calculate frame_num for recovery point "keyframe"
[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     h->sync            = h1->sync;
1205
1206     if(!s->current_picture_ptr) return 0;
1207
1208     if(!s->dropable) {
1209         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1210         h->prev_poc_msb     = h->poc_msb;
1211         h->prev_poc_lsb     = h->poc_lsb;
1212     }
1213     h->prev_frame_num_offset= h->frame_num_offset;
1214     h->prev_frame_num       = h->frame_num;
1215     h->outputed_poc         = h->next_outputed_poc;
1216
1217     return err;
1218 }
1219
1220 int ff_h264_frame_start(H264Context *h){
1221     MpegEncContext * const s = &h->s;
1222     int i;
1223     const int pixel_shift = h->pixel_shift;
1224     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1225
1226     if(MPV_frame_start(s, s->avctx) < 0)
1227         return -1;
1228     ff_er_frame_start(s);
1229     /*
1230      * MPV_frame_start uses pict_type to derive key_frame.
1231      * This is incorrect for H.264; IDR markings must be used.
1232      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1233      * See decode_nal_units().
1234      */
1235     s->current_picture_ptr->f.key_frame = 0;
1236     s->current_picture_ptr->mmco_reset= 0;
1237
1238     assert(s->linesize && s->uvlinesize);
1239
1240     for(i=0; i<16; i++){
1241         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1242         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1243     }
1244     for(i=0; i<16; i++){
1245         h->block_offset[16+i]=
1246         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1247         h->block_offset[48+16+i]=
1248         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1249     }
1250
1251     /* can't be in alloc_tables because linesize isn't known there.
1252      * FIXME: redo bipred weight to not require extra buffer? */
1253     for(i = 0; i < thread_count; i++)
1254         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1255             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1256
1257     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1258     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1259
1260 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1261
1262     // We mark the current picture as non-reference after allocating it, so
1263     // that if we break out due to an error it can be released automatically
1264     // in the next MPV_frame_start().
1265     // SVQ3 as well as most other codecs have only last/next/current and thus
1266     // get released even with set reference, besides SVQ3 and others do not
1267     // mark frames as reference later "naturally".
1268     if(s->codec_id != CODEC_ID_SVQ3)
1269         s->current_picture_ptr->f.reference = 0;
1270
1271     s->current_picture_ptr->field_poc[0]=
1272     s->current_picture_ptr->field_poc[1]= INT_MAX;
1273
1274     h->next_output_pic = NULL;
1275
1276     assert(s->current_picture_ptr->long_ref==0);
1277
1278     return 0;
1279 }
1280
1281 /**
1282   * Run setup operations that must be run after slice header decoding.
1283   * This includes finding the next displayed frame.
1284   *
1285   * @param h h264 master context
1286   * @param setup_finished enough NALs have been read that we can call
1287   * ff_thread_finish_setup()
1288   */
1289 static void decode_postinit(H264Context *h, int setup_finished){
1290     MpegEncContext * const s = &h->s;
1291     Picture *out = s->current_picture_ptr;
1292     Picture *cur = s->current_picture_ptr;
1293     int i, pics, out_of_order, out_idx;
1294
1295     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1296     s->current_picture_ptr->f.pict_type   = s->pict_type;
1297
1298     if (h->next_output_pic) return;
1299
1300     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1301         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1302         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1303         //to find this yet, so we assume the worst for now.
1304         //if (setup_finished)
1305         //    ff_thread_finish_setup(s->avctx);
1306         return;
1307     }
1308
1309     cur->f.interlaced_frame = 0;
1310     cur->f.repeat_pict      = 0;
1311
1312     /* Signal interlacing information externally. */
1313     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1314
1315     if(h->sps.pic_struct_present_flag){
1316         switch (h->sei_pic_struct)
1317         {
1318         case SEI_PIC_STRUCT_FRAME:
1319             break;
1320         case SEI_PIC_STRUCT_TOP_FIELD:
1321         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1322             cur->f.interlaced_frame = 1;
1323             break;
1324         case SEI_PIC_STRUCT_TOP_BOTTOM:
1325         case SEI_PIC_STRUCT_BOTTOM_TOP:
1326             if (FIELD_OR_MBAFF_PICTURE)
1327                 cur->f.interlaced_frame = 1;
1328             else
1329                 // try to flag soft telecine progressive
1330                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1331             break;
1332         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1333         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1334             // Signal the possibility of telecined film externally (pic_struct 5,6)
1335             // From these hints, let the applications decide if they apply deinterlacing.
1336             cur->f.repeat_pict = 1;
1337             break;
1338         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1339             // Force progressive here, as doubling interlaced frame is a bad idea.
1340             cur->f.repeat_pict = 2;
1341             break;
1342         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1343             cur->f.repeat_pict = 4;
1344             break;
1345         }
1346
1347         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1348             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1349     }else{
1350         /* Derive interlacing flag from used decoding process. */
1351         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1352     }
1353     h->prev_interlaced_frame = cur->f.interlaced_frame;
1354
1355     if (cur->field_poc[0] != cur->field_poc[1]){
1356         /* Derive top_field_first from field pocs. */
1357         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1358     }else{
1359         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1360             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1361             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1362               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1363                 cur->f.top_field_first = 1;
1364             else
1365                 cur->f.top_field_first = 0;
1366         }else{
1367             /* Most likely progressive */
1368             cur->f.top_field_first = 0;
1369         }
1370     }
1371
1372     //FIXME do something with unavailable reference frames
1373
1374     /* Sort B-frames into display order */
1375
1376     if(h->sps.bitstream_restriction_flag
1377        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1378         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1379         s->low_delay = 0;
1380     }
1381
1382     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1383        && !h->sps.bitstream_restriction_flag){
1384         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1385         s->low_delay= 0;
1386     }
1387
1388     pics = 0;
1389     while(h->delayed_pic[pics]) pics++;
1390
1391     av_assert0(pics <= MAX_DELAYED_PIC_COUNT);
1392
1393     h->delayed_pic[pics++] = cur;
1394     if (cur->f.reference == 0)
1395         cur->f.reference = DELAYED_PIC_REF;
1396
1397     out = h->delayed_pic[0];
1398     out_idx = 0;
1399     for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1400         if(h->delayed_pic[i]->poc < out->poc){
1401             out = h->delayed_pic[i];
1402             out_idx = i;
1403         }
1404     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1405         h->next_outputed_poc= INT_MIN;
1406     out_of_order = out->poc < h->next_outputed_poc;
1407
1408     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1409         { }
1410     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1411        || (s->low_delay &&
1412         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1413          || cur->f.pict_type == AV_PICTURE_TYPE_B)))
1414     {
1415         s->low_delay = 0;
1416         s->avctx->has_b_frames++;
1417     }
1418
1419     if(out_of_order || pics > s->avctx->has_b_frames){
1420         out->f.reference &= ~DELAYED_PIC_REF;
1421         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1422                          // or else the first thread can release the picture and reuse it unsafely
1423         for(i=out_idx; h->delayed_pic[i]; i++)
1424             h->delayed_pic[i] = h->delayed_pic[i+1];
1425     }
1426     if(!out_of_order && pics > s->avctx->has_b_frames){
1427         h->next_output_pic = out;
1428         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1429             h->next_outputed_poc = INT_MIN;
1430         } else
1431             h->next_outputed_poc = out->poc;
1432     }else{
1433         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1434     }
1435
1436     if (h->next_output_pic && h->next_output_pic->sync) {
1437         h->sync |= 2*!!h->next_output_pic->f.key_frame;
1438     }
1439
1440     if (setup_finished)
1441         ff_thread_finish_setup(s->avctx);
1442 }
1443
1444 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){
1445     MpegEncContext * const s = &h->s;
1446     uint8_t *top_border;
1447     int top_idx = 1;
1448     const int pixel_shift = h->pixel_shift;
1449     int chroma444 = CHROMA444;
1450     int chroma422 = CHROMA422;
1451
1452     src_y  -=   linesize;
1453     src_cb -= uvlinesize;
1454     src_cr -= uvlinesize;
1455
1456     if(!simple && FRAME_MBAFF){
1457         if(s->mb_y&1){
1458             if(!MB_MBAFF){
1459                 top_border = h->top_borders[0][s->mb_x];
1460                 AV_COPY128(top_border, src_y + 15*linesize);
1461                 if (pixel_shift)
1462                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1463                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1464                     if(chroma444){
1465                         if (pixel_shift){
1466                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1467                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1468                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1469                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1470                         } else {
1471                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1472                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1473                         }
1474                     } else if(chroma422){
1475                         if (pixel_shift) {
1476                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1477                             AV_COPY128(top_border+48, src_cr + 15*uvlinesize);
1478                         } else {
1479                             AV_COPY64(top_border+16, src_cb +  15*uvlinesize);
1480                             AV_COPY64(top_border+24, src_cr +  15*uvlinesize);
1481                         }
1482                     } else {
1483                         if (pixel_shift) {
1484                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1485                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1486                         } else {
1487                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1488                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1489                         }
1490                     }
1491                 }
1492             }
1493         }else if(MB_MBAFF){
1494             top_idx = 0;
1495         }else
1496             return;
1497     }
1498
1499     top_border = h->top_borders[top_idx][s->mb_x];
1500     // There are two lines saved, the line above the the top macroblock of a pair,
1501     // and the line above the bottom macroblock
1502     AV_COPY128(top_border, src_y + 16*linesize);
1503     if (pixel_shift)
1504         AV_COPY128(top_border+16, src_y+16*linesize+16);
1505
1506     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1507         if(chroma444){
1508             if (pixel_shift){
1509                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1510                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1511                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1512                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1513             } else {
1514                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1515                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1516             }
1517         } else if(chroma422) {
1518             if (pixel_shift) {
