Merge remote-tracking branch 'qatar/master'
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "libavutil/imgutils.h"
29 #include "internal.h"
30 #include "dsputil.h"
31 #include "avcodec.h"
32 #include "mpegvideo.h"
33 #include "h264.h"
34 #include "h264data.h"
35 #include "h264_mvpred.h"
36 #include "golomb.h"
37 #include "mathops.h"
38 #include "rectangle.h"
39 #include "thread.h"
40 #include "vdpau_internal.h"
41 #include "libavutil/avassert.h"
42
43 #include "cabac.h"
44
45 //#undef NDEBUG
46 #include <assert.h>
47
48 static const uint8_t rem6[QP_MAX_NUM+1]={
49 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
50 };
51
52 static const uint8_t div6[QP_MAX_NUM+1]={
53 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9,10,10,10,10,
54 };
55
56 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
57     PIX_FMT_DXVA2_VLD,
58     PIX_FMT_VAAPI_VLD,
59     PIX_FMT_YUVJ420P,
60     PIX_FMT_NONE
61 };
62
63 /**
64  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
65  */
66 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
67     MpegEncContext * const s = &h->s;
68     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
69     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
70     int i;
71
72     if(!(h->top_samples_available&0x8000)){
73         for(i=0; i<4; i++){
74             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
75             if(status<0){
76                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
77                 return -1;
78             } else if(status){
79                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
80             }
81         }
82     }
83
84     if((h->left_samples_available&0x8888)!=0x8888){
85         static const int mask[4]={0x8000,0x2000,0x80,0x20};
86         for(i=0; i<4; i++){
87             if(!(h->left_samples_available&mask[i])){
88                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
89                 if(status<0){
90                     av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
91                     return -1;
92                 } else if(status){
93                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
94                 }
95             }
96         }
97     }
98
99     return 0;
100 } //FIXME cleanup like 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)+64)<<pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize;
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_PTHREADS && (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);
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);
974     ff_h264_pred_init(&h->hpc, s->codec_id, 8);
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(decode_nal_units(h, p, nalsize) < 0) {
1009                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1010                 return -1;
1011             }
1012             p += nalsize;
1013         }
1014         // Decode pps from avcC
1015         cnt = *(p++); // Number of pps
1016         for (i = 0; i < cnt; i++) {
1017             nalsize = AV_RB16(p) + 2;
1018             if (decode_nal_units(h, p, nalsize) < 0) {
1019                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1020                 return -1;
1021             }
1022             p += nalsize;
1023         }
1024         // Now store right nal length size, that will be use to parse all other nals
1025         h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1026     } else {
1027         h->is_avc = 0;
1028         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1029             return -1;
1030     }
1031     return 0;
1032 }
1033
1034 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1035     H264Context *h= avctx->priv_data;
1036     MpegEncContext * const s = &h->s;
1037
1038     MPV_decode_defaults(s);
1039
1040     s->avctx = avctx;
1041     common_init(h);
1042
1043     s->out_format = FMT_H264;
1044     s->workaround_bugs= avctx->workaround_bugs;
1045
1046     // set defaults
1047 //    s->decode_mb= ff_h263_decode_mb;
1048     s->quarter_sample = 1;
1049     if(!avctx->has_b_frames)
1050     s->low_delay= 1;
1051
1052     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1053
1054     ff_h264_decode_init_vlc();
1055
1056     h->pixel_shift = 0;
1057     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1058
1059     h->thread_context[0] = h;
1060     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1061     h->prev_poc_msb= 1<<16;
1062     h->x264_build = -1;
1063     ff_h264_reset_sei(h);
1064     if(avctx->codec_id == CODEC_ID_H264){
1065         if(avctx->ticks_per_frame == 1){
1066             s->avctx->time_base.den *=2;
1067         }
1068         avctx->ticks_per_frame = 2;
1069     }
1070
1071     if(avctx->extradata_size > 0 && avctx->extradata &&
1072         ff_h264_decode_extradata(h))
1073         return -1;
1074
1075     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1076         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1077         s->low_delay = 0;
1078     }
1079
1080     return 0;
1081 }
1082
1083 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1084 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1085 {
1086     int i;
1087
1088     for (i=0; i<count; i++){
1089         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1090                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1091                 !from[i]));
1092         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1093     }
1094 }
1095
1096 static void copy_parameter_set(void **to, void **from, int count, int size)
1097 {
1098     int i;
1099
1100     for (i=0; i<count; i++){
1101         if (to[i] && !from[i]) av_freep(&to[i]);
1102         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1103
1104         if (from[i]) memcpy(to[i], from[i], size);
1105     }
1106 }
1107
1108 static int decode_init_thread_copy(AVCodecContext *avctx){
1109     H264Context *h= avctx->priv_data;
1110
1111     if (!avctx->is_copy) return 0;
1112     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1113     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1114
1115     return 0;
1116 }
1117
1118 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1119 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1120     H264Context *h= dst->priv_data, *h1= src->priv_data;
1121     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1122     int inited = s->context_initialized, err;
1123     int i;
1124
1125     if(dst == src || !s1->context_initialized) return 0;
1126
1127     err = ff_mpeg_update_thread_context(dst, src);
1128     if(err) return err;
1129
1130     //FIXME handle width/height changing
1131     if(!inited){
1132         for(i = 0; i < MAX_SPS_COUNT; i++)
1133             av_freep(h->sps_buffers + i);
1134
1135         for(i = 0; i < MAX_PPS_COUNT; i++)
1136             av_freep(h->pps_buffers + i);
1137
1138         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1139         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1140         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1141         ff_h264_alloc_tables(h);
1142         context_init(h);
1143
1144         for(i=0; i<2; i++){
1145             h->rbsp_buffer[i] = NULL;
1146             h->rbsp_buffer_size[i] = 0;
1147         }
1148
1149         h->thread_context[0] = h;
1150
1151         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1152         // so this has to be allocated here
1153         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1154
1155         s->dsp.clear_blocks(h->mb);
1156         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1157     }
1158
1159     //extradata/NAL handling
1160     h->is_avc          = h1->is_avc;
1161
1162     //SPS/PPS
1163     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1164     h->sps             = h1->sps;
1165     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1166     h->pps             = h1->pps;
1167
1168     //Dequantization matrices
1169     //FIXME these are big - can they be only copied when PPS changes?
1170     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1171
1172     for(i=0; i<6; i++)
1173         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1174
1175     for(i=0; i<6; i++)
1176         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1177
1178     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1179
1180     //POC timing
1181     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1182
1183     //reference lists
1184     copy_fields(h, h1, ref_count, list_count);
1185     copy_fields(h, h1, ref_list,  intra_gb);
1186     copy_fields(h, h1, short_ref, cabac_init_idc);
1187
1188     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1189     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1190     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1191
1192     h->last_slice_type = h1->last_slice_type;
1193
1194     if(!s->current_picture_ptr) return 0;
1195
1196     if(!s->dropable) {
1197         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1198         h->prev_poc_msb     = h->poc_msb;
1199         h->prev_poc_lsb     = h->poc_lsb;
1200     }
1201     h->prev_frame_num_offset= h->frame_num_offset;
1202     h->prev_frame_num       = h->frame_num;
1203     h->outputed_poc         = h->next_outputed_poc;
1204
1205     return err;
1206 }
1207
1208 int ff_h264_frame_start(H264Context *h){
1209     MpegEncContext * const s = &h->s;
1210     int i;
1211     const int pixel_shift = h->pixel_shift;
1212     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1213
1214     if(MPV_frame_start(s, s->avctx) < 0)
1215         return -1;
1216     ff_er_frame_start(s);
1217     /*
1218      * MPV_frame_start uses pict_type to derive key_frame.
1219      * This is incorrect for H.264; IDR markings must be used.
1220      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1221      * See decode_nal_units().
1222      */
1223     s->current_picture_ptr->f.key_frame = 0;
1224     s->current_picture_ptr->mmco_reset= 0;
1225
1226     assert(s->linesize && s->uvlinesize);
1227
1228     for(i=0; i<16; i++){
1229         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1230         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1231     }
1232     for(i=0; i<16; i++){
1233         h->block_offset[16+i]=
1234         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1235         h->block_offset[48+16+i]=
1236         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1237     }
1238
1239     /* can't be in alloc_tables because linesize isn't known there.
1240      * FIXME: redo bipred weight to not require extra buffer? */
1241     for(i = 0; i < thread_count; i++)
1242         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1243             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1244
1245     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1246     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1247
1248 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1249
1250     // We mark the current picture as non-reference after allocating it, so
1251     // that if we break out due to an error it can be released automatically
1252     // in the next MPV_frame_start().
1253     // SVQ3 as well as most other codecs have only last/next/current and thus
1254     // get released even with set reference, besides SVQ3 and others do not
1255     // mark frames as reference later "naturally".
1256     if(s->codec_id != CODEC_ID_SVQ3)
1257         s->current_picture_ptr->f.reference = 0;
1258
1259     s->current_picture_ptr->field_poc[0]=
1260     s->current_picture_ptr->field_poc[1]= INT_MAX;
1261
1262     h->next_output_pic = NULL;
1263
1264     assert(s->current_picture_ptr->long_ref==0);
1265
1266     return 0;
1267 }
1268
1269 /**
1270   * Run setup operations that must be run after slice header decoding.
