h264: cap max has_b_frames at MAX_DELAYED_PIC_COUNT - 1.
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of Libav.
6  *
7  * Libav is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * Libav is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with Libav; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "libavutil/imgutils.h"
29 #include "internal.h"
30 #include "dsputil.h"
31 #include "avcodec.h"
32 #include "mpegvideo.h"
33 #include "h264.h"
34 #include "h264data.h"
35 #include "h264_mvpred.h"
36 #include "golomb.h"
37 #include "mathops.h"
38 #include "rectangle.h"
39 #include "thread.h"
40 #include "vdpau_internal.h"
41 #include "libavutil/avassert.h"
42
43 #include "cabac.h"
44
45 //#undef NDEBUG
46 #include <assert.h>
47
48 static const uint8_t rem6[QP_MAX_NUM+1]={
49 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
50 };
51
52 static const uint8_t div6[QP_MAX_NUM+1]={
53 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9,10,10,10,10,
54 };
55
56 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
57     PIX_FMT_DXVA2_VLD,
58     PIX_FMT_VAAPI_VLD,
59     PIX_FMT_VDA_VLD,
60     PIX_FMT_YUVJ420P,
61     PIX_FMT_NONE
62 };
63
64 /**
65  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
66  */
67 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
68     MpegEncContext * const s = &h->s;
69     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
70     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
71     int i;
72
73     if(!(h->top_samples_available&0x8000)){
74         for(i=0; i<4; i++){
75             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
76             if(status<0){
77                 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);
78                 return -1;
79             } else if(status){
80                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
81             }
82         }
83     }
84
85     if((h->left_samples_available&0x8888)!=0x8888){
86         static const int mask[4]={0x8000,0x2000,0x80,0x20};
87         for(i=0; i<4; i++){
88             if(!(h->left_samples_available&mask[i])){
89                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
90                 if(status<0){
91                     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);
92                     return -1;
93                 } else if(status){
94                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
95                 }
96             }
97         }
98     }
99
100     return 0;
101 } //FIXME cleanup like ff_h264_check_intra_pred_mode
102
103 /**
104  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
105  */
106 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
107     MpegEncContext * const s = &h->s;
108     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
109     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
110
111     if(mode > 6U) {
112         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);
113         return -1;
114     }
115
116     if(!(h->top_samples_available&0x8000)){
117         mode= top[ mode ];
118         if(mode<0){
119             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);
120             return -1;
121         }
122     }
123
124     if((h->left_samples_available&0x8080) != 0x8080){
125         mode= left[ mode ];
126         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
127             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
128         }
129         if(mode<0){
130             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);
131             return -1;
132         }
133     }
134
135     return mode;
136 }
137
138 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
139     int i, si, di;
140     uint8_t *dst;
141     int bufidx;
142
143 //    src[0]&0x80;                //forbidden bit
144     h->nal_ref_idc= src[0]>>5;
145     h->nal_unit_type= src[0]&0x1F;
146
147     src++; length--;
148
149 #if HAVE_FAST_UNALIGNED
150 # if HAVE_FAST_64BIT
151 #   define RS 7
152     for(i=0; i+1<length; i+=9){
153         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
154 # else
155 #   define RS 3
156     for(i=0; i+1<length; i+=5){
157         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
158 # endif
159             continue;
160         if(i>0 && !src[i]) i--;
161         while(src[i]) i++;
162 #else
163 #   define RS 0
164     for(i=0; i+1<length; i+=2){
165         if(src[i]) continue;
166         if(i>0 && src[i-1]==0) i--;
167 #endif
168         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
169             if(src[i+2]!=3){
170                 /* startcode, so we must be past the end */
171                 length=i;
172             }
173             break;
174         }
175         i-= RS;
176     }
177
178     if(i>=length-1){ //no escaped 0
179         *dst_length= length;
180         *consumed= length+1; //+1 for the header
181         return src;
182     }
183
184     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
185     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
186     dst= h->rbsp_buffer[bufidx];
187
188     if (dst == NULL){
189         return NULL;
190     }
191
192 //printf("decoding esc\n");
193     memcpy(dst, src, i);
194     si=di=i;
195     while(si+2<length){
196         //remove escapes (very rare 1:2^22)
197         if(src[si+2]>3){
198             dst[di++]= src[si++];
199             dst[di++]= src[si++];
200         }else if(src[si]==0 && src[si+1]==0){
201             if(src[si+2]==3){ //escape
202                 dst[di++]= 0;
203                 dst[di++]= 0;
204                 si+=3;
205                 continue;
206             }else //next start code
207                 goto nsc;
208         }
209
210         dst[di++]= src[si++];
211     }
212     while(si<length)
213         dst[di++]= src[si++];
214 nsc:
215
216     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
217
218     *dst_length= di;
219     *consumed= si + 1;//+1 for the header
220 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
221     return dst;
222 }
223
224 /**
225  * Identify the exact end of the bitstream
226  * @return the length of the trailing, or 0 if damaged
227  */
228 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
229     int v= *src;
230     int r;
231
232     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
233
234     for(r=1; r<9; r++){
235         if(v&1) return r;
236         v>>=1;
237     }
238     return 0;
239 }
240
241 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
242                                  int y_offset, int list){
243     int raw_my= h->mv_cache[list][ scan8[n] ][1];
244     int filter_height= (raw_my&3) ? 2 : 0;
245     int full_my= (raw_my>>2) + y_offset;
246     int top = full_my - filter_height, bottom = full_my + height + filter_height;
247
248     return FFMAX(abs(top), bottom);
249 }
250
251 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
252                                int y_offset, int list0, int list1, int *nrefs){
253     MpegEncContext * const s = &h->s;
254     int my;
255
256     y_offset += 16*(s->mb_y >> MB_FIELD);
257
258     if(list0){
259         int ref_n = h->ref_cache[0][ scan8[n] ];
260         Picture *ref= &h->ref_list[0][ref_n];
261
262         // Error resilience puts the current picture in the ref list.
263         // Don't try to wait on these as it will cause a deadlock.
264         // Fields can wait on each other, though.
265         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
266            (ref->f.reference & 3) != s->picture_structure) {
267             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
268             if (refs[0][ref_n] < 0) nrefs[0] += 1;
269             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
270         }
271     }
272
273     if(list1){
274         int ref_n = h->ref_cache[1][ scan8[n] ];
275         Picture *ref= &h->ref_list[1][ref_n];
276
277         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
278            (ref->f.reference & 3) != s->picture_structure) {
279             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
280             if (refs[1][ref_n] < 0) nrefs[1] += 1;
281             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
282         }
283     }
284 }
285
286 /**
287  * Wait until all reference frames are available for MC operations.
288  *
289  * @param h the H264 context
290  */
291 static void await_references(H264Context *h){
292     MpegEncContext * const s = &h->s;
293     const int mb_xy= h->mb_xy;
294     const int mb_type = s->current_picture.f.mb_type[mb_xy];
295     int refs[2][48];
296     int nrefs[2] = {0};
297     int ref, list;
298
299     memset(refs, -1, sizeof(refs));
300
301     if(IS_16X16(mb_type)){
302         get_lowest_part_y(h, refs, 0, 16, 0,
303                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
304     }else if(IS_16X8(mb_type)){
305         get_lowest_part_y(h, refs, 0, 8, 0,
306                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
307         get_lowest_part_y(h, refs, 8, 8, 8,
308                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
309     }else if(IS_8X16(mb_type)){
310         get_lowest_part_y(h, refs, 0, 16, 0,
311                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
312         get_lowest_part_y(h, refs, 4, 16, 0,
313                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
314     }else{
315         int i;
316
317         assert(IS_8X8(mb_type));
318
319         for(i=0; i<4; i++){
320             const int sub_mb_type= h->sub_mb_type[i];
321             const int n= 4*i;
322             int y_offset= (i&2)<<2;
323
324             if(IS_SUB_8X8(sub_mb_type)){
325                 get_lowest_part_y(h, refs, n  , 8, y_offset,
326                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
327             }else if(IS_SUB_8X4(sub_mb_type)){
328                 get_lowest_part_y(h, refs, n  , 4, y_offset,
329                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
330                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
331                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
332             }else if(IS_SUB_4X8(sub_mb_type)){
333                 get_lowest_part_y(h, refs, n  , 8, y_offset,
334                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
335                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
336                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
337             }else{
338                 int j;
339                 assert(IS_SUB_4X4(sub_mb_type));
340                 for(j=0; j<4; j++){
341                     int sub_y_offset= y_offset + 2*(j&2);
342                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
343                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
344                 }
345             }
346         }
347     }
348
349     for(list=h->list_count-1; list>=0; list--){
350         for(ref=0; ref<48 && nrefs[list]; ref++){
351             int row = refs[list][ref];
352             if(row >= 0){
353                 Picture *ref_pic = &h->ref_list[list][ref];
354                 int ref_field = ref_pic->f.reference - 1;
355                 int ref_field_picture = ref_pic->field_picture;
356                 int pic_height = 16*s->mb_height >> ref_field_picture;
357
358                 row <<= MB_MBAFF;
359                 nrefs[list]--;
360
361                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
362                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
363                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
364                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
365                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
366                 }else if(FIELD_PICTURE){
367                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
368                 }else{
369                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
370                 }
371             }
372         }
373     }
374 }
375
376 #if 0
377 /**
378  * DCT transforms the 16 dc values.
379  * @param qp quantization parameter ??? FIXME
380  */
381 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
382 //    const int qmul= dequant_coeff[qp][0];
383     int i;
384     int temp[16]; //FIXME check if this is a good idea
385     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
386     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
387
388     for(i=0; i<4; i++){
389         const int offset= y_offset[i];
390         const int z0= block[offset+stride*0] + block[offset+stride*4];
391         const int z1= block[offset+stride*0] - block[offset+stride*4];
392         const int z2= block[offset+stride*1] - block[offset+stride*5];
393         const int z3= block[offset+stride*1] + block[offset+stride*5];
394
395         temp[4*i+0]= z0+z3;
396         temp[4*i+1]= z1+z2;
397         temp[4*i+2]= z1-z2;
398         temp[4*i+3]= z0-z3;
399     }
400
401     for(i=0; i<4; i++){
402         const int offset= x_offset[i];
403         const int z0= temp[4*0+i] + temp[4*2+i];
404         const int z1= temp[4*0+i] - temp[4*2+i];
405         const int z2= temp[4*1+i] - temp[4*3+i];
406         const int z3= temp[4*1+i] + temp[4*3+i];
407
408         block[stride*0 +offset]= (z0 + z3)>>1;
409         block[stride*2 +offset]= (z1 + z2)>>1;
410         block[stride*8 +offset]= (z1 - z2)>>1;
411         block[stride*10+offset]= (z0 - z3)>>1;
412     }
413 }
414 #endif
415
416 #undef xStride
417 #undef stride
418
419 #if 0
420 static void chroma_dc_dct_c(DCTELEM *block){
421     const int stride= 16*2;
422     const int xStride= 16;
423     int a,b,c,d,e;
424
425     a= block[stride*0 + xStride*0];
426     b= block[stride*0 + xStride*1];
427     c= block[stride*1 + xStride*0];
428     d= block[stride*1 + xStride*1];
429
430     e= a-b;
431     a= a+b;
432     b= c-d;
433     c= c+d;
434
435     block[stride*0 + xStride*0]= (a+c);
436     block[stride*0 + xStride*1]= (e+b);
437     block[stride*1 + xStride*0]= (a-c);
438     block[stride*1 + xStride*1]= (e-b);
439 }
440 #endif
441
442 static av_always_inline void
443 mc_dir_part(H264Context *h, Picture *pic, int n, int square,
444             int height, int delta, int list,
445             uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
446             int src_x_offset, int src_y_offset,
447             qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
448             int pixel_shift, int chroma_idc)
449 {
450     MpegEncContext * const s = &h->s;
451     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
452     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
453     const int luma_xy= (mx&3) + ((my&3)<<2);
454     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
455     uint8_t * src_y = pic->f.data[0] + offset;
456     uint8_t * src_cb, * src_cr;
457     int extra_width= h->emu_edge_width;
458     int extra_height= h->emu_edge_height;
459     int emu=0;
460     const int full_mx= mx>>2;
461     const int full_my= my>>2;
462     const int pic_width  = 16*s->mb_width;
463     const int pic_height = 16*s->mb_height >> MB_FIELD;
464     int ysh;
465
466     if(mx&7) extra_width -= 3;
467     if(my&7) extra_height -= 3;
468
469     if(   full_mx < 0-extra_width
470        || full_my < 0-extra_height
471        || full_mx + 16/*FIXME*/ > pic_width + extra_width
472        || full_my + 16/*FIXME*/ > pic_height + extra_height){
473         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_y - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
474                                 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
475             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
476         emu=1;
477     }
478
479     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
480     if(!square){
481         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
482     }
483
484     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
485
486     if(chroma_idc == 3 /* yuv444 */){
487         src_cb = pic->f.data[1] + offset;
488         if(emu){
489             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
490                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
491             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
492         }
493         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
494         if(!square){
495             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
496         }
497
498         src_cr = pic->f.data[2] + offset;
499         if(emu){
500             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
501                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
502             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
503         }
504         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
505         if(!square){
506             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
507         }
508         return;
509     }
510
511     ysh = 3 - (chroma_idc == 2 /* yuv422 */);
512     if(chroma_idc == 1 /* yuv420 */ && MB_FIELD){
513         // chroma offset when predicting from a field of opposite parity
514         my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
515         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
516     }
517
518     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) + (my >> ysh) * h->mb_uvlinesize;
519     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) + (my >> ysh) * h->mb_uvlinesize;
520
521     if(emu){
522         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
523                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
524                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
525             src_cb= s->edge_emu_buffer;
526     }
527     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
528               mx&7, (my << (chroma_idc == 2 /* yuv422 */)) &7);
529
530     if(emu){
531         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
532                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
533                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
534             src_cr= s->edge_emu_buffer;
535     }
536     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
537               mx&7, (my << (chroma_idc == 2 /* yuv422 */)) &7);
538 }
539
540 static av_always_inline void
541 mc_part_std(H264Context *h, int n, int square, int height, int delta,
542             uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
543             int x_offset, int y_offset,
544             qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
545             qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
546             int list0, int list1, int pixel_shift, int chroma_idc)
547 {
548     MpegEncContext * const s = &h->s;
549     qpel_mc_func *qpix_op=  qpix_put;
550     h264_chroma_mc_func chroma_op= chroma_put;
551
552     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
553     if (chroma_idc == 3 /* yuv444 */) {
554         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
555         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
556     } else if (chroma_idc == 2 /* yuv422 */) {
557         dest_cb += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
558         dest_cr += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
559     } else /* yuv420 */ {
560         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
561         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
562     }
563     x_offset += 8*s->mb_x;
564     y_offset += 8*(s->mb_y >> MB_FIELD);
565
566     if(list0){
567         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
