H.264: faster write_back_*
[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->thread_opaque != s->current_picture.thread_opaque ||
265            (ref->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->thread_opaque != s->current_picture.thread_opaque ||
277            (ref->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.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->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->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->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->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->reference - 1));
508         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
509     }
510     src_cb= pic->data[1] + ((mx>>3) << pixel_shift) + (my>>3)*h->mb_uvlinesize;
511     src_cr= pic->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].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.mb_type[mb_xy];
688
689     assert(IS_INTER(mb_type));
690
691     if(HAVE_PTHREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
692         await_references(h);
693     prefetch_motion(h, 0, pixel_shift, chroma444);
694
695     if(IS_16X16(mb_type)){
696         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
697                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
698                 weight_op, weight_avg,
699                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
700                 pixel_shift, chroma444);
701     }else if(IS_16X8(mb_type)){
702         mc_part(h, 0, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
703                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
704                 &weight_op[1], &weight_avg[1],
705                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
706                 pixel_shift, chroma444);
707         mc_part(h, 8, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
708                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
709                 &weight_op[1], &weight_avg[1],
710                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
711                 pixel_shift, chroma444);
712     }else if(IS_8X16(mb_type)){
713         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
714                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
715                 &weight_op[2], &weight_avg[2],
716                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
717                 pixel_shift, chroma444);
718         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
719                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
720                 &weight_op[2], &weight_avg[2],
721                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
722                 pixel_shift, chroma444);
723     }else{
724         int i;
725
726         assert(IS_8X8(mb_type));
727
728         for(i=0; i<4; i++){
729             const int sub_mb_type= h->sub_mb_type[i];
730             const int n= 4*i;
731             int x_offset= (i&1)<<2;
732             int y_offset= (i&2)<<1;
733
734             if(IS_SUB_8X8(sub_mb_type)){
735                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
736                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
737                     &weight_op[3], &weight_avg[3],
738                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
739                     pixel_shift, chroma444);
740             }else if(IS_SUB_8X4(sub_mb_type)){
741                 mc_part(h, n  , 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
742                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
743                     &weight_op[4], &weight_avg[4],
744                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
745                     pixel_shift, chroma444);
746                 mc_part(h, n+2, 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
747                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
748                     &weight_op[4], &weight_avg[4],
749                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
750                     pixel_shift, chroma444);
751             }else if(IS_SUB_4X8(sub_mb_type)){
752                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
753                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
754                     &weight_op[5], &weight_avg[5],
755                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
756                     pixel_shift, chroma444);
757                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
758                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
759                     &weight_op[5], &weight_avg[5],
760                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
761                     pixel_shift, chroma444);
762             }else{
763                 int j;
764                 assert(IS_SUB_4X4(sub_mb_type));
765                 for(j=0; j<4; j++){
766                     int sub_x_offset= x_offset + 2*(j&1);
767                     int sub_y_offset= y_offset +   (j&2);
768                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
769                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
770                         &weight_op[6], &weight_avg[6],
771                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
772                         pixel_shift, chroma444);
773                 }
774             }
775         }
776     }
777
778     prefetch_motion(h, 1, pixel_shift, chroma444);
779 }
780
781 #define hl_motion_fn(sh, bits) \
782 static av_always_inline void hl_motion_ ## bits(H264Context *h, \
783                                        uint8_t *dest_y, \
784                                        uint8_t *dest_cb, uint8_t *dest_cr, \
785                                        qpel_mc_func (*qpix_put)[16], \
786                                        h264_chroma_mc_func (*chroma_put), \
787                                        qpel_mc_func (*qpix_avg)[16], \
788                                        h264_chroma_mc_func (*chroma_avg), \
789                                        h264_weight_func *weight_op, \
790                                        h264_biweight_func *weight_avg, \
791                                        int chroma444) \
792 { \
793     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put, \
794               qpix_avg, chroma_avg, weight_op, weight_avg, sh, chroma444); \
795 }
796 hl_motion_fn(0, 8);
797 hl_motion_fn(1, 16);
798
799 static void free_tables(H264Context *h, int free_rbsp){
800     int i;
801     H264Context *hx;
802
803     av_freep(&h->intra4x4_pred_mode);
804     av_freep(&h->chroma_pred_mode_table);
805     av_freep(&h->cbp_table);
806     av_freep(&h->mvd_table[0]);
807     av_freep(&h->mvd_table[1]);
808     av_freep(&h->direct_table);
809     av_freep(&h->non_zero_count);
810     av_freep(&h->slice_table_base);
811     h->slice_table= NULL;
812     av_freep(&h->list_counts);
813
814     av_freep(&h->mb2b_xy);
815     av_freep(&h->mb2br_xy);
816
817     for(i = 0; i < MAX_THREADS; i++) {
818         hx = h->thread_context[i];
819         if(!hx) continue;
820         av_freep(&hx->top_borders[1]);
821         av_freep(&hx->top_borders[0]);
822         av_freep(&hx->s.obmc_scratchpad);
823         if (free_rbsp){
824             av_freep(&hx->rbsp_buffer[1]);
825             av_freep(&hx->rbsp_buffer[0]);
826             hx->rbsp_buffer_size[0] = 0;
827             hx->rbsp_buffer_size[1] = 0;
828         }
829         if (i) av_freep(&h->thread_context[i]);
830     }
831 }
832
833 static void init_dequant8_coeff_table(H264Context *h){
834     int i,j,q,x;
835     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
836
837     for(i=0; i<6; i++ ){
838         h->dequant8_coeff[i] = h->dequant8_buffer[i];
839         for(j=0; j<i; j++){
840             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
841                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
842                 break;
843             }
844         }
845         if(j<i)
846             continue;
847
848         for(q=0; q<max_qp+1; q++){
849             int shift = div6[q];
850             int idx = rem6[q];
851             for(x=0; x<64; x++)
852                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
853                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
854                     h->pps.scaling_matrix8[i][x]) << shift;
855         }
856     }
857 }
858
859 static void init_dequant4_coeff_table(H264Context *h){
860     int i,j,q,x;
861     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
862     for(i=0; i<6; i++ ){
863         h->dequant4_coeff[i] = h->dequant4_buffer[i];
864         for(j=0; j<i; j++){
865             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
866                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
867                 break;
868             }
869         }
870         if(j<i)
871             continue;
872
873         for(q=0; q<max_qp+1; q++){
874             int shift = div6[q] + 2;
875             int idx = rem6[q];
876             for(x=0; x<16; x++)
877                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
878                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
879                     h->pps.scaling_matrix4[i][x]) << shift;
880         }
881     }
882 }
883
884 static void init_dequant_tables(H264Context *h){
885     int i,x;
886     init_dequant4_coeff_table(h);
887     if(h->pps.transform_8x8_mode)
888         init_dequant8_coeff_table(h);
889     if(h->sps.transform_bypass){
890         for(i=0; i<6; i++)
891             for(x=0; x<16; x++)
892                 h->dequant4_coeff[i][0][x] = 1<<6;
893         if(h->pps.transform_8x8_mode)
894             for(i=0; i<6; i++)
895                 for(x=0; x<64; x++)
896                     h->dequant8_coeff[i][0][x] = 1<<6;
897     }
898 }
899
900
901 int ff_h264_alloc_tables(H264Context *h){
902     MpegEncContext * const s = &h->s;
903     const int big_mb_num= s->mb_stride * (s->mb_height+1);
904     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
905     int x,y;
906
907     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
908
909     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
910     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
911     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
912
913     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
914     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
915     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
916     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
917     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
918
919     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
920     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
921
922     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
923     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
924     for(y=0; y<s->mb_height; y++){
925         for(x=0; x<s->mb_width; x++){
926             const int mb_xy= x + y*s->mb_stride;
927             const int b_xy = 4*x + 4*y*h->b_stride;
928
929             h->mb2b_xy [mb_xy]= b_xy;
930             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
931         }
932     }
933
934     s->obmc_scratchpad = NULL;
935
936     if(!h->dequant4_coeff[0])
937         init_dequant_tables(h);
938
939     return 0;
940 fail:
941     free_tables(h, 1);
942     return -1;
943 }
944
945 /**
946  * Mimic alloc_tables(), but for every context thread.
947  */
948 static void clone_tables(H264Context *dst, H264Context *src, int i){
949     MpegEncContext * const s = &src->s;
950     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
951     dst->non_zero_count           = src->non_zero_count;
952     dst->slice_table              = src->slice_table;
953     dst->cbp_table                = src->cbp_table;
954     dst->mb2b_xy                  = src->mb2b_xy;
955     dst->mb2br_xy                 = src->mb2br_xy;
956     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
957     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
958     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
959     dst->direct_table             = src->direct_table;
960     dst->list_counts              = src->list_counts;
961
962     dst->s.obmc_scratchpad = NULL;
963     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma);
964 }
965
966 /**
967  * Init context
968  * Allocate buffers which are not shared amongst multiple threads.
