4:4:4 H.264 decoding support
[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 void ff_h264_write_back_intra_pred_mode(H264Context *h){
64     int8_t *mode= h->intra4x4_pred_mode + h->mb2br_xy[h->mb_xy];
65
66     AV_COPY32(mode, h->intra4x4_pred_mode_cache + 4 + 8*4);
67     mode[4]= h->intra4x4_pred_mode_cache[7+8*3];
68     mode[5]= h->intra4x4_pred_mode_cache[7+8*2];
69     mode[6]= h->intra4x4_pred_mode_cache[7+8*1];
70 }
71
72 /**
73  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
74  */
75 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
76     MpegEncContext * const s = &h->s;
77     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
78     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
79     int i;
80
81     if(!(h->top_samples_available&0x8000)){
82         for(i=0; i<4; i++){
83             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
84             if(status<0){
85                 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);
86                 return -1;
87             } else if(status){
88                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
89             }
90         }
91     }
92
93     if((h->left_samples_available&0x8888)!=0x8888){
94         static const int mask[4]={0x8000,0x2000,0x80,0x20};
95         for(i=0; i<4; i++){
96             if(!(h->left_samples_available&mask[i])){
97                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
98                 if(status<0){
99                     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);
100                     return -1;
101                 } else if(status){
102                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
103                 }
104             }
105         }
106     }
107
108     return 0;
109 } //FIXME cleanup like ff_h264_check_intra_pred_mode
110
111 /**
112  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
113  */
114 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
115     MpegEncContext * const s = &h->s;
116     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
117     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
118
119     if(mode > 6U) {
120         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);
121         return -1;
122     }
123
124     if(!(h->top_samples_available&0x8000)){
125         mode= top[ mode ];
126         if(mode<0){
127             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);
128             return -1;
129         }
130     }
131
132     if((h->left_samples_available&0x8080) != 0x8080){
133         mode= left[ mode ];
134         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
135             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
136         }
137         if(mode<0){
138             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);
139             return -1;
140         }
141     }
142
143     return mode;
144 }
145
146 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
147     int i, si, di;
148     uint8_t *dst;
149     int bufidx;
150
151 //    src[0]&0x80;                //forbidden bit
152     h->nal_ref_idc= src[0]>>5;
153     h->nal_unit_type= src[0]&0x1F;
154
155     src++; length--;
156
157 #if HAVE_FAST_UNALIGNED
158 # if HAVE_FAST_64BIT
159 #   define RS 7
160     for(i=0; i+1<length; i+=9){
161         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
162 # else
163 #   define RS 3
164     for(i=0; i+1<length; i+=5){
165         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
166 # endif
167             continue;
168         if(i>0 && !src[i]) i--;
169         while(src[i]) i++;
170 #else
171 #   define RS 0
172     for(i=0; i+1<length; i+=2){
173         if(src[i]) continue;
174         if(i>0 && src[i-1]==0) i--;
175 #endif
176         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
177             if(src[i+2]!=3){
178                 /* startcode, so we must be past the end */
179                 length=i;
180             }
181             break;
182         }
183         i-= RS;
184     }
185
186     if(i>=length-1){ //no escaped 0
187         *dst_length= length;
188         *consumed= length+1; //+1 for the header
189         return src;
190     }
191
192     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
193     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
194     dst= h->rbsp_buffer[bufidx];
195
196     if (dst == NULL){
197         return NULL;
198     }
199
200 //printf("decoding esc\n");
201     memcpy(dst, src, i);
202     si=di=i;
203     while(si+2<length){
204         //remove escapes (very rare 1:2^22)
205         if(src[si+2]>3){
206             dst[di++]= src[si++];
207             dst[di++]= src[si++];
208         }else if(src[si]==0 && src[si+1]==0){
209             if(src[si+2]==3){ //escape
210                 dst[di++]= 0;
211                 dst[di++]= 0;
212                 si+=3;
213                 continue;
214             }else //next start code
215                 goto nsc;
216         }
217
218         dst[di++]= src[si++];
219     }
220     while(si<length)
221         dst[di++]= src[si++];
222 nsc:
223
224     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
225
226     *dst_length= di;
227     *consumed= si + 1;//+1 for the header
228 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
229     return dst;
230 }
231
232 /**
233  * Identify the exact end of the bitstream
234  * @return the length of the trailing, or 0 if damaged
235  */
236 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
237     int v= *src;
238     int r;
239
240     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
241
242     for(r=1; r<9; r++){
243         if(v&1) return r;
244         v>>=1;
245     }
246     return 0;
247 }
248
249 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
250                                  int y_offset, int list){
251     int raw_my= h->mv_cache[list][ scan8[n] ][1];
252     int filter_height= (raw_my&3) ? 2 : 0;
253     int full_my= (raw_my>>2) + y_offset;
254     int top = full_my - filter_height, bottom = full_my + height + filter_height;
255
256     return FFMAX(abs(top), bottom);
257 }
258
259 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
260                                int y_offset, int list0, int list1, int *nrefs){
261     MpegEncContext * const s = &h->s;
262     int my;
263
264     y_offset += 16*(s->mb_y >> MB_FIELD);
265
266     if(list0){
267         int ref_n = h->ref_cache[0][ scan8[n] ];
268         Picture *ref= &h->ref_list[0][ref_n];
269
270         // Error resilience puts the current picture in the ref list.
271         // Don't try to wait on these as it will cause a deadlock.
272         // Fields can wait on each other, though.
273         if(ref->thread_opaque != s->current_picture.thread_opaque ||
274            (ref->reference&3) != s->picture_structure) {
275             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
276             if (refs[0][ref_n] < 0) nrefs[0] += 1;
277             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
278         }
279     }
280
281     if(list1){
282         int ref_n = h->ref_cache[1][ scan8[n] ];
283         Picture *ref= &h->ref_list[1][ref_n];
284
285         if(ref->thread_opaque != s->current_picture.thread_opaque ||
286            (ref->reference&3) != s->picture_structure) {
287             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
288             if (refs[1][ref_n] < 0) nrefs[1] += 1;
289             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
290         }
291     }
292 }
293
294 /**
295  * Wait until all reference frames are available for MC operations.
296  *
297  * @param h the H264 context
298  */
299 static void await_references(H264Context *h){
300     MpegEncContext * const s = &h->s;
301     const int mb_xy= h->mb_xy;
302     const int mb_type= s->current_picture.mb_type[mb_xy];
303     int refs[2][48];
304     int nrefs[2] = {0};
305     int ref, list;
306
307     memset(refs, -1, sizeof(refs));
308
309     if(IS_16X16(mb_type)){
310         get_lowest_part_y(h, refs, 0, 16, 0,
311                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
312     }else if(IS_16X8(mb_type)){
313         get_lowest_part_y(h, refs, 0, 8, 0,
314                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
315         get_lowest_part_y(h, refs, 8, 8, 8,
316                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
317     }else if(IS_8X16(mb_type)){
318         get_lowest_part_y(h, refs, 0, 16, 0,
319                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
320         get_lowest_part_y(h, refs, 4, 16, 0,
321                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
322     }else{
323         int i;
324
325         assert(IS_8X8(mb_type));
326
327         for(i=0; i<4; i++){
328             const int sub_mb_type= h->sub_mb_type[i];
329             const int n= 4*i;
330             int y_offset= (i&2)<<2;
331
332             if(IS_SUB_8X8(sub_mb_type)){
333                 get_lowest_part_y(h, refs, n  , 8, y_offset,
334                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
335             }else if(IS_SUB_8X4(sub_mb_type)){
336                 get_lowest_part_y(h, refs, n  , 4, y_offset,
337                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
338                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
339                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
340             }else if(IS_SUB_4X8(sub_mb_type)){
341                 get_lowest_part_y(h, refs, n  , 8, y_offset,
342                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
343                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
344                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
345             }else{
346                 int j;
347                 assert(IS_SUB_4X4(sub_mb_type));
348                 for(j=0; j<4; j++){
349                     int sub_y_offset= y_offset + 2*(j&2);
350                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
351                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
352                 }
353             }
354         }
355     }
356
357     for(list=h->list_count-1; list>=0; list--){
358         for(ref=0; ref<48 && nrefs[list]; ref++){
359             int row = refs[list][ref];
360             if(row >= 0){
361                 Picture *ref_pic = &h->ref_list[list][ref];
362                 int ref_field = ref_pic->reference - 1;
363                 int ref_field_picture = ref_pic->field_picture;
364                 int pic_height = 16*s->mb_height >> ref_field_picture;
365
366                 row <<= MB_MBAFF;
367                 nrefs[list]--;
368
369                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
370                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
371                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
372                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
373                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
374                 }else if(FIELD_PICTURE){
375                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
376                 }else{
377                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
378                 }
379             }
380         }
381     }
382 }
383
384 #if 0
385 /**
386  * DCT transforms the 16 dc values.
