mov: support eac3 audio
[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 FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "libavutil/imgutils.h"
29 #include "internal.h"
30 #include "dsputil.h"
31 #include "avcodec.h"
32 #include "mpegvideo.h"
33 #include "h264.h"
34 #include "h264data.h"
35 #include "h264_mvpred.h"
36 #include "golomb.h"
37 #include "mathops.h"
38 #include "rectangle.h"
39 #include "thread.h"
40 #include "vdpau_internal.h"
41 #include "libavutil/avassert.h"
42
43 #include "cabac.h"
44
45 //#undef NDEBUG
46 #include <assert.h>
47
48 static const uint8_t rem6[QP_MAX_NUM+1]={
49 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
50 };
51
52 static const uint8_t div6[QP_MAX_NUM+1]={
53 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9,10,10,10,10,
54 };
55
56 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
57     PIX_FMT_DXVA2_VLD,
58     PIX_FMT_VAAPI_VLD,
59     PIX_FMT_YUVJ420P,
60     PIX_FMT_NONE
61 };
62
63 /**
64  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
65  */
66 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
67     MpegEncContext * const s = &h->s;
68     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
69     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
70     int i;
71
72     if(!(h->top_samples_available&0x8000)){
73         for(i=0; i<4; i++){
74             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
75             if(status<0){
76                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
77                 return -1;
78             } else if(status){
79                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
80             }
81         }
82     }
83
84     if((h->left_samples_available&0x8888)!=0x8888){
85         static const int mask[4]={0x8000,0x2000,0x80,0x20};
86         for(i=0; i<4; i++){
87             if(!(h->left_samples_available&mask[i])){
88                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
89                 if(status<0){
90                     av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
91                     return -1;
92                 } else if(status){
93                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
94                 }
95             }
96         }
97     }
98
99     return 0;
100 } //FIXME cleanup like check_intra_pred_mode
101
102 static int check_intra_pred_mode(H264Context *h, int mode, int is_chroma){
103     MpegEncContext * const s = &h->s;
104     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
105     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
106
107     if(mode > 6U) {
108         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);
109         return -1;
110     }
111
112     if(!(h->top_samples_available&0x8000)){
113         mode= top[ mode ];
114         if(mode<0){
115             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);
116             return -1;
117         }
118     }
119
120     if((h->left_samples_available&0x8080) != 0x8080){
121         mode= left[ mode ];
122         if(is_chroma && (h->left_samples_available&0x8080)){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
123             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
124         }
125         if(mode<0){
126             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);
127             return -1;
128         }
129     }
130
131     return mode;
132 }
133
134 /**
135  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
136  */
137 int ff_h264_check_intra16x16_pred_mode(H264Context *h, int mode)
138 {
139     return check_intra_pred_mode(h, mode, 0);
140 }
141
142 /**
143  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
144  */
145 int ff_h264_check_intra_chroma_pred_mode(H264Context *h, int mode)
146 {
147     return check_intra_pred_mode(h, mode, 1);
148 }
149
150
151 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
152     int i, si, di;
153     uint8_t *dst;
154     int bufidx;
155
156 //    src[0]&0x80;                //forbidden bit
157     h->nal_ref_idc= src[0]>>5;
158     h->nal_unit_type= src[0]&0x1F;
159
160     src++; length--;
161
162 #if HAVE_FAST_UNALIGNED
163 # if HAVE_FAST_64BIT
164 #   define RS 7
165     for(i=0; i+1<length; i+=9){
166         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
167 # else
168 #   define RS 3
169     for(i=0; i+1<length; i+=5){
170         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
171 # endif
172             continue;
173         if(i>0 && !src[i]) i--;
174         while(src[i]) i++;
175 #else
176 #   define RS 0
177     for(i=0; i+1<length; i+=2){
178         if(src[i]) continue;
179         if(i>0 && src[i-1]==0) i--;
180 #endif
181         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
182             if(src[i+2]!=3){
183                 /* startcode, so we must be past the end */
184                 length=i;
185             }
186             break;
187         }
188         i-= RS;
189     }
190
191     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
192     si=h->rbsp_buffer_size[bufidx];
193     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE+MAX_MBPAIR_SIZE);
194     dst= h->rbsp_buffer[bufidx];
195     if(si != h->rbsp_buffer_size[bufidx])
196         memset(dst + length, 0, FF_INPUT_BUFFER_PADDING_SIZE+MAX_MBPAIR_SIZE);
197
198     if (dst == NULL){
199         return NULL;
200     }
201
202     if(i>=length-1){ //no escaped 0
203         *dst_length= length;
204         *consumed= length+1; //+1 for the header
205         if(h->s.avctx->flags2 & CODEC_FLAG2_FAST){
206             return src;
207         }else{
208             memcpy(dst, src, length);
209             return dst;
210         }
211     }
212
213 //printf("decoding esc\n");
214     memcpy(dst, src, i);
215     si=di=i;
216     while(si+2<length){
217         //remove escapes (very rare 1:2^22)
218         if(src[si+2]>3){
219             dst[di++]= src[si++];
220             dst[di++]= src[si++];
221         }else if(src[si]==0 && src[si+1]==0){
222             if(src[si+2]==3){ //escape
223                 dst[di++]= 0;
224                 dst[di++]= 0;
225                 si+=3;
226                 continue;
227             }else //next start code
228                 goto nsc;
229         }
230
231         dst[di++]= src[si++];
232     }
233     while(si<length)
234         dst[di++]= src[si++];
235 nsc:
236
237     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
238
239     *dst_length= di;
240     *consumed= si + 1;//+1 for the header
241 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
242     return dst;
243 }
244
245 /**
246  * Identify the exact end of the bitstream
247  * @return the length of the trailing, or 0 if damaged
248  */
249 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
250     int v= *src;
251     int r;
252
253     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
254
255     for(r=1; r<9; r++){
256         if(v&1) return r;
257         v>>=1;
258     }
259     return 0;
260 }
261
262 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
263                                  int y_offset, int list){
264     int raw_my= h->mv_cache[list][ scan8[n] ][1];
265     int filter_height= (raw_my&3) ? 2 : 0;
266     int full_my= (raw_my>>2) + y_offset;
267     int top = full_my - filter_height, bottom = full_my + height + filter_height;
268
269     return FFMAX(abs(top), bottom);
270 }
271
272 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
273                                int y_offset, int list0, int list1, int *nrefs){
274     MpegEncContext * const s = &h->s;
275     int my;
276
277     y_offset += 16*(s->mb_y >> MB_FIELD);
278
279     if(list0){
280         int ref_n = h->ref_cache[0][ scan8[n] ];
281         Picture *ref= &h->ref_list[0][ref_n];
282
283         // Error resilience puts the current picture in the ref list.
284         // Don't try to wait on these as it will cause a deadlock.
285         // Fields can wait on each other, though.
286         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
287            (ref->f.reference & 3) != s->picture_structure) {
288             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
289             if (refs[0][ref_n] < 0) nrefs[0] += 1;
290             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
291         }
292     }
293
294     if(list1){
295         int ref_n = h->ref_cache[1][ scan8[n] ];
296         Picture *ref= &h->ref_list[1][ref_n];
297
298         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
299            (ref->f.reference & 3) != s->picture_structure) {
300             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
301             if (refs[1][ref_n] < 0) nrefs[1] += 1;
302             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
303         }
304     }
305 }
306
307 /**
308  * Wait until all reference frames are available for MC operations.
309  *
310  * @param h the H264 context
311  */
312 static void await_references(H264Context *h){
313     MpegEncContext * const s = &h->s;
314     const int mb_xy= h->mb_xy;
315     const int mb_type = s->current_picture.f.mb_type[mb_xy];
316     int refs[2][48];
317     int nrefs[2] = {0};
318     int ref, list;
319
320     memset(refs, -1, sizeof(refs));
321
322     if(IS_16X16(mb_type)){
323         get_lowest_part_y(h, refs, 0, 16, 0,
324                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
325     }else if(IS_16X8(mb_type)){
326         get_lowest_part_y(h, refs, 0, 8, 0,
327                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
328         get_lowest_part_y(h, refs, 8, 8, 8,
329                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
330     }else if(IS_8X16(mb_type)){
331         get_lowest_part_y(h, refs, 0, 16, 0,
332                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
333         get_lowest_part_y(h, refs, 4, 16, 0,
334                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
335     }else{
336         int i;
337
338         assert(IS_8X8(mb_type));
339
340         for(i=0; i<4; i++){
341             const int sub_mb_type= h->sub_mb_type[i];
342             const int n= 4*i;
343             int y_offset= (i&2)<<2;
344
345             if(IS_SUB_8X8(sub_mb_type)){
346                 get_lowest_part_y(h, refs, n  , 8, y_offset,
347                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
348             }else if(IS_SUB_8X4(sub_mb_type)){
349                 get_lowest_part_y(h, refs, n  , 4, y_offset,
350                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
351                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
352                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
353             }else if(IS_SUB_4X8(sub_mb_type)){
354                 get_lowest_part_y(h, refs, n  , 8, y_offset,
355                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
356                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
357                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
358             }else{
359                 int j;
360                 assert(IS_SUB_4X4(sub_mb_type));
361                 for(j=0; j<4; j++){
362                     int sub_y_offset= y_offset + 2*(j&2);
363                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
364                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
365                 }
366             }
367         }
368     }
369
370     for(list=h->list_count-1; list>=0; list--){
371         for(ref=0; ref<48 && nrefs[list]; ref++){
372             int row = refs[list][ref];
373             if(row >= 0){
374                 Picture *ref_pic = &h->ref_list[list][ref];
375                 int ref_field = ref_pic->f.reference - 1;
376                 int ref_field_picture = ref_pic->field_picture;
377                 int pic_height = 16*s->mb_height >> ref_field_picture;
378
379                 row <<= MB_MBAFF;
380                 nrefs[list]--;
381
382                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
383                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
384                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
385                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
386                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
387                 }else if(FIELD_PICTURE){
388                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
389                 }else{
390                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
391                 }
392             }
393         }
394     }
395 }
396
397 #if 0
398 /**
399  * DCT transforms the 16 dc values.
