h264dec: Support CODEC_FLAG2_SHOW_ALL.
[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     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
178     si=h->rbsp_buffer_size[bufidx];
179     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE+MAX_MBPAIR_SIZE);
180     dst= h->rbsp_buffer[bufidx];
181     if(si != h->rbsp_buffer_size[bufidx])
182         memset(dst + length, 0, FF_INPUT_BUFFER_PADDING_SIZE+MAX_MBPAIR_SIZE);
183
184     if (dst == NULL){
185         return NULL;
186     }
187
188     if(i>=length-1){ //no escaped 0
189         *dst_length= length;
190         *consumed= length+1; //+1 for the header
191         if(h->s.avctx->flags2 & CODEC_FLAG2_FAST){
192             return src;
193         }else{
194             memcpy(dst, src, length);
195             return dst;
196         }
197     }
198
199 //printf("decoding esc\n");
200     memcpy(dst, src, i);
201     si=di=i;
202     while(si+2<length){
203         //remove escapes (very rare 1:2^22)
204         if(src[si+2]>3){
205             dst[di++]= src[si++];
206             dst[di++]= src[si++];
207         }else if(src[si]==0 && src[si+1]==0){
208             if(src[si+2]==3){ //escape
209                 dst[di++]= 0;
210                 dst[di++]= 0;
211                 si+=3;
212                 continue;
213             }else //next start code
214                 goto nsc;
215         }
216
217         dst[di++]= src[si++];
218     }
219     while(si<length)
220         dst[di++]= src[si++];
221 nsc:
222
223     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
224
225     *dst_length= di;
226     *consumed= si + 1;//+1 for the header
227 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
228     return dst;
229 }
230
231 /**
232  * Identify the exact end of the bitstream
233  * @return the length of the trailing, or 0 if damaged
234  */
235 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
236     int v= *src;
237     int r;
238
239     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
240
241     for(r=1; r<9; r++){
242         if(v&1) return r;
243         v>>=1;
244     }
245     return 0;
246 }
247
248 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
249                                  int y_offset, int list){
250     int raw_my= h->mv_cache[list][ scan8[n] ][1];
251     int filter_height= (raw_my&3) ? 2 : 0;
252     int full_my= (raw_my>>2) + y_offset;
253     int top = full_my - filter_height, bottom = full_my + height + filter_height;
254
255     return FFMAX(abs(top), bottom);
256 }
257
258 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
259                                int y_offset, int list0, int list1, int *nrefs){
260     MpegEncContext * const s = &h->s;
261     int my;
262
263     y_offset += 16*(s->mb_y >> MB_FIELD);
264
265     if(list0){
266         int ref_n = h->ref_cache[0][ scan8[n] ];
267         Picture *ref= &h->ref_list[0][ref_n];
268
269         // Error resilience puts the current picture in the ref list.
270         // Don't try to wait on these as it will cause a deadlock.
271         // Fields can wait on each other, though.
272         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
273            (ref->f.reference & 3) != s->picture_structure) {
274             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
275             if (refs[0][ref_n] < 0) nrefs[0] += 1;
276             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
277         }
278     }
279
280     if(list1){
281         int ref_n = h->ref_cache[1][ scan8[n] ];
282         Picture *ref= &h->ref_list[1][ref_n];
283
284         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
285            (ref->f.reference & 3) != s->picture_structure) {
286             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
287             if (refs[1][ref_n] < 0) nrefs[1] += 1;
288             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
289         }
290     }
291 }
292
293 /**
294  * Wait until all reference frames are available for MC operations.
295  *
296  * @param h the H264 context
297  */
298 static void await_references(H264Context *h){
299     MpegEncContext * const s = &h->s;
300     const int mb_xy= h->mb_xy;
301     const int mb_type = s->current_picture.f.mb_type[mb_xy];
302     int refs[2][48];
303     int nrefs[2] = {0};
304     int ref, list;
305
306     memset(refs, -1, sizeof(refs));
307
308     if(IS_16X16(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     }else if(IS_16X8(mb_type)){
312         get_lowest_part_y(h, refs, 0, 8, 0,
313                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
314         get_lowest_part_y(h, refs, 8, 8, 8,
315                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
316     }else if(IS_8X16(mb_type)){
317         get_lowest_part_y(h, refs, 0, 16, 0,
318                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
319         get_lowest_part_y(h, refs, 4, 16, 0,
320                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
321     }else{
322         int i;
323
324         assert(IS_8X8(mb_type));
325
326         for(i=0; i<4; i++){
327             const int sub_mb_type= h->sub_mb_type[i];
328             const int n= 4*i;
329             int y_offset= (i&2)<<2;
330
331             if(IS_SUB_8X8(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             }else if(IS_SUB_8X4(sub_mb_type)){
335                 get_lowest_part_y(h, refs, n  , 4, y_offset,
336                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
337                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
338                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
339             }else if(IS_SUB_4X8(sub_mb_type)){
340                 get_lowest_part_y(h, refs, n  , 8, y_offset,
341                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
342                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
343                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
344             }else{
345                 int j;
346                 assert(IS_SUB_4X4(sub_mb_type));
347                 for(j=0; j<4; j++){
348                     int sub_y_offset= y_offset + 2*(j&2);
349                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
350                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
351                 }
352             }
353         }
354     }
355
356     for(list=h->list_count-1; list>=0; list--){
357         for(ref=0; ref<48 && nrefs[list]; ref++){
358             int row = refs[list][ref];
359             if(row >= 0){
360                 Picture *ref_pic = &h->ref_list[list][ref];
361                 int ref_field = ref_pic->f.reference - 1;
362                 int ref_field_picture = ref_pic->field_picture;
363                 int pic_height = 16*s->mb_height >> ref_field_picture;
364
365                 row <<= MB_MBAFF;
366                 nrefs[list]--;
367
368                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
369                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
370                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
371                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
372                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
373                 }else if(FIELD_PICTURE){
374                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
375                 }else{
376                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
377                 }
378             }
379         }
380     }
381 }
382
383 #if 0
384 /**
385  * DCT transforms the 16 dc values.
386  * @param qp quantization parameter ??? FIXME
387  */
388 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
389 //    const int qmul= dequant_coeff[qp][0];
390     int i;
391     int temp[16]; //FIXME check if this is a good idea
392     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
393     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
394
395     for(i=0; i<4; i++){
396         const int offset= y_offset[i];
397         const int z0= block[offset+stride*0] + block[offset+stride*4];
398         const int z1= block[offset+stride*0] - block[offset+stride*4];
399         const int z2= block[offset+stride*1] - block[offset+stride*5];
400         const int z3= block[offset+stride*1] + block[offset+stride*5];
401
402         temp[4*i+0]= z0+z3;
403         temp[4*i+1]= z1+z2;
404         temp[4*i+2]= z1-z2;
405         temp[4*i+3]= z0-z3;
406     }
407
408     for(i=0; i<4; i++){
409         const int offset= x_offset[i];
410         const int z0= temp[4*0+i] + temp[4*2+i];
411         const int z1= temp[4*0+i] - temp[4*2+i];
412         const int z2= temp[4*1+i] - temp[4*3+i];
413         const int z3= temp[4*1+i] + temp[4*3+i];
414
415         block[stride*0 +offset]= (z0 + z3)>>1;
416         block[stride*2 +offset]= (z1 + z2)>>1;
417         block[stride*8 +offset]= (z1 - z2)>>1;
418         block[stride*10+offset]= (z0 - z3)>>1;
419     }
420 }
421 #endif
422
423 #undef xStride
424 #undef stride
425
426 #if 0
427 static void chroma_dc_dct_c(DCTELEM *block){
428     const int stride= 16*2;
429     const int xStride= 16;
430     int a,b,c,d,e;
431
432     a= block[stride*0 + xStride*0];
433     b= block[stride*0 + xStride*1];
434     c= block[stride*1 + xStride*0];
435     d= block[stride*1 + xStride*1];
436
437     e= a-b;
438     a= a+b;
439     b= c-d;
440     c= c+d;
441
442     block[stride*0 + xStride*0]= (a+c);
443     block[stride*0 + xStride*1]= (e+b);
444     block[stride*1 + xStride*0]= (a-c);
445     block[stride*1 + xStride*1]= (e-b);
446 }
447 #endif
448
449 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
450                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
451                            int src_x_offset, int src_y_offset,
452                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
453                            int pixel_shift, int chroma444){
454     MpegEncContext * const s = &h->s;
455     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
456     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
457     const int luma_xy= (mx&3) + ((my&3)<<2);
458     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
459     uint8_t * src_y = pic->f.data[0] + offset;
460     uint8_t * src_cb, * src_cr;
461     int extra_width= h->emu_edge_width;
462     int extra_height= h->emu_edge_height;
463     int emu=0;
464     const int full_mx= mx>>2;
465     const int full_my= my>>2;
466     const int pic_width  = 16*s->mb_width;
467     const int pic_height = 16*s->mb_height >> MB_FIELD;
468
469     if(mx&7) extra_width -= 3;
470     if(my&7) extra_height -= 3;
471
472     if(   full_mx < 0-extra_width
473        || full_my < 0-extra_height
474        || full_mx + 16/*FIXME*/ > pic_width + extra_width
475        || full_my + 16/*FIXME*/ > pic_height + extra_height){
476         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);
477             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
478         emu=1;
479     }
480
481     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
482     if(!square){
483         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
484     }
485
486     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
487
488     if(chroma444){
489         src_cb = pic->f.data[1] + offset;
490         if(emu){
491             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
492                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
493             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
494         }
495         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
496         if(!square){
497             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
498         }
499
500         src_cr = pic->f.data[2] + offset;
501         if(emu){
502             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
503                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
504             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
505         }
506         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
507         if(!square){
508             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
509         }
510         return;
511     }
512
513     if(MB_FIELD){
514         // chroma offset when predicting from a field of opposite parity
515         my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
516         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
517     }
518     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
519     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
520
521     if(emu){
522         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);
523             src_cb= s->edge_emu_buffer;
524     }
525     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
526
527     if(emu){
528         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);
529             src_cr= s->edge_emu_buffer;
530     }
531     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
532 }
533
534 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
535                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
536                            int x_offset, int y_offset,
537                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
538                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
539                            int list0, int list1, int pixel_shift, int chroma444){
540     MpegEncContext * const s = &h->s;
541     qpel_mc_func *qpix_op=  qpix_put;
542     h264_chroma_mc_func chroma_op= chroma_put;
543
544     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
545     if(chroma444){
546         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
547         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
548     }else{
549         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
550         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
551     }
552     x_offset += 8*s->mb_x;
553     y_offset += 8*(s->mb_y >> MB_FIELD);
554
555     if(list0){
556         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
557         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
558                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
559                            qpix_op, chroma_op, pixel_shift, chroma444);
560
561         qpix_op=  qpix_avg;
562         chroma_op= chroma_avg;
563     }
564
565     if(list1){
566         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
567         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
568                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
569                            qpix_op, chroma_op, pixel_shift, chroma444);
570     }
571 }
572
573 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
574                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
575                            int x_offset, int y_offset,
576                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
577                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
578                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
579                            int list0, int list1, int pixel_shift, int chroma444){
580     MpegEncContext * const s = &h->s;
581
582     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
583     if(chroma444){
584         chroma_weight_avg = luma_weight_avg;
585         chroma_weight_op = luma_weight_op;
586         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
587         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
588     }else{
589         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
590         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
591     }
592     x_offset += 8*s->mb_x;
593     y_offset += 8*(s->mb_y >> MB_FIELD);
594
595     if(list0 && list1){
596         /* don't optimize for luma-only case, since B-frames usually
