w32threads: support for frame multithreading
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of Libav.
6  *
7  * Libav is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * Libav is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with Libav; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "libavutil/imgutils.h"
29 #include "internal.h"
30 #include "dsputil.h"
31 #include "avcodec.h"
32 #include "mpegvideo.h"
33 #include "h264.h"
34 #include "h264data.h"
35 #include "h264_mvpred.h"
36 #include "golomb.h"
37 #include "mathops.h"
38 #include "rectangle.h"
39 #include "thread.h"
40 #include "vdpau_internal.h"
41 #include "libavutil/avassert.h"
42
43 #include "cabac.h"
44
45 //#undef NDEBUG
46 #include <assert.h>
47
48 static const uint8_t rem6[QP_MAX_NUM+1]={
49 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
50 };
51
52 static const uint8_t div6[QP_MAX_NUM+1]={
53 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9,10,10,10,10,
54 };
55
56 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
57     PIX_FMT_DXVA2_VLD,
58     PIX_FMT_VAAPI_VLD,
59     PIX_FMT_YUVJ420P,
60     PIX_FMT_NONE
61 };
62
63 /**
64  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
65  */
66 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
67     MpegEncContext * const s = &h->s;
68     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
69     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
70     int i;
71
72     if(!(h->top_samples_available&0x8000)){
73         for(i=0; i<4; i++){
74             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
75             if(status<0){
76                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
77                 return -1;
78             } else if(status){
79                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
80             }
81         }
82     }
83
84     if((h->left_samples_available&0x8888)!=0x8888){
85         static const int mask[4]={0x8000,0x2000,0x80,0x20};
86         for(i=0; i<4; i++){
87             if(!(h->left_samples_available&mask[i])){
88                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
89                 if(status<0){
90                     av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
91                     return -1;
92                 } else if(status){
93                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
94                 }
95             }
96         }
97     }
98
99     return 0;
100 } //FIXME cleanup like ff_h264_check_intra_pred_mode
101
102 /**
103  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
104  */
105 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
106     MpegEncContext * const s = &h->s;
107     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
108     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
109
110     if(mode > 6U) {
111         av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
112         return -1;
113     }
114
115     if(!(h->top_samples_available&0x8000)){
116         mode= top[ mode ];
117         if(mode<0){
118             av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
119             return -1;
120         }
121     }
122
123     if((h->left_samples_available&0x8080) != 0x8080){
124         mode= left[ mode ];
125         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
126             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
127         }
128         if(mode<0){
129             av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
130             return -1;
131         }
132     }
133
134     return mode;
135 }
136
137 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
138     int i, si, di;
139     uint8_t *dst;
140     int bufidx;
141
142 //    src[0]&0x80;                //forbidden bit
143     h->nal_ref_idc= src[0]>>5;
144     h->nal_unit_type= src[0]&0x1F;
145
146     src++; length--;
147
148 #if HAVE_FAST_UNALIGNED
149 # if HAVE_FAST_64BIT
150 #   define RS 7
151     for(i=0; i+1<length; i+=9){
152         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
153 # else
154 #   define RS 3
155     for(i=0; i+1<length; i+=5){
156         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
157 # endif
158             continue;
159         if(i>0 && !src[i]) i--;
160         while(src[i]) i++;
161 #else
162 #   define RS 0
163     for(i=0; i+1<length; i+=2){
164         if(src[i]) continue;
165         if(i>0 && src[i-1]==0) i--;
166 #endif
167         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
168             if(src[i+2]!=3){
169                 /* startcode, so we must be past the end */
170                 length=i;
171             }
172             break;
173         }
174         i-= RS;
175     }
176
177     if(i>=length-1){ //no escaped 0
178         *dst_length= length;
179         *consumed= length+1; //+1 for the header
180         return src;
181     }
182
183     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
184     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
185     dst= h->rbsp_buffer[bufidx];
186
187     if (dst == NULL){
188         return NULL;
189     }
190
191 //printf("decoding esc\n");
192     memcpy(dst, src, i);
193     si=di=i;
194     while(si+2<length){
195         //remove escapes (very rare 1:2^22)
196         if(src[si+2]>3){
197             dst[di++]= src[si++];
198             dst[di++]= src[si++];
199         }else if(src[si]==0 && src[si+1]==0){
200             if(src[si+2]==3){ //escape
201                 dst[di++]= 0;
202                 dst[di++]= 0;
203                 si+=3;
204                 continue;
205             }else //next start code
206                 goto nsc;
207         }
208
209         dst[di++]= src[si++];
210     }
211     while(si<length)
212         dst[di++]= src[si++];
213 nsc:
214
215     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
216
217     *dst_length= di;
218     *consumed= si + 1;//+1 for the header
219 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
220     return dst;
221 }
222
223 /**
224  * Identify the exact end of the bitstream
225  * @return the length of the trailing, or 0 if damaged
226  */
227 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
228     int v= *src;
229     int r;
230
231     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
232
233     for(r=1; r<9; r++){
234         if(v&1) return r;
235         v>>=1;
236     }
237     return 0;
238 }
239
240 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
241                                  int y_offset, int list){
242     int raw_my= h->mv_cache[list][ scan8[n] ][1];
243     int filter_height= (raw_my&3) ? 2 : 0;
244     int full_my= (raw_my>>2) + y_offset;
245     int top = full_my - filter_height, bottom = full_my + height + filter_height;
246
247     return FFMAX(abs(top), bottom);
248 }
249
250 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
251                                int y_offset, int list0, int list1, int *nrefs){
252     MpegEncContext * const s = &h->s;
253     int my;
254
255     y_offset += 16*(s->mb_y >> MB_FIELD);
256
257     if(list0){
258         int ref_n = h->ref_cache[0][ scan8[n] ];
259         Picture *ref= &h->ref_list[0][ref_n];
260
261         // Error resilience puts the current picture in the ref list.
262         // Don't try to wait on these as it will cause a deadlock.
263         // Fields can wait on each other, though.
264         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
265            (ref->f.reference & 3) != s->picture_structure) {
266             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
267             if (refs[0][ref_n] < 0) nrefs[0] += 1;
268             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
269         }
270     }
271
272     if(list1){
273         int ref_n = h->ref_cache[1][ scan8[n] ];
274         Picture *ref= &h->ref_list[1][ref_n];
275
276         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
277            (ref->f.reference & 3) != s->picture_structure) {
278             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
279             if (refs[1][ref_n] < 0) nrefs[1] += 1;
280             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
281         }
282     }
283 }
284
285 /**
286  * Wait until all reference frames are available for MC operations.
287  *
288  * @param h the H264 context
289  */
290 static void await_references(H264Context *h){
291     MpegEncContext * const s = &h->s;
292     const int mb_xy= h->mb_xy;
293     const int mb_type = s->current_picture.f.mb_type[mb_xy];
294     int refs[2][48];
295     int nrefs[2] = {0};
296     int ref, list;
297
298     memset(refs, -1, sizeof(refs));
299
300     if(IS_16X16(mb_type)){
301         get_lowest_part_y(h, refs, 0, 16, 0,
302                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
303     }else if(IS_16X8(mb_type)){
304         get_lowest_part_y(h, refs, 0, 8, 0,
305                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
306         get_lowest_part_y(h, refs, 8, 8, 8,
307                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
308     }else if(IS_8X16(mb_type)){
309         get_lowest_part_y(h, refs, 0, 16, 0,
310                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
311         get_lowest_part_y(h, refs, 4, 16, 0,
312                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
313     }else{
314         int i;
315
316         assert(IS_8X8(mb_type));
317
318         for(i=0; i<4; i++){
319             const int sub_mb_type= h->sub_mb_type[i];
320             const int n= 4*i;
321             int y_offset= (i&2)<<2;
322
323             if(IS_SUB_8X8(sub_mb_type)){
324                 get_lowest_part_y(h, refs, n  , 8, y_offset,
325                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
326             }else if(IS_SUB_8X4(sub_mb_type)){
327                 get_lowest_part_y(h, refs, n  , 4, y_offset,
328                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
329                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
330                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
331             }else if(IS_SUB_4X8(sub_mb_type)){
332                 get_lowest_part_y(h, refs, n  , 8, y_offset,
333                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
334                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
335                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
336             }else{
337                 int j;
338                 assert(IS_SUB_4X4(sub_mb_type));
339                 for(j=0; j<4; j++){
340                     int sub_y_offset= y_offset + 2*(j&2);
341                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
342                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
343                 }
344             }
345         }
346     }
347
348     for(list=h->list_count-1; list>=0; list--){
349         for(ref=0; ref<48 && nrefs[list]; ref++){
350             int row = refs[list][ref];
351             if(row >= 0){
352                 Picture *ref_pic = &h->ref_list[list][ref];
353                 int ref_field = ref_pic->f.reference - 1;
354                 int ref_field_picture = ref_pic->field_picture;
355                 int pic_height = 16*s->mb_height >> ref_field_picture;
356
357                 row <<= MB_MBAFF;
358                 nrefs[list]--;
359
360                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
361                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
362                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
363                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
364                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
365                 }else if(FIELD_PICTURE){
366                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
367                 }else{
368                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
369                 }
370             }
371         }
372     }
373 }
374
375 #if 0
376 /**
377  * DCT transforms the 16 dc values.
378  * @param qp quantization parameter ??? FIXME
379  */
380 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
381 //    const int qmul= dequant_coeff[qp][0];
382     int i;
383     int temp[16]; //FIXME check if this is a good idea
384     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
385     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
386
387     for(i=0; i<4; i++){
388         const int offset= y_offset[i];
389         const int z0= block[offset+stride*0] + block[offset+stride*4];
390         const int z1= block[offset+stride*0] - block[offset+stride*4];
391         const int z2= block[offset+stride*1] - block[offset+stride*5];
392         const int z3= block[offset+stride*1] + block[offset+stride*5];
393
394         temp[4*i+0]= z0+z3;
395         temp[4*i+1]= z1+z2;
396         temp[4*i+2]= z1-z2;
397         temp[4*i+3]= z0-z3;
398     }
399
400     for(i=0; i<4; i++){
401         const int offset= x_offset[i];
402         const int z0= temp[4*0+i] + temp[4*2+i];
403         const int z1= temp[4*0+i] - temp[4*2+i];
404         const int z2= temp[4*1+i] - temp[4*3+i];
405         const int z3= temp[4*1+i] + temp[4*3+i];
406
407         block[stride*0 +offset]= (z0 + z3)>>1;
408         block[stride*2 +offset]= (z1 + z2)>>1;
409         block[stride*8 +offset]= (z1 - z2)>>1;
410         block[stride*10+offset]= (z0 - z3)>>1;
411     }
412 }
413 #endif
414
415 #undef xStride
416 #undef stride
417
418 #if 0
419 static void chroma_dc_dct_c(DCTELEM *block){
420     const int stride= 16*2;
421     const int xStride= 16;
422     int a,b,c,d,e;
423
424     a= block[stride*0 + xStride*0];
425     b= block[stride*0 + xStride*1];
426     c= block[stride*1 + xStride*0];
427     d= block[stride*1 + xStride*1];
428
429     e= a-b;
430     a= a+b;
431     b= c-d;
432     c= c+d;
433
434     block[stride*0 + xStride*0]= (a+c);
435     block[stride*0 + xStride*1]= (e+b);
436     block[stride*1 + xStride*0]= (a-c);
437     block[stride*1 + xStride*1]= (e-b);
438 }
439 #endif
440
441 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
442                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
443                            int src_x_offset, int src_y_offset,
444                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
445                            int pixel_shift, int chroma444){
446     MpegEncContext * const s = &h->s;
447     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
448     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
449     const int luma_xy= (mx&3) + ((my&3)<<2);
450     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
451     uint8_t * src_y = pic->f.data[0] + offset;
452     uint8_t * src_cb, * src_cr;
453     int extra_width= h->emu_edge_width;
454     int extra_height= h->emu_edge_height;
455     int emu=0;
456     const int full_mx= mx>>2;
457     const int full_my= my>>2;
458     const int pic_width  = 16*s->mb_width;
459     const int pic_height = 16*s->mb_height >> MB_FIELD;
460
461     if(mx&7) extra_width -= 3;
462     if(my&7) extra_height -= 3;
463
464     if(   full_mx < 0-extra_width
465        || full_my < 0-extra_height
466        || full_mx + 16/*FIXME*/ > pic_width + extra_width
467        || full_my + 16/*FIXME*/ > pic_height + extra_height){
468         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_y - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize, 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
469             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
470         emu=1;
471     }
472
473     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
474     if(!square){
475         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
476     }
477
478     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
479
480     if(chroma444){
481         src_cb = pic->f.data[1] + offset;
482         if(emu){
483             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
484                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
485             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
486         }
487         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
488         if(!square){
489             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
490         }
491
492         src_cr = pic->f.data[2] + offset;
493         if(emu){
494             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
495                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
496             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
497         }
498         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
499         if(!square){
500             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
501         }
502         return;
503     }
504
505     if(MB_FIELD){
506         // chroma offset when predicting from a field of opposite parity
507         my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
508         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
509     }
510     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
511     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
512
513     if(emu){
514         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
515             src_cb= s->edge_emu_buffer;
516     }
517     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
518
519     if(emu){
520         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
521             src_cr= s->edge_emu_buffer;
522     }
523     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
524 }
525
526 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
527                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
528                            int x_offset, int y_offset,
529                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
530                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
531                            int list0, int list1, int pixel_shift, int chroma444){
532     MpegEncContext * const s = &h->s;
533     qpel_mc_func *qpix_op=  qpix_put;
534     h264_chroma_mc_func chroma_op= chroma_put;
535
536     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
537     if(chroma444){
538         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
539         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
540     }else{
541         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
542         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
543     }
544     x_offset += 8*s->mb_x;
545     y_offset += 8*(s->mb_y >> MB_FIELD);
546
547     if(list0){
548         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
549         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
550                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
551                            qpix_op, chroma_op, pixel_shift, chroma444);
552
553         qpix_op=  qpix_avg;
554         chroma_op= chroma_avg;
555     }
556
557     if(list1){
558         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
559         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
560                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
561                            qpix_op, chroma_op, pixel_shift, chroma444);
562     }
563 }
564
565 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
566                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
567                            int x_offset, int y_offset,
568                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
569                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
570                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
571                            int list0, int list1, int pixel_shift, int chroma444){
572     MpegEncContext * const s = &h->s;
573
574     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
575     if(chroma444){
576         chroma_weight_avg = luma_weight_avg;
577         chroma_weight_op = luma_weight_op;
578         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
579         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
580     }else{
581         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
582         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
583     }
584     x_offset += 8*s->mb_x;
585     y_offset += 8*(s->mb_y >> MB_FIELD);
586
587     if(list0 && list1){
588         /* don't optimize for luma-only case, since B-frames usually
589          * use implicit weights => chroma too. */
590         uint8_t *tmp_cb = s->obmc_scratchpad;
591         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
592         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
593         int refn0 = h->ref_cache[0][ scan8[n] ];
594         int refn1 = h->ref_cache[1][ scan8[n] ];
595
596         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
597                     dest_y, dest_cb, dest_cr,
598                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
599         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
600                     tmp_y, tmp_cb, tmp_cr,
601                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
602
603         if(h->use_weight == 2){
604             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
605             int weight1 = 64 - weight0;
606             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
607             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
608             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
609         }else{
610             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
611                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
612                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
613             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
614                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
615                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