1519                 AV_COPY128(top_border+32, src_cb+16*uvlinesize);
1520                 AV_COPY128(top_border+48, src_cr+16*uvlinesize);
1521             } else {
1522                 AV_COPY64(top_border+16, src_cb+16*uvlinesize);
1523                 AV_COPY64(top_border+24, src_cr+16*uvlinesize);
1524             }
1525         } else {
1526             if (pixel_shift) {
1527                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1528                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1529             } else {
1530                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1531                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1532             }
1533         }
1534     }
1535 }
1536
1537 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1538                                   uint8_t *src_cb, uint8_t *src_cr,
1539                                   int linesize, int uvlinesize,
1540                                   int xchg, int chroma444,
1541                                   int simple, int pixel_shift){
1542     MpegEncContext * const s = &h->s;
1543     int deblock_topleft;
1544     int deblock_top;
1545     int top_idx = 1;
1546     uint8_t *top_border_m1;
1547     uint8_t *top_border;
1548
1549     if(!simple && FRAME_MBAFF){
1550         if(s->mb_y&1){
1551             if(!MB_MBAFF)
1552                 return;
1553         }else{
1554             top_idx = MB_MBAFF ? 0 : 1;
1555         }
1556     }
1557
1558     if(h->deblocking_filter == 2) {
1559         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1560         deblock_top     = h->top_type;
1561     } else {
1562         deblock_topleft = (s->mb_x > 0);
1563         deblock_top     = (s->mb_y > !!MB_FIELD);
1564     }
1565
1566     src_y  -=   linesize + 1 + pixel_shift;
1567     src_cb -= uvlinesize + 1 + pixel_shift;
1568     src_cr -= uvlinesize + 1 + pixel_shift;
1569
1570     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1571     top_border    = h->top_borders[top_idx][s->mb_x];
1572
1573 #define XCHG(a,b,xchg)\
1574     if (pixel_shift) {\
1575         if (xchg) {\
1576             AV_SWAP64(b+0,a+0);\
1577             AV_SWAP64(b+8,a+8);\
1578         } else {\
1579             AV_COPY128(b,a); \
1580         }\
1581     } else \
1582 if (xchg) AV_SWAP64(b,a);\
1583 else      AV_COPY64(b,a);
1584
1585     if(deblock_top){
1586         if(deblock_topleft){
1587             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1588         }
1589         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1590         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1591         if(s->mb_x+1 < s->mb_width){
1592             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1593         }
1594     }
1595     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1596         if(chroma444){
1597             if(deblock_topleft){
1598                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1599                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1600             }
1601             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1602             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1603             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1604             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1605             if(s->mb_x+1 < s->mb_width){
1606                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1607                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1608             }
1609         } else {
1610             if(deblock_top){
1611                 if(deblock_topleft){
1612                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1613                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1614                 }
1615                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1616                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1617             }
1618         }
1619     }
1620 }
1621
1622 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1623     if (high_bit_depth) {
1624         return AV_RN32A(((int32_t*)mb) + index);
1625     } else
1626         return AV_RN16A(mb + index);
1627 }
1628
1629 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1630     if (high_bit_depth) {
1631         AV_WN32A(((int32_t*)mb) + index, value);
1632     } else
1633         AV_WN16A(mb + index, value);
1634 }
1635
1636 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1637                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1638 {
1639     MpegEncContext * const s = &h->s;
1640     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1641     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1642     int i;
1643     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1644     block_offset += 16*p;
1645     if(IS_INTRA4x4(mb_type)){
1646         if(simple || !s->encoding){
1647             if(IS_8x8DCT(mb_type)){
1648                 if(transform_bypass){
1649                     idct_dc_add =
1650                     idct_add    = s->dsp.add_pixels8;
1651                 }else{
1652                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1653                     idct_add    = h->h264dsp.h264_idct8_add;
1654                 }
1655                 for(i=0; i<16; i+=4){
1656                     uint8_t * const ptr= dest_y + block_offset[i];
1657                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1658                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1659                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1660                     }else{
1661                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1662                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1663                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1664                         if(nnz){
1665                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1666                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1667                             else
1668                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1669                         }
1670                     }
1671                 }
1672             }else{
1673                 if(transform_bypass){
1674                     idct_dc_add =
1675                     idct_add    = s->dsp.add_pixels4;
1676                 }else{
1677                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1678                     idct_add    = h->h264dsp.h264_idct_add;
1679                 }
1680                 for(i=0; i<16; i++){
1681                     uint8_t * const ptr= dest_y + block_offset[i];
1682                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1683
1684                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1685                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1686                     }else{
1687                         uint8_t *topright;
1688                         int nnz, tr;
1689                         uint64_t tr_high;
1690                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1691                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1692                             assert(s->mb_y || linesize <= block_offset[i]);
1693                             if(!topright_avail){
1694                                 if (pixel_shift) {
1695                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1696                                     topright= (uint8_t*) &tr_high;
1697                                 } else {
1698                                     tr= ptr[3 - linesize]*0x01010101;
1699                                     topright= (uint8_t*) &tr;
1700                                 }
1701                             }else
1702                                 topright= ptr + (4 << pixel_shift) - linesize;
1703                         }else
1704                             topright= NULL;
1705
1706                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1707                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1708                         if(nnz){
1709                             if(is_h264){
1710                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1711                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1712                                 else
1713                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1714                             }else
1715                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1716                         }
1717                     }
1718                 }
1719             }
1720         }
1721     }else{
1722         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1723         if(is_h264){
1724             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1725                 if(!transform_bypass)
1726                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1727                 else{
1728                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1729                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1730                     for(i = 0; i < 16; i++)
1731                         dctcoef_set(h->mb+p*256, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1732                 }
1733             }
1734         }else
1735             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1736     }
1737 }
1738
1739 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1740                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1741 {
1742     MpegEncContext * const s = &h->s;
1743     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1744     int i;
1745     block_offset += 16*p;
1746     if(!IS_INTRA4x4(mb_type)){
1747         if(is_h264){
1748             if(IS_INTRA16x16(mb_type)){
1749                 if(transform_bypass){
1750                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1751                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1752                     }else{
1753                         for(i=0; i<16; i++){
1754                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1755                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1756                         }
1757                     }
1758                 }else{
1759                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1760                 }
1761             }else if(h->cbp&15){
1762                 if(transform_bypass){
1763                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1764                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1765                     for(i=0; i<16; i+=di){
1766                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1767                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1768                         }
1769                     }
1770                 }else{
1771                     if(IS_8x8DCT(mb_type)){
1772                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1773                     }else{
1774                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1775                     }
1776                 }
1777             }
1778         }else{
1779             for(i=0; i<16; i++){
1780                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1781                     uint8_t * const ptr= dest_y + block_offset[i];
1782                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1783                 }
1784             }
1785         }
1786     }
1787 }
1788
1789 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1790     MpegEncContext * const s = &h->s;
1791     const int mb_x= s->mb_x;
1792     const int mb_y= s->mb_y;
1793     const int mb_xy= h->mb_xy;
1794     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1795     uint8_t  *dest_y, *dest_cb, *dest_cr;
1796     int linesize, uvlinesize /*dct_offset*/;
1797     int i, j;
1798     int *block_offset = &h->block_offset[0];
1799     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1800     /* is_h264 should always be true if SVQ3 is disabled. */
1801     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1802     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1803     const int block_h = 16>>s->chroma_y_shift;
1804
1805     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1806     dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1807     dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1808
1809     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1810     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1811
1812     h->list_counts[mb_xy]= h->list_count;
1813
1814     if (!simple && MB_FIELD) {
1815         linesize   = h->mb_linesize   = s->linesize * 2;
1816         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1817         block_offset = &h->block_offset[48];
1818         if(mb_y&1){ //FIXME move out of this function?