597          * use implicit weights => chroma too. */
598         uint8_t *tmp_cb = s->obmc_scratchpad;
599         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
600         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
601         int refn0 = h->ref_cache[0][ scan8[n] ];
602         int refn1 = h->ref_cache[1][ scan8[n] ];
603
604         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
605                     dest_y, dest_cb, dest_cr,
606                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
607         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
608                     tmp_y, tmp_cb, tmp_cr,
609                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
610
611         if(h->use_weight == 2){
612             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
613             int weight1 = 64 - weight0;
614             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
615             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
616             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
617         }else{
618             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
619                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
620                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
621             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
622                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
623                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
624             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
625                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
626                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
627         }
628     }else{
629         int list = list1 ? 1 : 0;
630         int refn = h->ref_cache[list][ scan8[n] ];
631         Picture *ref= &h->ref_list[list][refn];
632         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
633                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
634                     qpix_put, chroma_put, pixel_shift, chroma444);
635
636         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
637                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
638         if(h->use_weight_chroma){
639             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
640                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
641             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
642                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
643         }
644     }
645 }
646
647 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
648                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
649                            int x_offset, int y_offset,
650                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
651                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
652                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
653                            int list0, int list1, int pixel_shift, int chroma444){
654     if((h->use_weight==2 && list0 && list1
655         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
656        || h->use_weight==1)
657         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
658                          x_offset, y_offset, qpix_put, chroma_put,
659                          weight_op[0], weight_op[3], weight_avg[0],
660                          weight_avg[3], list0, list1, pixel_shift, chroma444);
661     else
662         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
663                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
664                     chroma_avg, list0, list1, pixel_shift, chroma444);
665 }
666
667 static inline void prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma444){
668     /* fetch pixels for estimated mv 4 macroblocks ahead
669      * optimized for 64byte cache lines */
670     MpegEncContext * const s = &h->s;
671     const int refn = h->ref_cache[list][scan8[0]];
672     if(refn >= 0){
673         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
674         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
675         uint8_t **src = h->ref_list[list][refn].f.data;
676         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
677         s->dsp.prefetch(src[0]+off, s->linesize, 4);
678         if(chroma444){
679             s->dsp.prefetch(src[1]+off, s->linesize, 4);
680             s->dsp.prefetch(src[2]+off, s->linesize, 4);
681         }else{
682             off= (((mx>>1)+64)<<pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize;
683             s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
684         }
685     }
686 }
687
688 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
689                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
690                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
691                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
692                       int pixel_shift, int chroma444){
693     MpegEncContext * const s = &h->s;
694     const int mb_xy= h->mb_xy;
695     const int mb_type = s->current_picture.f.mb_type[mb_xy];
696
697     assert(IS_INTER(mb_type));
698
699     if(HAVE_PTHREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
700         await_references(h);
701     prefetch_motion(h, 0, pixel_shift, chroma444);
702
703     if(IS_16X16(mb_type)){
704         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
705                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
706                 weight_op, weight_avg,
707                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
708                 pixel_shift, chroma444);
709     }else if(IS_16X8(mb_type)){
710         mc_part(h, 0, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
711                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
712                 &weight_op[1], &weight_avg[1],
713                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
714                 pixel_shift, chroma444);
715         mc_part(h, 8, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
716                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
717                 &weight_op[1], &weight_avg[1],
718                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
719                 pixel_shift, chroma444);
720     }else if(IS_8X16(mb_type)){
721         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
722                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
723                 &weight_op[2], &weight_avg[2],
724                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
725                 pixel_shift, chroma444);
726         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
727                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
728                 &weight_op[2], &weight_avg[2],
729                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
730                 pixel_shift, chroma444);
731     }else{
732         int i;
733
734         assert(IS_8X8(mb_type));
735
736         for(i=0; i<4; i++){
737             const int sub_mb_type= h->sub_mb_type[i];
738             const int n= 4*i;
739             int x_offset= (i&1)<<2;
740             int y_offset= (i&2)<<1;
741
742             if(IS_SUB_8X8(sub_mb_type)){
743                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
744                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
745                     &weight_op[3], &weight_avg[3],
746                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
747                     pixel_shift, chroma444);
748             }else if(IS_SUB_8X4(sub_mb_type)){
749                 mc_part(h, n  , 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
750                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
751                     &weight_op[4], &weight_avg[4],
752                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
753                     pixel_shift, chroma444);
754                 mc_part(h, n+2, 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
755                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
756                     &weight_op[4], &weight_avg[4],
757                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
758                     pixel_shift, chroma444);
759             }else if(IS_SUB_4X8(sub_mb_type)){
760                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
761                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
762                     &weight_op[5], &weight_avg[5],
763                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
764                     pixel_shift, chroma444);
765                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
766                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
767                     &weight_op[5], &weight_avg[5],
768                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
769                     pixel_shift, chroma444);
770             }else{
771                 int j;
772                 assert(IS_SUB_4X4(sub_mb_type));
773                 for(j=0; j<4; j++){
774                     int sub_x_offset= x_offset + 2*(j&1);
775                     int sub_y_offset= y_offset +   (j&2);
776                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
777                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
778                         &weight_op[6], &weight_avg[6],
779                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
780                         pixel_shift, chroma444);
781                 }
782             }
783         }
784     }
785
786     prefetch_motion(h, 1, pixel_shift, chroma444);
787 }
788
789 static void free_tables(H264Context *h, int free_rbsp){
790     int i;
791     H264Context *hx;
792
793     av_freep(&h->intra4x4_pred_mode);
794     av_freep(&h->chroma_pred_mode_table);
795     av_freep(&h->cbp_table);
796     av_freep(&h->mvd_table[0]);
797     av_freep(&h->mvd_table[1]);
798     av_freep(&h->direct_table);
799     av_freep(&h->non_zero_count);
800     av_freep(&h->slice_table_base);
801     h->slice_table= NULL;
802     av_freep(&h->list_counts);
803
804     av_freep(&h->mb2b_xy);
805     av_freep(&h->mb2br_xy);
806
807     for(i = 0; i < MAX_THREADS; i++) {
808         hx = h->thread_context[i];
809         if(!hx) continue;
810         av_freep(&hx->top_borders[1]);
811         av_freep(&hx->top_borders[0]);
812         av_freep(&hx->s.obmc_scratchpad);
813         if (free_rbsp){
814             av_freep(&hx->rbsp_buffer[1]);
815             av_freep(&hx->rbsp_buffer[0]);
816             hx->rbsp_buffer_size[0] = 0;
817             hx->rbsp_buffer_size[1] = 0;
818         }
819         if (i) av_freep(&h->thread_context[i]);
820     }
821 }
822
823 static void init_dequant8_coeff_table(H264Context *h){
824     int i,j,q,x;
825     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
826
827     for(i=0; i<6; i++ ){
828         h->dequant8_coeff[i] = h->dequant8_buffer[i];
829         for(j=0; j<i; j++){
830             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
831                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
832                 break;
833             }
834         }
835         if(j<i)
836             continue;
837
838         for(q=0; q<max_qp+1; q++){
839             int shift = div6[q];
840             int idx = rem6[q];
841             for(x=0; x<64; x++)
842                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
843                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
844                     h->pps.scaling_matrix8[i][x]) << shift;
845         }
846     }
847 }
848
849 static void init_dequant4_coeff_table(H264Context *h){
850     int i,j,q,x;
851     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
852     for(i=0; i<6; i++ ){
853         h->dequant4_coeff[i] = h->dequant4_buffer[i];
854         for(j=0; j<i; j++){
855             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
856                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
857                 break;
858             }
859         }
860         if(j<i)
861             continue;
862
863         for(q=0; q<max_qp+1; q++){
864             int shift = div6[q] + 2;
865             int idx = rem6[q];
866             for(x=0; x<16; x++)
867                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
868                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
869                     h->pps.scaling_matrix4[i][x]) << shift;
870         }
871     }
872 }
873
874 static void init_dequant_tables(H264Context *h){
875     int i,x;
876     init_dequant4_coeff_table(h);
877     if(h->pps.transform_8x8_mode)
878         init_dequant8_coeff_table(h);
879     if(h->sps.transform_bypass){
880         for(i=0; i<6; i++)
881             for(x=0; x<16; x++)
882                 h->dequant4_coeff[i][0][x] = 1<<6;
883         if(h->pps.transform_8x8_mode)
884             for(i=0; i<6; i++)
885                 for(x=0; x<64; x++)
886                     h->dequant8_coeff[i][0][x] = 1<<6;
887     }
888 }
889
890
891 int ff_h264_alloc_tables(H264Context *h){
892     MpegEncContext * const s = &h->s;
893     const int big_mb_num= s->mb_stride * (s->mb_height+1);
894     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
895     int x,y;
896
897     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
898
899     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
900     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
901     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
902
903     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
904     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
905     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
906     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
907     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
908
909     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
910     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
911
912     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
913     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
914     for(y=0; y<s->mb_height; y++){
915         for(x=0; x<s->mb_width; x++){
916             const int mb_xy= x + y*s->mb_stride;
917             const int b_xy = 4*x + 4*y*h->b_stride;
918
919             h->mb2b_xy [mb_xy]= b_xy;
920             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
921         }
922     }
923
924     s->obmc_scratchpad = NULL;
925
926     if(!h->dequant4_coeff[0])
927         init_dequant_tables(h);
928
929     return 0;
930 fail:
931     free_tables(h, 1);
932     return -1;
933 }
934
935 /**
936  * Mimic alloc_tables(), but for every context thread.
937  */
938 static void clone_tables(H264Context *dst, H264Context *src, int i){
939     MpegEncContext * const s = &src->s;
940     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
941     dst->non_zero_count           = src->non_zero_count;
942     dst->slice_table              = src->slice_table;
943     dst->cbp_table                = src->cbp_table;
944     dst->mb2b_xy                  = src->mb2b_xy;
945     dst->mb2br_xy                 = src->mb2br_xy;
946     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
947     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
948     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
949     dst->direct_table             = src->direct_table;
950     dst->list_counts              = src->list_counts;
951
952     dst->s.obmc_scratchpad = NULL;
953     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma, src->sps.chroma_format_idc);
954 }
955
956 /**
957  * Init context
958  * Allocate buffers which are not shared amongst multiple threads.