616             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
617                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
618                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
619         }
620     }else{
621         int list = list1 ? 1 : 0;
622         int refn = h->ref_cache[list][ scan8[n] ];
623         Picture *ref= &h->ref_list[list][refn];
624         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
625                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
626                     qpix_put, chroma_put, pixel_shift, chroma444);
627
628         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
629                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
630         if(h->use_weight_chroma){
631             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
632                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
633             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
634                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
635         }
636     }
637 }
638
639 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
640                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
641                            int x_offset, int y_offset,
642                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
643                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
644                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
645                            int list0, int list1, int pixel_shift, int chroma444){
646     if((h->use_weight==2 && list0 && list1
647         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
648        || h->use_weight==1)
649         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
650                          x_offset, y_offset, qpix_put, chroma_put,
651                          weight_op[0], weight_op[3], weight_avg[0],
652                          weight_avg[3], list0, list1, pixel_shift, chroma444);
653     else
654         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
655                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
656                     chroma_avg, list0, list1, pixel_shift, chroma444);
657 }
658
659 static inline void prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma444){
660     /* fetch pixels for estimated mv 4 macroblocks ahead
661      * optimized for 64byte cache lines */
662     MpegEncContext * const s = &h->s;
663     const int refn = h->ref_cache[list][scan8[0]];
664     if(refn >= 0){
665         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
666         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
667         uint8_t **src = h->ref_list[list][refn].f.data;
668         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
669         s->dsp.prefetch(src[0]+off, s->linesize, 4);
670         if(chroma444){
671             s->dsp.prefetch(src[1]+off, s->linesize, 4);
672             s->dsp.prefetch(src[2]+off, s->linesize, 4);
673         }else{
674             off= ((mx>>1) << pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + (64 << pixel_shift);
675             s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
676         }
677     }
678 }
679
680 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
681                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
682                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
683                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
684                       int pixel_shift, int chroma444){
685     MpegEncContext * const s = &h->s;
686     const int mb_xy= h->mb_xy;
687     const int mb_type = s->current_picture.f.mb_type[mb_xy];
688
689     assert(IS_INTER(mb_type));
690
691     if(HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
692         await_references(h);
693     prefetch_motion(h, 0, pixel_shift, chroma444);
694
695     if(IS_16X16(mb_type)){
696         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
697                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
698                 weight_op, weight_avg,
699                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
700                 pixel_shift, chroma444);
701     }else if(IS_16X8(mb_type)){
702         mc_part(h, 0, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
703                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
704                 &weight_op[1], &weight_avg[1],
705                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
706                 pixel_shift, chroma444);
707         mc_part(h, 8, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
708                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
709                 &weight_op[1], &weight_avg[1],
710                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
711                 pixel_shift, chroma444);
712     }else if(IS_8X16(mb_type)){
713         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
714                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
715                 &weight_op[2], &weight_avg[2],
716                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
717                 pixel_shift, chroma444);
718         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
719                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
720                 &weight_op[2], &weight_avg[2],
721                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
722                 pixel_shift, chroma444);
723     }else{
724         int i;
725
726         assert(IS_8X8(mb_type));
727
728         for(i=0; i<4; i++){
729             const int sub_mb_type= h->sub_mb_type[i];
730             const int n= 4*i;
731             int x_offset= (i&1)<<2;
732             int y_offset= (i&2)<<1;
733
734             if(IS_SUB_8X8(sub_mb_type)){
735                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
736                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
737                     &weight_op[3], &weight_avg[3],
738                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
739                     pixel_shift, chroma444);
740             }else if(IS_SUB_8X4(sub_mb_type)){
741                 mc_part(h, n  , 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
742                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
743                     &weight_op[4], &weight_avg[4],
744                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
745                     pixel_shift, chroma444);
746                 mc_part(h, n+2, 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
747                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
748                     &weight_op[4], &weight_avg[4],
749                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
750                     pixel_shift, chroma444);
751             }else if(IS_SUB_4X8(sub_mb_type)){
752                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
753                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
754                     &weight_op[5], &weight_avg[5],
755                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
756                     pixel_shift, chroma444);
757                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
758                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
759                     &weight_op[5], &weight_avg[5],
760                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
761                     pixel_shift, chroma444);
762             }else{
763                 int j;
764                 assert(IS_SUB_4X4(sub_mb_type));
765                 for(j=0; j<4; j++){
766                     int sub_x_offset= x_offset + 2*(j&1);
767                     int sub_y_offset= y_offset +   (j&2);
768                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
769                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
770                         &weight_op[6], &weight_avg[6],
771                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
772                         pixel_shift, chroma444);
773                 }
774             }
775         }
776     }
777
778     prefetch_motion(h, 1, pixel_shift, chroma444);
779 }
780
781 static void free_tables(H264Context *h, int free_rbsp){
782     int i;
783     H264Context *hx;
784
785     av_freep(&h->intra4x4_pred_mode);
786     av_freep(&h->chroma_pred_mode_table);
787     av_freep(&h->cbp_table);
788     av_freep(&h->mvd_table[0]);
789     av_freep(&h->mvd_table[1]);
790     av_freep(&h->direct_table);
791     av_freep(&h->non_zero_count);
792     av_freep(&h->slice_table_base);
793     h->slice_table= NULL;
794     av_freep(&h->list_counts);
795
796     av_freep(&h->mb2b_xy);
797     av_freep(&h->mb2br_xy);
798
799     for(i = 0; i < MAX_THREADS; i++) {
800         hx = h->thread_context[i];
801         if(!hx) continue;
802         av_freep(&hx->top_borders[1]);
803         av_freep(&hx->top_borders[0]);
804         av_freep(&hx->s.obmc_scratchpad);
805         if (free_rbsp){
806             av_freep(&hx->rbsp_buffer[1]);
807             av_freep(&hx->rbsp_buffer[0]);
808             hx->rbsp_buffer_size[0] = 0;
809             hx->rbsp_buffer_size[1] = 0;
810         }
811         if (i) av_freep(&h->thread_context[i]);
812     }
813 }
814
815 static void init_dequant8_coeff_table(H264Context *h){
816     int i,j,q,x;
817     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
818
819     for(i=0; i<6; i++ ){
820         h->dequant8_coeff[i] = h->dequant8_buffer[i];
821         for(j=0; j<i; j++){
822             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
823                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
824                 break;
825             }
826         }
827         if(j<i)
828             continue;
829
830         for(q=0; q<max_qp+1; q++){
831             int shift = div6[q];
832             int idx = rem6[q];
833             for(x=0; x<64; x++)
834                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
835                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
836                     h->pps.scaling_matrix8[i][x]) << shift;
837         }
838     }
839 }
840
841 static void init_dequant4_coeff_table(H264Context *h){
842     int i,j,q,x;
843     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
844     for(i=0; i<6; i++ ){
845         h->dequant4_coeff[i] = h->dequant4_buffer[i];
846         for(j=0; j<i; j++){
847             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
848                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
849                 break;
850             }
851         }
852         if(j<i)
853             continue;
854
855         for(q=0; q<max_qp+1; q++){
856             int shift = div6[q] + 2;
857             int idx = rem6[q];
858             for(x=0; x<16; x++)
859                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
860                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
861                     h->pps.scaling_matrix4[i][x]) << shift;
862         }
863     }
864 }
865
866 static void init_dequant_tables(H264Context *h){
867     int i,x;
868     init_dequant4_coeff_table(h);
869     if(h->pps.transform_8x8_mode)
870         init_dequant8_coeff_table(h);
871     if(h->sps.transform_bypass){
872         for(i=0; i<6; i++)
873             for(x=0; x<16; x++)
874                 h->dequant4_coeff[i][0][x] = 1<<6;
875         if(h->pps.transform_8x8_mode)
876             for(i=0; i<6; i++)
877                 for(x=0; x<64; x++)
878                     h->dequant8_coeff[i][0][x] = 1<<6;
879     }
880 }
881
882
883 int ff_h264_alloc_tables(H264Context *h){
884     MpegEncContext * const s = &h->s;
885     const int big_mb_num= s->mb_stride * (s->mb_height+1);
886     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
887     int x,y;
888
889     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
890
891     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
892     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
893     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
894
895     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
896     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
897     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
898     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
899     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
900
901     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
902     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
903
904     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
905     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
906     for(y=0; y<s->mb_height; y++){
907         for(x=0; x<s->mb_width; x++){
908             const int mb_xy= x + y*s->mb_stride;
909             const int b_xy = 4*x + 4*y*h->b_stride;
910
911             h->mb2b_xy [mb_xy]= b_xy;
912             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
913         }
914     }
915
916     s->obmc_scratchpad = NULL;
917
918     if(!h->dequant4_coeff[0])
919         init_dequant_tables(h);
920
921     return 0;
922 fail:
923     free_tables(h, 1);
924     return -1;
925 }
926
927 /**
928  * Mimic alloc_tables(), but for every context thread.
929  */
930 static void clone_tables(H264Context *dst, H264Context *src, int i){
931     MpegEncContext * const s = &src->s;
932     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
933     dst->non_zero_count           = src->non_zero_count;
934     dst->slice_table              = src->slice_table;
935     dst->cbp_table                = src->cbp_table;
936     dst->mb2b_xy                  = src->mb2b_xy;
937     dst->mb2br_xy                 = src->mb2br_xy;
938     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
939     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
940     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
941     dst->direct_table             = src->direct_table;
942     dst->list_counts              = src->list_counts;
943
944     dst->s.obmc_scratchpad = NULL;
945     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma);
946 }
947
948 /**
949  * Init context
950  * Allocate buffers which are not shared amongst multiple threads.
951  */
952 static int context_init(H264Context *h){
953     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
954     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
955
956     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
957     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
958
959     return 0;
960 fail:
961     return -1; // free_tables will clean up for us
962 }
963
964 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
965
966 static av_cold void common_init(H264Context *h){
967     MpegEncContext * const s = &h->s;
968
969     s->width = s->avctx->width;
970     s->height = s->avctx->height;
971     s->codec_id= s->avctx->codec->id;
972
973     ff_h264dsp_init(&h->h264dsp, 8);
974     ff_h264_pred_init(&h->hpc, s->codec_id, 8);
975
976     h->dequant_coeff_pps= -1;
977     s->unrestricted_mv=1;
978     s->decode=1; //FIXME
979
980     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
981
982     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
983     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
984 }
985
986 int ff_h264_decode_extradata(H264Context *h)
987 {
988     AVCodecContext *avctx = h->s.avctx;
989
990     if(avctx->extradata[0] == 1){
991         int i, cnt, nalsize;
992         unsigned char *p = avctx->extradata;
993
994         h->is_avc = 1;
995
996         if(avctx->extradata_size < 7) {
997             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
998             return -1;
999         }
1000         /* sps and pps in the avcC always have length coded with 2 bytes,
1001            so put a fake nal_length_size = 2 while parsing them */
1002         h->nal_length_size = 2;
1003         // Decode sps from avcC
1004         cnt = *(p+5) & 0x1f; // Number of sps
1005         p += 6;
1006         for (i = 0; i < cnt; i++) {
1007             nalsize = AV_RB16(p) + 2;
1008             if (p - avctx->extradata + nalsize > avctx->extradata_size)
1009                 return -1;
1010             if(decode_nal_units(h, p, nalsize) < 0) {
1011                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1012                 return -1;
1013             }
1014             p += nalsize;
1015         }
1016         // Decode pps from avcC
1017         cnt = *(p++); // Number of pps
1018         for (i = 0; i < cnt; i++) {
1019             nalsize = AV_RB16(p) + 2;
1020             if (p - avctx->extradata + nalsize > avctx->extradata_size)
1021                 return -1;
1022             if (decode_nal_units(h, p, nalsize) < 0) {
1023                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1024                 return -1;
1025             }
1026             p += nalsize;
1027         }
1028         // Now store right nal length size, that will be use to parse all other nals
1029         h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1030     } else {
1031         h->is_avc = 0;
1032         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1033             return -1;
1034     }
1035     return 0;
1036 }
1037
1038 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1039     H264Context *h= avctx->priv_data;
1040     MpegEncContext * const s = &h->s;
1041
1042     MPV_decode_defaults(s);
1043
1044     s->avctx = avctx;
1045     common_init(h);
1046
1047     s->out_format = FMT_H264;
1048     s->workaround_bugs= avctx->workaround_bugs;
1049
1050     // set defaults
1051 //    s->decode_mb= ff_h263_decode_mb;
1052     s->quarter_sample = 1;
1053     if(!avctx->has_b_frames)
1054     s->low_delay= 1;
1055
1056     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1057
1058     ff_h264_decode_init_vlc();
1059
1060     h->pixel_shift = 0;
1061     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1062
1063     h->thread_context[0] = h;
1064     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1065     h->prev_poc_msb= 1<<16;
1066     h->x264_build = -1;
1067     ff_h264_reset_sei(h);
1068     if(avctx->codec_id == CODEC_ID_H264){
1069         if(avctx->ticks_per_frame == 1){
1070             s->avctx->time_base.den *=2;
1071         }
1072         avctx->ticks_per_frame = 2;
1073     }
1074
1075     if(avctx->extradata_size > 0 && avctx->extradata &&
1076         ff_h264_decode_extradata(h))
1077         return -1;
1078
1079     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1080         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1081         s->low_delay = 0;
1082     }
1083
1084     return 0;
1085 }
1086
1087 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1088 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1089 {
1090     int i;
1091
1092     for (i=0; i<count; i++){
1093         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1094                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1095                 !from[i]));
1096         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1097     }
1098 }
1099
1100 static void copy_parameter_set(void **to, void **from, int count, int size)
1101 {
1102     int i;
1103
1104     for (i=0; i<count; i++){
1105         if (to[i] && !from[i]) av_freep(&to[i]);
1106         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1107
1108         if (from[i]) memcpy(to[i], from[i], size);
1109     }
1110 }
1111
1112 static int decode_init_thread_copy(AVCodecContext *avctx){
1113     H264Context *h= avctx->priv_data;
1114
1115     if (!avctx->is_copy) return 0;
1116     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1117     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1118
1119     return 0;
1120 }
1121
1122 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1123 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1124     H264Context *h= dst->priv_data, *h1= src->priv_data;
1125     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1126     int inited = s->context_initialized, err;
1127     int i;
1128
1129     if(dst == src || !s1->context_initialized) return 0;
1130
1131     err = ff_mpeg_update_thread_context(dst, src);
1132     if(err) return err;
1133
1134     //FIXME handle width/height changing
1135     if(!inited){
1136         for(i = 0; i < MAX_SPS_COUNT; i++)
1137             av_freep(h->sps_buffers + i);
1138
1139         for(i = 0; i < MAX_PPS_COUNT; i++)
1140             av_freep(h->pps_buffers + i);
1141
1142         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1143         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1144         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1145         if (ff_h264_alloc_tables(h) < 0) {
1146             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1147             return AVERROR(ENOMEM);
1148         }
1149         context_init(h);
1150
1151         for(i=0; i<2; i++){
1152             h->rbsp_buffer[i] = NULL;
1153             h->rbsp_buffer_size[i] = 0;
1154         }
1155
1156         h->thread_context[0] = h;
1157
1158         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1159         // so this has to be allocated here
1160         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1161
1162         s->dsp.clear_blocks(h->mb);
1163         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1164     }
1165
1166     //extradata/NAL handling
1167     h->is_avc          = h1->is_avc;
1168
1169     //SPS/PPS
1170     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1171     h->sps             = h1->sps;
1172     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1173     h->pps             = h1->pps;
1174
1175     //Dequantization matrices
1176     //FIXME these are big - can they be only copied when PPS changes?
