f61f524508900518ceb382bccb093a1959860711
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of Libav.
6  *
7  * Libav is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * Libav is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with Libav; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "libavutil/imgutils.h"
29 #include "internal.h"
30 #include "dsputil.h"
31 #include "avcodec.h"
32 #include "mpegvideo.h"
33 #include "h264.h"
34 #include "h264data.h"
35 #include "h264_mvpred.h"
36 #include "golomb.h"
37 #include "mathops.h"
38 #include "rectangle.h"
39 #include "thread.h"
40 #include "vdpau_internal.h"
41 #include "libavutil/avassert.h"
42
43 #include "cabac.h"
44
45 //#undef NDEBUG
46 #include <assert.h>
47
48 static const uint8_t rem6[QP_MAX_NUM+1]={
49 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
50 };
51
52 static const uint8_t div6[QP_MAX_NUM+1]={
53 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9,10,10,10,10,
54 };
55
56 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
57     PIX_FMT_DXVA2_VLD,
58     PIX_FMT_VAAPI_VLD,
59     PIX_FMT_YUVJ420P,
60     PIX_FMT_NONE
61 };
62
63 /**
64  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
65  */
66 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
67     MpegEncContext * const s = &h->s;
68     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
69     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
70     int i;
71
72     if(!(h->top_samples_available&0x8000)){
73         for(i=0; i<4; i++){
74             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
75             if(status<0){
76                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
77                 return -1;
78             } else if(status){
79                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
80             }
81         }
82     }
83
84     if((h->left_samples_available&0x8888)!=0x8888){
85         static const int mask[4]={0x8000,0x2000,0x80,0x20};
86         for(i=0; i<4; i++){
87             if(!(h->left_samples_available&mask[i])){
88                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
89                 if(status<0){
90                     av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
91                     return -1;
92                 } else if(status){
93                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
94                 }
95             }
96         }
97     }
98
99     return 0;
100 } //FIXME cleanup like ff_h264_check_intra_pred_mode
101
102 /**
103  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
104  */
105 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
106     MpegEncContext * const s = &h->s;
107     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
108     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
109
110     if(mode > 6U) {
111         av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
112         return -1;
113     }
114
115     if(!(h->top_samples_available&0x8000)){
116         mode= top[ mode ];
117         if(mode<0){
118             av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
119             return -1;
120         }
121     }
122
123     if((h->left_samples_available&0x8080) != 0x8080){
124         mode= left[ mode ];
125         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
126             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
127         }
128         if(mode<0){
129             av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
130             return -1;
131         }
132     }
133
134     return mode;
135 }
136
137 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
138     int i, si, di;
139     uint8_t *dst;
140     int bufidx;
141
142 //    src[0]&0x80;                //forbidden bit
143     h->nal_ref_idc= src[0]>>5;
144     h->nal_unit_type= src[0]&0x1F;
145
146     src++; length--;
147
148 #if HAVE_FAST_UNALIGNED
149 # if HAVE_FAST_64BIT
150 #   define RS 7
151     for(i=0; i+1<length; i+=9){
152         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
153 # else
154 #   define RS 3
155     for(i=0; i+1<length; i+=5){
156         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
157 # endif
158             continue;
159         if(i>0 && !src[i]) i--;
160         while(src[i]) i++;
161 #else
162 #   define RS 0
163     for(i=0; i+1<length; i+=2){
164         if(src[i]) continue;
165         if(i>0 && src[i-1]==0) i--;
166 #endif
167         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
168             if(src[i+2]!=3){
169                 /* startcode, so we must be past the end */
170                 length=i;
171             }
172             break;
173         }
174         i-= RS;
175     }
176
177     if(i>=length-1){ //no escaped 0
178         *dst_length= length;
179         *consumed= length+1; //+1 for the header
180         return src;
181     }
182
183     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
184     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
185     dst= h->rbsp_buffer[bufidx];
186
187     if (dst == NULL){
188         return NULL;
189     }
190
191 //printf("decoding esc\n");
192     memcpy(dst, src, i);
193     si=di=i;
194     while(si+2<length){
195         //remove escapes (very rare 1:2^22)
196         if(src[si+2]>3){
197             dst[di++]= src[si++];
198             dst[di++]= src[si++];
199         }else if(src[si]==0 && src[si+1]==0){
200             if(src[si+2]==3){ //escape
201                 dst[di++]= 0;
202                 dst[di++]= 0;
203                 si+=3;
204                 continue;
205             }else //next start code
206                 goto nsc;
207         }
208
209         dst[di++]= src[si++];
210     }
211     while(si<length)
212         dst[di++]= src[si++];
213 nsc:
214
215     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
216
217     *dst_length= di;
218     *consumed= si + 1;//+1 for the header
219 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
220     return dst;
221 }
222
223 /**
224  * Identify the exact end of the bitstream
225  * @return the length of the trailing, or 0 if damaged
226  */
227 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
228     int v= *src;
229     int r;
230
231     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
232
233     for(r=1; r<9; r++){
234         if(v&1) return r;
235         v>>=1;
236     }
237     return 0;
238 }
239
240 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
241                                  int y_offset, int list){
242     int raw_my= h->mv_cache[list][ scan8[n] ][1];
243     int filter_height= (raw_my&3) ? 2 : 0;
244     int full_my= (raw_my>>2) + y_offset;
245     int top = full_my - filter_height, bottom = full_my + height + filter_height;
246
247     return FFMAX(abs(top), bottom);
248 }
249
250 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
251                                int y_offset, int list0, int list1, int *nrefs){
252     MpegEncContext * const s = &h->s;
253     int my;
254
255     y_offset += 16*(s->mb_y >> MB_FIELD);
256
257     if(list0){
258         int ref_n = h->ref_cache[0][ scan8[n] ];
259         Picture *ref= &h->ref_list[0][ref_n];
260
261         // Error resilience puts the current picture in the ref list.
262         // Don't try to wait on these as it will cause a deadlock.
263         // Fields can wait on each other, though.
264         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
265            (ref->f.reference & 3) != s->picture_structure) {
266             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
267             if (refs[0][ref_n] < 0) nrefs[0] += 1;
268             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
269         }
270     }
271
272     if(list1){
273         int ref_n = h->ref_cache[1][ scan8[n] ];
274         Picture *ref= &h->ref_list[1][ref_n];
275
276         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
277            (ref->f.reference & 3) != s->picture_structure) {
278             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
279             if (refs[1][ref_n] < 0) nrefs[1] += 1;
280             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
281         }
282     }
283 }
284
285 /**
286  * Wait until all reference frames are available for MC operations.
287  *
288  * @param h the H264 context
289  */
290 static void await_references(H264Context *h){
291     MpegEncContext * const s = &h->s;
292     const int mb_xy= h->mb_xy;
293     const int mb_type = s->current_picture.f.mb_type[mb_xy];
294     int refs[2][48];
295     int nrefs[2] = {0};
296     int ref, list;
297
298     memset(refs, -1, sizeof(refs));
299
300     if(IS_16X16(mb_type)){
301         get_lowest_part_y(h, refs, 0, 16, 0,
302                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
303     }else if(IS_16X8(mb_type)){
304         get_lowest_part_y(h, refs, 0, 8, 0,
305                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
306         get_lowest_part_y(h, refs, 8, 8, 8,
307                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
308     }else if(IS_8X16(mb_type)){
309         get_lowest_part_y(h, refs, 0, 16, 0,
310                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
311         get_lowest_part_y(h, refs, 4, 16, 0,
312                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
313     }else{
314         int i;
315
316         assert(IS_8X8(mb_type));
317
318         for(i=0; i<4; i++){
319             const int sub_mb_type= h->sub_mb_type[i];
320             const int n= 4*i;
321             int y_offset= (i&2)<<2;
322
323             if(IS_SUB_8X8(sub_mb_type)){
324                 get_lowest_part_y(h, refs, n  , 8, y_offset,
325                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
326             }else if(IS_SUB_8X4(sub_mb_type)){
327                 get_lowest_part_y(h, refs, n  , 4, y_offset,
328                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
329                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
330                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
331             }else if(IS_SUB_4X8(sub_mb_type)){
332                 get_lowest_part_y(h, refs, n  , 8, y_offset,
333                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
334                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
335                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
336             }else{
337                 int j;
338                 assert(IS_SUB_4X4(sub_mb_type));
339                 for(j=0; j<4; j++){
340                     int sub_y_offset= y_offset + 2*(j&2);
341                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
342                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
343                 }
344             }
345         }
346     }
347
348     for(list=h->list_count-1; list>=0; list--){
349         for(ref=0; ref<48 && nrefs[list]; ref++){
350             int row = refs[list][ref];
351             if(row >= 0){
352                 Picture *ref_pic = &h->ref_list[list][ref];
353                 int ref_field = ref_pic->f.reference - 1;
354                 int ref_field_picture = ref_pic->field_picture;
355                 int pic_height = 16*s->mb_height >> ref_field_picture;
356
357                 row <<= MB_MBAFF;
358                 nrefs[list]--;
359
360                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
361                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
362                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
363                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
364                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
365                 }else if(FIELD_PICTURE){
366                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
367                 }else{
368                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
369                 }
370             }
371         }
372     }
373 }
374
375 #if 0
376 /**
377  * DCT transforms the 16 dc values.