959  */
960 static int context_init(H264Context *h){
961     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
962     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
963
964     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
965     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
966
967     return 0;
968 fail:
969     return -1; // free_tables will clean up for us
970 }
971
972 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
973
974 static av_cold void common_init(H264Context *h){
975     MpegEncContext * const s = &h->s;
976
977     s->width = s->avctx->width;
978     s->height = s->avctx->height;
979     s->codec_id= s->avctx->codec->id;
980
981     ff_h264dsp_init(&h->h264dsp, 8, 1);
982     ff_h264_pred_init(&h->hpc, s->codec_id, 8, 1);
983
984     h->dequant_coeff_pps= -1;
985     s->unrestricted_mv=1;
986     s->decode=1; //FIXME
987
988     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
989
990     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
991     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
992 }
993
994 int ff_h264_decode_extradata(H264Context *h)
995 {
996     AVCodecContext *avctx = h->s.avctx;
997
998     if(avctx->extradata[0] == 1){
999         int i, cnt, nalsize;
1000         unsigned char *p = avctx->extradata;
1001
1002         h->is_avc = 1;
1003
1004         if(avctx->extradata_size < 7) {
1005             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1006             return -1;
1007         }
1008         /* sps and pps in the avcC always have length coded with 2 bytes,
1009            so put a fake nal_length_size = 2 while parsing them */
1010         h->nal_length_size = 2;
1011         // Decode sps from avcC
1012         cnt = *(p+5) & 0x1f; // Number of sps
1013         p += 6;
1014         for (i = 0; i < cnt; i++) {
1015             nalsize = AV_RB16(p) + 2;
1016             if(decode_nal_units(h, p, nalsize) < 0) {
1017                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1018                 return -1;
1019             }
1020             p += nalsize;
1021         }
1022         // Decode pps from avcC
1023         cnt = *(p++); // Number of pps
1024         for (i = 0; i < cnt; i++) {
1025             nalsize = AV_RB16(p) + 2;
1026             if (decode_nal_units(h, p, nalsize) < 0) {
1027                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1028                 return -1;
1029             }
1030             p += nalsize;
1031         }
1032         // Now store right nal length size, that will be use to parse all other nals
1033         h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1034     } else {
1035         h->is_avc = 0;
1036         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1037             return -1;
1038     }
1039     return 0;
1040 }
1041
1042 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1043     H264Context *h= avctx->priv_data;
1044     MpegEncContext * const s = &h->s;
1045
1046     MPV_decode_defaults(s);
1047
1048     s->avctx = avctx;
1049     common_init(h);
1050
1051     s->out_format = FMT_H264;
1052     s->workaround_bugs= avctx->workaround_bugs;
1053
1054     // set defaults
1055 //    s->decode_mb= ff_h263_decode_mb;
1056     s->quarter_sample = 1;
1057     if(!avctx->has_b_frames)
1058     s->low_delay= 1;
1059
1060     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1061
1062     ff_h264_decode_init_vlc();
1063
1064     h->pixel_shift = 0;
1065     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1066
1067     h->thread_context[0] = h;
1068     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1069     h->prev_poc_msb= 1<<16;
1070     h->x264_build = -1;
1071     ff_h264_reset_sei(h);
1072     if(avctx->codec_id == CODEC_ID_H264){
1073         if(avctx->ticks_per_frame == 1){
1074             s->avctx->time_base.den *=2;
1075         }
1076         avctx->ticks_per_frame = 2;
1077     }
1078
1079     if(avctx->extradata_size > 0 && avctx->extradata &&
1080         ff_h264_decode_extradata(h))
1081         return -1;
1082
1083     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1084         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1085         s->low_delay = 0;
1086     }
1087
1088     return 0;
1089 }
1090
1091 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1092 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1093 {
1094     int i;
1095
1096     for (i=0; i<count; i++){
1097         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1098                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1099                 !from[i]));
1100         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1101     }
1102 }
1103
1104 static void copy_parameter_set(void **to, void **from, int count, int size)
1105 {
1106     int i;
1107
1108     for (i=0; i<count; i++){
1109         if (to[i] && !from[i]) av_freep(&to[i]);
1110         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1111
1112         if (from[i]) memcpy(to[i], from[i], size);
1113     }
1114 }
1115
1116 static int decode_init_thread_copy(AVCodecContext *avctx){
1117     H264Context *h= avctx->priv_data;
1118
1119     if (!avctx->is_copy) return 0;
1120     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1121     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1122
1123     return 0;
1124 }
1125
1126 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1127 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1128     H264Context *h= dst->priv_data, *h1= src->priv_data;
1129     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1130     int inited = s->context_initialized, err;
1131     int i;
1132
1133     if(dst == src || !s1->context_initialized) return 0;
1134
1135     err = ff_mpeg_update_thread_context(dst, src);
1136     if(err) return err;
1137
1138     //FIXME handle width/height changing
1139     if(!inited){
1140         for(i = 0; i < MAX_SPS_COUNT; i++)
1141             av_freep(h->sps_buffers + i);
1142
1143         for(i = 0; i < MAX_PPS_COUNT; i++)
1144             av_freep(h->pps_buffers + i);
1145
1146         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1147         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1148         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1149         if (ff_h264_alloc_tables(h) < 0) {
1150             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1151             return AVERROR(ENOMEM);
1152         }
1153         context_init(h);
1154
1155         for(i=0; i<2; i++){
1156             h->rbsp_buffer[i] = NULL;
1157             h->rbsp_buffer_size[i] = 0;
1158         }
1159
1160         h->thread_context[0] = h;
1161
1162         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1163         // so this has to be allocated here
1164         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1165
1166         s->dsp.clear_blocks(h->mb);
1167         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1168     }
1169
1170     //extradata/NAL handling
1171     h->is_avc          = h1->is_avc;
1172
1173     //SPS/PPS
1174     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1175     h->sps             = h1->sps;
1176     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1177     h->pps             = h1->pps;
1178
1179     //Dequantization matrices
1180     //FIXME these are big - can they be only copied when PPS changes?
1181     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1182
1183     for(i=0; i<6; i++)
1184         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1185
1186     for(i=0; i<6; i++)
1187         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1188
1189     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1190
1191     //POC timing
1192     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1193
1194     //reference lists
1195     copy_fields(h, h1, ref_count, list_count);
1196     copy_fields(h, h1, ref_list,  intra_gb);
1197     copy_fields(h, h1, short_ref, cabac_init_idc);
1198
1199     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1200     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1201     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1202
1203     h->last_slice_type = h1->last_slice_type;
1204
1205     if(!s->current_picture_ptr) return 0;
1206
1207     if(!s->dropable) {
1208         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1209         h->prev_poc_msb     = h->poc_msb;
1210         h->prev_poc_lsb     = h->poc_lsb;
1211     }
1212     h->prev_frame_num_offset= h->frame_num_offset;
1213     h->prev_frame_num       = h->frame_num;
1214     h->outputed_poc         = h->next_outputed_poc;
1215
1216     return err;
1217 }
1218
1219 int ff_h264_frame_start(H264Context *h){
1220     MpegEncContext * const s = &h->s;
1221     int i;
1222     const int pixel_shift = h->pixel_shift;
1223     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1224
1225     if(MPV_frame_start(s, s->avctx) < 0)
1226         return -1;
1227     ff_er_frame_start(s);
1228     /*
1229      * MPV_frame_start uses pict_type to derive key_frame.
1230      * This is incorrect for H.264; IDR markings must be used.
1231      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1232      * See decode_nal_units().
1233      */
1234     s->current_picture_ptr->f.key_frame = 0;
1235     s->current_picture_ptr->mmco_reset= 0;
1236
1237     assert(s->linesize && s->uvlinesize);
1238
1239     for(i=0; i<16; i++){
1240         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1241         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1242     }
1243     for(i=0; i<16; i++){
1244         h->block_offset[16+i]=
1245         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1246         h->block_offset[48+16+i]=
1247         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1248     }
1249
1250     /* can't be in alloc_tables because linesize isn't known there.
1251      * FIXME: redo bipred weight to not require extra buffer? */
1252     for(i = 0; i < thread_count; i++)
1253         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1254             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1255
1256     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1257     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1258
1259 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1260
1261     // We mark the current picture as non-reference after allocating it, so
1262     // that if we break out due to an error it can be released automatically
1263     // in the next MPV_frame_start().
1264     // SVQ3 as well as most other codecs have only last/next/current and thus
1265     // get released even with set reference, besides SVQ3 and others do not
1266     // mark frames as reference later "naturally".
1267     if(s->codec_id != CODEC_ID_SVQ3)
1268         s->current_picture_ptr->f.reference = 0;
1269
1270     s->current_picture_ptr->field_poc[0]=
1271     s->current_picture_ptr->field_poc[1]= INT_MAX;
1272
1273     h->next_output_pic = NULL;
1274
1275     assert(s->current_picture_ptr->long_ref==0);
1276
1277     return 0;
1278 }
1279
1280 /**
1281   * Run setup operations that must be run after slice header decoding.
1282   * This includes finding the next displayed frame.
1283   *
1284   * @param h h264 master context
1285   * @param setup_finished enough NALs have been read that we can call
1286   * ff_thread_finish_setup()
1287   */
1288 static void decode_postinit(H264Context *h, int setup_finished){
1289     MpegEncContext * const s = &h->s;
1290     Picture *out = s->current_picture_ptr;
1291     Picture *cur = s->current_picture_ptr;
1292     int i, pics, out_of_order, out_idx;
1293
1294     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1295     s->current_picture_ptr->f.pict_type   = s->pict_type;
1296
1297     if (h->next_output_pic) return;
1298
1299     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1300         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1301         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1302         //to find this yet, so we assume the worst for now.
1303         //if (setup_finished)
1304         //    ff_thread_finish_setup(s->avctx);
1305         return;
1306     }
1307
1308     cur->f.interlaced_frame = 0;
1309     cur->f.repeat_pict      = 0;
1310
1311     /* Signal interlacing information externally. */
1312     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1313
1314     if(h->sps.pic_struct_present_flag){
1315         switch (h->sei_pic_struct)
1316         {
1317         case SEI_PIC_STRUCT_FRAME:
1318             break;
1319         case SEI_PIC_STRUCT_TOP_FIELD:
1320         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1321             cur->f.interlaced_frame = 1;
1322             break;
1323         case SEI_PIC_STRUCT_TOP_BOTTOM:
1324         case SEI_PIC_STRUCT_BOTTOM_TOP:
1325             if (FIELD_OR_MBAFF_PICTURE)
1326                 cur->f.interlaced_frame = 1;
1327             else
1328                 // try to flag soft telecine progressive
1329                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1330             break;
1331         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1332         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1333             // Signal the possibility of telecined film externally (pic_struct 5,6)
1334             // From these hints, let the applications decide if they apply deinterlacing.
1335             cur->f.repeat_pict = 1;
1336             break;
1337         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1338             // Force progressive here, as doubling interlaced frame is a bad idea.