1271   * This includes finding the next displayed frame.
1272   *
1273   * @param h h264 master context
1274   * @param setup_finished enough NALs have been read that we can call
1275   * ff_thread_finish_setup()
1276   */
1277 static void decode_postinit(H264Context *h, int setup_finished){
1278     MpegEncContext * const s = &h->s;
1279     Picture *out = s->current_picture_ptr;
1280     Picture *cur = s->current_picture_ptr;
1281     int i, pics, out_of_order, out_idx;
1282
1283     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1284     s->current_picture_ptr->f.pict_type   = s->pict_type;
1285
1286     if (h->next_output_pic) return;
1287
1288     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1289         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1290         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1291         //to find this yet, so we assume the worst for now.
1292         //if (setup_finished)
1293         //    ff_thread_finish_setup(s->avctx);
1294         return;
1295     }
1296
1297     cur->f.interlaced_frame = 0;
1298     cur->f.repeat_pict      = 0;
1299
1300     /* Signal interlacing information externally. */
1301     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1302
1303     if(h->sps.pic_struct_present_flag){
1304         switch (h->sei_pic_struct)
1305         {
1306         case SEI_PIC_STRUCT_FRAME:
1307             break;
1308         case SEI_PIC_STRUCT_TOP_FIELD:
1309         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1310             cur->f.interlaced_frame = 1;
1311             break;
1312         case SEI_PIC_STRUCT_TOP_BOTTOM:
1313         case SEI_PIC_STRUCT_BOTTOM_TOP:
1314             if (FIELD_OR_MBAFF_PICTURE)
1315                 cur->f.interlaced_frame = 1;
1316             else
1317                 // try to flag soft telecine progressive
1318                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1319             break;
1320         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1321         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1322             // Signal the possibility of telecined film externally (pic_struct 5,6)
1323             // From these hints, let the applications decide if they apply deinterlacing.
1324             cur->f.repeat_pict = 1;
1325             break;
1326         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1327             // Force progressive here, as doubling interlaced frame is a bad idea.
1328             cur->f.repeat_pict = 2;
1329             break;
1330         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1331             cur->f.repeat_pict = 4;
1332             break;
1333         }
1334
1335         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1336             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1337     }else{
1338         /* Derive interlacing flag from used decoding process. */
1339         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1340     }
1341     h->prev_interlaced_frame = cur->f.interlaced_frame;
1342
1343     if (cur->field_poc[0] != cur->field_poc[1]){
1344         /* Derive top_field_first from field pocs. */
1345         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1346     }else{
1347         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1348             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1349             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1350               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1351                 cur->f.top_field_first = 1;
1352             else
1353                 cur->f.top_field_first = 0;
1354         }else{
1355             /* Most likely progressive */
1356             cur->f.top_field_first = 0;
1357         }
1358     }
1359
1360     //FIXME do something with unavailable reference frames
1361
1362     /* Sort B-frames into display order */
1363
1364     if(h->sps.bitstream_restriction_flag
1365        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1366         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1367         s->low_delay = 0;
1368     }
1369
1370     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1371        && !h->sps.bitstream_restriction_flag){
1372         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1373         s->low_delay= 0;
1374     }
1375
1376     pics = 0;
1377     while(h->delayed_pic[pics]) pics++;
1378
1379     assert(pics <= MAX_DELAYED_PIC_COUNT);
1380
1381     h->delayed_pic[pics++] = cur;
1382     if (cur->f.reference == 0)
1383         cur->f.reference = DELAYED_PIC_REF;
1384
1385     out = h->delayed_pic[0];
1386     out_idx = 0;
1387     for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1388         if(h->delayed_pic[i]->poc < out->poc){
1389             out = h->delayed_pic[i];
1390             out_idx = i;
1391         }
1392     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1393         h->next_outputed_poc= INT_MIN;
1394     out_of_order = out->poc < h->next_outputed_poc;
1395
1396     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1397         { }
1398     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1399        || (s->low_delay &&
1400         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1401          || cur->f.pict_type == AV_PICTURE_TYPE_B)))
1402     {
1403         s->low_delay = 0;
1404         s->avctx->has_b_frames++;
1405     }
1406
1407     if(out_of_order || pics > s->avctx->has_b_frames){
1408         out->f.reference &= ~DELAYED_PIC_REF;
1409         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1410                          // or else the first thread can release the picture and reuse it unsafely
1411         for(i=out_idx; h->delayed_pic[i]; i++)
1412             h->delayed_pic[i] = h->delayed_pic[i+1];
1413     }
1414     if(!out_of_order && pics > s->avctx->has_b_frames){
1415         h->next_output_pic = out;
1416         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1417             h->next_outputed_poc = INT_MIN;
1418         } else
1419             h->next_outputed_poc = out->poc;
1420     }else{
1421         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1422     }
1423
1424     if (setup_finished)
1425         ff_thread_finish_setup(s->avctx);
1426 }
1427
1428 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int chroma444, int simple){
1429     MpegEncContext * const s = &h->s;
1430     uint8_t *top_border;
1431     int top_idx = 1;
1432     const int pixel_shift = h->pixel_shift;
1433
1434     src_y  -=   linesize;
1435     src_cb -= uvlinesize;
1436     src_cr -= uvlinesize;
1437
1438     if(!simple && FRAME_MBAFF){
1439         if(s->mb_y&1){
1440             if(!MB_MBAFF){
1441                 top_border = h->top_borders[0][s->mb_x];
1442                 AV_COPY128(top_border, src_y + 15*linesize);
1443                 if (pixel_shift)
1444                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1445                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1446                     if(chroma444){
1447                         if (pixel_shift){
1448                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1449                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1450                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1451                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1452                         } else {
1453                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1454                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1455                         }
1456                     } else {
1457                         if (pixel_shift) {
1458                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1459                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1460                         } else {
1461                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1462                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1463                         }
1464                     }
1465                 }
1466             }
1467         }else if(MB_MBAFF){
1468             top_idx = 0;
1469         }else
1470             return;
1471     }
1472
1473     top_border = h->top_borders[top_idx][s->mb_x];
1474     // There are two lines saved, the line above the the top macroblock of a pair,
1475     // and the line above the bottom macroblock
1476     AV_COPY128(top_border, src_y + 16*linesize);
1477     if (pixel_shift)
1478         AV_COPY128(top_border+16, src_y+16*linesize+16);
1479
1480     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1481         if(chroma444){
1482             if (pixel_shift){
1483                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1484                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1485                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1486                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1487             } else {
1488                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1489                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1490             }
1491         } else {
1492             if (pixel_shift) {
1493                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1494                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1495             } else {
1496                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1497                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1498             }
1499         }
1500     }
1501 }
1502
1503 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1504                                   uint8_t *src_cb, uint8_t *src_cr,
1505                                   int linesize, int uvlinesize,
1506                                   int xchg, int chroma444,
1507                                   int simple, int pixel_shift){
1508     MpegEncContext * const s = &h->s;
1509     int deblock_topleft;
1510     int deblock_top;
1511     int top_idx = 1;
1512     uint8_t *top_border_m1;
1513     uint8_t *top_border;
1514
1515     if(!simple && FRAME_MBAFF){
1516         if(s->mb_y&1){
1517             if(!MB_MBAFF)
1518                 return;
1519         }else{
1520             top_idx = MB_MBAFF ? 0 : 1;
1521         }
1522     }
1523
1524     if(h->deblocking_filter == 2) {
1525         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1526         deblock_top     = h->top_type;
1527     } else {
1528         deblock_topleft = (s->mb_x > 0);
1529         deblock_top     = (s->mb_y > !!MB_FIELD);
1530     }
1531
1532     src_y  -=   linesize + 1 + pixel_shift;
1533     src_cb -= uvlinesize + 1 + pixel_shift;
1534     src_cr -= uvlinesize + 1 + pixel_shift;
1535
1536     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1537     top_border    = h->top_borders[top_idx][s->mb_x];
1538
1539 #define XCHG(a,b,xchg)\
1540     if (pixel_shift) {\
1541         if (xchg) {\
1542             AV_SWAP64(b+0,a+0);\
1543             AV_SWAP64(b+8,a+8);\
1544         } else {\
1545             AV_COPY128(b,a); \
1546         }\
1547     } else \
1548 if (xchg) AV_SWAP64(b,a);\
1549 else      AV_COPY64(b,a);
1550
1551     if(deblock_top){
1552         if(deblock_topleft){
1553             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1554         }
1555         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1556         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1557         if(s->mb_x+1 < s->mb_width){
1558             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1559         }
1560     }
1561     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1562         if(chroma444){
1563             if(deblock_topleft){
1564                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1565                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1566             }
1567             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1568             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1569             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1570             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1571             if(s->mb_x+1 < s->mb_width){
1572                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1573                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1574             }
1575         } else {
1576             if(deblock_top){
1577                 if(deblock_topleft){
1578                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1579                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1580                 }
1581                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1582                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1583             }
1584         }
1585     }
1586 }
1587
1588 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1589     if (high_bit_depth) {
1590         return AV_RN32A(((int32_t*)mb) + index);
1591     } else
1592         return AV_RN16A(mb + index);
1593 }
1594
1595 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1596     if (high_bit_depth) {
1597         AV_WN32A(((int32_t*)mb) + index, value);
1598     } else
1599         AV_WN16A(mb + index, value);
1600 }
1601
1602 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1603                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1604 {
1605     MpegEncContext * const s = &h->s;
1606     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1607     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1608     int i;
1609     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1610     block_offset += 16*p;
1611     if(IS_INTRA4x4(mb_type)){
1612         if(simple || !s->encoding){
1613             if(IS_8x8DCT(mb_type)){
1614                 if(transform_bypass){
1615                     idct_dc_add =
1616                     idct_add    = s->dsp.add_pixels8;
1617                 }else{
1618                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1619                     idct_add    = h->h264dsp.h264_idct8_add;
1620                 }
1621                 for(i=0; i<16; i+=4){
1622                     uint8_t * const ptr= dest_y + block_offset[i];
1623                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1624                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1625                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1626                     }else{
1627                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1628                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1629                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1630                         if(nnz){
1631                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1632                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1633                             else