387  * @param qp quantization parameter ??? FIXME
388  */
389 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
390 //    const int qmul= dequant_coeff[qp][0];
391     int i;
392     int temp[16]; //FIXME check if this is a good idea
393     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
394     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
395
396     for(i=0; i<4; i++){
397         const int offset= y_offset[i];
398         const int z0= block[offset+stride*0] + block[offset+stride*4];
399         const int z1= block[offset+stride*0] - block[offset+stride*4];
400         const int z2= block[offset+stride*1] - block[offset+stride*5];
401         const int z3= block[offset+stride*1] + block[offset+stride*5];
402
403         temp[4*i+0]= z0+z3;
404         temp[4*i+1]= z1+z2;
405         temp[4*i+2]= z1-z2;
406         temp[4*i+3]= z0-z3;
407     }
408
409     for(i=0; i<4; i++){
410         const int offset= x_offset[i];
411         const int z0= temp[4*0+i] + temp[4*2+i];
412         const int z1= temp[4*0+i] - temp[4*2+i];
413         const int z2= temp[4*1+i] - temp[4*3+i];
414         const int z3= temp[4*1+i] + temp[4*3+i];
415
416         block[stride*0 +offset]= (z0 + z3)>>1;
417         block[stride*2 +offset]= (z1 + z2)>>1;
418         block[stride*8 +offset]= (z1 - z2)>>1;
419         block[stride*10+offset]= (z0 - z3)>>1;
420     }
421 }
422 #endif
423
424 #undef xStride
425 #undef stride
426
427 #if 0
428 static void chroma_dc_dct_c(DCTELEM *block){
429     const int stride= 16*2;
430     const int xStride= 16;
431     int a,b,c,d,e;
432
433     a= block[stride*0 + xStride*0];
434     b= block[stride*0 + xStride*1];
435     c= block[stride*1 + xStride*0];
436     d= block[stride*1 + xStride*1];
437
438     e= a-b;
439     a= a+b;
440     b= c-d;
441     c= c+d;
442
443     block[stride*0 + xStride*0]= (a+c);
444     block[stride*0 + xStride*1]= (e+b);
445     block[stride*1 + xStride*0]= (a-c);
446     block[stride*1 + xStride*1]= (e-b);
447 }
448 #endif
449
450 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
451                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
452                            int src_x_offset, int src_y_offset,
453                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
454                            int pixel_shift, int chroma444){
455     MpegEncContext * const s = &h->s;
456     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
457     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
458     const int luma_xy= (mx&3) + ((my&3)<<2);
459     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
460     uint8_t * src_y = pic->data[0] + offset;
461     uint8_t * src_cb, * src_cr;
462     int extra_width= h->emu_edge_width;
463     int extra_height= h->emu_edge_height;
464     int emu=0;
465     const int full_mx= mx>>2;
466     const int full_my= my>>2;
467     const int pic_width  = 16*s->mb_width;
468     const int pic_height = 16*s->mb_height >> MB_FIELD;
469
470     if(mx&7) extra_width -= 3;
471     if(my&7) extra_height -= 3;
472
473     if(   full_mx < 0-extra_width
474        || full_my < 0-extra_height
475        || full_mx + 16/*FIXME*/ > pic_width + extra_width
476        || full_my + 16/*FIXME*/ > pic_height + extra_height){
477         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);
478             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
479         emu=1;
480     }
481
482     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
483     if(!square){
484         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
485     }
486
487     if(chroma444){
488         src_cb = pic->data[1] + offset;
489         if(emu){
490             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
491                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
492             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
493         }
494         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
495         if(!square){
496             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
497         }
498
499         src_cr = pic->data[2] + offset;
500         if(emu){
501             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
502                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
503             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
504         }
505         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
506         if(!square){
507             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
508         }
509         return;
510     }
511
512     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
513
514     if(MB_FIELD){
515         // chroma offset when predicting from a field of opposite parity
516         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
517         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
518     }
519     src_cb= pic->data[1] + ((mx>>3) << pixel_shift) + (my>>3)*h->mb_uvlinesize;
520     src_cr= pic->data[2] + ((mx>>3) << pixel_shift) + (my>>3)*h->mb_uvlinesize;
521
522     if(emu){
523         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);
524             src_cb= s->edge_emu_buffer;
525     }
526     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
527
528     if(emu){
529         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);
530             src_cr= s->edge_emu_buffer;
531     }
532     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
533 }
534
535 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
536                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
537                            int x_offset, int y_offset,
538                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
539                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
540                            int list0, int list1, int pixel_shift, int chroma444){
541     MpegEncContext * const s = &h->s;
542     qpel_mc_func *qpix_op=  qpix_put;
543     h264_chroma_mc_func chroma_op= chroma_put;
544
545     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
546     if(chroma444){
547         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
548         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
549     }else{
550         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
551         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
552     }
553     x_offset += 8*s->mb_x;
554     y_offset += 8*(s->mb_y >> MB_FIELD);
555
556     if(list0){
557         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
558         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
559                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
560                            qpix_op, chroma_op, pixel_shift, chroma444);
561
562         qpix_op=  qpix_avg;
563         chroma_op= chroma_avg;
564     }
565
566     if(list1){
567         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
568         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
569                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
570                            qpix_op, chroma_op, pixel_shift, chroma444);
571     }
572 }
573
574 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
575                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
576                            int x_offset, int y_offset,
577                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
578                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
579                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
580                            int list0, int list1, int pixel_shift, int chroma444){
581     MpegEncContext * const s = &h->s;
582
583     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
584     if(chroma444){
585         chroma_weight_avg = luma_weight_avg;
586         chroma_weight_op = luma_weight_op;
587         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
588         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
589     }else{
590         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
591         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
592     }
593     x_offset += 8*s->mb_x;
594     y_offset += 8*(s->mb_y >> MB_FIELD);
595
596     if(list0 && list1){
597         /* don't optimize for luma-only case, since B-frames usually
598          * use implicit weights => chroma too. */
599         uint8_t *tmp_cb = s->obmc_scratchpad;
600         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
601         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
602         int refn0 = h->ref_cache[0][ scan8[n] ];
603         int refn1 = h->ref_cache[1][ scan8[n] ];
604
605         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
606                     dest_y, dest_cb, dest_cr,
607                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
608         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
609                     tmp_y, tmp_cb, tmp_cr,
610                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
611
612         if(h->use_weight == 2){
613             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
614             int weight1 = 64 - weight0;
615             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
616             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
617             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
618         }else{
619             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
620                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
621                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
622             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
623                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
624                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
625             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
626                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
627                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
628         }
629     }else{
630         int list = list1 ? 1 : 0;
631         int refn = h->ref_cache[list][ scan8[n] ];
632         Picture *ref= &h->ref_list[list][refn];
633         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
634                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
635                     qpix_put, chroma_put, pixel_shift, chroma444);
636
637         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
638                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
639         if(h->use_weight_chroma){
640             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
641                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
642             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
643                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
644         }
645     }
646 }
647
648 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
649                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
650                            int x_offset, int y_offset,
651                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
652                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
653                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
654                            int list0, int list1, int pixel_shift, int chroma444){
655     if((h->use_weight==2 && list0 && list1
656         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
657        || h->use_weight==1)
658         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
659                          x_offset, y_offset, qpix_put, chroma_put,
660                          weight_op[0], weight_op[3], weight_avg[0],
661                          weight_avg[3], list0, list1, pixel_shift, chroma444);
662     else
663         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
664                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
665                     chroma_avg, list0, list1, pixel_shift, chroma444);
666 }
667
668 static inline void prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma444){
669     /* fetch pixels for estimated mv 4 macroblocks ahead
670      * optimized for 64byte cache lines */
671     MpegEncContext * const s = &h->s;
672     const int refn = h->ref_cache[list][scan8[0]];
673     if(refn >= 0){
674         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
675         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
676         uint8_t **src= h->ref_list[list][refn].data;
677         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
678         s->dsp.prefetch(src[0]+off, s->linesize, 4);
679         if(chroma444){
680             s->dsp.prefetch(src[1]+off, s->linesize, 4);
681             s->dsp.prefetch(src[2]+off, s->linesize, 4);
682         }else{
683             off= ((mx>>1) << pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + (64 << pixel_shift);
684             s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
685         }
686     }
687 }
688
689 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
690                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
691                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
692                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
693                       int pixel_shift, int chroma444){
694     MpegEncContext * const s = &h->s;
695     const int mb_xy= h->mb_xy;
696     const int mb_type= s->current_picture.mb_type[mb_xy];
697
698     assert(IS_INTER(mb_type));
699
700     if(HAVE_PTHREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
701         await_references(h);
702     prefetch_motion(h, 0, pixel_shift, chroma444);
703
704     if(IS_16X16(mb_type)){
705         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
706                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
707                 weight_op, weight_avg,
708                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
709                 pixel_shift, chroma444);
710     }else if(IS_16X8(mb_type)){
711         mc_part(h, 0, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
712                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
713                 &weight_op[1], &weight_avg[1],
714                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
715                 pixel_shift, chroma444);
716         mc_part(h, 8, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
717                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
718                 &weight_op[1], &weight_avg[1],
719                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
720                 pixel_shift, chroma444);
721     }else if(IS_8X16(mb_type)){
722         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
723                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
724                 &weight_op[2], &weight_avg[2],
725                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
726                 pixel_shift, chroma444);
727         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
728                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
729                 &weight_op[2], &weight_avg[2],
730                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
731                 pixel_shift, chroma444);
732     }else{
733         int i;
734
735         assert(IS_8X8(mb_type));
736
737         for(i=0; i<4; i++){
738             const int sub_mb_type= h->sub_mb_type[i];
739             const int n= 4*i;
740             int x_offset= (i&1)<<2;
741             int y_offset= (i&2)<<1;
742
743             if(IS_SUB_8X8(sub_mb_type)){
744                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
745                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
746                     &weight_op[3], &weight_avg[3],
747                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
748                     pixel_shift, chroma444);
749             }else if(IS_SUB_8X4(sub_mb_type)){
750                 mc_part(h, n  , 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