400  * @param qp quantization parameter ??? FIXME
401  */
402 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
403 //    const int qmul= dequant_coeff[qp][0];
404     int i;
405     int temp[16]; //FIXME check if this is a good idea
406     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
407     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
408
409     for(i=0; i<4; i++){
410         const int offset= y_offset[i];
411         const int z0= block[offset+stride*0] + block[offset+stride*4];
412         const int z1= block[offset+stride*0] - block[offset+stride*4];
413         const int z2= block[offset+stride*1] - block[offset+stride*5];
414         const int z3= block[offset+stride*1] + block[offset+stride*5];
415
416         temp[4*i+0]= z0+z3;
417         temp[4*i+1]= z1+z2;
418         temp[4*i+2]= z1-z2;
419         temp[4*i+3]= z0-z3;
420     }
421
422     for(i=0; i<4; i++){
423         const int offset= x_offset[i];
424         const int z0= temp[4*0+i] + temp[4*2+i];
425         const int z1= temp[4*0+i] - temp[4*2+i];
426         const int z2= temp[4*1+i] - temp[4*3+i];
427         const int z3= temp[4*1+i] + temp[4*3+i];
428
429         block[stride*0 +offset]= (z0 + z3)>>1;
430         block[stride*2 +offset]= (z1 + z2)>>1;
431         block[stride*8 +offset]= (z1 - z2)>>1;
432         block[stride*10+offset]= (z0 - z3)>>1;
433     }
434 }
435 #endif
436
437 #undef xStride
438 #undef stride
439
440 #if 0
441 static void chroma_dc_dct_c(DCTELEM *block){
442     const int stride= 16*2;
443     const int xStride= 16;
444     int a,b,c,d,e;
445
446     a= block[stride*0 + xStride*0];
447     b= block[stride*0 + xStride*1];
448     c= block[stride*1 + xStride*0];
449     d= block[stride*1 + xStride*1];
450
451     e= a-b;
452     a= a+b;
453     b= c-d;
454     c= c+d;
455
456     block[stride*0 + xStride*0]= (a+c);
457     block[stride*0 + xStride*1]= (e+b);
458     block[stride*1 + xStride*0]= (a-c);
459     block[stride*1 + xStride*1]= (e-b);
460 }
461 #endif
462
463 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
464                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
465                            int src_x_offset, int src_y_offset,
466                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
467                            int pixel_shift, int chroma444){
468     MpegEncContext * const s = &h->s;
469     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
470     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
471     const int luma_xy= (mx&3) + ((my&3)<<2);
472     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
473     uint8_t * src_y = pic->f.data[0] + offset;
474     uint8_t * src_cb, * src_cr;
475     int extra_width= h->emu_edge_width;
476     int extra_height= h->emu_edge_height;
477     int emu=0;
478     const int full_mx= mx>>2;
479     const int full_my= my>>2;
480     const int pic_width  = 16*s->mb_width;
481     const int pic_height = 16*s->mb_height >> MB_FIELD;
482
483     if(mx&7) extra_width -= 3;
484     if(my&7) extra_height -= 3;
485
486     if(   full_mx < 0-extra_width
487        || full_my < 0-extra_height
488        || full_mx + 16/*FIXME*/ > pic_width + extra_width
489        || full_my + 16/*FIXME*/ > pic_height + extra_height){
490         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_y - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize, 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
491             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
492         emu=1;
493     }
494
495     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
496     if(!square){
497         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
498     }
499
500     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
501
502     if(chroma444){
503         src_cb = pic->f.data[1] + offset;
504         if(emu){
505             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
506                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
507             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
508         }
509         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
510         if(!square){
511             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
512         }
513
514         src_cr = pic->f.data[2] + offset;
515         if(emu){
516             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
517                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
518             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
519         }
520         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
521         if(!square){
522             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
523         }
524         return;
525     }
526
527     if(MB_FIELD){
528         // chroma offset when predicting from a field of opposite parity
529         my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
530         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
531     }
532     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
533     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
534
535     if(emu){
536         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
537             src_cb= s->edge_emu_buffer;
538     }
539     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
540
541     if(emu){
542         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
543             src_cr= s->edge_emu_buffer;
544     }
545     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
546 }
547
548 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
549                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
550                            int x_offset, int y_offset,
551                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
552                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
553                            int list0, int list1, int pixel_shift, int chroma444){
554     MpegEncContext * const s = &h->s;
555     qpel_mc_func *qpix_op=  qpix_put;
556     h264_chroma_mc_func chroma_op= chroma_put;
557
558     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
559     if(chroma444){
560         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
561         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
562     }else{
563         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
564         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
565     }
566     x_offset += 8*s->mb_x;
567     y_offset += 8*(s->mb_y >> MB_FIELD);
568
569     if(list0){
570         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
571         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
572                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
573                            qpix_op, chroma_op, pixel_shift, chroma444);
574
575         qpix_op=  qpix_avg;
576         chroma_op= chroma_avg;
577     }
578
579     if(list1){
580         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
581         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
582                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
583                            qpix_op, chroma_op, pixel_shift, chroma444);
584     }
585 }
586
587 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
588                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
589                            int x_offset, int y_offset,
590                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
591                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
592                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
593                            int list0, int list1, int pixel_shift, int chroma444){
594     MpegEncContext * const s = &h->s;
595
596     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
597     if(chroma444){
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{
603         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
604         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
605     }
606     x_offset += 8*s->mb_x;
607     y_offset += 8*(s->mb_y >> MB_FIELD);
608
609     if(list0 && list1){
610         /* don't optimize for luma-only case, since B-frames usually
611          * use implicit weights => chroma too. */
612         uint8_t *tmp_cb = s->obmc_scratchpad;
613         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
614         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
615         int refn0 = h->ref_cache[0][ scan8[n] ];
616         int refn1 = h->ref_cache[1][ scan8[n] ];
617
618         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
619                     dest_y, dest_cb, dest_cr,
620                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
621         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
622                     tmp_y, tmp_cb, tmp_cr,
623                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
624
625         if(h->use_weight == 2){
626             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
627             int weight1 = 64 - weight0;
628             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
629             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
630             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
631         }else{
632             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
633                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
634                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
635             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
636                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
637                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
638             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
639                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
640                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
641         }
642     }else{
643         int list = list1 ? 1 : 0;
644         int refn = h->ref_cache[list][ scan8[n] ];
645         Picture *ref= &h->ref_list[list][refn];
646         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
647                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
648                     qpix_put, chroma_put, pixel_shift, chroma444);
649
650         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
651                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
652         if(h->use_weight_chroma){
653             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
654                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
655             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
656                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
657         }
658     }
659 }
660
661 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
662                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
663                            int x_offset, int y_offset,
664                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
665                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
666                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
667                            int list0, int list1, int pixel_shift, int chroma444){
668     if((h->use_weight==2 && list0 && list1
669         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
670        || h->use_weight==1)
671         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
672                          x_offset, y_offset, qpix_put, chroma_put,
673                          weight_op[0], weight_op[3], weight_avg[0],
674                          weight_avg[3], list0, list1, pixel_shift, chroma444);
675     else
676         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
677                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
678                     chroma_avg, list0, list1, pixel_shift, chroma444);
679 }
680
681 static inline void prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma444){
682     /* fetch pixels for estimated mv 4 macroblocks ahead
683      * optimized for 64byte cache lines */
684     MpegEncContext * const s = &h->s;
685     const int refn = h->ref_cache[list][scan8[0]];
686     if(refn >= 0){
687         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
688         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
689         uint8_t **src = h->ref_list[list][refn].f.data;
690         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
691         s->dsp.prefetch(src[0]+off, s->linesize, 4);
692         if(chroma444){
693             s->dsp.prefetch(src[1]+off, s->linesize, 4);
694             s->dsp.prefetch(src[2]+off, s->linesize, 4);
695         }else{
696             off= (((mx>>1)+64)<<pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize;
697             s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
698         }
699     }
700 }
701
702 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
703                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
704                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
705                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
706                       int pixel_shift, int chroma444){
707     MpegEncContext * const s = &h->s;
708     const int mb_xy= h->mb_xy;
709     const int mb_type = s->current_picture.f.mb_type[mb_xy];
710
711     assert(IS_INTER(mb_type));
712
713     if(HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
714         await_references(h);
715     prefetch_motion(h, 0, pixel_shift, chroma444);
716
717     if(IS_16X16(mb_type)){
718         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
719                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
720                 weight_op, weight_avg,
721                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
722                 pixel_shift, chroma444);
723     }else if(IS_16X8(mb_type)){
724         mc_part(h, 0, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
725                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
726                 &weight_op[1], &weight_avg[1],
727                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
728                 pixel_shift, chroma444);
729         mc_part(h, 8, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
730                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
731                 &weight_op[1], &weight_avg[1],
732                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
733                 pixel_shift, chroma444);
734     }else if(IS_8X16(mb_type)){
735         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
736                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
737                 &weight_op[2], &weight_avg[2],
738                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
739                 pixel_shift, chroma444);
740         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
741                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
742                 &weight_op[2], &weight_avg[2],
743                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
744                 pixel_shift, chroma444);
745     }else{
746         int i;
747
748         assert(IS_8X8(mb_type));
749
750         for(i=0; i<4; i++){
751             const int sub_mb_type= h->sub_mb_type[i];
752             const int n= 4*i;
753             int x_offset= (i&1)<<2;
754             int y_offset= (i&2)<<1;
755
756             if(IS_SUB_8X8(sub_mb_type)){
757                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
758                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
759                     &weight_op[3], &weight_avg[3],
760                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
761                     pixel_shift, chroma444);
762             }else if(IS_SUB_8X4(sub_mb_type)){
763                 mc_part(h, n  , 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
764                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
765                     &weight_op[4], &weight_avg[4],
766                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
767                     pixel_shift, chroma444);
768                 mc_part(h, n+2, 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
769                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
770                     &weight_op[4], &weight_avg[4],
771                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
772                     pixel_shift, chroma444);
773             }else if(IS_SUB_4X8(sub_mb_type)){
774                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
775                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
776                     &weight_op[5], &weight_avg[5],
777                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
778                     pixel_shift, chroma444);
779                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
780                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
781                     &weight_op[5], &weight_avg[5],
782                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
783                     pixel_shift, chroma444);
784             }else{
785                 int j;
786                 assert(IS_SUB_4X4(sub_mb_type));
787                 for(j=0; j<4; j++){
788                     int sub_x_offset= x_offset + 2*(j&1);
789                     int sub_y_offset= y_offset +   (j&2);
790                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
791                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
792                         &weight_op[6], &weight_avg[6],
793                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
794                         pixel_shift, chroma444);
795                 }
796             }
797         }
798     }
799
800     prefetch_motion(h, 1, pixel_shift, chroma444);
801 }
802
803 static void free_tables(H264Context *h, int free_rbsp){
804     int i;
805     H264Context *hx;
806
807     av_freep(&h->intra4x4_pred_mode);
808     av_freep(&h->chroma_pred_mode_table);
809     av_freep(&h->cbp_table);
810     av_freep(&h->mvd_table[0]);
811     av_freep(&h->mvd_table[1]);
812     av_freep(&h->direct_table);
813     av_freep(&h->non_zero_count);
814     av_freep(&h->slice_table_base);
815     h->slice_table= NULL;
816     av_freep(&h->list_counts);
817
818     av_freep(&h->mb2b_xy);
819     av_freep(&h->mb2br_xy);
820
821     for(i = 0; i < MAX_THREADS; i++) {
822         hx = h->thread_context[i];
823         if(!hx) continue;
824         av_freep(&hx->top_borders[1]);
825         av_freep(&hx->top_borders[0]);
826         av_freep(&hx->s.obmc_scratchpad);
827         if (free_rbsp){
828             av_freep(&hx->rbsp_buffer[1]);
829             av_freep(&hx->rbsp_buffer[0]);
830             hx->rbsp_buffer_size[0] = 0;
831             hx->rbsp_buffer_size[1] = 0;
832         }
833         if (i) av_freep(&h->thread_context[i]);
834     }
835 }
836
837 static void init_dequant8_coeff_table(H264Context *h){
838     int i,j,q,x;
839     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
840
841     for(i=0; i<6; i++ ){
842         h->dequant8_coeff[i] = h->dequant8_buffer[i];
843         for(j=0; j<i; j++){
844             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
845                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
846                 break;
847             }
848         }
849         if(j<i)
850             continue;
851
852         for(q=0; q<max_qp+1; q++){
853             int shift = div6[q];
854             int idx = rem6[q];
855             for(x=0; x<64; x++)
856                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
857                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
858                     h->pps.scaling_matrix8[i][x]) << shift;
859         }
860     }
861 }
862
863 static void init_dequant4_coeff_table(H264Context *h){
864     int i,j,q,x;
865     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
866     for(i=0; i<6; i++ ){
867         h->dequant4_coeff[i] = h->dequant4_buffer[i];
868         for(j=0; j<i; j++){
869             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
870                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
871                 break;
872             }
873         }
874         if(j<i)
875             continue;
876
877         for(q=0; q<max_qp+1; q++){
878             int shift = div6[q] + 2;
879             int idx = rem6[q];
880             for(x=0; x<16; x++)
881                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
882                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
883                     h->pps.scaling_matrix4[i][x]) << shift;
884         }
885     }
886 }
887
888 static void init_dequant_tables(H264Context *h){
889     int i,x;
890     init_dequant4_coeff_table(h);
891     if(h->pps.transform_8x8_mode)
892         init_dequant8_coeff_table(h);
893     if(h->sps.transform_bypass){
894         for(i=0; i<6; i++)
895             for(x=0; x<16; x++)
896                 h->dequant4_coeff[i][0][x] = 1<<6;
897         if(h->pps.transform_8x8_mode)
898             for(i=0; i<6; i++)
899                 for(x=0; x<64; x++)
900                     h->dequant8_coeff[i][0][x] = 1<<6;
901     }
902 }
903
904
905 int ff_h264_alloc_tables(H264Context *h){
906     MpegEncContext * const s = &h->s;
907     const int big_mb_num= s->mb_stride * (s->mb_height+1);
908     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
909     int x,y;
910
911     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
912
913     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
914     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
915     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
916
917     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
918     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
919     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
920     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
921     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
922
923     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
924     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
925
926     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
927     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
928     for(y=0; y<s->mb_height; y++){
929         for(x=0; x<s->mb_width; x++){
930             const int mb_xy= x + y*s->mb_stride;
931             const int b_xy = 4*x + 4*y*h->b_stride;
932
933             h->mb2b_xy [mb_xy]= b_xy;
934             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
935         }
936     }
937
938     s->obmc_scratchpad = NULL;
939
940     if(!h->dequant4_coeff[0])
941         init_dequant_tables(h);
942
943     return 0;
944 fail:
945     free_tables(h, 1);
946     return -1;
947 }
948
949 /**
950  * Mimic alloc_tables(), but for every context thread.