1177     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1178
1179     for(i=0; i<6; i++)
1180         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1181
1182     for(i=0; i<6; i++)
1183         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1184
1185     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1186
1187     //POC timing
1188     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1189
1190     //reference lists
1191     copy_fields(h, h1, ref_count, list_count);
1192     copy_fields(h, h1, ref_list,  intra_gb);
1193     copy_fields(h, h1, short_ref, cabac_init_idc);
1194
1195     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1196     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1197     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1198
1199     h->last_slice_type = h1->last_slice_type;
1200
1201     if(!s->current_picture_ptr) return 0;
1202
1203     if(!s->dropable) {
1204         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1205         h->prev_poc_msb     = h->poc_msb;
1206         h->prev_poc_lsb     = h->poc_lsb;
1207     }
1208     h->prev_frame_num_offset= h->frame_num_offset;
1209     h->prev_frame_num       = h->frame_num;
1210     h->outputed_poc         = h->next_outputed_poc;
1211
1212     return err;
1213 }
1214
1215 int ff_h264_frame_start(H264Context *h){
1216     MpegEncContext * const s = &h->s;
1217     int i;
1218     const int pixel_shift = h->pixel_shift;
1219     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1220
1221     if(MPV_frame_start(s, s->avctx) < 0)
1222         return -1;
1223     ff_er_frame_start(s);
1224     /*
1225      * MPV_frame_start uses pict_type to derive key_frame.
1226      * This is incorrect for H.264; IDR markings must be used.
1227      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1228      * See decode_nal_units().
1229      */
1230     s->current_picture_ptr->f.key_frame = 0;
1231     s->current_picture_ptr->mmco_reset= 0;
1232
1233     assert(s->linesize && s->uvlinesize);
1234
1235     for(i=0; i<16; i++){
1236         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1237         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1238     }
1239     for(i=0; i<16; i++){
1240         h->block_offset[16+i]=
1241         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1242         h->block_offset[48+16+i]=
1243         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1244     }
1245
1246     /* can't be in alloc_tables because linesize isn't known there.
1247      * FIXME: redo bipred weight to not require extra buffer? */
1248     for(i = 0; i < thread_count; i++)
1249         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1250             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1251
1252     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1253     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1254
1255 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1256
1257     // We mark the current picture as non-reference after allocating it, so
1258     // that if we break out due to an error it can be released automatically
1259     // in the next MPV_frame_start().
1260     // SVQ3 as well as most other codecs have only last/next/current and thus
1261     // get released even with set reference, besides SVQ3 and others do not
1262     // mark frames as reference later "naturally".
1263     if(s->codec_id != CODEC_ID_SVQ3)
1264         s->current_picture_ptr->f.reference = 0;
1265
1266     s->current_picture_ptr->field_poc[0]=
1267     s->current_picture_ptr->field_poc[1]= INT_MAX;
1268
1269     h->next_output_pic = NULL;
1270
1271     assert(s->current_picture_ptr->long_ref==0);
1272
1273     return 0;
1274 }
1275
1276 /**
1277   * Run setup operations that must be run after slice header decoding.
1278   * This includes finding the next displayed frame.
1279   *
1280   * @param h h264 master context
1281   * @param setup_finished enough NALs have been read that we can call
1282   * ff_thread_finish_setup()
1283   */
1284 static void decode_postinit(H264Context *h, int setup_finished){
1285     MpegEncContext * const s = &h->s;
1286     Picture *out = s->current_picture_ptr;
1287     Picture *cur = s->current_picture_ptr;
1288     int i, pics, out_of_order, out_idx;
1289
1290     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1291     s->current_picture_ptr->f.pict_type   = s->pict_type;
1292
1293     if (h->next_output_pic) return;
1294
1295     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1296         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1297         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1298         //to find this yet, so we assume the worst for now.
1299         //if (setup_finished)
1300         //    ff_thread_finish_setup(s->avctx);
1301         return;
1302     }
1303
1304     cur->f.interlaced_frame = 0;
1305     cur->f.repeat_pict      = 0;
1306
1307     /* Signal interlacing information externally. */
1308     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1309
1310     if(h->sps.pic_struct_present_flag){
1311         switch (h->sei_pic_struct)
1312         {
1313         case SEI_PIC_STRUCT_FRAME:
1314             break;
1315         case SEI_PIC_STRUCT_TOP_FIELD:
1316         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1317             cur->f.interlaced_frame = 1;
1318             break;
1319         case SEI_PIC_STRUCT_TOP_BOTTOM:
1320         case SEI_PIC_STRUCT_BOTTOM_TOP:
1321             if (FIELD_OR_MBAFF_PICTURE)
1322                 cur->f.interlaced_frame = 1;
1323             else
1324                 // try to flag soft telecine progressive
1325                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1326             break;
1327         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1328         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1329             // Signal the possibility of telecined film externally (pic_struct 5,6)
1330             // From these hints, let the applications decide if they apply deinterlacing.
1331             cur->f.repeat_pict = 1;
1332             break;
1333         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1334             // Force progressive here, as doubling interlaced frame is a bad idea.
1335             cur->f.repeat_pict = 2;
1336             break;
1337         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1338             cur->f.repeat_pict = 4;
1339             break;
1340         }
1341
1342         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1343             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1344     }else{
1345         /* Derive interlacing flag from used decoding process. */
1346         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1347     }
1348     h->prev_interlaced_frame = cur->f.interlaced_frame;
1349
1350     if (cur->field_poc[0] != cur->field_poc[1]){
1351         /* Derive top_field_first from field pocs. */
1352         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1353     }else{
1354         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1355             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1356             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1357               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1358                 cur->f.top_field_first = 1;
1359             else
1360                 cur->f.top_field_first = 0;
1361         }else{
1362             /* Most likely progressive */
1363             cur->f.top_field_first = 0;
1364         }
1365     }
1366
1367     //FIXME do something with unavailable reference frames
1368
1369     /* Sort B-frames into display order */
1370
1371     if(h->sps.bitstream_restriction_flag
1372        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1373         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1374         s->low_delay = 0;
1375     }
1376
1377     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1378        && !h->sps.bitstream_restriction_flag){
1379         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1380         s->low_delay= 0;
1381     }
1382
1383     pics = 0;
1384     while(h->delayed_pic[pics]) pics++;
1385
1386     assert(pics <= MAX_DELAYED_PIC_COUNT);
1387
1388     h->delayed_pic[pics++] = cur;
1389     if (cur->f.reference == 0)
1390         cur->f.reference = DELAYED_PIC_REF;
1391
1392     out = h->delayed_pic[0];
1393     out_idx = 0;
1394     for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1395         if(h->delayed_pic[i]->poc < out->poc){
1396             out = h->delayed_pic[i];
1397             out_idx = i;
1398         }
1399     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1400         h->next_outputed_poc= INT_MIN;
1401     out_of_order = out->poc < h->next_outputed_poc;
1402
1403     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1404         { }
1405     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1406        || (s->low_delay &&
1407         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1408          || cur->f.pict_type == AV_PICTURE_TYPE_B)))
1409     {
1410         s->low_delay = 0;
1411         s->avctx->has_b_frames++;
1412     }
1413
1414     if(out_of_order || pics > s->avctx->has_b_frames){
1415         out->f.reference &= ~DELAYED_PIC_REF;
1416         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1417                          // or else the first thread can release the picture and reuse it unsafely
1418         for(i=out_idx; h->delayed_pic[i]; i++)
1419             h->delayed_pic[i] = h->delayed_pic[i+1];
1420     }
1421     if(!out_of_order && pics > s->avctx->has_b_frames){
1422         h->next_output_pic = out;
1423         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1424             h->next_outputed_poc = INT_MIN;
1425         } else
1426             h->next_outputed_poc = out->poc;
1427     }else{
1428         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1429     }
1430
1431     if (setup_finished)
1432         ff_thread_finish_setup(s->avctx);
1433 }
1434
1435 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int chroma444, int simple){
1436     MpegEncContext * const s = &h->s;
1437     uint8_t *top_border;
1438     int top_idx = 1;
1439     const int pixel_shift = h->pixel_shift;
1440
1441     src_y  -=   linesize;
1442     src_cb -= uvlinesize;
1443     src_cr -= uvlinesize;
1444
1445     if(!simple && FRAME_MBAFF){
1446         if(s->mb_y&1){
1447             if(!MB_MBAFF){
1448                 top_border = h->top_borders[0][s->mb_x];
1449                 AV_COPY128(top_border, src_y + 15*linesize);
1450                 if (pixel_shift)
1451                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1452                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1453                     if(chroma444){
1454                         if (pixel_shift){
1455                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1456                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1457                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1458                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1459                         } else {
1460                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1461                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1462                         }
1463                     } else {
1464                         if (pixel_shift) {
1465                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1466                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1467                         } else {
1468                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1469                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1470                         }
1471                     }
1472                 }
1473             }
1474         }else if(MB_MBAFF){
1475             top_idx = 0;
1476         }else
1477             return;
1478     }
1479
1480     top_border = h->top_borders[top_idx][s->mb_x];
1481     // There are two lines saved, the line above the the top macroblock of a pair,
1482     // and the line above the bottom macroblock
1483     AV_COPY128(top_border, src_y + 16*linesize);
1484     if (pixel_shift)
1485         AV_COPY128(top_border+16, src_y+16*linesize+16);
1486
1487     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1488         if(chroma444){
1489             if (pixel_shift){
1490                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1491                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1492                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1493                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1494             } else {
1495                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1496                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1497             }
1498         } else {
1499             if (pixel_shift) {
1500                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1501                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1502             } else {
1503                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1504                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1505             }
1506         }
1507     }
1508 }
1509
1510 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1511                                   uint8_t *src_cb, uint8_t *src_cr,
1512                                   int linesize, int uvlinesize,
1513                                   int xchg, int chroma444,
1514                                   int simple, int pixel_shift){
1515     MpegEncContext * const s = &h->s;
1516     int deblock_topleft;
1517     int deblock_top;
1518     int top_idx = 1;
1519     uint8_t *top_border_m1;
1520     uint8_t *top_border;
1521
1522     if(!simple && FRAME_MBAFF){
1523         if(s->mb_y&1){
1524             if(!MB_MBAFF)
1525                 return;
1526         }else{
1527             top_idx = MB_MBAFF ? 0 : 1;
1528         }
1529     }
1530
1531     if(h->deblocking_filter == 2) {
1532         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1533         deblock_top     = h->top_type;
1534     } else {
1535         deblock_topleft = (s->mb_x > 0);
1536         deblock_top     = (s->mb_y > !!MB_FIELD);
1537     }
1538
1539     src_y  -=   linesize + 1 + pixel_shift;
1540     src_cb -= uvlinesize + 1 + pixel_shift;
1541     src_cr -= uvlinesize + 1 + pixel_shift;
1542
1543     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1544     top_border    = h->top_borders[top_idx][s->mb_x];
1545
1546 #define XCHG(a,b,xchg)\
1547     if (pixel_shift) {\
1548         if (xchg) {\
1549             AV_SWAP64(b+0,a+0);\
1550             AV_SWAP64(b+8,a+8);\
1551         } else {\
1552             AV_COPY128(b,a); \
1553         }\
1554     } else \
1555 if (xchg) AV_SWAP64(b,a);\
1556 else      AV_COPY64(b,a);
1557
1558     if(deblock_top){
1559         if(deblock_topleft){
1560             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1561         }
1562         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1563         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1564         if(s->mb_x+1 < s->mb_width){
1565             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1566         }
1567     }
1568     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1569         if(chroma444){
1570             if(deblock_topleft){
1571                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1572                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1573             }
1574             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1575             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1576             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1577             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1578             if(s->mb_x+1 < s->mb_width){
1579                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1580                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1581             }
1582         } else {
1583             if(deblock_top){
1584                 if(deblock_topleft){
1585                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1586                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1587                 }
1588                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1589                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1590             }
1591         }
1592     }
1593 }
1594
1595 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1596     if (high_bit_depth) {
1597         return AV_RN32A(((int32_t*)mb) + index);
1598     } else
1599         return AV_RN16A(mb + index);
1600 }
1601
1602 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1603     if (high_bit_depth) {
1604         AV_WN32A(((int32_t*)mb) + index, value);
1605     } else
1606         AV_WN16A(mb + index, value);
1607 }
1608
1609 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1610                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1611 {
1612     MpegEncContext * const s = &h->s;
1613     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1614     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1615     int i;
1616     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1617     block_offset += 16*p;
1618     if(IS_INTRA4x4(mb_type)){
1619         if(simple || !s->encoding){
1620             if(IS_8x8DCT(mb_type)){
1621                 if(transform_bypass){
1622                     idct_dc_add =
1623                     idct_add    = s->dsp.add_pixels8;
1624                 }else{
1625                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1626                     idct_add    = h->h264dsp.h264_idct8_add;
1627                 }
1628                 for(i=0; i<16; i+=4){
1629                     uint8_t * const ptr= dest_y + block_offset[i];
1630                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1631                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1632                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1633                     }else{
1634                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1635                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1636                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1637                         if(nnz){
1638                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1639                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1640                             else
1641                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1642                         }
1643                     }
1644                 }
1645             }else{
1646                 if(transform_bypass){
1647                     idct_dc_add =
1648                     idct_add    = s->dsp.add_pixels4;
1649                 }else{
1650                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1651                     idct_add    = h->h264dsp.h264_idct_add;
1652                 }
1653                 for(i=0; i<16; i++){
1654                     uint8_t * const ptr= dest_y + block_offset[i];
1655                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1656
1657                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1658                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1659                     }else{
1660                         uint8_t *topright;
1661                         int nnz, tr;
1662                         uint64_t tr_high;
1663                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1664                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1665                             assert(s->mb_y || linesize <= block_offset[i]);
1666                             if(!topright_avail){
1667                                 if (pixel_shift) {
1668                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1669                                     topright= (uint8_t*) &tr_high;
1670                                 } else {
1671                                     tr= ptr[3 - linesize]*0x01010101u;
1672                                     topright= (uint8_t*) &tr;
1673                                 }
1674                             }else
1675                                 topright= ptr + (4 << pixel_shift) - linesize;
1676                         }else
1677                             topright= NULL;
1678
1679                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1680                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1681                         if(nnz){
1682                             if(is_h264){
1683                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1684                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1685                                 else
1686                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1687                             }else
1688                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1689                         }
1690                     }
1691                 }
1692             }
1693         }
1694     }else{
1695         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1696         if(is_h264){
1697             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1698                 if(!transform_bypass)
1699                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1700                 else{
1701                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1702                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1703                     for(i = 0; i < 16; i++)
1704                         dctcoef_set(h->mb+p*256, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1705                 }
1706             }
1707         }else
1708             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1709     }
1710 }
1711
1712 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1713                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1714 {
1715     MpegEncContext * const s = &h->s;
1716     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1717     int i;
1718     block_offset += 16*p;
1719     if(!IS_INTRA4x4(mb_type)){
1720         if(is_h264){
1721             if(IS_INTRA16x16(mb_type)){
1722                 if(transform_bypass){
1723                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1724                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1725                     }else{
1726                         for(i=0; i<16; i++){
1727                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1728                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1729                         }
1730                     }
1731                 }else{
1732                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1733                 }
1734             }else if(h->cbp&15){
1735                 if(transform_bypass){
1736                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1737                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1738                     for(i=0; i<16; i+=di){
1739                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1740                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1741                         }
1742                     }
1743                 }else{
1744                     if(IS_8x8DCT(mb_type)){
1745                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1746                     }else{
1747                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1748                     }
1749                 }
1750             }
1751         }else{
1752             for(i=0; i<16; i++){
1753                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1754                     uint8_t * const ptr= dest_y + block_offset[i];
1755                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1756                 }
1757             }
1758         }
1759     }
1760 }
1761
1762 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1763     MpegEncContext * const s = &h->s;
1764     const int mb_x= s->mb_x;
1765     const int mb_y= s->mb_y;
1766     const int mb_xy= h->mb_xy;
1767     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1768     uint8_t  *dest_y, *dest_cb, *dest_cr;
1769     int linesize, uvlinesize /*dct_offset*/;
1770     int i, j;
1771     int *block_offset = &h->block_offset[0];
1772     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1773     /* is_h264 should always be true if SVQ3 is disabled. */
1774     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1775     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1776
1777     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1778     dest_cb = s->current_picture.f.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) *  8;
1779     dest_cr = s->current_picture.f.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) *  8;
1780
1781     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1782     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1783
1784     h->list_counts[mb_xy]= h->list_count;
1785
1786     if (!simple && MB_FIELD) {
1787         linesize   = h->mb_linesize   = s->linesize * 2;
1788         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1789         block_offset = &h->block_offset[48];
1790         if(mb_y&1){ //FIXME move out of this function?