378  * @param qp quantization parameter ??? FIXME
379  */
380 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
381 //    const int qmul= dequant_coeff[qp][0];
382     int i;
383     int temp[16]; //FIXME check if this is a good idea
384     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
385     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
386
387     for(i=0; i<4; i++){
388         const int offset= y_offset[i];
389         const int z0= block[offset+stride*0] + block[offset+stride*4];
390         const int z1= block[offset+stride*0] - block[offset+stride*4];
391         const int z2= block[offset+stride*1] - block[offset+stride*5];
392         const int z3= block[offset+stride*1] + block[offset+stride*5];
393
394         temp[4*i+0]= z0+z3;
395         temp[4*i+1]= z1+z2;
396         temp[4*i+2]= z1-z2;
397         temp[4*i+3]= z0-z3;
398     }
399
400     for(i=0; i<4; i++){
401         const int offset= x_offset[i];
402         const int z0= temp[4*0+i] + temp[4*2+i];
403         const int z1= temp[4*0+i] - temp[4*2+i];
404         const int z2= temp[4*1+i] - temp[4*3+i];
405         const int z3= temp[4*1+i] + temp[4*3+i];
406
407         block[stride*0 +offset]= (z0 + z3)>>1;
408         block[stride*2 +offset]= (z1 + z2)>>1;
409         block[stride*8 +offset]= (z1 - z2)>>1;
410         block[stride*10+offset]= (z0 - z3)>>1;
411     }
412 }
413 #endif
414
415 #undef xStride
416 #undef stride
417
418 #if 0
419 static void chroma_dc_dct_c(DCTELEM *block){
420     const int stride= 16*2;
421     const int xStride= 16;
422     int a,b,c,d,e;
423
424     a= block[stride*0 + xStride*0];
425     b= block[stride*0 + xStride*1];
426     c= block[stride*1 + xStride*0];
427     d= block[stride*1 + xStride*1];
428
429     e= a-b;
430     a= a+b;
431     b= c-d;
432     c= c+d;
433
434     block[stride*0 + xStride*0]= (a+c);
435     block[stride*0 + xStride*1]= (e+b);
436     block[stride*1 + xStride*0]= (a-c);
437     block[stride*1 + xStride*1]= (e-b);
438 }
439 #endif
440
441 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
442                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
443                            int src_x_offset, int src_y_offset,
444                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
445                            int pixel_shift, int chroma444){
446     MpegEncContext * const s = &h->s;
447     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
448     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
449     const int luma_xy= (mx&3) + ((my&3)<<2);
450     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
451     uint8_t * src_y = pic->f.data[0] + offset;
452     uint8_t * src_cb, * src_cr;
453     int extra_width= h->emu_edge_width;
454     int extra_height= h->emu_edge_height;
455     int emu=0;
456     const int full_mx= mx>>2;
457     const int full_my= my>>2;
458     const int pic_width  = 16*s->mb_width;
459     const int pic_height = 16*s->mb_height >> MB_FIELD;
460
461     if(mx&7) extra_width -= 3;
462     if(my&7) extra_height -= 3;
463
464     if(   full_mx < 0-extra_width
465        || full_my < 0-extra_height
466        || full_mx + 16/*FIXME*/ > pic_width + extra_width
467        || full_my + 16/*FIXME*/ > pic_height + extra_height){
468         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);
469             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
470         emu=1;
471     }
472
473     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
474     if(!square){
475         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
476     }
477
478     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
479
480     if(chroma444){
481         src_cb = pic->f.data[1] + offset;
482         if(emu){
483             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
484                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
485             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
486         }
487         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
488         if(!square){
489             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
490         }
491
492         src_cr = pic->f.data[2] + offset;
493         if(emu){
494             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
495                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
496             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
497         }
498         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
499         if(!square){
500             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
501         }
502         return;
503     }
504
505     if(MB_FIELD){
506         // chroma offset when predicting from a field of opposite parity
507         my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
508         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
509     }
510     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
511     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
512
513     if(emu){
514         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);
515             src_cb= s->edge_emu_buffer;
516     }
517     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
518
519     if(emu){
520         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);
521             src_cr= s->edge_emu_buffer;
522     }
523     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
524 }
525
526 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
527                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
528                            int x_offset, int y_offset,
529                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
530                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
531                            int list0, int list1, int pixel_shift, int chroma444){
532     MpegEncContext * const s = &h->s;
533     qpel_mc_func *qpix_op=  qpix_put;
534     h264_chroma_mc_func chroma_op= chroma_put;
535
536     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
537     if(chroma444){
538         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
539         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
540     }else{
541         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
542         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
543     }
544     x_offset += 8*s->mb_x;
545     y_offset += 8*(s->mb_y >> MB_FIELD);
546
547     if(list0){
548         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
549         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
550                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
551                            qpix_op, chroma_op, pixel_shift, chroma444);
552
553         qpix_op=  qpix_avg;
554         chroma_op= chroma_avg;
555     }
556
557     if(list1){
558         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
559         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
560                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
561                            qpix_op, chroma_op, pixel_shift, chroma444);
562     }
563 }
564
565 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
566                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
567                            int x_offset, int y_offset,
568                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
569                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
570                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
571                            int list0, int list1, int pixel_shift, int chroma444){
572     MpegEncContext * const s = &h->s;
573
574     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
575     if(chroma444){
576         chroma_weight_avg = luma_weight_avg;
577         chroma_weight_op = luma_weight_op;
578         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
579         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
580     }else{
581         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
582         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
583     }
584     x_offset += 8*s->mb_x;
585     y_offset += 8*(s->mb_y >> MB_FIELD);
586
587     if(list0 && list1){
588         /* don't optimize for luma-only case, since B-frames usually
589          * use implicit weights => chroma too. */
590         uint8_t *tmp_cb = s->obmc_scratchpad;
591         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
592         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
593         int refn0 = h->ref_cache[0][ scan8[n] ];
594         int refn1 = h->ref_cache[1][ scan8[n] ];
595
596         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
597                     dest_y, dest_cb, dest_cr,
598                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
599         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
600                     tmp_y, tmp_cb, tmp_cr,
601                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
602
603         if(h->use_weight == 2){
604             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
605             int weight1 = 64 - weight0;
606             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
607             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
608             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
609         }else{
610             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
611                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
612                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
613             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
614                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
615                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
616             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
617                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
618                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
619         }
620     }else{
621         int list = list1 ? 1 : 0;
622         int refn = h->ref_cache[list][ scan8[n] ];
623         Picture *ref= &h->ref_list[list][refn];
624         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
625                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
626                     qpix_put, chroma_put, pixel_shift, chroma444);
627
628         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
629                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
630         if(h->use_weight_chroma){
631             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
632                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
633             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
634                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
635         }
636     }
637 }
638
639 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
640                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
641                            int x_offset, int y_offset,
642                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
643                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
644                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
645                            int list0, int list1, int pixel_shift, int chroma444){
646     if((h->use_weight==2 && list0 && list1
647         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
648        || h->use_weight==1)
649         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
650                          x_offset, y_offset, qpix_put, chroma_put,
651                          weight_op[0], weight_op[3], weight_avg[0],
652                          weight_avg[3], list0, list1, pixel_shift, chroma444);
653     else
654         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
655                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
656                     chroma_avg, list0, list1, pixel_shift, chroma444);
657 }
658
659 static inline void prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma444){
660     /* fetch pixels for estimated mv 4 macroblocks ahead
661      * optimized for 64byte cache lines */
662     MpegEncContext * const s = &h->s;
663     const int refn = h->ref_cache[list][scan8[0]];
664     if(refn >= 0){
665         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
666         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
667         uint8_t **src = h->ref_list[list][refn].f.data;
668         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
669         s->dsp.prefetch(src[0]+off, s->linesize, 4);
670         if(chroma444){
671             s->dsp.prefetch(src[1]+off, s->linesize, 4);
672             s->dsp.prefetch(src[2]+off, s->linesize, 4);
673         }else{
674             off= ((mx>>1) << pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + (64 << pixel_shift);
675             s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
676         }
677     }
678 }
679
680 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
681                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
682                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
683                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
684                       int pixel_shift, int chroma444){
685     MpegEncContext * const s = &h->s;
686     const int mb_xy= h->mb_xy;
687     const int mb_type = s->current_picture.f.mb_type[mb_xy];
688
689     assert(IS_INTER(mb_type));
690
691     if(HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
692         await_references(h);
693     prefetch_motion(h, 0, pixel_shift, chroma444);
694
695     if(IS_16X16(mb_type)){
696         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
697                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
698                 weight_op, weight_avg,
699                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
700                 pixel_shift, chroma444);
701     }else if(IS_16X8(mb_type)){
702         mc_part(h, 0, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
703                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
704                 &weight_op[1], &weight_avg[1],
705                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
706                 pixel_shift, chroma444);
707         mc_part(h, 8, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
708                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
709                 &weight_op[1], &weight_avg[1],
710                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
711                 pixel_shift, chroma444);
712     }else if(IS_8X16(mb_type)){
713         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
714                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
715                 &weight_op[2], &weight_avg[2],
716                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
717                 pixel_shift, chroma444);
718         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
719                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
720                 &weight_op[2], &weight_avg[2],
721                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
722                 pixel_shift, chroma444);
723     }else{
724         int i;
725
726         assert(IS_8X8(mb_type));
727
728         for(i=0; i<4; i++){
729             const int sub_mb_type= h->sub_mb_type[i];
730             const int n= 4*i;
731             int x_offset= (i&1)<<2;
732             int y_offset= (i&2)<<1;
733
734             if(IS_SUB_8X8(sub_mb_type)){
735                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
736                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
737                     &weight_op[3], &weight_avg[3],
738                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
739                     pixel_shift, chroma444);
740             }else if(IS_SUB_8X4(sub_mb_type)){
741                 mc_part(h, n  , 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
742                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
743                     &weight_op[4], &weight_avg[4],
744                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
745                     pixel_shift, chroma444);
746                 mc_part(h, n+2, 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
747                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
748                     &weight_op[4], &weight_avg[4],
749                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
750                     pixel_shift, chroma444);
751             }else if(IS_SUB_4X8(sub_mb_type)){
752                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
753                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
754                     &weight_op[5], &weight_avg[5],
755                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
756                     pixel_shift, chroma444);
757                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
758                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
759                     &weight_op[5], &weight_avg[5],
760                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
761                     pixel_shift, chroma444);
762             }else{
763                 int j;
764                 assert(IS_SUB_4X4(sub_mb_type));
765                 for(j=0; j<4; j++){
766                     int sub_x_offset= x_offset + 2*(j&1);
767                     int sub_y_offset= y_offset +   (j&2);
768                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
769                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
770                         &weight_op[6], &weight_avg[6],
771                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
772                         pixel_shift, chroma444);
773                 }
774             }
775         }
776     }
777
778     prefetch_motion(h, 1, pixel_shift, chroma444);
779 }
780
781 static void free_tables(H264Context *h, int free_rbsp){
782     int i;
783     H264Context *hx;
784
785     av_freep(&h->intra4x4_pred_mode);
786     av_freep(&h->chroma_pred_mode_table);
787     av_freep(&h->cbp_table);
788     av_freep(&h->mvd_table[0]);
789     av_freep(&h->mvd_table[1]);
790     av_freep(&h->direct_table);
791     av_freep(&h->non_zero_count);
792     av_freep(&h->slice_table_base);
793     h->slice_table= NULL;
794     av_freep(&h->list_counts);
795
796     av_freep(&h->mb2b_xy);
797     av_freep(&h->mb2br_xy);
798
799     for(i = 0; i < MAX_THREADS; i++) {
800         hx = h->thread_context[i];
801         if(!hx) continue;
802         av_freep(&hx->top_borders[1]);
803         av_freep(&hx->top_borders[0]);
804         av_freep(&hx->s.obmc_scratchpad);
805         if (free_rbsp){
806             av_freep(&hx->rbsp_buffer[1]);
807             av_freep(&hx->rbsp_buffer[0]);
808             hx->rbsp_buffer_size[0] = 0;
809             hx->rbsp_buffer_size[1] = 0;
810         }
811         if (i) av_freep(&h->thread_context[i]);
812     }
813 }
814
815 static void init_dequant8_coeff_table(H264Context *h){
816     int i,j,q,x;
817     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
818
819     for(i=0; i<6; i++ ){
820         h->dequant8_coeff[i] = h->dequant8_buffer[i];
821         for(j=0; j<i; j++){
822             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
823                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
824                 break;
825             }
826         }
827         if(j<i)
828             continue;
829
830         for(q=0; q<max_qp+1; q++){
831             int shift = div6[q];
832             int idx = rem6[q];
833             for(x=0; x<64; x++)
834                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
835                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
836                     h->pps.scaling_matrix8[i][x]) << shift;
837         }
838     }
839 }
840
841 static void init_dequant4_coeff_table(H264Context *h){
842     int i,j,q,x;
843     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
844     for(i=0; i<6; i++ ){
845         h->dequant4_coeff[i] = h->dequant4_buffer[i];
846         for(j=0; j<i; j++){
847             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
848                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
849                 break;
850             }
851         }
852         if(j<i)
853             continue;
854
855         for(q=0; q<max_qp+1; q++){
856             int shift = div6[q] + 2;
857             int idx = rem6[q];
858             for(x=0; x<16; x++)
859                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
860                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
861                     h->pps.scaling_matrix4[i][x]) << shift;
862         }
863     }
864 }
865
866 static void init_dequant_tables(H264Context *h){
867     int i,x;
868     init_dequant4_coeff_table(h);
869     if(h->pps.transform_8x8_mode)
870         init_dequant8_coeff_table(h);
871     if(h->sps.transform_bypass){
872         for(i=0; i<6; i++)
873             for(x=0; x<16; x++)
874                 h->dequant4_coeff[i][0][x] = 1<<6;
875         if(h->pps.transform_8x8_mode)
876             for(i=0; i<6; i++)
877                 for(x=0; x<64; x++)
878                     h->dequant8_coeff[i][0][x] = 1<<6;
879     }
880 }
881
882
883 int ff_h264_alloc_tables(H264Context *h){
884     MpegEncContext * const s = &h->s;
885     const int big_mb_num= s->mb_stride * (s->mb_height+1);
886     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
887     int x,y;
888
889     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
890
891     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
892     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
893     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
894
895     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
896     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
897     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
898     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
899     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
900
901     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
902     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
903
904     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
905     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
906     for(y=0; y<s->mb_height; y++){
907         for(x=0; x<s->mb_width; x++){
908             const int mb_xy= x + y*s->mb_stride;
909             const int b_xy = 4*x + 4*y*h->b_stride;
910
911             h->mb2b_xy [mb_xy]= b_xy;
912             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
913         }
914     }
915
916     s->obmc_scratchpad = NULL;
917
918     if(!h->dequant4_coeff[0])
919         init_dequant_tables(h);
920
921     return 0;
922 fail:
923     free_tables(h, 1);
924     return -1;
925 }
926
927 /**
928  * Mimic alloc_tables(), but for every context thread.