951  */
952 static void clone_tables(H264Context *dst, H264Context *src, int i){
953     MpegEncContext * const s = &src->s;
954     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
955     dst->non_zero_count           = src->non_zero_count;
956     dst->slice_table              = src->slice_table;
957     dst->cbp_table                = src->cbp_table;
958     dst->mb2b_xy                  = src->mb2b_xy;
959     dst->mb2br_xy                 = src->mb2br_xy;
960     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
961     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
962     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
963     dst->direct_table             = src->direct_table;
964     dst->list_counts              = src->list_counts;
965
966     dst->s.obmc_scratchpad = NULL;
967     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma, src->sps.chroma_format_idc);
968 }
969
970 /**
971  * Init context
972  * Allocate buffers which are not shared amongst multiple threads.
973  */
974 static int context_init(H264Context *h){
975     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
976     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
977
978     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
979     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
980
981     return 0;
982 fail:
983     return -1; // free_tables will clean up for us
984 }
985
986 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
987
988 static av_cold void common_init(H264Context *h){
989     MpegEncContext * const s = &h->s;
990
991     s->width = s->avctx->width;
992     s->height = s->avctx->height;
993     s->codec_id= s->avctx->codec->id;
994
995     s->avctx->bits_per_raw_sample = 8;
996     h->cur_chroma_format_idc = 1;
997
998     ff_h264dsp_init(&h->h264dsp,
999                     s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1000     ff_h264_pred_init(&h->hpc, s->codec_id,
1001                       s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1002
1003     h->dequant_coeff_pps= -1;
1004     s->unrestricted_mv=1;
1005     s->decode=1; //FIXME
1006
1007     s->dsp.dct_bits = 16;
1008     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
1009
1010     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
1011     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
1012 }
1013
1014 int ff_h264_decode_extradata(H264Context *h, const uint8_t *buf, int size)
1015 {
1016     AVCodecContext *avctx = h->s.avctx;
1017
1018     if(!buf || size <= 0)
1019         return -1;
1020
1021     if(buf[0] == 1){
1022         int i, cnt, nalsize;
1023         const unsigned char *p = buf;
1024
1025         h->is_avc = 1;
1026
1027         if(size < 7) {
1028             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1029             return -1;
1030         }
1031         /* sps and pps in the avcC always have length coded with 2 bytes,
1032            so put a fake nal_length_size = 2 while parsing them */
1033         h->nal_length_size = 2;
1034         // Decode sps from avcC
1035         cnt = *(p+5) & 0x1f; // Number of sps
1036         p += 6;
1037         for (i = 0; i < cnt; i++) {
1038             nalsize = AV_RB16(p) + 2;
1039             if(nalsize > size - (p-buf))
1040                 return -1;
1041             if(decode_nal_units(h, p, nalsize) < 0) {
1042                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1043                 return -1;
1044             }
1045             p += nalsize;
1046         }
1047         // Decode pps from avcC
1048         cnt = *(p++); // Number of pps
1049         for (i = 0; i < cnt; i++) {
1050             nalsize = AV_RB16(p) + 2;
1051             if(nalsize > size - (p-buf))
1052                 return -1;
1053             if (decode_nal_units(h, p, nalsize) < 0) {
1054                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1055                 return -1;
1056             }
1057             p += nalsize;
1058         }
1059         // Now store right nal length size, that will be use to parse all other nals
1060         h->nal_length_size = (buf[4] & 0x03) + 1;
1061     } else {
1062         h->is_avc = 0;
1063         if(decode_nal_units(h, buf, size) < 0)
1064             return -1;
1065     }
1066     return 0;
1067 }
1068
1069 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1070     H264Context *h= avctx->priv_data;
1071     MpegEncContext * const s = &h->s;
1072
1073     MPV_decode_defaults(s);
1074
1075     s->avctx = avctx;
1076     common_init(h);
1077
1078     s->out_format = FMT_H264;
1079     s->workaround_bugs= avctx->workaround_bugs;
1080
1081     // set defaults
1082 //    s->decode_mb= ff_h263_decode_mb;
1083     s->quarter_sample = 1;
1084     if(!avctx->has_b_frames)
1085     s->low_delay= 1;
1086
1087     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1088
1089     ff_h264_decode_init_vlc();
1090
1091     h->pixel_shift = 0;
1092     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1093
1094     h->thread_context[0] = h;
1095     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1096     h->prev_poc_msb= 1<<16;
1097     h->x264_build = -1;
1098     ff_h264_reset_sei(h);
1099     if(avctx->codec_id == CODEC_ID_H264){
1100         if(avctx->ticks_per_frame == 1){
1101             s->avctx->time_base.den *=2;
1102         }
1103         avctx->ticks_per_frame = 2;
1104     }
1105
1106     if(avctx->extradata_size > 0 && avctx->extradata &&
1107         ff_h264_decode_extradata(h, avctx->extradata, avctx->extradata_size))
1108         return -1;
1109
1110     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1111         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1112         s->low_delay = 0;
1113     }
1114
1115     return 0;
1116 }
1117
1118 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1119 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1120 {
1121     int i;
1122
1123     for (i=0; i<count; i++){
1124         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1125                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1126                 !from[i]));
1127         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1128     }
1129 }
1130
1131 static void copy_parameter_set(void **to, void **from, int count, int size)
1132 {
1133     int i;
1134
1135     for (i=0; i<count; i++){
1136         if (to[i] && !from[i]) av_freep(&to[i]);
1137         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1138
1139         if (from[i]) memcpy(to[i], from[i], size);
1140     }
1141 }
1142
1143 static int decode_init_thread_copy(AVCodecContext *avctx){
1144     H264Context *h= avctx->priv_data;
1145
1146     if (!avctx->is_copy) return 0;
1147     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1148     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1149
1150     return 0;
1151 }
1152
1153 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1154 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1155     H264Context *h= dst->priv_data, *h1= src->priv_data;
1156     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1157     int inited = s->context_initialized, err;
1158     int i;
1159
1160     if(dst == src || !s1->context_initialized) return 0;
1161
1162     err = ff_mpeg_update_thread_context(dst, src);
1163     if(err) return err;
1164
1165     //FIXME handle width/height changing
1166     if(!inited){
1167         for(i = 0; i < MAX_SPS_COUNT; i++)
1168             av_freep(h->sps_buffers + i);
1169
1170         for(i = 0; i < MAX_PPS_COUNT; i++)
1171             av_freep(h->pps_buffers + i);
1172
1173         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1174         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1175         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1176         if (ff_h264_alloc_tables(h) < 0) {
1177             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1178             return AVERROR(ENOMEM);
1179         }
1180         context_init(h);
1181
1182         for(i=0; i<2; i++){
1183             h->rbsp_buffer[i] = NULL;
1184             h->rbsp_buffer_size[i] = 0;
1185         }
1186
1187         h->thread_context[0] = h;
1188
1189         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1190         // so this has to be allocated here
1191         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1192
1193         s->dsp.clear_blocks(h->mb);
1194         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1195     }
1196
1197     //extradata/NAL handling
1198     h->is_avc          = h1->is_avc;
1199
1200     //SPS/PPS
1201     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1202     h->sps             = h1->sps;
1203     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1204     h->pps             = h1->pps;
1205
1206     //Dequantization matrices
1207     //FIXME these are big - can they be only copied when PPS changes?
1208     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1209
1210     for(i=0; i<6; i++)
1211         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1212
1213     for(i=0; i<6; i++)
1214         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1215
1216     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1217
1218     //POC timing
1219     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1220
1221     //reference lists
1222     copy_fields(h, h1, ref_count, list_count);
1223     copy_fields(h, h1, ref_list,  intra_gb);
1224     copy_fields(h, h1, short_ref, cabac_init_idc);
1225
1226     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1227     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1228     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1229
1230     h->last_slice_type = h1->last_slice_type;
1231     h->sync            = h1->sync;
1232
1233     if(!s->current_picture_ptr) return 0;
1234
1235     if(!s->dropable) {
1236         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1237         h->prev_poc_msb     = h->poc_msb;
1238         h->prev_poc_lsb     = h->poc_lsb;
1239     }
1240     h->prev_frame_num_offset= h->frame_num_offset;
1241     h->prev_frame_num       = h->frame_num;
1242     h->outputed_poc         = h->next_outputed_poc;
1243
1244     return err;
1245 }
1246
1247 int ff_h264_frame_start(H264Context *h){
1248     MpegEncContext * const s = &h->s;
1249     int i;
1250     const int pixel_shift = h->pixel_shift;
1251     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1252
1253     if(MPV_frame_start(s, s->avctx) < 0)
1254         return -1;
1255     ff_er_frame_start(s);
1256     /*
1257      * MPV_frame_start uses pict_type to derive key_frame.
1258      * This is incorrect for H.264; IDR markings must be used.
1259      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1260      * See decode_nal_units().
1261      */
1262     s->current_picture_ptr->f.key_frame = 0;
1263     s->current_picture_ptr->mmco_reset= 0;
1264
1265     assert(s->linesize && s->uvlinesize);
1266
1267     for(i=0; i<16; i++){
1268         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1269         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1270     }
1271     for(i=0; i<16; i++){
1272         h->block_offset[16+i]=
1273         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1274         h->block_offset[48+16+i]=
1275         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1276     }
1277
1278     /* can't be in alloc_tables because linesize isn't known there.
1279      * FIXME: redo bipred weight to not require extra buffer? */
1280     for(i = 0; i < thread_count; i++)
1281         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1282             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1283
1284     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1285     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1286
1287 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1288
1289     // We mark the current picture as non-reference after allocating it, so
1290     // that if we break out due to an error it can be released automatically
1291     // in the next MPV_frame_start().
1292     // SVQ3 as well as most other codecs have only last/next/current and thus
1293     // get released even with set reference, besides SVQ3 and others do not
1294     // mark frames as reference later "naturally".
1295     if(s->codec_id != CODEC_ID_SVQ3)
1296         s->current_picture_ptr->f.reference = 0;
1297
1298     s->current_picture_ptr->field_poc[0]=
1299     s->current_picture_ptr->field_poc[1]= INT_MAX;
1300
1301     h->next_output_pic = NULL;
1302
1303     assert(s->current_picture_ptr->long_ref==0);
1304
1305     return 0;
1306 }
1307
1308 /**
1309   * Run setup operations that must be run after slice header decoding.
1310   * This includes finding the next displayed frame.
1311   *
1312   * @param h h264 master context
1313   * @param setup_finished enough NALs have been read that we can call
1314   * ff_thread_finish_setup()
1315   */
1316 static void decode_postinit(H264Context *h, int setup_finished){
1317     MpegEncContext * const s = &h->s;
1318     Picture *out = s->current_picture_ptr;
1319     Picture *cur = s->current_picture_ptr;
1320     int i, pics, out_of_order, out_idx;
1321
1322     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1323     s->current_picture_ptr->f.pict_type   = s->pict_type;
1324
1325     if (h->next_output_pic) return;
1326
1327     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1328         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1329         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1330         //to find this yet, so we assume the worst for now.
1331         //if (setup_finished)
1332         //    ff_thread_finish_setup(s->avctx);
1333         return;
1334     }
1335
1336     cur->f.interlaced_frame = 0;
1337     cur->f.repeat_pict      = 0;
1338
1339     /* Signal interlacing information externally. */
1340     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1341
1342     if(h->sps.pic_struct_present_flag){
1343         switch (h->sei_pic_struct)
1344         {
1345         case SEI_PIC_STRUCT_FRAME:
1346             break;
1347         case SEI_PIC_STRUCT_TOP_FIELD:
1348         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1349             cur->f.interlaced_frame = 1;
1350             break;
1351         case SEI_PIC_STRUCT_TOP_BOTTOM:
1352         case SEI_PIC_STRUCT_BOTTOM_TOP:
1353             if (FIELD_OR_MBAFF_PICTURE)
1354                 cur->f.interlaced_frame = 1;
1355             else
1356                 // try to flag soft telecine progressive
1357                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1358             break;
1359         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1360         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1361             // Signal the possibility of telecined film externally (pic_struct 5,6)
1362             // From these hints, let the applications decide if they apply deinterlacing.
1363             cur->f.repeat_pict = 1;
1364             break;
1365         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1366             // Force progressive here, as doubling interlaced frame is a bad idea.
