h);
1214
1215         for(i=0; i<2; i++){
1216             h->rbsp_buffer[i] = NULL;
1217             h->rbsp_buffer_size[i] = 0;
1218         }
1219
1220         h->thread_context[0] = h;
1221
1222         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1223         // so this has to be allocated here
1224         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1225
1226         s->dsp.clear_blocks(h->mb);
1227         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1228     }
1229
1230     //extradata/NAL handling
1231     h->is_avc          = h1->is_avc;
1232
1233     //SPS/PPS
1234     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1235     h->sps             = h1->sps;
1236     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1237     h->pps             = h1->pps;
1238
1239     //Dequantization matrices
1240     //FIXME these are big - can they be only copied when PPS changes?
1241     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1242
1243     for(i=0; i<6; i++)
1244         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1245
1246     for(i=0; i<6; i++)
1247         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1248
1249     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1250
1251     //POC timing
1252     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1253
1254     //reference lists
1255     copy_fields(h, h1, ref_count, list_count);
1256     copy_fields(h, h1, ref_list,  intra_gb);
1257     copy_fields(h, h1, short_ref, cabac_init_idc);
1258
1259     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1260     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1261     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1262
1263     h->last_slice_type = h1->last_slice_type;
1264
1265     if(!s->current_picture_ptr) return 0;
1266
1267     if(!s->dropable) {
1268         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1269         h->prev_poc_msb     = h->poc_msb;
1270         h->prev_poc_lsb     = h->poc_lsb;
1271     }
1272     h->prev_frame_num_offset= h->frame_num_offset;
1273     h->prev_frame_num       = h->frame_num;
1274     h->outputed_poc         = h->next_outputed_poc;
1275
1276     return err;
1277 }
1278
1279 int ff_h264_frame_start(H264Context *h){
1280     MpegEncContext * const s = &h->s;
1281     int i;
1282     const int pixel_shift = h->pixel_shift;
1283
1284     if(MPV_frame_start(s, s->avctx) < 0)
1285         return -1;
1286     ff_er_frame_start(s);
1287     /*
1288      * MPV_frame_start uses pict_type to derive key_frame.
1289      * This is incorrect for H.264; IDR markings must be used.
1290      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1291      * See decode_nal_units().
1292      */
1293     s->current_picture_ptr->f.key_frame = 0;
1294     s->current_picture_ptr->mmco_reset= 0;
1295
1296     assert(s->linesize && s->uvlinesize);
1297
1298     for(i=0; i<16; i++){
1299         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1300         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1301     }
1302     for(i=0; i<16; i++){
1303         h->block_offset[16+i]=
1304         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1305         h->block_offset[48+16+i]=
1306         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1307     }
1308
1309     /* can't be in alloc_tables because linesize isn't known there.
1310      * FIXME: redo bipred weight to not require extra buffer? */
1311     for(i = 0; i < s->slice_context_count; i++)
1312         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1313             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1314
1315     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1316     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1317
1318 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1319
1320     // We mark the current picture as non-reference after allocating it, so
1321     // that if we break out due to an error it can be released automatically
1322     // in the next MPV_frame_start().
1323     // SVQ3 as well as most other codecs have only last/next/current and thus
1324     // get released even with set reference, besides SVQ3 and others do not
1325     // mark frames as reference later "naturally".
1326     if(s->codec_id != CODEC_ID_SVQ3)
1327         s->current_picture_ptr->f.reference = 0;
1328
1329     s->current_picture_ptr->field_poc[0]=
1330     s->current_picture_ptr->field_poc[1]= INT_MAX;
1331
1332     h->next_output_pic = NULL;
1333
1334     assert(s->current_picture_ptr->long_ref==0);
1335
1336     return 0;
1337 }
1338
1339 /**
1340   * Run setup operations that must be run after slice header decoding.
1341   * This includes finding the next displayed frame.
1342   *
1343   * @param h h264 master context
1344   * @param setup_finished enough NALs have been read that we can call
1345   * ff_thread_finish_setup()
1346   */
1347 static void decode_postinit(H264Context *h, int setup_finished){
1348     MpegEncContext * const s = &h->s;
1349     Picture *out = s->current_picture_ptr;
1350     Picture *cur = s->current_picture_ptr;
1351     int i, pics, out_of_order, out_idx;
1352     int invalid = 0, cnt = 0;
1353
1354     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1355     s->current_picture_ptr->f.pict_type   = s->pict_type;
1356
1357     if (h->next_output_pic) return;
1358
1359     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1360         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1361         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1362         //to find this yet, so we assume the worst for now.
1363         //if (setup_finished)
1364         //    ff_thread_finish_setup(s->avctx);
1365         return;
1366     }
1367
1368     cur->f.interlaced_frame = 0;
1369     cur->f.repeat_pict      = 0;
1370
1371     /* Signal interlacing information externally. */
1372     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1373
1374     if(h->sps.pic_struct_present_flag){
1375         switch (h->sei_pic_struct)
1376         {
1377         case SEI_PIC_STRUCT_FRAME:
1378             break;
1379         case SEI_PIC_STRUCT_TOP_FIELD:
1380         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1381             cur->f.interlaced_frame = 1;
1382             break;
1383         case SEI_PIC_STRUCT_TOP_BOTTOM:
1384         case SEI_PIC_STRUCT_BOTTOM_TOP:
1385             if (FIELD_OR_MBAFF_PICTURE)
1386                 cur->f.interlaced_frame = 1;
1387             else
1388                 // try to flag soft telecine progressive
1389                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1390             break;
1391         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1392         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1393             // Signal the possibility of telecined film externally (pic_struct 5,6)
1394             // From these hints, let the applications decide if they apply deinterlacing.
1395             cur->f.repeat_pict = 1;
1396             break;
1397         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1398             // Force progressive here, as doubling interlaced frame is a bad idea.
1399             cur->f.repeat_pict = 2;
1400             break;
1401         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1402             cur->f.repeat_pict = 4;
1403             break;
1404         }
1405
1406         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1407             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1408     }else{
1409         /* Derive interlacing flag from used decoding process. */
1410         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1411     }
1412     h->prev_interlaced_frame = cur->f.interlaced_frame;
1413
1414     if (cur->field_poc[0] != cur->field_poc[1]){
1415         /* Derive top_field_first from field pocs. */
1416         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1417     }else{
1418         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1419             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1420             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1421               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1422                 cur->f.top_field_first = 1;
1423             else
1424                 cur->f.top_field_first = 0;
1425         }else{
1426             /* Most likely progressive */
1427             cur->f.top_field_first = 0;
1428         }
1429     }
1430
1431     //FIXME do something with unavailable reference frames
1432
1433     /* Sort B-frames into display order */
1434
1435     if(h->sps.bitstream_restriction_flag
1436        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1437         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1438         s->low_delay = 0;
1439     }
1440
1441     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1442        && !h->sps.bitstream_restriction_flag){
1443         s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1444         s->low_delay= 0;
1445     }
1446
1447     pics = 0;
1448     while(h->delayed_pic[pics]) pics++;
1449
1450     assert(pics <= MAX_DELAYED_PIC_COUNT);
1451
1452     h->delayed_pic[pics++] = cur;
1453     if (cur->f.reference == 0)
1454         cur->f.reference = DELAYED_PIC_REF;
1455
1456     /* Frame reordering. This code takes pictures from coding order and sorts
1457      * them by their incremental POC value into display order. It supports POC
1458      * gaps, MMCO reset codes and random resets.
1459      * A "display group" can start either with a IDR frame (f.key_frame = 1),
1460      * and/or can be closed down with a MMCO reset code. In sequences where
1461      * there is no delay, we can't detect that (since the frame was already
1462      * output to the user), so we also set h->mmco_reset to detect the MMCO
1463      * reset code.