751                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
752                     &weight_op[4], &weight_avg[4],
753                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
754                     pixel_shift, chroma444);
755                 mc_part(h, n+2, 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
756                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
757                     &weight_op[4], &weight_avg[4],
758                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
759                     pixel_shift, chroma444);
760             }else if(IS_SUB_4X8(sub_mb_type)){
761                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
762                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
763                     &weight_op[5], &weight_avg[5],
764                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
765                     pixel_shift, chroma444);
766                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
767                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
768                     &weight_op[5], &weight_avg[5],
769                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
770                     pixel_shift, chroma444);
771             }else{
772                 int j;
773                 assert(IS_SUB_4X4(sub_mb_type));
774                 for(j=0; j<4; j++){
775                     int sub_x_offset= x_offset + 2*(j&1);
776                     int sub_y_offset= y_offset +   (j&2);
777                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
778                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
779                         &weight_op[6], &weight_avg[6],
780                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
781                         pixel_shift, chroma444);
782                 }
783             }
784         }
785     }
786
787     prefetch_motion(h, 1, pixel_shift, chroma444);
788 }
789
790 #define hl_motion_fn(sh, bits) \
791 static av_always_inline void hl_motion_ ## bits(H264Context *h, \
792                                        uint8_t *dest_y, \
793                                        uint8_t *dest_cb, uint8_t *dest_cr, \
794                                        qpel_mc_func (*qpix_put)[16], \
795                                        h264_chroma_mc_func (*chroma_put), \
796                                        qpel_mc_func (*qpix_avg)[16], \
797                                        h264_chroma_mc_func (*chroma_avg), \
798                                        h264_weight_func *weight_op, \
799                                        h264_biweight_func *weight_avg, \
800                                        int chroma444) \
801 { \
802     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put, \
803               qpix_avg, chroma_avg, weight_op, weight_avg, sh, chroma444); \
804 }
805 hl_motion_fn(0, 8);
806 hl_motion_fn(1, 16);
807
808 static void free_tables(H264Context *h, int free_rbsp){
809     int i;
810     H264Context *hx;
811
812     av_freep(&h->intra4x4_pred_mode);
813     av_freep(&h->chroma_pred_mode_table);
814     av_freep(&h->cbp_table);
815     av_freep(&h->mvd_table[0]);
816     av_freep(&h->mvd_table[1]);
817     av_freep(&h->direct_table);
818     av_freep(&h->non_zero_count);
819     av_freep(&h->slice_table_base);
820     h->slice_table= NULL;
821     av_freep(&h->list_counts);
822
823     av_freep(&h->mb2b_xy);
824     av_freep(&h->mb2br_xy);
825
826     for(i = 0; i < MAX_THREADS; i++) {
827         hx = h->thread_context[i];
828         if(!hx) continue;
829         av_freep(&hx->top_borders[1]);
830         av_freep(&hx->top_borders[0]);
831         av_freep(&hx->s.obmc_scratchpad);
832         if (free_rbsp){
833             av_freep(&hx->rbsp_buffer[1]);
834             av_freep(&hx->rbsp_buffer[0]);
835             hx->rbsp_buffer_size[0] = 0;
836             hx->rbsp_buffer_size[1] = 0;
837         }
838         if (i) av_freep(&h->thread_context[i]);
839     }
840 }
841
842 static void init_dequant8_coeff_table(H264Context *h){
843     int i,j,q,x;
844     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
845
846     for(i=0; i<6; i++ ){
847         h->dequant8_coeff[i] = h->dequant8_buffer[i];
848         for(j=0; j<i; j++){
849             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
850                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
851                 break;
852             }
853         }
854         if(j<i)
855             continue;
856
857         for(q=0; q<max_qp+1; q++){
858             int shift = div6[q];
859             int idx = rem6[q];
860             for(x=0; x<64; x++)
861                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
862                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
863                     h->pps.scaling_matrix8[i][x]) << shift;
864         }
865     }
866 }
867
868 static void init_dequant4_coeff_table(H264Context *h){
869     int i,j,q,x;
870     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
871     for(i=0; i<6; i++ ){
872         h->dequant4_coeff[i] = h->dequant4_buffer[i];
873         for(j=0; j<i; j++){
874             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
875                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
876                 break;
877             }
878         }
879         if(j<i)
880             continue;
881
882         for(q=0; q<max_qp+1; q++){
883             int shift = div6[q] + 2;
884             int idx = rem6[q];
885             for(x=0; x<16; x++)
886                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
887                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
888                     h->pps.scaling_matrix4[i][x]) << shift;
889         }
890     }
891 }
892
893 static void init_dequant_tables(H264Context *h){
894     int i,x;
895     init_dequant4_coeff_table(h);
896     if(h->pps.transform_8x8_mode)
897         init_dequant8_coeff_table(h);
898     if(h->sps.transform_bypass){
899         for(i=0; i<6; i++)
900             for(x=0; x<16; x++)
901                 h->dequant4_coeff[i][0][x] = 1<<6;
902         if(h->pps.transform_8x8_mode)
903             for(i=0; i<6; i++)
904                 for(x=0; x<64; x++)
905                     h->dequant8_coeff[i][0][x] = 1<<6;
906     }
907 }
908
909
910 int ff_h264_alloc_tables(H264Context *h){
911     MpegEncContext * const s = &h->s;
912     const int big_mb_num= s->mb_stride * (s->mb_height+1);
913     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
914     int x,y;
915
916     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
917
918     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
919     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
920     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
921
922     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
923     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
924     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
925     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
926     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
927
928     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
929     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
930
931     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
932     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
933     for(y=0; y<s->mb_height; y++){
934         for(x=0; x<s->mb_width; x++){
935             const int mb_xy= x + y*s->mb_stride;
936             const int b_xy = 4*x + 4*y*h->b_stride;
937
938             h->mb2b_xy [mb_xy]= b_xy;
939             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
940         }
941     }
942
943     s->obmc_scratchpad = NULL;
944
945     if(!h->dequant4_coeff[0])
946         init_dequant_tables(h);
947
948     return 0;
949 fail:
950     free_tables(h, 1);
951     return -1;
952 }
953
954 /**
955  * Mimic alloc_tables(), but for every context thread.
956  */
957 static void clone_tables(H264Context *dst, H264Context *src, int i){
958     MpegEncContext * const s = &src->s;
959     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
960     dst->non_zero_count           = src->non_zero_count;
961     dst->slice_table              = src->slice_table;
962     dst->cbp_table                = src->cbp_table;
963     dst->mb2b_xy                  = src->mb2b_xy;
964     dst->mb2br_xy                 = src->mb2br_xy;
965     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
966     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
967     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
968     dst->direct_table             = src->direct_table;
969     dst->list_counts              = src->list_counts;
970
971     dst->s.obmc_scratchpad = NULL;
972     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma);
973 }
974
975 /**
976  * Init context
977  * Allocate buffers which are not shared amongst multiple threads.
978  */
979 static int context_init(H264Context *h){
980     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
981     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
982
983     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
984     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
985
986     return 0;
987 fail:
988     return -1; // free_tables will clean up for us
989 }
990
991 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
992
993 static av_cold void common_init(H264Context *h){
994     MpegEncContext * const s = &h->s;
995
996     s->width = s->avctx->width;
997     s->height = s->avctx->height;
998     s->codec_id= s->avctx->codec->id;
999
1000     ff_h264dsp_init(&h->h264dsp, 8);
1001     ff_h264_pred_init(&h->hpc, s->codec_id, 8);
1002
1003     h->dequant_coeff_pps= -1;
1004     s->unrestricted_mv=1;
1005     s->decode=1; //FIXME
1006
1007     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
1008
1009     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
1010     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
1011 }
1012
1013 int ff_h264_decode_extradata(H264Context *h)
1014 {
1015     AVCodecContext *avctx = h->s.avctx;
1016
1017     if(*(char *)avctx->extradata == 1){
1018         int i, cnt, nalsize;
1019         unsigned char *p = avctx->extradata;
1020
1021         h->is_avc = 1;
1022
1023         if(avctx->extradata_size < 7) {
1024             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1025             return -1;
1026         }
1027         /* sps and pps in the avcC always have length coded with 2 bytes,
1028            so put a fake nal_length_size = 2 while parsing them */
1029         h->nal_length_size = 2;
1030         // Decode sps from avcC
1031         cnt = *(p+5) & 0x1f; // Number of sps
1032         p += 6;
1033         for (i = 0; i < cnt; i++) {
1034             nalsize = AV_RB16(p) + 2;
1035             if(decode_nal_units(h, p, nalsize) < 0) {
1036                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1037                 return -1;
1038             }
1039             p += nalsize;
1040         }
1041         // Decode pps from avcC
1042         cnt = *(p++); // Number of pps
1043         for (i = 0; i < cnt; i++) {
1044             nalsize = AV_RB16(p) + 2;
1045             if (decode_nal_units(h, p, nalsize) < 0) {
1046                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1047                 return -1;
1048             }
1049             p += nalsize;
1050         }
1051         // Now store right nal length size, that will be use to parse all other nals
1052         h->nal_length_size = ((*(((char*)(avctx->extradata))+4))&0x03)+1;
1053     } else {
1054         h->is_avc = 0;
1055         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1056             return -1;
1057     }
1058     return 0;
1059 }
1060
1061 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1062     H264Context *h= avctx->priv_data;
1063     MpegEncContext * const s = &h->s;
1064
1065     MPV_decode_defaults(s);
1066
1067     s->avctx = avctx;
1068     common_init(h);
1069
1070     s->out_format = FMT_H264;
1071     s->workaround_bugs= avctx->workaround_bugs;
1072
1073     // set defaults
1074 //    s->decode_mb= ff_h263_decode_mb;
1075     s->quarter_sample = 1;
1076     if(!avctx->has_b_frames)
1077     s->low_delay= 1;
1078
1079     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1080
1081     ff_h264_decode_init_vlc();
1082
1083     h->pixel_shift = 0;
1084     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1085
1086     h->thread_context[0] = h;
1087     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1088     h->prev_poc_msb= 1<<16;
1089     h->x264_build = -1;
1090     ff_h264_reset_sei(h);
1091     if(avctx->codec_id == CODEC_ID_H264){
1092         if(avctx->ticks_per_frame == 1){
1093             s->avctx->time_base.den *=2;
1094         }
1095         avctx->ticks_per_frame = 2;
1096     }
1097
1098     if(avctx->extradata_size > 0 && avctx->extradata &&
1099         ff_h264_decode_extradata(h))
1100         return -1;
1101
1102     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1103         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1104         s->low_delay = 0;
1105     }
1106
1107     return 0;
1108 }
1109
1110 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1111 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1112 {
1113     int i;
1114
1115     for (i=0; i<count; i++){
1116         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1117                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1118                 !from[i]));
1119         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1120     }
1121 }
1122
1123 static void copy_parameter_set(void **to, void **from, int count, int size)
1124 {
1125     int i;
1126
1127     for (i=0; i<count; i++){
1128         if (to[i] && !from[i]) av_freep(&to[i]);
1129         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1130
1131         if (from[i]) memcpy(to[i], from[i], size);
1132     }
1133 }
1134
1135 static int decode_init_thread_copy(AVCodecContext *avctx){
1136     H264Context *h= avctx->priv_data;
1137
1138     if (!avctx->is_copy) return 0;
1139     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1140     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1141
1142     return 0;
1143 }
1144
1145 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1146 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1147     H264Context *h= dst->priv_data, *h1= src->priv_data;
1148     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1149     int inited = s->context_initialized, err;
1150     int i;
1151
1152     if(dst == src || !s1->context_initialized) return 0;
1153
1154     err = ff_mpeg_update_thread_context(dst, src);
1155     if(err) return err;
1156
1157     //FIXME handle width/height changing
1158     if(!inited){
1159         for(i = 0; i < MAX_SPS_COUNT; i++)
1160             av_freep(h->sps_buffers + i);
1161
1162         for(i = 0; i < MAX_PPS_COUNT; i++)
1163             av_freep(h->pps_buffers + i);
1164
1165         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1166         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1167         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1168         ff_h264_alloc_tables(h);
1169         context_init(h);
1170
1171         for(i=0; i<2; i++){
1172             h->rbsp_buffer[i] = NULL;
1173             h->rbsp_buffer_size[i] = 0;
1174         }
1175
1176         h->thread_context[0] = h;
1177
1178         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1179         // so this has to be allocated here
1180         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1181
1182         s->dsp.clear_blocks(h->mb);
1183         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1184     }
1185
1186     //extradata/NAL handling
1187     h->is_avc          = h1->is_avc;
1188
1189     //SPS/PPS
1190     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1191     h->sps             = h1->sps;
1192     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1193     h->pps             = h1->pps;
1194
1195     //Dequantization matrices
1196     //FIXME these are big - can they be only copied when PPS changes?