969  */
970 static int context_init(H264Context *h){
971     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
972     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
973
974     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
975     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
976
977     return 0;
978 fail:
979     return -1; // free_tables will clean up for us
980 }
981
982 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
983
984 static av_cold void common_init(H264Context *h){
985     MpegEncContext * const s = &h->s;
986
987     s->width = s->avctx->width;
988     s->height = s->avctx->height;
989     s->codec_id= s->avctx->codec->id;
990
991     ff_h264dsp_init(&h->h264dsp, 8);
992     ff_h264_pred_init(&h->hpc, s->codec_id, 8);
993
994     h->dequant_coeff_pps= -1;
995     s->unrestricted_mv=1;
996     s->decode=1; //FIXME
997
998     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
999
1000     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
1001     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
1002 }
1003
1004 int ff_h264_decode_extradata(H264Context *h)
1005 {
1006     AVCodecContext *avctx = h->s.avctx;
1007
1008     if(avctx->extradata[0] == 1){
1009         int i, cnt, nalsize;
1010         unsigned char *p = avctx->extradata;
1011
1012         h->is_avc = 1;
1013
1014         if(avctx->extradata_size < 7) {
1015             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1016             return -1;
1017         }
1018         /* sps and pps in the avcC always have length coded with 2 bytes,
1019            so put a fake nal_length_size = 2 while parsing them */
1020         h->nal_length_size = 2;
1021         // Decode sps from avcC
1022         cnt = *(p+5) & 0x1f; // Number of sps
1023         p += 6;
1024         for (i = 0; i < cnt; i++) {
1025             nalsize = AV_RB16(p) + 2;
1026             if(decode_nal_units(h, p, nalsize) < 0) {
1027                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1028                 return -1;
1029             }
1030             p += nalsize;
1031         }
1032         // Decode pps from avcC
1033         cnt = *(p++); // Number of pps
1034         for (i = 0; i < cnt; i++) {
1035             nalsize = AV_RB16(p) + 2;
1036             if (decode_nal_units(h, p, nalsize) < 0) {
1037                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1038                 return -1;
1039             }
1040             p += nalsize;
1041         }
1042         // Now store right nal length size, that will be use to parse all other nals
1043         h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1044     } else {
1045         h->is_avc = 0;
1046         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1047             return -1;
1048     }
1049     return 0;
1050 }
1051
1052 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1053     H264Context *h= avctx->priv_data;
1054     MpegEncContext * const s = &h->s;
1055
1056     MPV_decode_defaults(s);
1057
1058     s->avctx = avctx;
1059     common_init(h);
1060
1061     s->out_format = FMT_H264;
1062     s->workaround_bugs= avctx->workaround_bugs;
1063
1064     // set defaults
1065 //    s->decode_mb= ff_h263_decode_mb;
1066     s->quarter_sample = 1;
1067     if(!avctx->has_b_frames)
1068     s->low_delay= 1;
1069
1070     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1071
1072     ff_h264_decode_init_vlc();
1073
1074     h->pixel_shift = 0;
1075     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1076
1077     h->thread_context[0] = h;
1078     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1079     h->prev_poc_msb= 1<<16;
1080     h->x264_build = -1;
1081     ff_h264_reset_sei(h);
1082     if(avctx->codec_id == CODEC_ID_H264){
1083         if(avctx->ticks_per_frame == 1){
1084             s->avctx->time_base.den *=2;
1085         }
1086         avctx->ticks_per_frame = 2;
1087     }
1088
1089     if(avctx->extradata_size > 0 && avctx->extradata &&
1090         ff_h264_decode_extradata(h))
1091         return -1;
1092
1093     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1094         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1095         s->low_delay = 0;
1096     }
1097
1098     return 0;
1099 }
1100
1101 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1102 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1103 {
1104     int i;
1105
1106     for (i=0; i<count; i++){
1107         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1108                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1109                 !from[i]));
1110         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1111     }
1112 }
1113
1114 static void copy_parameter_set(void **to, void **from, int count, int size)
1115 {
1116     int i;
1117
1118     for (i=0; i<count; i++){
1119         if (to[i] && !from[i]) av_freep(&to[i]);
1120         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1121
1122         if (from[i]) memcpy(to[i], from[i], size);
1123     }
1124 }
1125
1126 static int decode_init_thread_copy(AVCodecContext *avctx){
1127     H264Context *h= avctx->priv_data;
1128
1129     if (!avctx->is_copy) return 0;
1130     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1131     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1132
1133     return 0;
1134 }
1135
1136 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1137 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1138     H264Context *h= dst->priv_data, *h1= src->priv_data;
1139     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1140     int inited = s->context_initialized, err;
1141     int i;
1142
1143     if(dst == src || !s1->context_initialized) return 0;
1144
1145     err = ff_mpeg_update_thread_context(dst, src);
1146     if(err) return err;
1147
1148     //FIXME handle width/height changing
1149     if(!inited){
1150         for(i = 0; i < MAX_SPS_COUNT; i++)
1151             av_freep(h->sps_buffers + i);
1152
1153         for(i = 0; i < MAX_PPS_COUNT; i++)
1154             av_freep(h->pps_buffers + i);
1155
1156         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1157         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1158         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1159         ff_h264_alloc_tables(h);
1160         context_init(h);
1161
1162         for(i=0; i<2; i++){
1163             h->rbsp_buffer[i] = NULL;
1164             h->rbsp_buffer_size[i] = 0;
1165         }
1166
1167         h->thread_context[0] = h;
1168
1169         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1170         // so this has to be allocated here
1171         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1172
1173         s->dsp.clear_blocks(h->mb);
1174         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1175     }
1176
1177     //extradata/NAL handling
1178     h->is_avc          = h1->is_avc;
1179
1180     //SPS/PPS
1181     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1182     h->sps             = h1->sps;
1183     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1184     h->pps             = h1->pps;
1185
1186     //Dequantization matrices
1187     //FIXME these are big - can they be only copied when PPS changes?
1188     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1189
1190     for(i=0; i<6; i++)
1191         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1192
1193     for(i=0; i<6; i++)
1194         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1195
1196     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1197
1198     //POC timing
1199     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1200
1201     //reference lists
1202     copy_fields(h, h1, ref_count, list_count);
1203     copy_fields(h, h1, ref_list,  intra_gb);
1204     copy_fields(h, h1, short_ref, cabac_init_idc);
1205
1206     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1207     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1208     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1209
1210     h->last_slice_type = h1->last_slice_type;
1211
1212     if(!s->current_picture_ptr) return 0;
1213
1214     if(!s->dropable) {
1215         ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1216         h->prev_poc_msb     = h->poc_msb;
1217         h->prev_poc_lsb     = h->poc_lsb;
1218     }
1219     h->prev_frame_num_offset= h->frame_num_offset;
1220     h->prev_frame_num       = h->frame_num;
1221     h->outputed_poc         = h->next_outputed_poc;
1222
1223     return 0;
1224 }
1225
1226 int ff_h264_frame_start(H264Context *h){
1227     MpegEncContext * const s = &h->s;
1228     int i;
1229     const int pixel_shift = h->pixel_shift;
1230     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1231
1232     if(MPV_frame_start(s, s->avctx) < 0)
1233         return -1;
1234     ff_er_frame_start(s);
1235     /*
1236      * MPV_frame_start uses pict_type to derive key_frame.
1237      * This is incorrect for H.264; IDR markings must be used.
1238      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1239      * See decode_nal_units().
1240      */
1241     s->current_picture_ptr->key_frame= 0;
1242     s->current_picture_ptr->mmco_reset= 0;
1243
1244     assert(s->linesize && s->uvlinesize);
1245
1246     for(i=0; i<16; i++){
1247         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1248         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1249     }
1250     for(i=0; i<16; i++){
1251         h->block_offset[16+i]=
1252         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1253         h->block_offset[48+16+i]=
1254         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1255     }
1256
1257     /* can't be in alloc_tables because linesize isn't known there.
1258      * FIXME: redo bipred weight to not require extra buffer? */
1259     for(i = 0; i < thread_count; i++)
1260         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1261             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1262
1263     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1264     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1265
1266 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
1267
1268     // We mark the current picture as non-reference after allocating it, so
1269     // that if we break out due to an error it can be released automatically
1270     // in the next MPV_frame_start().
1271     // SVQ3 as well as most other codecs have only last/next/current and thus
1272     // get released even with set reference, besides SVQ3 and others do not
1273     // mark frames as reference later "naturally".
1274     if(s->codec_id != CODEC_ID_SVQ3)
1275         s->current_picture_ptr->reference= 0;
1276
1277     s->current_picture_ptr->field_poc[0]=
1278     s->current_picture_ptr->field_poc[1]= INT_MAX;
1279
1280     h->next_output_pic = NULL;
1281
1282     assert(s->current_picture_ptr->long_ref==0);
1283
1284     return 0;
1285 }
1286
1287 /**
1288   * Run setup operations that must be run after slice header decoding.
1289   * This includes finding the next displayed frame.
1290   *
1291   * @param h h264 master context
1292   * @param setup_finished enough NALs have been read that we can call
1293   * ff_thread_finish_setup()
1294   */
1295 static void decode_postinit(H264Context *h, int setup_finished){
1296     MpegEncContext * const s = &h->s;
1297     Picture *out = s->current_picture_ptr;
1298     Picture *cur = s->current_picture_ptr;
1299     int i, pics, out_of_order, out_idx;
1300
1301     s->current_picture_ptr->qscale_type= FF_QSCALE_TYPE_H264;
1302     s->current_picture_ptr->pict_type= s->pict_type;
1303
1304     if (h->next_output_pic) return;
1305
1306     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1307         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1308         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1309         //to find this yet, so we assume the worst for now.
1310         //if (setup_finished)
1311         //    ff_thread_finish_setup(s->avctx);
1312         return;
1313     }
1314
1315     cur->interlaced_frame = 0;
1316     cur->repeat_pict = 0;
1317
1318     /* Signal interlacing information externally. */
1319     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1320
1321     if(h->sps.pic_struct_present_flag){
1322         switch (h->sei_pic_struct)
1323         {
1324         case SEI_PIC_STRUCT_FRAME:
1325             break;
1326         case SEI_PIC_STRUCT_TOP_FIELD:
1327         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1328             cur->interlaced_frame = 1;
1329             break;
1330         case SEI_PIC_STRUCT_TOP_BOTTOM:
1331         case SEI_PIC_STRUCT_BOTTOM_TOP:
1332             if (FIELD_OR_MBAFF_PICTURE)
1333                 cur->interlaced_frame = 1;
1334             else
1335                 // try to flag soft telecine progressive
1336                 cur->interlaced_frame = h->prev_interlaced_frame;
1337             break;
1338         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1339         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1340             // Signal the possibility of telecined film externally (pic_struct 5,6)
1341             // From these hints, let the applications decide if they apply deinterlacing.
1342             cur->repeat_pict = 1;
1343             break;
1344         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1345             // Force progressive here, as doubling interlaced frame is a bad idea.