568         mc_dir_part(h, ref, n, square, height, delta, 0,
569                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
570                            qpix_op, chroma_op, pixel_shift, chroma_idc);
571
572         qpix_op=  qpix_avg;
573         chroma_op= chroma_avg;
574     }
575
576     if(list1){
577         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
578         mc_dir_part(h, ref, n, square, height, delta, 1,
579                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
580                            qpix_op, chroma_op, pixel_shift, chroma_idc);
581     }
582 }
583
584 static av_always_inline void
585 mc_part_weighted(H264Context *h, int n, int square, int height, int delta,
586                  uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
587                  int x_offset, int y_offset,
588                  qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
589                  h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
590                  h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
591                  int list0, int list1, int pixel_shift, int chroma_idc){
592     MpegEncContext * const s = &h->s;
593     int chroma_height;
594
595     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
596     if (chroma_idc == 3 /* yuv444 */) {
597         chroma_height = height;
598         chroma_weight_avg = luma_weight_avg;
599         chroma_weight_op = luma_weight_op;
600         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
601         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
602     } else if (chroma_idc == 2 /* yuv422 */) {
603         chroma_height = height;
604         dest_cb += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
605         dest_cr += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
606     } else /* yuv420 */ {
607         chroma_height = height >> 1;
608         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
609         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
610     }
611     x_offset += 8*s->mb_x;
612     y_offset += 8*(s->mb_y >> MB_FIELD);
613
614     if(list0 && list1){
615         /* don't optimize for luma-only case, since B-frames usually
616          * use implicit weights => chroma too. */
617         uint8_t *tmp_cb = s->obmc_scratchpad;
618         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
619         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
620         int refn0 = h->ref_cache[0][ scan8[n] ];
621         int refn1 = h->ref_cache[1][ scan8[n] ];
622
623         mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
624                     dest_y, dest_cb, dest_cr,
625                     x_offset, y_offset, qpix_put, chroma_put,
626                     pixel_shift, chroma_idc);
627         mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
628                     tmp_y, tmp_cb, tmp_cr,
629                     x_offset, y_offset, qpix_put, chroma_put,
630                     pixel_shift, chroma_idc);
631
632         if(h->use_weight == 2){
633             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
634             int weight1 = 64 - weight0;
635             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize,
636                               height,        5, weight0, weight1, 0);
637             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
638                               chroma_height, 5, weight0, weight1, 0);
639             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
640                               chroma_height, 5, weight0, weight1, 0);
641         }else{
642             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height, h->luma_log2_weight_denom,
643                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
644                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
645             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
646                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
647                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
648             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
649                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
650                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
651         }
652     }else{
653         int list = list1 ? 1 : 0;
654         int refn = h->ref_cache[list][ scan8[n] ];
655         Picture *ref= &h->ref_list[list][refn];
656         mc_dir_part(h, ref, n, square, height, delta, list,
657                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
658                     qpix_put, chroma_put, pixel_shift, chroma_idc);
659
660         luma_weight_op(dest_y, h->mb_linesize, height, h->luma_log2_weight_denom,
661                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
662         if(h->use_weight_chroma){
663             chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
664                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
665             chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
666                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
667         }
668     }
669 }
670
671 static av_always_inline void
672 mc_part(H264Context *h, int n, int square, int height, int delta,
673         uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
674         int x_offset, int y_offset,
675         qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
676         qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
677         h264_weight_func *weight_op, h264_biweight_func *weight_avg,
678         int list0, int list1, int pixel_shift, int chroma_idc)
679 {
680     if((h->use_weight==2 && list0 && list1
681         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
682        || h->use_weight==1)
683         mc_part_weighted(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
684                          x_offset, y_offset, qpix_put, chroma_put,
685                          weight_op[0], weight_op[1], weight_avg[0],
686                          weight_avg[1], list0, list1, pixel_shift, chroma_idc);
687     else
688         mc_part_std(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
689                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
690                     chroma_avg, list0, list1, pixel_shift, chroma_idc);
691 }
692
693 static av_always_inline void
694 prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma_idc)
695 {
696     /* fetch pixels for estimated mv 4 macroblocks ahead
697      * optimized for 64byte cache lines */
698     MpegEncContext * const s = &h->s;
699     const int refn = h->ref_cache[list][scan8[0]];
700     if(refn >= 0){
701         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
702         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
703         uint8_t **src = h->ref_list[list][refn].f.data;
704         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
705         s->dsp.prefetch(src[0]+off, s->linesize, 4);
706         if (chroma_idc == 3 /* yuv444 */) {
707             s->dsp.prefetch(src[1]+off, s->linesize, 4);
708             s->dsp.prefetch(src[2]+off, s->linesize, 4);
709         }else{
710             off= ((mx>>1) << pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + (64 << pixel_shift);
711             s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
712         }
713     }
714 }
715
716 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
717                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
718                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
719                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
720                       int pixel_shift, int chroma_idc)
721 {
722     MpegEncContext * const s = &h->s;
723     const int mb_xy= h->mb_xy;
724     const int mb_type = s->current_picture.f.mb_type[mb_xy];
725
726     assert(IS_INTER(mb_type));
727
728     if(HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
729         await_references(h);
730     prefetch_motion(h, 0, pixel_shift, chroma_idc);
731
732     if(IS_16X16(mb_type)){
733         mc_part(h, 0, 1, 16, 0, dest_y, dest_cb, dest_cr, 0, 0,
734                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
735                 weight_op, weight_avg,
736                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
737                 pixel_shift, chroma_idc);
738     }else if(IS_16X8(mb_type)){
739         mc_part(h, 0, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
740                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
741                 weight_op, weight_avg,
742                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
743                 pixel_shift, chroma_idc);
744         mc_part(h, 8, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
745                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
746                 weight_op, weight_avg,
747                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
748                 pixel_shift, chroma_idc);
749     }else if(IS_8X16(mb_type)){
750         mc_part(h, 0, 0, 16, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
751                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
752                 &weight_op[1], &weight_avg[1],
753                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
754                 pixel_shift, chroma_idc);
755         mc_part(h, 4, 0, 16, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
756                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
757                 &weight_op[1], &weight_avg[1],
758                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
759                 pixel_shift, chroma_idc);
760     }else{
761         int i;
762
763         assert(IS_8X8(mb_type));
764
765         for(i=0; i<4; i++){
766             const int sub_mb_type= h->sub_mb_type[i];
767             const int n= 4*i;
768             int x_offset= (i&1)<<2;
769             int y_offset= (i&2)<<1;
770
771             if(IS_SUB_8X8(sub_mb_type)){
772                 mc_part(h, n, 1, 8, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
773                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
774                     &weight_op[1], &weight_avg[1],
775                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
776                     pixel_shift, chroma_idc);
777             }else if(IS_SUB_8X4(sub_mb_type)){
778                 mc_part(h, n  , 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
779                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
780                     &weight_op[1], &weight_avg[1],
781                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
782                     pixel_shift, chroma_idc);
783                 mc_part(h, n+2, 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
784                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
785                     &weight_op[1], &weight_avg[1],
786                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
787                     pixel_shift, chroma_idc);
788             }else if(IS_SUB_4X8(sub_mb_type)){
789                 mc_part(h, n  , 0, 8, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
790                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
791                     &weight_op[2], &weight_avg[2],
792                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
793                     pixel_shift, chroma_idc);
794                 mc_part(h, n+1, 0, 8, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
795                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
796                     &weight_op[2], &weight_avg[2],
797                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
798                     pixel_shift, chroma_idc);
799             }else{
800                 int j;
801                 assert(IS_SUB_4X4(sub_mb_type));
802                 for(j=0; j<4; j++){
803                     int sub_x_offset= x_offset + 2*(j&1);
804                     int sub_y_offset= y_offset +   (j&2);
805                     mc_part(h, n+j, 1, 4, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
806                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
807                         &weight_op[2], &weight_avg[2],
808                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
809                         pixel_shift, chroma_idc);
810                 }
811             }
812         }
813     }
814
815     prefetch_motion(h, 1, pixel_shift, chroma_idc);
816 }
817
818 static av_always_inline void
819 hl_motion_420(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
820               qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
821               qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
822               h264_weight_func *weight_op, h264_biweight_func *weight_avg,
823               int pixel_shift)
824 {
825     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
826               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 1);
827 }
828
829 static av_always_inline void
830 hl_motion_422(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
831               qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
832               qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
833               h264_weight_func *weight_op, h264_biweight_func *weight_avg,
834               int pixel_shift)
835 {
836     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
837               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 2);
838 }
839
840 static void free_tables(H264Context *h, int free_rbsp){
841     int i;
842     H264Context *hx;
843
844     av_freep(&h->intra4x4_pred_mode);
845     av_freep(&h->chroma_pred_mode_table);
846     av_freep(&h->cbp_table);
847     av_freep(&h->mvd_table[0]);
848     av_freep(&h->mvd_table[1]);
849     av_freep(&h->direct_table);
850     av_freep(&h->non_zero_count);
851     av_freep(&h->slice_table_base);
852     h->slice_table= NULL;
853     av_freep(&h->list_counts);
854
855     av_freep(&h->mb2b_xy);
856     av_freep(&h->mb2br_xy);
857
858     for(i = 0; i < MAX_THREADS; i++) {
859         hx = h->thread_context[i];
860         if(!hx) continue;
861         av_freep(&hx->top_borders[1]);
862         av_freep(&hx->top_borders[0]);
863         av_freep(&hx->s.obmc_scratchpad);
864         if (free_rbsp){
865             av_freep(&hx->rbsp_buffer[1]);
866             av_freep(&hx->rbsp_buffer[0]);
867             hx->rbsp_buffer_size[0] = 0;
868             hx->rbsp_buffer_size[1] = 0;
869         }
870         if (i) av_freep(&h->thread_context[i]);
871     }
872 }
873
874 static void init_dequant8_coeff_table(H264Context *h){
875     int i,j,q,x;
876     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
877
878     for(i=0; i<6; i++ ){
879         h->dequant8_coeff[i] = h->dequant8_buffer[i];
880         for(j=0; j<i; j++){
881             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
882                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
883                 break;
884             }
885         }
886         if(j<i)
887             continue;
888
889         for(q=0; q<max_qp+1; q++){
890             int shift = div6[q];
891             int idx = rem6[q];
892             for(x=0; x<64; x++)
893                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
894                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
895                     h->pps.scaling_matrix8[i][x]) << shift;
896         }
897     }
898 }
899
900 static void init_dequant4_coeff_table(H264Context *h){
901     int i,j,q,x;
902     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
903     for(i=0; i<6; i++ ){
904         h->dequant4_coeff[i] = h->dequant4_buffer[i];
905         for(j=0; j<i; j++){
906             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
907                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
908                 break;
909             }
910         }
911         if(j<i)
912             continue;
913
914         for(q=0; q<max_qp+1; q++){
915             int shift = div6[q] + 2;
916             int idx = rem6[q];
917             for(x=0; x<16; x++)
918                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
919                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
920                     h->pps.scaling_matrix4[i][x]) << shift;
921         }
922     }
923 }
924
925 static void init_dequant_tables(H264Context *h){
926     int i,x;
927     init_dequant4_coeff_table(h);
928     if(h->pps.transform_8x8_mode)
929         init_dequant8_coeff_table(h);
930     if(h->sps.transform_bypass){
931         for(i=0; i<6; i++)
932             for(x=0; x<16; x++)
933                 h->dequant4_coeff[i][0][x] = 1<<6;
934         if(h->pps.transform_8x8_mode)
935             for(i=0; i<6; i++)
936                 for(x=0; x<64; x++)
937                     h->dequant8_coeff[i][0][x] = 1<<6;
938     }
939 }
940
941
942 int ff_h264_alloc_tables(H264Context *h){
943     MpegEncContext * const s = &h->s;
944     const int big_mb_num= s->mb_stride * (s->mb_height+1);
945     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
946     int x,y;
947
948     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
949
950     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
951     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
952     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
953
954     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
955     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
956     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
957     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
958     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
959
960     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
961     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
962
963     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
964     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
965     for(y=0; y<s->mb_height; y++){
966         for(x=0; x<s->mb_width; x++){
967             const int mb_xy= x + y*s->mb_stride;
968             const int b_xy = 4*x + 4*y*h->b_stride;
969
970             h->mb2b_xy [mb_xy]= b_xy;
971             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
972         }
973     }
974
975     s->obmc_scratchpad = NULL;
976
977     if(!h->dequant4_coeff[0])
978         init_dequant_tables(h);
979
980     return 0;
981 fail:
982     free_tables(h, 1);
983     return -1;
984 }
985
986 /**
987  * Mimic alloc_tables(), but for every context thread.
988  */
989 static void clone_tables(H264Context *dst, H264Context *src, int i){
990     MpegEncContext * const s = &src->s;
991     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
992     dst->non_zero_count           = src->non_zero_count;
993     dst->slice_table              = src->slice_table;
994     dst->cbp_table                = src->cbp_table;
995     dst->mb2b_xy                  = src->mb2b_xy;
996     dst->mb2br_xy                 = src->mb2br_xy;
997     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
998     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
999     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
1000     dst->direct_table             = src->direct_table;
1001     dst->list_counts              = src->list_counts;
1002
1003     dst->s.obmc_scratchpad = NULL;
1004     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma, src->sps.chroma_format_idc);
1005 }
1006
1007 /**
1008  * Init context
1009  * Allocate buffers which are not shared amongst multiple threads.