1197     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1198
1199     for(i=0; i<6; i++)
1200         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1201
1202     for(i=0; i<6; i++)
1203         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1204
1205     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1206
1207     //POC timing
1208     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1209
1210     //reference lists
1211     copy_fields(h, h1, ref_count, list_count);
1212     copy_fields(h, h1, ref_list,  intra_gb);
1213     copy_fields(h, h1, short_ref, cabac_init_idc);
1214
1215     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1216     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1217     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1218
1219     h->last_slice_type = h1->last_slice_type;
1220
1221     if(!s->current_picture_ptr) return 0;
1222
1223     if(!s->dropable) {
1224         ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1225         h->prev_poc_msb     = h->poc_msb;
1226         h->prev_poc_lsb     = h->poc_lsb;
1227     }
1228     h->prev_frame_num_offset= h->frame_num_offset;
1229     h->prev_frame_num       = h->frame_num;
1230     h->outputed_poc         = h->next_outputed_poc;
1231
1232     return 0;
1233 }
1234
1235 int ff_h264_frame_start(H264Context *h){
1236     MpegEncContext * const s = &h->s;
1237     int i;
1238     const int pixel_shift = h->pixel_shift;
1239     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1240
1241     if(MPV_frame_start(s, s->avctx) < 0)
1242         return -1;
1243     ff_er_frame_start(s);
1244     /*
1245      * MPV_frame_start uses pict_type to derive key_frame.
1246      * This is incorrect for H.264; IDR markings must be used.
1247      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1248      * See decode_nal_units().
1249      */
1250     s->current_picture_ptr->key_frame= 0;
1251     s->current_picture_ptr->mmco_reset= 0;
1252
1253     assert(s->linesize && s->uvlinesize);
1254
1255     for(i=0; i<16; i++){
1256         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1257         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1258     }
1259     for(i=0; i<16; i++){
1260         h->block_offset[16+i]=
1261         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1262         h->block_offset[48+16+i]=
1263         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1264     }
1265
1266     /* can't be in alloc_tables because linesize isn't known there.
1267      * FIXME: redo bipred weight to not require extra buffer? */
1268     for(i = 0; i < thread_count; i++)
1269         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1270             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1271
1272     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1273     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1274
1275 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
1276
1277     // We mark the current picture as non-reference after allocating it, so
1278     // that if we break out due to an error it can be released automatically
1279     // in the next MPV_frame_start().
1280     // SVQ3 as well as most other codecs have only last/next/current and thus
1281     // get released even with set reference, besides SVQ3 and others do not
1282     // mark frames as reference later "naturally".
1283     if(s->codec_id != CODEC_ID_SVQ3)
1284         s->current_picture_ptr->reference= 0;
1285
1286     s->current_picture_ptr->field_poc[0]=
1287     s->current_picture_ptr->field_poc[1]= INT_MAX;
1288
1289     h->next_output_pic = NULL;
1290
1291     assert(s->current_picture_ptr->long_ref==0);
1292
1293     return 0;
1294 }
1295
1296 /**
1297   * Run setup operations that must be run after slice header decoding.
1298   * This includes finding the next displayed frame.
1299   *
1300   * @param h h264 master context
1301   * @param setup_finished enough NALs have been read that we can call
1302   * ff_thread_finish_setup()
1303   */
1304 static void decode_postinit(H264Context *h, int setup_finished){
1305     MpegEncContext * const s = &h->s;
1306     Picture *out = s->current_picture_ptr;
1307     Picture *cur = s->current_picture_ptr;
1308     int i, pics, out_of_order, out_idx;
1309
1310     s->current_picture_ptr->qscale_type= FF_QSCALE_TYPE_H264;
1311     s->current_picture_ptr->pict_type= s->pict_type;
1312
1313     if (h->next_output_pic) return;
1314
1315     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1316         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1317         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1318         //to find this yet, so we assume the worst for now.
1319         //if (setup_finished)
1320         //    ff_thread_finish_setup(s->avctx);
1321         return;
1322     }
1323
1324     cur->interlaced_frame = 0;
1325     cur->repeat_pict = 0;
1326
1327     /* Signal interlacing information externally. */
1328     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1329
1330     if(h->sps.pic_struct_present_flag){
1331         switch (h->sei_pic_struct)
1332         {
1333         case SEI_PIC_STRUCT_FRAME:
1334             break;
1335         case SEI_PIC_STRUCT_TOP_FIELD:
1336         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1337             cur->interlaced_frame = 1;
1338             break;
1339         case SEI_PIC_STRUCT_TOP_BOTTOM:
1340         case SEI_PIC_STRUCT_BOTTOM_TOP:
1341             if (FIELD_OR_MBAFF_PICTURE)
1342                 cur->interlaced_frame = 1;
1343             else
1344                 // try to flag soft telecine progressive
1345                 cur->interlaced_frame = h->prev_interlaced_frame;
1346             break;
1347         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1348         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1349             // Signal the possibility of telecined film externally (pic_struct 5,6)
1350             // From these hints, let the applications decide if they apply deinterlacing.
1351             cur->repeat_pict = 1;
1352             break;
1353         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1354             // Force progressive here, as doubling interlaced frame is a bad idea.