1634                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1635                         }
1636                     }
1637                 }
1638             }else{
1639                 if(transform_bypass){
1640                     idct_dc_add =
1641                     idct_add    = s->dsp.add_pixels4;
1642                 }else{
1643                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1644                     idct_add    = h->h264dsp.h264_idct_add;
1645                 }
1646                 for(i=0; i<16; i++){
1647                     uint8_t * const ptr= dest_y + block_offset[i];
1648                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1649
1650                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1651                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1652                     }else{
1653                         uint8_t *topright;
1654                         int nnz, tr;
1655                         uint64_t tr_high;
1656                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1657                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1658                             assert(s->mb_y || linesize <= block_offset[i]);
1659                             if(!topright_avail){
1660                                 if (pixel_shift) {
1661                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1662                                     topright= (uint8_t*) &tr_high;
1663                                 } else {
1664                                     tr= ptr[3 - linesize]*0x01010101;
1665                                     topright= (uint8_t*) &tr;
1666                                 }
1667                             }else
1668                                 topright= ptr + (4 << pixel_shift) - linesize;
1669                         }else
1670                             topright= NULL;
1671
1672                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1673                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1674                         if(nnz){
1675                             if(is_h264){
1676                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1677                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1678                                 else
1679                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1680                             }else
1681                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1682                         }
1683                     }
1684                 }
1685             }
1686         }
1687     }else{
1688         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1689         if(is_h264){
1690             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1691                 if(!transform_bypass)
1692                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1693                 else{
1694                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1695                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1696                     for(i = 0; i < 16; i++)
1697                         dctcoef_set(h->mb+p*256, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1698                 }
1699             }
1700         }else
1701             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1702     }
1703 }
1704
1705 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1706                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1707 {
1708     MpegEncContext * const s = &h->s;
1709     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1710     int i;
1711     block_offset += 16*p;
1712     if(!IS_INTRA4x4(mb_type)){
1713         if(is_h264){
1714             if(IS_INTRA16x16(mb_type)){
1715                 if(transform_bypass){
1716                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1717                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1718                     }else{
1719                         for(i=0; i<16; i++){
1720                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1721                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1722                         }
1723                     }
1724                 }else{
1725                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1726                 }
1727             }else if(h->cbp&15){
1728                 if(transform_bypass){
1729                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1730                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1731                     for(i=0; i<16; i+=di){
1732                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1733                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1734                         }
1735                     }
1736                 }else{
1737                     if(IS_8x8DCT(mb_type)){
1738                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1739                     }else{
1740                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1741                     }
1742                 }
1743             }
1744         }else{
1745             for(i=0; i<16; i++){
1746                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1747                     uint8_t * const ptr= dest_y + block_offset[i];
1748                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1749                 }
1750             }
1751         }
1752     }
1753 }
1754
1755 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1756     MpegEncContext * const s = &h->s;
1757     const int mb_x= s->mb_x;
1758     const int mb_y= s->mb_y;
1759     const int mb_xy= h->mb_xy;
1760     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1761     uint8_t  *dest_y, *dest_cb, *dest_cr;
1762     int linesize, uvlinesize /*dct_offset*/;
1763     int i, j;
1764     int *block_offset = &h->block_offset[0];
1765     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1766     /* is_h264 should always be true if SVQ3 is disabled. */
1767     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1768     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1769
1770     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1771     dest_cb = s->current_picture.f.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) *  8;
1772     dest_cr = s->current_picture.f.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) *  8;
1773
1774     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1775     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1776
1777     h->list_counts[mb_xy]= h->list_count;
1778
1779     if (!simple && MB_FIELD) {
1780         linesize   = h->mb_linesize   = s->linesize * 2;
1781         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1782         block_offset = &h->block_offset[48];
1783         if(mb_y&1){ //FIXME move out of this function?
1784             dest_y -= s->linesize*15;
1785             dest_cb-= s->uvlinesize*7;
1786             dest_cr-= s->uvlinesize*7;
1787         }
1788         if(FRAME_MBAFF) {
1789             int list;
1790             for(list=0; list<h->list_count; list++){
1791                 if(!USES_LIST(mb_type, list))
1792                     continue;
1793                 if(IS_16X16(mb_type)){
1794                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1795                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1796                 }else{
1797                     for(i=0; i<16; i+=4){
1798                         int ref = h->ref_cache[list][scan8[i]];
1799                         if(ref >= 0)
1800                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1801                     }
1802                 }
1803             }
1804         }
1805     } else {
1806         linesize   = h->mb_linesize   = s->linesize;
1807         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1808 //        dct_offset = s->linesize * 16;
1809     }
1810
1811     if (!simple && IS_INTRA_PCM(mb_type)) {
1812         const int bit_depth = h->sps.bit_depth_luma;
1813         if (pixel_shift) {
1814             int j;
1815             GetBitContext gb;
1816             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1817
1818             for (i = 0; i < 16; i++) {
1819                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1820                 for (j = 0; j < 16; j++)
1821                     tmp_y[j] = get_bits(&gb, bit_depth);
1822             }
1823             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1824                 if (!h->sps.chroma_format_idc) {
1825                     for (i = 0; i < 8; i++) {
1826                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1827                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1828                         for (j = 0; j < 8; j++) {
1829                             tmp_cb[j] = tmp_cr[j] = 1 << (bit_depth - 1);
1830                         }
1831                     }
1832                 } else {
1833                 for (i = 0; i < 8; i++) {
1834                     uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1835                     for (j = 0; j < 8; j++)
1836                         tmp_cb[j] = get_bits(&gb, bit_depth);
1837                 }
1838                 for (i = 0; i < 8; i++) {
1839                     uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1840                     for (j = 0; j < 8; j++)
1841                         tmp_cr[j] = get_bits(&gb, bit_depth);
1842                 }
1843                 }
1844             }
1845         } else {
1846             for (i=0; i<16; i++) {
1847                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1848             }
1849             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1850                 if (!h->sps.chroma_format_idc) {
1851                     for (i=0; i<8; i++) {
1852                         memset(dest_cb+ i*uvlinesize, 1 << (bit_depth - 1), 8);
1853                         memset(dest_cr+ i*uvlinesize, 1 << (bit_depth - 1), 8);
1854                     }
1855                 } else {
1856                 for (i=0; i<8; i++) {
1857                     memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
1858                     memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
1859                 }
1860                 }
1861             }
1862         }
1863     } else {
1864         if(IS_INTRA(mb_type)){
1865             if(h->deblocking_filter)
1866                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1867
1868             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1869                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1870                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1871             }
1872
1873             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1874
1875             if(h->deblocking_filter)
1876                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1877         }else if(is_h264){
1878             hl_motion(h, dest_y, dest_cb, dest_cr,
1879                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1880                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1881                       h->h264dsp.weight_h264_pixels_tab,
1882                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 0);
1883         }
1884
1885         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1886
1887         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1888             uint8_t *dest[2] = {dest_cb, dest_cr};
1889             if(transform_bypass){
1890                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1891                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
1892                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
1893                 }else{
1894                     idct_add = s->dsp.add_pixels4;
1895                     for(j=1; j<3; j++){
1896                         for(i=j*16; i<j*16+4; i++){
1897                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1898                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1899                         }
1900                     }
1901                 }
1902             }else{
1903                 if(is_h264){
1904                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1905                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*1 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1906                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1907                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*2 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
1908                     h->h264dsp.h264_idct_add8(dest, block_offset,
1909                                               h->mb, uvlinesize,
1910                                               h->non_zero_count_cache);
1911                 }
1912 #if CONFIG_SVQ3_DECODER
1913                 else{
1914                     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]);
1915                     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]);
1916                     for(j=1; j<3; j++){
1917                         for(i=j*16; i<j*16+4; i++){
1918                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1919                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
1920                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1921                             }
1922                         }
1923                     }
1924                 }
1925 #endif
1926             }
1927         }
1928     }
1929     if(h->cbp || IS_INTRA(mb_type))
1930     {
1931         s->dsp.clear_blocks(h->mb);
1932         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
1933     }
1934 }
1935
1936 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
1937     MpegEncContext * const s = &h->s;
1938     const int mb_x= s->mb_x;
1939     const int mb_y= s->mb_y;
1940     const int mb_xy= h->mb_xy;
1941     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1942     uint8_t  *dest[3];
1943     int linesize;
1944     int i, j, p;
1945     int *block_offset = &h->block_offset[0];
1946     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1947     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
1948
1949     for (p = 0; p < plane_count; p++)
1950     {
1951         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
1952         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1953     }
1954
1955     h->list_counts[mb_xy]= h->list_count;
1956
1957     if (!simple && MB_FIELD) {
1958         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
1959         block_offset = &h->block_offset[48];
1960         if(mb_y&1) //FIXME move out of this function?