1791             dest_y -= s->linesize*15;
1792             dest_cb-= s->uvlinesize*7;
1793             dest_cr-= s->uvlinesize*7;
1794         }
1795         if(FRAME_MBAFF) {
1796             int list;
1797             for(list=0; list<h->list_count; list++){
1798                 if(!USES_LIST(mb_type, list))
1799                     continue;
1800                 if(IS_16X16(mb_type)){
1801                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1802                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1803                 }else{
1804                     for(i=0; i<16; i+=4){
1805                         int ref = h->ref_cache[list][scan8[i]];
1806                         if(ref >= 0)
1807                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1808                     }
1809                 }
1810             }
1811         }
1812     } else {
1813         linesize   = h->mb_linesize   = s->linesize;
1814         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1815 //        dct_offset = s->linesize * 16;
1816     }
1817
1818     if (!simple && IS_INTRA_PCM(mb_type)) {
1819         if (pixel_shift) {
1820             const int bit_depth = h->sps.bit_depth_luma;
1821             int j;
1822             GetBitContext gb;
1823             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1824
1825             for (i = 0; i < 16; i++) {
1826                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1827                 for (j = 0; j < 16; j++)
1828                     tmp_y[j] = get_bits(&gb, bit_depth);
1829             }
1830             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1831                 if (!h->sps.chroma_format_idc) {
1832                     for (i = 0; i < 8; i++) {
1833                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1834                         for (j = 0; j < 8; j++) {
1835                             tmp_cb[j] = 1 << (bit_depth - 1);
1836                         }
1837                     }
1838                     for (i = 0; i < 8; i++) {
1839                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1840                         for (j = 0; j < 8; j++) {
1841                             tmp_cr[j] = 1 << (bit_depth - 1);
1842                         }
1843                     }
1844                 } else {
1845                     for (i = 0; i < 8; i++) {
1846                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1847                         for (j = 0; j < 8; j++)
1848                             tmp_cb[j] = get_bits(&gb, bit_depth);
1849                     }
1850                     for (i = 0; i < 8; i++) {
1851                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1852                         for (j = 0; j < 8; j++)
1853                             tmp_cr[j] = get_bits(&gb, bit_depth);
1854                     }
1855                 }
1856             }
1857         } else {
1858             for (i=0; i<16; i++) {
1859                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1860             }
1861             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1862                 if (!h->sps.chroma_format_idc) {
1863                     for (i = 0; i < 8; i++) {
1864                         memset(dest_cb + i*uvlinesize, 128, 8);
1865                         memset(dest_cr + i*uvlinesize, 128, 8);
1866                     }
1867                 } else {
1868                     for (i = 0; i < 8; i++) {
1869                         memcpy(dest_cb + i*uvlinesize, h->mb + 128 + i*4,  8);
1870                         memcpy(dest_cr + i*uvlinesize, h->mb + 160 + i*4,  8);
1871                     }
1872                 }
1873             }
1874         }
1875     } else {
1876         if(IS_INTRA(mb_type)){
1877             if(h->deblocking_filter)
1878                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1879
1880             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1881                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1882                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1883             }
1884
1885             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1886
1887             if(h->deblocking_filter)
1888                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1889         }else if(is_h264){
1890             hl_motion(h, dest_y, dest_cb, dest_cr,
1891                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1892                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1893                       h->h264dsp.weight_h264_pixels_tab,
1894                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 0);
1895         }
1896
1897         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1898
1899         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1900             uint8_t *dest[2] = {dest_cb, dest_cr};
1901             if(transform_bypass){
1902                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1903                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
1904                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
1905                 }else{
1906                     idct_add = s->dsp.add_pixels4;
1907                     for(j=1; j<3; j++){
1908                         for(i=j*16; i<j*16+4; i++){
1909                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1910                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1911                         }
1912                     }
1913                 }
1914             }else{
1915                 if(is_h264){
1916                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1917                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*1 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1918                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1919                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*2 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
1920                     h->h264dsp.h264_idct_add8(dest, block_offset,
1921                                               h->mb, uvlinesize,
1922                                               h->non_zero_count_cache);
1923                 }else{
1924                     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]);
1925                     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]);
1926                     for(j=1; j<3; j++){
1927                         for(i=j*16; i<j*16+4; i++){
1928                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1929                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
1930                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1931                             }
1932                         }
1933                     }
1934                 }
1935             }
1936         }
1937     }
1938     if(h->cbp || IS_INTRA(mb_type))
1939     {
1940         s->dsp.clear_blocks(h->mb);
1941         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
1942     }
1943 }
1944
1945 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
1946     MpegEncContext * const s = &h->s;
1947     const int mb_x= s->mb_x;
1948     const int mb_y= s->mb_y;
1949     const int mb_xy= h->mb_xy;
1950     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1951     uint8_t  *dest[3];
1952     int linesize;
1953     int i, j, p;
1954     int *block_offset = &h->block_offset[0];
1955     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1956     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
1957
1958     for (p = 0; p < plane_count; p++)
1959     {
1960         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
1961         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1962     }
1963
1964     h->list_counts[mb_xy]= h->list_count;
1965
1966     if (!simple && MB_FIELD) {
1967         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
1968         block_offset = &h->block_offset[48];
1969         if(mb_y&1) //FIXME move out of this function?
1970             for (p = 0; p < 3; p++)
1971                 dest[p] -= s->linesize*15;
1972         if(FRAME_MBAFF) {
1973             int list;
1974             for(list=0; list<h->list_count; list++){
1975                 if(!USES_LIST(mb_type, list))
1976                     continue;
1977                 if(IS_16X16(mb_type)){
1978                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1979                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1980                 }else{
1981                     for(i=0; i<16; i+=4){
1982                         int ref = h->ref_cache[list][scan8[i]];
1983                         if(ref >= 0)
1984                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1985                     }
1986                 }
1987             }
1988         }
1989     } else {
1990         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
1991     }
1992
1993     if (!simple && IS_INTRA_PCM(mb_type)) {
1994         if (pixel_shift) {
1995             const int bit_depth = h->sps.bit_depth_luma;
1996             GetBitContext gb;
1997             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
1998
1999             for (p = 0; p < plane_count; p++) {
2000                 for (i = 0; i < 16; i++) {
2001                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2002                     for (j = 0; j < 16; j++)
2003                         tmp[j] = get_bits(&gb, bit_depth);
2004                 }
2005             }
2006         } else {
2007             for (p = 0; p < plane_count; p++) {
2008                 for (i = 0; i < 16; i++) {
2009                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2010                 }
2011             }
2012         }
2013     } else {
2014         if(IS_INTRA(mb_type)){
2015             if(h->deblocking_filter)
2016                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2017
2018             for (p = 0; p < plane_count; p++)
2019                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2020
2021             if(h->deblocking_filter)
2022                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2023         }else{
2024             hl_motion(h, dest[0], dest[1], dest[2],
2025                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2026                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2027                       h->h264dsp.weight_h264_pixels_tab,
2028                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 1);
2029         }
2030
2031         for (p = 0; p < plane_count; p++)
2032             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2033     }
2034     if(h->cbp || IS_INTRA(mb_type))
2035     {
2036         s->dsp.clear_blocks(h->mb);
2037         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2038     }
2039 }
2040
2041 /**
2042  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2043  */
2044 #define hl_decode_mb_simple(sh, bits) \
2045 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2046     hl_decode_mb_internal(h, 1, sh); \
2047 }
2048 hl_decode_mb_simple(0, 8);
2049 hl_decode_mb_simple(1, 16);
2050
2051 /**
2052  * Process a macroblock; this handles edge cases, such as interlacing.
2053  */
2054 static void av_noinline hl_decode_mb_complex(H264Context *h){
2055     hl_decode_mb_internal(h, 0, h->pixel_shift);
2056 }
2057
2058 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2059     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2060 }
2061
2062 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2063     hl_decode_mb_444_internal(h, 1, 0);
2064 }
2065
2066 void ff_h264_hl_decode_mb(H264Context *h){
2067     MpegEncContext * const s = &h->s;
2068     const int mb_xy= h->mb_xy;
2069     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2070     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2071
2072     if (CHROMA444) {
2073         if(is_complex || h->pixel_shift)
2074             hl_decode_mb_444_complex(h);
2075         else
2076             hl_decode_mb_444_simple(h);
2077     } else if (is_complex) {
2078         hl_decode_mb_complex(h);
2079     } else if (h->pixel_shift) {
2080         hl_decode_mb_simple_16(h);
2081     } else
2082         hl_decode_mb_simple_8(h);
2083 }
2084
2085 static int pred_weight_table(H264Context *h){
2086     MpegEncContext * const s = &h->s;
2087     int list, i;
2088     int luma_def, chroma_def;
2089
2090     h->use_weight= 0;
2091     h->use_weight_chroma= 0;
2092     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2093     if(h->sps.chroma_format_idc)
2094         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2095     luma_def = 1<<h->luma_log2_weight_denom;
2096     chroma_def = 1<<h->chroma_log2_weight_denom;
2097
2098     for(list=0; list<2; list++){
2099         h->luma_weight_flag[list]   = 0;
2100         h->chroma_weight_flag[list] = 0;
2101         for(i=0; i<h->ref_count[list]; i++){
2102             int luma_weight_flag, chroma_weight_flag;
2103
2104             luma_weight_flag= get_bits1(&s->gb);
2105             if(luma_weight_flag){
2106                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2107                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2108                 if(   h->luma_weight[i][list][0] != luma_def
2109                    || h->luma_weight[i][list][1] != 0) {
2110                     h->use_weight= 1;
2111                     h->luma_weight_flag[list]= 1;
2112                 }
2113             }else{
2114                 h->luma_weight[i][list][0]= luma_def;
2115                 h->luma_weight[i][list][1]= 0;
2116             }
2117
2118             if(h->sps.chroma_format_idc){
2119                 chroma_weight_flag= get_bits1(&s->gb);
2120                 if(chroma_weight_flag){
2121                     int j;
2122                     for(j=0; j<2; j++){
2123                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2124                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2125                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2126                            || h->chroma_weight[i][list][j][1] != 0) {
2127                             h->use_weight_chroma= 1;
2128                             h->chroma_weight_flag[list]= 1;
2129                         }
2130                     }
2131                 }else{
2132                     int j;
2133                     for(j=0; j<2; j++){
2134                         h->chroma_weight[i][list][j][0]= chroma_def;
2135                         h->chroma_weight[i][list][j][1]= 0;
2136                     }
2137                 }
2138             }
2139         }
2140         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2141     }
2142     h->use_weight= h->use_weight || h->use_weight_chroma;
2143     return 0;
2144 }
2145
2146 /**
2147  * Initialize implicit_weight table.