1346             cur->repeat_pict = 2;
1347             break;
1348         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1349             cur->repeat_pict = 4;
1350             break;
1351         }
1352
1353         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1354             cur->interlaced_frame = (h->sei_ct_type & (1<<1)) != 0;
1355     }else{
1356         /* Derive interlacing flag from used decoding process. */
1357         cur->interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1358     }
1359     h->prev_interlaced_frame = cur->interlaced_frame;
1360
1361     if (cur->field_poc[0] != cur->field_poc[1]){
1362         /* Derive top_field_first from field pocs. */
1363         cur->top_field_first = cur->field_poc[0] < cur->field_poc[1];
1364     }else{
1365         if(cur->interlaced_frame || h->sps.pic_struct_present_flag){
1366             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1367             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1368               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1369                 cur->top_field_first = 1;
1370             else
1371                 cur->top_field_first = 0;
1372         }else{
1373             /* Most likely progressive */
1374             cur->top_field_first = 0;
1375         }
1376     }
1377
1378     //FIXME do something with unavailable reference frames
1379
1380     /* Sort B-frames into display order */
1381
1382     if(h->sps.bitstream_restriction_flag
1383        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1384         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1385         s->low_delay = 0;
1386     }
1387
1388     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1389        && !h->sps.bitstream_restriction_flag){
1390         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1391         s->low_delay= 0;
1392     }
1393
1394     pics = 0;
1395     while(h->delayed_pic[pics]) pics++;
1396
1397     assert(pics <= MAX_DELAYED_PIC_COUNT);
1398
1399     h->delayed_pic[pics++] = cur;
1400     if(cur->reference == 0)
1401         cur->reference = DELAYED_PIC_REF;
1402
1403     out = h->delayed_pic[0];
1404     out_idx = 0;
1405     for(i=1; h->delayed_pic[i] && !h->delayed_pic[i]->key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1406         if(h->delayed_pic[i]->poc < out->poc){
1407             out = h->delayed_pic[i];
1408             out_idx = i;
1409         }
1410     if(s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset))
1411         h->next_outputed_poc= INT_MIN;
1412     out_of_order = out->poc < h->next_outputed_poc;
1413
1414     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1415         { }
1416     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1417        || (s->low_delay &&
1418         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1419          || cur->pict_type == AV_PICTURE_TYPE_B)))
1420     {
1421         s->low_delay = 0;
1422         s->avctx->has_b_frames++;
1423     }
1424
1425     if(out_of_order || pics > s->avctx->has_b_frames){
1426         out->reference &= ~DELAYED_PIC_REF;
1427         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1428                          // or else the first thread can release the picture and reuse it unsafely
1429         for(i=out_idx; h->delayed_pic[i]; i++)
1430             h->delayed_pic[i] = h->delayed_pic[i+1];
1431     }
1432     if(!out_of_order && pics > s->avctx->has_b_frames){
1433         h->next_output_pic = out;
1434         if(out_idx==0 && h->delayed_pic[0] && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset)) {
1435             h->next_outputed_poc = INT_MIN;
1436         } else
1437             h->next_outputed_poc = out->poc;
1438     }else{
1439         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1440     }
1441
1442     if (setup_finished)
1443         ff_thread_finish_setup(s->avctx);
1444 }
1445
1446 static 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){
1447     MpegEncContext * const s = &h->s;
1448     uint8_t *top_border;
1449     int top_idx = 1;
1450     const int pixel_shift = h->pixel_shift;
1451
1452     src_y  -=   linesize;
1453     src_cb -= uvlinesize;
1454     src_cr -= uvlinesize;
1455
1456     if(!simple && FRAME_MBAFF){
1457         if(s->mb_y&1){
1458             if(!MB_MBAFF){
1459                 top_border = h->top_borders[0][s->mb_x];
1460                 AV_COPY128(top_border, src_y + 15*linesize);
1461                 if (pixel_shift)
1462                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1463                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1464                     if(chroma444){
1465                         if (pixel_shift){
1466                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1467                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1468                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1469                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1470                         } else {
1471                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1472                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1473                         }
1474                     } else {
1475                         if (pixel_shift) {
1476                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1477                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1478                         } else {
1479                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1480                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1481                         }
1482                     }
1483                 }
1484             }
1485         }else if(MB_MBAFF){
1486             top_idx = 0;
1487         }else
1488             return;
1489     }
1490
1491     top_border = h->top_borders[top_idx][s->mb_x];
1492     // There are two lines saved, the line above the the top macroblock of a pair,
1493     // and the line above the bottom macroblock
1494     AV_COPY128(top_border, src_y + 16*linesize);
1495     if (pixel_shift)
1496         AV_COPY128(top_border+16, src_y+16*linesize+16);
1497
1498     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1499         if(chroma444){
1500             if (pixel_shift){
1501                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1502                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1503                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1504                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1505             } else {
1506                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1507                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1508             }
1509         } else {
1510             if (pixel_shift) {
1511                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1512                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1513             } else {
1514                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1515                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1516             }
1517         }
1518     }
1519 }
1520
1521 static inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1522                                   uint8_t *src_cb, uint8_t *src_cr,
1523                                   int linesize, int uvlinesize,
1524                                   int xchg, int chroma444,
1525                                   int simple, int pixel_shift){
1526     MpegEncContext * const s = &h->s;
1527     int deblock_topleft;
1528     int deblock_top;
1529     int top_idx = 1;
1530     uint8_t *top_border_m1;
1531     uint8_t *top_border;
1532
1533     if(!simple && FRAME_MBAFF){
1534         if(s->mb_y&1){
1535             if(!MB_MBAFF)
1536                 return;
1537         }else{
1538             top_idx = MB_MBAFF ? 0 : 1;
1539         }
1540     }
1541
1542     if(h->deblocking_filter == 2) {
1543         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1544         deblock_top     = h->top_type;
1545     } else {
1546         deblock_topleft = (s->mb_x > 0);
1547         deblock_top     = (s->mb_y > !!MB_FIELD);
1548     }
1549
1550     src_y  -=   linesize + 1 + pixel_shift;
1551     src_cb -= uvlinesize + 1 + pixel_shift;
1552     src_cr -= uvlinesize + 1 + pixel_shift;
1553
1554     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1555     top_border    = h->top_borders[top_idx][s->mb_x];
1556
1557 #define XCHG(a,b,xchg)\
1558     if (pixel_shift) {\
1559         if (xchg) {\
1560             AV_SWAP64(b+0,a+0);\
1561             AV_SWAP64(b+8,a+8);\
1562         } else {\
1563             AV_COPY128(b,a); \
1564         }\
1565     } else \
1566 if (xchg) AV_SWAP64(b,a);\
1567 else      AV_COPY64(b,a);
1568
1569     if(deblock_top){
1570         if(deblock_topleft){
1571             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1572         }
1573         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1574         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1575         if(s->mb_x+1 < s->mb_width){
1576             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1577         }
1578     }
1579     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1580         if(chroma444){
1581             if(deblock_topleft){
1582                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1583                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1584             }
1585             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1586             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1587             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1588             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1589             if(s->mb_x+1 < s->mb_width){
1590                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1591                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1592             }
1593         } else {
1594             if(deblock_top){
1595                 if(deblock_topleft){
1596                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1597                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1598                 }
1599                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1600                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1601             }
1602         }
1603     }
1604 }
1605
1606 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1607     if (high_bit_depth) {
1608         return AV_RN32A(((int32_t*)mb) + index);
1609     } else
1610         return AV_RN16A(mb + index);
1611 }
1612
1613 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1614     if (high_bit_depth) {
1615         AV_WN32A(((int32_t*)mb) + index, value);
1616     } else
1617         AV_WN16A(mb + index, value);
1618 }
1619
1620 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1621                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1622 {
1623     MpegEncContext * const s = &h->s;
1624     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1625     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1626     int i;
1627     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1628     block_offset += 16*p;
1629     if(IS_INTRA4x4(mb_type)){
1630         if(simple || !s->encoding){
1631             if(IS_8x8DCT(mb_type)){
1632                 if(transform_bypass){
1633                     idct_dc_add =
1634                     idct_add    = s->dsp.add_pixels8;
1635                 }else{
1636                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1637                     idct_add    = h->h264dsp.h264_idct8_add;
1638                 }
1639                 for(i=0; i<16; i+=4){
1640                     uint8_t * const ptr= dest_y + block_offset[i];
1641                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1642                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1643                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1644                     }else{
1645                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1646                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1647                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1648                         if(nnz){
1649                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1650                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1651                             else
1652                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1653                         }
1654                     }
1655                 }
1656             }else{
1657                 if(transform_bypass){
1658                     idct_dc_add =
1659                     idct_add    = s->dsp.add_pixels4;
1660                 }else{
1661                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1662                     idct_add    = h->h264dsp.h264_idct_add;
1663                 }
1664                 for(i=0; i<16; i++){
1665                     uint8_t * const ptr= dest_y + block_offset[i];
1666                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1667
1668                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1669                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1670                     }else{
1671                         uint8_t *topright;
1672                         int nnz, tr;
1673                         uint64_t tr_high;
1674                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1675                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1676                             assert(s->mb_y || linesize <= block_offset[i]);
1677                             if(!topright_avail){
1678                                 if (pixel_shift) {
1679                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1680                                     topright= (uint8_t*) &tr_high;
1681                                 } else {
1682                                     tr= ptr[3 - linesize]*0x01010101;
1683                                     topright= (uint8_t*) &tr;
1684                                 }
1685                             }else
1686                                 topright= ptr + (4 << pixel_shift) - linesize;
1687                         }else
1688                             topright= NULL;
1689
1690                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1691                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1692                         if(nnz){
1693                             if(is_h264){
1694                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1695                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1696                                 else
1697                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1698                             }else
1699                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1700                         }
1701                     }
1702                 }
1703             }
1704         }
1705     }else{
1706         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1707         if(is_h264){
1708             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1709                 if(!transform_bypass)
1710                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1711                 else{
1712                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1713                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1714                     for(i = 0; i < 16; i++)
1715                         dctcoef_set(h->mb+p*256, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1716                 }
1717             }
1718         }else
1719             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1720     }
1721 }
1722
1723 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1724                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1725 {
1726     MpegEncContext * const s = &h->s;
1727     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1728     int i;
1729     block_offset += 16*p;
1730     if(!IS_INTRA4x4(mb_type)){
1731         if(is_h264){
1732             if(IS_INTRA16x16(mb_type)){
1733                 if(transform_bypass){
1734                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1735                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1736                     }else{
1737                         for(i=0; i<16; i++){
1738                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1739                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1740                         }
1741                     }
1742                 }else{
1743                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1744                 }
1745             }else if(h->cbp&15){
1746                 if(transform_bypass){
1747                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1748                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1749                     for(i=0; i<16; i+=di){
1750                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1751                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1752                         }
1753                     }
1754                 }else{
1755                     if(IS_8x8DCT(mb_type)){
1756                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1757                     }else{
1758                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1759                     }
1760                 }
1761             }
1762         }else{
1763             for(i=0; i<16; i++){
1764                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1765                     uint8_t * const ptr= dest_y + block_offset[i];
1766                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1767                 }
1768             }
1769         }
1770     }
1771 }
1772
1773 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1774     MpegEncContext * const s = &h->s;
1775     const int mb_x= s->mb_x;
1776     const int mb_y= s->mb_y;
1777     const int mb_xy= h->mb_xy;
1778     const int mb_type= s->current_picture.mb_type[mb_xy];
1779     uint8_t  *dest_y, *dest_cb, *dest_cr;
1780     int linesize, uvlinesize /*dct_offset*/;
1781     int i, j;
1782     int *block_offset = &h->block_offset[0];
1783     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1784     /* is_h264 should always be true if SVQ3 is disabled. */
1785     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1786     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1787
1788     dest_y  = s->current_picture.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1789     dest_cb = s->current_picture.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
1790     dest_cr = s->current_picture.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
1791
1792     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1793     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1794
1795     h->list_counts[mb_xy]= h->list_count;
1796
1797     if (!simple && MB_FIELD) {
1798         linesize   = h->mb_linesize   = s->linesize * 2;
1799         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1800         block_offset = &h->block_offset[48];
1801         if(mb_y&1){ //FIXME move out of this function?