1355             cur->repeat_pict = 2;
1356             break;
1357         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1358             cur->repeat_pict = 4;
1359             break;
1360         }
1361
1362         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1363             cur->interlaced_frame = (h->sei_ct_type & (1<<1)) != 0;
1364     }else{
1365         /* Derive interlacing flag from used decoding process. */
1366         cur->interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1367     }
1368     h->prev_interlaced_frame = cur->interlaced_frame;
1369
1370     if (cur->field_poc[0] != cur->field_poc[1]){
1371         /* Derive top_field_first from field pocs. */
1372         cur->top_field_first = cur->field_poc[0] < cur->field_poc[1];
1373     }else{
1374         if(cur->interlaced_frame || h->sps.pic_struct_present_flag){
1375             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1376             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1377               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1378                 cur->top_field_first = 1;
1379             else
1380                 cur->top_field_first = 0;
1381         }else{
1382             /* Most likely progressive */
1383             cur->top_field_first = 0;
1384         }
1385     }
1386
1387     //FIXME do something with unavailable reference frames
1388
1389     /* Sort B-frames into display order */
1390
1391     if(h->sps.bitstream_restriction_flag
1392        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1393         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1394         s->low_delay = 0;
1395     }
1396
1397     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1398        && !h->sps.bitstream_restriction_flag){
1399         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1400         s->low_delay= 0;
1401     }
1402
1403     pics = 0;
1404     while(h->delayed_pic[pics]) pics++;
1405
1406     assert(pics <= MAX_DELAYED_PIC_COUNT);
1407
1408     h->delayed_pic[pics++] = cur;
1409     if(cur->reference == 0)
1410         cur->reference = DELAYED_PIC_REF;
1411
1412     out = h->delayed_pic[0];
1413     out_idx = 0;
1414     for(i=1; h->delayed_pic[i] && !h->delayed_pic[i]->key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1415         if(h->delayed_pic[i]->poc < out->poc){
1416             out = h->delayed_pic[i];
1417             out_idx = i;
1418         }
1419     if(s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset))
1420         h->next_outputed_poc= INT_MIN;
1421     out_of_order = out->poc < h->next_outputed_poc;
1422
1423     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1424         { }
1425     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1426        || (s->low_delay &&
1427         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1428          || cur->pict_type == AV_PICTURE_TYPE_B)))
1429     {
1430         s->low_delay = 0;
1431         s->avctx->has_b_frames++;
1432     }
1433
1434     if(out_of_order || pics > s->avctx->has_b_frames){
1435         out->reference &= ~DELAYED_PIC_REF;
1436         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1437                          // or else the first thread can release the picture and reuse it unsafely
1438         for(i=out_idx; h->delayed_pic[i]; i++)
1439             h->delayed_pic[i] = h->delayed_pic[i+1];
1440     }
1441     if(!out_of_order && pics > s->avctx->has_b_frames){
1442         h->next_output_pic = out;
1443         if(out_idx==0 && h->delayed_pic[0] && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset)) {
1444             h->next_outputed_poc = INT_MIN;
1445         } else
1446             h->next_outputed_poc = out->poc;
1447     }else{
1448         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1449     }
1450
1451     if (setup_finished)
1452         ff_thread_finish_setup(s->avctx);
1453 }
1454
1455 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){
1456     MpegEncContext * const s = &h->s;
1457     uint8_t *top_border;
1458     int top_idx = 1;
1459     const int pixel_shift = h->pixel_shift;
1460
1461     src_y  -=   linesize;
1462     src_cb -= uvlinesize;
1463     src_cr -= uvlinesize;
1464
1465     if(!simple && FRAME_MBAFF){
1466         if(s->mb_y&1){
1467             if(!MB_MBAFF){
1468                 top_border = h->top_borders[0][s->mb_x];
1469                 AV_COPY128(top_border, src_y + 15*linesize);
1470                 if (pixel_shift)
1471                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1472                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1473                     if(chroma444){
1474                         if (pixel_shift){
1475                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1476                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1477                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1478                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1479                         } else {
1480                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1481                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1482                         }
1483                     } else {
1484                         if (pixel_shift) {
1485                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1486                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1487                         } else {
1488                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1489                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1490                         }
1491                     }
1492                 }
1493             }
1494         }else if(MB_MBAFF){
1495             top_idx = 0;
1496         }else
1497             return;
1498     }
1499
1500     top_border = h->top_borders[top_idx][s->mb_x];
1501     // There are two lines saved, the line above the the top macroblock of a pair,
1502     // and the line above the bottom macroblock
1503     AV_COPY128(top_border, src_y + 16*linesize);
1504     if (pixel_shift)
1505         AV_COPY128(top_border+16, src_y+16*linesize+16);
1506
1507     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1508         if(chroma444){
1509             if (pixel_shift){
1510                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1511                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1512                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1513                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1514             } else {
1515                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1516                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1517             }
1518         } else {
1519             if (pixel_shift) {
1520                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1521                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1522             } else {
1523                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1524                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1525             }
1526         }
1527     }
1528 }
1529
1530 static inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1531                                   uint8_t *src_cb, uint8_t *src_cr,
1532                                   int linesize, int uvlinesize,
1533                                   int xchg, int chroma444,
1534                                   int simple, int pixel_shift){
1535     MpegEncContext * const s = &h->s;
1536     int deblock_topleft;
1537     int deblock_top;
1538     int top_idx = 1;
1539     uint8_t *top_border_m1;
1540     uint8_t *top_border;
1541
1542     if(!simple && FRAME_MBAFF){
1543         if(s->mb_y&1){
1544             if(!MB_MBAFF)
1545                 return;
1546         }else{
1547             top_idx = MB_MBAFF ? 0 : 1;
1548         }
1549     }
1550
1551     if(h->deblocking_filter == 2) {
1552         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1553         deblock_top     = h->top_type;
1554     } else {
1555         deblock_topleft = (s->mb_x > 0);
1556         deblock_top     = (s->mb_y > !!MB_FIELD);
1557     }
1558
1559     src_y  -=   linesize + 1 + pixel_shift;
1560     src_cb -= uvlinesize + 1 + pixel_shift;
1561     src_cr -= uvlinesize + 1 + pixel_shift;
1562
1563     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1564     top_border    = h->top_borders[top_idx][s->mb_x];
1565
1566 #define XCHG(a,b,xchg)\
1567     if (pixel_shift) {\
1568         if (xchg) {\
1569             AV_SWAP64(b+0,a+0);\
1570             AV_SWAP64(b+8,a+8);\
1571         } else {\
1572             AV_COPY128(b,a); \
1573         }\
1574     } else \
1575 if (xchg) AV_SWAP64(b,a);\
1576 else      AV_COPY64(b,a);
1577
1578     if(deblock_top){
1579         if(deblock_topleft){
1580             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1581         }
1582         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1583         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1584         if(s->mb_x+1 < s->mb_width){
1585             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1586         }
1587     }
1588     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1589         if(chroma444){
1590             if(deblock_topleft){
1591                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1592                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1593             }
1594             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1595             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1596             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1597             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1598             if(s->mb_x+1 < s->mb_width){
1599                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1600                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1601             }
1602         } else {
1603             if(deblock_top){
1604                 if(deblock_topleft){
1605                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1606                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1607                 }
1608                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1609                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1610             }
1611         }
1612     }
1613 }
1614
1615 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1616     if (high_bit_depth) {
1617         return AV_RN32A(((int32_t*)mb) + index);
1618     } else
1619         return AV_RN16A(mb + index);
1620 }
1621
1622 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1623     if (high_bit_depth) {
1624         AV_WN32A(((int32_t*)mb) + index, value);
1625     } else
1626         AV_WN16A(mb + index, value);
1627 }
1628
1629 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1630                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1631 {
1632     MpegEncContext * const s = &h->s;
1633     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1634     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1635     int i;
1636     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1637     block_offset += 16*p;
1638     if(IS_INTRA4x4(mb_type)){
1639         if(simple || !s->encoding){
1640             if(IS_8x8DCT(mb_type)){
1641                 if(transform_bypass){
1642                     idct_dc_add =
1643                     idct_add    = s->dsp.add_pixels8;
1644                 }else{
1645                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1646                     idct_add    = h->h264dsp.h264_idct8_add;
1647                 }
1648                 for(i=0; i<16; i+=4){
1649                     uint8_t * const ptr= dest_y + block_offset[i];
1650                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1651                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1652                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1653                     }else{
1654                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1655                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1656                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1657                         if(nnz){
1658                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1659                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1660                             else
1661                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1662                         }
1663                     }
1664                 }
1665             }else{
1666                 if(transform_bypass){
1667                     idct_dc_add =
1668                     idct_add    = s->dsp.add_pixels4;
1669                 }else{
1670                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1671                     idct_add    = h->h264dsp.h264_idct_add;
1672                 }
1673                 for(i=0; i<16; i++){
1674                     uint8_t * const ptr= dest_y + block_offset[i];
1675                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1676
1677                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1678                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1679                     }else{
1680                         uint8_t *topright;
1681                         int nnz, tr;
1682                         uint64_t tr_high;
1683                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1684                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1685                             assert(mb_y || linesize <= block_offset[i]);
1686                             if(!topright_avail){
1687                                 if (pixel_shift) {
1688                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1689                                     topright= (uint8_t*) &tr_high;
1690                                 } else {
1691                                     tr= ptr[3 - linesize]*0x01010101;
1692                                     topright= (uint8_t*) &tr;
1693                                 }
1694                             }else
1695                                 topright= ptr + (4 << pixel_shift) - linesize;
1696                         }else
1697                             topright= NULL;
1698
1699                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1700                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1701                         if(nnz){
1702                             if(is_h264){
1703                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1704                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1705                                 else
1706                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1707                             }else
1708                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1709                         }
1710                     }
1711                 }
1712             }
1713         }
1714     }else{
1715         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1716         if(is_h264){
1717             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1718                 if(!transform_bypass)
1719                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1720                 else{
1721                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1722                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1723                     for(i = 0; i < 16; i++)
1724                         dctcoef_set(h->mb+p*256, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1725                 }
1726             }
1727         }else
1728             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1729     }
1730 }
1731
1732 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1733                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1734 {
1735     MpegEncContext * const s = &h->s;
1736     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1737     int i;
1738     block_offset += 16*p;
1739     if(!IS_INTRA4x4(mb_type)){
1740         if(is_h264){
1741             if(IS_INTRA16x16(mb_type)){
1742                 if(transform_bypass){
1743                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1744                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1745                     }else{
1746                         for(i=0; i<16; i++){
1747                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1748                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1749                         }
1750                     }
1751                 }else{
1752                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1753                 }
1754             }else if(h->cbp&15){
1755                 if(transform_bypass){
1756                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1757                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1758                     for(i=0; i<16; i+=di){
1759                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1760                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1761                         }
1762                     }
1763                 }else{
1764                     if(IS_8x8DCT(mb_type)){
1765                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1766                     }else{
1767                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1768                     }
1769                 }
1770             }
1771         }else{
1772             for(i=0; i<16; i++){
1773                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1774                     uint8_t * const ptr= dest_y + block_offset[i];
1775                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1776                 }
1777             }
1778         }
1779     }
1780 }
1781
1782 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1783     MpegEncContext * const s = &h->s;
1784     const int mb_x= s->mb_x;
1785     const int mb_y= s->mb_y;
1786     const int mb_xy= h->mb_xy;
1787     const int mb_type= s->current_picture.mb_type[mb_xy];
1788     uint8_t  *dest_y, *dest_cb, *dest_cr;
1789     int linesize, uvlinesize /*dct_offset*/;
1790     int i, j;
1791     int *block_offset = &h->block_offset[0];
1792     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1793     /* is_h264 should always be true if SVQ3 is disabled. */
1794     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1795     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1796
1797     dest_y  = s->current_picture.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1798     dest_cb = s->current_picture.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
1799     dest_cr = s->current_picture.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
1800
1801     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1802     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1803
1804     h->list_counts[mb_xy]= h->list_count;
1805
1806     if (!simple && MB_FIELD) {
1807         linesize   = h->mb_linesize   = s->linesize * 2;
1808         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1809         block_offset = &h->block_offset[48];
1810         if(mb_y&1){ //FIXME move out of this function?