2148  * @param field  0/1 initialize the weight for interlaced MBAFF
2149  *                -1 initializes the rest
2150  */
2151 static void implicit_weight_table(H264Context *h, int field){
2152     MpegEncContext * const s = &h->s;
2153     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2154
2155     for (i = 0; i < 2; i++) {
2156         h->luma_weight_flag[i]   = 0;
2157         h->chroma_weight_flag[i] = 0;
2158     }
2159
2160     if(field < 0){
2161         if (s->picture_structure == PICT_FRAME) {
2162             cur_poc = s->current_picture_ptr->poc;
2163         } else {
2164             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2165         }
2166     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2167        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2168         h->use_weight= 0;
2169         h->use_weight_chroma= 0;
2170         return;
2171     }
2172         ref_start= 0;
2173         ref_count0= h->ref_count[0];
2174         ref_count1= h->ref_count[1];
2175     }else{
2176         cur_poc = s->current_picture_ptr->field_poc[field];
2177         ref_start= 16;
2178         ref_count0= 16+2*h->ref_count[0];
2179         ref_count1= 16+2*h->ref_count[1];
2180     }
2181
2182     h->use_weight= 2;
2183     h->use_weight_chroma= 2;
2184     h->luma_log2_weight_denom= 5;
2185     h->chroma_log2_weight_denom= 5;
2186
2187     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2188         int poc0 = h->ref_list[0][ref0].poc;
2189         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2190             int w = 32;
2191             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2192                 int poc1 = h->ref_list[1][ref1].poc;
2193                 int td = av_clip(poc1 - poc0, -128, 127);
2194                 if(td){
2195                     int tb = av_clip(cur_poc - poc0, -128, 127);
2196                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2197                     int dist_scale_factor = (tb*tx + 32) >> 8;
2198                     if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2199                         w = 64 - dist_scale_factor;
2200                 }
2201             }
2202             if(field<0){
2203                 h->implicit_weight[ref0][ref1][0]=
2204                 h->implicit_weight[ref0][ref1][1]= w;
2205             }else{
2206                 h->implicit_weight[ref0][ref1][field]=w;
2207             }
2208         }
2209     }
2210 }
2211
2212 /**
2213  * instantaneous decoder refresh.
2214  */
2215 static void idr(H264Context *h){
2216     ff_h264_remove_all_refs(h);
2217     h->prev_frame_num= 0;
2218     h->prev_frame_num_offset= 0;
2219     h->prev_poc_msb=
2220     h->prev_poc_lsb= 0;
2221 }
2222
2223 /* forget old pics after a seek */
2224 static void flush_dpb(AVCodecContext *avctx){
2225     H264Context *h= avctx->priv_data;
2226     int i;
2227     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
2228         if(h->delayed_pic[i])
2229             h->delayed_pic[i]->f.reference = 0;
2230         h->delayed_pic[i]= NULL;
2231     }
2232     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2233     h->prev_interlaced_frame = 1;
2234     idr(h);
2235     if(h->s.current_picture_ptr)
2236         h->s.current_picture_ptr->f.reference = 0;
2237     h->s.first_field= 0;
2238     ff_h264_reset_sei(h);
2239     ff_mpeg_flush(avctx);
2240 }
2241
2242 static int init_poc(H264Context *h){
2243     MpegEncContext * const s = &h->s;
2244     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2245     int field_poc[2];
2246     Picture *cur = s->current_picture_ptr;
2247
2248     h->frame_num_offset= h->prev_frame_num_offset;
2249     if(h->frame_num < h->prev_frame_num)
2250         h->frame_num_offset += max_frame_num;
2251
2252     if(h->sps.poc_type==0){
2253         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2254
2255         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2256             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2257         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2258             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2259         else
2260             h->poc_msb = h->prev_poc_msb;
2261 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2262         field_poc[0] =
2263         field_poc[1] = h->poc_msb + h->poc_lsb;
2264         if(s->picture_structure == PICT_FRAME)
2265             field_poc[1] += h->delta_poc_bottom;
2266     }else if(h->sps.poc_type==1){
2267         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2268         int i;
2269
2270         if(h->sps.poc_cycle_length != 0)
2271             abs_frame_num = h->frame_num_offset + h->frame_num;
2272         else
2273             abs_frame_num = 0;
2274
2275         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2276             abs_frame_num--;
2277
2278         expected_delta_per_poc_cycle = 0;
2279         for(i=0; i < h->sps.poc_cycle_length; i++)
2280             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2281
2282         if(abs_frame_num > 0){
2283             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2284             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2285
2286             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2287             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2288                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2289         } else
2290             expectedpoc = 0;
2291
2292         if(h->nal_ref_idc == 0)
2293             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2294
2295         field_poc[0] = expectedpoc + h->delta_poc[0];
2296         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2297
2298         if(s->picture_structure == PICT_FRAME)
2299             field_poc[1] += h->delta_poc[1];
2300     }else{
2301         int poc= 2*(h->frame_num_offset + h->frame_num);
2302
2303         if(!h->nal_ref_idc)
2304             poc--;
2305
2306         field_poc[0]= poc;
2307         field_poc[1]= poc;
2308     }
2309
2310     if(s->picture_structure != PICT_BOTTOM_FIELD)
2311         s->current_picture_ptr->field_poc[0]= field_poc[0];
2312     if(s->picture_structure != PICT_TOP_FIELD)
2313         s->current_picture_ptr->field_poc[1]= field_poc[1];
2314     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2315
2316     return 0;
2317 }
2318
2319
2320 /**
2321  * initialize scan tables
2322  */
2323 static void init_scan_tables(H264Context *h){
2324     int i;
2325     for(i=0; i<16; i++){
2326 #define T(x) (x>>2) | ((x<<2) & 0xF)
2327         h->zigzag_scan[i] = T(zigzag_scan[i]);
2328         h-> field_scan[i] = T( field_scan[i]);
2329 #undef T
2330     }
2331     for(i=0; i<64; i++){
2332 #define T(x) (x>>3) | ((x&7)<<3)
2333         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2334         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2335         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2336         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2337 #undef T
2338     }
2339     if(h->sps.transform_bypass){ //FIXME same ugly
2340         h->zigzag_scan_q0          = zigzag_scan;
2341         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2342         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2343         h->field_scan_q0           = field_scan;
2344         h->field_scan8x8_q0        = field_scan8x8;
2345         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2346     }else{
2347         h->zigzag_scan_q0          = h->zigzag_scan;
2348         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2349         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2350         h->field_scan_q0           = h->field_scan;
2351         h->field_scan8x8_q0        = h->field_scan8x8;
2352         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2353     }
2354 }
2355
2356 static int field_end(H264Context *h, int in_setup){
2357     MpegEncContext * const s = &h->s;
2358     AVCodecContext * const avctx= s->avctx;
2359     int err = 0;
2360     s->mb_y= 0;
2361
2362     if (!in_setup && !s->dropable)
2363         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2364                                  s->picture_structure==PICT_BOTTOM_FIELD);
2365
2366     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2367         ff_vdpau_h264_set_reference_frames(s);
2368
2369     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2370         if(!s->dropable) {
2371             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2372             h->prev_poc_msb= h->poc_msb;
2373             h->prev_poc_lsb= h->poc_lsb;
2374         }
2375         h->prev_frame_num_offset= h->frame_num_offset;
2376         h->prev_frame_num= h->frame_num;
2377         h->outputed_poc = h->next_outputed_poc;
2378     }
2379
2380     if (avctx->hwaccel) {
2381         if (avctx->hwaccel->end_frame(avctx) < 0)
2382             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2383     }
2384
2385     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2386         ff_vdpau_h264_picture_complete(s);
2387
2388     /*
2389      * FIXME: Error handling code does not seem to support interlaced
2390      * when slices span multiple rows
2391      * The ff_er_add_slice calls don't work right for bottom
2392      * fields; they cause massive erroneous error concealing
2393      * Error marking covers both fields (top and bottom).
2394      * This causes a mismatched s->error_count
2395      * and a bad error table. Further, the error count goes to
2396      * INT_MAX when called for bottom field, because mb_y is
2397      * past end by one (callers fault) and resync_mb_y != 0
2398      * causes problems for the first MB line, too.
2399      */
2400     if (!FIELD_PICTURE)
2401         ff_er_frame_end(s);
2402
2403     MPV_frame_end(s);
2404
2405     h->current_slice=0;
2406
2407     return err;
2408 }
2409
2410 /**
2411  * Replicate H264 "master" context to thread contexts.
2412  */
2413 static void clone_slice(H264Context *dst, H264Context *src)
2414 {
2415     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2416     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2417     dst->s.current_picture      = src->s.current_picture;
2418     dst->s.linesize             = src->s.linesize;
2419     dst->s.uvlinesize           = src->s.uvlinesize;
2420     dst->s.first_field          = src->s.first_field;
2421
2422     dst->prev_poc_msb           = src->prev_poc_msb;
2423     dst->prev_poc_lsb           = src->prev_poc_lsb;
2424     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2425     dst->prev_frame_num         = src->prev_frame_num;
2426     dst->short_ref_count        = src->short_ref_count;
2427
2428     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2429     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2430     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2431     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2432
2433     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2434     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2435 }
2436
2437 /**
2438  * computes profile from profile_idc and constraint_set?_flags
2439  *
2440  * @param sps SPS
2441  *
2442  * @return profile as defined by FF_PROFILE_H264_*
2443  */
2444 int ff_h264_get_profile(SPS *sps)
2445 {
2446     int profile = sps->profile_idc;
2447
2448     switch(sps->profile_idc) {
2449     case FF_PROFILE_H264_BASELINE:
2450         // constraint_set1_flag set to 1
2451         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2452         break;
2453     case FF_PROFILE_H264_HIGH_10:
2454     case FF_PROFILE_H264_HIGH_422:
2455     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2456         // constraint_set3_flag set to 1
2457         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2458         break;
2459     }
2460
2461     return profile;
2462 }
2463
2464 /**
2465  * decodes a slice header.
2466  * This will also call MPV_common_init() and frame_start() as needed.
2467  *
2468  * @param h h264context
2469  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2470  *
2471  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2472  */
2473 static int decode_slice_header(H264Context *h, H264Context *h0){
2474     MpegEncContext * const s = &h->s;
2475     MpegEncContext * const s0 = &h0->s;
2476     unsigned int first_mb_in_slice;
2477     unsigned int pps_id;
2478     int num_ref_idx_active_override_flag;
2479     unsigned int slice_type, tmp, i, j;
2480     int default_ref_list_done = 0;
2481     int last_pic_structure;
2482
2483     s->dropable= h->nal_ref_idc == 0;
2484
2485     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2486     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2487         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2488         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2489     }else{
2490         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2491         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2492     }
2493
2494     first_mb_in_slice= get_ue_golomb(&s->gb);
2495
2496     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2497         if(h0->current_slice && FIELD_PICTURE){
2498             field_end(h, 1);
2499         }
2500
2501         h0->current_slice = 0;
2502         if (!s0->first_field)
2503             s->current_picture_ptr= NULL;
2504     }
2505
2506     slice_type= get_ue_golomb_31(&s->gb);
2507     if(slice_type > 9){
2508         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);
2509         return -1;
2510     }
2511     if(slice_type > 4){
2512         slice_type -= 5;
2513         h->slice_type_fixed=1;
2514     }else
2515         h->slice_type_fixed=0;
2516
2517     slice_type= golomb_to_pict_type[ slice_type ];
2518     if (slice_type == AV_PICTURE_TYPE_I
2519         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2520         default_ref_list_done = 1;
2521     }
2522     h->slice_type= slice_type;
2523     h->slice_type_nos= slice_type & 3;
2524
2525     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2526
2527     pps_id= get_ue_golomb(&s->gb);
2528     if(pps_id>=MAX_PPS_COUNT){
2529         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2530         return -1;
2531     }
2532     if(!h0->pps_buffers[pps_id]) {
2533         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2534         return -1;
2535     }
2536     h->pps= *h0->pps_buffers[pps_id];
2537
2538     if(!h0->sps_buffers[h->pps.sps_id]) {
2539         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2540         return -1;
2541     }
2542     h->sps = *h0->sps_buffers[h->pps.sps_id];
2543
2544     s->avctx->profile = ff_h264_get_profile(&h->sps);
2545     s->avctx->level   = h->sps.level_idc;
2546     s->avctx->refs    = h->sps.ref_frame_count;
2547
2548     if(h == h0 && h->dequant_coeff_pps != pps_id){
2549         h->dequant_coeff_pps = pps_id;
2550         init_dequant_tables(h);
2551     }
2552
2553     s->mb_width= h->sps.mb_width;
2554     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2555
2556     h->b_stride=  s->mb_width*4;
2557
2558     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2559     if(h->sps.frame_mbs_only_flag)
2560         s->height= 16*s->mb_height - (2>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2561     else
2562         s->height= 16*s->mb_height - (4>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2563
2564     if (s->context_initialized
2565         && (   s->width != s->avctx->width || s->height != s->avctx->height
2566             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2567         if(h != h0) {
2568             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2569             return -1;   // width / height changed during parallelized decoding
2570         }
2571         free_tables(h, 0);
2572         flush_dpb(s->avctx);
2573         MPV_common_end(s);
2574     }
2575     if (!s->context_initialized) {
2576         if (h != h0) {
2577             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2578             return -1;
2579         }
2580
2581         avcodec_set_dimensions(s->avctx, s->width, s->height);
2582         s->avctx->sample_aspect_ratio= h->sps.sar;
2583         av_assert0(s->avctx->sample_aspect_ratio.den);
2584
2585         if(h->sps.video_signal_type_present_flag){
2586             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2587             if(h->sps.colour_description_present_flag){
2588                 s->avctx->color_primaries = h->sps.color_primaries;
2589                 s->avctx->color_trc       = h->sps.color_trc;
2590                 s->avctx->colorspace      = h->sps.colorspace;
2591             }
2592         }
2593
2594         if(h->sps.timing_info_present_flag){
2595             int64_t den= h->sps.time_scale;
2596             if(h->x264_build < 44U)
2597                 den *= 2;
2598             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2599                       h->sps.num_units_in_tick, den, 1<<30);
2600         }
2601
2602         switch (h->sps.bit_depth_luma) {
2603             case 9 :
2604                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P9 : PIX_FMT_YUV420P9;
2605                 break;
2606             case 10 :
2607                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P10 : PIX_FMT_YUV420P10;
2608                 break;
2609             default:
2610                 if (CHROMA444){
2611                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2612                 }else{
2613                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2614                                                              s->avctx->codec->pix_fmts ?