1339             cur->f.repeat_pict = 2;
1340             break;
1341         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1342             cur->f.repeat_pict = 4;
1343             break;
1344         }
1345
1346         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1347             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1348     }else{
1349         /* Derive interlacing flag from used decoding process. */
1350         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1351     }
1352     h->prev_interlaced_frame = cur->f.interlaced_frame;
1353
1354     if (cur->field_poc[0] != cur->field_poc[1]){
1355         /* Derive top_field_first from field pocs. */
1356         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1357     }else{
1358         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1359             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1360             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1361               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1362                 cur->f.top_field_first = 1;
1363             else
1364                 cur->f.top_field_first = 0;
1365         }else{
1366             /* Most likely progressive */
1367             cur->f.top_field_first = 0;
1368         }
1369     }
1370
1371     //FIXME do something with unavailable reference frames
1372
1373     /* Sort B-frames into display order */
1374
1375     if(h->sps.bitstream_restriction_flag
1376        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1377         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1378         s->low_delay = 0;
1379     }
1380
1381     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1382        && !h->sps.bitstream_restriction_flag){
1383         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1384         s->low_delay= 0;
1385     }
1386
1387     pics = 0;
1388     while(h->delayed_pic[pics]) pics++;
1389
1390     av_assert0(pics <= MAX_DELAYED_PIC_COUNT);
1391
1392     h->delayed_pic[pics++] = cur;
1393     if (cur->f.reference == 0)
1394         cur->f.reference = DELAYED_PIC_REF;
1395
1396     out = h->delayed_pic[0];
1397     out_idx = 0;
1398     for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1399         if(h->delayed_pic[i]->poc < out->poc){
1400             out = h->delayed_pic[i];
1401             out_idx = i;
1402         }
1403     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1404         h->next_outputed_poc= INT_MIN;
1405     out_of_order = out->poc < h->next_outputed_poc;
1406
1407     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1408         { }
1409     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1410        || (s->low_delay &&
1411         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1412          || cur->f.pict_type == AV_PICTURE_TYPE_B)))
1413     {
1414         s->low_delay = 0;
1415         s->avctx->has_b_frames++;
1416     }
1417
1418     if(out_of_order || pics > s->avctx->has_b_frames){
1419         out->f.reference &= ~DELAYED_PIC_REF;
1420         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1421                          // or else the first thread can release the picture and reuse it unsafely
1422         for(i=out_idx; h->delayed_pic[i]; i++)
1423             h->delayed_pic[i] = h->delayed_pic[i+1];
1424     }
1425     if(!out_of_order && pics > s->avctx->has_b_frames){
1426         h->next_output_pic = out;
1427         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1428             h->next_outputed_poc = INT_MIN;
1429         } else
1430             h->next_outputed_poc = out->poc;
1431     }else{
1432         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1433     }
1434
1435     if (setup_finished)
1436         ff_thread_finish_setup(s->avctx);
1437 }
1438
1439 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int simple){
1440     MpegEncContext * const s = &h->s;
1441     uint8_t *top_border;
1442     int top_idx = 1;
1443     const int pixel_shift = h->pixel_shift;
1444     int chroma444 = CHROMA444;
1445     int chroma422 = CHROMA422;
1446
1447     src_y  -=   linesize;
1448     src_cb -= uvlinesize;
1449     src_cr -= uvlinesize;
1450
1451     if(!simple && FRAME_MBAFF){
1452         if(s->mb_y&1){
1453             if(!MB_MBAFF){
1454                 top_border = h->top_borders[0][s->mb_x];
1455                 AV_COPY128(top_border, src_y + 15*linesize);
1456                 if (pixel_shift)
1457                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1458                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1459                     if(chroma444){
1460                         if (pixel_shift){
1461                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1462                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1463                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1464                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1465                         } else {
1466                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1467                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1468                         }
1469                     } else if(chroma422){
1470                         if (pixel_shift) {
1471                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1472                             AV_COPY128(top_border+48, src_cr + 15*uvlinesize);
1473                         } else {
1474                             AV_COPY64(top_border+16, src_cb +  15*uvlinesize);
1475                             AV_COPY64(top_border+24, src_cr +  15*uvlinesize);
1476                         }
1477                     } else {
1478                         if (pixel_shift) {
1479                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1480                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1481                         } else {
1482                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1483                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1484                         }
1485                     }
1486                 }
1487             }
1488         }else if(MB_MBAFF){
1489             top_idx = 0;
1490         }else
1491             return;
1492     }
1493
1494     top_border = h->top_borders[top_idx][s->mb_x];
1495     // There are two lines saved, the line above the the top macroblock of a pair,
1496     // and the line above the bottom macroblock
1497     AV_COPY128(top_border, src_y + 16*linesize);
1498     if (pixel_shift)
1499         AV_COPY128(top_border+16, src_y+16*linesize+16);
1500
1501     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1502         if(chroma444){
1503             if (pixel_shift){
1504                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1505                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1506                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1507                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1508             } else {
1509                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1510                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1511             }
1512         } else if(chroma422) {
1513             if (pixel_shift) {
1514                 AV_COPY128(top_border+32, src_cb+16*uvlinesize);
1515                 AV_COPY128(top_border+48, src_cr+16*uvlinesize);
1516             } else {
1517                 AV_COPY64(top_border+16, src_cb+16*uvlinesize);
1518                 AV_COPY64(top_border+24, src_cr+16*uvlinesize);
1519             }
1520         } else {
1521             if (pixel_shift) {
1522                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1523                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1524             } else {
1525                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1526                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1527             }
1528         }
1529     }
1530 }
1531
1532 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1533                                   uint8_t *src_cb, uint8_t *src_cr,
1534                                   int linesize, int uvlinesize,
1535                                   int xchg, int chroma444,
1536                                   int simple, int pixel_shift){
1537     MpegEncContext * const s = &h->s;
1538     int deblock_topleft;
1539     int deblock_top;
1540     int top_idx = 1;
1541     uint8_t *top_border_m1;
1542     uint8_t *top_border;
1543
1544     if(!simple && FRAME_MBAFF){
1545         if(s->mb_y&1){
1546             if(!MB_MBAFF)
1547                 return;
1548         }else{
1549             top_idx = MB_MBAFF ? 0 : 1;
1550         }
1551     }
1552
1553     if(h->deblocking_filter == 2) {
1554         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1555         deblock_top     = h->top_type;
1556     } else {
1557         deblock_topleft = (s->mb_x > 0);
1558         deblock_top     = (s->mb_y > !!MB_FIELD);
1559     }
1560
1561     src_y  -=   linesize + 1 + pixel_shift;
1562     src_cb -= uvlinesize + 1 + pixel_shift;
1563     src_cr -= uvlinesize + 1 + pixel_shift;
1564
1565     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1566     top_border    = h->top_borders[top_idx][s->mb_x];
1567
1568 #define XCHG(a,b,xchg)\
1569     if (pixel_shift) {\
1570         if (xchg) {\
1571             AV_SWAP64(b+0,a+0);\
1572             AV_SWAP64(b+8,a+8);\
1573         } else {\
1574             AV_COPY128(b,a); \
1575         }\
1576     } else \
1577 if (xchg) AV_SWAP64(b,a);\
1578 else      AV_COPY64(b,a);
1579
1580     if(deblock_top){
1581         if(deblock_topleft){
1582             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1583         }
1584         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1585         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1586         if(s->mb_x+1 < s->mb_width){
1587             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1588         }
1589     }
1590     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1591         if(chroma444){
1592             if(deblock_topleft){
1593                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1594                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1595             }
1596             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1597             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1598             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1599             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1600             if(s->mb_x+1 < s->mb_width){
1601                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1602                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1603             }
1604         } else {
1605             if(deblock_top){
1606                 if(deblock_topleft){
1607                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1608                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1609                 }
1610                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1611                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1612             }
1613         }
1614     }
1615 }
1616
1617 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1618     if (high_bit_depth) {
1619         return AV_RN32A(((int32_t*)mb) + index);
1620     } else
1621         return AV_RN16A(mb + index);
1622 }
1623
1624 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1625     if (high_bit_depth) {
1626         AV_WN32A(((int32_t*)mb) + index, value);
1627     } else
1628         AV_WN16A(mb + index, value);
1629 }
1630
1631 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1632                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1633 {
1634     MpegEncContext * const s = &h->s;
1635     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1636     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1637     int i;
1638     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1639     block_offset += 16*p;
1640     if(IS_INTRA4x4(mb_type)){
1641         if(simple || !s->encoding){
1642             if(IS_8x8DCT(mb_type)){
1643                 if(transform_bypass){
1644                     idct_dc_add =
1645                     idct_add    = s->dsp.add_pixels8;
1646                 }else{
1647                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1648                     idct_add    = h->h264dsp.h264_idct8_add;
1649                 }
1650                 for(i=0; i<16; i+=4){
1651                     uint8_t * const ptr= dest_y + block_offset[i];
1652                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1653                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1654                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1655                     }else{
1656                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1657                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1658                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1659                         if(nnz){
1660                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1661                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1662                             else
1663                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1664                         }
1665                     }
1666                 }
1667             }else{
1668                 if(transform_bypass){
1669                     idct_dc_add =
1670                     idct_add    = s->dsp.add_pixels4;
1671                 }else{
1672                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1673                     idct_add    = h->h264dsp.h264_idct_add;
1674                 }
1675                 for(i=0; i<16; i++){
1676                     uint8_t * const ptr= dest_y + block_offset[i];
1677                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1678
1679                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1680                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1681                     }else{
1682                         uint8_t *topright;
1683                         int nnz, tr;
1684                         uint64_t tr_high;
1685                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1686                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1687                             assert(s->mb_y || linesize <= block_offset[i]);
1688                             if(!topright_avail){
1689                                 if (pixel_shift) {
1690                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1691                                     topright= (uint8_t*) &tr_high;
1692                                 } else {
1693                                     tr= ptr[3 - linesize]*0x01010101;
1694                                     topright= (uint8_t*) &tr;
1695                                 }
1696                             }else
1697                                 topright= ptr + (4 << pixel_shift) - linesize;
1698                         }else
1699                             topright= NULL;
1700
1701                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1702                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1703                         if(nnz){
1704                             if(is_h264){
1705                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1706                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1707                                 else
1708                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1709                             }else
1710                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1711                         }
1712                     }
1713                 }
1714             }
1715         }
1716     }else{
1717         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1718         if(is_h264){
1719             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1720                 if(!transform_bypass)
1721                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1722                 else{
1723                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1724                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1725                     for(i = 0; i < 16; i++)
1726                         dctcoef_set(h->mb+p*256, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1727                 }
1728             }
1729         }else
1730             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1731     }
1732 }
1733
1734 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1735                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1736 {
1737     MpegEncContext * const s = &h->s;
1738     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1739     int i;
1740     block_offset += 16*p;
1741     if(!IS_INTRA4x4(mb_type)){
1742         if(is_h264){
1743             if(IS_INTRA16x16(mb_type)){
1744                 if(transform_bypass){
1745                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1746                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1747                     }else{
1748                         for(i=0; i<16; i++){
1749                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1750                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1751                         }
1752                     }
1753                 }else{
1754                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1755                 }
1756             }else if(h->cbp&15){
1757                 if(transform_bypass){
1758                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1759                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1760                     for(i=0; i<16; i+=di){
1761                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1762                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1763                         }
1764                     }
1765                 }else{
1766                     if(IS_8x8DCT(mb_type)){
1767                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1768                     }else{
1769                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1770                     }
1771                 }
1772             }
1773         }else{
1774             for(i=0; i<16; i++){
1775                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1776                     uint8_t * const ptr= dest_y + block_offset[i];
1777                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1778                 }
1779             }
1780         }
1781     }
1782 }
1783
1784 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1785     MpegEncContext * const s = &h->s;
1786     const int mb_x= s->mb_x;
1787     const int mb_y= s->mb_y;
1788     const int mb_xy= h->mb_xy;
1789     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1790     uint8_t  *dest_y, *dest_cb, *dest_cr;
1791     int linesize, uvlinesize /*dct_offset*/;
1792     int i, j;
1793     int *block_offset = &h->block_offset[0];
1794     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1795     /* is_h264 should always be true if SVQ3 is disabled. */
1796     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1797     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1798     const int block_h = 16>>s->chroma_y_shift;
1799
1800     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1801     dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1802     dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1803
1804     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1805     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1806
1807     h->list_counts[mb_xy]= h->list_count;
1808
1809     if (!simple && MB_FIELD) {
1810         linesize   = h->mb_linesize   = s->linesize * 2;
1811         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1812         block_offset = &h->block_offset[48];
1813         if(mb_y&1){ //FIXME move out of this function?