1819             dest_y -= s->linesize*15;
1820             dest_cb-= s->uvlinesize*(block_h-1);
1821             dest_cr-= s->uvlinesize*(block_h-1);
1822         }
1823         if(FRAME_MBAFF) {
1824             int list;
1825             for(list=0; list<h->list_count; list++){
1826                 if(!USES_LIST(mb_type, list))
1827                     continue;
1828                 if(IS_16X16(mb_type)){
1829                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1830                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1831                 }else{
1832                     for(i=0; i<16; i+=4){
1833                         int ref = h->ref_cache[list][scan8[i]];
1834                         if(ref >= 0)
1835                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1836                     }
1837                 }
1838             }
1839         }
1840     } else {
1841         linesize   = h->mb_linesize   = s->linesize;
1842         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1843 //        dct_offset = s->linesize * 16;
1844     }
1845
1846     if (!simple && IS_INTRA_PCM(mb_type)) {
1847         const int bit_depth = h->sps.bit_depth_luma;
1848         if (pixel_shift) {
1849             int j;
1850             GetBitContext gb;
1851             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1852
1853             for (i = 0; i < 16; i++) {
1854                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1855                 for (j = 0; j < 16; j++)
1856                     tmp_y[j] = get_bits(&gb, bit_depth);
1857             }
1858             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1859                 if (!h->sps.chroma_format_idc) {
1860                     for (i = 0; i < 8; i++) {
1861                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1862                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1863                         for (j = 0; j < 8; j++) {
1864                             tmp_cb[j] = tmp_cr[j] = 1 << (bit_depth - 1);
1865                         }
1866                     }
1867                 } else {
1868                     for (i = 0; i < block_h; i++) {
1869                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1870                         for (j = 0; j < 8; j++)
1871                             tmp_cb[j] = get_bits(&gb, bit_depth);
1872                     }
1873                     for (i = 0; i < block_h; i++) {
1874                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1875                         for (j = 0; j < 8; j++)
1876                             tmp_cr[j] = get_bits(&gb, bit_depth);
1877                     }
1878                 }
1879             }
1880         } else {
1881             for (i=0; i<16; i++) {
1882                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1883             }
1884             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1885                 if (!h->sps.chroma_format_idc) {
1886                     for (i=0; i<8; i++) {
1887                         memset(dest_cb+ i*uvlinesize, 1 << (bit_depth - 1), 8);
1888                         memset(dest_cr+ i*uvlinesize, 1 << (bit_depth - 1), 8);
1889                     }
1890                 } else {
1891                     for (i=0; i<block_h; i++) {
1892                         memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
1893                         memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
1894                     }
1895                 }
1896             }
1897         }
1898     } else {
1899         if(IS_INTRA(mb_type)){
1900             if(h->deblocking_filter)
1901                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1902
1903             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1904                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1905                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1906             }
1907
1908             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1909
1910             if(h->deblocking_filter)
1911                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1912         }else if(is_h264){
1913             hl_motion(h, dest_y, dest_cb, dest_cr,
1914                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1915                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1916                       h->h264dsp.weight_h264_pixels_tab,
1917                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 0);
1918         }
1919
1920         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1921
1922         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1923             uint8_t *dest[2] = {dest_cb, dest_cr};
1924             if(transform_bypass){
1925                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1926                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
1927                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
1928                 }else{
1929                     idct_add = s->dsp.add_pixels4;
1930                     for(j=1; j<3; j++){
1931                         for(i=j*16; i<j*16+4; i++){
1932                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1933                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1934                         }
1935                     }
1936                 }
1937             }else{
1938                 if(is_h264){
1939                     int qp[2];
1940                     if (CHROMA422) {
1941                         qp[0] = h->chroma_qp[0]+3;
1942                         qp[1] = h->chroma_qp[1]+3;
1943                     } else {
1944                         qp[0] = h->chroma_qp[0];
1945                         qp[1] = h->chroma_qp[1];
1946                     }
1947                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1948                         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]);
1949                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1950                         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]);
1951                     h->h264dsp.h264_idct_add8(dest, block_offset,
1952                                               h->mb, uvlinesize,
1953                                               h->non_zero_count_cache);
1954                 }
1955 #if CONFIG_SVQ3_DECODER
1956                 else{
1957                     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]);
1958                     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]);
1959                     for(j=1; j<3; j++){
1960                         for(i=j*16; i<j*16+4; i++){
1961                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1962                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
1963                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1964                             }
1965                         }
1966                     }
1967                 }
1968 #endif
1969             }
1970         }
1971     }
1972     if(h->cbp || IS_INTRA(mb_type))
1973     {
1974         s->dsp.clear_blocks(h->mb);
1975         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
1976     }
1977 }
1978
1979 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
1980     MpegEncContext * const s = &h->s;
1981     const int mb_x= s->mb_x;
1982     const int mb_y= s->mb_y;
1983     const int mb_xy= h->mb_xy;
1984     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1985     uint8_t  *dest[3];
1986     int linesize;
1987     int i, j, p;
1988     int *block_offset = &h->block_offset[0];
1989     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1990     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
1991
1992     for (p = 0; p < plane_count; p++)
1993     {
1994         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
1995         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1996     }
1997
1998     h->list_counts[mb_xy]= h->list_count;
1999
2000     if (!simple && MB_FIELD) {
2001         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
2002         block_offset = &h->block_offset[48];
2003         if(mb_y&1) //FIXME move out of this function?
2004             for (p = 0; p < 3; p++)
2005                 dest[p] -= s->linesize*15;
2006         if(FRAME_MBAFF) {
2007             int list;
2008             for(list=0; list<h->list_count; list++){
2009                 if(!USES_LIST(mb_type, list))
2010                     continue;
2011                 if(IS_16X16(mb_type)){
2012                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2013                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2014                 }else{
2015                     for(i=0; i<16; i+=4){
2016                         int ref = h->ref_cache[list][scan8[i]];
2017                         if(ref >= 0)
2018                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2019                     }
2020                 }
2021             }
2022         }
2023     } else {
2024         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2025     }
2026
2027     if (!simple && IS_INTRA_PCM(mb_type)) {
2028         if (pixel_shift) {
2029             const int bit_depth = h->sps.bit_depth_luma;
2030             GetBitContext gb;
2031             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
2032
2033             for (p = 0; p < plane_count; p++) {
2034                 for (i = 0; i < 16; i++) {
2035                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2036                     for (j = 0; j < 16; j++)
2037                         tmp[j] = get_bits(&gb, bit_depth);
2038                 }
2039             }
2040         } else {
2041             for (p = 0; p < plane_count; p++) {
2042                 for (i = 0; i < 16; i++) {
2043                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2044                 }
2045             }
2046         }
2047     } else {
2048         if(IS_INTRA(mb_type)){
2049             if(h->deblocking_filter)
2050                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2051
2052             for (p = 0; p < plane_count; p++)
2053                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2054
2055             if(h->deblocking_filter)
2056                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2057         }else{
2058             hl_motion(h, dest[0], dest[1], dest[2],
2059                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2060                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2061                       h->h264dsp.weight_h264_pixels_tab,
2062                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 1);
2063         }
2064
2065         for (p = 0; p < plane_count; p++)
2066             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2067     }
2068     if(h->cbp || IS_INTRA(mb_type))
2069     {
2070         s->dsp.clear_blocks(h->mb);
2071         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2072     }
2073 }
2074
2075 /**
2076  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2077  */
2078 #define hl_decode_mb_simple(sh, bits) \
2079 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2080     hl_decode_mb_internal(h, 1, sh); \
2081 }
2082 hl_decode_mb_simple(0, 8);
2083 hl_decode_mb_simple(1, 16);
2084
2085 /**
2086  * Process a macroblock; this handles edge cases, such as interlacing.