1464      * FIXME: if we detect insufficient delays (as per s->avctx->has_b_frames),
1465      * we increase the delay between input and output. All frames affected by
1466      * the lag (e.g. those that should have been output before another frame
1467      * that we already returned to the user) will be dropped. This is a bug
1468      * that we will fix later. */
1469     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++) {
1470         cnt     += out->poc < h->last_pocs[i];
1471         invalid += out->poc == INT_MIN;
1472     }
1473     if (!h->mmco_reset && !cur->f.key_frame && cnt + invalid == MAX_DELAYED_PIC_COUNT && cnt > 0) {
1474         h->mmco_reset = 2;
1475         if (pics > 1)
1476             h->delayed_pic[pics - 2]->mmco_reset = 2;
1477     }
1478     if (h->mmco_reset || cur->f.key_frame) {
1479         for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1480             h->last_pocs[i] = INT_MIN;
1481         cnt     = 0;
1482         invalid = MAX_DELAYED_PIC_COUNT;
1483     }
1484     out = h->delayed_pic[0];
1485     out_idx = 0;
1486     for (i = 1; i < MAX_DELAYED_PIC_COUNT && h->delayed_pic[i] &&
1487          !h->delayed_pic[i-1]->mmco_reset && !h->delayed_pic[i]->f.key_frame; i++)
1488     {
1489         if(h->delayed_pic[i]->poc < out->poc){
1490             out = h->delayed_pic[i];
1491             out_idx = i;
1492         }
1493     }
1494     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->mmco_reset))
1495         h->next_outputed_poc = INT_MIN;
1496     out_of_order = !out->f.key_frame && !h->mmco_reset && (out->poc < h->next_outputed_poc);
1497
1498     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1499         { }
1500     else if (out_of_order && pics-1 == s->avctx->has_b_frames &&
1501              s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT) {
1502         if (invalid + cnt < MAX_DELAYED_PIC_COUNT) {
1503             s->avctx->has_b_frames = FFMAX(s->avctx->has_b_frames, cnt);
1504         }
1505         s->low_delay = 0;
1506     } else if (s->low_delay &&
1507                ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2) ||
1508                 cur->f.pict_type == AV_PICTURE_TYPE_B)) {
1509         s->low_delay = 0;
1510         s->avctx->has_b_frames++;
1511     }
1512
1513     if(pics > s->avctx->has_b_frames){
1514         out->f.reference &= ~DELAYED_PIC_REF;
1515         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1516                          // or else the first thread can release the picture and reuse it unsafely
1517         for(i=out_idx; h->delayed_pic[i]; i++)
1518             h->delayed_pic[i] = h->delayed_pic[i+1];
1519     }
1520     memmove(h->last_pocs, &h->last_pocs[1], sizeof(*h->last_pocs) * (MAX_DELAYED_PIC_COUNT - 1));
1521     h->last_pocs[MAX_DELAYED_PIC_COUNT - 1] = cur->poc;
1522     if(!out_of_order && pics > s->avctx->has_b_frames){
1523         h->next_output_pic = out;
1524         if (out->mmco_reset) {
1525             if (out_idx > 0) {
1526                 h->next_outputed_poc = out->poc;
1527                 h->delayed_pic[out_idx - 1]->mmco_reset = out->mmco_reset;
1528             } else {
1529                 h->next_outputed_poc = INT_MIN;
1530             }
1531         } else {
1532             if (out_idx == 0 && pics > 1 && h->delayed_pic[0]->f.key_frame) {
1533                 h->next_outputed_poc = INT_MIN;
1534             } else {
1535                 h->next_outputed_poc = out->poc;
1536             }
1537         }
1538         h->mmco_reset = 0;
1539     }else{
1540         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1541     }
1542
1543     if (setup_finished)
1544         ff_thread_finish_setup(s->avctx);
1545 }
1546
1547 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1548                                               uint8_t *src_cb, uint8_t *src_cr,
1549                                               int linesize, int uvlinesize, int simple)
1550 {
1551     MpegEncContext * const s = &h->s;
1552     uint8_t *top_border;
1553     int top_idx = 1;
1554     const int pixel_shift = h->pixel_shift;
1555     int chroma444 = CHROMA444;
1556     int chroma422 = CHROMA422;
1557
1558     src_y  -=   linesize;
1559     src_cb -= uvlinesize;
1560     src_cr -= uvlinesize;
1561
1562     if(!simple && FRAME_MBAFF){
1563         if(s->mb_y&1){
1564             if(!MB_MBAFF){
1565                 top_border = h->top_borders[0][s->mb_x];
1566                 AV_COPY128(top_border, src_y + 15*linesize);
1567                 if (pixel_shift)
1568                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1569                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1570                     if(chroma444){
1571                         if (pixel_shift){
1572                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1573                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1574                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1575                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1576                         } else {
1577                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1578                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1579                         }
1580                     } else if(chroma422) {
1581                         if (pixel_shift) {
1582                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1583                             AV_COPY128(top_border+48, src_cr + 15*uvlinesize);
1584                         } else {
1585                             AV_COPY64(top_border+16, src_cb +  15*uvlinesize);
1586                             AV_COPY64(top_border+24, src_cr +  15*uvlinesize);
1587                         }
1588                     } else {
1589                         if (pixel_shift) {
1590                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1591                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1592                         } else {
1593                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1594                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1595                         }
1596                     }
1597                 }
1598             }
1599         }else if(MB_MBAFF){
1600             top_idx = 0;
1601         }else
1602             return;
1603     }
1604
1605     top_border = h->top_borders[top_idx][s->mb_x];
1606     // There are two lines saved, the line above the the top macroblock of a pair,
1607     // and the line above the bottom macroblock
1608     AV_COPY128(top_border, src_y + 16*linesize);
1609     if (pixel_shift)
1610         AV_COPY128(top_border+16, src_y+16*linesize+16);
1611
1612     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1613         if(chroma444){
1614             if (pixel_shift){
1615                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1616                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1617                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1618                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1619             } else {
1620                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1621                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1622             }
1623         } else if(chroma422) {
1624             if (pixel_shift) {
1625                 AV_COPY128(top_border+32, src_cb+16*uvlinesize);
1626                 AV_COPY128(top_border+48, src_cr+16*uvlinesize);
1627             } else {
1628                 AV_COPY64(top_border+16, src_cb+16*uvlinesize);
1629                 AV_COPY64(top_border+24, src_cr+16*uvlinesize);
1630             }
1631         } else {
1632             if (pixel_shift) {
1633                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1634                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1635             } else {
1636                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1637                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1638             }
1639         }
1640     }
1641 }
1642
1643 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1644                                   uint8_t *src_cb, uint8_t *src_cr,
1645                                   int linesize, int uvlinesize,
1646                                   int xchg, int chroma444,
1647                                   int simple, int pixel_shift){
1648     MpegEncContext * const s = &h->s;
1649     int deblock_topleft;
1650     int deblock_top;
1651     int top_idx = 1;
1652     uint8_t *top_border_m1;
1653     uint8_t *top_border;
1654
1655     if(!simple && FRAME_MBAFF){
1656         if(s->mb_y&1){
1657             if(!MB_MBAFF)
1658                 return;
1659         }else{
1660             top_idx = MB_MBAFF ? 0 : 1;
1661         }
1662     }
1663
1664     if(h->deblocking_filter == 2) {
1665         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1666         deblock_top     = h->top_type;
1667     } else {
1668         deblock_topleft = (s->mb_x > 0);
1669         deblock_top     = (s->mb_y > !!MB_FIELD);
1670     }
1671
1672     src_y  -=   linesize + 1 + pixel_shift;
1673     src_cb -= uvlinesize + 1 + pixel_shift;
1674     src_cr -= uvlinesize + 1 + pixel_shift;
1675
1676     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1677     top_border    = h->top_borders[top_idx][s->mb_x];
1678
1679 #define XCHG(a,b,xchg)\
1680     if (pixel_shift) {\
1681         if (xchg) {\
1682             AV_SWAP64(b+0,a+0);\
1683             AV_SWAP64(b+8,a+8);\
1684         } else {\
1685             AV_COPY128(b,a); \
1686         }\
1687     } else \
1688 if (xchg) AV_SWAP64(b,a);\
1689 else      AV_COPY64(b,a);
1690
1691     if(deblock_top){
1692         if(deblock_topleft){
1693             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1694         }
1695         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1696         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1697         if(s->mb_x+1 < s->mb_width){
1698             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1699         }
1700     }
1701     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1702         if(chroma444){
1703             if(deblock_topleft){
1704                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1705                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1706             }
1707             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1708             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1709             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1710             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1711             if(s->mb_x+1 < s->mb_width){
1712                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1713                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1714             }
1715         } else {
1716             if(deblock_top){
1717                 if(deblock_topleft){
1718                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1719                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1720                 }
1721                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1722                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1723             }
1724         }
1725     }
1726 }
1727
1728 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1729     if (high_bit_depth) {
1730         return AV_RN32A(((int32_t*)mb) + index);
1731     } else
1732         return AV_RN16A(mb + index);
1733 }
1734
1735 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1736     if (high_bit_depth) {
1737         AV_WN32A(((int32_t*)mb) + index, value);
1738     } else
1739         AV_WN16A(mb + index, value);
1740 }
1741
1742 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1743                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1744 {
1745     MpegEncContext * const s = &h->s;
1746     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1747     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1748     int i;
1749     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1750     block_offset += 16*p;
1751     if(IS_INTRA4x4(mb_type)){
1752         if(simple || !s->encoding){
1753             if(IS_8x8DCT(mb_type)){
1754                 if(transform_bypass){
1755                     idct_dc_add =
1756                     idct_add    = s->dsp.add_pixels8;
1757                 }else{
1758                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1759                     idct_add    = h->h264dsp.h264_idct8_add;
1760                 }
1761                 for(i=0; i<16; i+=4){
1762                     uint8_t * const ptr= dest_y + block_offset[i];
1763                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1764                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1765                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1766                     }else{
1767                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1768                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1769                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1770                         if(nnz){
1771                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1772                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1773                             else
1774                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1775                         }
1776                     }
1777                 }
1778             }else{
1779                 if(transform_bypass){
1780                     idct_dc_add =
1781                     idct_add    = s->dsp.add_pixels4;
1782                 }else{
1783                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1784                     idct_add    = h->h264dsp.h264_idct_add;
1785                 }
1786                 for(i=0; i<16; i++){
1787                     uint8_t * const ptr= dest_y + block_offset[i];
1788                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1789
1790                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1791                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1792                     }else{
1793                         uint8_t *topright;
1794                         int nnz, tr;
1795                         uint64_t tr_high;
1796                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1797                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1798                             assert(s->mb_y || linesize <= block_offset[i]);
1799                             if(!topright_avail){
1800                                 if (pixel_shift) {
1801                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1802                                     topright= (uint8_t*) &tr_high;
1803                                 } else {
1804                                     tr= ptr[3 - linesize]*0x01010101u;
1805                                     topright= (uint8_t*) &tr;
1806                                 }
1807                             }else
1808                                 topright= ptr + (4 << pixel_shift) - linesize;
1809                         }else
1810                             topright= NULL;
1811
1812                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1813                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1814                         if(nnz){
1815                             if(is_h264){
1816                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1817                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1818                                 else
1819                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1820                             } else if (CONFIG_SVQ3_DECODER)
1821                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1822                         }
1823                     }
1824                 }
1825             }
1826         }
1827     }else{
1828         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1829         if(is_h264){
1830             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1831                 if(!transform_bypass)
1832                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1833                 else{
1834                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1835                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1836                     for(i = 0; i < 16; i++)
1837                         dctcoef_set(h->mb+(p*256 << pixel_shift), pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1838                 }
1839             }
1840         } else if (CONFIG_SVQ3_DECODER)
1841             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1842     }
1843 }
1844
1845 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1846                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1847 {
1848     MpegEncContext * const s = &h->s;
1849     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1850     int i;
1851     block_offset += 16*p;
1852     if(!IS_INTRA4x4(mb_type)){
1853         if(is_h264){
1854             if(IS_INTRA16x16(mb_type)){
1855                 if(transform_bypass){
1856                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1857                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1858                     }else{
1859                         for(i=0; i<16; i++){
1860                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1861                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1862                         }
1863                     }
1864                 }else{
1865                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1866                 }
1867             }else if(h->cbp&15){
1868                 if(transform_bypass){
1869                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1870                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1871                     for(i=0; i<16; i+=di){
1872                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1873                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1874                         }
1875                     }
1876                 }else{
1877                     if(IS_8x8DCT(mb_type)){
1878                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1879                     }else{
1880                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1881                     }
1882                 }
1883             }
1884         } else if (CONFIG_SVQ3_DECODER) {
1885             for(i=0; i<16; i++){
1886                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1887                     uint8_t * const ptr= dest_y + block_offset[i];
1888                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1889                 }
1890             }
1891         }
1892     }
1893 }
1894
1895 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift)
1896 {
1897     MpegEncContext * const s = &h->s;
1898     const int mb_x= s->mb_x;
1899     const int mb_y= s->mb_y;
1900     const int mb_xy= h->mb_xy;
1901     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1902     uint8_t  *dest_y, *dest_cb, *dest_cr;
1903     int linesize, uvlinesize /*dct_offset*/;
1904     int i, j;
1905     int *block_offset = &h->block_offset[0];
1906     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1907     /* is_h264 should always be true if SVQ3 is disabled. */
1908     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1909     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1910     const int block_h = 16 >> s->chroma_y_shift;
1911     const int chroma422 = CHROMA422;
1912
1913     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1914     dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1915     dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1916
1917     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1918     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1919
1920     h->list_counts[mb_xy]= h->list_count;
1921
1922     if (!simple && MB_FIELD) {
1923         linesize   = h->mb_linesize   = s->linesize * 2;
1924         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1925         block_offset = &h->block_offset[48];
1926         if(mb_y&1){ //FIXME move out of this function?