1802             dest_y -= s->linesize*15;
1803             dest_cb-= s->uvlinesize*7;
1804             dest_cr-= s->uvlinesize*7;
1805         }
1806         if(FRAME_MBAFF) {
1807             int list;
1808             for(list=0; list<h->list_count; list++){
1809                 if(!USES_LIST(mb_type, list))
1810                     continue;
1811                 if(IS_16X16(mb_type)){
1812                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1813                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1814                 }else{
1815                     for(i=0; i<16; i+=4){
1816                         int ref = h->ref_cache[list][scan8[i]];
1817                         if(ref >= 0)
1818                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1819                     }
1820                 }
1821             }
1822         }
1823     } else {
1824         linesize   = h->mb_linesize   = s->linesize;
1825         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1826 //        dct_offset = s->linesize * 16;
1827     }
1828
1829     if (!simple && IS_INTRA_PCM(mb_type)) {
1830         if (pixel_shift) {
1831             const int bit_depth = h->sps.bit_depth_luma;
1832             int j;
1833             GetBitContext gb;
1834             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1835
1836             for (i = 0; i < 16; i++) {
1837                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1838                 for (j = 0; j < 16; j++)
1839                     tmp_y[j] = get_bits(&gb, bit_depth);
1840             }
1841             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1842                 for (i = 0; i < 8; i++) {
1843                     uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1844                     for (j = 0; j < 8; j++)
1845                         tmp_cb[j] = get_bits(&gb, bit_depth);
1846                 }
1847                 for (i = 0; i < 8; i++) {
1848                     uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1849                     for (j = 0; j < 8; j++)
1850                         tmp_cr[j] = get_bits(&gb, bit_depth);
1851                 }
1852             }
1853         } else {
1854             for (i=0; i<16; i++) {
1855                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1856             }
1857             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1858                 for (i=0; i<8; i++) {
1859                     memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
1860                     memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
1861                 }
1862             }
1863         }
1864     } else {
1865         if(IS_INTRA(mb_type)){
1866             if(h->deblocking_filter)
1867                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1868
1869             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1870                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1871                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1872             }
1873
1874             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1875
1876             if(h->deblocking_filter)
1877                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1878         }else if(is_h264){
1879             if (pixel_shift) {
1880                 hl_motion_16(h, dest_y, dest_cb, dest_cr,
1881                              s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1882                              s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1883                              h->h264dsp.weight_h264_pixels_tab,
1884                              h->h264dsp.biweight_h264_pixels_tab, 0);
1885             } else
1886                 hl_motion_8(h, dest_y, dest_cb, dest_cr,
1887                             s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1888                             s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1889                             h->h264dsp.weight_h264_pixels_tab,
1890                             h->h264dsp.biweight_h264_pixels_tab, 0);
1891         }
1892
1893         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1894
1895         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1896             uint8_t *dest[2] = {dest_cb, dest_cr};
1897             if(transform_bypass){
1898                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1899                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
1900                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
1901                 }else{
1902                     idct_add = s->dsp.add_pixels4;
1903                     for(j=1; j<3; j++){
1904                         for(i=j*16; i<j*16+4; i++){
1905                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1906                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1907                         }
1908                     }
1909                 }
1910             }else{
1911                 if(is_h264){
1912                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1913                         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]);
1914                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1915                         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]);
1916                     h->h264dsp.h264_idct_add8(dest, block_offset,
1917                                               h->mb, uvlinesize,
1918                                               h->non_zero_count_cache);
1919                 }else{
1920                     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]);
1921                     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]);
1922                     for(j=1; j<3; j++){
1923                         for(i=j*16; i<j*16+4; i++){
1924                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1925                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
1926                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1927                             }
1928                         }
1929                     }
1930                 }
1931             }
1932         }
1933     }
1934     if(h->cbp || IS_INTRA(mb_type))
1935     {
1936         s->dsp.clear_blocks(h->mb);
1937         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
1938     }
1939 }
1940
1941 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
1942     MpegEncContext * const s = &h->s;
1943     const int mb_x= s->mb_x;
1944     const int mb_y= s->mb_y;
1945     const int mb_xy= h->mb_xy;
1946     const int mb_type= s->current_picture.mb_type[mb_xy];
1947     uint8_t  *dest[3];
1948     int linesize;
1949     int i, j, p;
1950     int *block_offset = &h->block_offset[0];
1951     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1952     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
1953
1954     for (p = 0; p < plane_count; p++)
1955     {
1956         dest[p] = s->current_picture.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
1957         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1958     }
1959
1960     h->list_counts[mb_xy]= h->list_count;
1961
1962     if (!simple && MB_FIELD) {
1963         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
1964         block_offset = &h->block_offset[48];
1965         if(mb_y&1) //FIXME move out of this function?
1966             for (p = 0; p < 3; p++)
1967                 dest[p] -= s->linesize*15;
1968         if(FRAME_MBAFF) {
1969             int list;
1970             for(list=0; list<h->list_count; list++){
1971                 if(!USES_LIST(mb_type, list))
1972                     continue;
1973                 if(IS_16X16(mb_type)){
1974                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1975                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1976                 }else{
1977                     for(i=0; i<16; i+=4){
1978                         int ref = h->ref_cache[list][scan8[i]];
1979                         if(ref >= 0)
1980                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1981                     }
1982                 }
1983             }
1984         }
1985     } else {
1986         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
1987     }
1988
1989     if (!simple && IS_INTRA_PCM(mb_type)) {
1990         if (pixel_shift) {
1991             const int bit_depth = h->sps.bit_depth_luma;
1992             GetBitContext gb;
1993             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
1994
1995             for (p = 0; p < plane_count; p++) {
1996                 for (i = 0; i < 16; i++) {
1997                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
1998                     for (j = 0; j < 16; j++)
1999                         tmp[j] = get_bits(&gb, bit_depth);
2000                 }
2001             }
2002         } else {
2003             for (p = 0; p < plane_count; p++) {
2004                 for (i = 0; i < 16; i++) {
2005                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2006                 }
2007             }
2008         }
2009     } else {
2010         if(IS_INTRA(mb_type)){
2011             if(h->deblocking_filter)
2012                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2013
2014             for (p = 0; p < plane_count; p++)
2015                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2016
2017             if(h->deblocking_filter)
2018                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2019         }else{
2020             if (pixel_shift) {
2021                 hl_motion_16(h, dest[0], dest[1], dest[2],
2022                              s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2023                              s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2024                              h->h264dsp.weight_h264_pixels_tab,
2025                              h->h264dsp.biweight_h264_pixels_tab, 1);
2026             } else
2027                 hl_motion_8(h, dest[0], dest[1], dest[2],
2028                             s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2029                             s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2030                             h->h264dsp.weight_h264_pixels_tab,
2031                             h->h264dsp.biweight_h264_pixels_tab, 1);
2032         }
2033
2034         for (p = 0; p < plane_count; p++)
2035             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2036     }
2037     if(h->cbp || IS_INTRA(mb_type))
2038     {
2039         s->dsp.clear_blocks(h->mb);
2040         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2041     }
2042 }
2043
2044 /**
2045  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2046  */
2047 #define hl_decode_mb_simple(sh, bits) \
2048 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2049     hl_decode_mb_internal(h, 1, sh); \
2050 }
2051 hl_decode_mb_simple(0, 8);
2052 hl_decode_mb_simple(1, 16);
2053
2054 /**
2055  * Process a macroblock; this handles edge cases, such as interlacing.
2056  */
2057 static void av_noinline hl_decode_mb_complex(H264Context *h){
2058     hl_decode_mb_internal(h, 0, h->pixel_shift);
2059 }
2060
2061 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2062     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2063 }
2064
2065 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2066     hl_decode_mb_444_internal(h, 1, 0);
2067 }
2068
2069 void ff_h264_hl_decode_mb(H264Context *h){
2070     MpegEncContext * const s = &h->s;
2071     const int mb_xy= h->mb_xy;
2072     const int mb_type= s->current_picture.mb_type[mb_xy];
2073     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2074
2075     if (CHROMA444) {
2076         if(is_complex || h->pixel_shift)
2077             hl_decode_mb_444_complex(h);
2078         else
2079             hl_decode_mb_444_simple(h);
2080     } else if (is_complex) {
2081         hl_decode_mb_complex(h);
2082     } else if (h->pixel_shift) {
2083         hl_decode_mb_simple_16(h);
2084     } else
2085         hl_decode_mb_simple_8(h);
2086 }
2087
2088 static int pred_weight_table(H264Context *h){
2089     MpegEncContext * const s = &h->s;
2090     int list, i;
2091     int luma_def, chroma_def;
2092
2093     h->use_weight= 0;
2094     h->use_weight_chroma= 0;
2095     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2096     if(h->sps.chroma_format_idc)
2097         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2098     luma_def = 1<<h->luma_log2_weight_denom;
2099     chroma_def = 1<<h->chroma_log2_weight_denom;
2100
2101     for(list=0; list<2; list++){
2102         h->luma_weight_flag[list]   = 0;
2103         h->chroma_weight_flag[list] = 0;
2104         for(i=0; i<h->ref_count[list]; i++){
2105             int luma_weight_flag, chroma_weight_flag;
2106
2107             luma_weight_flag= get_bits1(&s->gb);
2108             if(luma_weight_flag){
2109                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2110                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2111                 if(   h->luma_weight[i][list][0] != luma_def
2112                    || h->luma_weight[i][list][1] != 0) {
2113                     h->use_weight= 1;
2114                     h->luma_weight_flag[list]= 1;
2115                 }
2116             }else{
2117                 h->luma_weight[i][list][0]= luma_def;
2118                 h->luma_weight[i][list][1]= 0;
2119             }
2120
2121             if(h->sps.chroma_format_idc){
2122                 chroma_weight_flag= get_bits1(&s->gb);
2123                 if(chroma_weight_flag){
2124                     int j;
2125                     for(j=0; j<2; j++){
2126                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2127                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2128                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2129                            || h->chroma_weight[i][list][j][1] != 0) {
2130                             h->use_weight_chroma= 1;
2131                             h->chroma_weight_flag[list]= 1;
2132                         }
2133                     }
2134                 }else{
2135                     int j;
2136                     for(j=0; j<2; j++){
2137                         h->chroma_weight[i][list][j][0]= chroma_def;
2138                         h->chroma_weight[i][list][j][1]= 0;
2139                     }
2140                 }
2141             }
2142         }
2143         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2144     }
2145     h->use_weight= h->use_weight || h->use_weight_chroma;
2146     return 0;
2147 }
2148
2149 /**
2150  * Initialize implicit_weight table.