1367             cur->f.repeat_pict = 2;
1368             break;
1369         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1370             cur->f.repeat_pict = 4;
1371             break;
1372         }
1373
1374         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1375             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1376     }else{
1377         /* Derive interlacing flag from used decoding process. */
1378         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1379     }
1380     h->prev_interlaced_frame = cur->f.interlaced_frame;
1381
1382     if (cur->field_poc[0] != cur->field_poc[1]){
1383         /* Derive top_field_first from field pocs. */
1384         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1385     }else{
1386         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1387             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1388             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1389               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1390                 cur->f.top_field_first = 1;
1391             else
1392                 cur->f.top_field_first = 0;
1393         }else{
1394             /* Most likely progressive */
1395             cur->f.top_field_first = 0;
1396         }
1397     }
1398
1399     //FIXME do something with unavailable reference frames
1400
1401     /* Sort B-frames into display order */
1402
1403     if(h->sps.bitstream_restriction_flag
1404        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1405         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1406         s->low_delay = 0;
1407     }
1408
1409     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1410        && !h->sps.bitstream_restriction_flag){
1411         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1412         s->low_delay= 0;
1413     }
1414
1415     pics = 0;
1416     while(h->delayed_pic[pics]) pics++;
1417
1418     av_assert0(pics <= MAX_DELAYED_PIC_COUNT);
1419
1420     h->delayed_pic[pics++] = cur;
1421     if (cur->f.reference == 0)
1422         cur->f.reference = DELAYED_PIC_REF;
1423
1424     out = h->delayed_pic[0];
1425     out_idx = 0;
1426     for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1427         if(h->delayed_pic[i]->poc < out->poc){
1428             out = h->delayed_pic[i];
1429             out_idx = i;
1430         }
1431     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1432         h->next_outputed_poc= INT_MIN;
1433     out_of_order = out->poc < h->next_outputed_poc;
1434
1435     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1436         { }
1437     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1438        || (s->low_delay &&
1439         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1440          || cur->f.pict_type == AV_PICTURE_TYPE_B)))
1441     {
1442         s->low_delay = 0;
1443         s->avctx->has_b_frames++;
1444     }
1445
1446     if(out_of_order || pics > s->avctx->has_b_frames){
1447         out->f.reference &= ~DELAYED_PIC_REF;
1448         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1449                          // or else the first thread can release the picture and reuse it unsafely
1450         for(i=out_idx; h->delayed_pic[i]; i++)
1451             h->delayed_pic[i] = h->delayed_pic[i+1];
1452     }
1453     if(!out_of_order && pics > s->avctx->has_b_frames){
1454         h->next_output_pic = out;
1455         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1456             h->next_outputed_poc = INT_MIN;
1457         } else
1458             h->next_outputed_poc = out->poc;
1459     }else{
1460         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1461     }
1462
1463     if (h->next_output_pic && h->next_output_pic->sync) {
1464         h->sync |= 2;
1465     }
1466
1467     if (setup_finished)
1468         ff_thread_finish_setup(s->avctx);
1469 }
1470
1471 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int simple){
1472     MpegEncContext * const s = &h->s;
1473     uint8_t *top_border;
1474     int top_idx = 1;
1475     const int pixel_shift = h->pixel_shift;
1476     int chroma444 = CHROMA444;
1477     int chroma422 = CHROMA422;
1478
1479     src_y  -=   linesize;
1480     src_cb -= uvlinesize;
1481     src_cr -= uvlinesize;
1482
1483     if(!simple && FRAME_MBAFF){
1484         if(s->mb_y&1){
1485             if(!MB_MBAFF){
1486                 top_border = h->top_borders[0][s->mb_x];
1487                 AV_COPY128(top_border, src_y + 15*linesize);
1488                 if (pixel_shift)
1489                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1490                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1491                     if(chroma444){
1492                         if (pixel_shift){
1493                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1494                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1495                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1496                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1497                         } else {
1498                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1499                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1500                         }
1501                     } else if(chroma422){
1502                         if (pixel_shift) {
1503                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1504                             AV_COPY128(top_border+48, src_cr + 15*uvlinesize);
1505                         } else {
1506                             AV_COPY64(top_border+16, src_cb +  15*uvlinesize);
1507                             AV_COPY64(top_border+24, src_cr +  15*uvlinesize);
1508                         }
1509                     } else {
1510                         if (pixel_shift) {
1511                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1512                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1513                         } else {
1514                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1515                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1516                         }
1517                     }
1518                 }
1519             }
1520         }else if(MB_MBAFF){
1521             top_idx = 0;
1522         }else
1523             return;
1524     }
1525
1526     top_border = h->top_borders[top_idx][s->mb_x];
1527     // There are two lines saved, the line above the the top macroblock of a pair,
1528     // and the line above the bottom macroblock
1529     AV_COPY128(top_border, src_y + 16*linesize);
1530     if (pixel_shift)
1531         AV_COPY128(top_border+16, src_y+16*linesize+16);
1532
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 + 16*linesize);
1537                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1538                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1539                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1540             } else {
1541                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1542                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1543             }
1544         } else if(chroma422) {
1545             if (pixel_shift) {
1546                 AV_COPY128(top_border+32, src_cb+16*uvlinesize);
1547                 AV_COPY128(top_border+48, src_cr+16*uvlinesize);
1548             } else {
1549                 AV_COPY64(top_border+16, src_cb+16*uvlinesize);
1550                 AV_COPY64(top_border+24, src_cr+16*uvlinesize);
1551             }
1552         } else {
1553             if (pixel_shift) {
1554                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1555                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1556             } else {
1557                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1558                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1559             }
1560         }
1561     }
1562 }
1563
1564 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1565                                   uint8_t *src_cb, uint8_t *src_cr,
1566                                   int linesize, int uvlinesize,
1567                                   int xchg, int chroma444,
1568                                   int simple, int pixel_shift){
1569     MpegEncContext * const s = &h->s;
1570     int deblock_topleft;
1571     int deblock_top;
1572     int top_idx = 1;
1573     uint8_t *top_border_m1;
1574     uint8_t *top_border;
1575
1576     if(!simple && FRAME_MBAFF){
1577         if(s->mb_y&1){
1578             if(!MB_MBAFF)
1579                 return;
1580         }else{
1581             top_idx = MB_MBAFF ? 0 : 1;
1582         }
1583     }
1584
1585     if(h->deblocking_filter == 2) {
1586         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1587         deblock_top     = h->top_type;
1588     } else {
1589         deblock_topleft = (s->mb_x > 0);
1590         deblock_top     = (s->mb_y > !!MB_FIELD);
1591     }
1592
1593     src_y  -=   linesize + 1 + pixel_shift;
1594     src_cb -= uvlinesize + 1 + pixel_shift;
1595     src_cr -= uvlinesize + 1 + pixel_shift;
1596
1597     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1598     top_border    = h->top_borders[top_idx][s->mb_x];
1599
1600 #define XCHG(a,b,xchg)\
1601     if (pixel_shift) {\
1602         if (xchg) {\
1603             AV_SWAP64(b+0,a+0);\
1604             AV_SWAP64(b+8,a+8);\
1605         } else {\
1606             AV_COPY128(b,a); \
1607         }\
1608     } else \
1609 if (xchg) AV_SWAP64(b,a);\
1610 else      AV_COPY64(b,a);
1611
1612     if(deblock_top){
1613         if(deblock_topleft){
1614             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1615         }
1616         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1617         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1618         if(s->mb_x+1 < s->mb_width){
1619             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1620         }
1621     }
1622     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1623         if(chroma444){
1624             if(deblock_topleft){
1625                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1626                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1627             }
1628             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1629             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1630             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1631             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1632             if(s->mb_x+1 < s->mb_width){
1633                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1634                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1635             }
1636         } else {
1637             if(deblock_top){
1638                 if(deblock_topleft){
1639                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1640                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1641                 }
1642                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1643                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1644             }
1645         }
1646     }
1647 }
1648
1649 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1650     if (high_bit_depth) {
1651         return AV_RN32A(((int32_t*)mb) + index);
1652     } else
1653         return AV_RN16A(mb + index);
1654 }
1655
1656 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1657     if (high_bit_depth) {
1658         AV_WN32A(((int32_t*)mb) + index, value);
1659     } else
1660         AV_WN16A(mb + index, value);
1661 }
1662
1663 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1664                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1665 {
1666     MpegEncContext * const s = &h->s;
1667     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1668     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1669     int i;
1670     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1671     block_offset += 16*p;
1672     if(IS_INTRA4x4(mb_type)){
1673         if(simple || !s->encoding){
1674             if(IS_8x8DCT(mb_type)){
1675                 if(transform_bypass){
1676                     idct_dc_add =
1677                     idct_add    = s->dsp.add_pixels8;
1678                 }else{
1679                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1680                     idct_add    = h->h264dsp.h264_idct8_add;
1681                 }
1682                 for(i=0; i<16; i+=4){
1683                     uint8_t * const ptr= dest_y + block_offset[i];
1684                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1685                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1686                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1687                     }else{
1688                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1689                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1690                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1691                         if(nnz){
1692                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1693                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1694                             else
1695                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1696                         }
1697                     }
1698                 }
1699             }else{
1700                 if(transform_bypass){
1701                     idct_dc_add =
1702                     idct_add    = s->dsp.add_pixels4;
1703                 }else{
1704                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1705                     idct_add    = h->h264dsp.h264_idct_add;
1706                 }
1707                 for(i=0; i<16; i++){
1708                     uint8_t * const ptr= dest_y + block_offset[i];
1709                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1710
1711                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1712                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1713                     }else{
1714                         uint8_t *topright;
1715                         int nnz, tr;
1716                         uint64_t tr_high;
1717                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1718                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1719                             assert(s->mb_y || linesize <= block_offset[i]);
1720                             if(!topright_avail){
1721                                 if (pixel_shift) {
1722                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1723                                     topright= (uint8_t*) &tr_high;
1724                                 } else {
1725                                     tr= ptr[3 - linesize]*0x01010101u;
1726                                     topright= (uint8_t*) &tr;
1727                                 }
1728                             }else
1729                                 topright= ptr + (4 << pixel_shift) - linesize;
1730                         }else
1731                             topright= NULL;
1732
1733                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1734                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1735                         if(nnz){
1736                             if(is_h264){
1737                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1738                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1739                                 else
1740                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1741                             }else
1742                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1743                         }
1744                     }
1745                 }
1746             }
1747         }
1748     }else{
1749         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1750         if(is_h264){
1751             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1752                 if(!transform_bypass)
1753                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1754                 else{
1755                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1756                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1757                     for(i = 0; i < 16; i++)
1758                         dctcoef_set(h->mb+p*256, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1759                 }
1760             }
1761         }else
1762             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1763     }
1764 }
1765
1766 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1767                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1768 {
1769     MpegEncContext * const s = &h->s;
1770     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1771     int i;
1772     block_offset += 16*p;
1773     if(!IS_INTRA4x4(mb_type)){
1774         if(is_h264){
1775             if(IS_INTRA16x16(mb_type)){
1776                 if(transform_bypass){
1777                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1778                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1779                     }else{
1780                         for(i=0; i<16; i++){
1781                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1782                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1783                         }
1784                     }
1785                 }else{
1786                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1787                 }
1788             }else if(h->cbp&15){
1789                 if(transform_bypass){
1790                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1791                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1792                     for(i=0; i<16; i+=di){
1793                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1794                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1795                         }
1796                     }
1797                 }else{
1798                     if(IS_8x8DCT(mb_type)){
1799                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1800                     }else{
1801                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1802                     }
1803                 }
1804             }
1805         }else{
1806             for(i=0; i<16; i++){
1807                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1808                     uint8_t * const ptr= dest_y + block_offset[i];
1809                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1810                 }
1811             }
1812         }
1813     }
1814 }
1815
1816 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1817     MpegEncContext * const s = &h->s;
1818     const int mb_x= s->mb_x;
1819     const int mb_y= s->mb_y;
1820     const int mb_xy= h->mb_xy;
1821     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1822     uint8_t  *dest_y, *dest_cb, *dest_cr;
1823     int linesize, uvlinesize /*dct_offset*/;
1824     int i, j;
1825     int *block_offset = &h->block_offset[0];
1826     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1827     /* is_h264 should always be true if SVQ3 is disabled. */
1828     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1829     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1830     const int block_h = 16>>s->chroma_y_shift;
1831
1832     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1833     dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1834     dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1835
1836     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1837     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1838
1839     h->list_counts[mb_xy]= h->list_count;
1840
1841     if (!simple && MB_FIELD) {
1842         linesize   = h->mb_linesize   = s->linesize * 2;
1843         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1844         block_offset = &h->block_offset[48];
1845         if(mb_y&1){ //FIXME move out of this function?