1814             dest_y -= s->linesize*15;
1815             dest_cb-= s->uvlinesize*(block_h-1);
1816             dest_cr-= s->uvlinesize*(block_h-1);
1817         }
1818         if(FRAME_MBAFF) {
1819             int list;
1820             for(list=0; list<h->list_count; list++){
1821                 if(!USES_LIST(mb_type, list))
1822                     continue;
1823                 if(IS_16X16(mb_type)){
1824                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1825                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1826                 }else{
1827                     for(i=0; i<16; i+=4){
1828                         int ref = h->ref_cache[list][scan8[i]];
1829                         if(ref >= 0)
1830                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1831                     }
1832                 }
1833             }
1834         }
1835     } else {
1836         linesize   = h->mb_linesize   = s->linesize;
1837         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1838 //        dct_offset = s->linesize * 16;
1839     }
1840
1841     if (!simple && IS_INTRA_PCM(mb_type)) {
1842         const int bit_depth = h->sps.bit_depth_luma;
1843         if (pixel_shift) {
1844             int j;
1845             GetBitContext gb;
1846             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1847
1848             for (i = 0; i < 16; i++) {
1849                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1850                 for (j = 0; j < 16; j++)
1851                     tmp_y[j] = get_bits(&gb, bit_depth);
1852             }
1853             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1854                 if (!h->sps.chroma_format_idc) {
1855                     for (i = 0; i < 8; i++) {
1856                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1857                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1858                         for (j = 0; j < 8; j++) {
1859                             tmp_cb[j] = tmp_cr[j] = 1 << (bit_depth - 1);
1860                         }
1861                     }
1862                 } else {
1863                     for (i = 0; i < block_h; i++) {
1864                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1865                         for (j = 0; j < 8; j++)
1866                             tmp_cb[j] = get_bits(&gb, bit_depth);
1867                     }
1868                     for (i = 0; i < block_h; i++) {
1869                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1870                         for (j = 0; j < 8; j++)
1871                             tmp_cr[j] = get_bits(&gb, bit_depth);
1872                     }
1873                 }
1874             }
1875         } else {
1876             for (i=0; i<16; i++) {
1877                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1878             }
1879             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1880                 if (!h->sps.chroma_format_idc) {
1881                     for (i=0; i<8; i++) {
1882                         memset(dest_cb+ i*uvlinesize, 1 << (bit_depth - 1), 8);
1883                         memset(dest_cr+ i*uvlinesize, 1 << (bit_depth - 1), 8);
1884                     }
1885                 } else {
1886                     for (i=0; i<block_h; i++) {
1887                         memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
1888                         memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
1889                     }
1890                 }
1891             }
1892         }
1893     } else {
1894         if(IS_INTRA(mb_type)){
1895             if(h->deblocking_filter)
1896                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1897
1898             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1899                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1900                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1901             }
1902
1903             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1904
1905             if(h->deblocking_filter)
1906                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1907         }else if(is_h264){
1908             hl_motion(h, dest_y, dest_cb, dest_cr,
1909                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1910                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1911                       h->h264dsp.weight_h264_pixels_tab,
1912                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 0);
1913         }
1914
1915         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1916
1917         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1918             uint8_t *dest[2] = {dest_cb, dest_cr};
1919             if(transform_bypass){
1920                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1921                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
1922                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
1923                 }else{
1924                     idct_add = s->dsp.add_pixels4;
1925                     for(j=1; j<3; j++){
1926                         for(i=j*16; i<j*16+4; i++){
1927                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1928                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1929                         }
1930                     }
1931                 }
1932             }else{
1933                 if(is_h264){
1934                     int qp[2];
1935                     if (CHROMA422) {
1936                         qp[0] = h->chroma_qp[0]+3;
1937                         qp[1] = h->chroma_qp[1]+3;
1938                     } else {
1939                         qp[0] = h->chroma_qp[0];
1940                         qp[1] = h->chroma_qp[1];
1941                     }
1942                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1943                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*1 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][qp[0]][0]);
1944                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1945                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*2 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][qp[1]][0]);
1946                     h->h264dsp.h264_idct_add8(dest, block_offset,
1947                                               h->mb, uvlinesize,
1948                                               h->non_zero_count_cache);
1949                 }
1950 #if CONFIG_SVQ3_DECODER
1951                 else{
1952                     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]);
1953                     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]);
1954                     for(j=1; j<3; j++){
1955                         for(i=j*16; i<j*16+4; i++){
1956                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1957                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
1958                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1959                             }
1960                         }
1961                     }
1962                 }
1963 #endif
1964             }
1965         }
1966     }
1967     if(h->cbp || IS_INTRA(mb_type))
1968     {
1969         s->dsp.clear_blocks(h->mb);
1970         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
1971     }
1972 }
1973
1974 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
1975     MpegEncContext * const s = &h->s;
1976     const int mb_x= s->mb_x;
1977     const int mb_y= s->mb_y;
1978     const int mb_xy= h->mb_xy;
1979     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1980     uint8_t  *dest[3];
1981     int linesize;
1982     int i, j, p;
1983     int *block_offset = &h->block_offset[0];
1984     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1985     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
1986
1987     for (p = 0; p < plane_count; p++)
1988     {
1989         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
1990         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1991     }
1992
1993     h->list_counts[mb_xy]= h->list_count;
1994
1995     if (!simple && MB_FIELD) {
1996         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
1997         block_offset = &h->block_offset[48];
1998         if(mb_y&1) //FIXME move out of this function?
1999             for (p = 0; p < 3; p++)
2000                 dest[p] -= s->linesize*15;
2001         if(FRAME_MBAFF) {
2002             int list;
2003             for(list=0; list<h->list_count; list++){
2004                 if(!USES_LIST(mb_type, list))
2005                     continue;
2006                 if(IS_16X16(mb_type)){
2007                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2008                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2009                 }else{
2010                     for(i=0; i<16; i+=4){
2011                         int ref = h->ref_cache[list][scan8[i]];
2012                         if(ref >= 0)
2013                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2014                     }
2015                 }
2016             }
2017         }
2018     } else {
2019         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2020     }
2021
2022     if (!simple && IS_INTRA_PCM(mb_type)) {
2023         if (pixel_shift) {
2024             const int bit_depth = h->sps.bit_depth_luma;
2025             GetBitContext gb;
2026             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
2027
2028             for (p = 0; p < plane_count; p++) {
2029                 for (i = 0; i < 16; i++) {
2030                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2031                     for (j = 0; j < 16; j++)
2032                         tmp[j] = get_bits(&gb, bit_depth);
2033                 }
2034             }
2035         } else {
2036             for (p = 0; p < plane_count; p++) {
2037                 for (i = 0; i < 16; i++) {
2038                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2039                 }
2040             }
2041         }
2042     } else {
2043         if(IS_INTRA(mb_type)){
2044             if(h->deblocking_filter)
2045                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2046
2047             for (p = 0; p < plane_count; p++)
2048                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2049
2050             if(h->deblocking_filter)
2051                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2052         }else{
2053             hl_motion(h, dest[0], dest[1], dest[2],
2054                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2055                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2056                       h->h264dsp.weight_h264_pixels_tab,
2057                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 1);
2058         }
2059
2060         for (p = 0; p < plane_count; p++)
2061             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2062     }
2063     if(h->cbp || IS_INTRA(mb_type))
2064     {
2065         s->dsp.clear_blocks(h->mb);
2066         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2067     }
2068 }
2069
2070 /**
2071  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2072  */
2073 #define hl_decode_mb_simple(sh, bits) \
2074 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2075     hl_decode_mb_internal(h, 1, sh); \
2076 }
2077 hl_decode_mb_simple(0, 8);
2078 hl_decode_mb_simple(1, 16);
2079
2080 /**
2081  * Process a macroblock; this handles edge cases, such as interlacing.
2082  */
2083 static void av_noinline hl_decode_mb_complex(H264Context *h){
2084     hl_decode_mb_internal(h, 0, h->pixel_shift);
2085 }
2086
2087 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2088     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2089 }
2090
2091 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2092     hl_decode_mb_444_internal(h, 1, 0);
2093 }
2094
2095 void ff_h264_hl_decode_mb(H264Context *h){
2096     MpegEncContext * const s = &h->s;
2097     const int mb_xy= h->mb_xy;
2098     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2099     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2100
2101     if (CHROMA444) {
2102         if(is_complex || h->pixel_shift)
2103             hl_decode_mb_444_complex(h);
2104         else
2105             hl_decode_mb_444_simple(h);
2106     } else if (is_complex) {
2107         hl_decode_mb_complex(h);
2108     } else if (h->pixel_shift) {
2109         hl_decode_mb_simple_16(h);
2110     } else
2111         hl_decode_mb_simple_8(h);
2112 }
2113
2114 static int pred_weight_table(H264Context *h){
2115     MpegEncContext * const s = &h->s;
2116     int list, i;
2117     int luma_def, chroma_def;
2118
2119     h->use_weight= 0;
2120     h->use_weight_chroma= 0;
2121     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2122     if(h->sps.chroma_format_idc)
2123         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2124     luma_def = 1<<h->luma_log2_weight_denom;
2125     chroma_def = 1<<h->chroma_log2_weight_denom;
2126
2127     for(list=0; list<2; list++){
2128         h->luma_weight_flag[list]   = 0;
2129         h->chroma_weight_flag[list] = 0;
2130         for(i=0; i<h->ref_count[list]; i++){
2131             int luma_weight_flag, chroma_weight_flag;
2132
2133             luma_weight_flag= get_bits1(&s->gb);
2134             if(luma_weight_flag){
2135                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2136                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2137                 if(   h->luma_weight[i][list][0] != luma_def
2138                    || h->luma_weight[i][list][1] != 0) {
2139                     h->use_weight= 1;
2140                     h->luma_weight_flag[list]= 1;
2141                 }
2142             }else{
2143                 h->luma_weight[i][list][0]= luma_def;
2144                 h->luma_weight[i][list][1]= 0;
2145             }
2146
2147             if(h->sps.chroma_format_idc){
2148                 chroma_weight_flag= get_bits1(&s->gb);
2149                 if(chroma_weight_flag){
2150                     int j;
2151                     for(j=0; j<2; j++){
2152                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2153                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2154                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2155                            || h->chroma_weight[i][list][j][1] != 0) {
2156                             h->use_weight_chroma= 1;
2157                             h->chroma_weight_flag[list]= 1;
2158                         }
2159                     }
2160                 }else{
2161                     int j;
2162                     for(j=0; j<2; j++){
2163                         h->chroma_weight[i][list][j][0]= chroma_def;
2164                         h->chroma_weight[i][list][j][1]= 0;
2165                     }
2166                 }
2167             }
2168         }
2169         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2170     }
2171     h->use_weight= h->use_weight || h->use_weight_chroma;
2172     return 0;
2173 }
2174
2175 /**
2176  * Initialize implicit_weight table.
2177  * @param field  0/1 initialize the weight for interlaced MBAFF
2178  *                -1 initializes the rest
2179  */
2180 static void implicit_weight_table(H264Context *h, int field){
2181     MpegEncContext * const s = &h->s;
2182     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2183
2184     for (i = 0; i < 2; i++) {
2185         h->luma_weight_flag[i]   = 0;
2186         h->chroma_weight_flag[i] = 0;
2187     }
2188
2189     if(field < 0){
2190         cur_poc = s->current_picture_ptr->poc;
2191     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2192        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2193         h->use_weight= 0;
2194         h->use_weight_chroma= 0;
2195         return;
2196     }
2197         ref_start= 0;
2198         ref_count0= h->ref_count[0];
2199         ref_count1= h->ref_count[1];
2200     }else{
2201         cur_poc = s->current_picture_ptr->field_poc[field];
2202         ref_start= 16;
2203         ref_count0= 16+2*h->ref_count[0];
2204         ref_count1= 16+2*h->ref_count[1];
2205     }
2206
2207     h->use_weight= 2;
2208     h->use_weight_chroma= 2;
2209     h->luma_log2_weight_denom= 5;
2210     h->chroma_log2_weight_denom= 5;
2211
2212     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2213         int poc0 = h->ref_list[0][ref0].poc;
2214         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2215             int w = 32;
2216             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2217                 int poc1 = h->ref_list[1][ref1].poc;
2218                 int td = av_clip(poc1 - poc0, -128, 127);
2219                 if(td){
2220                     int tb = av_clip(cur_poc - poc0, -128, 127);
2221                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2222                     int dist_scale_factor = (tb*tx + 32) >> 8;
2223                     if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2224                         w = 64 - dist_scale_factor;
2225                 }
2226             }
2227             if(field<0){
2228                 h->implicit_weight[ref0][ref1][0]=
2229                 h->implicit_weight[ref0][ref1][1]= w;
2230             }else{
2231                 h->implicit_weight[ref0][ref1][field]=w;
2232             }
2233         }
2234     }
2235 }
2236
2237 /**
2238  * instantaneous decoder refresh.