2151  * @param field  0/1 initialize the weight for interlaced MBAFF
2152  *                -1 initializes the rest
2153  */
2154 static void implicit_weight_table(H264Context *h, int field){
2155     MpegEncContext * const s = &h->s;
2156     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2157
2158     for (i = 0; i < 2; i++) {
2159         h->luma_weight_flag[i]   = 0;
2160         h->chroma_weight_flag[i] = 0;
2161     }
2162
2163     if(field < 0){
2164         cur_poc = s->current_picture_ptr->poc;
2165     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2166        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2167         h->use_weight= 0;
2168         h->use_weight_chroma= 0;
2169         return;
2170     }
2171         ref_start= 0;
2172         ref_count0= h->ref_count[0];
2173         ref_count1= h->ref_count[1];
2174     }else{
2175         cur_poc = s->current_picture_ptr->field_poc[field];
2176         ref_start= 16;
2177         ref_count0= 16+2*h->ref_count[0];
2178         ref_count1= 16+2*h->ref_count[1];
2179     }
2180
2181     h->use_weight= 2;
2182     h->use_weight_chroma= 2;
2183     h->luma_log2_weight_denom= 5;
2184     h->chroma_log2_weight_denom= 5;
2185
2186     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2187         int poc0 = h->ref_list[0][ref0].poc;
2188         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2189             int poc1 = h->ref_list[1][ref1].poc;
2190             int td = av_clip(poc1 - poc0, -128, 127);
2191             int w= 32;
2192             if(td){
2193                 int tb = av_clip(cur_poc - poc0, -128, 127);
2194                 int tx = (16384 + (FFABS(td) >> 1)) / td;
2195                 int dist_scale_factor = (tb*tx + 32) >> 8;
2196                 if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2197                     w = 64 - dist_scale_factor;
2198             }
2199             if(field<0){
2200                 h->implicit_weight[ref0][ref1][0]=
2201                 h->implicit_weight[ref0][ref1][1]= w;
2202             }else{
2203                 h->implicit_weight[ref0][ref1][field]=w;
2204             }
2205         }
2206     }
2207 }
2208
2209 /**
2210  * instantaneous decoder refresh.
2211  */
2212 static void idr(H264Context *h){
2213     ff_h264_remove_all_refs(h);
2214     h->prev_frame_num= 0;
2215     h->prev_frame_num_offset= 0;
2216     h->prev_poc_msb=
2217     h->prev_poc_lsb= 0;
2218 }
2219
2220 /* forget old pics after a seek */
2221 static void flush_dpb(AVCodecContext *avctx){
2222     H264Context *h= avctx->priv_data;
2223     int i;
2224     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
2225         if(h->delayed_pic[i])
2226             h->delayed_pic[i]->reference= 0;
2227         h->delayed_pic[i]= NULL;
2228     }
2229     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2230     h->prev_interlaced_frame = 1;
2231     idr(h);
2232     if(h->s.current_picture_ptr)
2233         h->s.current_picture_ptr->reference= 0;
2234     h->s.first_field= 0;
2235     ff_h264_reset_sei(h);
2236     ff_mpeg_flush(avctx);
2237 }
2238
2239 static int init_poc(H264Context *h){
2240     MpegEncContext * const s = &h->s;
2241     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2242     int field_poc[2];
2243     Picture *cur = s->current_picture_ptr;
2244
2245     h->frame_num_offset= h->prev_frame_num_offset;
2246     if(h->frame_num < h->prev_frame_num)
2247         h->frame_num_offset += max_frame_num;
2248
2249     if(h->sps.poc_type==0){
2250         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2251
2252         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2253             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2254         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2255             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2256         else
2257             h->poc_msb = h->prev_poc_msb;
2258 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2259         field_poc[0] =
2260         field_poc[1] = h->poc_msb + h->poc_lsb;
2261         if(s->picture_structure == PICT_FRAME)
2262             field_poc[1] += h->delta_poc_bottom;
2263     }else if(h->sps.poc_type==1){
2264         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2265         int i;
2266
2267         if(h->sps.poc_cycle_length != 0)
2268             abs_frame_num = h->frame_num_offset + h->frame_num;
2269         else
2270             abs_frame_num = 0;
2271
2272         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2273             abs_frame_num--;
2274
2275         expected_delta_per_poc_cycle = 0;
2276         for(i=0; i < h->sps.poc_cycle_length; i++)
2277             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2278
2279         if(abs_frame_num > 0){
2280             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2281             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2282
2283             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2284             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2285                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2286         } else
2287             expectedpoc = 0;
2288
2289         if(h->nal_ref_idc == 0)
2290             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2291
2292         field_poc[0] = expectedpoc + h->delta_poc[0];
2293         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2294
2295         if(s->picture_structure == PICT_FRAME)
2296             field_poc[1] += h->delta_poc[1];
2297     }else{
2298         int poc= 2*(h->frame_num_offset + h->frame_num);
2299
2300         if(!h->nal_ref_idc)
2301             poc--;
2302
2303         field_poc[0]= poc;
2304         field_poc[1]= poc;
2305     }
2306
2307     if(s->picture_structure != PICT_BOTTOM_FIELD)
2308         s->current_picture_ptr->field_poc[0]= field_poc[0];
2309     if(s->picture_structure != PICT_TOP_FIELD)
2310         s->current_picture_ptr->field_poc[1]= field_poc[1];
2311     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2312
2313     return 0;
2314 }
2315
2316
2317 /**
2318  * initialize scan tables
2319  */
2320 static void init_scan_tables(H264Context *h){
2321     int i;
2322     for(i=0; i<16; i++){
2323 #define T(x) (x>>2) | ((x<<2) & 0xF)
2324         h->zigzag_scan[i] = T(zigzag_scan[i]);
2325         h-> field_scan[i] = T( field_scan[i]);
2326 #undef T
2327     }
2328     for(i=0; i<64; i++){
2329 #define T(x) (x>>3) | ((x&7)<<3)
2330         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2331         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2332         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2333         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2334 #undef T
2335     }
2336     if(h->sps.transform_bypass){ //FIXME same ugly
2337         h->zigzag_scan_q0          = zigzag_scan;
2338         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2339         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2340         h->field_scan_q0           = field_scan;
2341         h->field_scan8x8_q0        = field_scan8x8;
2342         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2343     }else{
2344         h->zigzag_scan_q0          = h->zigzag_scan;
2345         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2346         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2347         h->field_scan_q0           = h->field_scan;
2348         h->field_scan8x8_q0        = h->field_scan8x8;
2349         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2350     }
2351 }
2352
2353 static void field_end(H264Context *h, int in_setup){
2354     MpegEncContext * const s = &h->s;
2355     AVCodecContext * const avctx= s->avctx;
2356     s->mb_y= 0;
2357
2358     if (!in_setup && !s->dropable)
2359         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2360                                  s->picture_structure==PICT_BOTTOM_FIELD);
2361
2362     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2363         ff_vdpau_h264_set_reference_frames(s);
2364
2365     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2366         if(!s->dropable) {
2367             ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2368             h->prev_poc_msb= h->poc_msb;
2369             h->prev_poc_lsb= h->poc_lsb;
2370         }
2371         h->prev_frame_num_offset= h->frame_num_offset;
2372         h->prev_frame_num= h->frame_num;
2373         h->outputed_poc = h->next_outputed_poc;
2374     }
2375
2376     if (avctx->hwaccel) {
2377         if (avctx->hwaccel->end_frame(avctx) < 0)
2378             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2379     }
2380
2381     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2382         ff_vdpau_h264_picture_complete(s);
2383
2384     /*
2385      * FIXME: Error handling code does not seem to support interlaced
2386      * when slices span multiple rows
2387      * The ff_er_add_slice calls don't work right for bottom
2388      * fields; they cause massive erroneous error concealing
2389      * Error marking covers both fields (top and bottom).
2390      * This causes a mismatched s->error_count
2391      * and a bad error table. Further, the error count goes to
2392      * INT_MAX when called for bottom field, because mb_y is
2393      * past end by one (callers fault) and resync_mb_y != 0
2394      * causes problems for the first MB line, too.
2395      */
2396     if (!FIELD_PICTURE)
2397         ff_er_frame_end(s);
2398
2399     MPV_frame_end(s);
2400
2401     h->current_slice=0;
2402 }
2403
2404 /**
2405  * Replicate H264 "master" context to thread contexts.
2406  */
2407 static void clone_slice(H264Context *dst, H264Context *src)
2408 {
2409     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2410     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2411     dst->s.current_picture      = src->s.current_picture;
2412     dst->s.linesize             = src->s.linesize;
2413     dst->s.uvlinesize           = src->s.uvlinesize;
2414     dst->s.first_field          = src->s.first_field;
2415
2416     dst->prev_poc_msb           = src->prev_poc_msb;
2417     dst->prev_poc_lsb           = src->prev_poc_lsb;
2418     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2419     dst->prev_frame_num         = src->prev_frame_num;
2420     dst->short_ref_count        = src->short_ref_count;
2421
2422     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2423     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2424     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2425     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2426
2427     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2428     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2429 }
2430
2431 /**
2432  * computes profile from profile_idc and constraint_set?_flags
2433  *
2434  * @param sps SPS
2435  *
2436  * @return profile as defined by FF_PROFILE_H264_*
2437  */
2438 int ff_h264_get_profile(SPS *sps)
2439 {
2440     int profile = sps->profile_idc;
2441
2442     switch(sps->profile_idc) {
2443     case FF_PROFILE_H264_BASELINE:
2444         // constraint_set1_flag set to 1
2445         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2446         break;
2447     case FF_PROFILE_H264_HIGH_10:
2448     case FF_PROFILE_H264_HIGH_422:
2449     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2450         // constraint_set3_flag set to 1
2451         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2452         break;
2453     }
2454
2455     return profile;
2456 }
2457
2458 /**
2459  * decodes a slice header.
2460  * This will also call MPV_common_init() and frame_start() as needed.