1927             dest_y -= s->linesize*15;
1928             dest_cb-= s->uvlinesize * (block_h - 1);
1929             dest_cr-= s->uvlinesize * (block_h - 1);
1930         }
1931         if(FRAME_MBAFF) {
1932             int list;
1933             for(list=0; list<h->list_count; list++){
1934                 if(!USES_LIST(mb_type, list))
1935                     continue;
1936                 if(IS_16X16(mb_type)){
1937                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1938                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1939                 }else{
1940                     for(i=0; i<16; i+=4){
1941                         int ref = h->ref_cache[list][scan8[i]];
1942                         if(ref >= 0)
1943                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1944                     }
1945                 }
1946             }
1947         }
1948     } else {
1949         linesize   = h->mb_linesize   = s->linesize;
1950         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1951 //        dct_offset = s->linesize * 16;
1952     }
1953
1954     if (!simple && IS_INTRA_PCM(mb_type)) {
1955         if (pixel_shift) {
1956             const int bit_depth = h->sps.bit_depth_luma;
1957             int j;
1958             GetBitContext gb;
1959             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1960
1961             for (i = 0; i < 16; i++) {
1962                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1963                 for (j = 0; j < 16; j++)
1964                     tmp_y[j] = get_bits(&gb, bit_depth);
1965             }
1966             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1967                 if (!h->sps.chroma_format_idc) {
1968                     for (i = 0; i < block_h; i++) {
1969                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1970                         for (j = 0; j < 8; j++) {
1971                             tmp_cb[j] = 1 << (bit_depth - 1);
1972                         }
1973                     }
1974                     for (i = 0; i < block_h; i++) {
1975                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1976                         for (j = 0; j < 8; j++) {
1977                             tmp_cr[j] = 1 << (bit_depth - 1);
1978                         }
1979                     }
1980                 } else {
1981                     for (i = 0; i < block_h; i++) {
1982                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1983                         for (j = 0; j < 8; j++)
1984                             tmp_cb[j] = get_bits(&gb, bit_depth);
1985                     }
1986                     for (i = 0; i < block_h; i++) {
1987                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1988                         for (j = 0; j < 8; j++)
1989                             tmp_cr[j] = get_bits(&gb, bit_depth);
1990                     }
1991                 }
1992             }
1993         } else {
1994             for (i=0; i<16; i++) {
1995                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1996             }
1997             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1998                 if (!h->sps.chroma_format_idc) {
1999                     for (i = 0; i < block_h; i++) {
2000                         memset(dest_cb + i*uvlinesize, 128, 8);
2001                         memset(dest_cr + i*uvlinesize, 128, 8);
2002                     }
2003                 } else {
2004                     for (i = 0; i < block_h; i++) {
2005                         memcpy(dest_cb + i*uvlinesize, h->mb + 128 + i*4,  8);
2006                         memcpy(dest_cr + i*uvlinesize, h->mb + 160 + i*4,  8);
2007                     }
2008                 }
2009             }
2010         }
2011     } else {
2012         if(IS_INTRA(mb_type)){
2013             if(h->deblocking_filter)
2014                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
2015
2016             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2017                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2018                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2019             }
2020
2021             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
2022
2023             if(h->deblocking_filter)
2024                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
2025         }else if(is_h264){
2026             if (chroma422) {
2027                 hl_motion_422(h, dest_y, dest_cb, dest_cr,
2028                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2029                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2030                               h->h264dsp.weight_h264_pixels_tab,
2031                               h->h264dsp.biweight_h264_pixels_tab,
2032                               pixel_shift);
2033             } else {
2034                 hl_motion_420(h, dest_y, dest_cb, dest_cr,
2035                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2036                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2037                               h->h264dsp.weight_h264_pixels_tab,
2038                               h->h264dsp.biweight_h264_pixels_tab,
2039                               pixel_shift);
2040             }
2041         }
2042
2043         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
2044
2045         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
2046             uint8_t *dest[2] = {dest_cb, dest_cr};
2047             if(transform_bypass){
2048                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
2049                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
2050                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
2051                 }else{
2052                     idct_add = s->dsp.add_pixels4;
2053                     for(j=1; j<3; j++){
2054                         for(i=j*16; i<j*16+4; i++){
2055                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
2056                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
2057                         }
2058                         if (chroma422) {
2059                             for(i=j*16+4; i<j*16+8; i++){
2060                                 if(h->non_zero_count_cache[ scan8[i+4] ] || dctcoef_get(h->mb, pixel_shift, i*16))
2061                                     idct_add   (dest[j-1] + block_offset[i+4], h->mb + (i*16 << pixel_shift), uvlinesize);
2062                             }
2063                         }
2064                     }
2065                 }
2066             }else{
2067                 if(is_h264){
2068                     int qp[2];
2069                     if (chroma422) {
2070                         qp[0] = h->chroma_qp[0] + 3;
2071                         qp[1] = h->chroma_qp[1] + 3;
2072                     } else {
2073                         qp[0] = h->chroma_qp[0];
2074                         qp[1] = h->chroma_qp[1];
2075                     }
2076                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
2077                         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]);
2078                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
2079                         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]);
2080                     h->h264dsp.h264_idct_add8(dest, block_offset,
2081                                               h->mb, uvlinesize,
2082                                               h->non_zero_count_cache);
2083                 } else if (CONFIG_SVQ3_DECODER) {
2084                     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]);
2085                     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]);
2086                     for(j=1; j<3; j++){
2087                         for(i=j*16; i<j*16+4; i++){
2088                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2089                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
2090                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
2091                             }
2092                         }
2093                     }
2094                 }
2095             }
2096         }
2097     }
2098     if(h->cbp || IS_INTRA(mb_type))
2099     {
2100         s->dsp.clear_blocks(h->mb);
2101         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2102     }
2103 }
2104
2105 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
2106     MpegEncContext * const s = &h->s;
2107     const int mb_x= s->mb_x;
2108     const int mb_y= s->mb_y;
2109     const int mb_xy= h->mb_xy;
2110     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2111     uint8_t  *dest[3];
2112     int linesize;
2113     int i, j, p;
2114     int *block_offset = &h->block_offset[0];
2115     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2116     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
2117
2118     for (p = 0; p < plane_count; p++)
2119     {
2120         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
2121         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
2122     }
2123
2124     h->list_counts[mb_xy]= h->list_count;
2125
2126     if (!simple && MB_FIELD) {
2127         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
2128         block_offset = &h->block_offset[48];
2129         if(mb_y&1) //FIXME move out of this function?
2130             for (p = 0; p < 3; p++)
2131                 dest[p] -= s->linesize*15;
2132         if(FRAME_MBAFF) {
2133             int list;
2134             for(list=0; list<h->list_count; list++){
2135                 if(!USES_LIST(mb_type, list))
2136                     continue;
2137                 if(IS_16X16(mb_type)){
2138                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2139                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2140                 }else{
2141                     for(i=0; i<16; i+=4){
2142                         int ref = h->ref_cache[list][scan8[i]];
2143                         if(ref >= 0)
2144                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2145                     }
2146                 }
2147             }
2148         }
2149     } else {
2150         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2151     }
2152
2153     if (!simple && IS_INTRA_PCM(mb_type)) {
2154         if (pixel_shift) {
2155             const int bit_depth = h->sps.bit_depth_luma;
2156             GetBitContext gb;
2157             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
2158
2159             for (p = 0; p < plane_count; p++) {
2160                 for (i = 0; i < 16; i++) {
2161                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2162                     for (j = 0; j < 16; j++)
2163                         tmp[j] = get_bits(&gb, bit_depth);
2164                 }
2165             }
2166         } else {
2167             for (p = 0; p < plane_count; p++) {
2168                 for (i = 0; i < 16; i++) {
2169                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2170                 }
2171             }
2172         }
2173     } else {
2174         if(IS_INTRA(mb_type)){
2175             if(h->deblocking_filter)
2176                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2177
2178             for (p = 0; p < plane_count; p++)
2179                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2180
2181             if(h->deblocking_filter)
2182                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2183         }else{
2184             hl_motion(h, dest[0], dest[1], dest[2],
2185                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2186                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2187                       h->h264dsp.weight_h264_pixels_tab,
2188                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 3);
2189         }
2190
2191         for (p = 0; p < plane_count; p++)
2192             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2193     }
2194     if(h->cbp || IS_INTRA(mb_type))
2195     {
2196         s->dsp.clear_blocks(h->mb);
2197         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2198     }
2199 }
2200
2201 /**
2202  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2203  */
2204 #define hl_decode_mb_simple(sh, bits) \
2205 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2206     hl_decode_mb_internal(h, 1, sh); \
2207 }
2208 hl_decode_mb_simple(0, 8)
2209 hl_decode_mb_simple(1, 16)
2210
2211 /**
2212  * Process a macroblock; this handles edge cases, such as interlacing.
2213  */
2214 static void av_noinline hl_decode_mb_complex(H264Context *h){
2215     hl_decode_mb_internal(h, 0, h->pixel_shift);
2216 }
2217
2218 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2219     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2220 }
2221
2222 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2223     hl_decode_mb_444_internal(h, 1, 0);
2224 }
2225
2226 void ff_h264_hl_decode_mb(H264Context *h){
2227     MpegEncContext * const s = &h->s;
2228     const int mb_xy= h->mb_xy;
2229     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2230     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2231
2232     if (CHROMA444) {
2233         if(is_complex || h->pixel_shift)
2234             hl_decode_mb_444_complex(h);
2235         else
2236             hl_decode_mb_444_simple(h);
2237     } else if (is_complex) {
2238         hl_decode_mb_complex(h);
2239     } else if (h->pixel_shift) {
2240         hl_decode_mb_simple_16(h);
2241     } else
2242         hl_decode_mb_simple_8(h);
2243 }
2244
2245 static int pred_weight_table(H264Context *h){
2246     MpegEncContext * const s = &h->s;
2247     int list, i;
2248     int luma_def, chroma_def;
2249
2250     h->use_weight= 0;
2251     h->use_weight_chroma= 0;
2252     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2253     if(h->sps.chroma_format_idc)
2254         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2255     luma_def = 1<<h->luma_log2_weight_denom;
2256     chroma_def = 1<<h->chroma_log2_weight_denom;
2257
2258     for(list=0; list<2; list++){
2259         h->luma_weight_flag[list]   = 0;
2260         h->chroma_weight_flag[list] = 0;
2261         for(i=0; i<h->ref_count[list]; i++){
2262             int luma_weight_flag, chroma_weight_flag;
2263
2264             luma_weight_flag= get_bits1(&s->gb);
2265             if(luma_weight_flag){
2266                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2267                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2268                 if(   h->luma_weight[i][list][0] != luma_def
2269                    || h->luma_weight[i][list][1] != 0) {
2270                     h->use_weight= 1;
2271                     h->luma_weight_flag[list]= 1;
2272                 }
2273             }else{
2274                 h->luma_weight[i][list][0]= luma_def;
2275                 h->luma_weight[i][list][1]= 0;
2276             }
2277
2278             if(h->sps.chroma_format_idc){
2279                 chroma_weight_flag= get_bits1(&s->gb);
2280                 if(chroma_weight_flag){
2281                     int j;
2282                     for(j=0; j<2; j++){
2283                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2284                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2285                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2286                            || h->chroma_weight[i][list][j][1] != 0) {
2287                             h->use_weight_chroma= 1;
2288                             h->chroma_weight_flag[list]= 1;
2289                         }
2290                     }
2291                 }else{
2292                     int j;
2293                     for(j=0; j<2; j++){
2294                         h->chroma_weight[i][list][j][0]= chroma_def;
2295                         h->chroma_weight[i][list][j][1]= 0;
2296                     }
2297                 }
2298             }
2299         }
2300         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2301     }
2302     h->use_weight= h->use_weight || h->use_weight_chroma;
2303     return 0;
2304 }
2305
2306 /**
2307  * Initialize implicit_weight table.