1010  */
1011 static int context_init(H264Context *h){
1012     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
1013     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
1014
1015     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
1016     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
1017
1018     return 0;
1019 fail:
1020     return -1; // free_tables will clean up for us
1021 }
1022
1023 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
1024
1025 static av_cold void common_init(H264Context *h){
1026     MpegEncContext * const s = &h->s;
1027
1028     s->width = s->avctx->width;
1029     s->height = s->avctx->height;
1030     s->codec_id= s->avctx->codec->id;
1031
1032     ff_h264dsp_init(&h->h264dsp, 8, 1);
1033     ff_h264_pred_init(&h->hpc, s->codec_id, 8, 1);
1034
1035     h->dequant_coeff_pps= -1;
1036     s->unrestricted_mv=1;
1037
1038     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
1039
1040     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
1041     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
1042 }
1043
1044 int ff_h264_decode_extradata(H264Context *h)
1045 {
1046     AVCodecContext *avctx = h->s.avctx;
1047
1048     if(avctx->extradata[0] == 1){
1049         int i, cnt, nalsize;
1050         unsigned char *p = avctx->extradata;
1051
1052         h->is_avc = 1;
1053
1054         if(avctx->extradata_size < 7) {
1055             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1056             return -1;
1057         }
1058         /* sps and pps in the avcC always have length coded with 2 bytes,
1059            so put a fake nal_length_size = 2 while parsing them */
1060         h->nal_length_size = 2;
1061         // Decode sps from avcC
1062         cnt = *(p+5) & 0x1f; // Number of sps
1063         p += 6;
1064         for (i = 0; i < cnt; i++) {
1065             nalsize = AV_RB16(p) + 2;
1066             if (p - avctx->extradata + nalsize > avctx->extradata_size)
1067                 return -1;
1068             if(decode_nal_units(h, p, nalsize) < 0) {
1069                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1070                 return -1;
1071             }
1072             p += nalsize;
1073         }
1074         // Decode pps from avcC
1075         cnt = *(p++); // Number of pps
1076         for (i = 0; i < cnt; i++) {
1077             nalsize = AV_RB16(p) + 2;
1078             if (p - avctx->extradata + nalsize > avctx->extradata_size)
1079                 return -1;
1080             if (decode_nal_units(h, p, nalsize) < 0) {
1081                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1082                 return -1;
1083             }
1084             p += nalsize;
1085         }
1086         // Now store right nal length size, that will be use to parse all other nals
1087         h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1088     } else {
1089         h->is_avc = 0;
1090         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1091             return -1;
1092     }
1093     return 0;
1094 }
1095
1096 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1097     H264Context *h= avctx->priv_data;
1098     MpegEncContext * const s = &h->s;
1099     int i;
1100
1101     MPV_decode_defaults(s);
1102
1103     s->avctx = avctx;
1104     common_init(h);
1105
1106     s->out_format = FMT_H264;
1107     s->workaround_bugs= avctx->workaround_bugs;
1108
1109     // set defaults
1110 //    s->decode_mb= ff_h263_decode_mb;
1111     s->quarter_sample = 1;
1112     if(!avctx->has_b_frames)
1113     s->low_delay= 1;
1114
1115     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1116
1117     ff_h264_decode_init_vlc();
1118
1119     h->pixel_shift = 0;
1120     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1121
1122     h->thread_context[0] = h;
1123     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1124     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1125         h->last_pocs[i] = INT_MIN;
1126     h->prev_poc_msb= 1<<16;
1127     h->x264_build = -1;
1128     ff_h264_reset_sei(h);
1129     if(avctx->codec_id == CODEC_ID_H264){
1130         if(avctx->ticks_per_frame == 1){
1131             s->avctx->time_base.den *=2;
1132         }
1133         avctx->ticks_per_frame = 2;
1134     }
1135
1136     if(avctx->extradata_size > 0 && avctx->extradata &&
1137         ff_h264_decode_extradata(h))
1138         return -1;
1139
1140     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1141         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1142         s->low_delay = 0;
1143     }
1144
1145     return 0;
1146 }
1147
1148 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1149 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1150 {
1151     int i;
1152
1153     for (i=0; i<count; i++){
1154         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1155                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1156                 !from[i]));
1157         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1158     }
1159 }
1160
1161 static void copy_parameter_set(void **to, void **from, int count, int size)
1162 {
1163     int i;
1164
1165     for (i=0; i<count; i++){
1166         if (to[i] && !from[i]) av_freep(&to[i]);
1167         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1168
1169         if (from[i]) memcpy(to[i], from[i], size);
1170     }
1171 }
1172
1173 static int decode_init_thread_copy(AVCodecContext *avctx){
1174     H264Context *h= avctx->priv_data;
1175
1176     if (!avctx->internal->is_copy)
1177         return 0;
1178     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1179     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1180
1181     return 0;
1182 }
1183
1184 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1185 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1186     H264Context *h= dst->priv_data, *h1= src->priv_data;
1187     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1188     int inited = s->context_initialized, err;
1189     int i;
1190
1191     if(dst == src || !s1->context_initialized) return 0;
1192
1193     err = ff_mpeg_update_thread_context(dst, src);
1194     if(err) return err;
1195
1196     //FIXME handle width/height changing
1197     if(!inited){
1198         for(i = 0; i < MAX_SPS_COUNT; i++)
1199             av_freep(h->sps_buffers + i);
1200
1201         for(i = 0; i < MAX_PPS_COUNT; i++)
1202             av_freep(h->pps_buffers + i);
1203
1204         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1205         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1206         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1207         if (ff_h264_alloc_tables(h) < 0) {
1208             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1209             return AVERROR(ENOMEM);
1210         }
1211         context_init(h);
1212
1213         for(i=0; i<2; i++){
1214             h->rbsp_buffer[i] = NULL;
1215             h->rbsp_buffer_size[i] = 0;
1216         }
1217
1218         h->thread_context[0] = h;
1219
1220         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1221         // so this has to be allocated here
1222         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1223
1224         s->dsp.clear_blocks(h->mb);
1225         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1226     }
1227
1228     //extradata/NAL handling
1229     h->is_avc          = h1->is_avc;
1230
1231     //SPS/PPS
1232     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1233     h->sps             = h1->sps;
1234     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1235     h->pps             = h1->pps;
1236
1237     //Dequantization matrices
1238     //FIXME these are big - can they be only copied when PPS changes?
1239     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1240
1241     for(i=0; i<6; i++)
1242         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1243
1244     for(i=0; i<6; i++)
1245         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1246
1247     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1248
1249     //POC timing
1250     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1251
1252     //reference lists
1253     copy_fields(h, h1, ref_count, list_count);
1254     copy_fields(h, h1, ref_list,  intra_gb);
1255     copy_fields(h, h1, short_ref, cabac_init_idc);
1256
1257     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1258     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1259     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1260
1261     h->last_slice_type = h1->last_slice_type;
1262
1263     if(!s->current_picture_ptr) return 0;
1264
1265     if(!s->dropable) {
1266         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1267         h->prev_poc_msb     = h->poc_msb;
1268         h->prev_poc_lsb     = h->poc_lsb;
1269     }
1270     h->prev_frame_num_offset= h->frame_num_offset;
1271     h->prev_frame_num       = h->frame_num;
1272     h->outputed_poc         = h->next_outputed_poc;
1273
1274     return err;
1275 }
1276
1277 int ff_h264_frame_start(H264Context *h){
1278     MpegEncContext * const s = &h->s;
1279     int i;
1280     const int pixel_shift = h->pixel_shift;
1281     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1282
1283     if(MPV_frame_start(s, s->avctx) < 0)
1284         return -1;
1285     ff_er_frame_start(s);
1286     /*
1287      * MPV_frame_start uses pict_type to derive key_frame.
1288      * This is incorrect for H.264; IDR markings must be used.
1289      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1290      * See decode_nal_units().
1291      */
1292     s->current_picture_ptr->f.key_frame = 0;
1293     s->current_picture_ptr->mmco_reset= 0;
1294
1295     assert(s->linesize && s->uvlinesize);
1296
1297     for(i=0; i<16; i++){
1298         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1299         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1300     }
1301     for(i=0; i<16; i++){
1302         h->block_offset[16+i]=
1303         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1304         h->block_offset[48+16+i]=
1305         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1306     }
1307
1308     /* can't be in alloc_tables because linesize isn't known there.
1309      * FIXME: redo bipred weight to not require extra buffer? */
1310     for(i = 0; i < thread_count; i++)
1311         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1312             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1313
1314     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1315     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1316
1317 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1318
1319     // We mark the current picture as non-reference after allocating it, so
1320     // that if we break out due to an error it can be released automatically
1321     // in the next MPV_frame_start().
1322     // SVQ3 as well as most other codecs have only last/next/current and thus
1323     // get released even with set reference, besides SVQ3 and others do not
1324     // mark frames as reference later "naturally".
1325     if(s->codec_id != CODEC_ID_SVQ3)
1326         s->current_picture_ptr->f.reference = 0;
1327
1328     s->current_picture_ptr->field_poc[0]=
1329     s->current_picture_ptr->field_poc[1]= INT_MAX;
1330
1331     h->next_output_pic = NULL;
1332
1333     assert(s->current_picture_ptr->long_ref==0);
1334
1335     return 0;
1336 }
1337
1338 /**
1339   * Run setup operations that must be run after slice header decoding.
1340   * This includes finding the next displayed frame.
1341   *
1342   * @param h h264 master context
1343   * @param setup_finished enough NALs have been read that we can call
1344   * ff_thread_finish_setup()
1345   */
1346 static void decode_postinit(H264Context *h, int setup_finished){
1347     MpegEncContext * const s = &h->s;
1348     Picture *out = s->current_picture_ptr;
1349     Picture *cur = s->current_picture_ptr;
1350     int i, pics, out_of_order, out_idx;
1351
1352     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1353     s->current_picture_ptr->f.pict_type   = s->pict_type;
1354
1355     if (h->next_output_pic) return;
1356
1357     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1358         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1359         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1360         //to find this yet, so we assume the worst for now.
1361         //if (setup_finished)
1362         //    ff_thread_finish_setup(s->avctx);
1363         return;
1364     }
1365
1366     cur->f.interlaced_frame = 0;
1367     cur->f.repeat_pict      = 0;
1368
1369     /* Signal interlacing information externally. */
1370     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1371
1372     if(h->sps.pic_struct_present_flag){
1373         switch (h->sei_pic_struct)
1374         {
1375         case SEI_PIC_STRUCT_FRAME:
1376             break;
1377         case SEI_PIC_STRUCT_TOP_FIELD:
1378         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1379             cur->f.interlaced_frame = 1;
1380             break;
1381         case SEI_PIC_STRUCT_TOP_BOTTOM:
1382         case SEI_PIC_STRUCT_BOTTOM_TOP:
1383             if (FIELD_OR_MBAFF_PICTURE)
1384                 cur->f.interlaced_frame = 1;
1385             else
1386                 // try to flag soft telecine progressive
1387                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1388             break;
1389         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1390         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1391             // Signal the possibility of telecined film externally (pic_struct 5,6)
1392             // From these hints, let the applications decide if they apply deinterlacing.
1393             cur->f.repeat_pict = 1;
1394             break;
1395         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1396             // Force progressive here, as doubling interlaced frame is a bad idea.