2087  */
2088 static void av_noinline hl_decode_mb_complex(H264Context *h){
2089     hl_decode_mb_internal(h, 0, h->pixel_shift);
2090 }
2091
2092 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2093     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2094 }
2095
2096 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2097     hl_decode_mb_444_internal(h, 1, 0);
2098 }
2099
2100 void ff_h264_hl_decode_mb(H264Context *h){
2101     MpegEncContext * const s = &h->s;
2102     const int mb_xy= h->mb_xy;
2103     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2104     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2105
2106     if (CHROMA444) {
2107         if(is_complex || h->pixel_shift)
2108             hl_decode_mb_444_complex(h);
2109         else
2110             hl_decode_mb_444_simple(h);
2111     } else if (is_complex) {
2112         hl_decode_mb_complex(h);
2113     } else if (h->pixel_shift) {
2114         hl_decode_mb_simple_16(h);
2115     } else
2116         hl_decode_mb_simple_8(h);
2117 }
2118
2119 static int pred_weight_table(H264Context *h){
2120     MpegEncContext * const s = &h->s;
2121     int list, i;
2122     int luma_def, chroma_def;
2123
2124     h->use_weight= 0;
2125     h->use_weight_chroma= 0;
2126     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2127     if(h->sps.chroma_format_idc)
2128         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2129     luma_def = 1<<h->luma_log2_weight_denom;
2130     chroma_def = 1<<h->chroma_log2_weight_denom;
2131
2132     for(list=0; list<2; list++){
2133         h->luma_weight_flag[list]   = 0;
2134         h->chroma_weight_flag[list] = 0;
2135         for(i=0; i<h->ref_count[list]; i++){
2136             int luma_weight_flag, chroma_weight_flag;
2137
2138             luma_weight_flag= get_bits1(&s->gb);
2139             if(luma_weight_flag){
2140                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2141                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2142                 if(   h->luma_weight[i][list][0] != luma_def
2143                    || h->luma_weight[i][list][1] != 0) {
2144                     h->use_weight= 1;
2145                     h->luma_weight_flag[list]= 1;
2146                 }
2147             }else{
2148                 h->luma_weight[i][list][0]= luma_def;
2149                 h->luma_weight[i][list][1]= 0;
2150             }
2151
2152             if(h->sps.chroma_format_idc){
2153                 chroma_weight_flag= get_bits1(&s->gb);
2154                 if(chroma_weight_flag){
2155                     int j;
2156                     for(j=0; j<2; j++){
2157                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2158                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2159                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2160                            || h->chroma_weight[i][list][j][1] != 0) {
2161                             h->use_weight_chroma= 1;
2162                             h->chroma_weight_flag[list]= 1;
2163                         }
2164                     }
2165                 }else{
2166                     int j;
2167                     for(j=0; j<2; j++){
2168                         h->chroma_weight[i][list][j][0]= chroma_def;
2169                         h->chroma_weight[i][list][j][1]= 0;
2170                     }
2171                 }
2172             }
2173         }
2174         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2175     }
2176     h->use_weight= h->use_weight || h->use_weight_chroma;
2177     return 0;
2178 }
2179
2180 /**
2181  * Initialize implicit_weight table.
2182  * @param field  0/1 initialize the weight for interlaced MBAFF
2183  *                -1 initializes the rest
2184  */
2185 static void implicit_weight_table(H264Context *h, int field){
2186     MpegEncContext * const s = &h->s;
2187     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2188
2189     for (i = 0; i < 2; i++) {
2190         h->luma_weight_flag[i]   = 0;
2191         h->chroma_weight_flag[i] = 0;
2192     }
2193
2194     if(field < 0){
2195         cur_poc = s->current_picture_ptr->poc;
2196     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2197        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2198         h->use_weight= 0;
2199         h->use_weight_chroma= 0;
2200         return;
2201     }
2202         ref_start= 0;
2203         ref_count0= h->ref_count[0];
2204         ref_count1= h->ref_count[1];
2205     }else{
2206         cur_poc = s->current_picture_ptr->field_poc[field];
2207         ref_start= 16;
2208         ref_count0= 16+2*h->ref_count[0];
2209         ref_count1= 16+2*h->ref_count[1];
2210     }
2211
2212     h->use_weight= 2;
2213     h->use_weight_chroma= 2;
2214     h->luma_log2_weight_denom= 5;
2215     h->chroma_log2_weight_denom= 5;
2216
2217     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2218         int poc0 = h->ref_list[0][ref0].poc;
2219         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2220             int w = 32;
2221             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2222                 int poc1 = h->ref_list[1][ref1].poc;
2223                 int td = av_clip(poc1 - poc0, -128, 127);
2224                 if(td){
2225                     int tb = av_clip(cur_poc - poc0, -128, 127);
2226                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2227                     int dist_scale_factor = (tb*tx + 32) >> 8;
2228                     if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2229                         w = 64 - dist_scale_factor;
2230                 }
2231             }
2232             if(field<0){
2233                 h->implicit_weight[ref0][ref1][0]=
2234                 h->implicit_weight[ref0][ref1][1]= w;
2235             }else{
2236                 h->implicit_weight[ref0][ref1][field]=w;
2237             }
2238         }
2239     }
2240 }
2241
2242 /**
2243  * instantaneous decoder refresh.
2244  */
2245 static void idr(H264Context *h){
2246     ff_h264_remove_all_refs(h);
2247     h->prev_frame_num= 0;
2248     h->prev_frame_num_offset= 0;
2249     h->prev_poc_msb=
2250     h->prev_poc_lsb= 0;
2251 }
2252
2253 /* forget old pics after a seek */
2254 static void flush_dpb(AVCodecContext *avctx){
2255     H264Context *h= avctx->priv_data;
2256     int i;
2257     for(i=0; i<=MAX_DELAYED_PIC_COUNT; i++) {
2258         if(h->delayed_pic[i])
2259             h->delayed_pic[i]->f.reference = 0;
2260         h->delayed_pic[i]= NULL;
2261     }
2262     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2263     h->prev_interlaced_frame = 1;
2264     idr(h);
2265     if(h->s.current_picture_ptr)
2266         h->s.current_picture_ptr->f.reference = 0;
2267     h->s.first_field= 0;
2268     ff_h264_reset_sei(h);
2269     ff_mpeg_flush(avctx);
2270     h->sync= 0;
2271 }
2272
2273 static int init_poc(H264Context *h){
2274     MpegEncContext * const s = &h->s;
2275     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2276     int field_poc[2];
2277     Picture *cur = s->current_picture_ptr;
2278
2279     h->frame_num_offset= h->prev_frame_num_offset;
2280     if(h->frame_num < h->prev_frame_num)
2281         h->frame_num_offset += max_frame_num;
2282
2283     if(h->sps.poc_type==0){
2284         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2285
2286         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2287             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2288         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2289             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2290         else
2291             h->poc_msb = h->prev_poc_msb;
2292 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2293         field_poc[0] =
2294         field_poc[1] = h->poc_msb + h->poc_lsb;
2295         if(s->picture_structure == PICT_FRAME)
2296             field_poc[1] += h->delta_poc_bottom;
2297     }else if(h->sps.poc_type==1){
2298         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2299         int i;
2300
2301         if(h->sps.poc_cycle_length != 0)
2302             abs_frame_num = h->frame_num_offset + h->frame_num;
2303         else
2304             abs_frame_num = 0;
2305
2306         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2307             abs_frame_num--;
2308
2309         expected_delta_per_poc_cycle = 0;
2310         for(i=0; i < h->sps.poc_cycle_length; i++)
2311             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2312
2313         if(abs_frame_num > 0){
2314             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2315             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2316
2317             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2318             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2319                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2320         } else
2321             expectedpoc = 0;
2322
2323         if(h->nal_ref_idc == 0)
2324             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2325
2326         field_poc[0] = expectedpoc + h->delta_poc[0];
2327         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2328
2329         if(s->picture_structure == PICT_FRAME)
2330             field_poc[1] += h->delta_poc[1];
2331     }else{
2332         int poc= 2*(h->frame_num_offset + h->frame_num);
2333
2334         if(!h->nal_ref_idc)
2335             poc--;
2336
2337         field_poc[0]= poc;
2338         field_poc[1]= poc;
2339     }
2340
2341     if(s->picture_structure != PICT_BOTTOM_FIELD)
2342         s->current_picture_ptr->field_poc[0]= field_poc[0];
2343     if(s->picture_structure != PICT_TOP_FIELD)
2344         s->current_picture_ptr->field_poc[1]= field_poc[1];
2345     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2346
2347     return 0;
2348 }
2349
2350
2351 /**
2352  * initialize scan tables
2353  */
2354 static void init_scan_tables(H264Context *h){
2355     int i;
2356     for(i=0; i<16; i++){
2357 #define T(x) (x>>2) | ((x<<2) & 0xF)
2358         h->zigzag_scan[i] = T(zigzag_scan[i]);
2359         h-> field_scan[i] = T( field_scan[i]);
2360 #undef T
2361     }
2362     for(i=0; i<64; i++){
2363 #define T(x) (x>>3) | ((x&7)<<3)
2364         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2365         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2366         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2367         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2368 #undef T
2369     }
2370     if(h->sps.transform_bypass){ //FIXME same ugly
2371         h->zigzag_scan_q0          = zigzag_scan;
2372         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2373         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2374         h->field_scan_q0           = field_scan;