2615                                                              s->avctx->codec->pix_fmts :
2616                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2617                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2618                                                              ff_hwaccel_pixfmt_list_420);
2619                 }
2620         }
2621
2622         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2623
2624         if (MPV_common_init(s) < 0) {
2625             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2626             return -1;
2627         }
2628         s->first_field = 0;
2629         h->prev_interlaced_frame = 1;
2630
2631         init_scan_tables(h);
2632         if (ff_h264_alloc_tables(h) < 0) {
2633             av_log(h->s.avctx, AV_LOG_ERROR, "Could not allocate memory for h264\n");
2634             return AVERROR(ENOMEM);
2635         }
2636
2637         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2638             if (context_init(h) < 0) {
2639                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2640                 return -1;
2641             }
2642         } else {
2643             for(i = 1; i < s->avctx->thread_count; i++) {
2644                 H264Context *c;
2645                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2646                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2647                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2648                 c->h264dsp = h->h264dsp;
2649                 c->sps = h->sps;
2650                 c->pps = h->pps;
2651                 c->pixel_shift = h->pixel_shift;
2652                 init_scan_tables(c);
2653                 clone_tables(c, h, i);
2654             }
2655
2656             for(i = 0; i < s->avctx->thread_count; i++)
2657                 if (context_init(h->thread_context[i]) < 0) {
2658                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2659                     return -1;
2660                 }
2661         }
2662     }
2663
2664     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2665
2666     h->mb_mbaff = 0;
2667     h->mb_aff_frame = 0;
2668     last_pic_structure = s0->picture_structure;
2669     if(h->sps.frame_mbs_only_flag){
2670         s->picture_structure= PICT_FRAME;
2671     }else{
2672         if(get_bits1(&s->gb)) { //field_pic_flag
2673             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2674         } else {
2675             s->picture_structure= PICT_FRAME;
2676             h->mb_aff_frame = h->sps.mb_aff;
2677         }
2678     }
2679     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2680
2681     if(h0->current_slice == 0){
2682         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2683         if(h->frame_num != h->prev_frame_num) {
2684             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2685
2686             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2687
2688             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2689                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2690                 if (unwrap_prev_frame_num < 0)
2691                     unwrap_prev_frame_num += max_frame_num;
2692
2693                 h->prev_frame_num = unwrap_prev_frame_num;
2694             }
2695         }
2696
2697         while(h->frame_num !=  h->prev_frame_num &&
2698               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2699             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2700             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2701             if (ff_h264_frame_start(h) < 0)
2702                 return -1;
2703             h->prev_frame_num++;
2704             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2705             s->current_picture_ptr->frame_num= h->prev_frame_num;
2706             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2707             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2708             ff_generate_sliding_window_mmcos(h);
2709             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2710                 s->avctx->error_recognition >= FF_ER_EXPLODE)
2711                 return AVERROR_INVALIDDATA;
2712             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2713              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2714              * about there being no actual duplicates.
2715              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2716              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2717              * be fixed. */
2718             if (h->short_ref_count) {
2719                 if (prev) {
2720                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2721                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2722                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2723                     h->short_ref[0]->poc = prev->poc+2;
2724                 }
2725                 h->short_ref[0]->frame_num = h->prev_frame_num;
2726             }
2727         }
2728
2729         /* See if we have a decoded first field looking for a pair... */
2730         if (s0->first_field) {
2731             assert(s0->current_picture_ptr);
2732             assert(s0->current_picture_ptr->f.data[0]);
2733             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2734
2735             /* figure out if we have a complementary field pair */
2736             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2737                 /*
2738                  * Previous field is unmatched. Don't display it, but let it
2739                  * remain for reference if marked as such.
2740                  */
2741                 s0->current_picture_ptr = NULL;
2742                 s0->first_field = FIELD_PICTURE;
2743
2744             } else {
2745                 if (h->nal_ref_idc &&
2746                         s0->current_picture_ptr->f.reference &&
2747                         s0->current_picture_ptr->frame_num != h->frame_num) {
2748                     /*
2749                      * This and previous field were reference, but had
2750                      * different frame_nums. Consider this field first in
2751                      * pair. Throw away previous field except for reference
2752                      * purposes.
2753                      */
2754                     s0->first_field = 1;
2755                     s0->current_picture_ptr = NULL;
2756
2757                 } else {
2758                     /* Second field in complementary pair */
2759                     s0->first_field = 0;
2760                 }
2761             }
2762
2763         } else {
2764             /* Frame or first field in a potentially complementary pair */
2765             assert(!s0->current_picture_ptr);
2766             s0->first_field = FIELD_PICTURE;
2767         }
2768
2769         if(!FIELD_PICTURE || s0->first_field) {
2770             if (ff_h264_frame_start(h) < 0) {
2771                 s0->first_field = 0;
2772                 return -1;
2773             }
2774         } else {
2775             ff_release_unused_pictures(s, 0);
2776         }
2777     }
2778     if(h != h0)
2779         clone_slice(h, h0);
2780
2781     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2782
2783     assert(s->mb_num == s->mb_width * s->mb_height);
2784     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2785        first_mb_in_slice                    >= s->mb_num){
2786         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2787         return -1;
2788     }
2789     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2790     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2791     if (s->picture_structure == PICT_BOTTOM_FIELD)
2792         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2793     assert(s->mb_y < s->mb_height);
2794
2795     if(s->picture_structure==PICT_FRAME){
2796         h->curr_pic_num=   h->frame_num;
2797         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2798     }else{
2799         h->curr_pic_num= 2*h->frame_num + 1;
2800         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2801     }
2802
2803     if(h->nal_unit_type == NAL_IDR_SLICE){
2804         get_ue_golomb(&s->gb); /* idr_pic_id */
2805     }
2806
2807     if(h->sps.poc_type==0){
2808         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2809
2810         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2811             h->delta_poc_bottom= get_se_golomb(&s->gb);
2812         }
2813     }
2814
2815     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2816         h->delta_poc[0]= get_se_golomb(&s->gb);
2817
2818         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2819             h->delta_poc[1]= get_se_golomb(&s->gb);
2820     }
2821
2822     init_poc(h);
2823
2824     if(h->pps.redundant_pic_cnt_present){
2825         h->redundant_pic_count= get_ue_golomb(&s->gb);
2826     }
2827
2828     //set defaults, might be overridden a few lines later
2829     h->ref_count[0]= h->pps.ref_count[0];
2830     h->ref_count[1]= h->pps.ref_count[1];
2831
2832     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2833         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2834             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2835         }
2836         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2837
2838         if(num_ref_idx_active_override_flag){
2839             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2840             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2841                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2842
2843             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2844                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2845                 h->ref_count[0]= h->ref_count[1]= 1;
2846                 return -1;
2847             }
2848         }
2849         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2850             h->list_count= 2;
2851         else
2852             h->list_count= 1;
2853     }else
2854         h->list_count= 0;
2855
2856     if(!default_ref_list_done){
2857         ff_h264_fill_default_ref_list(h);
2858     }
2859
2860     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0) {
2861         h->ref_count[1]= h->ref_count[0]= 0;
2862         return -1;
2863     }
2864
2865     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2866         s->last_picture_ptr= &h->ref_list[0][0];
2867         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2868     }
2869     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2870         s->next_picture_ptr= &h->ref_list[1][0];
2871         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2872     }
2873
2874     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2875        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2876         pred_weight_table(h);
2877     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2878         implicit_weight_table(h, -1);
2879     }else {
2880         h->use_weight = 0;
2881         for (i = 0; i < 2; i++) {
2882             h->luma_weight_flag[i]   = 0;
2883             h->chroma_weight_flag[i] = 0;
2884         }
2885     }
2886
2887     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
2888        s->avctx->error_recognition >= FF_ER_EXPLODE)
2889         return AVERROR_INVALIDDATA;
2890
2891     if(FRAME_MBAFF){
2892         ff_h264_fill_mbaff_ref_list(h);
2893
2894         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2895             implicit_weight_table(h, 0);
2896             implicit_weight_table(h, 1);
2897         }
2898     }
2899
2900     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2901         ff_h264_direct_dist_scale_factor(h);
2902     ff_h264_direct_ref_list_init(h);
2903
2904     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2905         tmp = get_ue_golomb_31(&s->gb);
2906         if(tmp > 2){
2907             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2908             return -1;
2909         }
2910         h->cabac_init_idc= tmp;
2911     }
2912
2913     h->last_qscale_diff = 0;
2914     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2915     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
2916         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2917         return -1;
2918     }
2919     s->qscale= tmp;
2920     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2921     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2922     //FIXME qscale / qp ... stuff
2923     if(h->slice_type == AV_PICTURE_TYPE_SP){
2924         get_bits1(&s->gb); /* sp_for_switch_flag */
2925     }
2926     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
2927         get_se_golomb(&s->gb); /* slice_qs_delta */
2928     }
2929
2930     h->deblocking_filter = 1;
2931     h->slice_alpha_c0_offset = 52;
2932     h->slice_beta_offset = 52;
2933     if( h->pps.deblocking_filter_parameters_present ) {
2934         tmp= get_ue_golomb_31(&s->gb);
2935         if(tmp > 2){
2936             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2937             return -1;
2938         }
2939         h->deblocking_filter= tmp;
2940         if(h->deblocking_filter < 2)
2941             h->deblocking_filter^= 1; // 1<->0
2942
2943         if( h->deblocking_filter ) {
2944             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2945             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2946             if(   h->slice_alpha_c0_offset > 104U
2947                || h->slice_beta_offset     > 104U){
2948                 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);
2949                 return -1;
2950             }
2951         }
2952     }
2953
2954     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2955        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
2956        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
2957        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2958         h->deblocking_filter= 0;
2959
2960     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
2961         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
2962             /* Cheat slightly for speed:
2963                Do not bother to deblock across slices. */
2964             h->deblocking_filter = 2;
2965         } else {
2966             h0->max_contexts = 1;
2967             if(!h0->single_decode_warning) {
2968                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
2969                 h0->single_decode_warning = 1;
2970             }
2971             if (h != h0) {
2972                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
2973                 return 1;
2974             }
2975         }
2976     }
2977     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
2978                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
2979                  + 6 * (h->sps.bit_depth_luma - 8);
2980
2981 #if 0 //FMO
2982     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
2983         slice_group_change_cycle= get_bits(&s->gb, ?);
2984 #endif
2985
2986     h0->last_slice_type = slice_type;
2987     h->slice_num = ++h0->current_slice;
2988     if(h->slice_num >= MAX_SLICES){
2989         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
2990     }
2991
2992     for(j=0; j<2; j++){
2993         int id_list[16];
2994         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
2995         for(i=0; i<16; i++){
2996             id_list[i]= 60;
2997             if (h->ref_list[j][i].f.data[0]) {
2998                 int k;
2999                 uint8_t *base = h->ref_list[j][i].f.base[0];
3000                 for(k=0; k<h->short_ref_count; k++)
3001                     if (h->short_ref[k]->f.base[0] == base) {
3002                         id_list[i]= k;
3003                         break;
3004                     }
3005                 for(k=0; k<h->long_ref_count; k++)
3006                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3007                         id_list[i]= h->short_ref_count + k;
3008                         break;
3009                     }
3010             }
3011         }
3012
3013         ref2frm[0]=
3014         ref2frm[1]= -1;
3015         for(i=0; i<16; i++)
3016             ref2frm[i+2]= 4*id_list[i]
3017                           + (h->ref_list[j][i].f.reference & 3);
3018         ref2frm[18+0]=
3019         ref2frm[18+1]= -1;
3020         for(i=16; i<48; i++)
3021             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3022                           + (h->ref_list[j][i].f.reference & 3);
3023     }
3024
3025     //FIXME: fix draw_edges+PAFF+frame threads
3026     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3027     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3028
3029     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3030         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",
3031                h->slice_num,
3032                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3033                first_mb_in_slice,
3034                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3035                pps_id, h->frame_num,
3036                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3037                h->ref_count[0], h->ref_count[1],
3038                s->qscale,
3039                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3040                h->use_weight,
3041                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3042                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3043                );
3044     }
3045
3046     return 0;
3047 }
3048
3049 int ff_h264_get_slice_type(const H264Context *h)
3050 {
3051     switch (h->slice_type) {
3052     case AV_PICTURE_TYPE_P:  return 0;
3053     case AV_PICTURE_TYPE_B:  return 1;
3054     case AV_PICTURE_TYPE_I:  return 2;
3055     case AV_PICTURE_TYPE_SP: return 3;
3056     case AV_PICTURE_TYPE_SI: return 4;
3057     default:         return -1;
3058     }
3059 }
3060
3061 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3062                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3063 {
3064     int b_stride = h->b_stride;
3065     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3066     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3067     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3068         if(USES_LIST(top_type, list)){
3069             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3070             const int b8_xy= 4*top_xy + 2;
3071             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3072             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3073             ref_cache[0 - 1*8]=
3074             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3075             ref_cache[2 - 1*8]=
3076             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3077         }else{
3078             AV_ZERO128(mv_dst - 1*8);
3079             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3080         }
3081
3082         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3083             if(USES_LIST(left_type[LTOP], list)){
3084                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3085                 const int b8_xy= 4*left_xy[LTOP] + 1;
3086                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3087                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3088                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3089                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3090                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3091                 ref_cache[-1 +  0]=
3092                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3093                 ref_cache[-1 + 16]=