929  */
930 static void clone_tables(H264Context *dst, H264Context *src, int i){
931     MpegEncContext * const s = &src->s;
932     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
933     dst->non_zero_count           = src->non_zero_count;
934     dst->slice_table              = src->slice_table;
935     dst->cbp_table                = src->cbp_table;
936     dst->mb2b_xy                  = src->mb2b_xy;
937     dst->mb2br_xy                 = src->mb2br_xy;
938     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
939     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
940     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
941     dst->direct_table             = src->direct_table;
942     dst->list_counts              = src->list_counts;
943
944     dst->s.obmc_scratchpad = NULL;
945     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma, src->sps.chroma_format_idc);
946 }
947
948 /**
949  * Init context
950  * Allocate buffers which are not shared amongst multiple threads.
951  */
952 static int context_init(H264Context *h){
953     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
954     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
955
956     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
957     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
958
959     return 0;
960 fail:
961     return -1; // free_tables will clean up for us
962 }
963
964 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
965
966 static av_cold void common_init(H264Context *h){
967     MpegEncContext * const s = &h->s;
968
969     s->width = s->avctx->width;
970     s->height = s->avctx->height;
971     s->codec_id= s->avctx->codec->id;
972
973     ff_h264dsp_init(&h->h264dsp, 8, 1);
974     ff_h264_pred_init(&h->hpc, s->codec_id, 8, 1);
975
976     h->dequant_coeff_pps= -1;
977     s->unrestricted_mv=1;
978     s->decode=1; //FIXME
979
980     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
981
982     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
983     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
984 }
985
986 int ff_h264_decode_extradata(H264Context *h)
987 {
988     AVCodecContext *avctx = h->s.avctx;
989
990     if(avctx->extradata[0] == 1){
991         int i, cnt, nalsize;
992         unsigned char *p = avctx->extradata;
993
994         h->is_avc = 1;
995
996         if(avctx->extradata_size < 7) {
997             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
998             return -1;
999         }
1000         /* sps and pps in the avcC always have length coded with 2 bytes,
1001            so put a fake nal_length_size = 2 while parsing them */
1002         h->nal_length_size = 2;
1003         // Decode sps from avcC
1004         cnt = *(p+5) & 0x1f; // Number of sps
1005         p += 6;
1006         for (i = 0; i < cnt; i++) {
1007             nalsize = AV_RB16(p) + 2;
1008             if (p - avctx->extradata + nalsize > avctx->extradata_size)
1009                 return -1;
1010             if(decode_nal_units(h, p, nalsize) < 0) {
1011                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1012                 return -1;
1013             }
1014             p += nalsize;
1015         }
1016         // Decode pps from avcC
1017         cnt = *(p++); // Number of pps
1018         for (i = 0; i < cnt; i++) {
1019             nalsize = AV_RB16(p) + 2;
1020             if (p - avctx->extradata + nalsize > avctx->extradata_size)
1021                 return -1;
1022             if (decode_nal_units(h, p, nalsize) < 0) {
1023                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1024                 return -1;
1025             }
1026             p += nalsize;
1027         }
1028         // Now store right nal length size, that will be use to parse all other nals
1029         h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1030     } else {
1031         h->is_avc = 0;
1032         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1033             return -1;
1034     }
1035     return 0;
1036 }
1037
1038 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1039     H264Context *h= avctx->priv_data;
1040     MpegEncContext * const s = &h->s;
1041
1042     MPV_decode_defaults(s);
1043
1044     s->avctx = avctx;
1045     common_init(h);
1046
1047     s->out_format = FMT_H264;
1048     s->workaround_bugs= avctx->workaround_bugs;
1049
1050     // set defaults
1051 //    s->decode_mb= ff_h263_decode_mb;
1052     s->quarter_sample = 1;
1053     if(!avctx->has_b_frames)
1054     s->low_delay= 1;
1055
1056     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1057
1058     ff_h264_decode_init_vlc();
1059
1060     h->pixel_shift = 0;
1061     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1062
1063     h->thread_context[0] = h;
1064     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1065     h->prev_poc_msb= 1<<16;
1066     h->x264_build = -1;
1067     ff_h264_reset_sei(h);
1068     if(avctx->codec_id == CODEC_ID_H264){
1069         if(avctx->ticks_per_frame == 1){
1070             s->avctx->time_base.den *=2;
1071         }
1072         avctx->ticks_per_frame = 2;
1073     }
1074
1075     if(avctx->extradata_size > 0 && avctx->extradata &&
1076         ff_h264_decode_extradata(h))
1077         return -1;
1078
1079     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1080         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1081         s->low_delay = 0;
1082     }
1083
1084     return 0;
1085 }
1086
1087 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1088 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1089 {
1090     int i;
1091
1092     for (i=0; i<count; i++){
1093         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1094                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1095                 !from[i]));
1096         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1097     }
1098 }
1099
1100 static void copy_parameter_set(void **to, void **from, int count, int size)
1101 {
1102     int i;
1103
1104     for (i=0; i<count; i++){
1105         if (to[i] && !from[i]) av_freep(&to[i]);
1106         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1107
1108         if (from[i]) memcpy(to[i], from[i], size);
1109     }
1110 }
1111
1112 static int decode_init_thread_copy(AVCodecContext *avctx){
1113     H264Context *h= avctx->priv_data;
1114
1115     if (!avctx->is_copy) return 0;
1116     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1117     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1118
1119     return 0;
1120 }
1121
1122 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1123 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1124     H264Context *h= dst->priv_data, *h1= src->priv_data;
1125     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1126     int inited = s->context_initialized, err;
1127     int i;
1128
1129     if(dst == src || !s1->context_initialized) return 0;
1130
1131     err = ff_mpeg_update_thread_context(dst, src);
1132     if(err) return err;
1133
1134     //FIXME handle width/height changing
1135     if(!inited){
1136         for(i = 0; i < MAX_SPS_COUNT; i++)
1137             av_freep(h->sps_buffers + i);
1138
1139         for(i = 0; i < MAX_PPS_COUNT; i++)
1140             av_freep(h->pps_buffers + i);
1141
1142         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1143         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1144         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1145         if (ff_h264_alloc_tables(h) < 0) {
1146             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1147             return AVERROR(ENOMEM);
1148         }
1149         context_init(h);
1150
1151         for(i=0; i<2; i++){
1152             h->rbsp_buffer[i] = NULL;
1153             h->rbsp_buffer_size[i] = 0;
1154         }
1155
1156         h->thread_context[0] = h;
1157
1158         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1159         // so this has to be allocated here
1160         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1161
1162         s->dsp.clear_blocks(h->mb);
1163         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1164     }
1165
1166     //extradata/NAL handling
1167     h->is_avc          = h1->is_avc;
1168
1169     //SPS/PPS
1170     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1171     h->sps             = h1->sps;
1172     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1173     h->pps             = h1->pps;
1174
1175     //Dequantization matrices
1176     //FIXME these are big - can they be only copied when PPS changes?
1177     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1178
1179     for(i=0; i<6; i++)
1180         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1181
1182     for(i=0; i<6; i++)
1183         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1184
1185     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1186
1187     //POC timing
1188     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1189
1190     //reference lists
1191     copy_fields(h, h1, ref_count, list_count);
1192     copy_fields(h, h1, ref_list,  intra_gb);
1193     copy_fields(h, h1, short_ref, cabac_init_idc);
1194
1195     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1196     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1197     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1198
1199     h->last_slice_type = h1->last_slice_type;
1200
1201     if(!s->current_picture_ptr) return 0;
1202
1203     if(!s->dropable) {
1204         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1205         h->prev_poc_msb     = h->poc_msb;
1206         h->prev_poc_lsb     = h->poc_lsb;
1207     }
1208     h->prev_frame_num_offset= h->frame_num_offset;
1209     h->prev_frame_num       = h->frame_num;
1210     h->outputed_poc         = h->next_outputed_poc;
1211
1212     return err;
1213 }
1214
1215 int ff_h264_frame_start(H264Context *h){
1216     MpegEncContext * const s = &h->s;
1217     int i;
1218     const int pixel_shift = h->pixel_shift;
1219     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1220
1221     if(MPV_frame_start(s, s->avctx) < 0)
1222         return -1;
1223     ff_er_frame_start(s);
1224     /*
1225      * MPV_frame_start uses pict_type to derive key_frame.
1226      * This is incorrect for H.264; IDR markings must be used.
1227      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1228      * See decode_nal_units().
1229      */
1230     s->current_picture_ptr->f.key_frame = 0;
1231     s->current_picture_ptr->mmco_reset= 0;
1232
1233     assert(s->linesize && s->uvlinesize);
1234
1235     for(i=0; i<16; i++){
1236         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1237         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1238     }
1239     for(i=0; i<16; i++){
1240         h->block_offset[16+i]=
1241         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1242         h->block_offset[48+16+i]=
1243         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1244     }
1245
1246     /* can't be in alloc_tables because linesize isn't known there.
1247      * FIXME: redo bipred weight to not require extra buffer? */
1248     for(i = 0; i < thread_count; i++)
1249         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1250             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1251
1252     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1253     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1254
1255 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1256
1257     // We mark the current picture as non-reference after allocating it, so
1258     // that if we break out due to an error it can be released automatically
1259     // in the next MPV_frame_start().
1260     // SVQ3 as well as most other codecs have only last/next/current and thus
1261     // get released even with set reference, besides SVQ3 and others do not
1262     // mark frames as reference later "naturally".
1263     if(s->codec_id != CODEC_ID_SVQ3)
1264         s->current_picture_ptr->f.reference = 0;
1265
1266     s->current_picture_ptr->field_poc[0]=
1267     s->current_picture_ptr->field_poc[1]= INT_MAX;
1268
1269     h->next_output_pic = NULL;
1270
1271     assert(s->current_picture_ptr->long_ref==0);
1272
1273     return 0;
1274 }
1275
1276 /**
1277   * Run setup operations that must be run after slice header decoding.
1278   * This includes finding the next displayed frame.
1279   *
1280   * @param h h264 master context
1281   * @param setup_finished enough NALs have been read that we can call
1282   * ff_thread_finish_setup()
1283   */
1284 static void decode_postinit(H264Context *h, int setup_finished){
1285     MpegEncContext * const s = &h->s;
1286     Picture *out = s->current_picture_ptr;
1287     Picture *cur = s->current_picture_ptr;
1288     int i, pics, out_of_order, out_idx;
1289
1290     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1291     s->current_picture_ptr->f.pict_type   = s->pict_type;
1292
1293     if (h->next_output_pic) return;
1294
1295     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1296         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1297         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1298         //to find this yet, so we assume the worst for now.
1299         //if (setup_finished)
1300         //    ff_thread_finish_setup(s->avctx);
1301         return;
1302     }
1303
1304     cur->f.interlaced_frame = 0;
1305     cur->f.repeat_pict      = 0;
1306
1307     /* Signal interlacing information externally. */
1308     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1309
1310     if(h->sps.pic_struct_present_flag){
1311         switch (h->sei_pic_struct)
1312         {
1313         case SEI_PIC_STRUCT_FRAME:
1314             break;
1315         case SEI_PIC_STRUCT_TOP_FIELD:
1316         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1317             cur->f.interlaced_frame = 1;
1318             break;
1319         case SEI_PIC_STRUCT_TOP_BOTTOM:
1320         case SEI_PIC_STRUCT_BOTTOM_TOP:
1321             if (FIELD_OR_MBAFF_PICTURE)
1322                 cur->f.interlaced_frame = 1;
1323             else
1324                 // try to flag soft telecine progressive
1325                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1326             break;
1327         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1328         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1329             // Signal the possibility of telecined film externally (pic_struct 5,6)
1330             // From these hints, let the applications decide if they apply deinterlacing.
1331             cur->f.repeat_pict = 1;
1332             break;
1333         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1334             // Force progressive here, as doubling interlaced frame is a bad idea.