2375         h->field_scan8x8_q0        = field_scan8x8;
2376         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2377     }else{
2378         h->zigzag_scan_q0          = h->zigzag_scan;
2379         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2380         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2381         h->field_scan_q0           = h->field_scan;
2382         h->field_scan8x8_q0        = h->field_scan8x8;
2383         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2384     }
2385 }
2386
2387 static int field_end(H264Context *h, int in_setup){
2388     MpegEncContext * const s = &h->s;
2389     AVCodecContext * const avctx= s->avctx;
2390     int err = 0;
2391     s->mb_y= 0;
2392
2393     if (!in_setup && !s->dropable)
2394         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2395                                  s->picture_structure==PICT_BOTTOM_FIELD);
2396
2397     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2398         ff_vdpau_h264_set_reference_frames(s);
2399
2400     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2401         if(!s->dropable) {
2402             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2403             h->prev_poc_msb= h->poc_msb;
2404             h->prev_poc_lsb= h->poc_lsb;
2405         }
2406         h->prev_frame_num_offset= h->frame_num_offset;
2407         h->prev_frame_num= h->frame_num;
2408         h->outputed_poc = h->next_outputed_poc;
2409     }
2410
2411     if (avctx->hwaccel) {
2412         if (avctx->hwaccel->end_frame(avctx) < 0)
2413             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2414     }
2415
2416     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2417         ff_vdpau_h264_picture_complete(s);
2418
2419     /*
2420      * FIXME: Error handling code does not seem to support interlaced
2421      * when slices span multiple rows
2422      * The ff_er_add_slice calls don't work right for bottom
2423      * fields; they cause massive erroneous error concealing
2424      * Error marking covers both fields (top and bottom).
2425      * This causes a mismatched s->error_count
2426      * and a bad error table. Further, the error count goes to
2427      * INT_MAX when called for bottom field, because mb_y is
2428      * past end by one (callers fault) and resync_mb_y != 0
2429      * causes problems for the first MB line, too.
2430      */
2431     if (!FIELD_PICTURE)
2432         ff_er_frame_end(s);
2433
2434     MPV_frame_end(s);
2435
2436     h->current_slice=0;
2437
2438     return err;
2439 }
2440
2441 /**
2442  * Replicate H264 "master" context to thread contexts.
2443  */
2444 static void clone_slice(H264Context *dst, H264Context *src)
2445 {
2446     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2447     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2448     dst->s.current_picture      = src->s.current_picture;
2449     dst->s.linesize             = src->s.linesize;
2450     dst->s.uvlinesize           = src->s.uvlinesize;
2451     dst->s.first_field          = src->s.first_field;
2452
2453     dst->prev_poc_msb           = src->prev_poc_msb;
2454     dst->prev_poc_lsb           = src->prev_poc_lsb;
2455     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2456     dst->prev_frame_num         = src->prev_frame_num;
2457     dst->short_ref_count        = src->short_ref_count;
2458
2459     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2460     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2461     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2462     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2463
2464     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2465     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2466 }
2467
2468 /**
2469  * computes profile from profile_idc and constraint_set?_flags
2470  *
2471  * @param sps SPS
2472  *
2473  * @return profile as defined by FF_PROFILE_H264_*
2474  */
2475 int ff_h264_get_profile(SPS *sps)
2476 {
2477     int profile = sps->profile_idc;
2478
2479     switch(sps->profile_idc) {
2480     case FF_PROFILE_H264_BASELINE:
2481         // constraint_set1_flag set to 1
2482         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2483         break;
2484     case FF_PROFILE_H264_HIGH_10:
2485     case FF_PROFILE_H264_HIGH_422:
2486     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2487         // constraint_set3_flag set to 1
2488         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2489         break;
2490     }
2491
2492     return profile;
2493 }
2494
2495 /**
2496  * decodes a slice header.
2497  * This will also call MPV_common_init() and frame_start() as needed.
2498  *
2499  * @param h h264context
2500  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2501  *
2502  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2503  */
2504 static int decode_slice_header(H264Context *h, H264Context *h0){
2505     MpegEncContext * const s = &h->s;
2506     MpegEncContext * const s0 = &h0->s;
2507     unsigned int first_mb_in_slice;
2508     unsigned int pps_id;
2509     int num_ref_idx_active_override_flag;
2510     unsigned int slice_type, tmp, i, j;
2511     int default_ref_list_done = 0;
2512     int last_pic_structure;
2513
2514     s->dropable= h->nal_ref_idc == 0;
2515
2516     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2517     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2518         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2519         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2520     }else{
2521         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2522         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2523     }
2524
2525     first_mb_in_slice= get_ue_golomb(&s->gb);
2526
2527     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2528         if(h0->current_slice && FIELD_PICTURE){
2529             field_end(h, 1);
2530         }
2531
2532         h0->current_slice = 0;
2533         if (!s0->first_field)
2534             s->current_picture_ptr= NULL;
2535     }
2536
2537     slice_type= get_ue_golomb_31(&s->gb);
2538     if(slice_type > 9){
2539         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);
2540         return -1;
2541     }
2542     if(slice_type > 4){
2543         slice_type -= 5;
2544         h->slice_type_fixed=1;
2545     }else
2546         h->slice_type_fixed=0;
2547
2548     slice_type= golomb_to_pict_type[ slice_type ];
2549     if (slice_type == AV_PICTURE_TYPE_I
2550         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2551         default_ref_list_done = 1;
2552     }
2553     h->slice_type= slice_type;
2554     h->slice_type_nos= slice_type & 3;
2555
2556     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2557
2558     pps_id= get_ue_golomb(&s->gb);
2559     if(pps_id>=MAX_PPS_COUNT){
2560         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2561         return -1;
2562     }
2563     if(!h0->pps_buffers[pps_id]) {
2564         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2565         return -1;
2566     }
2567     h->pps= *h0->pps_buffers[pps_id];
2568
2569     if(!h0->sps_buffers[h->pps.sps_id]) {
2570         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2571         return -1;
2572     }
2573     h->sps = *h0->sps_buffers[h->pps.sps_id];
2574
2575     s->avctx->profile = ff_h264_get_profile(&h->sps);
2576     s->avctx->level   = h->sps.level_idc;
2577     s->avctx->refs    = h->sps.ref_frame_count;
2578
2579     if(h == h0 && h->dequant_coeff_pps != pps_id){
2580         h->dequant_coeff_pps = pps_id;
2581         init_dequant_tables(h);
2582     }
2583
2584     s->mb_width= h->sps.mb_width;
2585     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2586
2587     h->b_stride=  s->mb_width*4;
2588
2589     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2590
2591     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2592     if(h->sps.frame_mbs_only_flag)
2593         s->height= 16*s->mb_height - (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2594     else
2595         s->height= 16*s->mb_height - (2<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2596
2597     if (s->context_initialized
2598         && (   s->width != s->avctx->width || s->height != s->avctx->height
2599             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2600         if(h != h0) {
2601             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2602             return -1;   // width / height changed during parallelized decoding
2603         }
2604         free_tables(h, 0);
2605         flush_dpb(s->avctx);
2606         MPV_common_end(s);
2607     }
2608     if (!s->context_initialized) {
2609         if (h != h0) {
2610             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2611             return -1;
2612         }
2613
2614         avcodec_set_dimensions(s->avctx, s->width, s->height);
2615         s->avctx->sample_aspect_ratio= h->sps.sar;
2616         av_assert0(s->avctx->sample_aspect_ratio.den);
2617
2618         if(h->sps.video_signal_type_present_flag){
2619             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2620             if(h->sps.colour_description_present_flag){
2621                 s->avctx->color_primaries = h->sps.color_primaries;
2622                 s->avctx->color_trc       = h->sps.color_trc;
2623                 s->avctx->colorspace      = h->sps.colorspace;
2624             }
2625         }
2626
2627         if(h->sps.timing_info_present_flag){
2628             int64_t den= h->sps.time_scale;
2629             if(h->x264_build < 44U)
2630                 den *= 2;
2631             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2632                       h->sps.num_units_in_tick, den, 1<<30);
2633         }
2634
2635         switch (h->sps.bit_depth_luma) {
2636             case 9 :
2637                 if (CHROMA444)
2638                     s->avctx->pix_fmt = PIX_FMT_YUV444P9;
2639                 else
2640                     s->avctx->pix_fmt = PIX_FMT_YUV420P9;
2641                 break;
2642             case 10 :
2643                 if (CHROMA444)
2644                     s->avctx->pix_fmt = PIX_FMT_YUV444P10;
2645                 else if (CHROMA422)
2646                     s->avctx->pix_fmt = PIX_FMT_YUV422P10;
2647                 else
2648                     s->avctx->pix_fmt = PIX_FMT_YUV420P10;
2649                 break;
2650             default:
2651                 if (CHROMA444){
2652                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2653                 }else if (CHROMA422) {
2654                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P : PIX_FMT_YUV422P;
2655                 }else{
2656                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2657                                                              s->avctx->codec->pix_fmts ?