2308  * @param field  0/1 initialize the weight for interlaced MBAFF
2309  *                -1 initializes the rest
2310  */
2311 static void implicit_weight_table(H264Context *h, int field){
2312     MpegEncContext * const s = &h->s;
2313     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2314
2315     for (i = 0; i < 2; i++) {
2316         h->luma_weight_flag[i]   = 0;
2317         h->chroma_weight_flag[i] = 0;
2318     }
2319
2320     if(field < 0){
2321         if (s->picture_structure == PICT_FRAME) {
2322             cur_poc = s->current_picture_ptr->poc;
2323         } else {
2324             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2325         }
2326     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2327        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2328         h->use_weight= 0;
2329         h->use_weight_chroma= 0;
2330         return;
2331     }
2332         ref_start= 0;
2333         ref_count0= h->ref_count[0];
2334         ref_count1= h->ref_count[1];
2335     }else{
2336         cur_poc = s->current_picture_ptr->field_poc[field];
2337         ref_start= 16;
2338         ref_count0= 16+2*h->ref_count[0];
2339         ref_count1= 16+2*h->ref_count[1];
2340     }
2341
2342     h->use_weight= 2;
2343     h->use_weight_chroma= 2;
2344     h->luma_log2_weight_denom= 5;
2345     h->chroma_log2_weight_denom= 5;
2346
2347     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2348         int poc0 = h->ref_list[0][ref0].poc;
2349         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2350             int w = 32;
2351             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2352                 int poc1 = h->ref_list[1][ref1].poc;
2353                 int td = av_clip(poc1 - poc0, -128, 127);
2354                 if(td){
2355                     int tb = av_clip(cur_poc - poc0, -128, 127);
2356                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2357                     int dist_scale_factor = (tb*tx + 32) >> 8;
2358                     if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2359                         w = 64 - dist_scale_factor;
2360                 }
2361             }
2362             if(field<0){
2363                 h->implicit_weight[ref0][ref1][0]=
2364                 h->implicit_weight[ref0][ref1][1]= w;
2365             }else{
2366                 h->implicit_weight[ref0][ref1][field]=w;
2367             }
2368         }
2369     }
2370 }
2371
2372 /**
2373  * instantaneous decoder refresh.
2374  */
2375 static void idr(H264Context *h){
2376     ff_h264_remove_all_refs(h);
2377     h->prev_frame_num= 0;
2378     h->prev_frame_num_offset= 0;
2379     h->prev_poc_msb=
2380     h->prev_poc_lsb= 0;
2381 }
2382
2383 /* forget old pics after a seek */
2384 static void flush_dpb(AVCodecContext *avctx){
2385     H264Context *h= avctx->priv_data;
2386     int i;
2387     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
2388         if(h->delayed_pic[i])
2389             h->delayed_pic[i]->f.reference = 0;
2390         h->delayed_pic[i]= NULL;
2391     }
2392     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2393         h->last_pocs[i] = INT_MIN;
2394     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2395     h->prev_interlaced_frame = 1;
2396     idr(h);
2397     if(h->s.current_picture_ptr)
2398         h->s.current_picture_ptr->f.reference = 0;
2399     h->s.first_field= 0;
2400     ff_h264_reset_sei(h);
2401     ff_mpeg_flush(avctx);
2402 }
2403
2404 static int init_poc(H264Context *h){
2405     MpegEncContext * const s = &h->s;
2406     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2407     int field_poc[2];
2408     Picture *cur = s->current_picture_ptr;
2409
2410     h->frame_num_offset= h->prev_frame_num_offset;
2411     if(h->frame_num < h->prev_frame_num)
2412         h->frame_num_offset += max_frame_num;
2413
2414     if(h->sps.poc_type==0){
2415         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2416
2417         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2418             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2419         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2420             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2421         else
2422             h->poc_msb = h->prev_poc_msb;
2423 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2424         field_poc[0] =
2425         field_poc[1] = h->poc_msb + h->poc_lsb;
2426         if(s->picture_structure == PICT_FRAME)
2427             field_poc[1] += h->delta_poc_bottom;
2428     }else if(h->sps.poc_type==1){
2429         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2430         int i;
2431
2432         if(h->sps.poc_cycle_length != 0)
2433             abs_frame_num = h->frame_num_offset + h->frame_num;
2434         else
2435             abs_frame_num = 0;
2436
2437         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2438             abs_frame_num--;
2439
2440         expected_delta_per_poc_cycle = 0;
2441         for(i=0; i < h->sps.poc_cycle_length; i++)
2442             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2443
2444         if(abs_frame_num > 0){
2445             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2446             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2447
2448             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2449             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2450                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2451         } else
2452             expectedpoc = 0;
2453
2454         if(h->nal_ref_idc == 0)
2455             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2456
2457         field_poc[0] = expectedpoc + h->delta_poc[0];
2458         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2459
2460         if(s->picture_structure == PICT_FRAME)
2461             field_poc[1] += h->delta_poc[1];
2462     }else{
2463         int poc= 2*(h->frame_num_offset + h->frame_num);
2464
2465         if(!h->nal_ref_idc)
2466             poc--;
2467
2468         field_poc[0]= poc;
2469         field_poc[1]= poc;
2470     }
2471
2472     if(s->picture_structure != PICT_BOTTOM_FIELD)
2473         s->current_picture_ptr->field_poc[0]= field_poc[0];
2474     if(s->picture_structure != PICT_TOP_FIELD)
2475         s->current_picture_ptr->field_poc[1]= field_poc[1];
2476     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2477
2478     return 0;
2479 }
2480
2481
2482 /**
2483  * initialize scan tables
2484  */
2485 static void init_scan_tables(H264Context *h){
2486     int i;
2487     for(i=0; i<16; i++){
2488 #define T(x) (x>>2) | ((x<<2) & 0xF)
2489         h->zigzag_scan[i] = T(zigzag_scan[i]);
2490         h-> field_scan[i] = T( field_scan[i]);
2491 #undef T
2492     }
2493     for(i=0; i<64; i++){
2494 #define T(x) (x>>3) | ((x&7)<<3)
2495         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2496         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2497         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2498         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2499 #undef T
2500     }
2501     if(h->sps.transform_bypass){ //FIXME same ugly
2502         h->zigzag_scan_q0          = zigzag_scan;
2503         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2504         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2505         h->field_scan_q0           = field_scan;
2506         h->field_scan8x8_q0        = field_scan8x8;
2507         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2508     }else{
2509         h->zigzag_scan_q0          = h->zigzag_scan;
2510         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2511         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2512         h->field_scan_q0           = h->field_scan;
2513         h->field_scan8x8_q0        = h->field_scan8x8;
2514         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2515     }
2516 }
2517
2518 static int field_end(H264Context *h, int in_setup){
2519     MpegEncContext * const s = &h->s;
2520     AVCodecContext * const avctx= s->avctx;
2521     int err = 0;
2522     s->mb_y= 0;
2523
2524     if (!in_setup && !s->dropable)
2525         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2526                                  s->picture_structure==PICT_BOTTOM_FIELD);
2527
2528     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2529         ff_vdpau_h264_set_reference_frames(s);
2530
2531     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2532         if(!s->dropable) {
2533             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2534             h->prev_poc_msb= h->poc_msb;
2535             h->prev_poc_lsb= h->poc_lsb;
2536         }
2537         h->prev_frame_num_offset= h->frame_num_offset;
2538         h->prev_frame_num= h->frame_num;
2539         h->outputed_poc = h->next_outputed_poc;
2540     }
2541
2542     if (avctx->hwaccel) {
2543         if (avctx->hwaccel->end_frame(avctx) < 0)
2544             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2545     }
2546
2547     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2548         ff_vdpau_h264_picture_complete(s);
2549
2550     /*
2551      * FIXME: Error handling code does not seem to support interlaced
2552      * when slices span multiple rows
2553      * The ff_er_add_slice calls don't work right for bottom
2554      * fields; they cause massive erroneous error concealing
2555      * Error marking covers both fields (top and bottom).
2556      * This causes a mismatched s->error_count
2557      * and a bad error table. Further, the error count goes to
2558      * INT_MAX when called for bottom field, because mb_y is
2559      * past end by one (callers fault) and resync_mb_y != 0
2560      * causes problems for the first MB line, too.
2561      */
2562     if (!FIELD_PICTURE)
2563         ff_er_frame_end(s);
2564
2565     MPV_frame_end(s);
2566
2567     h->current_slice=0;
2568
2569     return err;
2570 }
2571
2572 /**
2573  * Replicate H264 "master" context to thread contexts.
2574  */
2575 static void clone_slice(H264Context *dst, H264Context *src)
2576 {
2577     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2578     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2579     dst->s.current_picture      = src->s.current_picture;
2580     dst->s.linesize             = src->s.linesize;
2581     dst->s.uvlinesize           = src->s.uvlinesize;
2582     dst->s.first_field          = src->s.first_field;
2583
2584     dst->prev_poc_msb           = src->prev_poc_msb;
2585     dst->prev_poc_lsb           = src->prev_poc_lsb;
2586     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2587     dst->prev_frame_num         = src->prev_frame_num;
2588     dst->short_ref_count        = src->short_ref_count;
2589
2590     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2591     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2592     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2593     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2594
2595     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2596     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2597 }
2598
2599 /**
2600  * Compute profile from profile_idc and constraint_set?_flags.