1335             cur->f.repeat_pict = 2;
1336             break;
1337         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1338             cur->f.repeat_pict = 4;
1339             break;
1340         }
1341
1342         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1343             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1344     }else{
1345         /* Derive interlacing flag from used decoding process. */
1346         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1347     }
1348     h->prev_interlaced_frame = cur->f.interlaced_frame;
1349
1350     if (cur->field_poc[0] != cur->field_poc[1]){
1351         /* Derive top_field_first from field pocs. */
1352         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1353     }else{
1354         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1355             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1356             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1357               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1358                 cur->f.top_field_first = 1;
1359             else
1360                 cur->f.top_field_first = 0;
1361         }else{
1362             /* Most likely progressive */
1363             cur->f.top_field_first = 0;
1364         }
1365     }
1366
1367     //FIXME do something with unavailable reference frames
1368
1369     /* Sort B-frames into display order */
1370
1371     if(h->sps.bitstream_restriction_flag
1372        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1373         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1374         s->low_delay = 0;
1375     }
1376
1377     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1378        && !h->sps.bitstream_restriction_flag){
1379         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1380         s->low_delay= 0;
1381     }
1382
1383     pics = 0;
1384     while(h->delayed_pic[pics]) pics++;
1385
1386     assert(pics <= MAX_DELAYED_PIC_COUNT);
1387
1388     h->delayed_pic[pics++] = cur;
1389     if (cur->f.reference == 0)
1390         cur->f.reference = DELAYED_PIC_REF;
1391
1392     out = h->delayed_pic[0];
1393     out_idx = 0;
1394     for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1395         if(h->delayed_pic[i]->poc < out->poc){
1396             out = h->delayed_pic[i];
1397             out_idx = i;
1398         }
1399     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1400         h->next_outputed_poc= INT_MIN;
1401     out_of_order = out->poc < h->next_outputed_poc;
1402
1403     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1404         { }
1405     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1406        || (s->low_delay &&
1407         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1408          || cur->f.pict_type == AV_PICTURE_TYPE_B)))
1409     {
1410         s->low_delay = 0;
1411         s->avctx->has_b_frames++;
1412     }
1413
1414     if(out_of_order || pics > s->avctx->has_b_frames){
1415         out->f.reference &= ~DELAYED_PIC_REF;
1416         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1417                          // or else the first thread can release the picture and reuse it unsafely
1418         for(i=out_idx; h->delayed_pic[i]; i++)
1419             h->delayed_pic[i] = h->delayed_pic[i+1];
1420     }
1421     if(!out_of_order && pics > s->avctx->has_b_frames){
1422         h->next_output_pic = out;
1423         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1424             h->next_outputed_poc = INT_MIN;
1425         } else
1426             h->next_outputed_poc = out->poc;
1427     }else{
1428         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1429     }
1430
1431     if (setup_finished)
1432         ff_thread_finish_setup(s->avctx);
1433 }
1434
1435 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1436                                               uint8_t *src_cb, uint8_t *src_cr,
1437                                               int linesize, int uvlinesize, int simple)
1438 {
1439     MpegEncContext * const s = &h->s;
1440     uint8_t *top_border;
1441     int top_idx = 1;
1442     const int pixel_shift = h->pixel_shift;
1443     int chroma444 = CHROMA444;
1444     int chroma422 = CHROMA422;
1445
1446     src_y  -=   linesize;
1447     src_cb -= uvlinesize;
1448     src_cr -= uvlinesize;
1449
1450     if(!simple && FRAME_MBAFF){
1451         if(s->mb_y&1){
1452             if(!MB_MBAFF){
1453                 top_border = h->top_borders[0][s->mb_x];
1454                 AV_COPY128(top_border, src_y + 15*linesize);
1455                 if (pixel_shift)
1456                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1457                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1458                     if(chroma444){
1459                         if (pixel_shift){
1460                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1461                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1462                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1463                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1464                         } else {
1465                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1466                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1467                         }
1468                     } else if(chroma422) {
1469                         if (pixel_shift) {
1470                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1471                             AV_COPY128(top_border+48, src_cr + 15*uvlinesize);
1472                         } else {
1473                             AV_COPY64(top_border+16, src_cb +  15*uvlinesize);
1474                             AV_COPY64(top_border+24, src_cr +  15*uvlinesize);
1475                         }
1476                     } else {
1477                         if (pixel_shift) {
1478                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1479                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1480                         } else {
1481                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1482                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1483                         }
1484                     }
1485                 }
1486             }
1487         }else if(MB_MBAFF){
1488             top_idx = 0;
1489         }else
1490             return;
1491     }
1492
1493     top_border = h->top_borders[top_idx][s->mb_x];
1494     // There are two lines saved, the line above the the top macroblock of a pair,
1495     // and the line above the bottom macroblock
1496     AV_COPY128(top_border, src_y + 16*linesize);
1497     if (pixel_shift)
1498         AV_COPY128(top_border+16, src_y+16*linesize+16);
1499
1500     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1501         if(chroma444){
1502             if (pixel_shift){
1503                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1504                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1505                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1506                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1507             } else {
1508                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1509                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1510             }
1511         } else if(chroma422) {
1512             if (pixel_shift) {
1513                 AV_COPY128(top_border+32, src_cb+16*uvlinesize);
1514                 AV_COPY128(top_border+48, src_cr+16*uvlinesize);
1515             } else {
1516                 AV_COPY64(top_border+16, src_cb+16*uvlinesize);
1517                 AV_COPY64(top_border+24, src_cr+16*uvlinesize);
1518             }
1519         } else {
1520             if (pixel_shift) {
1521                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1522                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1523             } else {
1524                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1525                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1526             }
1527         }
1528     }
1529 }
1530
1531 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1532                                   uint8_t *src_cb, uint8_t *src_cr,
1533                                   int linesize, int uvlinesize,
1534                                   int xchg, int chroma444,
1535                                   int simple, int pixel_shift){
1536     MpegEncContext * const s = &h->s;
1537     int deblock_topleft;
1538     int deblock_top;
1539     int top_idx = 1;
1540     uint8_t *top_border_m1;
1541     uint8_t *top_border;
1542
1543     if(!simple && FRAME_MBAFF){
1544         if(s->mb_y&1){
1545             if(!MB_MBAFF)
1546                 return;
1547         }else{
1548             top_idx = MB_MBAFF ? 0 : 1;
1549         }
1550     }
1551
1552     if(h->deblocking_filter == 2) {
1553         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1554         deblock_top     = h->top_type;
1555     } else {
1556         deblock_topleft = (s->mb_x > 0);
1557         deblock_top     = (s->mb_y > !!MB_FIELD);
1558     }
1559
1560     src_y  -=   linesize + 1 + pixel_shift;
1561     src_cb -= uvlinesize + 1 + pixel_shift;
1562     src_cr -= uvlinesize + 1 + pixel_shift;
1563
1564     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1565     top_border    = h->top_borders[top_idx][s->mb_x];
1566
1567 #define XCHG(a,b,xchg)\
1568     if (pixel_shift) {\
1569         if (xchg) {\
1570             AV_SWAP64(b+0,a+0);\
1571             AV_SWAP64(b+8,a+8);\
1572         } else {\
1573             AV_COPY128(b,a); \
1574         }\
1575     } else \
1576 if (xchg) AV_SWAP64(b,a);\
1577 else      AV_COPY64(b,a);
1578
1579     if(deblock_top){
1580         if(deblock_topleft){
1581             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1582         }
1583         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1584         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1585         if(s->mb_x+1 < s->mb_width){
1586             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1587         }
1588     }
1589     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1590         if(chroma444){
1591             if(deblock_topleft){
1592                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1593                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1594             }
1595             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1596             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1597             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1598             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1599             if(s->mb_x+1 < s->mb_width){
1600                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1601                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1602             }
1603         } else {
1604             if(deblock_top){
1605                 if(deblock_topleft){
1606                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1607                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1608                 }
1609                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1610                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1611             }
1612         }
1613     }
1614 }
1615
1616 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1617     if (high_bit_depth) {
1618         return AV_RN32A(((int32_t*)mb) + index);
1619     } else
1620         return AV_RN16A(mb + index);
1621 }
1622
1623 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1624     if (high_bit_depth) {
1625         AV_WN32A(((int32_t*)mb) + index, value);
1626     } else
1627         AV_WN16A(mb + index, value);
1628 }
1629
1630 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1631                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1632 {
1633     MpegEncContext * const s = &h->s;
1634     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1635     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1636     int i;
1637     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1638     block_offset += 16*p;
1639     if(IS_INTRA4x4(mb_type)){
1640         if(simple || !s->encoding){
1641             if(IS_8x8DCT(mb_type)){
1642                 if(transform_bypass){
1643                     idct_dc_add =
1644                     idct_add    = s->dsp.add_pixels8;
1645                 }else{
1646                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1647                     idct_add    = h->h264dsp.h264_idct8_add;
1648                 }
1649                 for(i=0; i<16; i+=4){
1650                     uint8_t * const ptr= dest_y + block_offset[i];
1651                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1652                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1653                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1654                     }else{
1655                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1656                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1657                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1658                         if(nnz){
1659                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1660                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1661                             else
1662                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1663                         }
1664                     }
1665                 }
1666             }else{
1667                 if(transform_bypass){
1668                     idct_dc_add =
1669                     idct_add    = s->dsp.add_pixels4;
1670                 }else{
1671                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1672                     idct_add    = h->h264dsp.h264_idct_add;
1673                 }
1674                 for(i=0; i<16; i++){
1675                     uint8_t * const ptr= dest_y + block_offset[i];
1676                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1677
1678                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1679                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1680                     }else{
1681                         uint8_t *topright;
1682                         int nnz, tr;
1683                         uint64_t tr_high;
1684                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1685                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1686                             assert(s->mb_y || linesize <= block_offset[i]);
1687                             if(!topright_avail){
1688                                 if (pixel_shift) {
1689                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1690                                     topright= (uint8_t*) &tr_high;
1691                                 } else {
1692                                     tr= ptr[3 - linesize]*0x01010101u;
1693                                     topright= (uint8_t*) &tr;
1694                                 }
1695                             }else
1696                                 topright= ptr + (4 << pixel_shift) - linesize;
1697                         }else
1698                             topright= NULL;
1699
1700                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1701                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1702                         if(nnz){
1703                             if(is_h264){
1704                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1705                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1706                                 else
1707                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1708                             }else
1709                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1710                         }
1711                     }
1712                 }
1713             }
1714         }
1715     }else{
1716         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1717         if(is_h264){
1718             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1719                 if(!transform_bypass)
1720                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1721                 else{
1722                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1723                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1724                     for(i = 0; i < 16; i++)
1725                         dctcoef_set(h->mb+p*256, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1726                 }
1727             }
1728         }else
1729             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1730     }
1731 }
1732
1733 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1734                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1735 {
1736     MpegEncContext * const s = &h->s;
1737     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1738     int i;
1739     block_offset += 16*p;
1740     if(!IS_INTRA4x4(mb_type)){
1741         if(is_h264){
1742             if(IS_INTRA16x16(mb_type)){
1743                 if(transform_bypass){
1744                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1745                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1746                     }else{
1747                         for(i=0; i<16; i++){
1748                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1749                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1750                         }
1751                     }
1752                 }else{
1753                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1754                 }
1755             }else if(h->cbp&15){
1756                 if(transform_bypass){
1757                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1758                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1759                     for(i=0; i<16; i+=di){
1760                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1761                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1762                         }
1763                     }
1764                 }else{
1765                     if(IS_8x8DCT(mb_type)){
1766                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1767                     }else{
1768                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1769                     }
1770                 }
1771             }
1772         }else{
1773             for(i=0; i<16; i++){
1774                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1775                     uint8_t * const ptr= dest_y + block_offset[i];
1776                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1777                 }
1778             }
1779         }
1780     }
1781 }
1782
1783 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1784     MpegEncContext * const s = &h->s;
1785     const int mb_x= s->mb_x;
1786     const int mb_y= s->mb_y;
1787     const int mb_xy= h->mb_xy;
1788     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1789     uint8_t  *dest_y, *dest_cb, *dest_cr;
1790     int linesize, uvlinesize /*dct_offset*/;
1791     int i, j;
1792     int *block_offset = &h->block_offset[0];
1793     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1794     /* is_h264 should always be true if SVQ3 is disabled. */
1795     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1796     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1797     const int block_h = 16 >> s->chroma_y_shift;
1798
1799     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1800     dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1801     dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1802
1803     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1804     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1805
1806     h->list_counts[mb_xy]= h->list_count;
1807
1808     if (!simple && MB_FIELD) {
1809         linesize   = h->mb_linesize   = s->linesize * 2;
1810         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1811         block_offset = &h->block_offset[48];
1812         if(mb_y&1){ //FIXME move out of this function?