1397             cur->f.repeat_pict = 2;
1398             break;
1399         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1400             cur->f.repeat_pict = 4;
1401             break;
1402         }
1403
1404         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1405             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1406     }else{
1407         /* Derive interlacing flag from used decoding process. */
1408         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1409     }
1410     h->prev_interlaced_frame = cur->f.interlaced_frame;
1411
1412     if (cur->field_poc[0] != cur->field_poc[1]){
1413         /* Derive top_field_first from field pocs. */
1414         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1415     }else{
1416         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1417             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1418             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1419               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1420                 cur->f.top_field_first = 1;
1421             else
1422                 cur->f.top_field_first = 0;
1423         }else{
1424             /* Most likely progressive */
1425             cur->f.top_field_first = 0;
1426         }
1427     }
1428
1429     //FIXME do something with unavailable reference frames
1430
1431     /* Sort B-frames into display order */
1432
1433     if(h->sps.bitstream_restriction_flag
1434        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1435         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1436         s->low_delay = 0;
1437     }
1438
1439     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1440        && !h->sps.bitstream_restriction_flag){
1441         s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1442         s->low_delay= 0;
1443     }
1444
1445     pics = 0;
1446     while(h->delayed_pic[pics]) pics++;
1447
1448     assert(pics <= MAX_DELAYED_PIC_COUNT);
1449
1450     h->delayed_pic[pics++] = cur;
1451     if (cur->f.reference == 0)
1452         cur->f.reference = DELAYED_PIC_REF;
1453
1454     out = h->delayed_pic[0];
1455     out_idx = 0;
1456     for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1457         if(h->delayed_pic[i]->poc < out->poc){
1458             out = h->delayed_pic[i];
1459             out_idx = i;
1460         }
1461     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1462         h->next_outputed_poc= INT_MIN;
1463     out_of_order = out->poc < h->next_outputed_poc;
1464
1465     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1466         { }
1467     else if (out_of_order && pics-1 == s->avctx->has_b_frames &&
1468              s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT) {
1469         int cnt = 0, invalid = 0;
1470         for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++) {
1471             cnt += out->poc < h->last_pocs[i];
1472             invalid += h->last_pocs[i] == INT_MIN;
1473         }
1474         if (invalid + cnt < MAX_DELAYED_PIC_COUNT) {
1475             s->avctx->has_b_frames = FFMAX(s->avctx->has_b_frames, cnt);
1476         } else if (cnt) {
1477             for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1478                 h->last_pocs[i] = INT_MIN;
1479         }
1480         s->low_delay = 0;
1481     } else if (s->low_delay &&
1482                ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2) ||
1483                 cur->f.pict_type == AV_PICTURE_TYPE_B)) {
1484         s->low_delay = 0;
1485         s->avctx->has_b_frames++;
1486     }
1487
1488     if(out_of_order || pics > s->avctx->has_b_frames){
1489         out->f.reference &= ~DELAYED_PIC_REF;
1490         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1491                          // or else the first thread can release the picture and reuse it unsafely
1492         for(i=out_idx; h->delayed_pic[i]; i++)
1493             h->delayed_pic[i] = h->delayed_pic[i+1];
1494     }
1495     memmove(h->last_pocs, &h->last_pocs[1], sizeof(*h->last_pocs) * (MAX_DELAYED_PIC_COUNT - 1));
1496     h->last_pocs[MAX_DELAYED_PIC_COUNT - 1] = out->poc;
1497     if(!out_of_order && pics > s->avctx->has_b_frames){
1498         h->next_output_pic = out;
1499         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1500             h->next_outputed_poc = INT_MIN;
1501         } else
1502             h->next_outputed_poc = out->poc;
1503     }else{
1504         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1505     }
1506
1507     if (setup_finished)
1508         ff_thread_finish_setup(s->avctx);
1509 }
1510
1511 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1512                                               uint8_t *src_cb, uint8_t *src_cr,
1513                                               int linesize, int uvlinesize, int simple)
1514 {
1515     MpegEncContext * const s = &h->s;
1516     uint8_t *top_border;
1517     int top_idx = 1;
1518     const int pixel_shift = h->pixel_shift;
1519     int chroma444 = CHROMA444;
1520     int chroma422 = CHROMA422;
1521
1522     src_y  -=   linesize;
1523     src_cb -= uvlinesize;
1524     src_cr -= uvlinesize;
1525
1526     if(!simple && FRAME_MBAFF){
1527         if(s->mb_y&1){
1528             if(!MB_MBAFF){
1529                 top_border = h->top_borders[0][s->mb_x];
1530                 AV_COPY128(top_border, src_y + 15*linesize);
1531                 if (pixel_shift)
1532                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1533                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1534                     if(chroma444){
1535                         if (pixel_shift){
1536                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1537                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1538                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1539                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1540                         } else {
1541                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1542                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1543                         }
1544                     } else if(chroma422) {
1545                         if (pixel_shift) {
1546                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1547                             AV_COPY128(top_border+48, src_cr + 15*uvlinesize);
1548                         } else {
1549                             AV_COPY64(top_border+16, src_cb +  15*uvlinesize);
1550                             AV_COPY64(top_border+24, src_cr +  15*uvlinesize);
1551                         }
1552                     } else {
1553                         if (pixel_shift) {
1554                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1555                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1556                         } else {
1557                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1558                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1559                         }
1560                     }
1561                 }
1562             }
1563         }else if(MB_MBAFF){
1564             top_idx = 0;
1565         }else
1566             return;
1567     }
1568
1569     top_border = h->top_borders[top_idx][s->mb_x];
1570     // There are two lines saved, the line above the the top macroblock of a pair,
1571     // and the line above the bottom macroblock
1572     AV_COPY128(top_border, src_y + 16*linesize);
1573     if (pixel_shift)
1574         AV_COPY128(top_border+16, src_y+16*linesize+16);
1575
1576     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1577         if(chroma444){
1578             if (pixel_shift){
1579                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1580                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1581                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1582                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1583             } else {
1584                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1585                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1586             }
1587         } else if(chroma422) {
1588             if (pixel_shift) {
1589                 AV_COPY128(top_border+32, src_cb+16*uvlinesize);
1590                 AV_COPY128(top_border+48, src_cr+16*uvlinesize);
1591             } else {
1592                 AV_COPY64(top_border+16, src_cb+16*uvlinesize);
1593                 AV_COPY64(top_border+24, src_cr+16*uvlinesize);
1594             }
1595         } else {
1596             if (pixel_shift) {
1597                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1598                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1599             } else {
1600                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1601                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1602             }
1603         }
1604     }
1605 }
1606
1607 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1608                                   uint8_t *src_cb, uint8_t *src_cr,
1609                                   int linesize, int uvlinesize,
1610                                   int xchg, int chroma444,
1611                                   int simple, int pixel_shift){
1612     MpegEncContext * const s = &h->s;
1613     int deblock_topleft;
1614     int deblock_top;
1615     int top_idx = 1;
1616     uint8_t *top_border_m1;
1617     uint8_t *top_border;
1618
1619     if(!simple && FRAME_MBAFF){
1620         if(s->mb_y&1){
1621             if(!MB_MBAFF)
1622                 return;
1623         }else{
1624             top_idx = MB_MBAFF ? 0 : 1;
1625         }
1626     }
1627
1628     if(h->deblocking_filter == 2) {
1629         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1630         deblock_top     = h->top_type;
1631     } else {
1632         deblock_topleft = (s->mb_x > 0);
1633         deblock_top     = (s->mb_y > !!MB_FIELD);
1634     }
1635
1636     src_y  -=   linesize + 1 + pixel_shift;
1637     src_cb -= uvlinesize + 1 + pixel_shift;
1638     src_cr -= uvlinesize + 1 + pixel_shift;
1639
1640     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1641     top_border    = h->top_borders[top_idx][s->mb_x];
1642
1643 #define XCHG(a,b,xchg)\
1644     if (pixel_shift) {\
1645         if (xchg) {\
1646             AV_SWAP64(b+0,a+0);\
1647             AV_SWAP64(b+8,a+8);\
1648         } else {\
1649             AV_COPY128(b,a); \
1650         }\
1651     } else \
1652 if (xchg) AV_SWAP64(b,a);\
1653 else      AV_COPY64(b,a);
1654
1655     if(deblock_top){
1656         if(deblock_topleft){
1657             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1658         }
1659         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1660         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1661         if(s->mb_x+1 < s->mb_width){
1662             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1663         }
1664     }
1665     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1666         if(chroma444){
1667             if(deblock_topleft){
1668                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1669                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1670             }
1671             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1672             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1673             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1674             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1675             if(s->mb_x+1 < s->mb_width){
1676                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1677                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1678             }
1679         } else {
1680             if(deblock_top){
1681                 if(deblock_topleft){
1682                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1683                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1684                 }
1685                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1686                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1687             }
1688         }
1689     }
1690 }
1691
1692 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1693     if (high_bit_depth) {
1694         return AV_RN32A(((int32_t*)mb) + index);
1695     } else
1696         return AV_RN16A(mb + index);
1697 }
1698
1699 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1700     if (high_bit_depth) {
1701         AV_WN32A(((int32_t*)mb) + index, value);
1702     } else
1703         AV_WN16A(mb + index, value);
1704 }
1705
1706 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1707                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1708 {
1709     MpegEncContext * const s = &h->s;
1710     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1711     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1712     int i;
1713     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1714     block_offset += 16*p;
1715     if(IS_INTRA4x4(mb_type)){
1716         if(simple || !s->encoding){
1717             if(IS_8x8DCT(mb_type)){
1718                 if(transform_bypass){
1719                     idct_dc_add =
1720                     idct_add    = s->dsp.add_pixels8;
1721                 }else{
1722                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1723                     idct_add    = h->h264dsp.h264_idct8_add;
1724                 }
1725                 for(i=0; i<16; i+=4){
1726                     uint8_t * const ptr= dest_y + block_offset[i];
1727                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1728                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1729                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1730                     }else{
1731                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1732                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1733                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1734                         if(nnz){
1735                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1736                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1737                             else
1738                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1739                         }
1740                     }
1741                 }
1742             }else{
1743                 if(transform_bypass){
1744                     idct_dc_add =
1745                     idct_add    = s->dsp.add_pixels4;
1746                 }else{
1747                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1748                     idct_add    = h->h264dsp.h264_idct_add;
1749                 }
1750                 for(i=0; i<16; i++){
1751                     uint8_t * const ptr= dest_y + block_offset[i];
1752                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1753
1754                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1755                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1756                     }else{
1757                         uint8_t *topright;
1758                         int nnz, tr;
1759                         uint64_t tr_high;
1760                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1761                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1762                             assert(s->mb_y || linesize <= block_offset[i]);
1763                             if(!topright_avail){
1764                                 if (pixel_shift) {
1765                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1766                                     topright= (uint8_t*) &tr_high;
1767                                 } else {
1768                                     tr= ptr[3 - linesize]*0x01010101u;
1769                                     topright= (uint8_t*) &tr;
1770                                 }
1771                             }else
1772                                 topright= ptr + (4 << pixel_shift) - linesize;
1773                         }else
1774                             topright= NULL;
1775
1776                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1777                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1778                         if(nnz){
1779                             if(is_h264){
1780                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1781                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1782                                 else
1783                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1784                             }else
1785                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1786                         }
1787                     }
1788                 }
1789             }
1790         }
1791     }else{
1792         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1793         if(is_h264){
1794             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1795                 if(!transform_bypass)
1796                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1797                 else{
1798                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1799                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1800                     for(i = 0; i < 16; i++)
1801                         dctcoef_set(h->mb+(p*256 << pixel_shift), pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1802                 }
1803             }
1804         }else
1805             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1806     }
1807 }
1808
1809 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1810                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1811 {
1812     MpegEncContext * const s = &h->s;
1813     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1814     int i;
1815     block_offset += 16*p;
1816     if(!IS_INTRA4x4(mb_type)){
1817         if(is_h264){
1818             if(IS_INTRA16x16(mb_type)){
1819                 if(transform_bypass){
1820                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1821                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1822                     }else{
1823                         for(i=0; i<16; i++){
1824                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1825                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1826                         }
1827                     }
1828                 }else{
1829                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1830                 }
1831             }else if(h->cbp&15){
1832                 if(transform_bypass){
1833                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1834                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1835                     for(i=0; i<16; i+=di){
1836                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1837                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1838                         }
1839                     }
1840                 }else{
1841                     if(IS_8x8DCT(mb_type)){
1842                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1843                     }else{
1844                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1845                     }
1846                 }
1847             }
1848         }else{
1849             for(i=0; i<16; i++){
1850                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1851                     uint8_t * const ptr= dest_y + block_offset[i];
1852                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1853                 }
1854             }
1855         }
1856     }
1857 }
1858
1859 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift)
1860 {
1861     MpegEncContext * const s = &h->s;
1862     const int mb_x= s->mb_x;
1863     const int mb_y= s->mb_y;
1864     const int mb_xy= h->mb_xy;
1865     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1866     uint8_t  *dest_y, *dest_cb, *dest_cr;
1867     int linesize, uvlinesize /*dct_offset*/;
1868     int i, j;
1869     int *block_offset = &h->block_offset[0];
1870     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1871     /* is_h264 should always be true if SVQ3 is disabled. */
1872     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1873     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1874     const int block_h = 16 >> s->chroma_y_shift;
1875     const int chroma422 = CHROMA422;
1876
1877     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1878     dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1879     dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1880
1881     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1882     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1883
1884     h->list_counts[mb_xy]= h->list_count;
1885
1886     if (!simple && MB_FIELD) {
1887         linesize   = h->mb_linesize   = s->linesize * 2;
1888         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1889         block_offset = &h->block_offset[48];
1890         if(mb_y&1){ //FIXME move out of this function?
1891             dest_y -= s->linesize*15;
1892             dest_cb-= s->uvlinesize * (block_h - 1);
1893             dest_cr-= s->uvlinesize * (block_h - 1);
1894         }
1895         if(FRAME_MBAFF) {
1896             int list;
1897             for(list=0; list<h->list_count; list++){
1898                 if(!USES_LIST(mb_type, list))
1899                     continue;
1900                 if(IS_16X16(mb_type)){
1901                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1902                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1903                 }else{
1904                     for(i=0; i<16; i+=4){
1905                         int ref = h->ref_cache[list][scan8[i]];
1906                         if(ref >= 0)
1907                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1908                     }
1909                 }
1910             }
1911         }
1912     } else {
1913         linesize   = h->mb_linesize   = s->linesize;
1914         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1915 //        dct_offset = s->linesize * 16;
1916     }
1917
1918     if (!simple && IS_INTRA_PCM(mb_type)) {
1919         if (pixel_shift) {
1920             const int bit_depth = h->sps.bit_depth_luma;
1921             int j;
1922             GetBitContext gb;
1923             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1924
1925             for (i = 0; i < 16; i++) {
1926                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1927                 for (j = 0; j < 16; j++)
1928                     tmp_y[j] = get_bits(&gb, bit_depth);
1929             }
1930             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1931                 if (!h->sps.chroma_format_idc) {
1932                     for (i = 0; i < block_h; i++) {
1933                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1934                         for (j = 0; j < 8; j++) {
1935                             tmp_cb[j] = 1 << (bit_depth - 1);
1936                         }
1937                     }
1938                     for (i = 0; i < block_h; i++) {
1939                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1940                         for (j = 0; j < 8; j++) {
1941                             tmp_cr[j] = 1 << (bit_depth - 1);
1942                         }
1943                     }
1944                 } else {
1945                     for (i = 0; i < block_h; i++) {
1946                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1947                         for (j = 0; j < 8; j++)
1948                             tmp_cb[j] = get_bits(&gb, bit_depth);
1949                     }
1950                     for (i = 0; i < block_h; i++) {
1951                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1952                         for (j = 0; j < 8; j++)
1953                             tmp_cr[j] = get_bits(&gb, bit_depth);
1954                     }
1955                 }
1956             }
1957         } else {
1958             for (i=0; i<16; i++) {
1959                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1960             }
1961             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1962                 if (!h->sps.chroma_format_idc) {
1963                     for (i = 0; i < block_h; i++) {
1964                         memset(dest_cb + i*uvlinesize, 128, 8);
1965                         memset(dest_cr + i*uvlinesize, 128, 8);
1966                     }
1967                 } else {
1968                     for (i = 0; i < block_h; i++) {
1969                         memcpy(dest_cb + i*uvlinesize, h->mb + 128 + i*4,  8);
1970                         memcpy(dest_cr + i*uvlinesize, h->mb + 160 + i*4,  8);
1971                     }
1972                 }
1973             }
1974         }
1975     } else {
1976         if(IS_INTRA(mb_type)){
1977             if(h->deblocking_filter)
1978                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1979
1980             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1981                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1982                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1983             }
1984
1985             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1986
1987             if(h->deblocking_filter)
1988                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1989         }else if(is_h264){
1990             if (chroma422) {
1991                 hl_motion_422(h, dest_y, dest_cb, dest_cr,
1992                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1993                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1994                               h->h264dsp.weight_h264_pixels_tab,
1995                               h->h264dsp.biweight_h264_pixels_tab,
1996                               pixel_shift);
1997             } else {
1998                 hl_motion_420(h, dest_y, dest_cb, dest_cr,
1999                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2000                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2001                               h->h264dsp.weight_h264_pixels_tab,
2002                               h->h264dsp.biweight_h264_pixels_tab,
2003                               pixel_shift);
2004             }
2005         }
2006
2007         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
2008
2009         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
2010             uint8_t *dest[2] = {dest_cb, dest_cr};
2011             if(transform_bypass){
2012                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
2013                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
2014                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
2015                 }else{
2016                     idct_add = s->dsp.add_pixels4;
2017                     for(j=1; j<3; j++){
2018                         for(i=j*16; i<j*16+4; i++){
2019                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
2020                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
2021                         }
2022                         if (chroma422) {
2023                             for(i=j*16+4; i<j*16+8; i++){
2024                                 if(h->non_zero_count_cache[ scan8[i+4] ] || dctcoef_get(h->mb, pixel_shift, i*16))
2025                                     idct_add   (dest[j-1] + block_offset[i+4], h->mb + (i*16 << pixel_shift), uvlinesize);
2026                             }
2027                         }
2028                     }
2029                 }
2030             }else{
2031                 if(is_h264){
2032                     int qp[2];
2033                     if (chroma422) {
2034                         qp[0] = h->chroma_qp[0] + 3;
2035                         qp[1] = h->chroma_qp[1] + 3;
2036                     } else {
2037                         qp[0] = h->chroma_qp[0];
2038                         qp[1] = h->chroma_qp[1];
2039                     }
2040                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
2041                         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]);
2042                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
2043                         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]);
2044                     h->h264dsp.h264_idct_add8(dest, block_offset,
2045                                               h->mb, uvlinesize,
2046                                               h->non_zero_count_cache);
2047                 }else{
2048                     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]);
2049                     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]);
2050                     for(j=1; j<3; j++){
2051                         for(i=j*16; i<j*16+4; i++){
2052                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2053                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
2054                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
2055                             }
2056                         }
2057                     }
2058                 }
2059             }
2060         }
2061     }
2062     if(h->cbp || IS_INTRA(mb_type))
2063     {
2064         s->dsp.clear_blocks(h->mb);
2065         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2066     }
2067 }
2068
2069 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
2070     MpegEncContext * const s = &h->s;
2071     const int mb_x= s->mb_x;
2072     const int mb_y= s->mb_y;
2073     const int mb_xy= h->mb_xy;
2074     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2075     uint8_t  *dest[3];
2076     int linesize;
2077     int i, j, p;
2078     int *block_offset = &h->block_offset[0];
2079     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2080     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
2081
2082     for (p = 0; p < plane_count; p++)
2083     {
2084         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
2085         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
2086     }
2087
2088     h->list_counts[mb_xy]= h->list_count;
2089
2090     if (!simple && MB_FIELD) {
2091         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
2092         block_offset = &h->block_offset[48];
2093         if(mb_y&1) //FIXME move out of this function?