2601  *
2602  * @param sps SPS
2603  *
2604  * @return profile as defined by FF_PROFILE_H264_*
2605  */
2606 int ff_h264_get_profile(SPS *sps)
2607 {
2608     int profile = sps->profile_idc;
2609
2610     switch(sps->profile_idc) {
2611     case FF_PROFILE_H264_BASELINE:
2612         // constraint_set1_flag set to 1
2613         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2614         break;
2615     case FF_PROFILE_H264_HIGH_10:
2616     case FF_PROFILE_H264_HIGH_422:
2617     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2618         // constraint_set3_flag set to 1
2619         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2620         break;
2621     }
2622
2623     return profile;
2624 }
2625
2626 /**
2627  * Decode a slice header.
2628  * This will also call MPV_common_init() and frame_start() as needed.
2629  *
2630  * @param h h264context
2631  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2632  *
2633  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2634  */
2635 static int decode_slice_header(H264Context *h, H264Context *h0){
2636     MpegEncContext * const s = &h->s;
2637     MpegEncContext * const s0 = &h0->s;
2638     unsigned int first_mb_in_slice;
2639     unsigned int pps_id;
2640     int num_ref_idx_active_override_flag;
2641     unsigned int slice_type, tmp, i, j;
2642     int default_ref_list_done = 0;
2643     int last_pic_structure;
2644
2645     s->dropable= h->nal_ref_idc == 0;
2646
2647     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2648     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2649         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2650         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2651     }else{
2652         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2653         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2654     }
2655
2656     first_mb_in_slice= get_ue_golomb(&s->gb);
2657
2658     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2659         if(h0->current_slice && FIELD_PICTURE){
2660             field_end(h, 1);
2661         }
2662
2663         h0->current_slice = 0;
2664         if (!s0->first_field)
2665             s->current_picture_ptr= NULL;
2666     }
2667
2668     slice_type= get_ue_golomb_31(&s->gb);
2669     if(slice_type > 9){
2670         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);
2671         return -1;
2672     }
2673     if(slice_type > 4){
2674         slice_type -= 5;
2675         h->slice_type_fixed=1;
2676     }else
2677         h->slice_type_fixed=0;
2678
2679     slice_type= golomb_to_pict_type[ slice_type ];
2680     if (slice_type == AV_PICTURE_TYPE_I
2681         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2682         default_ref_list_done = 1;
2683     }
2684     h->slice_type= slice_type;
2685     h->slice_type_nos= slice_type & 3;
2686
2687     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2688
2689     pps_id= get_ue_golomb(&s->gb);
2690     if(pps_id>=MAX_PPS_COUNT){
2691         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2692         return -1;
2693     }
2694     if(!h0->pps_buffers[pps_id]) {
2695         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2696         return -1;
2697     }
2698     h->pps= *h0->pps_buffers[pps_id];
2699
2700     if(!h0->sps_buffers[h->pps.sps_id]) {
2701         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2702         return -1;
2703     }
2704     h->sps = *h0->sps_buffers[h->pps.sps_id];
2705
2706     s->avctx->profile = ff_h264_get_profile(&h->sps);
2707     s->avctx->level   = h->sps.level_idc;
2708     s->avctx->refs    = h->sps.ref_frame_count;
2709
2710     if(h == h0 && h->dequant_coeff_pps != pps_id){
2711         h->dequant_coeff_pps = pps_id;
2712         init_dequant_tables(h);
2713     }
2714
2715     s->mb_width= h->sps.mb_width;
2716     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2717
2718     h->b_stride=  s->mb_width*4;
2719
2720     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2721
2722     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2723     if(h->sps.frame_mbs_only_flag)
2724         s->height= 16*s->mb_height - (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2725     else
2726         s->height= 16*s->mb_height - (2<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2727
2728     if (s->context_initialized
2729         && (   s->width != s->avctx->width || s->height != s->avctx->height
2730             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2731         if(h != h0) {
2732             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2733             return -1;   // width / height changed during parallelized decoding
2734         }
2735         free_tables(h, 0);
2736         flush_dpb(s->avctx);
2737         MPV_common_end(s);
2738     }
2739     if (!s->context_initialized) {
2740         if (h != h0) {
2741             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2742             return -1;
2743         }
2744
2745         avcodec_set_dimensions(s->avctx, s->width, s->height);
2746         s->avctx->sample_aspect_ratio= h->sps.sar;
2747         av_assert0(s->avctx->sample_aspect_ratio.den);
2748
2749         if(h->sps.video_signal_type_present_flag){
2750             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2751             if(h->sps.colour_description_present_flag){
2752                 s->avctx->color_primaries = h->sps.color_primaries;
2753                 s->avctx->color_trc       = h->sps.color_trc;
2754                 s->avctx->colorspace      = h->sps.colorspace;
2755             }
2756         }
2757
2758         if(h->sps.timing_info_present_flag){
2759             int64_t den= h->sps.time_scale;
2760             if(h->x264_build < 44U)
2761                 den *= 2;
2762             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2763                       h->sps.num_units_in_tick, den, 1<<30);
2764         }
2765
2766         switch (h->sps.bit_depth_luma) {
2767             case 9 :
2768                 if (CHROMA444) {
2769                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2770                         s->avctx->pix_fmt = PIX_FMT_GBRP9;
2771                     } else
2772                         s->avctx->pix_fmt = PIX_FMT_YUV444P9;
2773                 } else if (CHROMA422)
2774                     s->avctx->pix_fmt = PIX_FMT_YUV422P9;
2775                 else
2776                     s->avctx->pix_fmt = PIX_FMT_YUV420P9;
2777                 break;
2778             case 10 :
2779                 if (CHROMA444) {
2780                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2781                         s->avctx->pix_fmt = PIX_FMT_GBRP10;
2782                     } else
2783                         s->avctx->pix_fmt = PIX_FMT_YUV444P10;
2784                 } else if (CHROMA422)
2785                     s->avctx->pix_fmt = PIX_FMT_YUV422P10;
2786                 else
2787                     s->avctx->pix_fmt = PIX_FMT_YUV420P10;
2788                 break;
2789             default:
2790                 if (CHROMA444){
2791                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2792                         s->avctx->pix_fmt = PIX_FMT_GBRP;
2793                     } else
2794                         s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2795                 } else if (CHROMA422) {
2796                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P : PIX_FMT_YUV422P;
2797                 }else{
2798                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2799                                                              s->avctx->codec->pix_fmts ?
2800                                                              s->avctx->codec->pix_fmts :
2801                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2802                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2803                                                              ff_hwaccel_pixfmt_list_420);
2804                 }
2805         }
2806
2807         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2808
2809         if (MPV_common_init(s) < 0) {
2810             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2811             return -1;
2812         }
2813         s->first_field = 0;
2814         h->prev_interlaced_frame = 1;
2815
2816         init_scan_tables(h);
2817         if (ff_h264_alloc_tables(h) < 0) {
2818             av_log(h->s.avctx, AV_LOG_ERROR, "Could not allocate memory for h264\n");
2819             return AVERROR(ENOMEM);
2820         }
2821
2822         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2823             if (context_init(h) < 0) {
2824                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2825                 return -1;
2826             }
2827         } else {
2828             for(i = 1; i < s->slice_context_count; i++) {
2829                 H264Context *c;
2830                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2831                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2832                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2833                 c->h264dsp = h->h264dsp;
2834                 c->sps = h->sps;
2835                 c->pps = h->pps;
2836                 c->pixel_shift = h->pixel_shift;
2837                 init_scan_tables(c);
2838                 clone_tables(c, h, i);
2839             }
2840
2841             for(i = 0; i < s->slice_context_count; i++)
2842                 if (context_init(h->thread_context[i]) < 0) {
2843                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2844                     return -1;
2845                 }
2846         }
2847     }
2848
2849     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2850
2851     h->mb_mbaff = 0;
2852     h->mb_aff_frame = 0;
2853     last_pic_structure = s0->picture_structure;
2854     if(h->sps.frame_mbs_only_flag){
2855         s->picture_structure= PICT_FRAME;
2856     }else{
2857         if(get_bits1(&s->gb)) { //field_pic_flag
2858             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2859         } else {
2860             s->picture_structure= PICT_FRAME;
2861             h->mb_aff_frame = h->sps.mb_aff;
2862         }
2863     }
2864     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2865
2866     if(h0->current_slice == 0){
2867         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2868         if(h->frame_num != h->prev_frame_num) {
2869             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2870
2871             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2872
2873             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2874                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2875                 if (unwrap_prev_frame_num < 0)
2876                     unwrap_prev_frame_num += max_frame_num;
2877
2878                 h->prev_frame_num = unwrap_prev_frame_num;
2879             }
2880         }
2881
2882         while(h->frame_num !=  h->prev_frame_num &&
2883               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2884             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2885             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2886             if (ff_h264_frame_start(h) < 0)
2887                 return -1;
2888             h->prev_frame_num++;
2889             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2890             s->current_picture_ptr->frame_num= h->prev_frame_num;
2891             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2892             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2893             ff_generate_sliding_window_mmcos(h);
2894             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2895                 (s->avctx->err_recognition & AV_EF_EXPLODE))
2896                 return AVERROR_INVALIDDATA;
2897             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2898              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2899              * about there being no actual duplicates.
2900              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2901              * concealing a lost frame, this probably isn't noticeable by comparison, but it should
2902              * be fixed. */
2903             if (h->short_ref_count) {
2904                 if (prev) {
2905                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2906                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2907                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2908                     h->short_ref[0]->poc = prev->poc+2;
2909                 }
2910                 h->short_ref[0]->frame_num = h->prev_frame_num;
2911             }
2912         }
2913
2914         /* See if we have a decoded first field looking for a pair... */
2915         if (s0->first_field) {
2916             assert(s0->current_picture_ptr);
2917             assert(s0->current_picture_ptr->f.data[0]);
2918             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2919
2920             /* figure out if we have a complementary field pair */
2921             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2922                 /*
2923                  * Previous field is unmatched. Don't display it, but let it
2924                  * remain for reference if marked as such.
2925                  */
2926                 s0->current_picture_ptr = NULL;
2927                 s0->first_field = FIELD_PICTURE;
2928
2929             } else {
2930                 if (h->nal_ref_idc &&
2931                         s0->current_picture_ptr->f.reference &&
2932                         s0->current_picture_ptr->frame_num != h->frame_num) {
2933                     /*
2934                      * This and previous field were reference, but had
2935                      * different frame_nums. Consider this field first in
2936                      * pair. Throw away previous field except for reference
2937                      * purposes.