2239  */
2240 static void idr(H264Context *h){
2241     ff_h264_remove_all_refs(h);
2242     h->prev_frame_num= 0;
2243     h->prev_frame_num_offset= 0;
2244     h->prev_poc_msb=
2245     h->prev_poc_lsb= 0;
2246 }
2247
2248 /* forget old pics after a seek */
2249 static void flush_dpb(AVCodecContext *avctx){
2250     H264Context *h= avctx->priv_data;
2251     int i;
2252     for(i=0; i<=MAX_DELAYED_PIC_COUNT; i++) {
2253         if(h->delayed_pic[i])
2254             h->delayed_pic[i]->f.reference = 0;
2255         h->delayed_pic[i]= NULL;
2256     }
2257     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2258     h->prev_interlaced_frame = 1;
2259     idr(h);
2260     if(h->s.current_picture_ptr)
2261         h->s.current_picture_ptr->f.reference = 0;
2262     h->s.first_field= 0;
2263     ff_h264_reset_sei(h);
2264     ff_mpeg_flush(avctx);
2265     h->sync= 0;
2266 }
2267
2268 static int init_poc(H264Context *h){
2269     MpegEncContext * const s = &h->s;
2270     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2271     int field_poc[2];
2272     Picture *cur = s->current_picture_ptr;
2273
2274     h->frame_num_offset= h->prev_frame_num_offset;
2275     if(h->frame_num < h->prev_frame_num)
2276         h->frame_num_offset += max_frame_num;
2277
2278     if(h->sps.poc_type==0){
2279         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2280
2281         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2282             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2283         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2284             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2285         else
2286             h->poc_msb = h->prev_poc_msb;
2287 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2288         field_poc[0] =
2289         field_poc[1] = h->poc_msb + h->poc_lsb;
2290         if(s->picture_structure == PICT_FRAME)
2291             field_poc[1] += h->delta_poc_bottom;
2292     }else if(h->sps.poc_type==1){
2293         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2294         int i;
2295
2296         if(h->sps.poc_cycle_length != 0)
2297             abs_frame_num = h->frame_num_offset + h->frame_num;
2298         else
2299             abs_frame_num = 0;
2300
2301         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2302             abs_frame_num--;
2303
2304         expected_delta_per_poc_cycle = 0;
2305         for(i=0; i < h->sps.poc_cycle_length; i++)
2306             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2307
2308         if(abs_frame_num > 0){
2309             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2310             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2311
2312             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2313             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2314                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2315         } else
2316             expectedpoc = 0;
2317
2318         if(h->nal_ref_idc == 0)
2319             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2320
2321         field_poc[0] = expectedpoc + h->delta_poc[0];
2322         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2323
2324         if(s->picture_structure == PICT_FRAME)
2325             field_poc[1] += h->delta_poc[1];
2326     }else{
2327         int poc= 2*(h->frame_num_offset + h->frame_num);
2328
2329         if(!h->nal_ref_idc)
2330             poc--;
2331
2332         field_poc[0]= poc;
2333         field_poc[1]= poc;
2334     }
2335
2336     if(s->picture_structure != PICT_BOTTOM_FIELD)
2337         s->current_picture_ptr->field_poc[0]= field_poc[0];
2338     if(s->picture_structure != PICT_TOP_FIELD)
2339         s->current_picture_ptr->field_poc[1]= field_poc[1];
2340     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2341
2342     return 0;
2343 }
2344
2345
2346 /**
2347  * initialize scan tables
2348  */
2349 static void init_scan_tables(H264Context *h){
2350     int i;
2351     for(i=0; i<16; i++){
2352 #define T(x) (x>>2) | ((x<<2) & 0xF)
2353         h->zigzag_scan[i] = T(zigzag_scan[i]);
2354         h-> field_scan[i] = T( field_scan[i]);
2355 #undef T
2356     }
2357     for(i=0; i<64; i++){
2358 #define T(x) (x>>3) | ((x&7)<<3)
2359         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2360         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2361         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2362         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2363 #undef T
2364     }
2365     if(h->sps.transform_bypass){ //FIXME same ugly
2366         h->zigzag_scan_q0          = zigzag_scan;
2367         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2368         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2369         h->field_scan_q0           = field_scan;
2370         h->field_scan8x8_q0        = field_scan8x8;
2371         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2372     }else{
2373         h->zigzag_scan_q0          = h->zigzag_scan;
2374         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2375         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2376         h->field_scan_q0           = h->field_scan;
2377         h->field_scan8x8_q0        = h->field_scan8x8;
2378         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2379     }
2380 }
2381
2382 static int field_end(H264Context *h, int in_setup){
2383     MpegEncContext * const s = &h->s;
2384     AVCodecContext * const avctx= s->avctx;
2385     int err = 0;
2386     s->mb_y= 0;
2387
2388     if (!in_setup && !s->dropable)
2389         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2390                                  s->picture_structure==PICT_BOTTOM_FIELD);
2391
2392     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2393         ff_vdpau_h264_set_reference_frames(s);
2394
2395     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2396         if(!s->dropable) {
2397             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2398             h->prev_poc_msb= h->poc_msb;
2399             h->prev_poc_lsb= h->poc_lsb;
2400         }
2401         h->prev_frame_num_offset= h->frame_num_offset;
2402         h->prev_frame_num= h->frame_num;
2403         h->outputed_poc = h->next_outputed_poc;
2404     }
2405
2406     if (avctx->hwaccel) {
2407         if (avctx->hwaccel->end_frame(avctx) < 0)
2408             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2409     }
2410
2411     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2412         ff_vdpau_h264_picture_complete(s);
2413
2414     /*
2415      * FIXME: Error handling code does not seem to support interlaced
2416      * when slices span multiple rows
2417      * The ff_er_add_slice calls don't work right for bottom
2418      * fields; they cause massive erroneous error concealing
2419      * Error marking covers both fields (top and bottom).
2420      * This causes a mismatched s->error_count
2421      * and a bad error table. Further, the error count goes to
2422      * INT_MAX when called for bottom field, because mb_y is
2423      * past end by one (callers fault) and resync_mb_y != 0
2424      * causes problems for the first MB line, too.
2425      */
2426     if (!FIELD_PICTURE)
2427         ff_er_frame_end(s);
2428
2429     MPV_frame_end(s);
2430
2431     h->current_slice=0;
2432
2433     return err;
2434 }
2435
2436 /**
2437  * Replicate H264 "master" context to thread contexts.
2438  */
2439 static void clone_slice(H264Context *dst, H264Context *src)
2440 {
2441     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2442     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2443     dst->s.current_picture      = src->s.current_picture;
2444     dst->s.linesize             = src->s.linesize;
2445     dst->s.uvlinesize           = src->s.uvlinesize;
2446     dst->s.first_field          = src->s.first_field;
2447
2448     dst->prev_poc_msb           = src->prev_poc_msb;
2449     dst->prev_poc_lsb           = src->prev_poc_lsb;
2450     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2451     dst->prev_frame_num         = src->prev_frame_num;
2452     dst->short_ref_count        = src->short_ref_count;
2453
2454     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2455     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2456     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2457     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2458
2459     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2460     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2461 }
2462
2463 /**
2464  * computes profile from profile_idc and constraint_set?_flags
2465  *
2466  * @param sps SPS
2467  *
2468  * @return profile as defined by FF_PROFILE_H264_*
2469  */
2470 int ff_h264_get_profile(SPS *sps)
2471 {
2472     int profile = sps->profile_idc;
2473
2474     switch(sps->profile_idc) {
2475     case FF_PROFILE_H264_BASELINE:
2476         // constraint_set1_flag set to 1
2477         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2478         break;
2479     case FF_PROFILE_H264_HIGH_10:
2480     case FF_PROFILE_H264_HIGH_422:
2481     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2482         // constraint_set3_flag set to 1
2483         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2484         break;
2485     }
2486
2487     return profile;
2488 }
2489
2490 /**
2491  * decodes a slice header.
2492  * This will also call MPV_common_init() and frame_start() as needed.
2493  *
2494  * @param h h264context
2495  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2496  *
2497  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2498  */
2499 static int decode_slice_header(H264Context *h, H264Context *h0){
2500     MpegEncContext * const s = &h->s;
2501     MpegEncContext * const s0 = &h0->s;
2502     unsigned int first_mb_in_slice;
2503     unsigned int pps_id;
2504     int num_ref_idx_active_override_flag;
2505     unsigned int slice_type, tmp, i, j;
2506     int default_ref_list_done = 0;
2507     int last_pic_structure;
2508
2509     s->dropable= h->nal_ref_idc == 0;
2510
2511     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2512     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2513         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2514         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2515     }else{
2516         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2517         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2518     }
2519
2520     first_mb_in_slice= get_ue_golomb(&s->gb);
2521
2522     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2523         if(h0->current_slice && FIELD_PICTURE){
2524             field_end(h, 1);
2525         }
2526
2527         h0->current_slice = 0;
2528         if (!s0->first_field)
2529             s->current_picture_ptr= NULL;
2530     }
2531
2532     slice_type= get_ue_golomb_31(&s->gb);
2533     if(slice_type > 9){
2534         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);
2535         return -1;
2536     }
2537     if(slice_type > 4){
2538         slice_type -= 5;
2539         h->slice_type_fixed=1;
2540     }else
2541         h->slice_type_fixed=0;
2542
2543     slice_type= golomb_to_pict_type[ slice_type ];
2544     if (slice_type == AV_PICTURE_TYPE_I
2545         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2546         default_ref_list_done = 1;
2547     }
2548     h->slice_type= slice_type;
2549     h->slice_type_nos= slice_type & 3;
2550
2551     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2552
2553     pps_id= get_ue_golomb(&s->gb);
2554     if(pps_id>=MAX_PPS_COUNT){
2555         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2556         return -1;
2557     }
2558     if(!h0->pps_buffers[pps_id]) {
2559         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2560         return -1;
2561     }
2562     h->pps= *h0->pps_buffers[pps_id];
2563
2564     if(!h0->sps_buffers[h->pps.sps_id]) {
2565         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2566         return -1;
2567     }
2568     h->sps = *h0->sps_buffers[h->pps.sps_id];
2569
2570     s->avctx->profile = ff_h264_get_profile(&h->sps);
2571     s->avctx->level   = h->sps.level_idc;
2572     s->avctx->refs    = h->sps.ref_frame_count;
2573
2574     if(h == h0 && h->dequant_coeff_pps != pps_id){
2575         h->dequant_coeff_pps = pps_id;
2576         init_dequant_tables(h);
2577     }
2578
2579     s->mb_width= h->sps.mb_width;
2580     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2581
2582     h->b_stride=  s->mb_width*4;
2583
2584     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2585
2586     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2587     if(h->sps.frame_mbs_only_flag)
2588         s->height= 16*s->mb_height - (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2589     else
2590         s->height= 16*s->mb_height - (2<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2591
2592     if (s->context_initialized
2593         && (   s->width != s->avctx->width || s->height != s->avctx->height
2594             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2595         if(h != h0) {
2596             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2597             return -1;   // width / height changed during parallelized decoding
2598         }
2599         free_tables(h, 0);
2600         flush_dpb(s->avctx);
2601         MPV_common_end(s);
2602     }
2603     if (!s->context_initialized) {
2604         if (h != h0) {
2605             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2606             return -1;
2607         }
2608
2609         avcodec_set_dimensions(s->avctx, s->width, s->height);
2610         s->avctx->sample_aspect_ratio= h->sps.sar;
2611         av_assert0(s->avctx->sample_aspect_ratio.den);
2612
2613         if(h->sps.video_signal_type_present_flag){
2614             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2615             if(h->sps.colour_description_present_flag){
2616                 s->avctx->color_primaries = h->sps.color_primaries;
2617                 s->avctx->color_trc       = h->sps.color_trc;
2618                 s->avctx->colorspace      = h->sps.colorspace;
2619             }
2620         }
2621
2622         if(h->sps.timing_info_present_flag){
2623             int64_t den= h->sps.time_scale;
2624             if(h->x264_build < 44U)
2625                 den *= 2;
2626             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2627                       h->sps.num_units_in_tick, den, 1<<30);
2628         }
2629
2630         switch (h->sps.bit_depth_luma) {
2631             case 9 :
2632                 if (CHROMA444)
2633                     s->avctx->pix_fmt = PIX_FMT_YUV444P9;
2634                 else
2635                     s->avctx->pix_fmt = PIX_FMT_YUV420P9;
2636                 break;
2637             case 10 :
2638                 if (CHROMA444)
2639                     s->avctx->pix_fmt = PIX_FMT_YUV444P10;
2640                 else if (CHROMA422)
2641                     s->avctx->pix_fmt = PIX_FMT_YUV422P10;
2642                 else
2643                     s->avctx->pix_fmt = PIX_FMT_YUV420P10;
2644                 break;
2645             default:
2646                 if (CHROMA444){
2647                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2648                 }else if (CHROMA422) {
2649                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P : PIX_FMT_YUV422P;
2650                 }else{
2651                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2652                                                              s->avctx->codec->pix_fmts ?