2658                                                              s->avctx->codec->pix_fmts :
2659                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2660                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2661                                                              ff_hwaccel_pixfmt_list_420);
2662                 }
2663         }
2664
2665         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2666
2667         if (MPV_common_init(s) < 0) {
2668             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2669             return -1;
2670         }
2671         s->first_field = 0;
2672         h->prev_interlaced_frame = 1;
2673
2674         init_scan_tables(h);
2675         if (ff_h264_alloc_tables(h) < 0) {
2676             av_log(h->s.avctx, AV_LOG_ERROR, "Could not allocate memory for h264\n");
2677             return AVERROR(ENOMEM);
2678         }
2679
2680         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2681             if (context_init(h) < 0) {
2682                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2683                 return -1;
2684             }
2685         } else {
2686             for(i = 1; i < s->avctx->thread_count; i++) {
2687                 H264Context *c;
2688                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2689                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2690                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2691                 c->h264dsp = h->h264dsp;
2692                 c->sps = h->sps;
2693                 c->pps = h->pps;
2694                 c->pixel_shift = h->pixel_shift;
2695                 init_scan_tables(c);
2696                 clone_tables(c, h, i);
2697             }
2698
2699             for(i = 0; i < s->avctx->thread_count; i++)
2700                 if (context_init(h->thread_context[i]) < 0) {
2701                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2702                     return -1;
2703                 }
2704         }
2705     }
2706
2707     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2708
2709     h->mb_mbaff = 0;
2710     h->mb_aff_frame = 0;
2711     last_pic_structure = s0->picture_structure;
2712     if(h->sps.frame_mbs_only_flag){
2713         s->picture_structure= PICT_FRAME;
2714     }else{
2715         if(get_bits1(&s->gb)) { //field_pic_flag
2716             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2717         } else {
2718             s->picture_structure= PICT_FRAME;
2719             h->mb_aff_frame = h->sps.mb_aff;
2720         }
2721     }
2722     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2723
2724     if(h0->current_slice == 0){
2725         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2726         if(h->frame_num != h->prev_frame_num) {
2727             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2728
2729             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2730
2731             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2732                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2733                 if (unwrap_prev_frame_num < 0)
2734                     unwrap_prev_frame_num += max_frame_num;
2735
2736                 h->prev_frame_num = unwrap_prev_frame_num;
2737             }
2738         }
2739
2740         while(h->frame_num !=  h->prev_frame_num &&
2741               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2742             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2743             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2744             if (ff_h264_frame_start(h) < 0)
2745                 return -1;
2746             h->prev_frame_num++;
2747             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2748             s->current_picture_ptr->frame_num= h->prev_frame_num;
2749             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2750             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2751             ff_generate_sliding_window_mmcos(h);
2752             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2753                 s->avctx->error_recognition >= FF_ER_EXPLODE)
2754                 return AVERROR_INVALIDDATA;
2755             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2756              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2757              * about there being no actual duplicates.
2758              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2759              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2760              * be fixed. */
2761             if (h->short_ref_count) {
2762                 if (prev) {
2763                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2764                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2765                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2766                     h->short_ref[0]->poc = prev->poc+2;
2767                 }
2768                 h->short_ref[0]->frame_num = h->prev_frame_num;
2769             }
2770         }
2771
2772         /* See if we have a decoded first field looking for a pair... */
2773         if (s0->first_field) {
2774             assert(s0->current_picture_ptr);
2775             assert(s0->current_picture_ptr->f.data[0]);
2776             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
2777
2778             /* figure out if we have a complementary field pair */
2779             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2780                 /*
2781                  * Previous field is unmatched. Don't display it, but let it
2782                  * remain for reference if marked as such.
2783                  */
2784                 s0->current_picture_ptr = NULL;
2785                 s0->first_field = FIELD_PICTURE;
2786
2787             } else {
2788                 if (h->nal_ref_idc &&
2789                         s0->current_picture_ptr->f.reference &&
2790                         s0->current_picture_ptr->frame_num != h->frame_num) {
2791                     /*
2792                      * This and previous field were reference, but had
2793                      * different frame_nums. Consider this field first in
2794                      * pair. Throw away previous field except for reference
2795                      * purposes.
2796                      */
2797                     s0->first_field = 1;
2798                     s0->current_picture_ptr = NULL;
2799
2800                 } else {
2801                     /* Second field in complementary pair */
2802                     s0->first_field = 0;
2803                 }
2804             }
2805
2806         } else {
2807             /* Frame or first field in a potentially complementary pair */
2808             assert(!s0->current_picture_ptr);
2809             s0->first_field = FIELD_PICTURE;
2810         }
2811
2812         if(!FIELD_PICTURE || s0->first_field) {
2813             if (ff_h264_frame_start(h) < 0) {
2814                 s0->first_field = 0;
2815                 return -1;
2816             }
2817         } else {
2818             ff_release_unused_pictures(s, 0);
2819         }
2820     }
2821     if(h != h0)
2822         clone_slice(h, h0);
2823
2824     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2825
2826     assert(s->mb_num == s->mb_width * s->mb_height);
2827     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2828        first_mb_in_slice                    >= s->mb_num){
2829         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2830         return -1;
2831     }
2832     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2833     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2834     if (s->picture_structure == PICT_BOTTOM_FIELD)
2835         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2836     assert(s->mb_y < s->mb_height);
2837
2838     if(s->picture_structure==PICT_FRAME){
2839         h->curr_pic_num=   h->frame_num;
2840         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2841     }else{
2842         h->curr_pic_num= 2*h->frame_num + 1;
2843         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2844     }
2845
2846     if(h->nal_unit_type == NAL_IDR_SLICE){
2847         get_ue_golomb(&s->gb); /* idr_pic_id */
2848     }
2849
2850     if(h->sps.poc_type==0){
2851         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2852
2853         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2854             h->delta_poc_bottom= get_se_golomb(&s->gb);
2855         }
2856     }
2857
2858     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2859         h->delta_poc[0]= get_se_golomb(&s->gb);
2860
2861         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2862             h->delta_poc[1]= get_se_golomb(&s->gb);
2863     }
2864
2865     init_poc(h);
2866
2867     if(h->pps.redundant_pic_cnt_present){
2868         h->redundant_pic_count= get_ue_golomb(&s->gb);
2869     }
2870
2871     //set defaults, might be overridden a few lines later
2872     h->ref_count[0]= h->pps.ref_count[0];
2873     h->ref_count[1]= h->pps.ref_count[1];
2874
2875     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2876         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2877             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2878         }
2879         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2880
2881         if(num_ref_idx_active_override_flag){
2882             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2883             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2884                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2885
2886             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2887                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2888                 h->ref_count[0]= h->ref_count[1]= 1;
2889                 return -1;
2890             }
2891         }
2892         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2893             h->list_count= 2;
2894         else
2895             h->list_count= 1;
2896     }else
2897         h->list_count= 0;
2898
2899     if(!default_ref_list_done){
2900         ff_h264_fill_default_ref_list(h);
2901     }
2902
2903     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0)
2904         return -1;
2905
2906     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2907         s->last_picture_ptr= &h->ref_list[0][0];
2908         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2909     }
2910     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2911         s->next_picture_ptr= &h->ref_list[1][0];
2912         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2913     }
2914
2915     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2916        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2917         pred_weight_table(h);
2918     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2919         implicit_weight_table(h, -1);
2920     }else {
2921         h->use_weight = 0;
2922         for (i = 0; i < 2; i++) {