2938                      */
2939                     s0->first_field = 1;
2940                     s0->current_picture_ptr = NULL;
2941
2942                 } else {
2943                     /* Second field in complementary pair */
2944                     s0->first_field = 0;
2945                 }
2946             }
2947
2948         } else {
2949             /* Frame or first field in a potentially complementary pair */
2950             assert(!s0->current_picture_ptr);
2951             s0->first_field = FIELD_PICTURE;
2952         }
2953
2954         if(!FIELD_PICTURE || s0->first_field) {
2955             if (ff_h264_frame_start(h) < 0) {
2956                 s0->first_field = 0;
2957                 return -1;
2958             }
2959         } else {
2960             ff_release_unused_pictures(s, 0);
2961         }
2962     }
2963     if(h != h0)
2964         clone_slice(h, h0);
2965
2966     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2967
2968     assert(s->mb_num == s->mb_width * s->mb_height);
2969     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2970        first_mb_in_slice                    >= s->mb_num){
2971         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2972         return -1;
2973     }
2974     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2975     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2976     if (s->picture_structure == PICT_BOTTOM_FIELD)
2977         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2978     assert(s->mb_y < s->mb_height);
2979
2980     if(s->picture_structure==PICT_FRAME){
2981         h->curr_pic_num=   h->frame_num;
2982         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2983     }else{
2984         h->curr_pic_num= 2*h->frame_num + 1;
2985         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2986     }
2987
2988     if(h->nal_unit_type == NAL_IDR_SLICE){
2989         get_ue_golomb(&s->gb); /* idr_pic_id */
2990     }
2991
2992     if(h->sps.poc_type==0){
2993         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2994
2995         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2996             h->delta_poc_bottom= get_se_golomb(&s->gb);
2997         }
2998     }
2999
3000     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
3001         h->delta_poc[0]= get_se_golomb(&s->gb);
3002
3003         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
3004             h->delta_poc[1]= get_se_golomb(&s->gb);
3005     }
3006
3007     init_poc(h);
3008
3009     if(h->pps.redundant_pic_cnt_present){
3010         h->redundant_pic_count= get_ue_golomb(&s->gb);
3011     }
3012
3013     //set defaults, might be overridden a few lines later
3014     h->ref_count[0]= h->pps.ref_count[0];
3015     h->ref_count[1]= h->pps.ref_count[1];
3016
3017     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
3018         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
3019             h->direct_spatial_mv_pred= get_bits1(&s->gb);
3020         }
3021         num_ref_idx_active_override_flag= get_bits1(&s->gb);
3022
3023         if(num_ref_idx_active_override_flag){
3024             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
3025             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
3026                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
3027
3028             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
3029                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3030                 h->ref_count[0]= h->ref_count[1]= 1;
3031                 return -1;
3032             }
3033         }
3034         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
3035             h->list_count= 2;
3036         else
3037             h->list_count= 1;
3038     }else
3039         h->list_count= 0;
3040
3041     if(!default_ref_list_done){
3042         ff_h264_fill_default_ref_list(h);
3043     }
3044
3045     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0) {
3046         h->ref_count[1]= h->ref_count[0]= 0;
3047         return -1;
3048     }
3049
3050     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
3051         s->last_picture_ptr= &h->ref_list[0][0];
3052         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
3053     }
3054     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
3055         s->next_picture_ptr= &h->ref_list[1][0];
3056         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
3057     }
3058
3059     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
3060        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
3061         pred_weight_table(h);
3062     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
3063         implicit_weight_table(h, -1);
3064     }else {
3065         h->use_weight = 0;
3066         for (i = 0; i < 2; i++) {
3067             h->luma_weight_flag[i]   = 0;
3068             h->chroma_weight_flag[i] = 0;
3069         }
3070     }
3071
3072     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
3073        (s->avctx->err_recognition & AV_EF_EXPLODE))
3074         return AVERROR_INVALIDDATA;
3075
3076     if(FRAME_MBAFF){
3077         ff_h264_fill_mbaff_ref_list(h);
3078
3079         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
3080             implicit_weight_table(h, 0);
3081             implicit_weight_table(h, 1);
3082         }
3083     }
3084
3085     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
3086         ff_h264_direct_dist_scale_factor(h);
3087     ff_h264_direct_ref_list_init(h);
3088
3089     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
3090         tmp = get_ue_golomb_31(&s->gb);
3091         if(tmp > 2){
3092             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3093             return -1;
3094         }
3095         h->cabac_init_idc= tmp;
3096     }
3097
3098     h->last_qscale_diff = 0;
3099     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3100     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
3101         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3102         return -1;
3103     }
3104     s->qscale= tmp;
3105     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3106     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3107     //FIXME qscale / qp ... stuff
3108     if(h->slice_type == AV_PICTURE_TYPE_SP){
3109         get_bits1(&s->gb); /* sp_for_switch_flag */
3110     }
3111     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
3112         get_se_golomb(&s->gb); /* slice_qs_delta */
3113     }
3114
3115     h->deblocking_filter = 1;
3116     h->slice_alpha_c0_offset = 52;
3117     h->slice_beta_offset = 52;
3118     if( h->pps.deblocking_filter_parameters_present ) {
3119         tmp= get_ue_golomb_31(&s->gb);
3120         if(tmp > 2){
3121             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
3122             return -1;
3123         }
3124         h->deblocking_filter= tmp;
3125         if(h->deblocking_filter < 2)
3126             h->deblocking_filter^= 1; // 1<->0
3127
3128         if( h->deblocking_filter ) {
3129             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
3130             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
3131             if(   h->slice_alpha_c0_offset > 104U
3132                || h->slice_beta_offset     > 104U){
3133                 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);
3134                 return -1;
3135             }
3136         }
3137     }
3138
3139     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
3140        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
3141        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
3142        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
3143         h->deblocking_filter= 0;
3144
3145     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
3146         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
3147             /* Cheat slightly for speed:
3148                Do not bother to deblock across slices. */
3149             h->deblocking_filter = 2;
3150         } else {
3151             h0->max_contexts = 1;
3152             if(!h0->single_decode_warning) {
3153                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3154                 h0->single_decode_warning = 1;
3155             }
3156             if (h != h0) {
3157                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
3158                 return 1;
3159             }
3160         }
3161     }
3162     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
3163                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
3164                  + 6 * (h->sps.bit_depth_luma - 8);
3165
3166 #if 0 //FMO
3167     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
3168         slice_group_change_cycle= get_bits(&s->gb, ?);
3169 #endif
3170
3171     h0->last_slice_type = slice_type;
3172     h->slice_num = ++h0->current_slice;
3173     if(h->slice_num >= MAX_SLICES){
3174         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
3175     }
3176
3177     for(j=0; j<2; j++){
3178         int id_list[16];
3179         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
3180         for(i=0; i<16; i++){
3181             id_list[i]= 60;
3182             if (h->ref_list[j][i].f.data[0]) {
3183                 int k;
3184                 uint8_t *base = h->ref_list[j][i].f.base[0];
3185                 for(k=0; k<h->short_ref_count; k++)
3186                     if (h->short_ref[k]->f.base[0] == base) {
3187                         id_list[i]= k;
3188                         break;
3189                     }
3190                 for(k=0; k<h->long_ref_count; k++)
3191                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3192                         id_list[i]= h->short_ref_count + k;
3193                         break;
3194                     }
3195             }
3196         }
3197
3198         ref2frm[0]=
3199         ref2frm[1]= -1;
3200         for(i=0; i<16; i++)
3201             ref2frm[i+2]= 4*id_list[i]
3202                           + (h->ref_list[j][i].f.reference & 3);
3203         ref2frm[18+0]=
3204         ref2frm[18+1]= -1;
3205         for(i=16; i<48; i++)
3206             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3207                           + (h->ref_list[j][i].f.reference & 3);
3208     }
3209
3210     //FIXME: fix draw_edges+PAFF+frame threads
3211     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3212     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3213
3214     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3215         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",
3216                h->slice_num,
3217                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3218                first_mb_in_slice,
3219                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3220                pps_id, h->frame_num,
3221                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3222                h->ref_count[0], h->ref_count[1],
3223                s->qscale,
3224                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3225                h->use_weight,
3226                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3227                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3228                );
3229     }
3230
3231     return 0;
3232 }
3233
3234 int ff_h264_get_slice_type(const H264Context *h)
3235 {
3236     switch (h->slice_type) {
3237     case AV_PICTURE_TYPE_P:  return 0;
3238     case AV_PICTURE_TYPE_B:  return 1;
3239     case AV_PICTURE_TYPE_I:  return 2;
3240     case AV_PICTURE_TYPE_SP: return 3;
3241     case AV_PICTURE_TYPE_SI: return 4;
3242     default:         return -1;
3243     }
3244 }
3245
3246 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3247                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3248 {
3249     int b_stride = h->b_stride;
3250     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3251     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3252     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3253         if(USES_LIST(top_type, list)){
3254             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3255             const int b8_xy= 4*top_xy + 2;
3256             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3257             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3258             ref_cache[0 - 1*8]=
3259             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3260             ref_cache[2 - 1*8]=
3261             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3262         }else{
3263             AV_ZERO128(mv_dst - 1*8);
3264             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3265         }
3266
3267         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3268             if(USES_LIST(left_type[LTOP], list)){
3269                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3270                 const int b8_xy= 4*left_xy[LTOP] + 1;
3271                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3272                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3273                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3274                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3275                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3276                 ref_cache[-1 +  0]=
3277                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3278                 ref_cache[-1 + 16]=
3279                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3280             }else{
3281                 AV_ZERO32(mv_dst - 1 + 0);
3282                 AV_ZERO32(mv_dst - 1 + 8);
3283                 AV_ZERO32(mv_dst - 1 +16);
3284                 AV_ZERO32(mv_dst - 1 +24);
3285                 ref_cache[-1 +  0]=
3286                 ref_cache[-1 +  8]=
3287                 ref_cache[-1 + 16]=
3288                 ref_cache[-1 + 24]= LIST_NOT_USED;
3289             }
3290         }
3291     }
3292
3293     if(!USES_LIST(mb_type, list)){
3294         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3295         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3296         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3297         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3298         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3299         return;
3300     }
3301
3302     {
3303         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3304         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3305         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3306         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3307         AV_WN32A(&ref_cache[0*8], ref01);
3308         AV_WN32A(&ref_cache[1*8], ref01);
3309         AV_WN32A(&ref_cache[2*8], ref23);
3310         AV_WN32A(&ref_cache[3*8], ref23);
3311     }
3312
3313     {