1811             dest_y -= s->linesize*15;
1812             dest_cb-= s->uvlinesize*7;
1813             dest_cr-= s->uvlinesize*7;
1814         }
1815         if(FRAME_MBAFF) {
1816             int list;
1817             for(list=0; list<h->list_count; list++){
1818                 if(!USES_LIST(mb_type, list))
1819                     continue;
1820                 if(IS_16X16(mb_type)){
1821                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1822                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1823                 }else{
1824                     for(i=0; i<16; i+=4){
1825                         int ref = h->ref_cache[list][scan8[i]];
1826                         if(ref >= 0)
1827                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1828                     }
1829                 }
1830             }
1831         }
1832     } else {
1833         linesize   = h->mb_linesize   = s->linesize;
1834         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1835 //        dct_offset = s->linesize * 16;
1836     }
1837
1838     if (!simple && IS_INTRA_PCM(mb_type)) {
1839         if (pixel_shift) {
1840             const int bit_depth = h->sps.bit_depth_luma;
1841             int j;
1842             GetBitContext gb;
1843             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1844
1845             for (i = 0; i < 16; i++) {
1846                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1847                 for (j = 0; j < 16; j++)
1848                     tmp_y[j] = get_bits(&gb, bit_depth);
1849             }
1850             for (i = 0; i < 8; i++) {
1851                 uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1852                 for (j = 0; j < 8; j++)
1853                     tmp_cb[j] = get_bits(&gb, bit_depth);
1854             }
1855             for (i = 0; i < 8; i++) {
1856                 uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1857                 for (j = 0; j < 8; j++)
1858                     tmp_cr[j] = get_bits(&gb, bit_depth);
1859             }
1860         } else {
1861         for (i=0; i<16; i++) {
1862             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1863         }
1864         for (i=0; i<8; i++) {
1865             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
1866             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
1867         }
1868         }
1869     } else {
1870         if(IS_INTRA(mb_type)){
1871             if(h->deblocking_filter)
1872                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1873
1874             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1875                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1876                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1877             }
1878
1879             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1880
1881             if(h->deblocking_filter)
1882                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1883         }else if(is_h264){
1884             if (pixel_shift) {
1885                 hl_motion_16(h, dest_y, dest_cb, dest_cr,
1886                              s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1887                              s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1888                              h->h264dsp.weight_h264_pixels_tab,
1889                              h->h264dsp.biweight_h264_pixels_tab, 0);
1890             } else
1891                 hl_motion_8(h, dest_y, dest_cb, dest_cr,
1892                             s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1893                             s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1894                             h->h264dsp.weight_h264_pixels_tab,
1895                             h->h264dsp.biweight_h264_pixels_tab, 0);
1896         }
1897
1898         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1899
1900         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1901             uint8_t *dest[2] = {dest_cb, dest_cr};
1902             if(transform_bypass){
1903                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1904                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
1905                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
1906                 }else{
1907                     idct_add = s->dsp.add_pixels4;
1908                     for(j=1; j<3; j++){
1909                         for(i=j*16; i<j*16+4; i++){
1910                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1911                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1912                         }
1913                     }
1914                 }
1915             }else{
1916                 if(is_h264){
1917                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1918                         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]);
1919                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1920                         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]);
1921                     h->h264dsp.h264_idct_add8(dest, block_offset,
1922                                               h->mb, uvlinesize,
1923                                               h->non_zero_count_cache);
1924                 }else{
1925                     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]);
1926                     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]);
1927                     for(j=1; j<3; j++){
1928                         for(i=j*16; i<j*16+4; i++){
1929                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1930                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
1931                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1932                             }
1933                         }
1934                     }
1935                 }
1936             }
1937         }
1938     }
1939     if(h->cbp || IS_INTRA(mb_type))
1940     {
1941         s->dsp.clear_blocks(h->mb);
1942         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
1943     }
1944 }
1945
1946 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
1947     MpegEncContext * const s = &h->s;
1948     const int mb_x= s->mb_x;
1949     const int mb_y= s->mb_y;
1950     const int mb_xy= h->mb_xy;
1951     const int mb_type= s->current_picture.mb_type[mb_xy];
1952     uint8_t  *dest[3];
1953     int linesize;
1954     int i, j, p;
1955     int *block_offset = &h->block_offset[0];
1956     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1957
1958     for (p = 0; p < 3; p++)
1959     {
1960         dest[p] = s->current_picture.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
1961         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1962     }
1963
1964     h->list_counts[mb_xy]= h->list_count;
1965
1966     if (!simple && MB_FIELD) {
1967         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
1968         block_offset = &h->block_offset[48];
1969         if(mb_y&1) //FIXME move out of this function?
1970             for (p = 0; p < 3; p++)
1971                 dest[p] -= s->linesize*15;
1972         if(FRAME_MBAFF) {
1973             int list;
1974             for(list=0; list<h->list_count; list++){
1975                 if(!USES_LIST(mb_type, list))
1976                     continue;
1977                 if(IS_16X16(mb_type)){
1978                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1979                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1980                 }else{
1981                     for(i=0; i<16; i+=4){
1982                         int ref = h->ref_cache[list][scan8[i]];
1983                         if(ref >= 0)
1984                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1985                     }
1986                 }
1987             }
1988         }
1989     } else {
1990         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
1991     }
1992
1993     if (!simple && IS_INTRA_PCM(mb_type)) {
1994         if (pixel_shift) {
1995             const int bit_depth = h->sps.bit_depth_luma;
1996             GetBitContext gb;
1997             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
1998
1999             for (p = 0; p < 3; p++) {
2000                 for (i = 0; i < 16; i++) {
2001                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2002                     for (j = 0; j < 16; j++)
2003                         tmp[j] = get_bits(&gb, bit_depth);
2004                 }
2005             }
2006         } else {
2007             for (p = 0; p < 3; p++) {
2008                 for (i = 0; i < 16; i++) {
2009                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2010                 }
2011             }
2012         }
2013     } else {
2014         if(IS_INTRA(mb_type)){
2015             if(h->deblocking_filter)
2016                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2017
2018             for (p = 0; p < 3; p++)
2019                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2020
2021             if(h->deblocking_filter)
2022                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2023         }else{
2024             if (pixel_shift) {
2025                 hl_motion_16(h, dest[0], dest[1], dest[2],
2026                              s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2027                              s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2028                              h->h264dsp.weight_h264_pixels_tab,
2029                              h->h264dsp.biweight_h264_pixels_tab, 1);
2030             } else
2031                 hl_motion_8(h, dest[0], dest[1], dest[2],
2032                             s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2033                             s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2034                             h->h264dsp.weight_h264_pixels_tab,
2035                             h->h264dsp.biweight_h264_pixels_tab, 1);
2036         }
2037
2038         for (p = 0; p < 3; p++)
2039             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2040     }
2041     if(h->cbp || IS_INTRA(mb_type))
2042     {
2043         s->dsp.clear_blocks(h->mb);
2044         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2045     }
2046 }
2047
2048 /**
2049  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2050  */
2051 #define hl_decode_mb_simple(sh, bits) \
2052 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2053     hl_decode_mb_internal(h, 1, sh); \
2054 }
2055 hl_decode_mb_simple(0, 8);
2056 hl_decode_mb_simple(1, 16);
2057
2058 /**
2059  * Process a macroblock; this handles edge cases, such as interlacing.
2060  */
2061 static void av_noinline hl_decode_mb_complex(H264Context *h){
2062     hl_decode_mb_internal(h, 0, h->pixel_shift);
2063 }
2064
2065 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2066     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2067 }
2068
2069 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2070     hl_decode_mb_444_internal(h, 1, 0);
2071 }
2072
2073 void ff_h264_hl_decode_mb(H264Context *h){
2074     MpegEncContext * const s = &h->s;
2075     const int mb_xy= h->mb_xy;
2076     const int mb_type= s->current_picture.mb_type[mb_xy];
2077     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2078
2079     if (CHROMA444) {
2080         if(is_complex || h->pixel_shift)
2081             hl_decode_mb_444_complex(h);
2082         else
2083             hl_decode_mb_444_simple(h);
2084     } else if (is_complex) {
2085         hl_decode_mb_complex(h);
2086     } else if (h->pixel_shift) {
2087         hl_decode_mb_simple_16(h);
2088     } else
2089         hl_decode_mb_simple_8(h);
2090 }
2091
2092 static int pred_weight_table(H264Context *h){
2093     MpegEncContext * const s = &h->s;
2094     int list, i;
2095     int luma_def, chroma_def;
2096
2097     h->use_weight= 0;
2098     h->use_weight_chroma= 0;
2099     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2100     if(h->sps.chroma_format_idc)
2101         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2102     luma_def = 1<<h->luma_log2_weight_denom;
2103     chroma_def = 1<<h->chroma_log2_weight_denom;
2104
2105     for(list=0; list<2; list++){
2106         h->luma_weight_flag[list]   = 0;
2107         h->chroma_weight_flag[list] = 0;
2108         for(i=0; i<h->ref_count[list]; i++){
2109             int luma_weight_flag, chroma_weight_flag;
2110
2111             luma_weight_flag= get_bits1(&s->gb);
2112             if(luma_weight_flag){
2113                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2114                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2115                 if(   h->luma_weight[i][list][0] != luma_def
2116                    || h->luma_weight[i][list][1] != 0) {
2117                     h->use_weight= 1;
2118                     h->luma_weight_flag[list]= 1;
2119                 }
2120             }else{
2121                 h->luma_weight[i][list][0]= luma_def;
2122                 h->luma_weight[i][list][1]= 0;
2123             }
2124
2125             if(h->sps.chroma_format_idc){
2126                 chroma_weight_flag= get_bits1(&s->gb);
2127                 if(chroma_weight_flag){
2128                     int j;
2129                     for(j=0; j<2; j++){
2130                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2131                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2132                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2133                            || h->chroma_weight[i][list][j][1] != 0) {
2134                             h->use_weight_chroma= 1;
2135                             h->chroma_weight_flag[list]= 1;
2136                         }
2137                     }
2138                 }else{
2139                     int j;
2140                     for(j=0; j<2; j++){
2141                         h->chroma_weight[i][list][j][0]= chroma_def;
2142                         h->chroma_weight[i][list][j][1]= 0;
2143                     }
2144                 }
2145             }
2146         }
2147         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2148     }
2149     h->use_weight= h->use_weight || h->use_weight_chroma;
2150     return 0;
2151 }
2152
2153 /**
2154  * Initialize implicit_weight table.