2653                                                              s->avctx->codec->pix_fmts :
2654                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2655                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2656                                                              ff_hwaccel_pixfmt_list_420);
2657                 }
2658         }
2659
2660         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2661
2662         if (MPV_common_init(s) < 0) {
2663             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2664             return -1;
2665         }
2666         s->first_field = 0;
2667         h->prev_interlaced_frame = 1;
2668
2669         init_scan_tables(h);
2670         if (ff_h264_alloc_tables(h) < 0) {
2671             av_log(h->s.avctx, AV_LOG_ERROR, "Could not allocate memory for h264\n");
2672             return AVERROR(ENOMEM);
2673         }
2674
2675         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2676             if (context_init(h) < 0) {
2677                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2678                 return -1;
2679             }
2680         } else {
2681             for(i = 1; i < s->avctx->thread_count; i++) {
2682                 H264Context *c;
2683                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2684                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2685                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2686                 c->h264dsp = h->h264dsp;
2687                 c->sps = h->sps;
2688                 c->pps = h->pps;
2689                 c->pixel_shift = h->pixel_shift;
2690                 init_scan_tables(c);
2691                 clone_tables(c, h, i);
2692             }
2693
2694             for(i = 0; i < s->avctx->thread_count; i++)
2695                 if (context_init(h->thread_context[i]) < 0) {
2696                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2697                     return -1;
2698                 }
2699         }
2700     }
2701
2702     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2703
2704     h->mb_mbaff = 0;
2705     h->mb_aff_frame = 0;
2706     last_pic_structure = s0->picture_structure;
2707     if(h->sps.frame_mbs_only_flag){
2708         s->picture_structure= PICT_FRAME;
2709     }else{
2710         if(get_bits1(&s->gb)) { //field_pic_flag
2711             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2712         } else {
2713             s->picture_structure= PICT_FRAME;
2714             h->mb_aff_frame = h->sps.mb_aff;
2715         }
2716     }
2717     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2718
2719     if(h0->current_slice == 0){
2720         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2721         if(h->frame_num != h->prev_frame_num) {
2722             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2723
2724             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2725
2726             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2727                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2728                 if (unwrap_prev_frame_num < 0)
2729                     unwrap_prev_frame_num += max_frame_num;
2730
2731                 h->prev_frame_num = unwrap_prev_frame_num;
2732             }
2733         }
2734
2735         while(h->frame_num !=  h->prev_frame_num &&
2736               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2737             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2738             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2739             if (ff_h264_frame_start(h) < 0)
2740                 return -1;
2741             h->prev_frame_num++;
2742             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2743             s->current_picture_ptr->frame_num= h->prev_frame_num;
2744             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2745             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2746             ff_generate_sliding_window_mmcos(h);
2747             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2748                 s->avctx->error_recognition >= FF_ER_EXPLODE)
2749                 return AVERROR_INVALIDDATA;
2750             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2751              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2752              * about there being no actual duplicates.
2753              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2754              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2755              * be fixed. */
2756             if (h->short_ref_count) {
2757                 if (prev) {
2758                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2759                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2760                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2761                     h->short_ref[0]->poc = prev->poc+2;
2762                 }
2763                 h->short_ref[0]->frame_num = h->prev_frame_num;
2764             }
2765         }
2766
2767         /* See if we have a decoded first field looking for a pair... */
2768         if (s0->first_field) {
2769             assert(s0->current_picture_ptr);
2770             assert(s0->current_picture_ptr->f.data[0]);
2771             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
2772
2773             /* figure out if we have a complementary field pair */
2774             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2775                 /*
2776                  * Previous field is unmatched. Don't display it, but let it
2777                  * remain for reference if marked as such.
2778                  */
2779                 s0->current_picture_ptr = NULL;
2780                 s0->first_field = FIELD_PICTURE;
2781
2782             } else {
2783                 if (h->nal_ref_idc &&
2784                         s0->current_picture_ptr->f.reference &&
2785                         s0->current_picture_ptr->frame_num != h->frame_num) {
2786                     /*
2787                      * This and previous field were reference, but had
2788                      * different frame_nums. Consider this field first in
2789                      * pair. Throw away previous field except for reference
2790                      * purposes.
2791                      */
2792                     s0->first_field = 1;
2793                     s0->current_picture_ptr = NULL;
2794
2795                 } else {
2796                     /* Second field in complementary pair */
2797                     s0->first_field = 0;
2798                 }
2799             }
2800
2801         } else {
2802             /* Frame or first field in a potentially complementary pair */
2803             assert(!s0->current_picture_ptr);
2804             s0->first_field = FIELD_PICTURE;
2805         }
2806
2807         if(!FIELD_PICTURE || s0->first_field) {
2808             if (ff_h264_frame_start(h) < 0) {
2809                 s0->first_field = 0;
2810                 return -1;
2811             }
2812         } else {
2813             ff_release_unused_pictures(s, 0);
2814         }
2815     }
2816     if(h != h0)
2817         clone_slice(h, h0);
2818
2819     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2820
2821     assert(s->mb_num == s->mb_width * s->mb_height);
2822     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2823        first_mb_in_slice                    >= s->mb_num){
2824         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2825         return -1;
2826     }
2827     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2828     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2829     if (s->picture_structure == PICT_BOTTOM_FIELD)
2830         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2831     assert(s->mb_y < s->mb_height);
2832
2833     if(s->picture_structure==PICT_FRAME){
2834         h->curr_pic_num=   h->frame_num;
2835         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2836     }else{
2837         h->curr_pic_num= 2*h->frame_num + 1;
2838         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2839     }
2840
2841     if(h->nal_unit_type == NAL_IDR_SLICE){
2842         get_ue_golomb(&s->gb); /* idr_pic_id */
2843     }
2844
2845     if(h->sps.poc_type==0){
2846         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2847
2848         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2849             h->delta_poc_bottom= get_se_golomb(&s->gb);
2850         }
2851     }
2852
2853     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2854         h->delta_poc[0]= get_se_golomb(&s->gb);
2855
2856         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2857             h->delta_poc[1]= get_se_golomb(&s->gb);
2858     }
2859
2860     init_poc(h);
2861
2862     if(h->pps.redundant_pic_cnt_present){
2863         h->redundant_pic_count= get_ue_golomb(&s->gb);
2864     }
2865
2866     //set defaults, might be overridden a few lines later
2867     h->ref_count[0]= h->pps.ref_count[0];
2868     h->ref_count[1]= h->pps.ref_count[1];
2869
2870     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2871         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2872             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2873         }
2874         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2875
2876         if(num_ref_idx_active_override_flag){
2877             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2878             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2879                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2880
2881             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2882                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2883                 h->ref_count[0]= h->ref_count[1]= 1;
2884                 return -1;
2885             }
2886         }
2887         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2888             h->list_count= 2;
2889         else
2890             h->list_count= 1;
2891     }else
2892         h->list_count= 0;
2893
2894     if(!default_ref_list_done){
2895         ff_h264_fill_default_ref_list(h);
2896     }
2897
2898     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0)
2899         return -1;
2900
2901     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2902         s->last_picture_ptr= &h->ref_list[0][0];
2903         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2904     }
2905     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2906         s->next_picture_ptr= &h->ref_list[1][0];
2907         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2908     }
2909
2910     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2911        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2912         pred_weight_table(h);
2913     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2914         implicit_weight_table(h, -1);
2915     }else {
2916         h->use_weight = 0;
2917         for (i = 0; i < 2; i++) {
2918             h->luma_weight_flag[i]   = 0;
2919             h->chroma_weight_flag[i] = 0;
2920         }
2921     }
2922
2923     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
2924        s->avctx->error_recognition >= FF_ER_EXPLODE)
2925         return AVERROR_INVALIDDATA;
2926
2927     if(FRAME_MBAFF){
2928         ff_h264_fill_mbaff_ref_list(h);
2929
2930         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2931             implicit_weight_table(h, 0);
2932             implicit_weight_table(h, 1);
2933         }
2934     }
2935
2936     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2937         ff_h264_direct_dist_scale_factor(h);
2938     ff_h264_direct_ref_list_init(h);
2939
2940     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2941         tmp = get_ue_golomb_31(&s->gb);
2942         if(tmp > 2){
2943             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2944             return -1;
2945         }
2946         h->cabac_init_idc= tmp;
2947     }
2948
2949     h->last_qscale_diff = 0;
2950     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2951     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
2952         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2953         return -1;
2954     }
2955     s->qscale= tmp;
2956     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2957     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2958     //FIXME qscale / qp ... stuff
2959     if(h->slice_type == AV_PICTURE_TYPE_SP){
2960         get_bits1(&s->gb); /* sp_for_switch_flag */
2961     }
2962     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
2963         get_se_golomb(&s->gb); /* slice_qs_delta */
2964     }
2965
2966     h->deblocking_filter = 1;
2967     h->slice_alpha_c0_offset = 52;
2968     h->slice_beta_offset = 52;
2969     if( h->pps.deblocking_filter_parameters_present ) {
2970         tmp= get_ue_golomb_31(&s->gb);
2971         if(tmp > 2){
2972             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2973             return -1;
2974         }
2975         h->deblocking_filter= tmp;
2976         if(h->deblocking_filter < 2)
2977             h->deblocking_filter^= 1; // 1<->0
2978
2979         if( h->deblocking_filter ) {
2980             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2981             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2982             if(   h->slice_alpha_c0_offset > 104U
2983                || h->slice_beta_offset     > 104U){
2984                 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);
2985                 return -1;
2986             }
2987         }
2988     }
2989
2990     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2991        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
2992        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
2993        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2994         h->deblocking_filter= 0;
2995
2996     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
2997         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
2998             /* Cheat slightly for speed:
2999                Do not bother to deblock across slices. */
3000             h->deblocking_filter = 2;
3001         } else {
3002             h0->max_contexts = 1;
3003             if(!h0->single_decode_warning) {
3004                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3005                 h0->single_decode_warning = 1;
3006             }
3007             if (h != h0) {
3008                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
3009                 return 1;
3010             }
3011         }
3012     }
3013     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
3014                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
3015                  + 6 * (h->sps.bit_depth_luma - 8);
3016
3017 #if 0 //FMO
3018     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
3019         slice_group_change_cycle= get_bits(&s->gb, ?);
3020 #endif
3021
3022     h0->last_slice_type = slice_type;
3023     h->slice_num = ++h0->current_slice;
3024
3025     if(h->slice_num)
3026         h0->slice_row[(h->slice_num-1)&(MAX_SLICES-1)]= s->resync_mb_y;
3027     if (   h0->slice_row[h->slice_num&(MAX_SLICES-1)] + 3 >= s->resync_mb_y
3028         && h0->slice_row[h->slice_num&(MAX_SLICES-1)] <= s->resync_mb_y
3029         && h->slice_num >= MAX_SLICES) {
3030         //in case of ASO this check needs to be updated depending on how we decide to assign slice numbers in this case
3031         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);
3032     }
3033
3034     for(j=0; j<2; j++){
3035         int id_list[16];
3036         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
3037         for(i=0; i<16; i++){
3038             id_list[i]= 60;
3039             if (h->ref_list[j][i].f.data[0]) {
3040                 int k;
3041                 uint8_t *base = h->ref_list[j][i].f.base[0];
3042                 for(k=0; k<h->short_ref_count; k++)
3043                     if (h->short_ref[k]->f.base[0] == base) {
3044                         id_list[i]= k;
3045                         break;
3046                     }
3047                 for(k=0; k<h->long_ref_count; k++)
3048                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3049                         id_list[i]= h->short_ref_count + k;
3050                         break;
3051                     }
3052             }
3053         }
3054
3055         ref2frm[0]=
3056         ref2frm[1]= -1;
3057         for(i=0; i<16; i++)
3058             ref2frm[i+2]= 4*id_list[i]