2094             for (p = 0; p < 3; p++)
2095                 dest[p] -= s->linesize*15;
2096         if(FRAME_MBAFF) {
2097             int list;
2098             for(list=0; list<h->list_count; list++){
2099                 if(!USES_LIST(mb_type, list))
2100                     continue;
2101                 if(IS_16X16(mb_type)){
2102                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2103                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2104                 }else{
2105                     for(i=0; i<16; i+=4){
2106                         int ref = h->ref_cache[list][scan8[i]];
2107                         if(ref >= 0)
2108                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2109                     }
2110                 }
2111             }
2112         }
2113     } else {
2114         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2115     }
2116
2117     if (!simple && IS_INTRA_PCM(mb_type)) {
2118         if (pixel_shift) {
2119             const int bit_depth = h->sps.bit_depth_luma;
2120             GetBitContext gb;
2121             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
2122
2123             for (p = 0; p < plane_count; p++) {
2124                 for (i = 0; i < 16; i++) {
2125                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2126                     for (j = 0; j < 16; j++)
2127                         tmp[j] = get_bits(&gb, bit_depth);
2128                 }
2129             }
2130         } else {
2131             for (p = 0; p < plane_count; p++) {
2132                 for (i = 0; i < 16; i++) {
2133                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2134                 }
2135             }
2136         }
2137     } else {
2138         if(IS_INTRA(mb_type)){
2139             if(h->deblocking_filter)
2140                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2141
2142             for (p = 0; p < plane_count; p++)
2143                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2144
2145             if(h->deblocking_filter)
2146                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2147         }else{
2148             hl_motion(h, dest[0], dest[1], dest[2],
2149                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2150                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2151                       h->h264dsp.weight_h264_pixels_tab,
2152                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 3);
2153         }
2154
2155         for (p = 0; p < plane_count; p++)
2156             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2157     }
2158     if(h->cbp || IS_INTRA(mb_type))
2159     {
2160         s->dsp.clear_blocks(h->mb);
2161         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2162     }
2163 }
2164
2165 /**
2166  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2167  */
2168 #define hl_decode_mb_simple(sh, bits) \
2169 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2170     hl_decode_mb_internal(h, 1, sh); \
2171 }
2172 hl_decode_mb_simple(0, 8);
2173 hl_decode_mb_simple(1, 16);
2174
2175 /**
2176  * Process a macroblock; this handles edge cases, such as interlacing.
2177  */
2178 static void av_noinline hl_decode_mb_complex(H264Context *h){
2179     hl_decode_mb_internal(h, 0, h->pixel_shift);
2180 }
2181
2182 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2183     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2184 }
2185
2186 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2187     hl_decode_mb_444_internal(h, 1, 0);
2188 }
2189
2190 void ff_h264_hl_decode_mb(H264Context *h){
2191     MpegEncContext * const s = &h->s;
2192     const int mb_xy= h->mb_xy;
2193     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2194     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2195
2196     if (CHROMA444) {
2197         if(is_complex || h->pixel_shift)
2198             hl_decode_mb_444_complex(h);
2199         else
2200             hl_decode_mb_444_simple(h);
2201     } else if (is_complex) {
2202         hl_decode_mb_complex(h);
2203     } else if (h->pixel_shift) {
2204         hl_decode_mb_simple_16(h);
2205     } else
2206         hl_decode_mb_simple_8(h);
2207 }
2208
2209 static int pred_weight_table(H264Context *h){
2210     MpegEncContext * const s = &h->s;
2211     int list, i;
2212     int luma_def, chroma_def;
2213
2214     h->use_weight= 0;
2215     h->use_weight_chroma= 0;
2216     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2217     if(h->sps.chroma_format_idc)
2218         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2219     luma_def = 1<<h->luma_log2_weight_denom;
2220     chroma_def = 1<<h->chroma_log2_weight_denom;
2221
2222     for(list=0; list<2; list++){
2223         h->luma_weight_flag[list]   = 0;
2224         h->chroma_weight_flag[list] = 0;
2225         for(i=0; i<h->ref_count[list]; i++){
2226             int luma_weight_flag, chroma_weight_flag;
2227
2228             luma_weight_flag= get_bits1(&s->gb);
2229             if(luma_weight_flag){
2230                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2231                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2232                 if(   h->luma_weight[i][list][0] != luma_def
2233                    || h->luma_weight[i][list][1] != 0) {
2234                     h->use_weight= 1;
2235                     h->luma_weight_flag[list]= 1;
2236                 }
2237             }else{
2238                 h->luma_weight[i][list][0]= luma_def;
2239                 h->luma_weight[i][list][1]= 0;
2240             }
2241
2242             if(h->sps.chroma_format_idc){
2243                 chroma_weight_flag= get_bits1(&s->gb);
2244                 if(chroma_weight_flag){
2245                     int j;
2246                     for(j=0; j<2; j++){
2247                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2248                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2249                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2250                            || h->chroma_weight[i][list][j][1] != 0) {
2251                             h->use_weight_chroma= 1;
2252                             h->chroma_weight_flag[list]= 1;
2253                         }
2254                     }
2255                 }else{
2256                     int j;
2257                     for(j=0; j<2; j++){
2258                         h->chroma_weight[i][list][j][0]= chroma_def;
2259                         h->chroma_weight[i][list][j][1]= 0;
2260                     }
2261                 }
2262             }
2263         }
2264         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2265     }
2266     h->use_weight= h->use_weight || h->use_weight_chroma;
2267     return 0;
2268 }
2269
2270 /**
2271  * Initialize implicit_weight table.
2272  * @param field  0/1 initialize the weight for interlaced MBAFF
2273  *                -1 initializes the rest
2274  */
2275 static void implicit_weight_table(H264Context *h, int field){
2276     MpegEncContext * const s = &h->s;
2277     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2278
2279     for (i = 0; i < 2; i++) {
2280         h->luma_weight_flag[i]   = 0;
2281         h->chroma_weight_flag[i] = 0;
2282     }
2283
2284     if(field < 0){
2285         if (s->picture_structure == PICT_FRAME) {
2286             cur_poc = s->current_picture_ptr->poc;
2287         } else {
2288             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2289         }
2290     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2291        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2292         h->use_weight= 0;
2293         h->use_weight_chroma= 0;
2294         return;
2295     }
2296         ref_start= 0;
2297         ref_count0= h->ref_count[0];
2298         ref_count1= h->ref_count[1];
2299     }else{
2300         cur_poc = s->current_picture_ptr->field_poc[field];
2301         ref_start= 16;
2302         ref_count0= 16+2*h->ref_count[0];
2303         ref_count1= 16+2*h->ref_count[1];
2304     }
2305
2306     h->use_weight= 2;
2307     h->use_weight_chroma= 2;
2308     h->luma_log2_weight_denom= 5;
2309     h->chroma_log2_weight_denom= 5;
2310
2311     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2312         int poc0 = h->ref_list[0][ref0].poc;
2313         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2314             int w = 32;
2315             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2316                 int poc1 = h->ref_list[1][ref1].poc;
2317                 int td = av_clip(poc1 - poc0, -128, 127);
2318                 if(td){
2319                     int tb = av_clip(cur_poc - poc0, -128, 127);
2320                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2321                     int dist_scale_factor = (tb*tx + 32) >> 8;
2322                     if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2323                         w = 64 - dist_scale_factor;
2324                 }
2325             }
2326             if(field<0){
2327                 h->implicit_weight[ref0][ref1][0]=
2328                 h->implicit_weight[ref0][ref1][1]= w;
2329             }else{
2330                 h->implicit_weight[ref0][ref1][field]=w;
2331             }
2332         }
2333     }
2334 }
2335
2336 /**
2337  * instantaneous decoder refresh.
2338  */
2339 static void idr(H264Context *h){
2340     ff_h264_remove_all_refs(h);
2341     h->prev_frame_num= 0;
2342     h->prev_frame_num_offset= 0;
2343     h->prev_poc_msb=
2344     h->prev_poc_lsb= 0;
2345 }
2346
2347 /* forget old pics after a seek */
2348 static void flush_dpb(AVCodecContext *avctx){
2349     H264Context *h= avctx->priv_data;
2350     int i;
2351     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
2352         if(h->delayed_pic[i])
2353             h->delayed_pic[i]->f.reference = 0;
2354         h->delayed_pic[i]= NULL;
2355     }
2356     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2357     h->prev_interlaced_frame = 1;
2358     idr(h);
2359     if(h->s.current_picture_ptr)
2360         h->s.current_picture_ptr->f.reference = 0;
2361     h->s.first_field= 0;
2362     ff_h264_reset_sei(h);
2363     ff_mpeg_flush(avctx);
2364 }
2365
2366 static int init_poc(H264Context *h){
2367     MpegEncContext * const s = &h->s;
2368     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2369     int field_poc[2];
2370     Picture *cur = s->current_picture_ptr;
2371
2372     h->frame_num_offset= h->prev_frame_num_offset;
2373     if(h->frame_num < h->prev_frame_num)
2374         h->frame_num_offset += max_frame_num;
2375
2376     if(h->sps.poc_type==0){
2377         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2378
2379         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2380             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2381         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2382             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2383         else
2384             h->poc_msb = h->prev_poc_msb;
2385 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2386         field_poc[0] =
2387         field_poc[1] = h->poc_msb + h->poc_lsb;
2388         if(s->picture_structure == PICT_FRAME)
2389             field_poc[1] += h->delta_poc_bottom;
2390     }else if(h->sps.poc_type==1){
2391         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2392         int i;
2393
2394         if(h->sps.poc_cycle_length != 0)
2395             abs_frame_num = h->frame_num_offset + h->frame_num;
2396         else
2397             abs_frame_num = 0;
2398
2399         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2400             abs_frame_num--;
2401
2402         expected_delta_per_poc_cycle = 0;
2403         for(i=0; i < h->sps.poc_cycle_length; i++)
2404             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2405
2406         if(abs_frame_num > 0){
2407             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2408             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2409
2410             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2411             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2412                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2413         } else
2414             expectedpoc = 0;
2415
2416         if(h->nal_ref_idc == 0)
2417             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2418
2419         field_poc[0] = expectedpoc + h->delta_poc[0];
2420         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2421
2422         if(s->picture_structure == PICT_FRAME)
2423             field_poc[1] += h->delta_poc[1];
2424     }else{
2425         int poc= 2*(h->frame_num_offset + h->frame_num);
2426
2427         if(!h->nal_ref_idc)
2428             poc--;
2429
2430         field_poc[0]= poc;
2431         field_poc[1]= poc;
2432     }
2433
2434     if(s->picture_structure != PICT_BOTTOM_FIELD)
2435         s->current_picture_ptr->field_poc[0]= field_poc[0];
2436     if(s->picture_structure != PICT_TOP_FIELD)
2437         s->current_picture_ptr->field_poc[1]= field_poc[1];
2438     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2439
2440     return 0;
2441 }
2442
2443
2444 /**
2445  * initialize scan tables
2446  */
2447 static void init_scan_tables(H264Context *h){
2448     int i;
2449     for(i=0; i<16; i++){
2450 #define T(x) (x>>2) | ((x<<2) & 0xF)
2451         h->zigzag_scan[i] = T(zigzag_scan[i]);
2452         h-> field_scan[i] = T( field_scan[i]);
2453 #undef T
2454     }
2455     for(i=0; i<64; i++){
2456 #define T(x) (x>>3) | ((x&7)<<3)
2457         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2458         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2459         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2460         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2461 #undef T
2462     }
2463     if(h->sps.transform_bypass){ //FIXME same ugly
2464         h->zigzag_scan_q0          = zigzag_scan;
2465         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2466         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2467         h->field_scan_q0           = field_scan;
2468         h->field_scan8x8_q0        = field_scan8x8;
2469         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2470     }else{
2471         h->zigzag_scan_q0          = h->zigzag_scan;
2472         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2473         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2474         h->field_scan_q0           = h->field_scan;
2475         h->field_scan8x8_q0        = h->field_scan8x8;
2476         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2477     }
2478 }
2479
2480 static int field_end(H264Context *h, int in_setup){
2481     MpegEncContext * const s = &h->s;
2482     AVCodecContext * const avctx= s->avctx;
2483     int err = 0;
2484     s->mb_y= 0;
2485
2486     if (!in_setup && !s->dropable)
2487         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2488                                  s->picture_structure==PICT_BOTTOM_FIELD);
2489
2490     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2491         ff_vdpau_h264_set_reference_frames(s);
2492
2493     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2494         if(!s->dropable) {
2495             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2496             h->prev_poc_msb= h->poc_msb;
2497             h->prev_poc_lsb= h->poc_lsb;
2498         }
2499         h->prev_frame_num_offset= h->frame_num_offset;
2500         h->prev_frame_num= h->frame_num;
2501         h->outputed_poc = h->next_outputed_poc;
2502     }
2503
2504     if (avctx->hwaccel) {
2505         if (avctx->hwaccel->end_frame(avctx) < 0)
2506             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2507     }
2508
2509     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2510         ff_vdpau_h264_picture_complete(s);
2511
2512     /*
2513      * FIXME: Error handling code does not seem to support interlaced
2514      * when slices span multiple rows
2515      * The ff_er_add_slice calls don't work right for bottom
2516      * fields; they cause massive erroneous error concealing
2517      * Error marking covers both fields (top and bottom).
2518      * This causes a mismatched s->error_count
2519      * and a bad error table. Further, the error count goes to
2520      * INT_MAX when called for bottom field, because mb_y is
2521      * past end by one (callers fault) and resync_mb_y != 0
2522      * causes problems for the first MB line, too.