3094                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3095             }else{
3096                 AV_ZERO32(mv_dst - 1 + 0);
3097                 AV_ZERO32(mv_dst - 1 + 8);
3098                 AV_ZERO32(mv_dst - 1 +16);
3099                 AV_ZERO32(mv_dst - 1 +24);
3100                 ref_cache[-1 +  0]=
3101                 ref_cache[-1 +  8]=
3102                 ref_cache[-1 + 16]=
3103                 ref_cache[-1 + 24]= LIST_NOT_USED;
3104             }
3105         }
3106     }
3107
3108     if(!USES_LIST(mb_type, list)){
3109         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3110         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3111         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3112         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3113         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3114         return;
3115     }
3116
3117     {
3118         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3119         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3120         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3121         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3122         AV_WN32A(&ref_cache[0*8], ref01);
3123         AV_WN32A(&ref_cache[1*8], ref01);
3124         AV_WN32A(&ref_cache[2*8], ref23);
3125         AV_WN32A(&ref_cache[3*8], ref23);
3126     }
3127
3128     {
3129         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3130         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3131         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3132         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3133         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3134     }
3135 }
3136
3137 /**
3138  *
3139  * @return non zero if the loop filter can be skiped
3140  */
3141 static int fill_filter_caches(H264Context *h, int mb_type){
3142     MpegEncContext * const s = &h->s;
3143     const int mb_xy= h->mb_xy;
3144     int top_xy, left_xy[LEFT_MBS];
3145     int top_type, left_type[LEFT_MBS];
3146     uint8_t *nnz;
3147     uint8_t *nnz_cache;
3148
3149     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3150
3151     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3152      * stuff, I can't imagine that these complex rules are worth it. */
3153
3154     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3155     if(FRAME_MBAFF){
3156         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3157         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3158         if(s->mb_y&1){
3159             if (left_mb_field_flag != curr_mb_field_flag) {
3160                 left_xy[LTOP] -= s->mb_stride;
3161             }
3162         }else{
3163             if(curr_mb_field_flag){
3164                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3165             }
3166             if (left_mb_field_flag != curr_mb_field_flag) {
3167                 left_xy[LBOT] += s->mb_stride;
3168             }
3169         }
3170     }
3171
3172     h->top_mb_xy = top_xy;
3173     h->left_mb_xy[LTOP] = left_xy[LTOP];
3174     h->left_mb_xy[LBOT] = left_xy[LBOT];
3175     {
3176         //for sufficiently low qp, filtering wouldn't do anything
3177         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3178         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3179         int qp = s->current_picture.f.qscale_table[mb_xy];
3180         if(qp <= qp_thresh
3181            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3182            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3183             if(!FRAME_MBAFF)
3184                 return 1;
3185             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3186                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3187                 return 1;
3188         }
3189     }
3190
3191     top_type        = s->current_picture.f.mb_type[top_xy];
3192     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3193     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3194     if(h->deblocking_filter == 2){
3195         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3196         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3197     }else{
3198         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3199         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3200     }
3201     h->top_type       = top_type;
3202     h->left_type[LTOP]= left_type[LTOP];
3203     h->left_type[LBOT]= left_type[LBOT];
3204
3205     if(IS_INTRA(mb_type))
3206         return 0;
3207
3208     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3209     if(h->list_count == 2)
3210         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3211
3212     nnz = h->non_zero_count[mb_xy];
3213     nnz_cache = h->non_zero_count_cache;
3214     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3215     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3216     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3217     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3218     h->cbp= h->cbp_table[mb_xy];
3219
3220     if(top_type){
3221         nnz = h->non_zero_count[top_xy];
3222         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3223     }
3224
3225     if(left_type[LTOP]){
3226         nnz = h->non_zero_count[left_xy[LTOP]];
3227         nnz_cache[3+8*1]= nnz[3+0*4];
3228         nnz_cache[3+8*2]= nnz[3+1*4];
3229         nnz_cache[3+8*3]= nnz[3+2*4];
3230         nnz_cache[3+8*4]= nnz[3+3*4];
3231     }
3232
3233     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3234     if(!CABAC && h->pps.transform_8x8_mode){
3235         if(IS_8x8DCT(top_type)){
3236             nnz_cache[4+8*0]=
3237             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3238             nnz_cache[6+8*0]=
3239             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3240         }
3241         if(IS_8x8DCT(left_type[LTOP])){
3242             nnz_cache[3+8*1]=
3243             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3244         }
3245         if(IS_8x8DCT(left_type[LBOT])){
3246             nnz_cache[3+8*3]=
3247             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3248         }
3249
3250         if(IS_8x8DCT(mb_type)){
3251             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3252             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3253
3254             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3255             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3256
3257             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3258             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3259
3260             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3261             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3262         }
3263     }
3264
3265     return 0;
3266 }
3267
3268 static void loop_filter(H264Context *h, int start_x, int end_x){
3269     MpegEncContext * const s = &h->s;
3270     uint8_t  *dest_y, *dest_cb, *dest_cr;
3271     int linesize, uvlinesize, mb_x, mb_y;
3272     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3273     const int old_slice_type= h->slice_type;
3274     const int pixel_shift = h->pixel_shift;
3275
3276     if(h->deblocking_filter) {
3277         for(mb_x= start_x; mb_x<end_x; mb_x++){
3278             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3279                 int mb_xy, mb_type;
3280                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3281                 h->slice_num= h->slice_table[mb_xy];
3282                 mb_type = s->current_picture.f.mb_type[mb_xy];
3283                 h->list_count= h->list_counts[mb_xy];
3284
3285                 if(FRAME_MBAFF)
3286                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3287
3288                 s->mb_x= mb_x;
3289                 s->mb_y= mb_y;
3290                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3291                 dest_cb = s->current_picture.f.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * (8 << CHROMA444);
3292                 dest_cr = s->current_picture.f.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * (8 << CHROMA444);
3293                     //FIXME simplify above
3294
3295                 if (MB_FIELD) {
3296                     linesize   = h->mb_linesize   = s->linesize * 2;
3297                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3298                     if(mb_y&1){ //FIXME move out of this function?
3299                         dest_y -= s->linesize*15;
3300                         dest_cb-= s->uvlinesize*((8 << CHROMA444)-1);
3301                         dest_cr-= s->uvlinesize*((8 << CHROMA444)-1);
3302                     }
3303                 } else {
3304                     linesize   = h->mb_linesize   = s->linesize;
3305                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3306                 }
3307                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, CHROMA444, 0);
3308                 if(fill_filter_caches(h, mb_type))
3309                     continue;
3310                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3311                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3312
3313                 if (FRAME_MBAFF) {
3314                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3315                 } else {
3316                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3317                 }
3318             }
3319         }
3320     }
3321     h->slice_type= old_slice_type;
3322     s->mb_x= end_x;
3323     s->mb_y= end_mb_y - FRAME_MBAFF;
3324     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3325     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3326 }
3327
3328 static void predict_field_decoding_flag(H264Context *h){
3329     MpegEncContext * const s = &h->s;
3330     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3331     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3332                 ? s->current_picture.f.mb_type[mb_xy - 1]
3333                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3334                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3335                 : 0;
3336     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3337 }
3338
3339 /**
3340  * Draw edges and report progress for the last MB row.
3341  */
3342 static void decode_finish_row(H264Context *h){
3343     MpegEncContext * const s = &h->s;
3344     int top = 16*(s->mb_y >> FIELD_PICTURE);
3345     int height = 16 << FRAME_MBAFF;
3346     int deblock_border = (16 + 4) << FRAME_MBAFF;
3347     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3348
3349     if (h->deblocking_filter) {
3350         if((top + height) >= pic_height)
3351             height += deblock_border;
3352
3353         top -= deblock_border;
3354     }
3355
3356     if (top >= pic_height || (top + height) < h->emu_edge_height)
3357         return;
3358
3359     height = FFMIN(height, pic_height - top);
3360     if (top < h->emu_edge_height) {
3361         height = top+height;
3362         top = 0;
3363     }
3364
3365     ff_draw_horiz_band(s, top, height);
3366
3367     if (s->dropable) return;
3368
3369     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3370                              s->picture_structure==PICT_BOTTOM_FIELD);
3371 }
3372
3373 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3374     H264Context *h = *(void**)arg;
3375     MpegEncContext * const s = &h->s;
3376     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3377     int lf_x_start = s->mb_x;
3378
3379     s->mb_skip_run= -1;
3380
3381     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3382                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3383
3384     if( h->pps.cabac ) {
3385         /* realign */
3386         align_get_bits( &s->gb );
3387
3388         /* init cabac */
3389         ff_init_cabac_states( &h->cabac);
3390         ff_init_cabac_decoder( &h->cabac,
3391                                s->gb.buffer + get_bits_count(&s->gb)/8,
3392                                (get_bits_left(&s->gb) + 7)/8);
3393
3394         ff_h264_init_cabac_states(h);
3395
3396         for(;;){
3397 //START_TIMER
3398             int ret = ff_h264_decode_mb_cabac(h);
3399             int eos;
3400 //STOP_TIMER("decode_mb_cabac")
3401
3402             if(ret>=0) ff_h264_hl_decode_mb(h);
3403
3404             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3405                 s->mb_y++;
3406
3407                 ret = ff_h264_decode_mb_cabac(h);
3408
3409                 if(ret>=0) ff_h264_hl_decode_mb(h);
3410                 s->mb_y--;
3411             }
3412             eos = get_cabac_terminate( &h->cabac );
3413
3414             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3415                 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);
3416                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3417                 return 0;
3418             }
3419             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3420                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d, bytestream (%td)\n", s->mb_x, s->mb_y, h->cabac.bytestream_end - h->cabac.bytestream);
3421                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3422                 return -1;
3423             }
3424
3425             if( ++s->mb_x >= s->mb_width ) {
3426                 loop_filter(h, lf_x_start, s->mb_x);
3427                 s->mb_x = lf_x_start = 0;
3428                 decode_finish_row(h);
3429                 ++s->mb_y;
3430                 if(FIELD_OR_MBAFF_PICTURE) {
3431                     ++s->mb_y;
3432                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3433                         predict_field_decoding_flag(h);
3434                 }
3435             }
3436
3437             if( eos || s->mb_y >= s->mb_height ) {
3438                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3439                 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);
3440                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3441                 return 0;
3442             }
3443         }
3444
3445     } else {
3446         for(;;){
3447             int ret = ff_h264_decode_mb_cavlc(h);
3448
3449             if(ret>=0) ff_h264_hl_decode_mb(h);
3450
3451             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3452                 s->mb_y++;
3453                 ret = ff_h264_decode_mb_cavlc(h);
3454
3455                 if(ret>=0) ff_h264_hl_decode_mb(h);
3456                 s->mb_y--;
3457             }
3458
3459             if(ret<0){
3460                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3461                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3462                 return -1;
3463             }
3464
3465             if(++s->mb_x >= s->mb_width){
3466                 loop_filter(h, lf_x_start, s->mb_x);
3467                 s->mb_x = lf_x_start = 0;
3468                 decode_finish_row(h);
3469                 ++s->mb_y;
3470                 if(FIELD_OR_MBAFF_PICTURE) {
3471                     ++s->mb_y;
3472                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3473                         predict_field_decoding_flag(h);
3474                 }
3475                 if(s->mb_y >= s->mb_height){
3476                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3477
3478                     if(get_bits_count(&s->gb) == s->gb.size_in_bits ) {
3479                         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);
3480
3481                         return 0;
3482                     }else{
3483                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3484
3485                         return -1;
3486                     }
3487                 }
3488             }
3489
3490             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
3491                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3492                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
3493                     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);
3494                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3495
3496                     return 0;
3497                 }else{
3498                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3499
3500                     return -1;
3501                 }
3502             }
3503         }
3504     }
3505 }
3506
3507 /**
3508  * Call decode_slice() for each context.
3509  *
3510  * @param h h264 master context
3511  * @param context_count number of contexts to execute
3512  */
3513 static int execute_decode_slices(H264Context *h, int context_count){
3514     MpegEncContext * const s = &h->s;
3515     AVCodecContext * const avctx= s->avctx;
3516     H264Context *hx;
3517     int i;
3518
3519     if (s->avctx->hwaccel || s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3520         return 0;
3521     if(context_count == 1) {
3522         return decode_slice(avctx, &h);
3523     } else {
3524         for(i = 1; i < context_count; i++) {
3525             hx = h->thread_context[i];
3526             hx->s.error_recognition = avctx->error_recognition;
3527             hx->s.error_count = 0;
3528         }
3529
3530         avctx->execute(avctx, (void *)decode_slice,
3531                        h->thread_context, NULL, context_count, sizeof(void*));
3532
3533         /* pull back stuff from slices to master context */
3534         hx = h->thread_context[context_count - 1];
3535         s->mb_x = hx->s.mb_x;
3536         s->mb_y = hx->s.mb_y;
3537         s->dropable = hx->s.dropable;
3538         s->picture_structure = hx->s.picture_structure;
3539         for(i = 1; i < context_count; i++)
3540             h->s.error_count += h->thread_context[i]->s.error_count;
3541     }
3542
3543     return 0;
3544 }
3545
3546
3547 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
3548     MpegEncContext * const s = &h->s;
3549     AVCodecContext * const avctx= s->avctx;
3550     H264Context *hx; ///< thread context
3551     int buf_index;
3552     int context_count;
3553     int next_avc;
3554     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3555     int nals_needed=0; ///< number of NALs that need decoding before the next frame thread starts
3556     int nal_index;
3557
3558     h->max_contexts = (HAVE_THREADS && (s->avctx->active_thread_type&FF_THREAD_SLICE)) ? avctx->thread_count : 1;
3559     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
3560         h->current_slice = 0;
3561         if (!s->first_field)
3562             s->current_picture_ptr= NULL;
3563         ff_h264_reset_sei(h);
3564     }
3565
3566     for(;pass <= 1;pass++){
3567         buf_index = 0;
3568         context_count = 0;
3569         next_avc = h->is_avc ? 0 : buf_size;
3570         nal_index = 0;
3571     for(;;){
3572         int consumed;
3573         int dst_length;
3574         int bit_length;
3575         const uint8_t *ptr;
3576         int i, nalsize = 0;
3577         int err;
3578
3579         if(buf_index >= next_avc) {
3580             if(buf_index >= buf_size) break;
3581             nalsize = 0;
3582             for(i = 0; i < h->nal_length_size; i++)
3583                 nalsize = (nalsize << 8) | buf[buf_index++];
3584             if(nalsize <= 0 || nalsize > buf_size - buf_index){
3585                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
3586                 break;
3587             }
3588             next_avc= buf_index + nalsize;
3589         } else {
3590             // start code prefix search
3591             for(; buf_index + 3 < next_avc; buf_index++){
3592                 // This should always succeed in the first iteration.