1813             dest_y -= s->linesize*15;
1814             dest_cb-= s->uvlinesize * (block_h - 1);
1815             dest_cr-= s->uvlinesize * (block_h - 1);
1816         }
1817         if(FRAME_MBAFF) {
1818             int list;
1819             for(list=0; list<h->list_count; list++){
1820                 if(!USES_LIST(mb_type, list))
1821                     continue;
1822                 if(IS_16X16(mb_type)){
1823                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1824                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1825                 }else{
1826                     for(i=0; i<16; i+=4){
1827                         int ref = h->ref_cache[list][scan8[i]];
1828                         if(ref >= 0)
1829                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1830                     }
1831                 }
1832             }
1833         }
1834     } else {
1835         linesize   = h->mb_linesize   = s->linesize;
1836         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1837 //        dct_offset = s->linesize * 16;
1838     }
1839
1840     if (!simple && IS_INTRA_PCM(mb_type)) {
1841         if (pixel_shift) {
1842             const int bit_depth = h->sps.bit_depth_luma;
1843             int j;
1844             GetBitContext gb;
1845             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1846
1847             for (i = 0; i < 16; i++) {
1848                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1849                 for (j = 0; j < 16; j++)
1850                     tmp_y[j] = get_bits(&gb, bit_depth);
1851             }
1852             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1853                 if (!h->sps.chroma_format_idc) {
1854                     for (i = 0; i < 8; i++) {
1855                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1856                         for (j = 0; j < 8; j++) {
1857                             tmp_cb[j] = 1 << (bit_depth - 1);
1858                         }
1859                     }
1860                     for (i = 0; i < 8; i++) {
1861                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1862                         for (j = 0; j < 8; j++) {
1863                             tmp_cr[j] = 1 << (bit_depth - 1);
1864                         }
1865                     }
1866                 } else {
1867                     for (i = 0; i < block_h; i++) {
1868                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1869                         for (j = 0; j < 8; j++)
1870                             tmp_cb[j] = get_bits(&gb, bit_depth);
1871                     }
1872                     for (i = 0; i < block_h; i++) {
1873                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1874                         for (j = 0; j < 8; j++)
1875                             tmp_cr[j] = get_bits(&gb, bit_depth);
1876                     }
1877                 }
1878             }
1879         } else {
1880             for (i=0; i<16; i++) {
1881                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1882             }
1883             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1884                 if (!h->sps.chroma_format_idc) {
1885                     for (i = 0; i < 8; i++) {
1886                         memset(dest_cb + i*uvlinesize, 128, 8);
1887                         memset(dest_cr + i*uvlinesize, 128, 8);
1888                     }
1889                 } else {
1890                     for (i = 0; i < block_h; i++) {
1891                         memcpy(dest_cb + i*uvlinesize, h->mb + 128 + i*4,  8);
1892                         memcpy(dest_cr + i*uvlinesize, h->mb + 160 + i*4,  8);
1893                     }
1894                 }
1895             }
1896         }
1897     } else {
1898         if(IS_INTRA(mb_type)){
1899             if(h->deblocking_filter)
1900                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1901
1902             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1903                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1904                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1905             }
1906
1907             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1908
1909             if(h->deblocking_filter)
1910                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1911         }else if(is_h264){
1912             hl_motion(h, dest_y, dest_cb, dest_cr,
1913                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1914                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1915                       h->h264dsp.weight_h264_pixels_tab,
1916                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 0);
1917         }
1918
1919         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1920
1921         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1922             uint8_t *dest[2] = {dest_cb, dest_cr};
1923             if(transform_bypass){
1924                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1925                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
1926                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
1927                 }else{
1928                     idct_add = s->dsp.add_pixels4;
1929                     for(j=1; j<3; j++){
1930                         for(i=j*16; i<j*16+4; i++){
1931                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1932                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1933                         }
1934                     }
1935                 }
1936             }else{
1937                 if(is_h264){
1938                     int qp[2];
1939                     if (CHROMA422) {
1940                         qp[0] = h->chroma_qp[0] + 3;
1941                         qp[1] = h->chroma_qp[1] + 3;
1942                     } else {
1943                         qp[0] = h->chroma_qp[0];
1944                         qp[1] = h->chroma_qp[1];
1945                     }
1946                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1947                         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]);
1948                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1949                         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]);
1950                     h->h264dsp.h264_idct_add8(dest, block_offset,
1951                                               h->mb, uvlinesize,
1952                                               h->non_zero_count_cache);
1953                 }else{
1954                     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]);
1955                     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]);
1956                     for(j=1; j<3; j++){
1957                         for(i=j*16; i<j*16+4; i++){
1958                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1959                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
1960                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1961                             }
1962                         }
1963                     }
1964                 }
1965             }
1966         }
1967     }
1968     if(h->cbp || IS_INTRA(mb_type))
1969     {
1970         s->dsp.clear_blocks(h->mb);
1971         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
1972     }
1973 }
1974
1975 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
1976     MpegEncContext * const s = &h->s;
1977     const int mb_x= s->mb_x;
1978     const int mb_y= s->mb_y;
1979     const int mb_xy= h->mb_xy;
1980     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1981     uint8_t  *dest[3];
1982     int linesize;
1983     int i, j, p;
1984     int *block_offset = &h->block_offset[0];
1985     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1986     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
1987
1988     for (p = 0; p < plane_count; p++)
1989     {
1990         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
1991         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1992     }
1993
1994     h->list_counts[mb_xy]= h->list_count;
1995
1996     if (!simple && MB_FIELD) {
1997         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
1998         block_offset = &h->block_offset[48];
1999         if(mb_y&1) //FIXME move out of this function?
2000             for (p = 0; p < 3; p++)
2001                 dest[p] -= s->linesize*15;
2002         if(FRAME_MBAFF) {
2003             int list;
2004             for(list=0; list<h->list_count; list++){
2005                 if(!USES_LIST(mb_type, list))
2006                     continue;
2007                 if(IS_16X16(mb_type)){
2008                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2009                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2010                 }else{
2011                     for(i=0; i<16; i+=4){
2012                         int ref = h->ref_cache[list][scan8[i]];
2013                         if(ref >= 0)
2014                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2015                     }
2016                 }
2017             }
2018         }
2019     } else {
2020         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2021     }
2022
2023     if (!simple && IS_INTRA_PCM(mb_type)) {
2024         if (pixel_shift) {
2025             const int bit_depth = h->sps.bit_depth_luma;
2026             GetBitContext gb;
2027             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
2028
2029             for (p = 0; p < plane_count; p++) {
2030                 for (i = 0; i < 16; i++) {
2031                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2032                     for (j = 0; j < 16; j++)
2033                         tmp[j] = get_bits(&gb, bit_depth);
2034                 }
2035             }
2036         } else {
2037             for (p = 0; p < plane_count; p++) {
2038                 for (i = 0; i < 16; i++) {
2039                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2040                 }
2041             }
2042         }
2043     } else {
2044         if(IS_INTRA(mb_type)){
2045             if(h->deblocking_filter)
2046                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2047
2048             for (p = 0; p < plane_count; p++)
2049                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2050
2051             if(h->deblocking_filter)
2052                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2053         }else{
2054             hl_motion(h, dest[0], dest[1], dest[2],
2055                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2056                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2057                       h->h264dsp.weight_h264_pixels_tab,
2058                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 1);
2059         }
2060
2061         for (p = 0; p < plane_count; p++)
2062             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2063     }
2064     if(h->cbp || IS_INTRA(mb_type))
2065     {
2066         s->dsp.clear_blocks(h->mb);
2067         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2068     }
2069 }
2070
2071 /**
2072  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2073  */
2074 #define hl_decode_mb_simple(sh, bits) \
2075 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2076     hl_decode_mb_internal(h, 1, sh); \
2077 }
2078 hl_decode_mb_simple(0, 8);
2079 hl_decode_mb_simple(1, 16);
2080
2081 /**
2082  * Process a macroblock; this handles edge cases, such as interlacing.
2083  */
2084 static void av_noinline hl_decode_mb_complex(H264Context *h){
2085     hl_decode_mb_internal(h, 0, h->pixel_shift);
2086 }
2087
2088 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2089     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2090 }
2091
2092 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2093     hl_decode_mb_444_internal(h, 1, 0);
2094 }
2095
2096 void ff_h264_hl_decode_mb(H264Context *h){
2097     MpegEncContext * const s = &h->s;
2098     const int mb_xy= h->mb_xy;
2099     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2100     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2101
2102     if (CHROMA444) {
2103         if(is_complex || h->pixel_shift)
2104             hl_decode_mb_444_complex(h);
2105         else
2106             hl_decode_mb_444_simple(h);
2107     } else if (is_complex) {
2108         hl_decode_mb_complex(h);
2109     } else if (h->pixel_shift) {
2110         hl_decode_mb_simple_16(h);
2111     } else
2112         hl_decode_mb_simple_8(h);
2113 }
2114
2115 static int pred_weight_table(H264Context *h){
2116     MpegEncContext * const s = &h->s;
2117     int list, i;
2118     int luma_def, chroma_def;
2119
2120     h->use_weight= 0;
2121     h->use_weight_chroma= 0;
2122     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2123     if(h->sps.chroma_format_idc)
2124         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2125     luma_def = 1<<h->luma_log2_weight_denom;
2126     chroma_def = 1<<h->chroma_log2_weight_denom;
2127
2128     for(list=0; list<2; list++){
2129         h->luma_weight_flag[list]   = 0;
2130         h->chroma_weight_flag[list] = 0;
2131         for(i=0; i<h->ref_count[list]; i++){
2132             int luma_weight_flag, chroma_weight_flag;
2133
2134             luma_weight_flag= get_bits1(&s->gb);
2135             if(luma_weight_flag){
2136                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2137                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2138                 if(   h->luma_weight[i][list][0] != luma_def
2139                    || h->luma_weight[i][list][1] != 0) {
2140                     h->use_weight= 1;
2141                     h->luma_weight_flag[list]= 1;
2142                 }
2143             }else{
2144                 h->luma_weight[i][list][0]= luma_def;
2145                 h->luma_weight[i][list][1]= 0;
2146             }
2147
2148             if(h->sps.chroma_format_idc){
2149                 chroma_weight_flag= get_bits1(&s->gb);
2150                 if(chroma_weight_flag){
2151                     int j;
2152                     for(j=0; j<2; j++){
2153                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2154                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2155                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2156                            || h->chroma_weight[i][list][j][1] != 0) {
2157                             h->use_weight_chroma= 1;
2158                             h->chroma_weight_flag[list]= 1;
2159                         }
2160                     }
2161                 }else{
2162                     int j;
2163                     for(j=0; j<2; j++){
2164                         h->chroma_weight[i][list][j][0]= chroma_def;
2165                         h->chroma_weight[i][list][j][1]= 0;
2166                     }
2167                 }
2168             }
2169         }
2170         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2171     }
2172     h->use_weight= h->use_weight || h->use_weight_chroma;
2173     return 0;
2174 }
2175
2176 /**
2177  * Initialize implicit_weight table.
2178  * @param field  0/1 initialize the weight for interlaced MBAFF
2179  *                -1 initializes the rest
2180  */
2181 static void implicit_weight_table(H264Context *h, int field){
2182     MpegEncContext * const s = &h->s;
2183     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2184
2185     for (i = 0; i < 2; i++) {
2186         h->luma_weight_flag[i]   = 0;
2187         h->chroma_weight_flag[i] = 0;
2188     }
2189
2190     if(field < 0){
2191         if (s->picture_structure == PICT_FRAME) {
2192             cur_poc = s->current_picture_ptr->poc;
2193         } else {
2194             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2195         }
2196     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2197        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2198         h->use_weight= 0;
2199         h->use_weight_chroma= 0;
2200         return;
2201     }
2202         ref_start= 0;
2203         ref_count0= h->ref_count[0];
2204         ref_count1= h->ref_count[1];
2205     }else{
2206         cur_poc = s->current_picture_ptr->field_poc[field];
2207         ref_start= 16;
2208         ref_count0= 16+2*h->ref_count[0];
2209         ref_count1= 16+2*h->ref_count[1];
2210     }
2211
2212     h->use_weight= 2;
2213     h->use_weight_chroma= 2;
2214     h->luma_log2_weight_denom= 5;
2215     h->chroma_log2_weight_denom= 5;
2216
2217     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2218         int poc0 = h->ref_list[0][ref0].poc;
2219         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2220             int w = 32;
2221             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2222                 int poc1 = h->ref_list[1][ref1].poc;
2223                 int td = av_clip(poc1 - poc0, -128, 127);
2224                 if(td){
2225                     int tb = av_clip(cur_poc - poc0, -128, 127);
2226                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2227                     int dist_scale_factor = (tb*tx + 32) >> 8;
2228                     if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2229                         w = 64 - dist_scale_factor;
2230                 }
2231             }
2232             if(field<0){
2233                 h->implicit_weight[ref0][ref1][0]=
2234                 h->implicit_weight[ref0][ref1][1]= w;
2235             }else{
2236                 h->implicit_weight[ref0][ref1][field]=w;
2237             }
2238         }
2239     }
2240 }
2241
2242 /**
2243  * instantaneous decoder refresh.