1961             for (p = 0; p < 3; p++)
1962                 dest[p] -= s->linesize*15;
1963         if(FRAME_MBAFF) {
1964             int list;
1965             for(list=0; list<h->list_count; list++){
1966                 if(!USES_LIST(mb_type, list))
1967                     continue;
1968                 if(IS_16X16(mb_type)){
1969                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1970                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1971                 }else{
1972                     for(i=0; i<16; i+=4){
1973                         int ref = h->ref_cache[list][scan8[i]];
1974                         if(ref >= 0)
1975                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1976                     }
1977                 }
1978             }
1979         }
1980     } else {
1981         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
1982     }
1983
1984     if (!simple && IS_INTRA_PCM(mb_type)) {
1985         if (pixel_shift) {
1986             const int bit_depth = h->sps.bit_depth_luma;
1987             GetBitContext gb;
1988             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
1989
1990             for (p = 0; p < plane_count; p++) {
1991                 for (i = 0; i < 16; i++) {
1992                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
1993                     for (j = 0; j < 16; j++)
1994                         tmp[j] = get_bits(&gb, bit_depth);
1995                 }
1996             }
1997         } else {
1998             for (p = 0; p < plane_count; p++) {
1999                 for (i = 0; i < 16; i++) {
2000                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2001                 }
2002             }
2003         }
2004     } else {
2005         if(IS_INTRA(mb_type)){
2006             if(h->deblocking_filter)
2007                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2008
2009             for (p = 0; p < plane_count; p++)
2010                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2011
2012             if(h->deblocking_filter)
2013                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2014         }else{
2015             hl_motion(h, dest[0], dest[1], dest[2],
2016                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2017                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2018                       h->h264dsp.weight_h264_pixels_tab,
2019                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 1);
2020         }
2021
2022         for (p = 0; p < plane_count; p++)
2023             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2024     }
2025     if(h->cbp || IS_INTRA(mb_type))
2026     {
2027         s->dsp.clear_blocks(h->mb);
2028         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2029     }
2030 }
2031
2032 /**
2033  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2034  */
2035 #define hl_decode_mb_simple(sh, bits) \
2036 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2037     hl_decode_mb_internal(h, 1, sh); \
2038 }
2039 hl_decode_mb_simple(0, 8);
2040 hl_decode_mb_simple(1, 16);
2041
2042 /**
2043  * Process a macroblock; this handles edge cases, such as interlacing.
2044  */
2045 static void av_noinline hl_decode_mb_complex(H264Context *h){
2046     hl_decode_mb_internal(h, 0, h->pixel_shift);
2047 }
2048
2049 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2050     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2051 }
2052
2053 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2054     hl_decode_mb_444_internal(h, 1, 0);
2055 }
2056
2057 void ff_h264_hl_decode_mb(H264Context *h){
2058     MpegEncContext * const s = &h->s;
2059     const int mb_xy= h->mb_xy;
2060     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2061     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2062
2063     if (CHROMA444) {
2064         if(is_complex || h->pixel_shift)
2065             hl_decode_mb_444_complex(h);
2066         else
2067             hl_decode_mb_444_simple(h);
2068     } else if (is_complex) {
2069         hl_decode_mb_complex(h);
2070     } else if (h->pixel_shift) {
2071         hl_decode_mb_simple_16(h);
2072     } else
2073         hl_decode_mb_simple_8(h);
2074 }
2075
2076 static int pred_weight_table(H264Context *h){
2077     MpegEncContext * const s = &h->s;
2078     int list, i;
2079     int luma_def, chroma_def;
2080
2081     h->use_weight= 0;
2082     h->use_weight_chroma= 0;
2083     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2084     if(h->sps.chroma_format_idc)
2085         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2086     luma_def = 1<<h->luma_log2_weight_denom;
2087     chroma_def = 1<<h->chroma_log2_weight_denom;
2088
2089     for(list=0; list<2; list++){
2090         h->luma_weight_flag[list]   = 0;
2091         h->chroma_weight_flag[list] = 0;
2092         for(i=0; i<h->ref_count[list]; i++){
2093             int luma_weight_flag, chroma_weight_flag;
2094
2095             luma_weight_flag= get_bits1(&s->gb);
2096             if(luma_weight_flag){
2097                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2098                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2099                 if(   h->luma_weight[i][list][0] != luma_def
2100                    || h->luma_weight[i][list][1] != 0) {
2101                     h->use_weight= 1;
2102                     h->luma_weight_flag[list]= 1;
2103                 }
2104             }else{
2105                 h->luma_weight[i][list][0]= luma_def;
2106                 h->luma_weight[i][list][1]= 0;
2107             }
2108
2109             if(h->sps.chroma_format_idc){
2110                 chroma_weight_flag= get_bits1(&s->gb);
2111                 if(chroma_weight_flag){
2112                     int j;
2113                     for(j=0; j<2; j++){
2114                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2115                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2116                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2117                            || h->chroma_weight[i][list][j][1] != 0) {
2118                             h->use_weight_chroma= 1;
2119                             h->chroma_weight_flag[list]= 1;
2120                         }
2121                     }
2122                 }else{
2123                     int j;
2124                     for(j=0; j<2; j++){
2125                         h->chroma_weight[i][list][j][0]= chroma_def;
2126                         h->chroma_weight[i][list][j][1]= 0;
2127                     }
2128                 }
2129             }
2130         }
2131         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2132     }
2133     h->use_weight= h->use_weight || h->use_weight_chroma;
2134     return 0;
2135 }
2136
2137 /**
2138  * Initialize implicit_weight table.
2139  * @param field  0/1 initialize the weight for interlaced MBAFF
2140  *                -1 initializes the rest
2141  */
2142 static void implicit_weight_table(H264Context *h, int field){
2143     MpegEncContext * const s = &h->s;
2144     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2145
2146     for (i = 0; i < 2; i++) {
2147         h->luma_weight_flag[i]   = 0;
2148         h->chroma_weight_flag[i] = 0;
2149     }
2150
2151     if(field < 0){
2152         cur_poc = s->current_picture_ptr->poc;
2153     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2154        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2155         h->use_weight= 0;
2156         h->use_weight_chroma= 0;
2157         return;
2158     }
2159         ref_start= 0;
2160         ref_count0= h->ref_count[0];
2161         ref_count1= h->ref_count[1];
2162     }else{
2163         cur_poc = s->current_picture_ptr->field_poc[field];
2164         ref_start= 16;
2165         ref_count0= 16+2*h->ref_count[0];
2166         ref_count1= 16+2*h->ref_count[1];
2167     }
2168
2169     h->use_weight= 2;
2170     h->use_weight_chroma= 2;
2171     h->luma_log2_weight_denom= 5;
2172     h->chroma_log2_weight_denom= 5;
2173
2174     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2175         int poc0 = h->ref_list[0][ref0].poc;
2176         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2177             int w= 32;
2178             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref){
2179             int poc1 = h->ref_list[1][ref1].poc;
2180             int td = av_clip(poc1 - poc0, -128, 127);
2181             if(td){
2182                 int tb = av_clip(cur_poc - poc0, -128, 127);
2183                 int tx = (16384 + (FFABS(td) >> 1)) / td;
2184                 int dist_scale_factor = (tb*tx + 32) >> 8;
2185                 if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2186                     w = 64 - dist_scale_factor;
2187             }
2188             }
2189             if(field<0){
2190                 h->implicit_weight[ref0][ref1][0]=
2191                 h->implicit_weight[ref0][ref1][1]= w;
2192             }else{
2193                 h->implicit_weight[ref0][ref1][field]=w;
2194             }
2195         }
2196     }
2197 }
2198
2199 /**
2200  * instantaneous decoder refresh.