2923             h->luma_weight_flag[i]   = 0;
2924             h->chroma_weight_flag[i] = 0;
2925         }
2926     }
2927
2928     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
2929        s->avctx->error_recognition >= FF_ER_EXPLODE)
2930         return AVERROR_INVALIDDATA;
2931
2932     if(FRAME_MBAFF){
2933         ff_h264_fill_mbaff_ref_list(h);
2934
2935         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2936             implicit_weight_table(h, 0);
2937             implicit_weight_table(h, 1);
2938         }
2939     }
2940
2941     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2942         ff_h264_direct_dist_scale_factor(h);
2943     ff_h264_direct_ref_list_init(h);
2944
2945     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2946         tmp = get_ue_golomb_31(&s->gb);
2947         if(tmp > 2){
2948             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2949             return -1;
2950         }
2951         h->cabac_init_idc= tmp;
2952     }
2953
2954     h->last_qscale_diff = 0;
2955     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2956     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
2957         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2958         return -1;
2959     }
2960     s->qscale= tmp;
2961     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2962     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2963     //FIXME qscale / qp ... stuff
2964     if(h->slice_type == AV_PICTURE_TYPE_SP){
2965         get_bits1(&s->gb); /* sp_for_switch_flag */
2966     }
2967     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
2968         get_se_golomb(&s->gb); /* slice_qs_delta */
2969     }
2970
2971     h->deblocking_filter = 1;
2972     h->slice_alpha_c0_offset = 52;
2973     h->slice_beta_offset = 52;
2974     if( h->pps.deblocking_filter_parameters_present ) {
2975         tmp= get_ue_golomb_31(&s->gb);
2976         if(tmp > 2){
2977             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2978             return -1;
2979         }
2980         h->deblocking_filter= tmp;
2981         if(h->deblocking_filter < 2)
2982             h->deblocking_filter^= 1; // 1<->0
2983
2984         if( h->deblocking_filter ) {
2985             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2986             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2987             if(   h->slice_alpha_c0_offset > 104U
2988                || h->slice_beta_offset     > 104U){
2989                 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);
2990                 return -1;
2991             }
2992         }
2993     }
2994
2995     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2996        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
2997        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
2998        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2999         h->deblocking_filter= 0;
3000
3001     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
3002         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
3003             /* Cheat slightly for speed:
3004                Do not bother to deblock across slices. */
3005             h->deblocking_filter = 2;
3006         } else {
3007             h0->max_contexts = 1;
3008             if(!h0->single_decode_warning) {
3009                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3010                 h0->single_decode_warning = 1;
3011             }
3012             if (h != h0) {
3013                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
3014                 return 1;
3015             }
3016         }
3017     }
3018     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
3019                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
3020                  + 6 * (h->sps.bit_depth_luma - 8);
3021
3022 #if 0 //FMO
3023     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
3024         slice_group_change_cycle= get_bits(&s->gb, ?);
3025 #endif
3026
3027     h0->last_slice_type = slice_type;
3028     h->slice_num = ++h0->current_slice;
3029
3030     if(h->slice_num)
3031         h0->slice_row[(h->slice_num-1)&(MAX_SLICES-1)]= s->resync_mb_y;
3032     if (   h0->slice_row[h->slice_num&(MAX_SLICES-1)] + 3 >= s->resync_mb_y
3033         && h0->slice_row[h->slice_num&(MAX_SLICES-1)] <= s->resync_mb_y
3034         && h->slice_num >= MAX_SLICES) {
3035         //in case of ASO this check needs to be updated depending on how we decide to assign slice numbers in this case
3036         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);
3037     }
3038
3039     for(j=0; j<2; j++){
3040         int id_list[16];
3041         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
3042         for(i=0; i<16; i++){
3043             id_list[i]= 60;
3044             if (h->ref_list[j][i].f.data[0]) {
3045                 int k;
3046                 uint8_t *base = h->ref_list[j][i].f.base[0];
3047                 for(k=0; k<h->short_ref_count; k++)
3048                     if (h->short_ref[k]->f.base[0] == base) {
3049                         id_list[i]= k;
3050                         break;
3051                     }
3052                 for(k=0; k<h->long_ref_count; k++)
3053                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3054                         id_list[i]= h->short_ref_count + k;
3055                         break;
3056                     }
3057             }
3058         }
3059
3060         ref2frm[0]=
3061         ref2frm[1]= -1;
3062         for(i=0; i<16; i++)
3063             ref2frm[i+2]= 4*id_list[i]
3064                           + (h->ref_list[j][i].f.reference & 3);
3065         ref2frm[18+0]=
3066         ref2frm[18+1]= -1;
3067         for(i=16; i<48; i++)
3068             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3069                           + (h->ref_list[j][i].f.reference & 3);
3070     }
3071
3072     //FIXME: fix draw_edges+PAFF+frame threads
3073     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3074     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3075
3076     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3077         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",
3078                h->slice_num,
3079                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3080                first_mb_in_slice,
3081                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3082                pps_id, h->frame_num,
3083                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3084                h->ref_count[0], h->ref_count[1],
3085                s->qscale,
3086                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3087                h->use_weight,
3088                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3089                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3090                );
3091     }
3092
3093     return 0;
3094 }
3095
3096 int ff_h264_get_slice_type(const H264Context *h)
3097 {
3098     switch (h->slice_type) {
3099     case AV_PICTURE_TYPE_P:  return 0;
3100     case AV_PICTURE_TYPE_B:  return 1;
3101     case AV_PICTURE_TYPE_I:  return 2;
3102     case AV_PICTURE_TYPE_SP: return 3;
3103     case AV_PICTURE_TYPE_SI: return 4;
3104     default:         return -1;
3105     }
3106 }
3107
3108 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3109                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3110 {
3111     int b_stride = h->b_stride;
3112     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3113     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3114     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3115         if(USES_LIST(top_type, list)){
3116             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3117             const int b8_xy= 4*top_xy + 2;
3118             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3119             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3120             ref_cache[0 - 1*8]=
3121             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3122             ref_cache[2 - 1*8]=
3123             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3124         }else{
3125             AV_ZERO128(mv_dst - 1*8);
3126             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3127         }
3128
3129         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3130             if(USES_LIST(left_type[LTOP], list)){
3131                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3132                 const int b8_xy= 4*left_xy[LTOP] + 1;
3133                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3134                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3135                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3136                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3137                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3138                 ref_cache[-1 +  0]=
3139                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3140                 ref_cache[-1 + 16]=
3141                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3142             }else{
3143                 AV_ZERO32(mv_dst - 1 + 0);
3144                 AV_ZERO32(mv_dst - 1 + 8);
3145                 AV_ZERO32(mv_dst - 1 +16);
3146                 AV_ZERO32(mv_dst - 1 +24);
3147                 ref_cache[-1 +  0]=
3148                 ref_cache[-1 +  8]=
3149                 ref_cache[-1 + 16]=
3150                 ref_cache[-1 + 24]= LIST_NOT_USED;
3151             }
3152         }
3153     }
3154
3155     if(!USES_LIST(mb_type, list)){
3156         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3157         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3158         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3159         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3160         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3161         return;
3162     }
3163
3164     {
3165         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3166         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3167         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3168         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3169         AV_WN32A(&ref_cache[0*8], ref01);
3170         AV_WN32A(&ref_cache[1*8], ref01);
3171         AV_WN32A(&ref_cache[2*8], ref23);
3172         AV_WN32A(&ref_cache[3*8], ref23);
3173     }
3174
3175     {
3176         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3177         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3178         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3179         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3180         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3181     }
3182 }
3183
3184 /**
3185  *
3186  * @return non zero if the loop filter can be skiped
3187  */
3188 static int fill_filter_caches(H264Context *h, int mb_type){
3189     MpegEncContext * const s = &h->s;