2523      */
2524     if (!FIELD_PICTURE)
2525         ff_er_frame_end(s);
2526
2527     MPV_frame_end(s);
2528
2529     h->current_slice=0;
2530
2531     return err;
2532 }
2533
2534 /**
2535  * Replicate H264 "master" context to thread contexts.
2536  */
2537 static void clone_slice(H264Context *dst, H264Context *src)
2538 {
2539     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2540     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2541     dst->s.current_picture      = src->s.current_picture;
2542     dst->s.linesize             = src->s.linesize;
2543     dst->s.uvlinesize           = src->s.uvlinesize;
2544     dst->s.first_field          = src->s.first_field;
2545
2546     dst->prev_poc_msb           = src->prev_poc_msb;
2547     dst->prev_poc_lsb           = src->prev_poc_lsb;
2548     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2549     dst->prev_frame_num         = src->prev_frame_num;
2550     dst->short_ref_count        = src->short_ref_count;
2551
2552     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2553     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2554     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2555     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2556
2557     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2558     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2559 }
2560
2561 /**
2562  * computes profile from profile_idc and constraint_set?_flags
2563  *
2564  * @param sps SPS
2565  *
2566  * @return profile as defined by FF_PROFILE_H264_*
2567  */
2568 int ff_h264_get_profile(SPS *sps)
2569 {
2570     int profile = sps->profile_idc;
2571
2572     switch(sps->profile_idc) {
2573     case FF_PROFILE_H264_BASELINE:
2574         // constraint_set1_flag set to 1
2575         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2576         break;
2577     case FF_PROFILE_H264_HIGH_10:
2578     case FF_PROFILE_H264_HIGH_422:
2579     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2580         // constraint_set3_flag set to 1
2581         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2582         break;
2583     }
2584
2585     return profile;
2586 }
2587
2588 /**
2589  * decodes a slice header.
2590  * This will also call MPV_common_init() and frame_start() as needed.
2591  *
2592  * @param h h264context
2593  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2594  *
2595  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2596  */
2597 static int decode_slice_header(H264Context *h, H264Context *h0){
2598     MpegEncContext * const s = &h->s;
2599     MpegEncContext * const s0 = &h0->s;
2600     unsigned int first_mb_in_slice;
2601     unsigned int pps_id;
2602     int num_ref_idx_active_override_flag;
2603     unsigned int slice_type, tmp, i, j;
2604     int default_ref_list_done = 0;
2605     int last_pic_structure;
2606
2607     s->dropable= h->nal_ref_idc == 0;
2608
2609     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2610     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2611         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2612         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2613     }else{
2614         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2615         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2616     }
2617
2618     first_mb_in_slice= get_ue_golomb(&s->gb);
2619
2620     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2621         if(h0->current_slice && FIELD_PICTURE){
2622             field_end(h, 1);
2623         }
2624
2625         h0->current_slice = 0;
2626         if (!s0->first_field)
2627             s->current_picture_ptr= NULL;
2628     }
2629
2630     slice_type= get_ue_golomb_31(&s->gb);
2631     if(slice_type > 9){
2632         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);
2633         return -1;
2634     }
2635     if(slice_type > 4){
2636         slice_type -= 5;
2637         h->slice_type_fixed=1;
2638     }else
2639         h->slice_type_fixed=0;
2640
2641     slice_type= golomb_to_pict_type[ slice_type ];
2642     if (slice_type == AV_PICTURE_TYPE_I
2643         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2644         default_ref_list_done = 1;
2645     }
2646     h->slice_type= slice_type;
2647     h->slice_type_nos= slice_type & 3;
2648
2649     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2650
2651     pps_id= get_ue_golomb(&s->gb);
2652     if(pps_id>=MAX_PPS_COUNT){
2653         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2654         return -1;
2655     }
2656     if(!h0->pps_buffers[pps_id]) {
2657         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2658         return -1;
2659     }
2660     h->pps= *h0->pps_buffers[pps_id];
2661
2662     if(!h0->sps_buffers[h->pps.sps_id]) {
2663         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2664         return -1;
2665     }
2666     h->sps = *h0->sps_buffers[h->pps.sps_id];
2667
2668     s->avctx->profile = ff_h264_get_profile(&h->sps);
2669     s->avctx->level   = h->sps.level_idc;
2670     s->avctx->refs    = h->sps.ref_frame_count;
2671
2672     if(h == h0 && h->dequant_coeff_pps != pps_id){
2673         h->dequant_coeff_pps = pps_id;
2674         init_dequant_tables(h);
2675     }
2676
2677     s->mb_width= h->sps.mb_width;
2678     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2679
2680     h->b_stride=  s->mb_width*4;
2681
2682     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2683
2684     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2685     if(h->sps.frame_mbs_only_flag)
2686         s->height= 16*s->mb_height - (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2687     else
2688         s->height= 16*s->mb_height - (2<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2689
2690     if (s->context_initialized
2691         && (   s->width != s->avctx->width || s->height != s->avctx->height
2692             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2693         if(h != h0) {
2694             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2695             return -1;   // width / height changed during parallelized decoding
2696         }
2697         free_tables(h, 0);
2698         flush_dpb(s->avctx);
2699         MPV_common_end(s);
2700     }
2701     if (!s->context_initialized) {
2702         if (h != h0) {
2703             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2704             return -1;
2705         }
2706
2707         avcodec_set_dimensions(s->avctx, s->width, s->height);
2708         s->avctx->sample_aspect_ratio= h->sps.sar;
2709         av_assert0(s->avctx->sample_aspect_ratio.den);
2710
2711         if(h->sps.video_signal_type_present_flag){
2712             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2713             if(h->sps.colour_description_present_flag){
2714                 s->avctx->color_primaries = h->sps.color_primaries;
2715                 s->avctx->color_trc       = h->sps.color_trc;
2716                 s->avctx->colorspace      = h->sps.colorspace;
2717             }
2718         }
2719
2720         if(h->sps.timing_info_present_flag){
2721             int64_t den= h->sps.time_scale;
2722             if(h->x264_build < 44U)
2723                 den *= 2;
2724             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2725                       h->sps.num_units_in_tick, den, 1<<30);
2726         }
2727
2728         switch (h->sps.bit_depth_luma) {
2729             case 9 :
2730                 if (CHROMA444) {
2731                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2732                         s->avctx->pix_fmt = PIX_FMT_GBRP9;
2733                     } else
2734                         s->avctx->pix_fmt = PIX_FMT_YUV444P9;
2735                 } else if (CHROMA422)
2736                     s->avctx->pix_fmt = PIX_FMT_YUV422P9;
2737                 else
2738                     s->avctx->pix_fmt = PIX_FMT_YUV420P9;
2739                 break;
2740             case 10 :
2741                 if (CHROMA444) {
2742                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2743                         s->avctx->pix_fmt = PIX_FMT_GBRP10;
2744                     } else
2745                         s->avctx->pix_fmt = PIX_FMT_YUV444P10;
2746                 } else if (CHROMA422)
2747                     s->avctx->pix_fmt = PIX_FMT_YUV422P10;
2748                 else
2749                     s->avctx->pix_fmt = PIX_FMT_YUV420P10;
2750                 break;
2751             default:
2752                 if (CHROMA444){
2753                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2754                         s->avctx->pix_fmt = PIX_FMT_GBRP;
2755                     } else
2756                         s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2757                 } else if (CHROMA422) {
2758                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P : PIX_FMT_YUV422P;
2759                 }else{
2760                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2761                                                              s->avctx->codec->pix_fmts ?
2762                                                              s->avctx->codec->pix_fmts :
2763                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2764                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2765                                                              ff_hwaccel_pixfmt_list_420);
2766                 }
2767         }
2768
2769         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2770
2771         if (MPV_common_init(s) < 0) {
2772             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2773             return -1;
2774         }
2775         s->first_field = 0;
2776         h->prev_interlaced_frame = 1;
2777
2778         init_scan_tables(h);
2779         if (ff_h264_alloc_tables(h) < 0) {
2780             av_log(h->s.avctx, AV_LOG_ERROR, "Could not allocate memory for h264\n");
2781             return AVERROR(ENOMEM);
2782         }
2783
2784         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2785             if (context_init(h) < 0) {
2786                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2787                 return -1;
2788             }
2789         } else {
2790             for(i = 1; i < s->avctx->thread_count; i++) {
2791                 H264Context *c;
2792                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2793                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2794                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2795                 c->h264dsp = h->h264dsp;
2796                 c->sps = h->sps;
2797                 c->pps = h->pps;
2798                 c->pixel_shift = h->pixel_shift;
2799                 init_scan_tables(c);
2800                 clone_tables(c, h, i);
2801             }
2802
2803             for(i = 0; i < s->avctx->thread_count; i++)
2804                 if (context_init(h->thread_context[i]) < 0) {
2805                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2806                     return -1;
2807                 }
2808         }
2809     }
2810
2811     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2812
2813     h->mb_mbaff = 0;
2814     h->mb_aff_frame = 0;
2815     last_pic_structure = s0->picture_structure;
2816     if(h->sps.frame_mbs_only_flag){
2817         s->picture_structure= PICT_FRAME;
2818     }else{
2819         if(get_bits1(&s->gb)) { //field_pic_flag
2820             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2821         } else {
2822             s->picture_structure= PICT_FRAME;
2823             h->mb_aff_frame = h->sps.mb_aff;
2824         }
2825     }
2826     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2827
2828     if(h0->current_slice == 0){
2829         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2830         if(h->frame_num != h->prev_frame_num) {
2831             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2832
2833             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2834
2835             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2836                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2837                 if (unwrap_prev_frame_num < 0)
2838                     unwrap_prev_frame_num += max_frame_num;
2839
2840                 h->prev_frame_num = unwrap_prev_frame_num;
2841             }
2842         }
2843
2844         while(h->frame_num !=  h->prev_frame_num &&
2845               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2846             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2847             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2848             if (ff_h264_frame_start(h) < 0)
2849                 return -1;
2850             h->prev_frame_num++;
2851             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2852             s->current_picture_ptr->frame_num= h->prev_frame_num;
2853             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2854             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2855             ff_generate_sliding_window_mmcos(h);
2856             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2857                 (s->avctx->err_recognition & AV_EF_EXPLODE))
2858                 return AVERROR_INVALIDDATA;
2859             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2860              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2861              * about there being no actual duplicates.
2862              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2863              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2864              * be fixed. */
2865             if (h->short_ref_count) {
2866                 if (prev) {
2867                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2868                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2869                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2870                     h->short_ref[0]->poc = prev->poc+2;
2871                 }
2872                 h->short_ref[0]->frame_num = h->prev_frame_num;
2873             }
2874         }
2875
2876         /* See if we have a decoded first field looking for a pair... */
2877         if (s0->first_field) {
2878             assert(s0->current_picture_ptr);
2879             assert(s0->current_picture_ptr->f.data[0]);
2880             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2881
2882             /* figure out if we have a complementary field pair */
2883             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2884                 /*
2885                  * Previous field is unmatched. Don't display it, but let it
2886                  * remain for reference if marked as such.
2887                  */
2888                 s0->current_picture_ptr = NULL;
2889                 s0->first_field = FIELD_PICTURE;
2890
2891             } else {
2892                 if (h->nal_ref_idc &&
2893                         s0->current_picture_ptr->f.reference &&
2894                         s0->current_picture_ptr->frame_num != h->frame_num) {
2895                     /*
2896                      * This and previous field were reference, but had
2897                      * different frame_nums. Consider this field first in
2898                      * pair. Throw away previous field except for reference
2899                      * purposes.