3314         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3315         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3316         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3317         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3318         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3319     }
3320 }
3321
3322 /**
3323  *
3324  * @return non zero if the loop filter can be skipped
3325  */
3326 static int fill_filter_caches(H264Context *h, int mb_type){
3327     MpegEncContext * const s = &h->s;
3328     const int mb_xy= h->mb_xy;
3329     int top_xy, left_xy[LEFT_MBS];
3330     int top_type, left_type[LEFT_MBS];
3331     uint8_t *nnz;
3332     uint8_t *nnz_cache;
3333
3334     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3335
3336     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3337      * stuff, I can't imagine that these complex rules are worth it. */
3338
3339     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3340     if(FRAME_MBAFF){
3341         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3342         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3343         if(s->mb_y&1){
3344             if (left_mb_field_flag != curr_mb_field_flag) {
3345                 left_xy[LTOP] -= s->mb_stride;
3346             }
3347         }else{
3348             if(curr_mb_field_flag){
3349                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3350             }
3351             if (left_mb_field_flag != curr_mb_field_flag) {
3352                 left_xy[LBOT] += s->mb_stride;
3353             }
3354         }
3355     }
3356
3357     h->top_mb_xy = top_xy;
3358     h->left_mb_xy[LTOP] = left_xy[LTOP];
3359     h->left_mb_xy[LBOT] = left_xy[LBOT];
3360     {
3361         //for sufficiently low qp, filtering wouldn't do anything
3362         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3363         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3364         int qp = s->current_picture.f.qscale_table[mb_xy];
3365         if(qp <= qp_thresh
3366            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3367            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3368             if(!FRAME_MBAFF)
3369                 return 1;
3370             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3371                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3372                 return 1;
3373         }
3374     }
3375
3376     top_type        = s->current_picture.f.mb_type[top_xy];
3377     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3378     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3379     if(h->deblocking_filter == 2){
3380         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3381         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3382     }else{
3383         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3384         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3385     }
3386     h->top_type       = top_type;
3387     h->left_type[LTOP]= left_type[LTOP];
3388     h->left_type[LBOT]= left_type[LBOT];
3389
3390     if(IS_INTRA(mb_type))
3391         return 0;
3392
3393     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3394     if(h->list_count == 2)
3395         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3396
3397     nnz = h->non_zero_count[mb_xy];
3398     nnz_cache = h->non_zero_count_cache;
3399     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3400     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3401     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3402     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3403     h->cbp= h->cbp_table[mb_xy];
3404
3405     if(top_type){
3406         nnz = h->non_zero_count[top_xy];
3407         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3408     }
3409
3410     if(left_type[LTOP]){
3411         nnz = h->non_zero_count[left_xy[LTOP]];
3412         nnz_cache[3+8*1]= nnz[3+0*4];
3413         nnz_cache[3+8*2]= nnz[3+1*4];
3414         nnz_cache[3+8*3]= nnz[3+2*4];
3415         nnz_cache[3+8*4]= nnz[3+3*4];
3416     }
3417
3418     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3419     if(!CABAC && h->pps.transform_8x8_mode){
3420         if(IS_8x8DCT(top_type)){
3421             nnz_cache[4+8*0]=
3422             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3423             nnz_cache[6+8*0]=
3424             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3425         }
3426         if(IS_8x8DCT(left_type[LTOP])){
3427             nnz_cache[3+8*1]=
3428             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3429         }
3430         if(IS_8x8DCT(left_type[LBOT])){
3431             nnz_cache[3+8*3]=
3432             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3433         }
3434
3435         if(IS_8x8DCT(mb_type)){
3436             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3437             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3438
3439             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3440             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3441
3442             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3443             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3444
3445             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3446             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3447         }
3448     }
3449
3450     return 0;
3451 }
3452
3453 static void loop_filter(H264Context *h, int start_x, int end_x){
3454     MpegEncContext * const s = &h->s;
3455     uint8_t  *dest_y, *dest_cb, *dest_cr;
3456     int linesize, uvlinesize, mb_x, mb_y;
3457     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3458     const int old_slice_type= h->slice_type;
3459     const int pixel_shift = h->pixel_shift;
3460     const int block_h = 16 >> s->chroma_y_shift;
3461
3462     if(h->deblocking_filter) {
3463         for(mb_x= start_x; mb_x<end_x; mb_x++){
3464             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3465                 int mb_xy, mb_type;
3466                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3467                 h->slice_num= h->slice_table[mb_xy];
3468                 mb_type = s->current_picture.f.mb_type[mb_xy];
3469                 h->list_count= h->list_counts[mb_xy];
3470
3471                 if(FRAME_MBAFF)
3472                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3473
3474                 s->mb_x= mb_x;
3475                 s->mb_y= mb_y;
3476                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3477                 dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3478                 dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3479                     //FIXME simplify above
3480
3481                 if (MB_FIELD) {
3482                     linesize   = h->mb_linesize   = s->linesize * 2;
3483                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3484                     if(mb_y&1){ //FIXME move out of this function?
3485                         dest_y -= s->linesize*15;
3486                         dest_cb-= s->uvlinesize * (block_h - 1);
3487                         dest_cr-= s->uvlinesize * (block_h - 1);
3488                     }
3489                 } else {
3490                     linesize   = h->mb_linesize   = s->linesize;
3491                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3492                 }
3493                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
3494                 if(fill_filter_caches(h, mb_type))
3495                     continue;
3496                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3497                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3498
3499                 if (FRAME_MBAFF) {
3500                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3501                 } else {
3502                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3503                 }
3504             }
3505         }
3506     }
3507     h->slice_type= old_slice_type;
3508     s->mb_x= end_x;
3509     s->mb_y= end_mb_y - FRAME_MBAFF;
3510     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3511     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3512 }
3513
3514 static void predict_field_decoding_flag(H264Context *h){
3515     MpegEncContext * const s = &h->s;
3516     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3517     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3518                 ? s->current_picture.f.mb_type[mb_xy - 1]
3519                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3520                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3521                 : 0;
3522     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3523 }
3524
3525 /**
3526  * Draw edges and report progress for the last MB row.
3527  */
3528 static void decode_finish_row(H264Context *h){
3529     MpegEncContext * const s = &h->s;
3530     int top = 16*(s->mb_y >> FIELD_PICTURE);
3531     int height = 16 << FRAME_MBAFF;
3532     int deblock_border = (16 + 4) << FRAME_MBAFF;
3533     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3534
3535     if (h->deblocking_filter) {
3536         if((top + height) >= pic_height)
3537             height += deblock_border;
3538
3539         top -= deblock_border;
3540     }
3541
3542     if (top >= pic_height || (top + height) < h->emu_edge_height)
3543         return;
3544
3545     height = FFMIN(height, pic_height - top);
3546     if (top < h->emu_edge_height) {
3547         height = top+height;
3548         top = 0;
3549     }
3550
3551     ff_draw_horiz_band(s, top, height);
3552
3553     if (s->dropable) return;
3554
3555     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3556                              s->picture_structure==PICT_BOTTOM_FIELD);
3557 }
3558
3559 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3560     H264Context *h = *(void**)arg;
3561     MpegEncContext * const s = &h->s;
3562     const int part_mask= s->partitioned_frame ? (ER_AC_END|ER_AC_ERROR) : 0x7F;
3563     int lf_x_start = s->mb_x;
3564
3565     s->mb_skip_run= -1;
3566
3567     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3568                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3569
3570     if( h->pps.cabac ) {
3571         /* realign */
3572         align_get_bits( &s->gb );
3573
3574         /* init cabac */
3575         ff_init_cabac_states( &h->cabac);
3576         ff_init_cabac_decoder( &h->cabac,
3577                                s->gb.buffer + get_bits_count(&s->gb)/8,
3578                                (get_bits_left(&s->gb) + 7)/8);
3579
3580         ff_h264_init_cabac_states(h);
3581
3582         for(;;){
3583 //START_TIMER
3584             int ret = ff_h264_decode_mb_cabac(h);
3585             int eos;
3586 //STOP_TIMER("decode_mb_cabac")
3587
3588             if(ret>=0) ff_h264_hl_decode_mb(h);
3589
3590             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3591                 s->mb_y++;
3592
3593                 ret = ff_h264_decode_mb_cabac(h);
3594
3595                 if(ret>=0) ff_h264_hl_decode_mb(h);
3596                 s->mb_y--;
3597             }
3598             eos = get_cabac_terminate( &h->cabac );
3599
3600             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3601                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, ER_MB_END&part_mask);
3602                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3603                 return 0;
3604             }
3605             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3606                 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);
3607                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, ER_MB_ERROR&part_mask);
3608                 return -1;
3609             }
3610
3611             if( ++s->mb_x >= s->mb_width ) {
3612                 loop_filter(h, lf_x_start, s->mb_x);
3613                 s->mb_x = lf_x_start = 0;
3614                 decode_finish_row(h);
3615                 ++s->mb_y;
3616                 if(FIELD_OR_MBAFF_PICTURE) {
3617                     ++s->mb_y;
3618                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3619                         predict_field_decoding_flag(h);
3620                 }
3621             }
3622
3623             if( eos || s->mb_y >= s->mb_height ) {
3624                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3625                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, ER_MB_END&part_mask);
3626                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3627                 return 0;
3628             }
3629         }
3630
3631     } else {
3632         for(;;){
3633             int ret = ff_h264_decode_mb_cavlc(h);
3634
3635             if(ret>=0) ff_h264_hl_decode_mb(h);
3636
3637             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3638                 s->mb_y++;
3639                 ret = ff_h264_decode_mb_cavlc(h);
3640
3641                 if(ret>=0) ff_h264_hl_decode_mb(h);
3642                 s->mb_y--;
3643             }
3644
3645             if(ret<0){
3646                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3647                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, ER_MB_ERROR&part_mask);
3648                 return -1;
3649             }
3650
3651             if(++s->mb_x >= s->mb_width){
3652                 loop_filter(h, lf_x_start, s->mb_x);
3653                 s->mb_x = lf_x_start = 0;
3654                 decode_finish_row(h);
3655                 ++s->mb_y;
3656                 if(FIELD_OR_MBAFF_PICTURE) {
3657                     ++s->mb_y;
3658                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3659                         predict_field_decoding_flag(h);
3660                 }
3661                 if(s->mb_y >= s->mb_height){
3662                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3663
3664                     if(get_bits_count(&s->gb) == s->gb.size_in_bits ) {
3665                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, ER_MB_END&part_mask);
3666
3667                         return 0;
3668                     }else{
3669                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, ER_MB_END&part_mask);
3670
3671                         return -1;
3672                     }
3673                 }
3674             }
3675
3676             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
3677                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3678                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
3679                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, ER_MB_END&part_mask);
3680                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3681
3682                     return 0;
3683                 }else{
3684                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, ER_MB_ERROR&part_mask);
3685
3686                     return -1;
3687                 }
3688             }
3689         }
3690     }
3691 }
3692
3693 /**
3694  * Call decode_slice() for each context.