2244  */
2245 static void idr(H264Context *h){
2246     ff_h264_remove_all_refs(h);
2247     h->prev_frame_num= 0;
2248     h->prev_frame_num_offset= 0;
2249     h->prev_poc_msb=
2250     h->prev_poc_lsb= 0;
2251 }
2252
2253 /* forget old pics after a seek */
2254 static void flush_dpb(AVCodecContext *avctx){
2255     H264Context *h= avctx->priv_data;
2256     int i;
2257     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
2258         if(h->delayed_pic[i])
2259             h->delayed_pic[i]->f.reference = 0;
2260         h->delayed_pic[i]= NULL;
2261     }
2262     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2263     h->prev_interlaced_frame = 1;
2264     idr(h);
2265     if(h->s.current_picture_ptr)
2266         h->s.current_picture_ptr->f.reference = 0;
2267     h->s.first_field= 0;
2268     ff_h264_reset_sei(h);
2269     ff_mpeg_flush(avctx);
2270 }
2271
2272 static int init_poc(H264Context *h){
2273     MpegEncContext * const s = &h->s;
2274     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2275     int field_poc[2];
2276     Picture *cur = s->current_picture_ptr;
2277
2278     h->frame_num_offset= h->prev_frame_num_offset;
2279     if(h->frame_num < h->prev_frame_num)
2280         h->frame_num_offset += max_frame_num;
2281
2282     if(h->sps.poc_type==0){
2283         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2284
2285         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2286             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2287         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2288             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2289         else
2290             h->poc_msb = h->prev_poc_msb;
2291 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2292         field_poc[0] =
2293         field_poc[1] = h->poc_msb + h->poc_lsb;
2294         if(s->picture_structure == PICT_FRAME)
2295             field_poc[1] += h->delta_poc_bottom;
2296     }else if(h->sps.poc_type==1){
2297         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2298         int i;
2299
2300         if(h->sps.poc_cycle_length != 0)
2301             abs_frame_num = h->frame_num_offset + h->frame_num;
2302         else
2303             abs_frame_num = 0;
2304
2305         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2306             abs_frame_num--;
2307
2308         expected_delta_per_poc_cycle = 0;
2309         for(i=0; i < h->sps.poc_cycle_length; i++)
2310             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2311
2312         if(abs_frame_num > 0){
2313             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2314             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2315
2316             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2317             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2318                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2319         } else
2320             expectedpoc = 0;
2321
2322         if(h->nal_ref_idc == 0)
2323             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2324
2325         field_poc[0] = expectedpoc + h->delta_poc[0];
2326         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2327
2328         if(s->picture_structure == PICT_FRAME)
2329             field_poc[1] += h->delta_poc[1];
2330     }else{
2331         int poc= 2*(h->frame_num_offset + h->frame_num);
2332
2333         if(!h->nal_ref_idc)
2334             poc--;
2335
2336         field_poc[0]= poc;
2337         field_poc[1]= poc;
2338     }
2339
2340     if(s->picture_structure != PICT_BOTTOM_FIELD)
2341         s->current_picture_ptr->field_poc[0]= field_poc[0];
2342     if(s->picture_structure != PICT_TOP_FIELD)
2343         s->current_picture_ptr->field_poc[1]= field_poc[1];
2344     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2345
2346     return 0;
2347 }
2348
2349
2350 /**
2351  * initialize scan tables
2352  */
2353 static void init_scan_tables(H264Context *h){
2354     int i;
2355     for(i=0; i<16; i++){
2356 #define T(x) (x>>2) | ((x<<2) & 0xF)
2357         h->zigzag_scan[i] = T(zigzag_scan[i]);
2358         h-> field_scan[i] = T( field_scan[i]);
2359 #undef T
2360     }
2361     for(i=0; i<64; i++){
2362 #define T(x) (x>>3) | ((x&7)<<3)
2363         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2364         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2365         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2366         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2367 #undef T
2368     }
2369     if(h->sps.transform_bypass){ //FIXME same ugly
2370         h->zigzag_scan_q0          = zigzag_scan;
2371         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2372         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2373         h->field_scan_q0           = field_scan;
2374         h->field_scan8x8_q0        = field_scan8x8;
2375         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2376     }else{
2377         h->zigzag_scan_q0          = h->zigzag_scan;
2378         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2379         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2380         h->field_scan_q0           = h->field_scan;
2381         h->field_scan8x8_q0        = h->field_scan8x8;
2382         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2383     }
2384 }
2385
2386 static int field_end(H264Context *h, int in_setup){
2387     MpegEncContext * const s = &h->s;
2388     AVCodecContext * const avctx= s->avctx;
2389     int err = 0;
2390     s->mb_y= 0;
2391
2392     if (!in_setup && !s->dropable)
2393         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2394                                  s->picture_structure==PICT_BOTTOM_FIELD);
2395
2396     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2397         ff_vdpau_h264_set_reference_frames(s);
2398
2399     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2400         if(!s->dropable) {
2401             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2402             h->prev_poc_msb= h->poc_msb;
2403             h->prev_poc_lsb= h->poc_lsb;
2404         }
2405         h->prev_frame_num_offset= h->frame_num_offset;
2406         h->prev_frame_num= h->frame_num;
2407         h->outputed_poc = h->next_outputed_poc;
2408     }
2409
2410     if (avctx->hwaccel) {
2411         if (avctx->hwaccel->end_frame(avctx) < 0)
2412             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2413     }
2414
2415     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2416         ff_vdpau_h264_picture_complete(s);
2417
2418     /*
2419      * FIXME: Error handling code does not seem to support interlaced
2420      * when slices span multiple rows
2421      * The ff_er_add_slice calls don't work right for bottom
2422      * fields; they cause massive erroneous error concealing
2423      * Error marking covers both fields (top and bottom).
2424      * This causes a mismatched s->error_count
2425      * and a bad error table. Further, the error count goes to
2426      * INT_MAX when called for bottom field, because mb_y is
2427      * past end by one (callers fault) and resync_mb_y != 0
2428      * causes problems for the first MB line, too.
2429      */
2430     if (!FIELD_PICTURE)
2431         ff_er_frame_end(s);
2432
2433     MPV_frame_end(s);
2434
2435     h->current_slice=0;
2436
2437     return err;
2438 }
2439
2440 /**
2441  * Replicate H264 "master" context to thread contexts.
2442  */
2443 static void clone_slice(H264Context *dst, H264Context *src)
2444 {
2445     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2446     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2447     dst->s.current_picture      = src->s.current_picture;
2448     dst->s.linesize             = src->s.linesize;
2449     dst->s.uvlinesize           = src->s.uvlinesize;
2450     dst->s.first_field          = src->s.first_field;
2451
2452     dst->prev_poc_msb           = src->prev_poc_msb;
2453     dst->prev_poc_lsb           = src->prev_poc_lsb;
2454     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2455     dst->prev_frame_num         = src->prev_frame_num;
2456     dst->short_ref_count        = src->short_ref_count;
2457
2458     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2459     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2460     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2461     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2462
2463     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2464     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2465 }
2466
2467 /**
2468  * computes profile from profile_idc and constraint_set?_flags
2469  *
2470  * @param sps SPS
2471  *
2472  * @return profile as defined by FF_PROFILE_H264_*
2473  */
2474 int ff_h264_get_profile(SPS *sps)
2475 {
2476     int profile = sps->profile_idc;
2477
2478     switch(sps->profile_idc) {
2479     case FF_PROFILE_H264_BASELINE:
2480         // constraint_set1_flag set to 1
2481         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2482         break;
2483     case FF_PROFILE_H264_HIGH_10:
2484     case FF_PROFILE_H264_HIGH_422:
2485     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2486         // constraint_set3_flag set to 1
2487         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2488         break;
2489     }
2490
2491     return profile;
2492 }
2493
2494 /**
2495  * decodes a slice header.
2496  * This will also call MPV_common_init() and frame_start() as needed.
2497  *
2498  * @param h h264context
2499  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2500  *
2501  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2502  */
2503 static int decode_slice_header(H264Context *h, H264Context *h0){
2504     MpegEncContext * const s = &h->s;
2505     MpegEncContext * const s0 = &h0->s;
2506     unsigned int first_mb_in_slice;
2507     unsigned int pps_id;
2508     int num_ref_idx_active_override_flag;
2509     unsigned int slice_type, tmp, i, j;
2510     int default_ref_list_done = 0;
2511     int last_pic_structure;
2512
2513     s->dropable= h->nal_ref_idc == 0;
2514
2515     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2516     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2517         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2518         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2519     }else{
2520         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2521         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2522     }
2523
2524     first_mb_in_slice= get_ue_golomb(&s->gb);
2525
2526     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2527         if(h0->current_slice && FIELD_PICTURE){
2528             field_end(h, 1);
2529         }
2530
2531         h0->current_slice = 0;
2532         if (!s0->first_field)
2533             s->current_picture_ptr= NULL;
2534     }
2535
2536     slice_type= get_ue_golomb_31(&s->gb);
2537     if(slice_type > 9){
2538         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);
2539         return -1;
2540     }
2541     if(slice_type > 4){
2542         slice_type -= 5;
2543         h->slice_type_fixed=1;
2544     }else
2545         h->slice_type_fixed=0;
2546
2547     slice_type= golomb_to_pict_type[ slice_type ];
2548     if (slice_type == AV_PICTURE_TYPE_I
2549         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2550         default_ref_list_done = 1;
2551     }
2552     h->slice_type= slice_type;
2553     h->slice_type_nos= slice_type & 3;
2554
2555     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2556
2557     pps_id= get_ue_golomb(&s->gb);
2558     if(pps_id>=MAX_PPS_COUNT){
2559         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2560         return -1;
2561     }
2562     if(!h0->pps_buffers[pps_id]) {
2563         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2564         return -1;
2565     }
2566     h->pps= *h0->pps_buffers[pps_id];
2567
2568     if(!h0->sps_buffers[h->pps.sps_id]) {
2569         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2570         return -1;
2571     }
2572     h->sps = *h0->sps_buffers[h->pps.sps_id];
2573
2574     s->avctx->profile = ff_h264_get_profile(&h->sps);
2575     s->avctx->level   = h->sps.level_idc;
2576     s->avctx->refs    = h->sps.ref_frame_count;
2577
2578     if(h == h0 && h->dequant_coeff_pps != pps_id){
2579         h->dequant_coeff_pps = pps_id;
2580         init_dequant_tables(h);
2581     }
2582
2583     s->mb_width= h->sps.mb_width;
2584     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2585
2586     h->b_stride=  s->mb_width*4;
2587
2588     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2589
2590     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2591     if(h->sps.frame_mbs_only_flag)
2592         s->height= 16*s->mb_height - (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2593     else
2594         s->height= 16*s->mb_height - (2<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2595
2596     if (s->context_initialized
2597         && (   s->width != s->avctx->width || s->height != s->avctx->height
2598             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2599         if(h != h0) {
2600             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2601             return -1;   // width / height changed during parallelized decoding
2602         }
2603         free_tables(h, 0);
2604         flush_dpb(s->avctx);
2605         MPV_common_end(s);
2606     }
2607     if (!s->context_initialized) {
2608         if (h != h0) {
2609             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2610             return -1;
2611         }
2612
2613         avcodec_set_dimensions(s->avctx, s->width, s->height);
2614         s->avctx->sample_aspect_ratio= h->sps.sar;
2615         av_assert0(s->avctx->sample_aspect_ratio.den);
2616
2617         if(h->sps.video_signal_type_present_flag){
2618             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2619             if(h->sps.colour_description_present_flag){
2620                 s->avctx->color_primaries = h->sps.color_primaries;
2621                 s->avctx->color_trc       = h->sps.color_trc;
2622                 s->avctx->colorspace      = h->sps.colorspace;
2623             }
2624         }
2625
2626         if(h->sps.timing_info_present_flag){
2627             int64_t den= h->sps.time_scale;
2628             if(h->x264_build < 44U)
2629                 den *= 2;
2630             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2631                       h->sps.num_units_in_tick, den, 1<<30);
2632         }
2633
2634         switch (h->sps.bit_depth_luma) {
2635             case 9 :
2636                 if (CHROMA444)
2637                     s->avctx->pix_fmt = PIX_FMT_YUV444P9;
2638                 else if (CHROMA422)
2639                     s->avctx->pix_fmt = PIX_FMT_YUV422P9;
2640                 else
2641                     s->avctx->pix_fmt = PIX_FMT_YUV420P9;
2642                 break;
2643             case 10 :
2644                 if (CHROMA444)
2645                     s->avctx->pix_fmt = PIX_FMT_YUV444P10;
2646                 else if (CHROMA422)
2647                     s->avctx->pix_fmt = PIX_FMT_YUV422P10;
2648                 else
2649                     s->avctx->pix_fmt = PIX_FMT_YUV420P10;
2650                 break;
2651             default:
2652                 if (CHROMA444){
2653                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2654                 } else if (CHROMA422) {
2655                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P : PIX_FMT_YUV422P;
2656                 }else{
2657                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2658                                                              s->avctx->codec->pix_fmts ?