3593                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
3594                     break;
3595             }
3596
3597             if(buf_index+3 >= buf_size) break;
3598
3599             buf_index+=3;
3600             if(buf_index >= next_avc) continue;
3601         }
3602
3603         hx = h->thread_context[context_count];
3604
3605         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
3606         if (ptr==NULL || dst_length < 0){
3607             return -1;
3608         }
3609         i= buf_index + consumed;
3610         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
3611            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
3612             s->workaround_bugs |= FF_BUG_TRUNCATED;
3613
3614         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
3615         while(ptr[dst_length - 1] == 0 && dst_length > 0)
3616             dst_length--;
3617         }
3618         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
3619
3620         if(s->avctx->debug&FF_DEBUG_STARTCODE){
3621             av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d at %d/%d length %d\n", hx->nal_unit_type, buf_index, buf_size, dst_length);
3622         }
3623
3624         if (h->is_avc && (nalsize != consumed) && nalsize){
3625             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
3626         }
3627
3628         buf_index += consumed;
3629         nal_index++;
3630
3631         if(pass == 0) {
3632             // packets can sometimes contain multiple PPS/SPS
3633             // e.g. two PAFF field pictures in one packet, or a demuxer which splits NALs strangely
3634             // if so, when frame threading we can't start the next thread until we've read all of them
3635             switch (hx->nal_unit_type) {
3636                 case NAL_SPS:
3637                 case NAL_PPS:
3638                     nals_needed = nal_index;
3639                     break;
3640                 case NAL_IDR_SLICE:
3641                 case NAL_SLICE:
3642                     init_get_bits(&hx->s.gb, ptr, bit_length);
3643                     if (!get_ue_golomb(&hx->s.gb))
3644                         nals_needed = nal_index;
3645             }
3646             continue;
3647         }
3648
3649         //FIXME do not discard SEI id
3650         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
3651             continue;
3652
3653       again:
3654         err = 0;
3655         switch(hx->nal_unit_type){
3656         case NAL_IDR_SLICE:
3657             if (h->nal_unit_type != NAL_IDR_SLICE) {
3658                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
3659                 return -1;
3660             }
3661             idr(h); //FIXME ensure we don't loose some frames if there is reordering
3662         case NAL_SLICE:
3663             init_get_bits(&hx->s.gb, ptr, bit_length);
3664             hx->intra_gb_ptr=
3665             hx->inter_gb_ptr= &hx->s.gb;
3666             hx->s.data_partitioning = 0;
3667
3668             if((err = decode_slice_header(hx, h)))
3669                break;
3670
3671             s->current_picture_ptr->f.key_frame |=
3672                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
3673                     (h->sei_recovery_frame_cnt >= 0);
3674
3675             if (h->current_slice == 1) {
3676                 if(!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3677                     decode_postinit(h, nal_index >= nals_needed);
3678                 }
3679
3680                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3681                     return -1;
3682                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3683                     ff_vdpau_h264_picture_start(s);
3684             }
3685
3686             if(hx->redundant_pic_count==0
3687                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3688                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3689                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3690                && avctx->skip_frame < AVDISCARD_ALL){
3691                 if(avctx->hwaccel) {
3692                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
3693                         return -1;
3694                 }else
3695                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
3696                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
3697                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
3698                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
3699                 }else
3700                     context_count++;
3701             }
3702             break;
3703         case NAL_DPA:
3704             init_get_bits(&hx->s.gb, ptr, bit_length);
3705             hx->intra_gb_ptr=
3706             hx->inter_gb_ptr= NULL;
3707
3708             if ((err = decode_slice_header(hx, h)) < 0)
3709                 break;
3710
3711             hx->s.data_partitioning = 1;
3712
3713             break;
3714         case NAL_DPB:
3715             init_get_bits(&hx->intra_gb, ptr, bit_length);
3716             hx->intra_gb_ptr= &hx->intra_gb;
3717             break;
3718         case NAL_DPC:
3719             init_get_bits(&hx->inter_gb, ptr, bit_length);
3720             hx->inter_gb_ptr= &hx->inter_gb;
3721
3722             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
3723                && s->context_initialized
3724                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3725                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3726                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3727                && avctx->skip_frame < AVDISCARD_ALL)
3728                 context_count++;
3729             break;
3730         case NAL_SEI:
3731             init_get_bits(&s->gb, ptr, bit_length);
3732             ff_h264_decode_sei(h);
3733             break;
3734         case NAL_SPS:
3735             init_get_bits(&s->gb, ptr, bit_length);
3736             ff_h264_decode_seq_parameter_set(h);
3737
3738             if (s->flags& CODEC_FLAG_LOW_DELAY ||
3739                 (h->sps.bitstream_restriction_flag && !h->sps.num_reorder_frames))
3740                 s->low_delay=1;
3741
3742             if(avctx->has_b_frames < 2)
3743                 avctx->has_b_frames= !s->low_delay;
3744
3745             if (avctx->bits_per_raw_sample != h->sps.bit_depth_luma) {
3746                 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
3747                     avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3748                     h->pixel_shift = h->sps.bit_depth_luma > 8;
3749
3750                     ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma);
3751                     ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma);
3752                     s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
3753                     dsputil_init(&s->dsp, s->avctx);
3754                 } else {
3755                     av_log(avctx, AV_LOG_ERROR, "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3756                     return -1;
3757                 }
3758             }
3759             break;
3760         case NAL_PPS:
3761             init_get_bits(&s->gb, ptr, bit_length);
3762
3763             ff_h264_decode_picture_parameter_set(h, bit_length);
3764
3765             break;
3766         case NAL_AUD:
3767         case NAL_END_SEQUENCE:
3768         case NAL_END_STREAM:
3769         case NAL_FILLER_DATA:
3770         case NAL_SPS_EXT:
3771         case NAL_AUXILIARY_SLICE:
3772             break;
3773         default:
3774             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
3775         }
3776
3777         if(context_count == h->max_contexts) {
3778             execute_decode_slices(h, context_count);
3779             context_count = 0;
3780         }
3781
3782         if (err < 0)
3783             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3784         else if(err == 1) {
3785             /* Slice could not be decoded in parallel mode, copy down
3786              * NAL unit stuff to context 0 and restart. Note that
3787              * rbsp_buffer is not transferred, but since we no longer
3788              * run in parallel mode this should not be an issue. */
3789             h->nal_unit_type = hx->nal_unit_type;
3790             h->nal_ref_idc   = hx->nal_ref_idc;
3791             hx = h;
3792             goto again;
3793         }
3794     }
3795     }
3796     if(context_count)
3797         execute_decode_slices(h, context_count);
3798     return buf_index;
3799 }
3800
3801 /**
3802  * returns the number of bytes consumed for building the current frame
3803  */
3804 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
3805         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
3806         if(pos+10>buf_size) pos=buf_size; // oops ;)
3807
3808         return pos;
3809 }
3810
3811 static int decode_frame(AVCodecContext *avctx,
3812                              void *data, int *data_size,
3813                              AVPacket *avpkt)
3814 {
3815     const uint8_t *buf = avpkt->data;
3816     int buf_size = avpkt->size;
3817     H264Context *h = avctx->priv_data;
3818     MpegEncContext *s = &h->s;
3819     AVFrame *pict = data;
3820     int buf_index;
3821
3822     s->flags= avctx->flags;
3823     s->flags2= avctx->flags2;
3824
3825    /* end of stream, output what is still in the buffers */
3826  out:
3827     if (buf_size == 0) {
3828         Picture *out;
3829         int i, out_idx;
3830
3831         s->current_picture_ptr = NULL;
3832
3833 //FIXME factorize this with the output code below
3834         out = h->delayed_pic[0];
3835         out_idx = 0;
3836         for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3837             if(h->delayed_pic[i]->poc < out->poc){
3838                 out = h->delayed_pic[i];
3839                 out_idx = i;
3840             }
3841
3842         for(i=out_idx; h->delayed_pic[i]; i++)
3843             h->delayed_pic[i] = h->delayed_pic[i+1];
3844
3845         if(out){
3846             *data_size = sizeof(AVFrame);
3847             *pict= *(AVFrame*)out;
3848         }
3849
3850         return 0;
3851     }
3852
3853     buf_index=decode_nal_units(h, buf, buf_size);
3854     if(buf_index < 0)
3855         return -1;
3856
3857     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
3858         buf_size = 0;
3859         goto out;
3860     }
3861
3862     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
3863         if (avctx->skip_frame >= AVDISCARD_NONREF)
3864             return 0;
3865         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
3866         return -1;
3867     }
3868
3869     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
3870
3871         if(s->flags2 & CODEC_FLAG2_CHUNKS) decode_postinit(h, 1);
3872
3873         field_end(h, 0);
3874
3875         if (!h->next_output_pic) {
3876             /* Wait for second field. */
3877             *data_size = 0;
3878
3879         } else {
3880             *data_size = sizeof(AVFrame);
3881             *pict = *(AVFrame*)h->next_output_pic;
3882         }
3883     }
3884
3885     assert(pict->data[0] || !*data_size);
3886     ff_print_debug_info(s, pict);
3887 //printf("out %d\n", (int)pict->data[0]);
3888
3889     return get_consumed_bytes(s, buf_index, buf_size);
3890 }
3891 #if 0
3892 static inline void fill_mb_avail(H264Context *h){
3893     MpegEncContext * const s = &h->s;
3894     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3895
3896     if(s->mb_y){
3897         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
3898         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
3899         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
3900     }else{
3901         h->mb_avail[0]=
3902         h->mb_avail[1]=
3903         h->mb_avail[2]= 0;
3904     }
3905     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
3906     h->mb_avail[4]= 1; //FIXME move out
3907     h->mb_avail[5]= 0; //FIXME move out
3908 }
3909 #endif
3910
3911 #ifdef TEST
3912 #undef printf
3913 #undef random
3914 #define COUNT 8000
3915 #define SIZE (COUNT*40)
3916 int main(void){
3917     int i;
3918     uint8_t temp[SIZE];
3919     PutBitContext pb;
3920     GetBitContext gb;
3921 //    int int_temp[10000];
3922     DSPContext dsp;
3923     AVCodecContext avctx;
3924
3925     dsputil_init(&dsp, &avctx);
3926
3927     init_put_bits(&pb, temp, SIZE);
3928     printf("testing unsigned exp golomb\n");
3929     for(i=0; i<COUNT; i++){
3930         START_TIMER
3931         set_ue_golomb(&pb, i);
3932         STOP_TIMER("set_ue_golomb");
3933     }
3934     flush_put_bits(&pb);
3935
3936     init_get_bits(&gb, temp, 8*SIZE);
3937     for(i=0; i<COUNT; i++){
3938         int j, s;
3939
3940         s= show_bits(&gb, 24);
3941
3942         START_TIMER
3943         j= get_ue_golomb(&gb);
3944         if(j != i){
3945             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3946 //            return -1;
3947         }
3948         STOP_TIMER("get_ue_golomb");
3949     }
3950
3951
3952     init_put_bits(&pb, temp, SIZE);
3953     printf("testing signed exp golomb\n");
3954     for(i=0; i<COUNT; i++){
3955         START_TIMER
3956         set_se_golomb(&pb, i - COUNT/2);
3957         STOP_TIMER("set_se_golomb");
3958     }
3959     flush_put_bits(&pb);
3960
3961     init_get_bits(&gb, temp, 8*SIZE);
3962     for(i=0; i<COUNT; i++){
3963         int j, s;
3964
3965         s= show_bits(&gb, 24);
3966
3967         START_TIMER
3968         j= get_se_golomb(&gb);
3969         if(j != i - COUNT/2){
3970             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3971 //            return -1;
3972         }
3973         STOP_TIMER("get_se_golomb");
3974     }
3975
3976     printf("Testing RBSP\n");
3977
3978
3979     return 0;
3980 }
3981 #endif /* TEST */
3982
3983
3984 av_cold void ff_h264_free_context(H264Context *h)
3985 {
3986     int i;
3987
3988     free_tables(h, 1); //FIXME cleanup init stuff perhaps
3989
3990     for(i = 0; i < MAX_SPS_COUNT; i++)
3991         av_freep(h->sps_buffers + i);
3992
3993     for(i = 0; i < MAX_PPS_COUNT; i++)
3994         av_freep(h->pps_buffers + i);
3995 }
3996
3997 av_cold int ff_h264_decode_end(AVCodecContext *avctx)
3998 {
3999     H264Context *h = avctx->priv_data;
4000     MpegEncContext *s = &h->s;
4001
4002     ff_h264_free_context(h);
4003
4004     MPV_common_end(s);
4005
4006 //    memset(h, 0, sizeof(H264Context));
4007
4008     return 0;
4009 }
4010
4011 static const AVProfile profiles[] = {
4012     { FF_PROFILE_H264_BASELINE,             "Baseline"              },
4013     { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline"  },
4014     { FF_PROFILE_H264_MAIN,                 "Main"                  },
4015     { FF_PROFILE_H264_EXTENDED,             "Extended"              },
4016     { FF_PROFILE_H264_HIGH,                 "High"                  },
4017     { FF_PROFILE_H264_HIGH_10,              "High 10"               },
4018     { FF_PROFILE_H264_HIGH_10_INTRA,        "High 10 Intra"         },
4019     { FF_PROFILE_H264_HIGH_422,             "High 4:2:2"            },
4020     { FF_PROFILE_H264_HIGH_422_INTRA,       "High 4:2:2 Intra"      },
4021     { FF_PROFILE_H264_HIGH_444,             "High 4:4:4"            },
4022     { FF_PROFILE_H264_HIGH_444_PREDICTIVE,  "High 4:4:4 Predictive" },
4023     { FF_PROFILE_H264_HIGH_444_INTRA,       "High 4:4:4 Intra"      },
4024     { FF_PROFILE_H264_CAVLC_444,            "CAVLC 4:4:4"           },
4025     { FF_PROFILE_UNKNOWN },
4026 };
4027
4028 AVCodec ff_h264_decoder = {
4029     .name           = "h264",
4030     .type           = AVMEDIA_TYPE_VIDEO,
4031     .id             = CODEC_ID_H264,
4032     .priv_data_size = sizeof(H264Context),
4033     .init           = ff_h264_decode_init,
4034     .close          = ff_h264_decode_end,
4035     .decode         = decode_frame,
4036     .capabilities   = /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 | CODEC_CAP_DELAY |
4037                       CODEC_CAP_SLICE_THREADS | CODEC_CAP_FRAME_THREADS,
4038     .flush= flush_dpb,
4039     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4040     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4041     .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4042     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4043 };
4044
4045 #if CONFIG_H264_VDPAU_DECODER
4046 AVCodec ff_h264_vdpau_decoder = {
4047     .name           = "h264_vdpau",
4048     .type           = AVMEDIA_TYPE_VIDEO,
4049     .id             = CODEC_ID_H264,
4050     .priv_data_size = sizeof(H264Context),
4051     .init           = ff_h264_decode_init,
4052     .close          = ff_h264_decode_end,
4053     .decode         = decode_frame,
4054     .capabilities   = CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4055     .flush= flush_dpb,
4056     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4057     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_VDPAU_H264, PIX_FMT_NONE},
4058     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4059 };
4060 #endif
</