2155  * @param field  0/1 initialize the weight for interlaced MBAFF
2156  *                -1 initializes the rest
2157  */
2158 static void implicit_weight_table(H264Context *h, int field){
2159     MpegEncContext * const s = &h->s;
2160     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2161
2162     for (i = 0; i < 2; i++) {
2163         h->luma_weight_flag[i]   = 0;
2164         h->chroma_weight_flag[i] = 0;
2165     }
2166
2167     if(field < 0){
2168         cur_poc = s->current_picture_ptr->poc;
2169     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2170        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2171         h->use_weight= 0;
2172         h->use_weight_chroma= 0;
2173         return;
2174     }
2175         ref_start= 0;
2176         ref_count0= h->ref_count[0];
2177         ref_count1= h->ref_count[1];
2178     }else{
2179         cur_poc = s->current_picture_ptr->field_poc[field];
2180         ref_start= 16;
2181         ref_count0= 16+2*h->ref_count[0];
2182         ref_count1= 16+2*h->ref_count[1];
2183     }
2184
2185     h->use_weight= 2;
2186     h->use_weight_chroma= 2;
2187     h->luma_log2_weight_denom= 5;
2188     h->chroma_log2_weight_denom= 5;
2189
2190     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2191         int poc0 = h->ref_list[0][ref0].poc;
2192         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2193             int poc1 = h->ref_list[1][ref1].poc;
2194             int td = av_clip(poc1 - poc0, -128, 127);
2195             int w= 32;
2196             if(td){
2197                 int tb = av_clip(cur_poc - poc0, -128, 127);
2198                 int tx = (16384 + (FFABS(td) >> 1)) / td;
2199                 int dist_scale_factor = (tb*tx + 32) >> 8;
2200                 if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2201                     w = 64 - dist_scale_factor;
2202             }
2203             if(field<0){
2204                 h->implicit_weight[ref0][ref1][0]=
2205                 h->implicit_weight[ref0][ref1][1]= w;
2206             }else{
2207                 h->implicit_weight[ref0][ref1][field]=w;
2208             }
2209         }
2210     }
2211 }
2212
2213 /**
2214  * instantaneous decoder refresh.
2215  */
2216 static void idr(H264Context *h){
2217     ff_h264_remove_all_refs(h);
2218     h->prev_frame_num= 0;
2219     h->prev_frame_num_offset= 0;
2220     h->prev_poc_msb=
2221     h->prev_poc_lsb= 0;
2222 }
2223
2224 /* forget old pics after a seek */
2225 static void flush_dpb(AVCodecContext *avctx){
2226     H264Context *h= avctx->priv_data;
2227     int i;
2228     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
2229         if(h->delayed_pic[i])
2230             h->delayed_pic[i]->reference= 0;
2231         h->delayed_pic[i]= NULL;
2232     }
2233     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2234     h->prev_interlaced_frame = 1;
2235     idr(h);
2236     if(h->s.current_picture_ptr)
2237         h->s.current_picture_ptr->reference= 0;
2238     h->s.first_field= 0;
2239     ff_h264_reset_sei(h);
2240     ff_mpeg_flush(avctx);
2241 }
2242
2243 static int init_poc(H264Context *h){
2244     MpegEncContext * const s = &h->s;
2245     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2246     int field_poc[2];
2247     Picture *cur = s->current_picture_ptr;
2248
2249     h->frame_num_offset= h->prev_frame_num_offset;
2250     if(h->frame_num < h->prev_frame_num)
2251         h->frame_num_offset += max_frame_num;
2252
2253     if(h->sps.poc_type==0){
2254         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2255
2256         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2257             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2258         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2259             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2260         else
2261             h->poc_msb = h->prev_poc_msb;
2262 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2263         field_poc[0] =
2264         field_poc[1] = h->poc_msb + h->poc_lsb;
2265         if(s->picture_structure == PICT_FRAME)
2266             field_poc[1] += h->delta_poc_bottom;
2267     }else if(h->sps.poc_type==1){
2268         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2269         int i;
2270
2271         if(h->sps.poc_cycle_length != 0)
2272             abs_frame_num = h->frame_num_offset + h->frame_num;
2273         else
2274             abs_frame_num = 0;
2275
2276         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2277             abs_frame_num--;
2278
2279         expected_delta_per_poc_cycle = 0;
2280         for(i=0; i < h->sps.poc_cycle_length; i++)
2281             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2282
2283         if(abs_frame_num > 0){
2284             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2285             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2286
2287             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2288             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2289                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2290         } else
2291             expectedpoc = 0;
2292
2293         if(h->nal_ref_idc == 0)
2294             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2295
2296         field_poc[0] = expectedpoc + h->delta_poc[0];
2297         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2298
2299         if(s->picture_structure == PICT_FRAME)
2300             field_poc[1] += h->delta_poc[1];
2301     }else{
2302         int poc= 2*(h->frame_num_offset + h->frame_num);
2303
2304         if(!h->nal_ref_idc)
2305             poc--;
2306
2307         field_poc[0]= poc;
2308         field_poc[1]= poc;
2309     }
2310
2311     if(s->picture_structure != PICT_BOTTOM_FIELD)
2312         s->current_picture_ptr->field_poc[0]= field_poc[0];
2313     if(s->picture_structure != PICT_TOP_FIELD)
2314         s->current_picture_ptr->field_poc[1]= field_poc[1];
2315     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2316
2317     return 0;
2318 }
2319
2320
2321 /**
2322  * initialize scan tables
2323  */
2324 static void init_scan_tables(H264Context *h){
2325     int i;
2326     for(i=0; i<16; i++){
2327 #define T(x) (x>>2) | ((x<<2) & 0xF)
2328         h->zigzag_scan[i] = T(zigzag_scan[i]);
2329         h-> field_scan[i] = T( field_scan[i]);
2330 #undef T
2331     }
2332     for(i=0; i<64; i++){
2333 #define T(x) (x>>3) | ((x&7)<<3)
2334         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2335         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2336         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2337         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2338 #undef T
2339     }
2340     if(h->sps.transform_bypass){ //FIXME same ugly
2341         h->zigzag_scan_q0          = zigzag_scan;
2342         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2343         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2344         h->field_scan_q0           = field_scan;
2345         h->field_scan8x8_q0        = field_scan8x8;
2346         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2347     }else{
2348         h->zigzag_scan_q0          = h->zigzag_scan;
2349         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2350         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2351         h->field_scan_q0           = h->field_scan;
2352         h->field_scan8x8_q0        = h->field_scan8x8;
2353         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2354     }
2355 }
2356
2357 static void field_end(H264Context *h, int in_setup){
2358     MpegEncContext * const s = &h->s;
2359     AVCodecContext * const avctx= s->avctx;
2360     s->mb_y= 0;
2361
2362     if (!in_setup && !s->dropable)
2363         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2364                                  s->picture_structure==PICT_BOTTOM_FIELD);
2365
2366     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2367         ff_vdpau_h264_set_reference_frames(s);
2368
2369     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2370         if(!s->dropable) {
2371             ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2372             h->prev_poc_msb= h->poc_msb;
2373             h->prev_poc_lsb= h->poc_lsb;
2374         }
2375         h->prev_frame_num_offset= h->frame_num_offset;
2376         h->prev_frame_num= h->frame_num;
2377         h->outputed_poc = h->next_outputed_poc;
2378     }
2379
2380     if (avctx->hwaccel) {
2381         if (avctx->hwaccel->end_frame(avctx) < 0)
2382             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2383     }
2384
2385     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2386         ff_vdpau_h264_picture_complete(s);
2387
2388     /*
2389      * FIXME: Error handling code does not seem to support interlaced
2390      * when slices span multiple rows
2391      * The ff_er_add_slice calls don't work right for bottom
2392      * fields; they cause massive erroneous error concealing
2393      * Error marking covers both fields (top and bottom).
2394      * This causes a mismatched s->error_count
2395      * and a bad error table. Further, the error count goes to
2396      * INT_MAX when called for bottom field, because mb_y is
2397      * past end by one (callers fault) and resync_mb_y != 0
2398      * causes problems for the first MB line, too.
2399      */
2400     if (!FIELD_PICTURE)
2401         ff_er_frame_end(s);
2402
2403     MPV_frame_end(s);
2404
2405     h->current_slice=0;
2406 }
2407
2408 /**
2409  * Replicate H264 "master" context to thread contexts.