2659                                                              s->avctx->codec->pix_fmts :
2660                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2661                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2662                                                              ff_hwaccel_pixfmt_list_420);
2663                 }
2664         }
2665
2666         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2667
2668         if (MPV_common_init(s) < 0) {
2669             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2670             return -1;
2671         }
2672         s->first_field = 0;
2673         h->prev_interlaced_frame = 1;
2674
2675         init_scan_tables(h);
2676         if (ff_h264_alloc_tables(h) < 0) {
2677             av_log(h->s.avctx, AV_LOG_ERROR, "Could not allocate memory for h264\n");
2678             return AVERROR(ENOMEM);
2679         }
2680
2681         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2682             if (context_init(h) < 0) {
2683                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2684                 return -1;
2685             }
2686         } else {
2687             for(i = 1; i < s->avctx->thread_count; i++) {
2688                 H264Context *c;
2689                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2690                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2691                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2692                 c->h264dsp = h->h264dsp;
2693                 c->sps = h->sps;
2694                 c->pps = h->pps;
2695                 c->pixel_shift = h->pixel_shift;
2696                 init_scan_tables(c);
2697                 clone_tables(c, h, i);
2698             }
2699
2700             for(i = 0; i < s->avctx->thread_count; i++)
2701                 if (context_init(h->thread_context[i]) < 0) {
2702                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2703                     return -1;
2704                 }
2705         }
2706     }
2707
2708     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2709
2710     h->mb_mbaff = 0;
2711     h->mb_aff_frame = 0;
2712     last_pic_structure = s0->picture_structure;
2713     if(h->sps.frame_mbs_only_flag){
2714         s->picture_structure= PICT_FRAME;
2715     }else{
2716         if(get_bits1(&s->gb)) { //field_pic_flag
2717             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2718         } else {
2719             s->picture_structure= PICT_FRAME;
2720             h->mb_aff_frame = h->sps.mb_aff;
2721         }
2722     }
2723     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2724
2725     if(h0->current_slice == 0){
2726         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2727         if(h->frame_num != h->prev_frame_num) {
2728             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2729
2730             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2731
2732             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2733                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2734                 if (unwrap_prev_frame_num < 0)
2735                     unwrap_prev_frame_num += max_frame_num;
2736
2737                 h->prev_frame_num = unwrap_prev_frame_num;
2738             }
2739         }
2740
2741         while(h->frame_num !=  h->prev_frame_num &&
2742               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2743             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2744             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2745             if (ff_h264_frame_start(h) < 0)
2746                 return -1;
2747             h->prev_frame_num++;
2748             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2749             s->current_picture_ptr->frame_num= h->prev_frame_num;
2750             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2751             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2752             ff_generate_sliding_window_mmcos(h);
2753             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2754                 s->avctx->error_recognition >= FF_ER_EXPLODE)
2755                 return AVERROR_INVALIDDATA;
2756             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2757              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2758              * about there being no actual duplicates.
2759              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2760              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2761              * be fixed. */
2762             if (h->short_ref_count) {
2763                 if (prev) {
2764                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2765                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2766                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2767                     h->short_ref[0]->poc = prev->poc+2;
2768                 }
2769                 h->short_ref[0]->frame_num = h->prev_frame_num;
2770             }
2771         }
2772
2773         /* See if we have a decoded first field looking for a pair... */
2774         if (s0->first_field) {
2775             assert(s0->current_picture_ptr);
2776             assert(s0->current_picture_ptr->f.data[0]);
2777             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2778
2779             /* figure out if we have a complementary field pair */
2780             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2781                 /*
2782                  * Previous field is unmatched. Don't display it, but let it
2783                  * remain for reference if marked as such.
2784                  */
2785                 s0->current_picture_ptr = NULL;
2786                 s0->first_field = FIELD_PICTURE;
2787
2788             } else {
2789                 if (h->nal_ref_idc &&
2790                         s0->current_picture_ptr->f.reference &&
2791                         s0->current_picture_ptr->frame_num != h->frame_num) {
2792                     /*
2793                      * This and previous field were reference, but had
2794                      * different frame_nums. Consider this field first in
2795                      * pair. Throw away previous field except for reference
2796                      * purposes.
2797                      */
2798                     s0->first_field = 1;
2799                     s0->current_picture_ptr = NULL;
2800
2801                 } else {
2802                     /* Second field in complementary pair */
2803                     s0->first_field = 0;
2804                 }
2805             }
2806
2807         } else {
2808             /* Frame or first field in a potentially complementary pair */
2809             assert(!s0->current_picture_ptr);
2810             s0->first_field = FIELD_PICTURE;
2811         }
2812
2813         if(!FIELD_PICTURE || s0->first_field) {
2814             if (ff_h264_frame_start(h) < 0) {
2815                 s0->first_field = 0;
2816                 return -1;
2817             }
2818         } else {
2819             ff_release_unused_pictures(s, 0);
2820         }
2821     }
2822     if(h != h0)
2823         clone_slice(h, h0);
2824
2825     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2826
2827     assert(s->mb_num == s->mb_width * s->mb_height);
2828     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2829        first_mb_in_slice                    >= s->mb_num){
2830         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2831         return -1;
2832     }
2833     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2834     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2835     if (s->picture_structure == PICT_BOTTOM_FIELD)
2836         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2837     assert(s->mb_y < s->mb_height);
2838
2839     if(s->picture_structure==PICT_FRAME){
2840         h->curr_pic_num=   h->frame_num;
2841         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2842     }else{
2843         h->curr_pic_num= 2*h->frame_num + 1;
2844         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2845     }
2846
2847     if(h->nal_unit_type == NAL_IDR_SLICE){
2848         get_ue_golomb(&s->gb); /* idr_pic_id */
2849     }
2850
2851     if(h->sps.poc_type==0){
2852         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2853
2854         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2855             h->delta_poc_bottom= get_se_golomb(&s->gb);
2856         }
2857     }
2858
2859     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2860         h->delta_poc[0]= get_se_golomb(&s->gb);
2861
2862         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2863             h->delta_poc[1]= get_se_golomb(&s->gb);
2864     }
2865
2866     init_poc(h);
2867
2868     if(h->pps.redundant_pic_cnt_present){
2869         h->redundant_pic_count= get_ue_golomb(&s->gb);
2870     }
2871
2872     //set defaults, might be overridden a few lines later
2873     h->ref_count[0]= h->pps.ref_count[0];
2874     h->ref_count[1]= h->pps.ref_count[1];
2875
2876     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2877         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2878             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2879         }
2880         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2881
2882         if(num_ref_idx_active_override_flag){
2883             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2884             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2885                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2886
2887             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2888                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2889                 h->ref_count[0]= h->ref_count[1]= 1;
2890                 return -1;
2891             }
2892         }
2893         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2894             h->list_count= 2;
2895         else
2896             h->list_count= 1;
2897     }else
2898         h->list_count= 0;
2899
2900     if(!default_ref_list_done){
2901         ff_h264_fill_default_ref_list(h);
2902     }
2903
2904     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0) {
2905         h->ref_count[1]= h->ref_count[0]= 0;
2906         return -1;
2907     }
2908
2909     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2910         s->last_picture_ptr= &h->ref_list[0][0];
2911         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2912     }
2913     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2914         s->next_picture_ptr= &h->ref_list[1][0];
2915         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2916     }
2917
2918     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2919        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2920         pred_weight_table(h);
2921     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2922         implicit_weight_table(h, -1);
2923     }else {
2924         h->use_weight = 0;
2925         for (i = 0; i < 2; i++) {
2926             h->luma_weight_flag[i]   = 0;
2927             h->chroma_weight_flag[i] = 0;
2928         }
2929     }
2930
2931     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
2932        s->avctx->error_recognition >= FF_ER_EXPLODE)
2933         return AVERROR_INVALIDDATA;
2934
2935     if(FRAME_MBAFF){
2936         ff_h264_fill_mbaff_ref_list(h);
2937
2938         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2939             implicit_weight_table(h, 0);
2940             implicit_weight_table(h, 1);
2941         }
2942     }
2943
2944     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2945         ff_h264_direct_dist_scale_factor(h);
2946     ff_h264_direct_ref_list_init(h);
2947
2948     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2949         tmp = get_ue_golomb_31(&s->gb);
2950         if(tmp > 2){
2951             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2952             return -1;
2953         }
2954         h->cabac_init_idc= tmp;
2955     }
2956
2957     h->last_qscale_diff = 0;
2958     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2959     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
2960         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2961         return -1;
2962     }
2963     s->qscale= tmp;
2964     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2965     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2966     //FIXME qscale / qp ... stuff
2967     if(h->slice_type == AV_PICTURE_TYPE_SP){
2968         get_bits1(&s->gb); /* sp_for_switch_flag */
2969     }
2970     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
2971         get_se_golomb(&s->gb); /* slice_qs_delta */
2972     }
2973
2974     h->deblocking_filter = 1;
2975     h->slice_alpha_c0_offset = 52;
2976     h->slice_beta_offset = 52;
2977     if( h->pps.deblocking_filter_parameters_present ) {
2978         tmp= get_ue_golomb_31(&s->gb);
2979         if(tmp > 2){
2980             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2981             return -1;
2982         }
2983         h->deblocking_filter= tmp;
2984         if(h->deblocking_filter < 2)
2985             h->deblocking_filter^= 1; // 1<->0
2986
2987         if( h->deblocking_filter ) {
2988             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2989             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2990             if(   h->slice_alpha_c0_offset > 104U
2991                || h->slice_beta_offset     > 104U){
2992                 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);
2993                 return -1;
2994             }
2995         }
2996     }
2997
2998     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2999        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
3000        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
3001        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
3002         h->deblocking_filter= 0;
3003
3004     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
3005         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
3006             /* Cheat slightly for speed:
3007                Do not bother to deblock across slices. */
3008             h->deblocking_filter = 2;
3009         } else {
3010             h0->max_contexts = 1;
3011             if(!h0->single_decode_warning) {
3012                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3013                 h0->single_decode_warning = 1;
3014             }
3015             if (h != h0) {
3016                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
3017                 return 1;
3018             }
3019         }
3020     }
3021     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
3022                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
3023                  + 6 * (h->sps.bit_depth_luma - 8);
3024
3025 #if 0 //FMO
3026     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
3027         slice_group_change_cycle= get_bits(&s->gb, ?);
3028 #endif
3029
3030     h0->last_slice_type = slice_type;
3031     h->slice_num = ++h0->current_slice;
3032     if(h->slice_num >= MAX_SLICES){
3033         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
3034     }
3035
3036     for(j=0; j<2; j++){
3037         int id_list[16];
3038         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
3039         for(i=0; i<16; i++){
3040             id_list[i]= 60;
3041             if (h->ref_list[j][i].f.data[0]) {
3042                 int k;
3043                 uint8_t *base = h->ref_list[j][i].f.base[0];
3044                 for(k=0; k<h->short_ref_count; k++)
3045                     if (h->short_ref[k]->f.base[0] == base) {
3046                         id_list[i]= k;
3047                         break;
3048                     }
3049                 for(k=0; k<h->long_ref_count; k++)
3050                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3051                         id_list[i]= h->short_ref_count + k;
3052                         break;
3053                     }
3054             }
3055         }
3056
3057         ref2frm[0]=
3058         ref2frm[1]= -1;
3059         for(i=0; i<16; i++)
3060             ref2frm[i+2]= 4*id_list[i]
3061                           + (h->ref_list[j][i].f.reference & 3);
3062         ref2frm[18+0]=