2900                      */
2901                     s0->first_field = 1;
2902                     s0->current_picture_ptr = NULL;
2903
2904                 } else {
2905                     /* Second field in complementary pair */
2906                     s0->first_field = 0;
2907                 }
2908             }
2909
2910         } else {
2911             /* Frame or first field in a potentially complementary pair */
2912             assert(!s0->current_picture_ptr);
2913             s0->first_field = FIELD_PICTURE;
2914         }
2915
2916         if(!FIELD_PICTURE || s0->first_field) {
2917             if (ff_h264_frame_start(h) < 0) {
2918                 s0->first_field = 0;
2919                 return -1;
2920             }
2921         } else {
2922             ff_release_unused_pictures(s, 0);
2923         }
2924     }
2925     if(h != h0)
2926         clone_slice(h, h0);
2927
2928     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2929
2930     assert(s->mb_num == s->mb_width * s->mb_height);
2931     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2932        first_mb_in_slice                    >= s->mb_num){
2933         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2934         return -1;
2935     }
2936     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2937     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2938     if (s->picture_structure == PICT_BOTTOM_FIELD)
2939         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2940     assert(s->mb_y < s->mb_height);
2941
2942     if(s->picture_structure==PICT_FRAME){
2943         h->curr_pic_num=   h->frame_num;
2944         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2945     }else{
2946         h->curr_pic_num= 2*h->frame_num + 1;
2947         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2948     }
2949
2950     if(h->nal_unit_type == NAL_IDR_SLICE){
2951         get_ue_golomb(&s->gb); /* idr_pic_id */
2952     }
2953
2954     if(h->sps.poc_type==0){
2955         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2956
2957         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2958             h->delta_poc_bottom= get_se_golomb(&s->gb);
2959         }
2960     }
2961
2962     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2963         h->delta_poc[0]= get_se_golomb(&s->gb);
2964
2965         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2966             h->delta_poc[1]= get_se_golomb(&s->gb);
2967     }
2968
2969     init_poc(h);
2970
2971     if(h->pps.redundant_pic_cnt_present){
2972         h->redundant_pic_count= get_ue_golomb(&s->gb);
2973     }
2974
2975     //set defaults, might be overridden a few lines later
2976     h->ref_count[0]= h->pps.ref_count[0];
2977     h->ref_count[1]= h->pps.ref_count[1];
2978
2979     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2980         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2981             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2982         }
2983         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2984
2985         if(num_ref_idx_active_override_flag){
2986             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2987             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2988                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2989
2990             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2991                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2992                 h->ref_count[0]= h->ref_count[1]= 1;
2993                 return -1;
2994             }
2995         }
2996         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2997             h->list_count= 2;
2998         else
2999             h->list_count= 1;
3000     }else
3001         h->list_count= 0;
3002
3003     if(!default_ref_list_done){
3004         ff_h264_fill_default_ref_list(h);
3005     }
3006
3007     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0) {
3008         h->ref_count[1]= h->ref_count[0]= 0;
3009         return -1;
3010     }
3011
3012     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
3013         s->last_picture_ptr= &h->ref_list[0][0];
3014         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
3015     }
3016     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
3017         s->next_picture_ptr= &h->ref_list[1][0];
3018         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
3019     }
3020
3021     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
3022        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
3023         pred_weight_table(h);
3024     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
3025         implicit_weight_table(h, -1);
3026     }else {
3027         h->use_weight = 0;
3028         for (i = 0; i < 2; i++) {
3029             h->luma_weight_flag[i]   = 0;
3030             h->chroma_weight_flag[i] = 0;
3031         }
3032     }
3033
3034     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
3035        (s->avctx->err_recognition & AV_EF_EXPLODE))
3036         return AVERROR_INVALIDDATA;
3037
3038     if(FRAME_MBAFF){
3039         ff_h264_fill_mbaff_ref_list(h);
3040
3041         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
3042             implicit_weight_table(h, 0);
3043             implicit_weight_table(h, 1);
3044         }
3045     }
3046
3047     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
3048         ff_h264_direct_dist_scale_factor(h);
3049     ff_h264_direct_ref_list_init(h);
3050
3051     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
3052         tmp = get_ue_golomb_31(&s->gb);
3053         if(tmp > 2){
3054             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3055             return -1;
3056         }
3057         h->cabac_init_idc= tmp;
3058     }
3059
3060     h->last_qscale_diff = 0;
3061     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3062     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
3063         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3064         return -1;
3065     }
3066     s->qscale= tmp;
3067     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3068     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3069     //FIXME qscale / qp ... stuff
3070     if(h->slice_type == AV_PICTURE_TYPE_SP){
3071         get_bits1(&s->gb); /* sp_for_switch_flag */
3072     }
3073     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
3074         get_se_golomb(&s->gb); /* slice_qs_delta */
3075     }
3076
3077     h->deblocking_filter = 1;
3078     h->slice_alpha_c0_offset = 52;
3079     h->slice_beta_offset = 52;
3080     if( h->pps.deblocking_filter_parameters_present ) {
3081         tmp= get_ue_golomb_31(&s->gb);
3082         if(tmp > 2){
3083             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
3084             return -1;
3085         }
3086         h->deblocking_filter= tmp;
3087         if(h->deblocking_filter < 2)
3088             h->deblocking_filter^= 1; // 1<->0
3089
3090         if( h->deblocking_filter ) {
3091             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
3092             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
3093             if(   h->slice_alpha_c0_offset > 104U
3094                || h->slice_beta_offset     > 104U){
3095                 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);
3096                 return -1;
3097             }
3098         }
3099     }
3100
3101     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
3102        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
3103        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
3104        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
3105         h->deblocking_filter= 0;
3106
3107     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
3108         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
3109             /* Cheat slightly for speed:
3110                Do not bother to deblock across slices. */
3111             h->deblocking_filter = 2;
3112         } else {
3113             h0->max_contexts = 1;
3114             if(!h0->single_decode_warning) {
3115                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3116                 h0->single_decode_warning = 1;
3117             }
3118             if (h != h0) {
3119                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
3120                 return 1;
3121             }
3122         }
3123     }
3124     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
3125                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
3126                  + 6 * (h->sps.bit_depth_luma - 8);
3127
3128 #if 0 //FMO
3129     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
3130         slice_group_change_cycle= get_bits(&s->gb, ?);
3131 #endif
3132
3133     h0->last_slice_type = slice_type;
3134     h->slice_num = ++h0->current_slice;
3135     if(h->slice_num >= MAX_SLICES){
3136         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
3137     }
3138
3139     for(j=0; j<2; j++){
3140         int id_list[16];
3141         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
3142         for(i=0; i<16; i++){
3143             id_list[i]= 60;
3144             if (h->ref_list[j][i].f.data[0]) {
3145                 int k;
3146                 uint8_t *base = h->ref_list[j][i].f.base[0];
3147                 for(k=0; k<h->short_ref_count; k++)
3148                     if (h->short_ref[k]->f.base[0] == base) {
3149                         id_list[i]= k;
3150                         break;
3151                     }
3152                 for(k=0; k<h->long_ref_count; k++)
3153                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3154                         id_list[i]= h->short_ref_count + k;
3155                         break;
3156                     }
3157             }
3158         }
3159
3160         ref2frm[0]=
3161         ref2frm[1]= -1;
3162         for(i=0; i<16; i++)
3163             ref2frm[i+2]= 4*id_list[i]
3164                           + (h->ref_list[j][i].f.reference & 3);
3165         ref2frm[18+0]=
3166         ref2frm[18+1]= -1;
3167         for(i=16; i<48; i++)
3168             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3169                           + (h->ref_list[j][i].f.reference & 3);
3170     }
3171
3172     //FIXME: fix draw_edges+PAFF+frame threads
3173     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3174     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3175
3176     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3177         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",
3178                h->slice_num,
3179                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3180                first_mb_in_slice,
3181                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3182                pps_id, h->frame_num,
3183                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3184                h->ref_count[0], h->ref_count[1],
3185                s->qscale,
3186                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3187                h->use_weight,
3188                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3189                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3190                );
3191     }
3192
3193     return 0;
3194 }
3195
3196 int ff_h264_get_slice_type(const H264Context *h)
3197 {
3198     switch (h->slice_type) {
3199     case AV_PICTURE_TYPE_P:  return 0;
3200     case AV_PICTURE_TYPE_B:  return 1;
3201     case AV_PICTURE_TYPE_I:  return 2;
3202     case AV_PICTURE_TYPE_SP: return 3;
3203     case AV_PICTURE_TYPE_SI: return 4;
3204     default:         return -1;
3205     }
3206 }
3207
3208 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3209                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3210 {
3211     int b_stride = h->b_stride;
3212     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3213     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3214     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3215         if(USES_LIST(top_type, list)){
3216             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3217             const int b8_xy= 4*top_xy + 2;
3218             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3219             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3220             ref_cache[0 - 1*8]=
3221             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3222             ref_cache[2 - 1*8]=
3223             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3224         }else{
3225             AV_ZERO128(mv_dst - 1*8);
3226             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3227         }
3228
3229         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3230             if(USES_LIST(left_type[LTOP], list)){
3231                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3232                 const int b8_xy= 4*left_xy[LTOP] + 1;
3233                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3234                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3235                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3236                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3237                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3238                 ref_cache[-1 +  0]=
3239                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3240                 ref_cache[-1 + 16]=
3241                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3242             }else{
3243                 AV_ZERO32(mv_dst - 1 + 0);
3244                 AV_ZERO32(mv_dst - 1 + 8);
3245                 AV_ZERO32(mv_dst - 1 +16);
3246                 AV_ZERO32(mv_dst - 1 +24);
3247                 ref_cache[-1 +  0]=
3248                 ref_cache[-1 +  8]=
3249                 ref_cache[-1 + 16]=
3250                 ref_cache[-1 + 24]= LIST_NOT_USED;
3251             }
3252         }
3253     }
3254
3255     if(!USES_LIST(mb_type, list)){
3256         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3257         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3258         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3259         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3260         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3261         return;
3262     }
3263
3264     {
3265         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3266         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3267         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3268         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3269         AV_WN32A(&ref_cache[0*8], ref01);
3270         AV_WN32A(&ref_cache[1*8], ref01);
3271         AV_WN32A(&ref_cache[2*8], ref23);
3272         AV_WN32A(&ref_cache[3*8], ref23);
3273     }
3274
3275     {
3276         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3277         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3278         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3279         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3280         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3281     }
3282 }
3283
3284 /**
3285  *
3286  * @return non zero if the loop filter can be skiped
3287  */
3288 static int fill_filter_caches(H264Context *h, int mb_type){
3289     MpegEncContext * const s = &h->s;
3290     const int mb_xy= h->mb_xy;
3291     int top_xy, left_xy[LEFT_MBS];
3292     int top_type, left_type[LEFT_MBS];
3293     uint8_t *nnz;
3294     uint8_t *nnz_cache;
3295
3296     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3297
3298     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3299      * stuff, I can't imagine that these complex rules are worth it. */
3300
3301     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3302     if(FRAME_MBAFF){
3303         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3304         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3305         if(s->mb_y&1){
3306             if (left_mb_field_flag != curr_mb_field_flag) {
3307                 left_xy[LTOP] -= s->mb_stride;
3308             }
3309         }else{
3310             if(curr_mb_field_flag){
3311                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3312             }
3313             if (left_mb_field_flag != curr_mb_field_flag) {
3314                 left_xy[LBOT] += s->mb_stride;
3315             }
3316         }
3317     }
3318
3319     h->top_mb_xy = top_xy;
3320     h->left_mb_xy[LTOP] = left_xy[LTOP];
3321     h->left_mb_xy[LBOT] = left_xy[LBOT];
3322     {
3323         //for sufficiently low qp, filtering wouldn't do anything
3324         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3325         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3326         int qp = s->current_picture.f.qscale_table[mb_xy];
3327         if(qp <= qp_thresh
3328            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3329            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3330             if(!FRAME_MBAFF)
3331                 return 1;
3332             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3333                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3334                 return 1;
3335         }
3336     }
3337
3338     top_type        = s->current_picture.f.mb_type[top_xy];
3339     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3340     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3341     if(h->deblocking_filter == 2){
3342         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3343         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3344     }else{
3345         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3346         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3347     }
3348     h->top_type       = top_type;
3349     h->left_type[LTOP]= left_type[LTOP];
3350     h->left_type[LBOT]= left_type[LBOT];
3351
3352     if(IS_INTRA(mb_type))
3353         return 0;
3354
3355     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3356     if(h->list_count == 2)
3357         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3358
3359     nnz = h->non_zero_count[mb_xy];
3360     nnz_cache = h->non_zero_count_cache;
3361     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3362     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3363     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3364     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3365     h->cbp= h->cbp_table[mb_xy];
3366
3367     if(top_type){
3368         nnz = h->non_zero_count[top_xy];
3369         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3370     }
3371
3372     if(left_type[LTOP]){
3373         nnz = h->non_zero_count[left_xy[LTOP]];
3374         nnz_cache[3+8*1]= nnz[3+0*4];
3375         nnz_cache[3+8*2]= nnz[3+1*4];
3376         nnz_cache[3+8*3]= nnz[3+2*4];
3377         nnz_cache[3+8*4]= nnz[3+3*4];
3378     }
3379
3380     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3381     if(!CABAC && h->pps.transform_8x8_mode){
3382         if(IS_8x8DCT(top_type)){
3383             nnz_cache[4+8*0]=
3384             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3385             nnz_cache[6+8*0]=
3386             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3387         }
3388         if(IS_8x8DCT(left_type[LTOP])){
3389             nnz_cache[3+8*1]=
3390             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3391         }
3392         if(IS_8x8DCT(left_type[LBOT])){
3393             nnz_cache[3+8*3]=
3394             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3395         }
3396
3397         if(IS_8x8DCT(mb_type)){
3398             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3399             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3400
3401             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3402             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3403
3404             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3405             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3406
3407             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3408             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3409         }
3410     }
3411
3412     return 0;
3413 }
3414
3415 static void loop_filter(H264Context *h, int start_x, int end_x){
3416     MpegEncContext * const s = &h->s;
3417     uint8_t  *dest_y, *dest_cb, *dest_cr;
3418     int linesize, uvlinesize, mb_x, mb_y;
3419     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3420     const int old_slice_type= h->slice_type;
3421     const int pixel_shift = h->pixel_shift;
3422     const int block_h = 16 >> s->chroma_y_shift;
3423
3424     if(h->deblocking_filter) {
3425         for(mb_x= start_x; mb_x<end_x; mb_x++){
3426             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3427                 int mb_xy, mb_type;
3428                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3429                 h->slice_num= h->slice_table[mb_xy];
3430                 mb_type = s->current_picture.f.mb_type[mb_xy];
3431                 h->list_count= h->list_counts[mb_xy];
3432
3433                 if(FRAME_MBAFF)
3434                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3435
3436                 s->mb_x= mb_x;
3437                 s->mb_y= mb_y;
3438                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3439                 dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3440                 dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3441                     //FIXME simplify above
3442
3443                 if (MB_FIELD) {
3444                     linesize   = h->mb_linesize   = s->linesize * 2;
3445                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3446                     if(mb_y&1){ //FIXME move out of this function?
3447                         dest_y -= s->linesize*15;
3448                         dest_cb-= s->uvlinesize * (block_h - 1);
3449                         dest_cr-= s->uvlinesize * (block_h - 1);
3450                     }
3451                 } else {
3452                     linesize   = h->mb_linesize   = s->linesize;
3453                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3454                 }
3455                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
3456                 if(fill_filter_caches(h, mb_type))
3457                     continue;
3458                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3459                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3460
3461                 if (FRAME_MBAFF) {
3462                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3463                 } else {
3464                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3465                 }
3466             }
3467         }
3468     }
3469     h->slice_type= old_slice_type;
3470     s->mb_x= end_x;
3471     s->mb_y= end_mb_y - FRAME_MBAFF;
3472     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3473     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3474 }
3475
3476 static void predict_field_decoding_flag(H264Context *h){
3477     MpegEncContext * const s = &h->s;
3478     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3479     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3480                 ? s->current_picture.f.mb_type[mb_xy - 1]
3481                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3482                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3483                 : 0;
3484     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3485 }
3486
3487 /**
3488  * Draw edges and report progress for the last MB row.