2461  *
2462  * @param h h264context
2463  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2464  *
2465  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2466  */
2467 static int decode_slice_header(H264Context *h, H264Context *h0){
2468     MpegEncContext * const s = &h->s;
2469     MpegEncContext * const s0 = &h0->s;
2470     unsigned int first_mb_in_slice;
2471     unsigned int pps_id;
2472     int num_ref_idx_active_override_flag;
2473     unsigned int slice_type, tmp, i, j;
2474     int default_ref_list_done = 0;
2475     int last_pic_structure;
2476
2477     s->dropable= h->nal_ref_idc == 0;
2478
2479     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2480     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2481         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2482         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2483     }else{
2484         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2485         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2486     }
2487
2488     first_mb_in_slice= get_ue_golomb(&s->gb);
2489
2490     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2491         if(h0->current_slice && FIELD_PICTURE){
2492             field_end(h, 1);
2493         }
2494
2495         h0->current_slice = 0;
2496         if (!s0->first_field)
2497             s->current_picture_ptr= NULL;
2498     }
2499
2500     slice_type= get_ue_golomb_31(&s->gb);
2501     if(slice_type > 9){
2502         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);
2503         return -1;
2504     }
2505     if(slice_type > 4){
2506         slice_type -= 5;
2507         h->slice_type_fixed=1;
2508     }else
2509         h->slice_type_fixed=0;
2510
2511     slice_type= golomb_to_pict_type[ slice_type ];
2512     if (slice_type == AV_PICTURE_TYPE_I
2513         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2514         default_ref_list_done = 1;
2515     }
2516     h->slice_type= slice_type;
2517     h->slice_type_nos= slice_type & 3;
2518
2519     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2520
2521     pps_id= get_ue_golomb(&s->gb);
2522     if(pps_id>=MAX_PPS_COUNT){
2523         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2524         return -1;
2525     }
2526     if(!h0->pps_buffers[pps_id]) {
2527         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2528         return -1;
2529     }
2530     h->pps= *h0->pps_buffers[pps_id];
2531
2532     if(!h0->sps_buffers[h->pps.sps_id]) {
2533         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2534         return -1;
2535     }
2536     h->sps = *h0->sps_buffers[h->pps.sps_id];
2537
2538     s->avctx->profile = ff_h264_get_profile(&h->sps);
2539     s->avctx->level   = h->sps.level_idc;
2540     s->avctx->refs    = h->sps.ref_frame_count;
2541
2542     if(h == h0 && h->dequant_coeff_pps != pps_id){
2543         h->dequant_coeff_pps = pps_id;
2544         init_dequant_tables(h);
2545     }
2546
2547     s->mb_width= h->sps.mb_width;
2548     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2549
2550     h->b_stride=  s->mb_width*4;
2551
2552     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2553     if(h->sps.frame_mbs_only_flag)
2554         s->height= 16*s->mb_height - (2>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2555     else
2556         s->height= 16*s->mb_height - (4>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2557
2558     if (s->context_initialized
2559         && (   s->width != s->avctx->width || s->height != s->avctx->height
2560             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2561         if(h != h0) {
2562             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2563             return -1;   // width / height changed during parallelized decoding
2564         }
2565         free_tables(h, 0);
2566         flush_dpb(s->avctx);
2567         MPV_common_end(s);
2568     }
2569     if (!s->context_initialized) {
2570         if (h != h0) {
2571             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2572             return -1;
2573         }
2574
2575         avcodec_set_dimensions(s->avctx, s->width, s->height);
2576         s->avctx->sample_aspect_ratio= h->sps.sar;
2577         av_assert0(s->avctx->sample_aspect_ratio.den);
2578
2579         h->s.avctx->coded_width = 16*s->mb_width;
2580         h->s.avctx->coded_height = 16*s->mb_height;
2581
2582         if(h->sps.video_signal_type_present_flag){
2583             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2584             if(h->sps.colour_description_present_flag){
2585                 s->avctx->color_primaries = h->sps.color_primaries;
2586                 s->avctx->color_trc       = h->sps.color_trc;
2587                 s->avctx->colorspace      = h->sps.colorspace;
2588             }
2589         }
2590
2591         if(h->sps.timing_info_present_flag){
2592             int64_t den= h->sps.time_scale;
2593             if(h->x264_build < 44U)
2594                 den *= 2;
2595             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2596                       h->sps.num_units_in_tick, den, 1<<30);
2597         }
2598
2599         switch (h->sps.bit_depth_luma) {
2600             case 9 :
2601                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P9 : PIX_FMT_YUV420P9;
2602                 break;
2603             case 10 :
2604                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P10 : PIX_FMT_YUV420P10;
2605                 break;
2606             default:
2607                 if (CHROMA444){
2608                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2609                 }else{
2610                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2611                                                              s->avctx->codec->pix_fmts ?
2612                                                              s->avctx->codec->pix_fmts :
2613                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2614                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2615                                                              ff_hwaccel_pixfmt_list_420);
2616                 }
2617         }
2618
2619         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2620
2621         if (MPV_common_init(s) < 0) {
2622             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2623             return -1;
2624         }
2625         s->first_field = 0;
2626         h->prev_interlaced_frame = 1;
2627
2628         init_scan_tables(h);
2629         ff_h264_alloc_tables(h);
2630
2631         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2632             if (context_init(h) < 0) {
2633                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2634                 return -1;
2635             }
2636         } else {
2637             for(i = 1; i < s->avctx->thread_count; i++) {
2638                 H264Context *c;
2639                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2640                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2641                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2642                 c->h264dsp = h->h264dsp;
2643                 c->sps = h->sps;
2644                 c->pps = h->pps;
2645                 c->pixel_shift = h->pixel_shift;
2646                 init_scan_tables(c);
2647                 clone_tables(c, h, i);
2648             }
2649
2650             for(i = 0; i < s->avctx->thread_count; i++)
2651                 if (context_init(h->thread_context[i]) < 0) {
2652                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2653                     return -1;
2654                 }
2655         }
2656     }
2657
2658     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2659
2660     h->mb_mbaff = 0;
2661     h->mb_aff_frame = 0;
2662     last_pic_structure = s0->picture_structure;
2663     if(h->sps.frame_mbs_only_flag){
2664         s->picture_structure= PICT_FRAME;
2665     }else{
2666         if(get_bits1(&s->gb)) { //field_pic_flag
2667             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2668         } else {
2669             s->picture_structure= PICT_FRAME;
2670             h->mb_aff_frame = h->sps.mb_aff;
2671         }
2672     }
2673     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2674
2675     if(h0->current_slice == 0){
2676         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2677         if(h->frame_num != h->prev_frame_num) {
2678             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2679
2680             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2681
2682             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2683                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2684                 if (unwrap_prev_frame_num < 0)
2685                     unwrap_prev_frame_num += max_frame_num;
2686
2687                 h->prev_frame_num = unwrap_prev_frame_num;
2688             }
2689         }
2690
2691         while(h->frame_num !=  h->prev_frame_num &&
2692               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2693             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2694             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2695             if (ff_h264_frame_start(h) < 0)
2696                 return -1;
2697             h->prev_frame_num++;
2698             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2699             s->current_picture_ptr->frame_num= h->prev_frame_num;
2700             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2701             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2702             ff_generate_sliding_window_mmcos(h);
2703             ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2704             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2705              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2706              * about there being no actual duplicates.
2707              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2708              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2709              * be fixed. */
2710             if (h->short_ref_count) {
2711                 if (prev) {
2712                     av_image_copy(h->short_ref[0]->data, h->short_ref[0]->linesize,
2713                                   (const uint8_t**)prev->data, prev->linesize,
2714                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2715                     h->short_ref[0]->poc = prev->poc+2;
2716                 }
2717                 h->short_ref[0]->frame_num = h->prev_frame_num;
2718             }
2719         }
2720
2721         /* See if we have a decoded first field looking for a pair... */
2722         if (s0->first_field) {
2723             assert(s0->current_picture_ptr);
2724             assert(s0->current_picture_ptr->data[0]);
2725             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
2726
2727             /* figure out if we have a complementary field pair */
2728             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2729                 /*
2730                  * Previous field is unmatched. Don't display it, but let it
2731                  * remain for reference if marked as such.
2732                  */
2733                 s0->current_picture_ptr = NULL;
2734                 s0->first_field = FIELD_PICTURE;
2735
2736             } else {
2737                 if (h->nal_ref_idc &&
2738                         s0->current_picture_ptr->reference &&
2739                         s0->current_picture_ptr->frame_num != h->frame_num) {
2740                     /*
2741                      * This and previous field were reference, but had
2742                      * different frame_nums. Consider this field first in
2743                      * pair. Throw away previous field except for reference
2744                      * purposes.