3059                           + (h->ref_list[j][i].f.reference & 3);
3060         ref2frm[18+0]=
3061         ref2frm[18+1]= -1;
3062         for(i=16; i<48; i++)
3063             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3064                           + (h->ref_list[j][i].f.reference & 3);
3065     }
3066
3067     //FIXME: fix draw_edges+PAFF+frame threads
3068     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3069     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3070
3071     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3072         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",
3073                h->slice_num,
3074                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3075                first_mb_in_slice,
3076                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3077                pps_id, h->frame_num,
3078                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3079                h->ref_count[0], h->ref_count[1],
3080                s->qscale,
3081                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3082                h->use_weight,
3083                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3084                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3085                );
3086     }
3087
3088     return 0;
3089 }
3090
3091 int ff_h264_get_slice_type(const H264Context *h)
3092 {
3093     switch (h->slice_type) {
3094     case AV_PICTURE_TYPE_P:  return 0;
3095     case AV_PICTURE_TYPE_B:  return 1;
3096     case AV_PICTURE_TYPE_I:  return 2;
3097     case AV_PICTURE_TYPE_SP: return 3;
3098     case AV_PICTURE_TYPE_SI: return 4;
3099     default:         return -1;
3100     }
3101 }
3102
3103 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3104                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3105 {
3106     int b_stride = h->b_stride;
3107     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3108     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3109     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3110         if(USES_LIST(top_type, list)){
3111             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3112             const int b8_xy= 4*top_xy + 2;
3113             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3114             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3115             ref_cache[0 - 1*8]=
3116             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3117             ref_cache[2 - 1*8]=
3118             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3119         }else{
3120             AV_ZERO128(mv_dst - 1*8);
3121             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3122         }
3123
3124         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3125             if(USES_LIST(left_type[LTOP], list)){
3126                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3127                 const int b8_xy= 4*left_xy[LTOP] + 1;
3128                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3129                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3130                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3131                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3132                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3133                 ref_cache[-1 +  0]=
3134                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3135                 ref_cache[-1 + 16]=
3136                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3137             }else{
3138                 AV_ZERO32(mv_dst - 1 + 0);
3139                 AV_ZERO32(mv_dst - 1 + 8);
3140                 AV_ZERO32(mv_dst - 1 +16);
3141                 AV_ZERO32(mv_dst - 1 +24);
3142                 ref_cache[-1 +  0]=
3143                 ref_cache[-1 +  8]=
3144                 ref_cache[-1 + 16]=
3145                 ref_cache[-1 + 24]= LIST_NOT_USED;
3146             }
3147         }
3148     }
3149
3150     if(!USES_LIST(mb_type, list)){
3151         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3152         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3153         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3154         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3155         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3156         return;
3157     }
3158
3159     {
3160         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3161         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3162         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3163         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3164         AV_WN32A(&ref_cache[0*8], ref01);
3165         AV_WN32A(&ref_cache[1*8], ref01);
3166         AV_WN32A(&ref_cache[2*8], ref23);
3167         AV_WN32A(&ref_cache[3*8], ref23);
3168     }
3169
3170     {
3171         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3172         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3173         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3174         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3175         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3176     }
3177 }
3178
3179 /**
3180  *
3181  * @return non zero if the loop filter can be skiped
3182  */
3183 static int fill_filter_caches(H264Context *h, int mb_type){
3184     MpegEncContext * const s = &h->s;
3185     const int mb_xy= h->mb_xy;
3186     int top_xy, left_xy[LEFT_MBS];
3187     int top_type, left_type[LEFT_MBS];
3188     uint8_t *nnz;
3189     uint8_t *nnz_cache;
3190
3191     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3192
3193     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3194      * stuff, I can't imagine that these complex rules are worth it. */
3195
3196     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3197     if(FRAME_MBAFF){
3198         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3199         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3200         if(s->mb_y&1){
3201             if (left_mb_field_flag != curr_mb_field_flag) {
3202                 left_xy[LTOP] -= s->mb_stride;
3203             }
3204         }else{
3205             if(curr_mb_field_flag){
3206                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3207             }
3208             if (left_mb_field_flag != curr_mb_field_flag) {
3209                 left_xy[LBOT] += s->mb_stride;
3210             }
3211         }
3212     }
3213
3214     h->top_mb_xy = top_xy;
3215     h->left_mb_xy[LTOP] = left_xy[LTOP];
3216     h->left_mb_xy[LBOT] = left_xy[LBOT];
3217     {
3218         //for sufficiently low qp, filtering wouldn't do anything
3219         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3220         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3221         int qp = s->current_picture.f.qscale_table[mb_xy];
3222         if(qp <= qp_thresh
3223            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3224            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3225             if(!FRAME_MBAFF)
3226                 return 1;
3227             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3228                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3229                 return 1;
3230         }
3231     }
3232
3233     top_type        = s->current_picture.f.mb_type[top_xy];
3234     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3235     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3236     if(h->deblocking_filter == 2){
3237         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3238         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3239     }else{
3240         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3241         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3242     }
3243     h->top_type       = top_type;
3244     h->left_type[LTOP]= left_type[LTOP];
3245     h->left_type[LBOT]= left_type[LBOT];
3246
3247     if(IS_INTRA(mb_type))
3248         return 0;
3249
3250     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3251     if(h->list_count == 2)
3252         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3253
3254     nnz = h->non_zero_count[mb_xy];
3255     nnz_cache = h->non_zero_count_cache;
3256     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3257     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3258     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3259     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3260     h->cbp= h->cbp_table[mb_xy];
3261
3262     if(top_type){
3263         nnz = h->non_zero_count[top_xy];
3264         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3265     }
3266
3267     if(left_type[LTOP]){
3268         nnz = h->non_zero_count[left_xy[LTOP]];
3269         nnz_cache[3+8*1]= nnz[3+0*4];
3270         nnz_cache[3+8*2]= nnz[3+1*4];
3271         nnz_cache[3+8*3]= nnz[3+2*4];
3272         nnz_cache[3+8*4]= nnz[3+3*4];
3273     }
3274
3275     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3276     if(!CABAC && h->pps.transform_8x8_mode){
3277         if(IS_8x8DCT(top_type)){
3278             nnz_cache[4+8*0]=
3279             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3280             nnz_cache[6+8*0]=
3281             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3282         }
3283         if(IS_8x8DCT(left_type[LTOP])){
3284             nnz_cache[3+8*1]=
3285             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3286         }
3287         if(IS_8x8DCT(left_type[LBOT])){
3288             nnz_cache[3+8*3]=
3289             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3290         }
3291
3292         if(IS_8x8DCT(mb_type)){
3293             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3294             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3295
3296             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3297             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3298
3299             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3300             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3301
3302             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3303             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3304         }
3305     }
3306
3307     return 0;
3308 }
3309
3310 static void loop_filter(H264Context *h, int start_x, int end_x){
3311     MpegEncContext * const s = &h->s;
3312     uint8_t  *dest_y, *dest_cb, *dest_cr;
3313     int linesize, uvlinesize, mb_x, mb_y;
3314     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3315     const int old_slice_type= h->slice_type;
3316     const int pixel_shift = h->pixel_shift;
3317     const int block_h = 16>>s->chroma_y_shift;
3318
3319     if(h->deblocking_filter) {
3320         for(mb_x= start_x; mb_x<end_x; mb_x++){
3321             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3322                 int mb_xy, mb_type;
3323                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3324                 h->slice_num= h->slice_table[mb_xy];
3325                 mb_type = s->current_picture.f.mb_type[mb_xy];
3326                 h->list_count= h->list_counts[mb_xy];
3327
3328                 if(FRAME_MBAFF)
3329                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3330
3331                 s->mb_x= mb_x;
3332                 s->mb_y= mb_y;
3333                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3334                 dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*(8<<CHROMA444) + mb_y * s->uvlinesize * block_h;
3335                 dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*(8<<CHROMA444) + mb_y * s->uvlinesize * block_h;
3336                     //FIXME simplify above
3337
3338                 if (MB_FIELD) {
3339                     linesize   = h->mb_linesize   = s->linesize * 2;
3340                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3341                     if(mb_y&1){ //FIXME move out of this function?
3342                         dest_y -= s->linesize*15;
3343                         dest_cb-= s->uvlinesize*(block_h-1);
3344                         dest_cr-= s->uvlinesize*(block_h-1);
3345                     }
3346                 } else {
3347                     linesize   = h->mb_linesize   = s->linesize;
3348                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3349                 }
3350                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
3351                 if(fill_filter_caches(h, mb_type))
3352                     continue;
3353                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3354                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3355
3356                 if (FRAME_MBAFF) {
3357                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3358                 } else {
3359                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3360                 }
3361             }
3362         }
3363     }
3364     h->slice_type= old_slice_type;
3365     s->mb_x= end_x;
3366     s->mb_y= end_mb_y - FRAME_MBAFF;
3367     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3368     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3369 }
3370
3371 static void predict_field_decoding_flag(H264Context *h){
3372     MpegEncContext * const s = &h->s;
3373     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3374     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3375                 ? s->current_picture.f.mb_type[mb_xy - 1]
3376                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3377                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3378                 : 0;
3379     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3380 }
3381
3382 /**
3383  * Draw edges and report progress for the last MB row.
3384  */
3385 static void decode_finish_row(H264Context *h){
3386     MpegEncContext * const s = &h->s;
3387     int top = 16*(s->mb_y >> FIELD_PICTURE);
3388     int height = 16 << FRAME_MBAFF;
3389     int deblock_border = (16 + 4) << FRAME_MBAFF;
3390     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3391
3392     if (h->deblocking_filter) {
3393         if((top + height) >= pic_height)
3394             height += deblock_border;
3395
3396         top -= deblock_border;
3397     }
3398
3399     if (top >= pic_height || (top + height) < h->emu_edge_height)
3400         return;
3401
3402     height = FFMIN(height, pic_height - top);
3403     if (top < h->emu_edge_height) {
3404         height = top+height;
3405         top = 0;
3406     }
3407
3408     ff_draw_horiz_band(s, top, height);
3409
3410     if (s->dropable) return;
3411
3412     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3413                              s->picture_structure==PICT_BOTTOM_FIELD);
3414 }
3415
3416 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3417     H264Context *h = *(void**)arg;
3418     MpegEncContext * const s = &h->s;
3419     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3420     int lf_x_start = s->mb_x;
3421
3422     s->mb_skip_run= -1;
3423
3424     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3425                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3426
3427     if( h->pps.cabac ) {
3428         /* realign */
3429         align_get_bits( &s->gb );
3430
3431         /* init cabac */
3432         ff_init_cabac_states( &h->cabac);
3433         ff_init_cabac_decoder( &h->cabac,
3434                                s->gb.buffer + get_bits_count(&s->gb)/8,
3435                                (get_bits_left(&s->gb) + 7)/8);
3436
3437         ff_h264_init_cabac_states(h);
3438
3439         for(;;){
3440 //START_TIMER
3441             int ret = ff_h264_decode_mb_cabac(h);
3442             int eos;
3443 //STOP_TIMER("decode_mb_cabac")
3444
3445             if(ret>=0) ff_h264_hl_decode_mb(h);
3446
3447             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3448                 s->mb_y++;
3449
3450                 ret = ff_h264_decode_mb_cabac(h);
3451
3452                 if(ret>=0) ff_h264_hl_decode_mb(h);
3453                 s->mb_y--;
3454             }
3455             eos = get_cabac_terminate( &h->cabac );
3456
3457             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3458                 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);
3459                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3460                 return 0;
3461             }
3462             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3463                 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);