1846             dest_y -= s->linesize*15;
1847             dest_cb-= s->uvlinesize*(block_h-1);
1848             dest_cr-= s->uvlinesize*(block_h-1);
1849         }
1850         if(FRAME_MBAFF) {
1851             int list;
1852             for(list=0; list<h->list_count; list++){
1853                 if(!USES_LIST(mb_type, list))
1854                     continue;
1855                 if(IS_16X16(mb_type)){
1856                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1857                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1858                 }else{
1859                     for(i=0; i<16; i+=4){
1860                         int ref = h->ref_cache[list][scan8[i]];
1861                         if(ref >= 0)
1862                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1863                     }
1864                 }
1865             }
1866         }
1867     } else {
1868         linesize   = h->mb_linesize   = s->linesize;
1869         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1870 //        dct_offset = s->linesize * 16;
1871     }
1872
1873     if (!simple && IS_INTRA_PCM(mb_type)) {
1874         const int bit_depth = h->sps.bit_depth_luma;
1875         if (pixel_shift) {
1876             int j;
1877             GetBitContext gb;
1878             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1879
1880             for (i = 0; i < 16; i++) {
1881                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1882                 for (j = 0; j < 16; j++)
1883                     tmp_y[j] = get_bits(&gb, bit_depth);
1884             }
1885             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1886                 if (!h->sps.chroma_format_idc) {
1887                     for (i = 0; i < 8; i++) {
1888                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1889                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1890                         for (j = 0; j < 8; j++) {
1891                             tmp_cb[j] = tmp_cr[j] = 1 << (bit_depth - 1);
1892                         }
1893                     }
1894                 } else {
1895                     for (i = 0; i < block_h; i++) {
1896                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1897                         for (j = 0; j < 8; j++)
1898                             tmp_cb[j] = get_bits(&gb, bit_depth);
1899                     }
1900                     for (i = 0; i < block_h; i++) {
1901                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1902                         for (j = 0; j < 8; j++)
1903                             tmp_cr[j] = get_bits(&gb, bit_depth);
1904                     }
1905                 }
1906             }
1907         } else {
1908             for (i=0; i<16; i++) {
1909                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1910             }
1911             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1912                 if (!h->sps.chroma_format_idc) {
1913                     for (i=0; i<8; i++) {
1914                         memset(dest_cb+ i*uvlinesize, 1 << (bit_depth - 1), 8);
1915                         memset(dest_cr+ i*uvlinesize, 1 << (bit_depth - 1), 8);
1916                     }
1917                 } else {
1918                     for (i=0; i<block_h; i++) {
1919                         memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
1920                         memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
1921                     }
1922                 }
1923             }
1924         }
1925     } else {
1926         if(IS_INTRA(mb_type)){
1927             if(h->deblocking_filter)
1928                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1929
1930             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1931                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1932                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1933             }
1934
1935             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1936
1937             if(h->deblocking_filter)
1938                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1939         }else if(is_h264){
1940             hl_motion(h, dest_y, dest_cb, dest_cr,
1941                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1942                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1943                       h->h264dsp.weight_h264_pixels_tab,
1944                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 0);
1945         }
1946
1947         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1948
1949         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1950             uint8_t *dest[2] = {dest_cb, dest_cr};
1951             if(transform_bypass){
1952                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1953                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
1954                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
1955                 }else{
1956                     idct_add = s->dsp.add_pixels4;
1957                     for(j=1; j<3; j++){
1958                         for(i=j*16; i<j*16+4; i++){
1959                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1960                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1961                         }
1962                     }
1963                 }
1964             }else{
1965                 if(is_h264){
1966                     int qp[2];
1967                     if (CHROMA422) {
1968                         qp[0] = h->chroma_qp[0]+3;
1969                         qp[1] = h->chroma_qp[1]+3;
1970                     } else {
1971                         qp[0] = h->chroma_qp[0];
1972                         qp[1] = h->chroma_qp[1];
1973                     }
1974                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1975                         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]);
1976                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1977                         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]);
1978                     h->h264dsp.h264_idct_add8(dest, block_offset,
1979                                               h->mb, uvlinesize,
1980                                               h->non_zero_count_cache);
1981                 }
1982 #if CONFIG_SVQ3_DECODER
1983                 else{
1984                     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]);
1985                     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]);
1986                     for(j=1; j<3; j++){
1987                         for(i=j*16; i<j*16+4; i++){
1988                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1989                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
1990                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1991                             }
1992                         }
1993                     }
1994                 }
1995 #endif
1996             }
1997         }
1998     }
1999     if(h->cbp || IS_INTRA(mb_type))
2000     {
2001         s->dsp.clear_blocks(h->mb);
2002         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2003     }
2004 }
2005
2006 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
2007     MpegEncContext * const s = &h->s;
2008     const int mb_x= s->mb_x;
2009     const int mb_y= s->mb_y;
2010     const int mb_xy= h->mb_xy;
2011     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2012     uint8_t  *dest[3];
2013     int linesize;
2014     int i, j, p;
2015     int *block_offset = &h->block_offset[0];
2016     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2017     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
2018
2019     for (p = 0; p < plane_count; p++)
2020     {
2021         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
2022         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
2023     }
2024
2025     h->list_counts[mb_xy]= h->list_count;
2026
2027     if (!simple && MB_FIELD) {
2028         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
2029         block_offset = &h->block_offset[48];
2030         if(mb_y&1) //FIXME move out of this function?
2031             for (p = 0; p < 3; p++)
2032                 dest[p] -= s->linesize*15;
2033         if(FRAME_MBAFF) {
2034             int list;
2035             for(list=0; list<h->list_count; list++){
2036                 if(!USES_LIST(mb_type, list))
2037                     continue;
2038                 if(IS_16X16(mb_type)){
2039                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2040                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2041                 }else{
2042                     for(i=0; i<16; i+=4){
2043                         int ref = h->ref_cache[list][scan8[i]];
2044                         if(ref >= 0)
2045                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2046                     }
2047                 }
2048             }
2049         }
2050     } else {
2051         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2052     }
2053
2054     if (!simple && IS_INTRA_PCM(mb_type)) {
2055         if (pixel_shift) {
2056             const int bit_depth = h->sps.bit_depth_luma;
2057             GetBitContext gb;
2058             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
2059
2060             for (p = 0; p < plane_count; p++) {
2061                 for (i = 0; i < 16; i++) {
2062                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2063                     for (j = 0; j < 16; j++)
2064                         tmp[j] = get_bits(&gb, bit_depth);
2065                 }
2066             }
2067         } else {
2068             for (p = 0; p < plane_count; p++) {
2069                 for (i = 0; i < 16; i++) {
2070                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2071                 }
2072             }
2073         }
2074     } else {
2075         if(IS_INTRA(mb_type)){
2076             if(h->deblocking_filter)
2077                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2078
2079             for (p = 0; p < plane_count; p++)
2080                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2081
2082             if(h->deblocking_filter)
2083                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2084         }else{
2085             hl_motion(h, dest[0], dest[1], dest[2],
2086                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2087                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2088                       h->h264dsp.weight_h264_pixels_tab,
2089                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 1);
2090         }
2091
2092         for (p = 0; p < plane_count; p++)
2093             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2094     }
2095     if(h->cbp || IS_INTRA(mb_type))
2096     {
2097         s->dsp.clear_blocks(h->mb);
2098         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2099     }
2100 }
2101
2102 /**
2103  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2104  */
2105 #define hl_decode_mb_simple(sh, bits) \
2106 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2107     hl_decode_mb_internal(h, 1, sh); \
2108 }
2109 hl_decode_mb_simple(0, 8);
2110 hl_decode_mb_simple(1, 16);
2111
2112 /**
2113  * Process a macroblock; this handles edge cases, such as interlacing.
2114  */
2115 static void av_noinline hl_decode_mb_complex(H264Context *h){
2116     hl_decode_mb_internal(h, 0, h->pixel_shift);
2117 }
2118
2119 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2120     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2121 }
2122
2123 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2124     hl_decode_mb_444_internal(h, 1, 0);
2125 }
2126
2127 void ff_h264_hl_decode_mb(H264Context *h){
2128     MpegEncContext * const s = &h->s;
2129     const int mb_xy= h->mb_xy;
2130     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2131     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2132
2133     if (CHROMA444) {
2134         if(is_complex || h->pixel_shift)
2135             hl_decode_mb_444_complex(h);
2136         else
2137             hl_decode_mb_444_simple(h);
2138     } else if (is_complex) {
2139         hl_decode_mb_complex(h);
2140     } else if (h->pixel_shift) {
2141         hl_decode_mb_simple_16(h);
2142     } else
2143         hl_decode_mb_simple_8(h);
2144 }
2145
2146 static int pred_weight_table(H264Context *h){
2147     MpegEncContext * const s = &h->s;
2148     int list, i;
2149     int luma_def, chroma_def;
2150
2151     h->use_weight= 0;
2152     h->use_weight_chroma= 0;
2153     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2154     if(h->sps.chroma_format_idc)
2155         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2156     luma_def = 1<<h->luma_log2_weight_denom;
2157     chroma_def = 1<<h->chroma_log2_weight_denom;
2158
2159     for(list=0; list<2; list++){
2160         h->luma_weight_flag[list]   = 0;
2161         h->chroma_weight_flag[list] = 0;
2162         for(i=0; i<h->ref_count[list]; i++){
2163             int luma_weight_flag, chroma_weight_flag;
2164
2165             luma_weight_flag= get_bits1(&s->gb);
2166             if(luma_weight_flag){
2167                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2168                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2169                 if(   h->luma_weight[i][list][0] != luma_def
2170                    || h->luma_weight[i][list][1] != 0) {
2171                     h->use_weight= 1;
2172                     h->luma_weight_flag[list]= 1;
2173                 }
2174             }else{
2175                 h->luma_weight[i][list][0]= luma_def;
2176                 h->luma_weight[i][list][1]= 0;
2177             }
2178
2179             if(h->sps.chroma_format_idc){
2180                 chroma_weight_flag= get_bits1(&s->gb);
2181                 if(chroma_weight_flag){
2182                     int j;
2183                     for(j=0; j<2; j++){
2184                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2185                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2186                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2187                            || h->chroma_weight[i][list][j][1] != 0) {
2188                             h->use_weight_chroma= 1;
2189                             h->chroma_weight_flag[list]= 1;
2190                         }
2191                     }
2192                 }else{
2193                     int j;
2194                     for(j=0; j<2; j++){
2195                         h->chroma_weight[i][list][j][0]= chroma_def;
2196                         h->chroma_weight[i][list][j][1]= 0;
2197                     }
2198                 }
2199             }
2200         }
2201         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2202     }
2203     h->use_weight= h->use_weight || h->use_weight_chroma;
2204     return 0;
2205 }
2206
2207 /**
2208  * Initialize implicit_weight table.
2209  * @param field  0/1 initialize the weight for interlaced MBAFF
2210  *                -1 initializes the rest
2211  */
2212 static void implicit_weight_table(H264Context *h, int field){
2213     MpegEncContext * const s = &h->s;
2214     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2215
2216     for (i = 0; i < 2; i++) {
2217         h->luma_weight_flag[i]   = 0;
2218         h->chroma_weight_flag[i] = 0;
2219     }
2220
2221     if(field < 0){
2222         if (s->picture_structure == PICT_FRAME) {
2223             cur_poc = s->current_picture_ptr->poc;
2224         } else {
2225             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2226         }
2227     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2228        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2229         h->use_weight= 0;
2230         h->use_weight_chroma= 0;
2231         return;
2232     }
2233         ref_start= 0;
2234         ref_count0= h->ref_count[0];
2235         ref_count1= h->ref_count[1];
2236     }else{
2237         cur_poc = s->current_picture_ptr->field_poc[field];
2238         ref_start= 16;
2239         ref_count0= 16+2*h->ref_count[0];
2240         ref_count1= 16+2*h->ref_count[1];
2241     }
2242
2243     h->use_weight= 2;
2244     h->use_weight_chroma= 2;
2245     h->luma_log2_weight_denom= 5;
2246     h->chroma_log2_weight_denom= 5;
2247
2248     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2249         int poc0 = h->ref_list[0][ref0].poc;
2250         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2251             int w = 32;
2252             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2253                 int poc1 = h->ref_list[1][ref1].poc;
2254                 int td = av_clip(poc1 - poc0, -128, 127);
2255                 if(td){
2256                     int tb = av_clip(cur_poc - poc0, -128, 127);
2257                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2258                     int dist_scale_factor = (tb*tx + 32) >> 8;
2259                     if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2260                         w = 64 - dist_scale_factor;
2261                 }
2262             }
2263             if(field<0){
2264                 h->implicit_weight[ref0][ref1][0]=
2265                 h->implicit_weight[ref0][ref1][1]= w;
2266             }else{
2267                 h->implicit_weight[ref0][ref1][field]=w;
2268             }
2269         }
2270     }
2271 }
2272
2273 /**
2274  * instantaneous decoder refresh.