2201  */
2202 static void idr(H264Context *h){
2203     ff_h264_remove_all_refs(h);
2204     h->prev_frame_num= 0;
2205     h->prev_frame_num_offset= 0;
2206     h->prev_poc_msb=
2207     h->prev_poc_lsb= 0;
2208 }
2209
2210 /* forget old pics after a seek */
2211 static void flush_dpb(AVCodecContext *avctx){
2212     H264Context *h= avctx->priv_data;
2213     int i;
2214     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
2215         if(h->delayed_pic[i])
2216             h->delayed_pic[i]->f.reference = 0;
2217         h->delayed_pic[i]= NULL;
2218     }
2219     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2220     h->prev_interlaced_frame = 1;
2221     idr(h);
2222     if(h->s.current_picture_ptr)
2223         h->s.current_picture_ptr->f.reference = 0;
2224     h->s.first_field= 0;
2225     ff_h264_reset_sei(h);
2226     ff_mpeg_flush(avctx);
2227 }
2228
2229 static int init_poc(H264Context *h){
2230     MpegEncContext * const s = &h->s;
2231     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2232     int field_poc[2];
2233     Picture *cur = s->current_picture_ptr;
2234
2235     h->frame_num_offset= h->prev_frame_num_offset;
2236     if(h->frame_num < h->prev_frame_num)
2237         h->frame_num_offset += max_frame_num;
2238
2239     if(h->sps.poc_type==0){
2240         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2241
2242         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2243             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2244         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2245             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2246         else
2247             h->poc_msb = h->prev_poc_msb;
2248 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2249         field_poc[0] =
2250         field_poc[1] = h->poc_msb + h->poc_lsb;
2251         if(s->picture_structure == PICT_FRAME)
2252             field_poc[1] += h->delta_poc_bottom;
2253     }else if(h->sps.poc_type==1){
2254         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2255         int i;
2256
2257         if(h->sps.poc_cycle_length != 0)
2258             abs_frame_num = h->frame_num_offset + h->frame_num;
2259         else
2260             abs_frame_num = 0;
2261
2262         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2263             abs_frame_num--;
2264
2265         expected_delta_per_poc_cycle = 0;
2266         for(i=0; i < h->sps.poc_cycle_length; i++)
2267             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2268
2269         if(abs_frame_num > 0){
2270             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2271             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2272
2273             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2274             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2275                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2276         } else
2277             expectedpoc = 0;
2278
2279         if(h->nal_ref_idc == 0)
2280             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2281
2282         field_poc[0] = expectedpoc + h->delta_poc[0];
2283         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2284
2285         if(s->picture_structure == PICT_FRAME)
2286             field_poc[1] += h->delta_poc[1];
2287     }else{
2288         int poc= 2*(h->frame_num_offset + h->frame_num);
2289
2290         if(!h->nal_ref_idc)
2291             poc--;
2292
2293         field_poc[0]= poc;
2294         field_poc[1]= poc;
2295     }
2296
2297     if(s->picture_structure != PICT_BOTTOM_FIELD)
2298         s->current_picture_ptr->field_poc[0]= field_poc[0];
2299     if(s->picture_structure != PICT_TOP_FIELD)
2300         s->current_picture_ptr->field_poc[1]= field_poc[1];
2301     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2302
2303     return 0;
2304 }
2305
2306
2307 /**
2308  * initialize scan tables
2309  */
2310 static void init_scan_tables(H264Context *h){
2311     int i;
2312     for(i=0; i<16; i++){
2313 #define T(x) (x>>2) | ((x<<2) & 0xF)
2314         h->zigzag_scan[i] = T(zigzag_scan[i]);
2315         h-> field_scan[i] = T( field_scan[i]);
2316 #undef T
2317     }
2318     for(i=0; i<64; i++){
2319 #define T(x) (x>>3) | ((x&7)<<3)
2320         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2321         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2322         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2323         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2324 #undef T
2325     }
2326     if(h->sps.transform_bypass){ //FIXME same ugly
2327         h->zigzag_scan_q0          = zigzag_scan;
2328         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2329         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2330         h->field_scan_q0           = field_scan;
2331         h->field_scan8x8_q0        = field_scan8x8;
2332         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2333     }else{
2334         h->zigzag_scan_q0          = h->zigzag_scan;
2335         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2336         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2337         h->field_scan_q0           = h->field_scan;
2338         h->field_scan8x8_q0        = h->field_scan8x8;
2339         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2340     }
2341 }
2342
2343 static int field_end(H264Context *h, int in_setup){
2344     MpegEncContext * const s = &h->s;
2345     AVCodecContext * const avctx= s->avctx;
2346     int err = 0;
2347     s->mb_y= 0;
2348
2349     if (!in_setup && !s->dropable)
2350         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2351                                  s->picture_structure==PICT_BOTTOM_FIELD);
2352
2353     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2354         ff_vdpau_h264_set_reference_frames(s);
2355
2356     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2357         if(!s->dropable) {
2358             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2359             h->prev_poc_msb= h->poc_msb;
2360             h->prev_poc_lsb= h->poc_lsb;
2361         }
2362         h->prev_frame_num_offset= h->frame_num_offset;
2363         h->prev_frame_num= h->frame_num;
2364         h->outputed_poc = h->next_outputed_poc;
2365     }
2366
2367     if (avctx->hwaccel) {
2368         if (avctx->hwaccel->end_frame(avctx) < 0)
2369             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2370     }
2371
2372     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2373         ff_vdpau_h264_picture_complete(s);
2374
2375     /*
2376      * FIXME: Error handling code does not seem to support interlaced
2377      * when slices span multiple rows
2378      * The ff_er_add_slice calls don't work right for bottom
2379      * fields; they cause massive erroneous error concealing
2380      * Error marking covers both fields (top and bottom).
2381      * This causes a mismatched s->error_count
2382      * and a bad error table. Further, the error count goes to
2383      * INT_MAX when called for bottom field, because mb_y is
2384      * past end by one (callers fault) and resync_mb_y != 0
2385      * causes problems for the first MB line, too.
2386      */
2387     if (!FIELD_PICTURE)
2388         ff_er_frame_end(s);
2389
2390     MPV_frame_end(s);
2391
2392     h->current_slice=0;
2393
2394     return err;
2395 }
2396
2397 /**
2398  * Replicate H264 "master" context to thread contexts.
2399  */
2400 static void clone_slice(H264Context *dst, H264Context *src)
2401 {
2402     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2403     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2404     dst->s.current_picture      = src->s.current_picture;
2405     dst->s.linesize             = src->s.linesize;
2406     dst->s.uvlinesize           = src->s.uvlinesize;
2407     dst->s.first_field          = src->s.first_field;
2408
2409     dst->prev_poc_msb           = src->prev_poc_msb;
2410     dst->prev_poc_lsb           = src->prev_poc_lsb;
2411     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2412     dst->prev_frame_num         = src->prev_frame_num;
2413     dst->short_ref_count        = src->short_ref_count;
2414
2415     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2416     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2417     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2418     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2419
2420     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2421     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2422 }
2423
2424 /**
2425  * computes profile from profile_idc and constraint_set?_flags
2426  *
2427  * @param sps SPS
2428  *
2429  * @return profile as defined by FF_PROFILE_H264_*
2430  */
2431 int ff_h264_get_profile(SPS *sps)
2432 {
2433     int profile = sps->profile_idc;
2434
2435     switch(sps->profile_idc) {
2436     case FF_PROFILE_H264_BASELINE:
2437         // constraint_set1_flag set to 1
2438         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2439         break;
2440     case FF_PROFILE_H264_HIGH_10:
2441     case FF_PROFILE_H264_HIGH_422:
2442     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2443         // constraint_set3_flag set to 1
2444         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2445         break;
2446     }
2447
2448     return profile;
2449 }
2450
2451 /**
2452  * decodes a slice header.
2453  * This will also call MPV_common_init() and frame_start() as needed.