2745                      */
2746                     s0->first_field = 1;
2747                     s0->current_picture_ptr = NULL;
2748
2749                 } else {
2750                     /* Second field in complementary pair */
2751                     s0->first_field = 0;
2752                 }
2753             }
2754
2755         } else {
2756             /* Frame or first field in a potentially complementary pair */
2757             assert(!s0->current_picture_ptr);
2758             s0->first_field = FIELD_PICTURE;
2759         }
2760
2761         if(!FIELD_PICTURE || s0->first_field) {
2762             if (ff_h264_frame_start(h) < 0) {
2763                 s0->first_field = 0;
2764                 return -1;
2765             }
2766         } else {
2767             ff_release_unused_pictures(s, 0);
2768         }
2769     }
2770     if(h != h0)
2771         clone_slice(h, h0);
2772
2773     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2774
2775     assert(s->mb_num == s->mb_width * s->mb_height);
2776     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2777        first_mb_in_slice                    >= s->mb_num){
2778         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2779         return -1;
2780     }
2781     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2782     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2783     if (s->picture_structure == PICT_BOTTOM_FIELD)
2784         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2785     assert(s->mb_y < s->mb_height);
2786
2787     if(s->picture_structure==PICT_FRAME){
2788         h->curr_pic_num=   h->frame_num;
2789         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2790     }else{
2791         h->curr_pic_num= 2*h->frame_num + 1;
2792         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2793     }
2794
2795     if(h->nal_unit_type == NAL_IDR_SLICE){
2796         get_ue_golomb(&s->gb); /* idr_pic_id */
2797     }
2798
2799     if(h->sps.poc_type==0){
2800         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2801
2802         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2803             h->delta_poc_bottom= get_se_golomb(&s->gb);
2804         }
2805     }
2806
2807     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2808         h->delta_poc[0]= get_se_golomb(&s->gb);
2809
2810         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2811             h->delta_poc[1]= get_se_golomb(&s->gb);
2812     }
2813
2814     init_poc(h);
2815
2816     if(h->pps.redundant_pic_cnt_present){
2817         h->redundant_pic_count= get_ue_golomb(&s->gb);
2818     }
2819
2820     //set defaults, might be overridden a few lines later
2821     h->ref_count[0]= h->pps.ref_count[0];
2822     h->ref_count[1]= h->pps.ref_count[1];
2823
2824     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2825         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2826             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2827         }
2828         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2829
2830         if(num_ref_idx_active_override_flag){
2831             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2832             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2833                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2834
2835             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2836                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2837                 h->ref_count[0]= h->ref_count[1]= 1;
2838                 return -1;
2839             }
2840         }
2841         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2842             h->list_count= 2;
2843         else
2844             h->list_count= 1;
2845     }else
2846         h->list_count= 0;
2847
2848     if(!default_ref_list_done){
2849         ff_h264_fill_default_ref_list(h);
2850     }
2851
2852     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0)
2853         return -1;
2854
2855     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2856         s->last_picture_ptr= &h->ref_list[0][0];
2857         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2858     }
2859     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2860         s->next_picture_ptr= &h->ref_list[1][0];
2861         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2862     }
2863
2864     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2865        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2866         pred_weight_table(h);
2867     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2868         implicit_weight_table(h, -1);
2869     }else {
2870         h->use_weight = 0;
2871         for (i = 0; i < 2; i++) {
2872             h->luma_weight_flag[i]   = 0;
2873             h->chroma_weight_flag[i] = 0;
2874         }
2875     }
2876
2877     if(h->nal_ref_idc)
2878         ff_h264_decode_ref_pic_marking(h0, &s->gb);
2879
2880     if(FRAME_MBAFF){
2881         ff_h264_fill_mbaff_ref_list(h);
2882
2883         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2884             implicit_weight_table(h, 0);
2885             implicit_weight_table(h, 1);
2886         }
2887     }
2888
2889     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2890         ff_h264_direct_dist_scale_factor(h);
2891     ff_h264_direct_ref_list_init(h);
2892
2893     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2894         tmp = get_ue_golomb_31(&s->gb);
2895         if(tmp > 2){
2896             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2897             return -1;
2898         }
2899         h->cabac_init_idc= tmp;
2900     }
2901
2902     h->last_qscale_diff = 0;
2903     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2904     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
2905         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2906         return -1;
2907     }
2908     s->qscale= tmp;
2909     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2910     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2911     //FIXME qscale / qp ... stuff
2912     if(h->slice_type == AV_PICTURE_TYPE_SP){
2913         get_bits1(&s->gb); /* sp_for_switch_flag */
2914     }
2915     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
2916         get_se_golomb(&s->gb); /* slice_qs_delta */
2917     }
2918
2919     h->deblocking_filter = 1;
2920     h->slice_alpha_c0_offset = 52;
2921     h->slice_beta_offset = 52;
2922     if( h->pps.deblocking_filter_parameters_present ) {
2923         tmp= get_ue_golomb_31(&s->gb);
2924         if(tmp > 2){
2925             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2926             return -1;
2927         }
2928         h->deblocking_filter= tmp;
2929         if(h->deblocking_filter < 2)
2930             h->deblocking_filter^= 1; // 1<->0
2931
2932         if( h->deblocking_filter ) {
2933             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2934             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2935             if(   h->slice_alpha_c0_offset > 104U
2936                || h->slice_beta_offset     > 104U){
2937                 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);
2938                 return -1;
2939             }
2940         }
2941     }
2942
2943     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2944        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
2945        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
2946        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2947         h->deblocking_filter= 0;
2948
2949     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
2950         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
2951             /* Cheat slightly for speed:
2952                Do not bother to deblock across slices. */
2953             h->deblocking_filter = 2;
2954         } else {
2955             h0->max_contexts = 1;
2956             if(!h0->single_decode_warning) {
2957                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
2958                 h0->single_decode_warning = 1;
2959             }
2960             if (h != h0) {
2961                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
2962                 return 1;
2963             }
2964         }
2965     }
2966     h->qp_thresh= 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1]);
2967
2968 #if 0 //FMO
2969     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
2970         slice_group_change_cycle= get_bits(&s->gb, ?);
2971 #endif
2972
2973     h0->last_slice_type = slice_type;
2974     h->slice_num = ++h0->current_slice;
2975     if(h->slice_num >= MAX_SLICES){
2976         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
2977     }
2978
2979     for(j=0; j<2; j++){
2980         int id_list[16];
2981         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
2982         for(i=0; i<16; i++){
2983             id_list[i]= 60;
2984             if(h->ref_list[j][i].data[0]){
2985                 int k;
2986                 uint8_t *base= h->ref_list[j][i].base[0];
2987                 for(k=0; k<h->short_ref_count; k++)
2988                     if(h->short_ref[k]->base[0] == base){
2989                         id_list[i]= k;
2990                         break;
2991                     }
2992                 for(k=0; k<h->long_ref_count; k++)
2993                     if(h->long_ref[k] && h->long_ref[k]->base[0] == base){
2994                         id_list[i]= h->short_ref_count + k;
2995                         break;
2996                     }
2997             }
2998         }
2999
3000         ref2frm[0]=
3001         ref2frm[1]= -1;
3002         for(i=0; i<16; i++)
3003             ref2frm[i+2]= 4*id_list[i]
3004                           +(h->ref_list[j][i].reference&3);
3005         ref2frm[18+0]=
3006         ref2frm[18+1]= -1;
3007         for(i=16; i<48; i++)
3008             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3009                           +(h->ref_list[j][i].reference&3);
3010     }
3011
3012     //FIXME: fix draw_edges+PAFF+frame threads
3013     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3014     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3015
3016     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3017         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",
3018                h->slice_num,
3019                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3020                first_mb_in_slice,
3021                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3022                pps_id, h->frame_num,
3023                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3024                h->ref_count[0], h->ref_count[1],
3025                s->qscale,
3026                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3027                h->use_weight,
3028                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3029                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3030                );
3031     }
3032
3033     return 0;
3034 }
3035
3036 int ff_h264_get_slice_type(const H264Context *h)
3037 {
3038     switch (h->slice_type) {
3039     case AV_PICTURE_TYPE_P:  return 0;
3040     case AV_PICTURE_TYPE_B:  return 1;
3041     case AV_PICTURE_TYPE_I:  return 2;
3042     case AV_PICTURE_TYPE_SP: return 3;
3043     case AV_PICTURE_TYPE_SI: return 4;
3044     default:         return -1;
3045     }
3046 }
3047
3048 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3049                                                       int left_xy[2], int top_type, int left_type[2], int mb_xy, int list)
3050 {
3051     int b_stride = h->b_stride;
3052     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3053     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3054     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3055         if(USES_LIST(top_type, list)){
3056             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3057             const int b8_xy= 4*top_xy + 2;
3058             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3059             AV_COPY128(mv_dst - 1*8, s->current_picture.motion_val[list][b_xy + 0]);
3060             ref_cache[0 - 1*8]=
3061             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 0]];
3062             ref_cache[2 - 1*8]=
3063             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 1]];
3064         }else{
3065             AV_ZERO128(mv_dst - 1*8);
3066             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3067         }
3068
3069         if(!IS_INTERLACED(mb_type^left_type[0])){
3070             if(USES_LIST(left_type[0], list)){
3071                 const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
3072                 const int b8_xy= 4*left_xy[0] + 1;
3073                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[0]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3074                 AV_COPY32(mv_dst - 1 + 0, s->current_picture.motion_val[list][b_xy + b_stride*0]);
3075                 AV_COPY32(mv_dst - 1 + 8, s->current_picture.motion_val[list][b_xy + b_stride*1]);
3076                 AV_COPY32(mv_dst - 1 +16, s->current_picture.motion_val[list][b_xy + b_stride*2]);
3077                 AV_COPY32(mv_dst - 1 +24, s->current_picture.motion_val[list][b_xy + b_stride*3]);
3078                 ref_cache[-1 +  0]=
3079                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*0]];
3080                 ref_cache[-1 + 16]=
3081                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*1]];
3082             }else{
3083                 AV_ZERO32(mv_dst - 1 + 0);
3084                 AV_ZERO32(mv_dst - 1 + 8);
3085                 AV_ZERO32(mv_dst - 1 +16);
3086                 AV_ZERO32(mv_dst - 1 +24);
3087                 ref_cache[-1 +  0]=
3088                 ref_cache[-1 +  8]=
3089                 ref_cache[-1 + 16]=
3090                 ref_cache[-1 + 24]= LIST_NOT_USED;
3091             }
3092         }
3093     }
3094
3095     if(!USES_LIST(mb_type, list)){
3096         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3097         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3098         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3099         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3100         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3101         return;
3102     }
3103
3104     {
3105         int8_t *ref = &s->current_picture.ref_index[list][4*mb_xy];
3106         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3107         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3108         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3109         AV_WN32A(&ref_cache[0*8], ref01);