2275  */
2276 static void idr(H264Context *h){
2277     ff_h264_remove_all_refs(h);
2278     h->prev_frame_num= 0;
2279     h->prev_frame_num_offset= 0;
2280     h->prev_poc_msb=
2281     h->prev_poc_lsb= 0;
2282 }
2283
2284 /* forget old pics after a seek */
2285 static void flush_dpb(AVCodecContext *avctx){
2286     H264Context *h= avctx->priv_data;
2287     int i;
2288     for(i=0; i<=MAX_DELAYED_PIC_COUNT; i++) {
2289         if(h->delayed_pic[i])
2290             h->delayed_pic[i]->f.reference = 0;
2291         h->delayed_pic[i]= NULL;
2292     }
2293     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2294     h->prev_interlaced_frame = 1;
2295     idr(h);
2296     if(h->s.current_picture_ptr)
2297         h->s.current_picture_ptr->f.reference = 0;
2298     h->s.first_field= 0;
2299     ff_h264_reset_sei(h);
2300     ff_mpeg_flush(avctx);
2301     h->recovery_frame= -1;
2302     h->sync= 0;
2303 }
2304
2305 static int init_poc(H264Context *h){
2306     MpegEncContext * const s = &h->s;
2307     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2308     int field_poc[2];
2309     Picture *cur = s->current_picture_ptr;
2310
2311     h->frame_num_offset= h->prev_frame_num_offset;
2312     if(h->frame_num < h->prev_frame_num)
2313         h->frame_num_offset += max_frame_num;
2314
2315     if(h->sps.poc_type==0){
2316         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2317
2318         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2319             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2320         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2321             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2322         else
2323             h->poc_msb = h->prev_poc_msb;
2324 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2325         field_poc[0] =
2326         field_poc[1] = h->poc_msb + h->poc_lsb;
2327         if(s->picture_structure == PICT_FRAME)
2328             field_poc[1] += h->delta_poc_bottom;
2329     }else if(h->sps.poc_type==1){
2330         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2331         int i;
2332
2333         if(h->sps.poc_cycle_length != 0)
2334             abs_frame_num = h->frame_num_offset + h->frame_num;
2335         else
2336             abs_frame_num = 0;
2337
2338         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2339             abs_frame_num--;
2340
2341         expected_delta_per_poc_cycle = 0;
2342         for(i=0; i < h->sps.poc_cycle_length; i++)
2343             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2344
2345         if(abs_frame_num > 0){
2346             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2347             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2348
2349             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2350             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2351                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2352         } else
2353             expectedpoc = 0;
2354
2355         if(h->nal_ref_idc == 0)
2356             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2357
2358         field_poc[0] = expectedpoc + h->delta_poc[0];
2359         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2360
2361         if(s->picture_structure == PICT_FRAME)
2362             field_poc[1] += h->delta_poc[1];
2363     }else{
2364         int poc= 2*(h->frame_num_offset + h->frame_num);
2365
2366         if(!h->nal_ref_idc)
2367             poc--;
2368
2369         field_poc[0]= poc;
2370         field_poc[1]= poc;
2371     }
2372
2373     if(s->picture_structure != PICT_BOTTOM_FIELD)
2374         s->current_picture_ptr->field_poc[0]= field_poc[0];
2375     if(s->picture_structure != PICT_TOP_FIELD)
2376         s->current_picture_ptr->field_poc[1]= field_poc[1];
2377     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2378
2379     return 0;
2380 }
2381
2382
2383 /**
2384  * initialize scan tables
2385  */
2386 static void init_scan_tables(H264Context *h){
2387     int i;
2388     for(i=0; i<16; i++){
2389 #define T(x) (x>>2) | ((x<<2) & 0xF)
2390         h->zigzag_scan[i] = T(zigzag_scan[i]);
2391         h-> field_scan[i] = T( field_scan[i]);
2392 #undef T
2393     }
2394     for(i=0; i<64; i++){
2395 #define T(x) (x>>3) | ((x&7)<<3)
2396         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2397         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2398         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2399         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2400 #undef T
2401     }
2402     if(h->sps.transform_bypass){ //FIXME same ugly
2403         h->zigzag_scan_q0          = zigzag_scan;
2404         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2405         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2406         h->field_scan_q0           = field_scan;
2407         h->field_scan8x8_q0        = field_scan8x8;
2408         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2409     }else{
2410         h->zigzag_scan_q0          = h->zigzag_scan;
2411         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2412         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2413         h->field_scan_q0           = h->field_scan;
2414         h->field_scan8x8_q0        = h->field_scan8x8;
2415         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2416     }
2417 }
2418
2419 static int field_end(H264Context *h, int in_setup){
2420     MpegEncContext * const s = &h->s;
2421     AVCodecContext * const avctx= s->avctx;
2422     int err = 0;
2423     s->mb_y= 0;
2424
2425     if (!in_setup && !s->dropable)
2426         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2427                                  s->picture_structure==PICT_BOTTOM_FIELD);
2428
2429     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2430         ff_vdpau_h264_set_reference_frames(s);
2431
2432     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2433         if(!s->dropable) {
2434             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2435             h->prev_poc_msb= h->poc_msb;
2436             h->prev_poc_lsb= h->poc_lsb;
2437         }
2438         h->prev_frame_num_offset= h->frame_num_offset;
2439         h->prev_frame_num= h->frame_num;
2440         h->outputed_poc = h->next_outputed_poc;
2441     }
2442
2443     if (avctx->hwaccel) {
2444         if (avctx->hwaccel->end_frame(avctx) < 0)
2445             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2446     }
2447
2448     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2449         ff_vdpau_h264_picture_complete(s);
2450
2451     /*
2452      * FIXME: Error handling code does not seem to support interlaced
2453      * when slices span multiple rows
2454      * The ff_er_add_slice calls don't work right for bottom
2455      * fields; they cause massive erroneous error concealing
2456      * Error marking covers both fields (top and bottom).
2457      * This causes a mismatched s->error_count
2458      * and a bad error table. Further, the error count goes to
2459      * INT_MAX when called for bottom field, because mb_y is
2460      * past end by one (callers fault) and resync_mb_y != 0
2461      * causes problems for the first MB line, too.
2462      */
2463     if (!FIELD_PICTURE)
2464         ff_er_frame_end(s);
2465
2466     MPV_frame_end(s);
2467
2468     h->current_slice=0;
2469
2470     return err;
2471 }
2472
2473 /**
2474  * Replicate H264 "master" context to thread contexts.
2475  */
2476 static void clone_slice(H264Context *dst, H264Context *src)
2477 {
2478     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2479     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2480     dst->s.current_picture      = src->s.current_picture;
2481     dst->s.linesize             = src->s.linesize;
2482     dst->s.uvlinesize           = src->s.uvlinesize;
2483     dst->s.first_field          = src->s.first_field;
2484
2485     dst->prev_poc_msb           = src->prev_poc_msb;
2486     dst->prev_poc_lsb           = src->prev_poc_lsb;
2487     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2488     dst->prev_frame_num         = src->prev_frame_num;
2489     dst->short_ref_count        = src->short_ref_count;
2490
2491     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2492     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2493     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2494     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2495
2496     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2497     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2498 }
2499
2500 /**
2501  * computes profile from profile_idc and constraint_set?_flags
2502  *
2503  * @param sps SPS
2504  *
2505  * @return profile as defined by FF_PROFILE_H264_*
2506  */
2507 int ff_h264_get_profile(SPS *sps)
2508 {
2509     int profile = sps->profile_idc;
2510
2511     switch(sps->profile_idc) {
2512     case FF_PROFILE_H264_BASELINE:
2513         // constraint_set1_flag set to 1
2514         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2515         break;
2516     case FF_PROFILE_H264_HIGH_10:
2517     case FF_PROFILE_H264_HIGH_422:
2518     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2519         // constraint_set3_flag set to 1
2520         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2521         break;
2522     }
2523
2524     return profile;
2525 }
2526
2527 /**
2528  * decodes a slice header.
2529  * This will also call MPV_common_init() and frame_start() as needed.
2530  *
2531  * @param h h264context
2532  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2533  *
2534  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2535  */
2536 static int decode_slice_header(H264Context *h, H264Context *h0){
2537     MpegEncContext * const s = &h->s;
2538     MpegEncContext * const s0 = &h0->s;
2539     unsigned int first_mb_in_slice;
2540     unsigned int pps_id;
2541     int num_ref_idx_active_override_flag;
2542     unsigned int slice_type, tmp, i, j;
2543     int default_ref_list_done = 0;
2544     int last_pic_structure;
2545
2546     s->dropable= h->nal_ref_idc == 0;
2547
2548     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2549     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2550         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2551         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2552     }else{
2553         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2554         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2555     }
2556
2557     first_mb_in_slice= get_ue_golomb(&s->gb);
2558
2559     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2560         if(h0->current_slice && FIELD_PICTURE){
2561             field_end(h, 1);
2562         }
2563
2564         h0->current_slice = 0;
2565         if (!s0->first_field)
2566             s->current_picture_ptr= NULL;
2567     }
2568
2569     slice_type= get_ue_golomb_31(&s->gb);
2570     if(slice_type > 9){
2571         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);
2572         return -1;
2573     }
2574     if(slice_type > 4){
2575         slice_type -= 5;
2576         h->slice_type_fixed=1;
2577     }else
2578         h->slice_type_fixed=0;
2579
2580     slice_type= golomb_to_pict_type[ slice_type ];
2581     if (slice_type == AV_PICTURE_TYPE_I
2582         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2583         default_ref_list_done = 1;
2584     }
2585     h->slice_type= slice_type;
2586     h->slice_type_nos= slice_type & 3;
2587
2588     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2589
2590     pps_id= get_ue_golomb(&s->gb);
2591     if(pps_id>=MAX_PPS_COUNT){
2592         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2593         return -1;
2594     }
2595     if(!h0->pps_buffers[pps_id]) {
2596         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2597         return -1;
2598     }
2599     h->pps= *h0->pps_buffers[pps_id];
2600
2601     if(!h0->sps_buffers[h->pps.sps_id]) {
2602         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2603         return -1;
2604     }
2605     h->sps = *h0->sps_buffers[h->pps.sps_id];
2606
2607     s->avctx->profile = ff_h264_get_profile(&h->sps);
2608     s->avctx->level   = h->sps.level_idc;
2609     s->avctx->refs    = h->sps.ref_frame_count;
2610
2611     if(h == h0 && h->dequant_coeff_pps != pps_id){
2612         h->dequant_coeff_pps = pps_id;
2613         init_dequant_tables(h);
2614     }
2615
2616     s->mb_width= h->sps.mb_width;
2617     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2618
2619     h->b_stride=  s->mb_width*4;
2620
2621     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2622
2623     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2624     if(h->sps.frame_mbs_only_flag)
2625         s->height= 16*s->mb_height - (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2626     else
2627         s->height= 16*s->mb_height - (2<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2628
2629     if (s->context_initialized
2630         && (   s->width != s->avctx->width || s->height != s->avctx->height
2631             || s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma
2632             || h->cur_chroma_format_idc != h->sps.chroma_format_idc
2633             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2634         if(h != h0) {
2635             av_log_missing_feature(s->avctx, "Width/height/bit depth/chroma idc changing with threads is", 0);
2636             return -1;   // width / height changed during parallelized decoding
2637         }
2638         free_tables(h, 0);
2639         flush_dpb(s->avctx);
2640         MPV_common_end(s);
2641         h->list_count = 0;
2642     }
2643     if (!s->context_initialized) {
2644         if (h != h0) {
2645             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2646             return -1;
2647         }
2648
2649         avcodec_set_dimensions(s->avctx, s->width, s->height);
2650         s->avctx->sample_aspect_ratio= h->sps.sar;
2651         av_assert0(s->avctx->sample_aspect_ratio.den);
2652
2653         if (s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
2654             h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
2655             if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10 &&
2656                 (h->sps.bit_depth_luma != 9 || !CHROMA422)) {
2657                 s->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
2658                 h->cur_chroma_format_idc = h->sps.chroma_format_idc;
2659                 h->pixel_shift = h->sps.bit_depth_luma > 8;
2660
2661                 ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2662                 ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2663                 s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
2664                 dsputil_init(&s->dsp, s->avctx);
2665             } else {
2666                 av_log(s->avctx, AV_LOG_DEBUG, "Unsupported bit depth: %d chroma_idc: %d\n",
2667                        h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2668                 return -1;
2669             }
2670         }
2671
2672         if(h->sps.video_signal_type_present_flag){
2673             s->avctx->color_range = h->sps.full_range>0 ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2674             if(h->sps.colour_description_present_flag){
2675                 s->avctx->color_primaries = h->sps.color_primaries;
2676                 s->avctx->color_trc       = h->sps.color_trc;
2677                 s->avctx->colorspace      = h->sps.colorspace;
2678             }
2679         }
2680
2681         if(h->sps.timing_info_present_flag){
2682             int64_t den= h->sps.time_scale;
2683             if(h->x264_build < 44U)
2684                 den *= 2;
2685             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2686                       h->sps.num_units_in_tick, den, 1<<30);
2687         }
2688
2689         switch (h->sps.bit_depth_luma) {
2690             case 9 :
2691                 if (CHROMA444)
2692                     s->avctx->pix_fmt = PIX_FMT_YUV444P9;
2693                 else
2694                     s->avctx->pix_fmt = PIX_FMT_YUV420P9;
2695                 break;
2696             case 10 :
2697                 if (CHROMA444)
2698                     s->avctx->pix_fmt = PIX_FMT_YUV444P10;
2699                 else if (CHROMA422)
2700                     s->avctx->pix_fmt = PIX_FMT_YUV422P10;
2701                 else
2702                     s->avctx->pix_fmt = PIX_FMT_YUV420P10;
2703                 break;
2704             default:
2705                 if (CHROMA444){
2706                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2707                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2708                        s->avctx->pix_fmt = PIX_FMT_GBR24P;
2709                        av_log(h->s.avctx, AV_LOG_DEBUG, "Detected GBR colorspace.\n");
2710                     }
2711                 }else if (CHROMA422) {
2712                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P : PIX_FMT_YUV422P;
2713                 }else{
2714                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2715                                                              s->avctx->codec->pix_fmts ?