2454  *
2455  * @param h h264context
2456  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2457  *
2458  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2459  */
2460 static int decode_slice_header(H264Context *h, H264Context *h0){
2461     MpegEncContext * const s = &h->s;
2462     MpegEncContext * const s0 = &h0->s;
2463     unsigned int first_mb_in_slice;
2464     unsigned int pps_id;
2465     int num_ref_idx_active_override_flag;
2466     unsigned int slice_type, tmp, i, j;
2467     int default_ref_list_done = 0;
2468     int last_pic_structure;
2469
2470     s->dropable= h->nal_ref_idc == 0;
2471
2472     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2473     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2474         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2475         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2476     }else{
2477         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2478         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2479     }
2480
2481     first_mb_in_slice= get_ue_golomb(&s->gb);
2482
2483     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2484         if(h0->current_slice && FIELD_PICTURE){
2485             field_end(h, 1);
2486         }
2487
2488         h0->current_slice = 0;
2489         if (!s0->first_field)
2490             s->current_picture_ptr= NULL;
2491     }
2492
2493     slice_type= get_ue_golomb_31(&s->gb);
2494     if(slice_type > 9){
2495         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);
2496         return -1;
2497     }
2498     if(slice_type > 4){
2499         slice_type -= 5;
2500         h->slice_type_fixed=1;
2501     }else
2502         h->slice_type_fixed=0;
2503
2504     slice_type= golomb_to_pict_type[ slice_type ];
2505     if (slice_type == AV_PICTURE_TYPE_I
2506         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2507         default_ref_list_done = 1;
2508     }
2509     h->slice_type= slice_type;
2510     h->slice_type_nos= slice_type & 3;
2511
2512     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2513
2514     pps_id= get_ue_golomb(&s->gb);
2515     if(pps_id>=MAX_PPS_COUNT){
2516         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2517         return -1;
2518     }
2519     if(!h0->pps_buffers[pps_id]) {
2520         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2521         return -1;
2522     }
2523     h->pps= *h0->pps_buffers[pps_id];
2524
2525     if(!h0->sps_buffers[h->pps.sps_id]) {
2526         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2527         return -1;
2528     }
2529     h->sps = *h0->sps_buffers[h->pps.sps_id];
2530
2531     s->avctx->profile = ff_h264_get_profile(&h->sps);
2532     s->avctx->level   = h->sps.level_idc;
2533     s->avctx->refs    = h->sps.ref_frame_count;
2534
2535     if(h == h0 && h->dequant_coeff_pps != pps_id){
2536         h->dequant_coeff_pps = pps_id;
2537         init_dequant_tables(h);
2538     }
2539
2540     s->mb_width= h->sps.mb_width;
2541     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2542
2543     h->b_stride=  s->mb_width*4;
2544
2545     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2546     if(h->sps.frame_mbs_only_flag)
2547         s->height= 16*s->mb_height - (2>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2548     else
2549         s->height= 16*s->mb_height - (4>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2550
2551     if (s->context_initialized
2552         && (   s->width != s->avctx->width || s->height != s->avctx->height
2553             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2554         if(h != h0) {
2555             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2556             return -1;   // width / height changed during parallelized decoding
2557         }
2558         free_tables(h, 0);
2559         flush_dpb(s->avctx);
2560         MPV_common_end(s);
2561     }
2562     if (!s->context_initialized) {
2563         if (h != h0) {
2564             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2565             return -1;
2566         }
2567
2568         avcodec_set_dimensions(s->avctx, s->width, s->height);
2569         s->avctx->sample_aspect_ratio= h->sps.sar;
2570         av_assert0(s->avctx->sample_aspect_ratio.den);
2571
2572         h->s.avctx->coded_width = 16*s->mb_width;
2573         h->s.avctx->coded_height = 16*s->mb_height;
2574
2575         if(h->sps.video_signal_type_present_flag){
2576             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2577             if(h->sps.colour_description_present_flag){
2578                 s->avctx->color_primaries = h->sps.color_primaries;
2579                 s->avctx->color_trc       = h->sps.color_trc;
2580                 s->avctx->colorspace      = h->sps.colorspace;
2581             }
2582         }
2583
2584         if(h->sps.timing_info_present_flag){
2585             int64_t den= h->sps.time_scale;
2586             if(h->x264_build < 44U)
2587                 den *= 2;
2588             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2589                       h->sps.num_units_in_tick, den, 1<<30);
2590         }
2591
2592         switch (h->sps.bit_depth_luma) {
2593             case 9 :
2594                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P9 : PIX_FMT_YUV420P9;
2595                 break;
2596             case 10 :
2597                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P10 : PIX_FMT_YUV420P10;
2598                 break;
2599             default:
2600                 if (CHROMA444){
2601                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2602                 }else{
2603                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2604                                                              s->avctx->codec->pix_fmts ?
2605                                                              s->avctx->codec->pix_fmts :
2606                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2607                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2608                                                              ff_hwaccel_pixfmt_list_420);
2609                 }
2610         }
2611
2612         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2613
2614         if (MPV_common_init(s) < 0) {
2615             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2616             return -1;
2617         }
2618         s->first_field = 0;
2619         h->prev_interlaced_frame = 1;
2620
2621         init_scan_tables(h);
2622         ff_h264_alloc_tables(h);
2623
2624         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2625             if (context_init(h) < 0) {
2626                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2627                 return -1;
2628             }
2629         } else {
2630             for(i = 1; i < s->avctx->thread_count; i++) {
2631                 H264Context *c;
2632                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2633                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2634                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2635                 c->h264dsp = h->h264dsp;
2636                 c->sps = h->sps;
2637                 c->pps = h->pps;
2638                 c->pixel_shift = h->pixel_shift;
2639                 init_scan_tables(c);
2640                 clone_tables(c, h, i);
2641             }
2642
2643             for(i = 0; i < s->avctx->thread_count; i++)
2644                 if (context_init(h->thread_context[i]) < 0) {
2645                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2646                     return -1;
2647                 }
2648         }
2649     }
2650
2651     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2652
2653     h->mb_mbaff = 0;
2654     h->mb_aff_frame = 0;
2655     last_pic_structure = s0->picture_structure;
2656     if(h->sps.frame_mbs_only_flag){
2657         s->picture_structure= PICT_FRAME;
2658     }else{
2659         if(get_bits1(&s->gb)) { //field_pic_flag
2660             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2661         } else {
2662             s->picture_structure= PICT_FRAME;
2663             h->mb_aff_frame = h->sps.mb_aff;
2664         }
2665     }
2666     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2667
2668     if(h0->current_slice == 0){
2669         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2670         if(h->frame_num != h->prev_frame_num) {
2671             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2672
2673             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2674
2675             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2676                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2677                 if (unwrap_prev_frame_num < 0)
2678                     unwrap_prev_frame_num += max_frame_num;
2679
2680                 h->prev_frame_num = unwrap_prev_frame_num;
2681             }
2682         }
2683
2684         while(h->frame_num !=  h->prev_frame_num &&
2685               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2686             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2687             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2688             if (ff_h264_frame_start(h) < 0)
2689                 return -1;
2690             h->prev_frame_num++;
2691             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2692             s->current_picture_ptr->frame_num= h->prev_frame_num;
2693             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2694             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2695             ff_generate_sliding_window_mmcos(h);
2696             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2697                 s->avctx->error_recognition >= FF_ER_EXPLODE)
2698                 return AVERROR_INVALIDDATA;
2699             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2700              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2701              * about there being no actual duplicates.
2702              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2703              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2704              * be fixed. */
2705             if (h->short_ref_count) {
2706                 if (prev) {
2707                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2708                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2709                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2710                     h->short_ref[0]->poc = prev->poc+2;
2711                 }
2712                 h->short_ref[0]->frame_num = h->prev_frame_num;
2713             }
2714         }
2715
2716         /* See if we have a decoded first field looking for a pair... */
2717         if (s0->first_field) {
2718             assert(s0->current_picture_ptr);
2719             assert(s0->current_picture_ptr->f.data[0]);
2720             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
2721
2722             /* figure out if we have a complementary field pair */
2723             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2724                 /*
2725                  * Previous field is unmatched. Don't display it, but let it
2726                  * remain for reference if marked as such.
2727                  */
2728                 s0->current_picture_ptr = NULL;
2729                 s0->first_field = FIELD_PICTURE;
2730
2731             } else {
2732                 if (h->nal_ref_idc &&
2733                         s0->current_picture_ptr->f.reference &&
2734                         s0->current_picture_ptr->frame_num != h->frame_num) {
2735                     /*
2736                      * This and previous field were reference, but had
2737                      * different frame_nums. Consider this field first in
2738                      * pair. Throw away previous field except for reference
2739                      * purposes.