3190     const int mb_xy= h->mb_xy;
3191     int top_xy, left_xy[LEFT_MBS];
3192     int top_type, left_type[LEFT_MBS];
3193     uint8_t *nnz;
3194     uint8_t *nnz_cache;
3195
3196     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3197
3198     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3199      * stuff, I can't imagine that these complex rules are worth it. */
3200
3201     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3202     if(FRAME_MBAFF){
3203         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3204         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3205         if(s->mb_y&1){
3206             if (left_mb_field_flag != curr_mb_field_flag) {
3207                 left_xy[LTOP] -= s->mb_stride;
3208             }
3209         }else{
3210             if(curr_mb_field_flag){
3211                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3212             }
3213             if (left_mb_field_flag != curr_mb_field_flag) {
3214                 left_xy[LBOT] += s->mb_stride;
3215             }
3216         }
3217     }
3218
3219     h->top_mb_xy = top_xy;
3220     h->left_mb_xy[LTOP] = left_xy[LTOP];
3221     h->left_mb_xy[LBOT] = left_xy[LBOT];
3222     {
3223         //for sufficiently low qp, filtering wouldn't do anything
3224         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3225         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3226         int qp = s->current_picture.f.qscale_table[mb_xy];
3227         if(qp <= qp_thresh
3228            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3229            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3230             if(!FRAME_MBAFF)
3231                 return 1;
3232             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3233                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3234                 return 1;
3235         }
3236     }
3237
3238     top_type        = s->current_picture.f.mb_type[top_xy];
3239     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3240     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3241     if(h->deblocking_filter == 2){
3242         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3243         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3244     }else{
3245         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3246         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3247     }
3248     h->top_type       = top_type;
3249     h->left_type[LTOP]= left_type[LTOP];
3250     h->left_type[LBOT]= left_type[LBOT];
3251
3252     if(IS_INTRA(mb_type))
3253         return 0;
3254
3255     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3256     if(h->list_count == 2)
3257         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3258
3259     nnz = h->non_zero_count[mb_xy];
3260     nnz_cache = h->non_zero_count_cache;
3261     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3262     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3263     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3264     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3265     h->cbp= h->cbp_table[mb_xy];
3266
3267     if(top_type){
3268         nnz = h->non_zero_count[top_xy];
3269         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3270     }
3271
3272     if(left_type[LTOP]){
3273         nnz = h->non_zero_count[left_xy[LTOP]];
3274         nnz_cache[3+8*1]= nnz[3+0*4];
3275         nnz_cache[3+8*2]= nnz[3+1*4];
3276         nnz_cache[3+8*3]= nnz[3+2*4];
3277         nnz_cache[3+8*4]= nnz[3+3*4];
3278     }
3279
3280     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3281     if(!CABAC && h->pps.transform_8x8_mode){
3282         if(IS_8x8DCT(top_type)){
3283             nnz_cache[4+8*0]=
3284             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3285             nnz_cache[6+8*0]=
3286             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3287         }
3288         if(IS_8x8DCT(left_type[LTOP])){
3289             nnz_cache[3+8*1]=
3290             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3291         }
3292         if(IS_8x8DCT(left_type[LBOT])){
3293             nnz_cache[3+8*3]=
3294             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3295         }
3296
3297         if(IS_8x8DCT(mb_type)){
3298             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3299             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3300
3301             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3302             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3303
3304             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3305             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3306
3307             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3308             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3309         }
3310     }
3311
3312     return 0;
3313 }
3314
3315 static void loop_filter(H264Context *h, int start_x, int end_x){
3316     MpegEncContext * const s = &h->s;
3317     uint8_t  *dest_y, *dest_cb, *dest_cr;
3318     int linesize, uvlinesize, mb_x, mb_y;
3319     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3320     const int old_slice_type= h->slice_type;
3321     const int pixel_shift = h->pixel_shift;
3322     const int block_h = 16>>s->chroma_y_shift;
3323
3324     if(h->deblocking_filter) {
3325         for(mb_x= start_x; mb_x<end_x; mb_x++){
3326             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3327                 int mb_xy, mb_type;
3328                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3329                 h->slice_num= h->slice_table[mb_xy];
3330                 mb_type = s->current_picture.f.mb_type[mb_xy];
3331                 h->list_count= h->list_counts[mb_xy];
3332
3333                 if(FRAME_MBAFF)
3334                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3335
3336                 s->mb_x= mb_x;
3337                 s->mb_y= mb_y;
3338                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3339                 dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*(8<<CHROMA444) + mb_y * s->uvlinesize * block_h;
3340                 dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*(8<<CHROMA444) + mb_y * s->uvlinesize * block_h;
3341                     //FIXME simplify above
3342
3343                 if (MB_FIELD) {
3344                     linesize   = h->mb_linesize   = s->linesize * 2;
3345                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3346                     if(mb_y&1){ //FIXME move out of this function?
3347                         dest_y -= s->linesize*15;
3348                         dest_cb-= s->uvlinesize*(block_h-1);
3349                         dest_cr-= s->uvlinesize*(block_h-1);
3350                     }
3351                 } else {
3352                     linesize   = h->mb_linesize   = s->linesize;
3353                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3354                 }
3355                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
3356                 if(fill_filter_caches(h, mb_type))
3357                     continue;
3358                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3359                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3360
3361                 if (FRAME_MBAFF) {
3362                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3363                 } else {
3364                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3365                 }
3366             }
3367         }
3368     }
3369     h->slice_type= old_slice_type;
3370     s->mb_x= end_x;
3371     s->mb_y= end_mb_y - FRAME_MBAFF;
3372     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3373     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3374 }
3375
3376 static void predict_field_decoding_flag(H264Context *h){
3377     MpegEncContext * const s = &h->s;
3378     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3379     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3380                 ? s->current_picture.f.mb_type[mb_xy - 1]
3381                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3382                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3383                 : 0;
3384     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3385 }
3386
3387 /**
3388  * Draw edges and report progress for the last MB row.
3389  */
3390 static void decode_finish_row(H264Context *h){
3391     MpegEncContext * const s = &h->s;
3392     int top = 16*(s->mb_y >> FIELD_PICTURE);
3393     int height = 16 << FRAME_MBAFF;
3394     int deblock_border = (16 + 4) << FRAME_MBAFF;
3395     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3396
3397     if (h->deblocking_filter) {
3398         if((top + height) >= pic_height)
3399             height += deblock_border;
3400
3401         top -= deblock_border;
3402     }
3403
3404     if (top >= pic_height || (top + height) < h->emu_edge_height)
3405         return;
3406
3407     height = FFMIN(height, pic_height - top);
3408     if (top < h->emu_edge_height) {
3409         height = top+height;
3410         top = 0;
3411     }
3412
3413     ff_draw_horiz_band(s, top, height);
3414
3415     if (s->dropable) return;
3416
3417     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3418                              s->picture_structure==PICT_BOTTOM_FIELD);
3419 }
3420
3421 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3422     H264Context *h = *(void**)arg;
3423     MpegEncContext * const s = &h->s;
3424     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3425     int lf_x_start = s->mb_x;
3426
3427     s->mb_skip_run= -1;
3428
3429     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3430                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3431
3432     if( h->pps.cabac ) {
3433         /* realign */
3434         align_get_bits( &s->gb );
3435
3436         /* init cabac */
3437         ff_init_cabac_states( &h->cabac);
3438         ff_init_cabac_decoder( &h->cabac,
3439                                s->gb.buffer + get_bits_count(&s->gb)/8,
3440                                (get_bits_left(&s->gb) + 7)/8);
3441
3442         ff_h264_init_cabac_states(h);
3443
3444         for(;;){
3445 //START_TIMER
3446             int ret = ff_h264_decode_mb_cabac(h);
3447             int eos;
3448 //STOP_TIMER("decode_mb_cabac")
3449
3450             if(ret>=0) ff_h264_hl_decode_mb(h);
3451
3452             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3453                 s->mb_y++;
3454
3455                 ret = ff_h264_decode_mb_cabac(h);
3456
3457                 if(ret>=0) ff_h264_hl_decode_mb(h);
3458                 s->mb_y--;
3459             }
3460             eos = get_cabac_terminate( &h->cabac );
3461
3462             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3463                 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);
3464                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);