2716                                                              s->avctx->codec->pix_fmts :
2717                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2718                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2719                                                              ff_hwaccel_pixfmt_list_420);
2720                 }
2721         }
2722
2723         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2724
2725         if (MPV_common_init(s) < 0) {
2726             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2727             return -1;
2728         }
2729         s->first_field = 0;
2730         h->prev_interlaced_frame = 1;
2731
2732         init_scan_tables(h);
2733         if (ff_h264_alloc_tables(h) < 0) {
2734             av_log(h->s.avctx, AV_LOG_ERROR, "Could not allocate memory for h264\n");
2735             return AVERROR(ENOMEM);
2736         }
2737
2738         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2739             if (context_init(h) < 0) {
2740                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2741                 return -1;
2742             }
2743         } else {
2744             for(i = 1; i < s->avctx->thread_count; i++) {
2745                 H264Context *c;
2746                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2747                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2748                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2749                 c->h264dsp = h->h264dsp;
2750                 c->sps = h->sps;
2751                 c->pps = h->pps;
2752                 c->pixel_shift = h->pixel_shift;
2753                 init_scan_tables(c);
2754                 clone_tables(c, h, i);
2755             }
2756
2757             for(i = 0; i < s->avctx->thread_count; i++)
2758                 if (context_init(h->thread_context[i]) < 0) {
2759                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2760                     return -1;
2761                 }
2762         }
2763     }
2764
2765     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2766
2767     h->mb_mbaff = 0;
2768     h->mb_aff_frame = 0;
2769     last_pic_structure = s0->picture_structure;
2770     if(h->sps.frame_mbs_only_flag){
2771         s->picture_structure= PICT_FRAME;
2772     }else{
2773         if(!h->sps.direct_8x8_inference_flag && slice_type == AV_PICTURE_TYPE_B){
2774             av_log(h->s.avctx, AV_LOG_ERROR, "This stream was generated by a broken encoder, invalid 8x8 inference\n");
2775             return -1;
2776         }
2777         if(get_bits1(&s->gb)) { //field_pic_flag
2778             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2779         } else {
2780             s->picture_structure= PICT_FRAME;
2781             h->mb_aff_frame = h->sps.mb_aff;
2782         }
2783     }
2784     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2785
2786     if(h0->current_slice == 0){
2787         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2788         if(h->frame_num != h->prev_frame_num) {
2789             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2790
2791             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2792
2793             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2794                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2795                 if (unwrap_prev_frame_num < 0)
2796                     unwrap_prev_frame_num += max_frame_num;
2797
2798                 h->prev_frame_num = unwrap_prev_frame_num;
2799             }
2800         }
2801
2802         while(h->frame_num !=  h->prev_frame_num &&
2803               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2804             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2805             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2806             if (ff_h264_frame_start(h) < 0)
2807                 return -1;
2808             h->prev_frame_num++;
2809             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2810             s->current_picture_ptr->frame_num= h->prev_frame_num;
2811             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2812             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2813             ff_generate_sliding_window_mmcos(h);
2814             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2815                 s->avctx->error_recognition >= FF_ER_EXPLODE)
2816                 return AVERROR_INVALIDDATA;
2817             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2818              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2819              * about there being no actual duplicates.
2820              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2821              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2822              * be fixed. */
2823             if (h->short_ref_count) {
2824                 if (prev) {
2825                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2826                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2827                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2828                     h->short_ref[0]->poc = prev->poc+2;
2829                 }
2830                 h->short_ref[0]->frame_num = h->prev_frame_num;
2831             }
2832         }
2833
2834         /* See if we have a decoded first field looking for a pair... */
2835         if (s0->first_field) {
2836             assert(s0->current_picture_ptr);
2837             assert(s0->current_picture_ptr->f.data[0]);
2838             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2839
2840             /* figure out if we have a complementary field pair */
2841             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2842                 /*
2843                  * Previous field is unmatched. Don't display it, but let it
2844                  * remain for reference if marked as such.
2845                  */
2846                 s0->current_picture_ptr = NULL;
2847                 s0->first_field = FIELD_PICTURE;
2848
2849             } else {
2850                 if (h->nal_ref_idc &&
2851                         s0->current_picture_ptr->f.reference &&
2852                         s0->current_picture_ptr->frame_num != h->frame_num) {
2853                     /*
2854                      * This and previous field were reference, but had
2855                      * different frame_nums. Consider this field first in
2856                      * pair. Throw away previous field except for reference
2857                      * purposes.
2858                      */
2859                     s0->first_field = 1;
2860                     s0->current_picture_ptr = NULL;
2861
2862                 } else {
2863                     /* Second field in complementary pair */
2864                     s0->first_field = 0;
2865                 }
2866             }
2867
2868         } else {
2869             /* Frame or first field in a potentially complementary pair */
2870             assert(!s0->current_picture_ptr);
2871             s0->first_field = FIELD_PICTURE;
2872         }
2873
2874         if(!FIELD_PICTURE || s0->first_field) {
2875             if (ff_h264_frame_start(h) < 0) {
2876                 s0->first_field = 0;
2877                 return -1;
2878             }
2879         } else {
2880             ff_release_unused_pictures(s, 0);
2881         }
2882     }
2883     if(h != h0)
2884         clone_slice(h, h0);
2885
2886     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2887
2888     assert(s->mb_num == s->mb_width * s->mb_height);
2889     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2890        first_mb_in_slice                    >= s->mb_num){
2891         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2892         return -1;
2893     }
2894     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2895     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2896     if (s->picture_structure == PICT_BOTTOM_FIELD)
2897         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2898     assert(s->mb_y < s->mb_height);
2899
2900     if(s->picture_structure==PICT_FRAME){
2901         h->curr_pic_num=   h->frame_num;
2902         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2903     }else{
2904         h->curr_pic_num= 2*h->frame_num + 1;
2905         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2906     }
2907
2908     if(h->nal_unit_type == NAL_IDR_SLICE){
2909         get_ue_golomb(&s->gb); /* idr_pic_id */
2910     }
2911
2912     if(h->sps.poc_type==0){
2913         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2914
2915         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2916             h->delta_poc_bottom= get_se_golomb(&s->gb);
2917         }
2918     }
2919
2920     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2921         h->delta_poc[0]= get_se_golomb(&s->gb);
2922
2923         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2924             h->delta_poc[1]= get_se_golomb(&s->gb);
2925     }
2926
2927     init_poc(h);
2928
2929     if(h->pps.redundant_pic_cnt_present){
2930         h->redundant_pic_count= get_ue_golomb(&s->gb);
2931     }
2932
2933     //set defaults, might be overridden a few lines later
2934     h->ref_count[0]= h->pps.ref_count[0];
2935     h->ref_count[1]= h->pps.ref_count[1];
2936
2937     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2938         unsigned max= (16<<(s->picture_structure != PICT_FRAME))-1;
2939         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2940             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2941         }
2942         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2943
2944         if(num_ref_idx_active_override_flag){
2945             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2946             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2947                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2948
2949         }
2950         if(h->ref_count[0]-1 > max || h->ref_count[1]-1 > max){
2951             av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2952             h->ref_count[0]= h->ref_count[1]= 1;
2953             return -1;
2954         }
2955         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2956             h->list_count= 2;
2957         else
2958             h->list_count= 1;
2959     }else
2960         h->ref_count[1]= h->ref_count[0]= h->list_count= 0;
2961
2962     if(!default_ref_list_done){
2963         ff_h264_fill_default_ref_list(h);
2964     }
2965
2966     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0) {
2967         h->ref_count[1]= h->ref_count[0]= 0;
2968         return -1;
2969     }
2970
2971     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2972         s->last_picture_ptr= &h->ref_list[0][0];
2973         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2974     }
2975     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2976         s->next_picture_ptr= &h->ref_list[1][0];
2977         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2978     }
2979
2980     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2981        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2982         pred_weight_table(h);
2983     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2984         implicit_weight_table(h, -1);
2985     }else {
2986         h->use_weight = 0;
2987         for (i = 0; i < 2; i++) {
2988             h->luma_weight_flag[i]   = 0;
2989             h->chroma_weight_flag[i] = 0;
2990         }
2991     }
2992
2993     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
2994        s->avctx->error_recognition >= FF_ER_EXPLODE)
2995         return AVERROR_INVALIDDATA;
2996
2997     if(FRAME_MBAFF){
2998         ff_h264_fill_mbaff_ref_list(h);
2999
3000         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
3001             implicit_weight_table(h, 0);
3002             implicit_weight_table(h, 1);
3003         }
3004     }
3005
3006     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
3007         ff_h264_direct_dist_scale_factor(h);
3008     ff_h264_direct_ref_list_init(h);
3009
3010     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
3011         tmp = get_ue_golomb_31(&s->gb);
3012         if(tmp > 2){
3013             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3014             return -1;
3015         }
3016         h->cabac_init_idc= tmp;
3017     }
3018
3019     h->last_qscale_diff = 0;
3020     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3021     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
3022         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3023         return -1;
3024     }
3025     s->qscale= tmp;
3026     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3027     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3028     //FIXME qscale / qp ... stuff
3029     if(h->slice_type == AV_PICTURE_TYPE_SP){
3030         get_bits1(&s->gb); /* sp_for_switch_flag */
3031     }
3032     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
3033         get_se_golomb(&s->gb); /* slice_qs_delta */
3034     }
3035
3036     h->deblocking_filter = 1;
3037     h->slice_alpha_c0_offset = 52;
3038     h->slice_beta_offset = 52;
3039     if( h->pps.deblocking_filter_parameters_present ) {
3040         tmp= get_ue_golomb_31(&s->gb);
3041         if(tmp > 2){
3042             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
3043             return -1;
3044         }
3045         h->deblocking_filter= tmp;
3046         if(h->deblocking_filter < 2)
3047             h->deblocking_filter^= 1; // 1<->0
3048
3049         if( h->deblocking_filter ) {
3050             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
3051             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
3052             if(   h->slice_alpha_c0_offset > 104U
3053                || h->slice_beta_offset     > 104U){
3054                 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);
3055                 return -1;
3056             }
3057         }
3058     }
3059
3060     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
3061        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
3062        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
3063        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
3064         h->deblocking_filter= 0;
3065
3066     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
3067         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
3068             /* Cheat slightly for speed:
3069                Do not bother to deblock across slices. */
3070             h->deblocking_filter = 2;
3071         } else {
3072             h0->max_contexts = 1;
3073             if(!h0->single_decode_warning) {
3074                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3075                 h0->single_decode_warning = 1;
3076             }
3077             if (h != h0) {
3078                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
3079                 return 1;
3080             }
3081         }
3082     }
3083     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
3084                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
3085                  + 6 * (h->sps.bit_depth_luma - 8);
3086
3087 #if 0 //FMO
3088     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
3089         slice_group_change_cycle= get_bits(&s->gb, ?);
3090 #endif
3091
3092     h0->last_slice_type = slice_type;
3093     h->slice_num = ++h0->current_slice;
3094
3095     if(h->slice_num)
3096         h0->slice_row[(h->slice_num-1)&(MAX_SLICES-1)]= s->resync_mb_y;