2740                      */
2741                     s0->first_field = 1;
2742                     s0->current_picture_ptr = NULL;
2743
2744                 } else {
2745                     /* Second field in complementary pair */
2746                     s0->first_field = 0;
2747                 }
2748             }
2749
2750         } else {
2751             /* Frame or first field in a potentially complementary pair */
2752             assert(!s0->current_picture_ptr);
2753             s0->first_field = FIELD_PICTURE;
2754         }
2755
2756         if(!FIELD_PICTURE || s0->first_field) {
2757             if (ff_h264_frame_start(h) < 0) {
2758                 s0->first_field = 0;
2759                 return -1;
2760             }
2761         } else {
2762             ff_release_unused_pictures(s, 0);
2763         }
2764     }
2765     if(h != h0)
2766         clone_slice(h, h0);
2767
2768     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2769
2770     assert(s->mb_num == s->mb_width * s->mb_height);
2771     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2772        first_mb_in_slice                    >= s->mb_num){
2773         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2774         return -1;
2775     }
2776     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2777     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2778     if (s->picture_structure == PICT_BOTTOM_FIELD)
2779         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2780     assert(s->mb_y < s->mb_height);
2781
2782     if(s->picture_structure==PICT_FRAME){
2783         h->curr_pic_num=   h->frame_num;
2784         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2785     }else{
2786         h->curr_pic_num= 2*h->frame_num + 1;
2787         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2788     }
2789
2790     if(h->nal_unit_type == NAL_IDR_SLICE){
2791         get_ue_golomb(&s->gb); /* idr_pic_id */
2792     }
2793
2794     if(h->sps.poc_type==0){
2795         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2796
2797         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2798             h->delta_poc_bottom= get_se_golomb(&s->gb);
2799         }
2800     }
2801
2802     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2803         h->delta_poc[0]= get_se_golomb(&s->gb);
2804
2805         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2806             h->delta_poc[1]= get_se_golomb(&s->gb);
2807     }
2808
2809     init_poc(h);
2810
2811     if(h->pps.redundant_pic_cnt_present){
2812         h->redundant_pic_count= get_ue_golomb(&s->gb);
2813     }
2814
2815     //set defaults, might be overridden a few lines later
2816     h->ref_count[0]= h->pps.ref_count[0];
2817     h->ref_count[1]= h->pps.ref_count[1];
2818
2819     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2820         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2821             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2822         }
2823         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2824
2825         if(num_ref_idx_active_override_flag){
2826             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2827             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2828                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2829
2830             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2831                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2832                 h->ref_count[0]= h->ref_count[1]= 1;
2833                 return -1;
2834             }
2835         }
2836         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2837             h->list_count= 2;
2838         else
2839             h->list_count= 1;
2840     }else
2841         h->list_count= 0;
2842
2843     if(!default_ref_list_done){
2844         ff_h264_fill_default_ref_list(h);
2845     }
2846
2847     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0)
2848         return -1;
2849
2850     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2851         s->last_picture_ptr= &h->ref_list[0][0];
2852         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2853     }
2854     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2855         s->next_picture_ptr= &h->ref_list[1][0];
2856         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2857     }
2858
2859     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2860        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2861         pred_weight_table(h);
2862     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2863         implicit_weight_table(h, -1);
2864     }else {
2865         h->use_weight = 0;
2866         for (i = 0; i < 2; i++) {
2867             h->luma_weight_flag[i]   = 0;
2868             h->chroma_weight_flag[i] = 0;
2869         }
2870     }
2871
2872     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
2873        s->avctx->error_recognition >= FF_ER_EXPLODE)
2874         return AVERROR_INVALIDDATA;
2875
2876     if(FRAME_MBAFF){
2877         ff_h264_fill_mbaff_ref_list(h);
2878
2879         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2880             implicit_weight_table(h, 0);
2881             implicit_weight_table(h, 1);
2882         }
2883     }
2884
2885     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2886         ff_h264_direct_dist_scale_factor(h);
2887     ff_h264_direct_ref_list_init(h);
2888
2889     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2890         tmp = get_ue_golomb_31(&s->gb);
2891         if(tmp > 2){
2892             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2893             return -1;
2894         }
2895         h->cabac_init_idc= tmp;
2896     }
2897
2898     h->last_qscale_diff = 0;
2899     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2900     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
2901         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2902         return -1;
2903     }
2904     s->qscale= tmp;
2905     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2906     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2907     //FIXME qscale / qp ... stuff
2908     if(h->slice_type == AV_PICTURE_TYPE_SP){
2909         get_bits1(&s->gb); /* sp_for_switch_flag */
2910     }
2911     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
2912         get_se_golomb(&s->gb); /* slice_qs_delta */
2913     }
2914
2915     h->deblocking_filter = 1;
2916     h->slice_alpha_c0_offset = 52;
2917     h->slice_beta_offset = 52;
2918     if( h->pps.deblocking_filter_parameters_present ) {
2919         tmp= get_ue_golomb_31(&s->gb);
2920         if(tmp > 2){
2921             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2922             return -1;
2923         }
2924         h->deblocking_filter= tmp;
2925         if(h->deblocking_filter < 2)
2926             h->deblocking_filter^= 1; // 1<->0
2927
2928         if( h->deblocking_filter ) {
2929             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2930             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2931             if(   h->slice_alpha_c0_offset > 104U
2932                || h->slice_beta_offset     > 104U){
2933                 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);
2934                 return -1;
2935             }
2936         }
2937     }
2938
2939     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2940        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
2941        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
2942        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2943         h->deblocking_filter= 0;
2944
2945     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
2946         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
2947             /* Cheat slightly for speed:
2948                Do not bother to deblock across slices. */
2949             h->deblocking_filter = 2;
2950         } else {
2951             h0->max_contexts = 1;
2952             if(!h0->single_decode_warning) {
2953                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
2954                 h0->single_decode_warning = 1;
2955             }
2956             if (h != h0) {
2957                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
2958                 return 1;
2959             }
2960         }
2961     }
2962     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
2963                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
2964                  + 6 * (h->sps.bit_depth_luma - 8);
2965
2966 #if 0 //FMO
2967     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
2968         slice_group_change_cycle= get_bits(&s->gb, ?);
2969 #endif
2970
2971     h0->last_slice_type = slice_type;
2972     h->slice_num = ++h0->current_slice;
2973     if(h->slice_num >= MAX_SLICES){
2974         av_log(s->avctx, AV_LOG_WARNING, "Possibly too many slices (%d >= %d), increase MAX_SLICES and recompile if there are artifacts\n", h->slice_num, MAX_SLICES);
2975     }
2976
2977     for(j=0; j<2; j++){
2978         int id_list[16];
2979         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
2980         for(i=0; i<16; i++){
2981             id_list[i]= 60;
2982             if (h->ref_list[j][i].f.data[0]) {
2983                 int k;
2984                 uint8_t *base = h->ref_list[j][i].f.base[0];
2985                 for(k=0; k<h->short_ref_count; k++)
2986                     if (h->short_ref[k]->f.base[0] == base) {
2987                         id_list[i]= k;
2988                         break;
2989                     }
2990                 for(k=0; k<h->long_ref_count; k++)
2991                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
2992                         id_list[i]= h->short_ref_count + k;
2993                         break;
2994                     }
2995             }
2996         }
2997
2998         ref2frm[0]=
2999         ref2frm[1]= -1;
3000         for(i=0; i<16; i++)
3001             ref2frm[i+2]= 4*id_list[i]
3002                           + (h->ref_list[j][i].f.reference & 3);
3003         ref2frm[18+0]=
3004         ref2frm[18+1]= -1;
3005         for(i=16; i<48; i++)
3006             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3007                           + (h->ref_list[j][i].f.reference & 3);
3008     }
3009
3010     //FIXME: fix draw_edges+PAFF+frame threads
3011     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3012     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3013
3014     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3015         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",
3016                h->slice_num,
3017                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3018                first_mb_in_slice,
3019                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3020                pps_id, h->frame_num,
3021                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3022                h->ref_count[0], h->ref_count[1],
3023                s->qscale,
3024                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3025                h->use_weight,
3026                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3027                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3028                );
3029     }
3030
3031     return 0;
3032 }
3033
3034 int ff_h264_get_slice_type(const H264Context *h)
3035 {
3036     switch (h->slice_type) {
3037     case AV_PICTURE_TYPE_P:  return 0;
3038     case AV_PICTURE_TYPE_B:  return 1;
3039     case AV_PICTURE_TYPE_I:  return 2;
3040     case AV_PICTURE_TYPE_SP: return 3;
3041     case AV_PICTURE_TYPE_SI: return 4;
3042     default:         return -1;
3043     }
3044 }
3045
3046 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3047                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3048 {
3049     int b_stride = h->b_stride;
3050     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3051     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3052     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3053         if(USES_LIST(top_type, list)){
3054             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3055             const int b8_xy= 4*top_xy + 2;
3056             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3057             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3058             ref_cache[0 - 1*8]=
3059             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3060             ref_cache[2 - 1*8]=
3061             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3062         }else{
3063             AV_ZERO128(mv_dst - 1*8);
3064             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3065         }
3066
3067         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3068             if(USES_LIST(left_type[LTOP], list)){
3069                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3070                 const int b8_xy= 4*left_xy[LTOP] + 1;
3071                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3072                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3073                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3074                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3075                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3076                 ref_cache[-1 +  0]=
3077                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3078                 ref_cache[-1 + 16]=
3079                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3080             }else{
3081                 AV_ZERO32(mv_dst - 1 + 0);
3082                 AV_ZERO32(mv_dst - 1 + 8);
3083                 AV_ZERO32(mv_dst - 1 +16);
3084                 AV_ZERO32(mv_dst - 1 +24);
3085                 ref_cache[-1 +  0]=
3086                 ref_cache[-1 +  8]=
3087                 ref_cache[-1 + 16]=
3088                 ref_cache[-1 + 24]= LIST_NOT_USED;
3089             }
3090         }
3091     }
3092
3093     if(!USES_LIST(mb_type, list)){
3094         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3095         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3096         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3097         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3098         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3099         return;
3100     }
3101
3102     {
3103         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3104         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3105         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3106         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3107         AV_WN32A(&ref_cache[0*8], ref01);