3695  *
3696  * @param h h264 master context
3697  * @param context_count number of contexts to execute
3698  */
3699 static int execute_decode_slices(H264Context *h, int context_count){
3700     MpegEncContext * const s = &h->s;
3701     AVCodecContext * const avctx= s->avctx;
3702     H264Context *hx;
3703     int i;
3704
3705     if (s->avctx->hwaccel || s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3706         return 0;
3707     if(context_count == 1) {
3708         return decode_slice(avctx, &h);
3709     } else {
3710         for(i = 1; i < context_count; i++) {
3711             hx = h->thread_context[i];
3712             hx->s.err_recognition = avctx->err_recognition;
3713             hx->s.error_count = 0;
3714         }
3715
3716         avctx->execute(avctx, decode_slice,
3717                        h->thread_context, NULL, context_count, sizeof(void*));
3718
3719         /* pull back stuff from slices to master context */
3720         hx = h->thread_context[context_count - 1];
3721         s->mb_x = hx->s.mb_x;
3722         s->mb_y = hx->s.mb_y;
3723         s->dropable = hx->s.dropable;
3724         s->picture_structure = hx->s.picture_structure;
3725         for(i = 1; i < context_count; i++)
3726             h->s.error_count += h->thread_context[i]->s.error_count;
3727     }
3728
3729     return 0;
3730 }
3731
3732
3733 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
3734     MpegEncContext * const s = &h->s;
3735     AVCodecContext * const avctx= s->avctx;
3736     H264Context *hx; ///< thread context
3737     int buf_index;
3738     int context_count;
3739     int next_avc;
3740     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3741     int nals_needed=0; ///< number of NALs that need decoding before the next frame thread starts
3742     int nal_index;
3743
3744     h->max_contexts = s->slice_context_count;
3745     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
3746         h->current_slice = 0;
3747         if (!s->first_field)
3748             s->current_picture_ptr= NULL;
3749         ff_h264_reset_sei(h);
3750     }
3751
3752     for(;pass <= 1;pass++){
3753         buf_index = 0;
3754         context_count = 0;
3755         next_avc = h->is_avc ? 0 : buf_size;
3756         nal_index = 0;
3757     for(;;){
3758         int consumed;
3759         int dst_length;
3760         int bit_length;
3761         uint8_t *ptr;
3762         int i, nalsize = 0;
3763         int err;
3764
3765         if(buf_index >= next_avc) {
3766             if (buf_index >= buf_size - h->nal_length_size) break;
3767             nalsize = 0;
3768             for(i = 0; i < h->nal_length_size; i++)
3769                 nalsize = (nalsize << 8) | buf[buf_index++];
3770             if(nalsize <= 0 || nalsize > buf_size - buf_index){
3771                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
3772                 break;
3773             }
3774             next_avc= buf_index + nalsize;
3775         } else {
3776             // start code prefix search
3777             for(; buf_index + 3 < next_avc; buf_index++){
3778                 // This should always succeed in the first iteration.
3779                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
3780                     break;
3781             }
3782
3783             if(buf_index+3 >= buf_size) break;
3784
3785             buf_index+=3;
3786             if(buf_index >= next_avc) continue;
3787         }
3788
3789         hx = h->thread_context[context_count];
3790
3791         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
3792         if (ptr==NULL || dst_length < 0){
3793             return -1;
3794         }
3795         i= buf_index + consumed;
3796         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
3797            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
3798             s->workaround_bugs |= FF_BUG_TRUNCATED;
3799
3800         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
3801         while(ptr[dst_length - 1] == 0 && dst_length > 0)
3802             dst_length--;
3803         }
3804         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
3805
3806         if(s->avctx->debug&FF_DEBUG_STARTCODE){
3807             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);
3808         }
3809
3810         if (h->is_avc && (nalsize != consumed) && nalsize){
3811             // set trailing bits in the last partial byte to zero
3812             if (bit_length & 7)
3813                 ptr[bit_length >> 3] = ptr[bit_length >> 3] & (0xff << 8 - (bit_length & 7));
3814             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
3815         }
3816
3817         buf_index += consumed;
3818         nal_index++;
3819
3820         if(pass == 0) {
3821             // packets can sometimes contain multiple PPS/SPS
3822             // e.g. two PAFF field pictures in one packet, or a demuxer which splits NALs strangely
3823             // if so, when frame threading we can't start the next thread until we've read all of them
3824             switch (hx->nal_unit_type) {
3825                 case NAL_SPS:
3826                 case NAL_PPS:
3827                     nals_needed = nal_index;
3828                     break;
3829                 case NAL_IDR_SLICE:
3830                 case NAL_SLICE:
3831                     init_get_bits(&hx->s.gb, ptr, bit_length);
3832                     if (!get_ue_golomb(&hx->s.gb))
3833                         nals_needed = nal_index;
3834             }
3835             continue;
3836         }
3837
3838         //FIXME do not discard SEI id
3839         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
3840             continue;
3841
3842       again:
3843         err = 0;
3844         switch(hx->nal_unit_type){
3845         case NAL_IDR_SLICE:
3846             if (h->nal_unit_type != NAL_IDR_SLICE) {
3847                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
3848                 return -1;
3849             }
3850             idr(h); // FIXME ensure we don't lose some frames if there is reordering
3851         case NAL_SLICE:
3852             init_get_bits(&hx->s.gb, ptr, bit_length);
3853             hx->intra_gb_ptr=
3854             hx->inter_gb_ptr= &hx->s.gb;
3855             hx->s.data_partitioning = 0;
3856
3857             if((err = decode_slice_header(hx, h)))
3858                break;
3859
3860             s->current_picture_ptr->f.key_frame |=
3861                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
3862                     (h->sei_recovery_frame_cnt >= 0);
3863
3864             if (h->current_slice == 1) {
3865                 if(!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3866                     decode_postinit(h, nal_index >= nals_needed);
3867                 }
3868
3869                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3870                     return -1;
3871                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3872                     ff_vdpau_h264_picture_start(s);
3873             }
3874
3875             if(hx->redundant_pic_count==0
3876                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3877                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3878                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3879                && avctx->skip_frame < AVDISCARD_ALL){
3880                 if(avctx->hwaccel) {
3881                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
3882                         return -1;
3883                 }else
3884                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
3885                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
3886                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
3887                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
3888                 }else
3889                     context_count++;
3890             }
3891             break;
3892         case NAL_DPA:
3893             init_get_bits(&hx->s.gb, ptr, bit_length);
3894             hx->intra_gb_ptr=
3895             hx->inter_gb_ptr= NULL;
3896
3897             if ((err = decode_slice_header(hx, h)) < 0)
3898                 break;
3899
3900             hx->s.data_partitioning = 1;
3901
3902             break;
3903         case NAL_DPB:
3904             init_get_bits(&hx->intra_gb, ptr, bit_length);
3905             hx->intra_gb_ptr= &hx->intra_gb;
3906             break;
3907         case NAL_DPC:
3908             init_get_bits(&hx->inter_gb, ptr, bit_length);
3909             hx->inter_gb_ptr= &hx->inter_gb;
3910
3911             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
3912                && s->context_initialized
3913                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3914                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3915                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3916                && avctx->skip_frame < AVDISCARD_ALL)
3917                 context_count++;
3918             break;
3919         case NAL_SEI:
3920             init_get_bits(&s->gb, ptr, bit_length);
3921             ff_h264_decode_sei(h);
3922             break;
3923         case NAL_SPS:
3924             init_get_bits(&s->gb, ptr, bit_length);
3925             ff_h264_decode_seq_parameter_set(h);
3926
3927             if (s->flags& CODEC_FLAG_LOW_DELAY ||
3928                 (h->sps.bitstream_restriction_flag && !h->sps.num_reorder_frames))
3929                 s->low_delay=1;
3930
3931             if(avctx->has_b_frames < 2)
3932                 avctx->has_b_frames= !s->low_delay;
3933
3934             if (avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
3935                 h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
3936                 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
3937                     avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3938                     h->cur_chroma_format_idc = h->sps.chroma_format_idc;
3939                     h->pixel_shift = h->sps.bit_depth_luma > 8;
3940
3941                     ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3942                     ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3943                     s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
3944                     dsputil_init(&s->dsp, s->avctx);
3945                 } else {
3946                     av_log(avctx, AV_LOG_ERROR, "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3947                     return -1;
3948                 }
3949             }
3950             break;
3951         case NAL_PPS:
3952             init_get_bits(&s->gb, ptr, bit_length);
3953
3954             ff_h264_decode_picture_parameter_set(h, bit_length);
3955
3956             break;
3957         case NAL_AUD:
3958         case NAL_END_SEQUENCE:
3959         case NAL_END_STREAM:
3960         case NAL_FILLER_DATA:
3961         case NAL_SPS_EXT:
3962         case NAL_AUXILIARY_SLICE:
3963             break;
3964         default:
3965             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
3966         }
3967
3968         if(context_count == h->max_contexts) {
3969             execute_decode_slices(h, context_count);
3970             context_count = 0;
3971         }
3972
3973         if (err < 0)
3974             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3975         else if(err == 1) {
3976             /* Slice could not be decoded in parallel mode, copy down
3977              * NAL unit stuff to context 0 and restart. Note that
3978              * rbsp_buffer is not transferred, but since we no longer
3979              * run in parallel mode this should not be an issue. */
3980             h->nal_unit_type = hx->nal_unit_type;
3981             h->nal_ref_idc   = hx->nal_ref_idc;
3982             hx = h;
3983             goto again;
3984         }
3985     }
3986     }
3987     if(context_count)
3988         execute_decode_slices(h, context_count);
3989     return buf_index;
3990 }
3991
3992 /**
3993  * Return the number of bytes consumed for building the current frame.
3994  */
3995 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
3996         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
3997         if(pos+10>buf_size) pos=buf_size; // oops ;)
3998
3999         return pos;
4000 }
4001
4002 static int decode_frame(AVCodecContext *avctx,
4003                              void *data, int *data_size,
4004                              AVPacket *avpkt)
4005 {
4006     const uint8_t *buf = avpkt->data;
4007     int buf_size = avpkt->size;
4008     H264Context *h = avctx->priv_data;
4009     MpegEncContext *s = &h->s;
4010     AVFrame *pict = data;
4011     int buf_index = 0;
4012
4013     s->flags= avctx->flags;
4014     s->flags2= avctx->flags2;
4015
4016    /* end of stream, output what is still in the buffers */
4017  out:
4018     if (buf_size == 0) {
4019         Picture *out;
4020         int i, out_idx;
4021