3110         AV_WN32A(&ref_cache[1*8], ref01);
3111         AV_WN32A(&ref_cache[2*8], ref23);
3112         AV_WN32A(&ref_cache[3*8], ref23);
3113     }
3114
3115     {
3116         int16_t (*mv_src)[2] = &s->current_picture.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3117         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3118         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3119         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3120         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3121     }
3122 }
3123
3124 /**
3125  *
3126  * @return non zero if the loop filter can be skiped
3127  */
3128 static int fill_filter_caches(H264Context *h, int mb_type){
3129     MpegEncContext * const s = &h->s;
3130     const int mb_xy= h->mb_xy;
3131     int top_xy, left_xy[2];
3132     int top_type, left_type[2];
3133     uint8_t *nnz;
3134     uint8_t *nnz_cache;
3135
3136     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3137
3138     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3139      * stuff, I can't imagine that these complex rules are worth it. */
3140
3141     left_xy[1] = left_xy[0] = mb_xy-1;
3142     if(FRAME_MBAFF){
3143         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[mb_xy-1]);
3144         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3145         if(s->mb_y&1){
3146             if (left_mb_field_flag != curr_mb_field_flag) {
3147                 left_xy[0] -= s->mb_stride;
3148             }
3149         }else{
3150             if(curr_mb_field_flag){
3151                 top_xy      += s->mb_stride & (((s->current_picture.mb_type[top_xy    ]>>7)&1)-1);
3152             }
3153             if (left_mb_field_flag != curr_mb_field_flag) {
3154                 left_xy[1] += s->mb_stride;
3155             }
3156         }
3157     }
3158
3159     h->top_mb_xy = top_xy;
3160     h->left_mb_xy[0] = left_xy[0];
3161     h->left_mb_xy[1] = left_xy[1];
3162     {
3163         //for sufficiently low qp, filtering wouldn't do anything
3164         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3165         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3166         int qp = s->current_picture.qscale_table[mb_xy];
3167         if(qp <= qp_thresh
3168            && (left_xy[0]<0 || ((qp + s->current_picture.qscale_table[left_xy[0]] + 1)>>1) <= qp_thresh)
3169            && (top_xy   < 0 || ((qp + s->current_picture.qscale_table[top_xy    ] + 1)>>1) <= qp_thresh)){
3170             if(!FRAME_MBAFF)
3171                 return 1;
3172             if(   (left_xy[0]< 0            || ((qp + s->current_picture.qscale_table[left_xy[1]             ] + 1)>>1) <= qp_thresh)
3173                && (top_xy    < s->mb_stride || ((qp + s->current_picture.qscale_table[top_xy    -s->mb_stride] + 1)>>1) <= qp_thresh))
3174                 return 1;
3175         }
3176     }
3177
3178     top_type     = s->current_picture.mb_type[top_xy]    ;
3179     left_type[0] = s->current_picture.mb_type[left_xy[0]];
3180     left_type[1] = s->current_picture.mb_type[left_xy[1]];
3181     if(h->deblocking_filter == 2){
3182         if(h->slice_table[top_xy     ] != h->slice_num) top_type= 0;
3183         if(h->slice_table[left_xy[0] ] != h->slice_num) left_type[0]= left_type[1]= 0;
3184     }else{
3185         if(h->slice_table[top_xy     ] == 0xFFFF) top_type= 0;
3186         if(h->slice_table[left_xy[0] ] == 0xFFFF) left_type[0]= left_type[1] =0;
3187     }
3188     h->top_type    = top_type    ;
3189     h->left_type[0]= left_type[0];
3190     h->left_type[1]= left_type[1];
3191
3192     if(IS_INTRA(mb_type))
3193         return 0;
3194
3195     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3196     if(h->list_count == 2)
3197         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3198
3199     nnz = h->non_zero_count[mb_xy];
3200     nnz_cache = h->non_zero_count_cache;
3201     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3202     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3203     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3204     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3205     h->cbp= h->cbp_table[mb_xy];
3206
3207     if(top_type){
3208         nnz = h->non_zero_count[top_xy];
3209         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3210     }
3211
3212     if(left_type[0]){
3213         nnz = h->non_zero_count[left_xy[0]];
3214         nnz_cache[3+8*1]= nnz[3+0*4];
3215         nnz_cache[3+8*2]= nnz[3+1*4];
3216         nnz_cache[3+8*3]= nnz[3+2*4];
3217         nnz_cache[3+8*4]= nnz[3+3*4];
3218     }
3219
3220     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3221     if(!CABAC && h->pps.transform_8x8_mode){
3222         if(IS_8x8DCT(top_type)){
3223             nnz_cache[4+8*0]=
3224             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3225             nnz_cache[6+8*0]=
3226             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3227         }
3228         if(IS_8x8DCT(left_type[0])){
3229             nnz_cache[3+8*1]=
3230             nnz_cache[3+8*2]= (h->cbp_table[left_xy[0]]&0x2000) >> 12; //FIXME check MBAFF
3231         }
3232         if(IS_8x8DCT(left_type[1])){
3233             nnz_cache[3+8*3]=
3234             nnz_cache[3+8*4]= (h->cbp_table[left_xy[1]]&0x8000) >> 12; //FIXME check MBAFF
3235         }
3236
3237         if(IS_8x8DCT(mb_type)){
3238             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3239             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3240
3241             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3242             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3243
3244             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3245             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3246
3247             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3248             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3249         }
3250     }
3251
3252     return 0;
3253 }
3254
3255 static void loop_filter(H264Context *h, int start_x, int end_x){
3256     MpegEncContext * const s = &h->s;
3257     uint8_t  *dest_y, *dest_cb, *dest_cr;
3258     int linesize, uvlinesize, mb_x, mb_y;
3259     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3260     const int old_slice_type= h->slice_type;
3261     const int pixel_shift = h->pixel_shift;
3262
3263     if(h->deblocking_filter) {
3264         for(mb_x= start_x; mb_x<end_x; mb_x++){
3265             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3266                 int mb_xy, mb_type;
3267                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3268                 h->slice_num= h->slice_table[mb_xy];
3269                 mb_type= s->current_picture.mb_type[mb_xy];
3270                 h->list_count= h->list_counts[mb_xy];
3271
3272                 if(FRAME_MBAFF)
3273                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3274
3275                 s->mb_x= mb_x;
3276                 s->mb_y= mb_y;
3277                 dest_y  = s->current_picture.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3278                 dest_cb = s->current_picture.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * (8 << CHROMA444);
3279                 dest_cr = s->current_picture.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * (8 << CHROMA444);
3280                     //FIXME simplify above
3281
3282                 if (MB_FIELD) {
3283                     linesize   = h->mb_linesize   = s->linesize * 2;
3284                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3285                     if(mb_y&1){ //FIXME move out of this function?
3286                         dest_y -= s->linesize*15;
3287                         dest_cb-= s->uvlinesize*((8 << CHROMA444)-1);
3288                         dest_cr-= s->uvlinesize*((8 << CHROMA444)-1);
3289                     }
3290                 } else {
3291                     linesize   = h->mb_linesize   = s->linesize;
3292                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3293                 }
3294                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, CHROMA444, 0);
3295                 if(fill_filter_caches(h, mb_type))
3296                     continue;
3297                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
3298                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
3299
3300                 if (FRAME_MBAFF) {
3301                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3302                 } else {
3303                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3304                 }
3305             }
3306         }
3307     }
3308     h->slice_type= old_slice_type;
3309     s->mb_x= end_x;
3310     s->mb_y= end_mb_y - FRAME_MBAFF;
3311     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3312     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3313 }
3314
3315 static void predict_field_decoding_flag(H264Context *h){
3316     MpegEncContext * const s = &h->s;
3317     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3318     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3319                 ? s->current_picture.mb_type[mb_xy-1]
3320                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3321                 ? s->current_picture.mb_type[mb_xy-s->mb_stride]
3322                 : 0;
3323     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3324 }
3325
3326 /**
3327  * Draw edges and report progress for the last MB row.
3328  */
3329 static void decode_finish_row(H264Context *h){
3330     MpegEncContext * const s = &h->s;
3331     int top = 16*(s->mb_y >> FIELD_PICTURE);
3332     int height = 16 << FRAME_MBAFF;
3333     int deblock_border = (16 + 4) << FRAME_MBAFF;
3334     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3335
3336     if (h->deblocking_filter) {
3337         if((top + height) >= pic_height)
3338             height += deblock_border;
3339
3340         top -= deblock_border;
3341     }
3342
3343     if (top >= pic_height || (top + height) < h->emu_edge_height)
3344         return;
3345
3346     height = FFMIN(height, pic_height - top);
3347     if (top < h->emu_edge_height) {
3348         height = top+height;
3349         top = 0;
3350     }
3351
3352     ff_draw_horiz_band(s, top, height);
3353
3354     if (s->dropable) return;
3355
3356     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3357                              s->picture_structure==PICT_BOTTOM_FIELD);
3358 }
3359
3360 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3361     H264Context *h = *(void**)arg;
3362     MpegEncContext * const s = &h->s;
3363     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3364     int lf_x_start = s->mb_x;
3365
3366     s->mb_skip_run= -1;
3367
3368     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3369                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3370
3371     if( h->pps.cabac ) {
3372         /* realign */
3373         align_get_bits( &s->gb );
3374
3375         /* init cabac */
3376         ff_init_cabac_states( &h->cabac);
3377         ff_init_cabac_decoder( &h->cabac,
3378                                s->gb.buffer + get_bits_count(&s->gb)/8,
3379                                (get_bits_left(&s->gb) + 7)/8);
3380
3381         ff_h264_init_cabac_states(h);
3382
3383         for(;;){
3384 //START_TIMER
3385             int ret = ff_h264_decode_mb_cabac(h);
3386             int eos;
3387 //STOP_TIMER("decode_mb_cabac")
3388
3389             if(ret>=0) ff_h264_hl_decode_mb(h);
3390
3391             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3392                 s->mb_y++;
3393
3394                 ret = ff_h264_decode_mb_cabac(h);
3395
3396                 if(ret>=0) ff_h264_hl_decode_mb(h);
3397                 s->mb_y--;
3398             }
3399             eos = get_cabac_terminate( &h->cabac );
3400
3401             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3402                 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);
3403                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3404                 return 0;
3405             }
3406             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3407                 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);
3408                 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);
3409                 return -1;
3410             }
3411
3412             if( ++s->mb_x >= s->mb_width ) {
3413                 loop_filter(h, lf_x_start, s->mb_x);
3414                 s->mb_x = lf_x_start = 0;
3415                 decode_finish_row(h);
3416                 ++s->mb_y;
3417                 if(FIELD_OR_MBAFF_PICTURE) {
3418                     ++s->mb_y;
3419                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3420                         predict_field_decoding_flag(h);
3421                 }
3422             }
3423
3424             if( eos || s->mb_y >= s->mb_height ) {
3425                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3426                 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);
3427                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3428                 return 0;
3429             }
3430         }
3431
3432     } else {
3433         for(;;){
3434             int ret = ff_h264_decode_mb_cavlc(h);
3435
3436             if(ret>=0) ff_h264_hl_decode_mb(h);
3437
3438             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3439                 s->mb_y++;
3440                 ret = ff_h264_decode_mb_cavlc(h);
3441
3442                 if(ret>=0) ff_h264_hl_decode_mb(h);
3443                 s->mb_y--;
3444             }
3445
3446             if(ret<0){
3447                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3448                 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);
3449                 return -1;
3450             }
3451
3452             if(++s->mb_x >= s->mb_width){
3453                 loop_filter(h, lf_x_start, s->mb_x);
3454                 s->mb_x = lf_x_start = 0;
3455                 decode_finish_row(h);
3456                 ++s->mb_y;
3457                 if(FIELD_OR_MBAFF_PICTURE) {
3458                     ++s->mb_y;
3459                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3460                         predict_field_decoding_flag(h);
3461                 }
3462                 if(s->mb_y >= s->mb_height){
3463                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3464
3465                     if(get_bits_count(&s->gb) == s->gb.size_in_bits ) {
3466                         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);