3097     if (   h0->slice_row[h->slice_num&(MAX_SLICES-1)] + 3 >= s->resync_mb_y
3098         && h0->slice_row[h->slice_num&(MAX_SLICES-1)] <= s->resync_mb_y
3099         && h->slice_num >= MAX_SLICES) {
3100         //in case of ASO this check needs to be updated depending on how we decide to assign slice numbers in this case
3101         av_log(s->avctx, AV_LOG_WARNING, "Possibly too many slices (%d >= %d), increase MAX_SLICES and recompile if there are artifacts\n", h->slice_num, MAX_SLICES);
3102     }
3103
3104     for(j=0; j<2; j++){
3105         int id_list[16];
3106         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
3107         for(i=0; i<16; i++){
3108             id_list[i]= 60;
3109             if (h->ref_list[j][i].f.data[0]) {
3110                 int k;
3111                 uint8_t *base = h->ref_list[j][i].f.base[0];
3112                 for(k=0; k<h->short_ref_count; k++)
3113                     if (h->short_ref[k]->f.base[0] == base) {
3114                         id_list[i]= k;
3115                         break;
3116                     }
3117                 for(k=0; k<h->long_ref_count; k++)
3118                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3119                         id_list[i]= h->short_ref_count + k;
3120                         break;
3121                     }
3122             }
3123         }
3124
3125         ref2frm[0]=
3126         ref2frm[1]= -1;
3127         for(i=0; i<16; i++)
3128             ref2frm[i+2]= 4*id_list[i]
3129                           + (h->ref_list[j][i].f.reference & 3);
3130         ref2frm[18+0]=
3131         ref2frm[18+1]= -1;
3132         for(i=16; i<48; i++)
3133             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3134                           + (h->ref_list[j][i].f.reference & 3);
3135     }
3136
3137     //FIXME: fix draw_edges+PAFF+frame threads
3138     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3139     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3140
3141     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3142         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",
3143                h->slice_num,
3144                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3145                first_mb_in_slice,
3146                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3147                pps_id, h->frame_num,
3148                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3149                h->ref_count[0], h->ref_count[1],
3150                s->qscale,
3151                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3152                h->use_weight,
3153                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3154                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3155                );
3156     }
3157
3158     return 0;
3159 }
3160
3161 int ff_h264_get_slice_type(const H264Context *h)
3162 {
3163     switch (h->slice_type) {
3164     case AV_PICTURE_TYPE_P:  return 0;
3165     case AV_PICTURE_TYPE_B:  return 1;
3166     case AV_PICTURE_TYPE_I:  return 2;
3167     case AV_PICTURE_TYPE_SP: return 3;
3168     case AV_PICTURE_TYPE_SI: return 4;
3169     default:         return -1;
3170     }
3171 }
3172
3173 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3174                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3175 {
3176     int b_stride = h->b_stride;
3177     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3178     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3179     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3180         if(USES_LIST(top_type, list)){
3181             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3182             const int b8_xy= 4*top_xy + 2;
3183             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3184             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3185             ref_cache[0 - 1*8]=
3186             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3187             ref_cache[2 - 1*8]=
3188             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3189         }else{
3190             AV_ZERO128(mv_dst - 1*8);
3191             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3192         }
3193
3194         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3195             if(USES_LIST(left_type[LTOP], list)){
3196                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3197                 const int b8_xy= 4*left_xy[LTOP] + 1;
3198                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3199                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3200                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3201                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3202                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3203                 ref_cache[-1 +  0]=
3204                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3205                 ref_cache[-1 + 16]=
3206                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3207             }else{
3208                 AV_ZERO32(mv_dst - 1 + 0);
3209                 AV_ZERO32(mv_dst - 1 + 8);
3210                 AV_ZERO32(mv_dst - 1 +16);
3211                 AV_ZERO32(mv_dst - 1 +24);
3212                 ref_cache[-1 +  0]=
3213                 ref_cache[-1 +  8]=
3214                 ref_cache[-1 + 16]=
3215                 ref_cache[-1 + 24]= LIST_NOT_USED;
3216             }
3217         }
3218     }
3219
3220     if(!USES_LIST(mb_type, list)){
3221         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3222         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3223         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3224         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3225         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3226         return;
3227     }
3228
3229     {
3230         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3231         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3232         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3233         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3234         AV_WN32A(&ref_cache[0*8], ref01);
3235         AV_WN32A(&ref_cache[1*8], ref01);
3236         AV_WN32A(&ref_cache[2*8], ref23);
3237         AV_WN32A(&ref_cache[3*8], ref23);
3238     }
3239
3240     {
3241         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3242         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3243         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3244         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3245         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3246     }
3247 }
3248
3249 /**
3250  *
3251  * @return non zero if the loop filter can be skiped
3252  */
3253 static int fill_filter_caches(H264Context *h, int mb_type){
3254     MpegEncContext * const s = &h->s;
3255     const int mb_xy= h->mb_xy;
3256     int top_xy, left_xy[LEFT_MBS];
3257     int top_type, left_type[LEFT_MBS];
3258     uint8_t *nnz;
3259     uint8_t *nnz_cache;
3260
3261     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3262
3263     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3264      * stuff, I can't imagine that these complex rules are worth it. */
3265
3266     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3267     if(FRAME_MBAFF){
3268         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3269         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3270         if(s->mb_y&1){
3271             if (left_mb_field_flag != curr_mb_field_flag) {
3272                 left_xy[LTOP] -= s->mb_stride;
3273             }
3274         }else{
3275             if(curr_mb_field_flag){
3276                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3277             }
3278             if (left_mb_field_flag != curr_mb_field_flag) {
3279                 left_xy[LBOT] += s->mb_stride;
3280             }
3281         }
3282     }
3283
3284     h->top_mb_xy = top_xy;
3285     h->left_mb_xy[LTOP] = left_xy[LTOP];
3286     h->left_mb_xy[LBOT] = left_xy[LBOT];
3287     {
3288         //for sufficiently low qp, filtering wouldn't do anything
3289         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3290         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3291         int qp = s->current_picture.f.qscale_table[mb_xy];
3292         if(qp <= qp_thresh
3293            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3294            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3295             if(!FRAME_MBAFF)
3296                 return 1;
3297             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3298                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3299                 return 1;
3300         }
3301     }
3302
3303     top_type        = s->current_picture.f.mb_type[top_xy];
3304     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3305     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3306     if(h->deblocking_filter == 2){
3307         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3308         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3309     }else{
3310         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3311         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3312     }
3313     h->top_type       = top_type;
3314     h->left_type[LTOP]= left_type[LTOP];
3315     h->left_type[LBOT]= left_type[LBOT];
3316
3317     if(IS_INTRA(mb_type))
3318         return 0;
3319
3320     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3321     if(h->list_count == 2)
3322         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3323
3324     nnz = h->non_zero_count[mb_xy];
3325     nnz_cache = h->non_zero_count_cache;
3326     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3327     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3328     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3329     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3330     h->cbp= h->cbp_table[mb_xy];
3331
3332     if(top_type){
3333         nnz = h->non_zero_count[top_xy];
3334         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3335     }
3336
3337     if(left_type[LTOP]){
3338         nnz = h->non_zero_count[left_xy[LTOP]];
3339         nnz_cache[3+8*1]= nnz[3+0*4];
3340         nnz_cache[3+8*2]= nnz[3+1*4];
3341         nnz_cache[3+8*3]= nnz[3+2*4];
3342         nnz_cache[3+8*4]= nnz[3+3*4];
3343     }
3344
3345     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3346     if(!CABAC && h->pps.transform_8x8_mode){
3347         if(IS_8x8DCT(top_type)){
3348             nnz_cache[4+8*0]=
3349             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3350             nnz_cache[6+8*0]=
3351             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3352         }
3353         if(IS_8x8DCT(left_type[LTOP])){
3354             nnz_cache[3+8*1]=
3355             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3356         }
3357         if(IS_8x8DCT(left_type[LBOT])){
3358             nnz_cache[3+8*3]=
3359             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3360         }
3361
3362         if(IS_8x8DCT(mb_type)){
3363             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3364             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3365
3366             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3367             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3368
3369             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3370             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3371
3372             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3373             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3374         }
3375     }
3376
3377     return 0;
3378 }
3379
3380 static void loop_filter(H264Context *h, int start_x, int end_x){
3381     MpegEncContext * const s = &h->s;
3382     uint8_t  *dest_y, *dest_cb, *dest_cr;
3383     int linesize, uvlinesize, mb_x, mb_y;
3384     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3385     const int old_slice_type= h->slice_type;
3386     const int pixel_shift = h->pixel_shift;
3387     const int block_h = 16>>s->chroma_y_shift;
3388
3389     if(h->deblocking_filter) {
3390         for(mb_x= start_x; mb_x<end_x; mb_x++){
3391             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3392                 int mb_xy, mb_type;
3393                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3394                 h->slice_num= h->slice_table[mb_xy];
3395                 mb_type = s->current_picture.f.mb_type[mb_xy];
3396                 h->list_count= h->list_counts[mb_xy];
3397
3398                 if(FRAME_MBAFF)
3399                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3400
3401                 s->mb_x= mb_x;
3402                 s->mb_y= mb_y;
3403                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3404                 dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*(8<<CHROMA444) + mb_y * s->uvlinesize * block_h;
3405                 dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*(8<<CHROMA444) + mb_y * s->uvlinesize * block_h;
3406                     //FIXME simplify above
3407
3408                 if (MB_FIELD) {
3409                     linesize   = h->mb_linesize   = s->linesize * 2;
3410                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3411                     if(mb_y&1){ //FIXME move out of this function?
3412                         dest_y -= s->linesize*15;
3413                         dest_cb-= s->uvlinesize*(block_h-1);
3414                         dest_cr-= s->uvlinesize*(block_h-1);
3415                     }
3416                 } else {
3417                     linesize   = h->mb_linesize   = s->linesize;
3418                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3419                 }
3420                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
3421                 if(fill_filter_caches(h, mb_type))
3422                     continue;
3423                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3424                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3425
3426                 if (FRAME_MBAFF) {
3427                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3428                 } else {
3429                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3430                 }
3431             }
3432         }
3433     }
3434     h->slice_type= old_slice_type;
3435     s->mb_x= end_x;
3436     s->mb_y= end_mb_y - FRAME_MBAFF;
3437     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3438     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3439 }
3440
3441 static void predict_field_decoding_flag(H264Context *h){
3442     MpegEncContext * const s = &h->s;
3443     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3444     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3445                 ? s->current_picture.f.mb_type[mb_xy - 1]
3446                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3447                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3448                 : 0;
3449     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3450 }
3451
3452 /**
3453  * Draw edges and report progress for the last MB row.
3454  */
3455 static void decode_finish_row(H264Context *h){
3456     MpegEncContext * const s = &h->s;
3457     int top = 16*(s->mb_y >> FIELD_PICTURE);
3458     int height = 16 << FRAME_MBAFF;
3459     int deb