3108         AV_WN32A(&ref_cache[1*8], ref01);
3109         AV_WN32A(&ref_cache[2*8], ref23);
3110         AV_WN32A(&ref_cache[3*8], ref23);
3111     }
3112
3113     {
3114         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3115         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3116         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3117         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3118         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3119     }
3120 }
3121
3122 /**
3123  *
3124  * @return non zero if the loop filter can be skiped
3125  */
3126 static int fill_filter_caches(H264Context *h, int mb_type){
3127     MpegEncContext * const s = &h->s;
3128     const int mb_xy= h->mb_xy;
3129     int top_xy, left_xy[LEFT_MBS];
3130     int top_type, left_type[LEFT_MBS];
3131     uint8_t *nnz;
3132     uint8_t *nnz_cache;
3133
3134     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3135
3136     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3137      * stuff, I can't imagine that these complex rules are worth it. */
3138
3139     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3140     if(FRAME_MBAFF){
3141         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3142         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3143         if(s->mb_y&1){
3144             if (left_mb_field_flag != curr_mb_field_flag) {
3145                 left_xy[LTOP] -= s->mb_stride;
3146             }
3147         }else{
3148             if(curr_mb_field_flag){
3149                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3150             }
3151             if (left_mb_field_flag != curr_mb_field_flag) {
3152                 left_xy[LBOT] += s->mb_stride;
3153             }
3154         }
3155     }
3156
3157     h->top_mb_xy = top_xy;
3158     h->left_mb_xy[LTOP] = left_xy[LTOP];
3159     h->left_mb_xy[LBOT] = left_xy[LBOT];
3160     {
3161         //for sufficiently low qp, filtering wouldn't do anything
3162         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3163         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3164         int qp = s->current_picture.f.qscale_table[mb_xy];
3165         if(qp <= qp_thresh
3166            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3167            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3168             if(!FRAME_MBAFF)
3169                 return 1;
3170             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3171                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3172                 return 1;
3173         }
3174     }
3175
3176     top_type        = s->current_picture.f.mb_type[top_xy];
3177     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3178     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3179     if(h->deblocking_filter == 2){
3180         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3181         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3182     }else{
3183         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3184         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3185     }
3186     h->top_type       = top_type;
3187     h->left_type[LTOP]= left_type[LTOP];
3188     h->left_type[LBOT]= left_type[LBOT];
3189
3190     if(IS_INTRA(mb_type))
3191         return 0;
3192
3193     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3194     if(h->list_count == 2)
3195         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3196
3197     nnz = h->non_zero_count[mb_xy];
3198     nnz_cache = h->non_zero_count_cache;
3199     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3200     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3201     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3202     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3203     h->cbp= h->cbp_table[mb_xy];
3204
3205     if(top_type){
3206         nnz = h->non_zero_count[top_xy];
3207         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3208     }
3209
3210     if(left_type[LTOP]){
3211         nnz = h->non_zero_count[left_xy[LTOP]];
3212         nnz_cache[3+8*1]= nnz[3+0*4];
3213         nnz_cache[3+8*2]= nnz[3+1*4];
3214         nnz_cache[3+8*3]= nnz[3+2*4];
3215         nnz_cache[3+8*4]= nnz[3+3*4];
3216     }
3217
3218     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3219     if(!CABAC && h->pps.transform_8x8_mode){
3220         if(IS_8x8DCT(top_type)){
3221             nnz_cache[4+8*0]=
3222             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3223             nnz_cache[6+8*0]=
3224             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3225         }
3226         if(IS_8x8DCT(left_type[LTOP])){
3227             nnz_cache[3+8*1]=
3228             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3229         }
3230         if(IS_8x8DCT(left_type[LBOT])){
3231             nnz_cache[3+8*3]=
3232             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3233         }
3234
3235         if(IS_8x8DCT(mb_type)){
3236             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3237             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3238
3239             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3240             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3241
3242             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3243             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3244
3245             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3246             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3247         }
3248     }
3249
3250     return 0;
3251 }
3252
3253 static void loop_filter(H264Context *h, int start_x, int end_x){
3254     MpegEncContext * const s = &h->s;
3255     uint8_t  *dest_y, *dest_cb, *dest_cr;
3256     int linesize, uvlinesize, mb_x, mb_y;
3257     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3258     const int old_slice_type= h->slice_type;
3259     const int pixel_shift = h->pixel_shift;
3260
3261     if(h->deblocking_filter) {
3262         for(mb_x= start_x; mb_x<end_x; mb_x++){
3263             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3264                 int mb_xy, mb_type;
3265                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3266                 h->slice_num= h->slice_table[mb_xy];
3267                 mb_type = s->current_picture.f.mb_type[mb_xy];
3268                 h->list_count= h->list_counts[mb_xy];
3269
3270                 if(FRAME_MBAFF)
3271                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3272
3273                 s->mb_x= mb_x;
3274                 s->mb_y= mb_y;
3275                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3276                 dest_cb = s->current_picture.f.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * (8 << CHROMA444);
3277                 dest_cr = s->current_picture.f.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * (8 << CHROMA444);
3278                     //FIXME simplify above
3279
3280                 if (MB_FIELD) {
3281                     linesize   = h->mb_linesize   = s->linesize * 2;
3282                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3283                     if(mb_y&1){ //FIXME move out of this function?
3284                         dest_y -= s->linesize*15;
3285                         dest_cb-= s->uvlinesize*((8 << CHROMA444)-1);
3286                         dest_cr-= s->uvlinesize*((8 << CHROMA444)-1);
3287                     }
3288                 } else {
3289                     linesize   = h->mb_linesize   = s->linesize;
3290                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3291                 }
3292                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, CHROMA444, 0);
3293                 if(fill_filter_caches(h, mb_type))
3294                     continue;
3295                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3296                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3297
3298                 if (FRAME_MBAFF) {
3299                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3300                 } else {
3301                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3302                 }
3303             }
3304         }
3305     }
3306     h->slice_type= old_slice_type;
3307     s->mb_x= end_x;
3308     s->mb_y= end_mb_y - FRAME_MBAFF;
3309     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3310     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3311 }
3312
3313 static void predict_field_decoding_flag(H264Context *h){
3314     MpegEncContext * const s = &h->s;
3315     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3316     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3317                 ? s->current_picture.f.mb_type[mb_xy - 1]
3318                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3319                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3320                 : 0;
3321     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3322 }
3323
3324 /**
3325  * Draw edges and report progress for the last MB row.
3326  */
3327 static void decode_finish_row(H264Context *h){
3328     MpegEncContext * const s = &h->s;
3329     int top = 16*(s->mb_y >> FIELD_PICTURE);
3330     int height = 16 << FRAME_MBAFF;
3331     int deblock_border = (16 + 4) << FRAME_MBAFF;
3332     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3333
3334     if (h->deblocking_filter) {
3335         if((top + height) >= pic_height)
3336             height += deblock_border;
3337
3338         top -= deblock_border;
3339     }
3340
3341     if (top >= pic_height || (top + height) < h->emu_edge_height)
3342         return;
3343
3344     height = FFMIN(height, pic_height - top);
3345     if (top < h->emu_edge_height) {
3346         height = top+height;
3347         top = 0;
3348     }
3349
3350     ff_draw_horiz_band(s, top, height);
3351
3352     if (s->dropable) return;
3353
3354     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3355                              s->picture_structure==PICT_BOTTOM_FIELD);
3356 }
3357
3358 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3359     H264Context *h = *(void**)arg;
3360     MpegEncContext * const s = &h->s;
3361     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3362     int lf_x_start = s->mb_x;
3363
3364     s->mb_skip_run= -1;
3365
3366     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3367                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3368
3369     if( h->pps.cabac ) {
3370         /* realign */
3371         align_get_bits( &s->gb );
3372
3373         /* init cabac */
3374         ff_init_cabac_states( &h->cabac);
3375         ff_init_cabac_decoder( &h->cabac,
3376                                s->gb.buffer + get_bits_count(&s->gb)/8,
3377                                (get_bits_left(&s->gb) + 7)/8);
3378
3379         ff_h264_init_cabac_states(h);
3380
3381         for(;;){
3382 //START_TIMER
3383             int ret = ff_h264_decode_mb_cabac(h);
3384             int eos;
3385 //STOP_TIMER("decode_mb_cabac")
3386
3387             if(ret>=0) ff_h264_hl_decode_mb(h);
3388
3389             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3390                 s->mb_y++;
3391
3392                 ret = ff_h264_decode_mb_cabac(h);
3393
3394                 if(ret>=0) ff_h264_hl_decode_mb(h);
3395                 s->mb_y--;
3396             }
3397             eos = get_cabac_terminate( &h->cabac );
3398
3399             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3400                 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);
3401                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3402                 return 0;
3403             }
3404             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3405                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d, bytestream (%td)\n", s->mb_x, s->mb_y, h->cabac.bytestream_end - h->cabac.bytestream);
3406                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3407                 return -1;
3408             }
3409
3410             if( ++s->mb_x >= s->mb_width ) {
3411                 loop_filter(h, lf_x_start, s->mb_x);
3412                 s->mb_x = lf_x_start = 0;
3413                 decode_finish_row(h);
3414                 ++s->mb_y;
3415                 if(FIELD_OR_MBAFF_PICTURE) {
3416                     ++s->mb_y;
3417                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3418                         predict_field_decoding_flag(h);
3419                 }
3420             }
3421
3422             if( eos || s->mb_y >= s->mb_height ) {
3423                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3424                 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);
3425                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3426                 return 0;
3427             }
3428         }
3429
3430     } else {
3431         for(;;){
3432             int ret = ff_h264_decode_mb_cavlc(h);
3433
3434             if(ret>=0) ff_h264_hl_decode_mb(h);
3435
3436             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3437                 s->mb_y++;
3438                 ret = ff_h264_decode_mb_cavlc(h);
3439
3440                 if(ret>=0) ff_h264_hl_decode_mb(h);
3441                 s->mb_y--;
3442             }
3443
3444             if(ret<0){
3445                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3446                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3447                 return -1;
3448             }
3449
3450             if(++s->mb_x >= s->mb_width){
3451                 loop_filter(h, lf_x_start, s->mb_x);
3452                 s->mb_x = lf_x_start = 0;
3453                 decode_finish_row(h);
3454                 ++s->mb_y;
3455                 if(FIELD_OR_MBAFF_PICTURE) {
3456                     ++s->mb_y;
3457                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3458                         predict_field_decoding_flag(h);
3459                 }
3460                 if(s->mb_y >= s->mb_height){
3461                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3462
3463                     if(   get_bits_count(&s->gb) == s->gb.size_in_bits
3464                        || get_bits_count(&s->gb) <  s->gb.size_in_bits && s->avctx->error_recognition < FF_ER_AGGRESSIVE) {
3465                         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);
3466
3467                         r