3489  */
3490 static void decode_finish_row(H264Context *h){
3491     MpegEncContext * const s = &h->s;
3492     int top = 16*(s->mb_y >> FIELD_PICTURE);
3493     int height = 16 << FRAME_MBAFF;
3494     int deblock_border = (16 + 4) << FRAME_MBAFF;
3495     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3496
3497     if (h->deblocking_filter) {
3498         if((top + height) >= pic_height)
3499             height += deblock_border;
3500
3501         top -= deblock_border;
3502     }
3503
3504     if (top >= pic_height || (top + height) < h->emu_edge_height)
3505         return;
3506
3507     height = FFMIN(height, pic_height - top);
3508     if (top < h->emu_edge_height) {
3509         height = top+height;
3510         top = 0;
3511     }
3512
3513     ff_draw_horiz_band(s, top, height);
3514
3515     if (s->dropable) return;
3516
3517     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3518                              s->picture_structure==PICT_BOTTOM_FIELD);
3519 }
3520
3521 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3522     H264Context *h = *(void**)arg;
3523     MpegEncContext * const s = &h->s;
3524     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3525     int lf_x_start = s->mb_x;
3526
3527     s->mb_skip_run= -1;
3528
3529     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3530                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3531
3532     if( h->pps.cabac ) {
3533         /* realign */
3534         align_get_bits( &s->gb );
3535
3536         /* init cabac */
3537         ff_init_cabac_states( &h->cabac);
3538         ff_init_cabac_decoder( &h->cabac,
3539                                s->gb.buffer + get_bits_count(&s->gb)/8,
3540                                (get_bits_left(&s->gb) + 7)/8);
3541
3542         ff_h264_init_cabac_states(h);
3543
3544         for(;;){
3545 //START_TIMER
3546             int ret = ff_h264_decode_mb_cabac(h);
3547             int eos;
3548 //STOP_TIMER("decode_mb_cabac")
3549
3550             if(ret>=0) ff_h264_hl_decode_mb(h);
3551
3552             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3553                 s->mb_y++;
3554
3555                 ret = ff_h264_decode_mb_cabac(h);
3556
3557                 if(ret>=0) ff_h264_hl_decode_mb(h);
3558                 s->mb_y--;
3559             }
3560             eos = get_cabac_terminate( &h->cabac );
3561
3562             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3563                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3564                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3565                 return 0;
3566             }
3567             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3568                 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);
3569                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3570                 return -1;
3571             }
3572
3573             if( ++s->mb_x >= s->mb_width ) {
3574                 loop_filter(h, lf_x_start, s->mb_x);
3575                 s->mb_x = lf_x_start = 0;
3576                 decode_finish_row(h);
3577                 ++s->mb_y;
3578                 if(FIELD_OR_MBAFF_PICTURE) {
3579                     ++s->mb_y;
3580                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3581                         predict_field_decoding_flag(h);
3582                 }
3583             }
3584
3585             if( eos || s->mb_y >= s->mb_height ) {
3586                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3587                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3588                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3589                 return 0;
3590             }
3591         }
3592
3593     } else {
3594         for(;;){
3595             int ret = ff_h264_decode_mb_cavlc(h);
3596
3597             if(ret>=0) ff_h264_hl_decode_mb(h);
3598
3599             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3600                 s->mb_y++;
3601                 ret = ff_h264_decode_mb_cavlc(h);
3602
3603                 if(ret>=0) ff_h264_hl_decode_mb(h);
3604                 s->mb_y--;
3605             }
3606
3607             if(ret<0){
3608                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3609                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3610                 return -1;
3611             }
3612
3613             if(++s->mb_x >= s->mb_width){
3614                 loop_filter(h, lf_x_start, s->mb_x);
3615                 s->mb_x = lf_x_start = 0;
3616                 decode_finish_row(h);
3617                 ++s->mb_y;
3618                 if(FIELD_OR_MBAFF_PICTURE) {
3619                     ++s->mb_y;
3620                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3621                         predict_field_decoding_flag(h);
3622                 }
3623                 if(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
3626                     if(get_bits_count(&s->gb) == s->gb.size_in_bits ) {
3627                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3628
3629                         return 0;
3630                     }else{
3631                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3632
3633                         return -1;
3634                     }
3635                 }
3636             }
3637
3638             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
3639                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3640                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
3641                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3642                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3643
3644                     return 0;
3645                 }else{
3646                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3647
3648                     return -1;
3649                 }
3650             }
3651         }
3652     }
3653 }
3654
3655 /**
3656  * Call decode_slice() for each context.
3657  *
3658  * @param h h264 master context
3659  * @param context_count number of contexts to execute
3660  */
3661 static int execute_decode_slices(H264Context *h, int context_count){
3662     MpegEncContext * const s = &h->s;
3663     AVCodecContext * const avctx= s->avctx;
3664     H264Context *hx;
3665     int i;
3666
3667     if (s->avctx->hwaccel || s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3668         return 0;
3669     if(context_count == 1) {
3670         return decode_slice(avctx, &h);
3671     } else {
3672         for(i = 1; i < context_count; i++) {
3673             hx = h->thread_context[i];
3674             hx->s.error_recognition = avctx->error_recognition;
3675             hx->s.error_count = 0;
3676         }
3677
3678         avctx->execute(avctx, (void *)decode_slice,
3679                        h->thread_context, NULL, context_count, sizeof(void*));
3680
3681         /* pull back stuff from slices to master context */
3682         hx = h->thread_context[context_count - 1];
3683         s->mb_x = hx->s.mb_x;
3684         s->mb_y = hx->s.mb_y;
3685         s->dropable = hx->s.dropable;
3686         s->picture_structure = hx->s.picture_structure;
3687         for(i = 1; i < context_count; i++)
3688             h->s.error_count += h->thread_context[i]->s.error_count;
3689     }
3690
3691     return 0;
3692 }
3693
3694
3695 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
3696     MpegEncContext * const s = &h->s;
3697     AVCodecContext * const avctx= s->avctx;
3698     H264Context *hx; ///< thread context
3699     int buf_index;
3700     int context_count;
3701     int next_avc;
3702     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3703     int nals_needed=0; ///< number of NALs that need decoding before the next frame thread starts
3704     int nal_index;
3705
3706     h->max_contexts = (HAVE_THREADS && (s->avctx->active_thread_type&FF_THREAD_SLICE)) ? avctx->thread_count : 1;
3707     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
3708         h->current_slice = 0;
3709         if (!s->first_field)
3710             s->current_picture_ptr= NULL;
3711         ff_h264_reset_sei(h);
3712     }
3713
3714     for(;pass <= 1;pass++){
3715         buf_index = 0;
3716         context_count = 0;
3717         next_avc = h->is_avc ? 0 : buf_size;
3718         nal_index = 0;
3719     for(;;){
3720         int consumed;
3721         int dst_length;
3722         int bit_length;
3723         const uint8_t *ptr;
3724         int i, nalsize = 0;
3725         int err;
3726
3727         if(buf_index >= next_avc) {
3728             if(buf_index >= buf_size) break;
3729             nalsize = 0;
3730             for(i = 0; i < h->nal_length_size; i++)
3731                 nalsize = (nalsize << 8) | buf[buf_index++];
3732             if(nalsize <= 0 || nalsize > buf_size - buf_index){
3733                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
3734                 break;
3735             }
3736             next_avc= buf_index + nalsize;
3737         } else {
3738             // start code prefix search
3739             for(; buf_index + 3 < next_avc; buf_index++){
3740                 // This should always succeed in the first iteration.
3741                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
3742                     break;
3743             }
3744
3745             if(buf_index+3 >= buf_size) break;
3746
3747             buf_index+=3;
3748             if(buf_index >= next_avc) continue;
3749         }
3750
3751         hx = h->thread_context[context_count];
3752
3753         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
3754         if (ptr==NULL || dst_length < 0){
3755             return -1;
3756         }
3757         i= buf_index + consumed;
3758         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
3759            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
3760             s->workaround_bugs |= FF_BUG_TRUNCATED;
3761
3762         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
3763         while(ptr[dst_length - 1] == 0 && dst_length > 0)
3764             dst_length--;
3765         }
3766         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
3767
3768         if(s->avctx->debug&FF_DEBUG_STARTCODE){
3769             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);
3770         }
3771
3772         if (h->is_avc && (nalsize != consumed) && nalsize){
3773             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
3774         }
3775
3776         buf_index += consumed;
3777         nal_index++;
3778
3779         if(pass == 0) {
3780             // packets can sometimes contain multiple PPS/SPS
3781             // e.g. two PAFF field pictures in one packet, or a demuxer which splits NALs strangely
3782             // if so, when frame threading we can't start the next thread until we've read all of them
3783             switch (hx->nal_unit_type) {
3784                 case NAL_SPS:
3785                 case NAL_PPS:
3786                     nals_needed = nal_index;
3787                     break;
3788                 case NAL_IDR_SLICE:
3789                 case NAL_SLICE:
3790                     init_get_bits(&hx->s.gb, ptr, bit_length);
3791                     if (!get_ue_golomb(&hx->s.gb))
3792                         nals_needed = nal_index;
3793             }
3794             continue;
3795         }
3796
3797         //FIXME do not discard SEI id
3798         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
3799             continue;
3800
3801       again:
3802         err = 0;
3803         switch(hx->nal_unit_type){
3804         case NAL_IDR_SLICE:
3805             if (h->nal_unit_type != NAL_IDR_SLICE) {
3806                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
3807                 return -1;
3808             }
3809             idr(h); //FIXME ensure we don't loose some frames if there is reordering
3810         case NAL_SLICE:
3811             init_get_bits(&hx->s.gb, ptr, bit_length);
3812             hx->intra_gb_ptr=
3813             hx->inter_gb_ptr= &hx->s.gb;
3814             hx->s.data_partitioning = 0;
3815
3816             if((err = decode_slice_header(hx, h)))
3817                break;
3818
3819             s->current_picture_ptr->f.key_frame |=
3820                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
3821                     (h->sei_recovery_frame_cnt >= 0);
3822
3823             if (h->current_slice == 1) {
3824                 if(!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3825                     decode_postinit(h, nal_index >= nals_needed);
3826                 }
3827
3828                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3829                     return -1;
3830                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3831                     ff_vdpau_h264_picture_start(s);
3832             }
3833
3834             if(hx->redundant_pic_count==0
3835                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3836                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3837                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3838                && avctx->skip_frame < AVDISCARD_ALL){
3839                 if(avctx->hwaccel) {
3840                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
3841                         return -1;
3842                 }else
3843                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
3844                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
3845                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
3846                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
3847                 }else
3848                     context_count++;
3849             }
3850             break;
3851         case NAL_DPA:
3852             init_get_bits(&hx->s.gb, ptr, bit_length);
3853             hx->intra_gb_ptr=
3854             hx->inter_gb_ptr= NULL;
3855
3856             if ((err = decode_slice_header(hx, h)) < 0)
3857                 break;
3858
3859             hx->s.data_partitioning = 1;
3860
3861             break;
3862         case NAL_DPB:
3863             init_get_bits(&hx->intra_gb, ptr, bit_length);
3864             hx->intra_gb_ptr= &hx->intra_gb;
3865             break;
3866         case NAL_DPC:
3867             init_get_bits(&hx->inter_gb, ptr, bit_length);
3868             hx->inter_gb_ptr= &hx->inter_gb;
3869
3870             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
3871                && s->context_initialized
3872                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3873                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3874                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3875                && avctx->skip_frame < AVDISCARD_ALL)
3876                 context_count++;
3877             break;
3878         case NAL_SEI:
3879             init_get_bits(&s->gb, ptr, bit_length);
3880             ff_h264_decode_sei(h);
3881             break;
3882         case NAL_SPS:
3883             init_get_bits(&s->gb, ptr, bit_length);
3884             ff_h264_decode_seq_parameter_set(h);
3885
3886             if (s->flags& CODEC_FLAG_LOW_DELAY ||
3887                 (h->sps.bitstream_restriction_flag && !h->sps.num_reorder_frames))
3888                 s->low_delay=1;
3889
3890             if(avctx->has_b_frames < 2)
3891                 avctx->has_b_frames= !s->low_delay;
3892
3893             if (avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
3894                 h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
3895                 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
3896                     avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3897                     h->cur_chroma_format_idc = h->sps.chroma_format_idc;
3898                     h->pixel_shift = h->sps.bit_depth_luma > 8;
3899
3900                     ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3901                     ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3902                     s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
3903                     dsputil_init(&s->dsp, s->avctx);
3904                 } else {
3905                     av_log(avctx, AV_LOG_ERROR, "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3906                     return -1;
3907                 }
3908             }
3909             break;
3910         case NAL_PPS:
3911             init_get_bits(&s->gb, ptr, bit_length);
3912
3913             ff_h264_decode_picture_parameter_set(h, bit_length);
3914
3915             break;
3916         case NAL_AUD:
3917         case NAL_END_SEQUENCE:
3918         case NAL_END_STREAM:
3919         case NAL_FILLER_DATA:
3920         case NAL_SPS_EXT:
3921         case NAL_AUXILIARY_SLICE:
3922             break;
3923         default:
3924             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
3925         }
3926
3927         if(context_count == h->max_contexts) {
3928             execute_decode_slices(h, context_count);
3929             context_count = 0;
3930         }
3931
3932         if (err < 0)
3933             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3934         else if(err == 1) {
3935             /* Slice could not be decoded in parallel mode, copy down
3936              * NAL unit stuff to context 0 and restart. Note that
3937              * rbsp_buffer is not transferred, but since we no longer
3938              * run in parallel mode this should not be an issue. */
3939             h->nal_unit_type = hx->nal_unit_type;
3940             h->nal_ref_idc   = hx->nal_ref_idc;
3941             hx = h;
3942             goto again;
3943         }
3944     }
3945     }
3946     if(context_count)
3947         execute_decode_slices(h, context_count);
3948     return buf_index;
3949 }
3950
3951 /**
3952  * returns the number of bytes consumed for building the current frame
3953  */
3954 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
3955         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
3956         if(pos+10>buf_size) pos=buf_size; // oops ;)
3957
3958         return pos;
3959 }
3960
3961 static int decode_frame(AVCodecContext *avctx,
3962                              void *data, int *data_size,
3963                              AVPacket *avpkt)
3964 {
3965     const uint8_t *buf = avpkt->data;
3966     int buf_size = avpkt->size;
3967     H264Context *h = avctx->priv_data;
3968     MpegEncContext *s = &h->s;
3969     AVFrame *pict = data;
3970     int buf_index;
3971
3972     s->flags= avctx->flags;
3973     s->flags2= avctx->flags2;
3974
3975    /* end of stream, output what is still in the buffers */
3976  out:
3977     if (buf_size == 0) {
3978         Picture *out;
3979         int i, out_idx;
3980
3981         s->current_picture_ptr = NULL;
3982
3983 //FIXME factorize this with the output code below
3984         out = h->delayed_pic[0];
3985         out_idx = 0;
3986         for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3987             if(h->delayed_pic[i]->poc < out->poc){
3988                 out = h->delayed_pic[i];
3989                 out_idx = i;
3990             }
3991
3992         for(i=out_idx; h->delayed_pic[i]; i++)
3993             h->delayed_pic[i] = h->delayed_pic[i+1];
3994
3995         if(out){
3996             *data_size = sizeof(AVFrame);
3997             *pict= *(AVFrame*)out;
3998         }
3999
4000         return 0;
4001     }
4002
4003     buf_index=decode_nal_units(h, buf, buf_size);
4004     if(buf_index < 0)
4005         return -1;
4006
4007     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
4008         buf_size = 0;
4009         goto out;
4010     }
4011
4012     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
4013         if (avctx->skip_frame >= AVDISCARD_NONREF)
4014             return 0;
4015         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
4016         return -1;
4017     }
4018
4019     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
4020
4021         if(s->flags2 & CODEC_FLAG2_CHUNKS) decode_postinit(h, 1);
4022
4023         field_end(h, 0);
4024
4025         if (!h->next_output_pic) {
4026             /* Wait for second field. */
4027             *data_size = 0;
4028