3063         ref2frm[18+1]= -1;
3064         for(i=16; i<48; i++)
3065             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3066                           + (h->ref_list[j][i].f.reference & 3);
3067     }
3068
3069     //FIXME: fix draw_edges+PAFF+frame threads
3070     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3071     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3072
3073     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3074         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",
3075                h->slice_num,
3076                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3077                first_mb_in_slice,
3078                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3079                pps_id, h->frame_num,
3080                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3081                h->ref_count[0], h->ref_count[1],
3082                s->qscale,
3083                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3084                h->use_weight,
3085                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3086                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3087                );
3088     }
3089
3090     return 0;
3091 }
3092
3093 int ff_h264_get_slice_type(const H264Context *h)
3094 {
3095     switch (h->slice_type) {
3096     case AV_PICTURE_TYPE_P:  return 0;
3097     case AV_PICTURE_TYPE_B:  return 1;
3098     case AV_PICTURE_TYPE_I:  return 2;
3099     case AV_PICTURE_TYPE_SP: return 3;
3100     case AV_PICTURE_TYPE_SI: return 4;
3101     default:         return -1;
3102     }
3103 }
3104
3105 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3106                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3107 {
3108     int b_stride = h->b_stride;
3109     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3110     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3111     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3112         if(USES_LIST(top_type, list)){
3113             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3114             const int b8_xy= 4*top_xy + 2;
3115             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3116             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3117             ref_cache[0 - 1*8]=
3118             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3119             ref_cache[2 - 1*8]=
3120             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3121         }else{
3122             AV_ZERO128(mv_dst - 1*8);
3123             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3124         }
3125
3126         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3127             if(USES_LIST(left_type[LTOP], list)){
3128                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3129                 const int b8_xy= 4*left_xy[LTOP] + 1;
3130                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3131                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3132                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3133                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3134                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3135                 ref_cache[-1 +  0]=
3136                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3137                 ref_cache[-1 + 16]=
3138                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3139             }else{
3140                 AV_ZERO32(mv_dst - 1 + 0);
3141                 AV_ZERO32(mv_dst - 1 + 8);
3142                 AV_ZERO32(mv_dst - 1 +16);
3143                 AV_ZERO32(mv_dst - 1 +24);
3144                 ref_cache[-1 +  0]=
3145                 ref_cache[-1 +  8]=
3146                 ref_cache[-1 + 16]=
3147                 ref_cache[-1 + 24]= LIST_NOT_USED;
3148             }
3149         }
3150     }
3151
3152     if(!USES_LIST(mb_type, list)){
3153         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3154         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3155         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3156         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3157         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3158         return;
3159     }
3160
3161     {
3162         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3163         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3164         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3165         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3166         AV_WN32A(&ref_cache[0*8], ref01);
3167         AV_WN32A(&ref_cache[1*8], ref01);
3168         AV_WN32A(&ref_cache[2*8], ref23);
3169         AV_WN32A(&ref_cache[3*8], ref23);
3170     }
3171
3172     {
3173         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3174         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3175         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3176         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3177         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3178     }
3179 }
3180
3181 /**
3182  *
3183  * @return non zero if the loop filter can be skiped
3184  */
3185 static int fill_filter_caches(H264Context *h, int mb_type){
3186     MpegEncContext * const s = &h->s;
3187     const int mb_xy= h->mb_xy;
3188     int top_xy, left_xy[LEFT_MBS];
3189     int top_type, left_type[LEFT_MBS];
3190     uint8_t *nnz;
3191     uint8_t *nnz_cache;
3192
3193     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3194
3195     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3196      * stuff, I can't imagine that these complex rules are worth it. */
3197
3198     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3199     if(FRAME_MBAFF){
3200         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3201         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3202         if(s->mb_y&1){
3203             if (left_mb_field_flag != curr_mb_field_flag) {
3204                 left_xy[LTOP] -= s->mb_stride;
3205             }
3206         }else{
3207             if(curr_mb_field_flag){
3208                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3209             }
3210             if (left_mb_field_flag != curr_mb_field_flag) {
3211                 left_xy[LBOT] += s->mb_stride;
3212             }
3213         }
3214     }
3215
3216     h->top_mb_xy = top_xy;
3217     h->left_mb_xy[LTOP] = left_xy[LTOP];
3218     h->left_mb_xy[LBOT] = left_xy[LBOT];
3219     {
3220         //for sufficiently low qp, filtering wouldn't do anything
3221         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3222         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3223         int qp = s->current_picture.f.qscale_table[mb_xy];
3224         if(qp <= qp_thresh
3225            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3226            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3227             if(!FRAME_MBAFF)
3228                 return 1;
3229             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3230                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3231                 return 1;
3232         }
3233     }
3234
3235     top_type        = s->current_picture.f.mb_type[top_xy];
3236     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3237     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3238     if(h->deblocking_filter == 2){
3239         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3240         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3241     }else{
3242         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3243         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3244     }
3245     h->top_type       = top_type;
3246     h->left_type[LTOP]= left_type[LTOP];
3247     h->left_type[LBOT]= left_type[LBOT];
3248
3249     if(IS_INTRA(mb_type))
3250         return 0;
3251
3252     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3253     if(h->list_count == 2)
3254         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3255
3256     nnz = h->non_zero_count[mb_xy];
3257     nnz_cache = h->non_zero_count_cache;
3258     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3259     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3260     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3261     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3262     h->cbp= h->cbp_table[mb_xy];
3263
3264     if(top_type){
3265         nnz = h->non_zero_count[top_xy];
3266         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3267     }
3268
3269     if(left_type[LTOP]){
3270         nnz = h->non_zero_count[left_xy[LTOP]];
3271         nnz_cache[3+8*1]= nnz[3+0*4];
3272         nnz_cache[3+8*2]= nnz[3+1*4];
3273         nnz_cache[3+8*3]= nnz[3+2*4];
3274         nnz_cache[3+8*4]= nnz[3+3*4];
3275     }
3276
3277     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3278     if(!CABAC && h->pps.transform_8x8_mode){
3279         if(IS_8x8DCT(top_type)){
3280             nnz_cache[4+8*0]=
3281             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3282             nnz_cache[6+8*0]=
3283             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3284         }
3285         if(IS_8x8DCT(left_type[LTOP])){
3286             nnz_cache[3+8*1]=
3287             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3288         }
3289         if(IS_8x8DCT(left_type[LBOT])){
3290             nnz_cache[3+8*3]=
3291             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3292         }
3293
3294         if(IS_8x8DCT(mb_type)){
3295             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3296             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3297
3298             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3299             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3300
3301             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3302             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3303
3304             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3305             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3306         }
3307     }
3308
3309     return 0;
3310 }
3311
3312 static void loop_filter(H264Context *h, int start_x, int end_x){
3313     MpegEncContext * const s = &h->s;
3314     uint8_t  *dest_y, *dest_cb, *dest_cr;
3315     int linesize, uvlinesize, mb_x, mb_y;
3316     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3317     const int old_slice_type= h->slice_type;
3318     const int pixel_shift = h->pixel_shift;
3319     const int block_h = 16 >> s->chroma_y_shift;
3320
3321     if(h->deblocking_filter) {
3322         for(mb_x= start_x; mb_x<end_x; mb_x++){
3323             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3324                 int mb_xy, mb_type;
3325                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3326                 h->slice_num= h->slice_table[mb_xy];
3327                 mb_type = s->current_picture.f.mb_type[mb_xy];
3328                 h->list_count= h->list_counts[mb_xy];
3329
3330                 if(FRAME_MBAFF)
3331                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3332
3333                 s->mb_x= mb_x;
3334                 s->mb_y= mb_y;
3335                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3336                 dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3337                 dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3338                     //FIXME simplify above
3339
3340                 if (MB_FIELD) {
3341                     linesize   = h->mb_linesize   = s->linesize * 2;
3342                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3343                     if(mb_y&1){ //FIXME move out of this function?
3344                         dest_y -= s->linesize*15;
3345                         dest_cb-= s->uvlinesize * (block_h - 1);
3346                         dest_cr-= s->uvlinesize * (block_h - 1);
3347                     }
3348                 } else {
3349                     linesize   = h->mb_linesize   = s->linesize;
3350                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3351                 }
3352                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
3353                 if(fill_filter_caches(h, mb_type))
3354                     continue;
3355                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3356                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3357
3358                 if (FRAME_MBAFF) {
3359                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3360                 } else {
3361                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3362                 }
3363             }
3364         }
3365     }
3366     h->slice_type= old_slice_type;
3367     s->mb_x= end_x;
3368     s->mb_y= end_mb_y - FRAME_MBAFF;
3369     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3370     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3371 }
3372
3373 static void predict_field_decoding_flag(H264Context *h){
3374     MpegEncContext * const s = &h->s;
3375     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3376     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3377                 ? s->current_picture.f.mb_type[mb_xy - 1]
3378                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3379                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3380                 : 0;
3381     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3382 }
3383
3384 /**
3385  * Draw edges and report progress for the last MB row.
3386  */
3387 static void decode_finish_row(H264Context *h){
3388     MpegEncContext * const s = &h->s;
3389     int top = 16*(s->mb_y >> FIELD_PICTURE);
3390     int height = 16 << FRAME_MBAFF;
3391     int deblock_border = (16 + 4) << FRAME_MBAFF;
3392     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3393
3394     if (h->deblocking_filter) {
3395         if((top + height) >= pic_height)
3396             height += deblock_border;
3397
3398         top -= deblock_border;
3399     }
3400
3401     if (top >= pic_height || (top + height) < h->emu_edge_height)
3402         return;
3403
3404     height = FFMIN(height, pic_height - top);
3405     if (top < h->emu_edge_height) {
3406         height = top+height;
3407         top = 0;
3408     }
3409
3410     ff_draw_horiz_band(s, top, height);
3411
3412     if (s->dropable) return;
3413
3414     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3415                              s->picture_structure==PICT_BOTTOM_FIELD);
3416 }
3417
3418 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3419     H264Context *h = *(void**)arg;
3420     MpegEncContext * const s = &h->s;
3421     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3422     int lf_x_start = s->mb_x;
3423
3424     s->mb_skip_run= -1;
3425
3426     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3427                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3428
3429     if( h->pps.cabac ) {
3430         /* realign */
3431         align_get_bits( &s->gb );
3432
3433         /* init cabac */
3434         ff_init_cabac_states( &h->cabac);
3435         ff_init_cabac_decoder( &h->cabac,
3436                                s->gb.buffer + get_bits_count(&s->gb)/8,
3437                                (get_bits_left(&s->gb) + 7)/8);
3438
3439         ff_h264_init_cabac_states(h);
3440
3441         for(;;){
3442 //START_TIMER
3443             int ret = ff_h264_decode_mb_cabac(h);
3444             int eos;
3445 //STOP_TIMER("decode_mb_cabac")
3446
3447             if(ret>=0) ff_h264_hl_decode_mb(h);
3448
3449             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3450                 s->mb_y++;
3451
3452                 ret = ff_h264_decode_mb_cabac(h);
3453
3454                 if(ret>=0) ff_h264_hl_decode_mb(h);
3455                 s->mb_y--;
3456             }
3457             eos = get_cabac_terminate( &h->cabac );
3458
3459             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3460                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3461                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3462                 return 0;
3463             }
3464    &nbs