2410  */
2411 static void clone_slice(H264Context *dst, H264Context *src)
2412 {
2413     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2414     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2415     dst->s.current_picture      = src->s.current_picture;
2416     dst->s.linesize             = src->s.linesize;
2417     dst->s.uvlinesize           = src->s.uvlinesize;
2418     dst->s.first_field          = src->s.first_field;
2419
2420     dst->prev_poc_msb           = src->prev_poc_msb;
2421     dst->prev_poc_lsb           = src->prev_poc_lsb;
2422     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2423     dst->prev_frame_num         = src->prev_frame_num;
2424     dst->short_ref_count        = src->short_ref_count;
2425
2426     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2427     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2428     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2429     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2430
2431     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2432     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2433 }
2434
2435 /**
2436  * computes profile from profile_idc and constraint_set?_flags
2437  *
2438  * @param sps SPS
2439  *
2440  * @return profile as defined by FF_PROFILE_H264_*
2441  */
2442 int ff_h264_get_profile(SPS *sps)
2443 {
2444     int profile = sps->profile_idc;
2445
2446     switch(sps->profile_idc) {
2447     case FF_PROFILE_H264_BASELINE:
2448         // constraint_set1_flag set to 1
2449         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2450         break;
2451     case FF_PROFILE_H264_HIGH_10:
2452     case FF_PROFILE_H264_HIGH_422:
2453     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2454         // constraint_set3_flag set to 1
2455         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2456         break;
2457     }
2458
2459     return profile;
2460 }
2461
2462 /**
2463  * decodes a slice header.
2464  * This will also call MPV_common_init() and frame_start() as needed.
2465  *
2466  * @param h h264context
2467  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2468  *
2469  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2470  */
2471 static int decode_slice_header(H264Context *h, H264Context *h0){
2472     MpegEncContext * const s = &h->s;
2473     MpegEncContext * const s0 = &h0->s;
2474     unsigned int first_mb_in_slice;
2475     unsigned int pps_id;
2476     int num_ref_idx_active_override_flag;
2477     unsigned int slice_type, tmp, i, j;
2478     int default_ref_list_done = 0;
2479     int last_pic_structure;
2480
2481     s->dropable= h->nal_ref_idc == 0;
2482
2483     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc){
2484         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2485         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2486     }else{
2487         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2488         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2489     }
2490
2491     first_mb_in_slice= get_ue_golomb(&s->gb);
2492
2493     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2494         if(h0->current_slice && FIELD_PICTURE){
2495             field_end(h, 1);
2496         }
2497
2498         h0->current_slice = 0;
2499         if (!s0->first_field)
2500             s->current_picture_ptr= NULL;
2501     }
2502
2503     slice_type= get_ue_golomb_31(&s->gb);
2504     if(slice_type > 9){
2505         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);
2506         return -1;
2507     }
2508     if(slice_type > 4){
2509         slice_type -= 5;
2510         h->slice_type_fixed=1;
2511     }else
2512         h->slice_type_fixed=0;
2513
2514     slice_type= golomb_to_pict_type[ slice_type ];
2515     if (slice_type == AV_PICTURE_TYPE_I
2516         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2517         default_ref_list_done = 1;
2518     }
2519     h->slice_type= slice_type;
2520     h->slice_type_nos= slice_type & 3;
2521
2522     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2523
2524     pps_id= get_ue_golomb(&s->gb);
2525     if(pps_id>=MAX_PPS_COUNT){
2526         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2527         return -1;
2528     }
2529     if(!h0->pps_buffers[pps_id]) {
2530         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2531         return -1;
2532     }
2533     h->pps= *h0->pps_buffers[pps_id];
2534
2535     if(!h0->sps_buffers[h->pps.sps_id]) {
2536         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2537         return -1;
2538     }
2539     h->sps = *h0->sps_buffers[h->pps.sps_id];
2540
2541     s->avctx->profile = ff_h264_get_profile(&h->sps);
2542     s->avctx->level   = h->sps.level_idc;
2543     s->avctx->refs    = h->sps.ref_frame_count;
2544
2545     if(h == h0 && h->dequant_coeff_pps != pps_id){
2546         h->dequant_coeff_pps = pps_id;
2547         init_dequant_tables(h);
2548     }
2549
2550     s->mb_width= h->sps.mb_width;
2551     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2552
2553     h->b_stride=  s->mb_width*4;
2554
2555     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2556     if(h->sps.frame_mbs_only_flag)
2557         s->height= 16*s->mb_height - (2>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2558     else
2559         s->height= 16*s->mb_height - (4>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2560
2561     if (s->context_initialized
2562         && (   s->width != s->avctx->width || s->height != s->avctx->height
2563             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2564         if(h != h0) {
2565             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2566             return -1;   // width / height changed during parallelized decoding
2567         }
2568         free_tables(h, 0);
2569         flush_dpb(s->avctx);
2570         MPV_common_end(s);
2571     }
2572     if (!s->context_initialized) {
2573         if (h != h0) {
2574             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2575             return -1;
2576         }
2577
2578         avcodec_set_dimensions(s->avctx, s->width, s->height);
2579         s->avctx->sample_aspect_ratio= h->sps.sar;
2580         av_assert0(s->avctx->sample_aspect_ratio.den);
2581
2582         h->s.avctx->coded_width = 16*s->mb_width;
2583         h->s.avctx->coded_height = 16*s->mb_height;
2584
2585         if(h->sps.video_signal_type_present_flag){
2586             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2587             if(h->sps.colour_description_present_flag){
2588                 s->avctx->color_primaries = h->sps.color_primaries;
2589                 s->avctx->color_trc       = h->sps.color_trc;
2590                 s->avctx->colorspace      = h->sps.colorspace;
2591             }
2592         }
2593
2594         if(h->sps.timing_info_present_flag){
2595             int64_t den= h->sps.time_scale;
2596             if(h->x264_build < 44U)
2597                 den *= 2;
2598             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2599                       h->sps.num_units_in_tick, den, 1<<30);
2600         }
2601
2602         switch (h->sps.bit_depth_luma) {
2603             case 9 :
2604                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P9 : PIX_FMT_YUV420P9;
2605                 break;
2606             case 10 :
2607                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P10 : PIX_FMT_YUV420P10;
2608                 break;
2609             default:
2610                 if (CHROMA444){
2611                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2612                 }else{
2613                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2614                                                              s->avctx->codec->pix_fmts ?
2615                                                              s->avctx->codec->pix_fmts :
2616                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2617                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2618                                                              ff_hwaccel_pixfmt_list_420);
2619                 }
2620         }
2621
2622         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2623
2624         if (MPV_common_init(s) < 0) {
2625             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2626             return -1;
2627         }
2628         s->first_field = 0;
2629         h->prev_interlaced_frame = 1;
2630
2631         init_scan_tables(h);
2632         ff_h264_alloc_tables(h);
2633
2634         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2635             if (context_init(h) < 0) {
2636                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2637                 return -1;
2638             }
2639         } else {
2640             for(i = 1; i < s->avctx->thread_count; i++) {
2641                 H264Context *c;
2642                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2643                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2644                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2645                 c->h264dsp = h->h264dsp;
2646                 c->sps = h->sps;
2647                 c->pps = h->pps;
2648                 c->pixel_shift = h->pixel_shift;
2649                 init_scan_tables(c);
2650                 clone_tables(c, h, i);
2651             }
2652
2653             for(i = 0; i < s->avctx->thread_count; i++)
2654                 if (context_init(h->thread_context[i]) < 0) {
2655                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2656                     return -1;
2657                 }
2658         }
2659     }
2660
2661     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2662
2663     h->mb_mbaff = 0;
2664     h->mb_aff_frame = 0;
2665     last_pic_structure = s0->picture_structure;
2666     if(h->sps.frame_mbs_only_flag){
2667         s->picture_structure= PICT_FRAME;
2668     }else{
2669         if(get_bits1(&s->gb)) { //field_pic_flag
2670             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2671         } else {
2672             s->picture_structure= PICT_FRAME;
2673             h->mb_aff_frame = h->sps.mb_aff;
2674         }
2675     }
2676     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2677
2678     if(h0->current_slice == 0){
2679         if(h->frame_num != h->prev_frame_num &&
2680           (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num) < (h->frame_num - h->sps.ref_frame_count))
2681             h->prev_frame_num = h->frame_num - h->sps.ref_frame_count - 1;
2682
2683         while(h->frame_num !=  h->prev_frame_num &&
2684               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2685             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2686             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2687             if (ff_h264_frame_start(h) < 0)
2688                 return -1;
2689             h->prev_frame_num++;
2690             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2691             s->current_picture_ptr->frame_num= h->prev_frame_num;
2692             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2693             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2694             ff_generate_sliding_window_mmcos(h);
2695             ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2696             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2697              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2698              * about there being no actual duplicates.
2699              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2700              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2701              * be fixed. */
2702             if (h->short_ref_count) {
2703                 if (prev) {
2704                     av_image_copy(h->short_ref[0]->data, h->short_ref[0]->linesize,
2705                                   (const uint8_t**)prev->data, prev->linesize,
2706                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2707                     h->short_ref[0]->poc = prev->poc+2;
2708                 }
2709                 h->short_ref[0]->frame_num = h->prev_frame_num;
2710             }
2711         }
2712
2713         /* See if we have a decoded first field looking for a pair... */
2714         if (s0->first_field) {
2715             assert(s0->current_picture_ptr);
2716             assert(s0->current_picture_ptr->data[0]);
2717             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
2718
2719             /* figure out if we have a complementary field pair */
2720             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2721                 /*
2722                  * Previous field is unmatched. Don't display it, but let it
2723                  * remain for reference if marked as such.
2724                  */
2725                 s0->current_picture_ptr = NULL;
2726                 s0->first_field = FIELD_PICTURE;
2727
2728             } else {
2729                 if (h->nal_ref_idc &&
2730                         s0->current_picture_ptr->reference &&
2731                         s0->current_picture_ptr->frame_num != h->frame_num) {
2732                     /*
2733                      * This and previous field were reference, but had
2734                      * different frame_nums. Consider this field first in
2735                      * pair. Throw away previous field except for reference
2736                      * purposes.
2737                      */
2738                     s0->first_field = 1;
2739                     s0->current_picture_ptr = NULL;
2740
2741                 } else {
2742                     /* Second field in complementary pair */
2743                     s0->first_field = 0;
2744                 }
2745             }
2746
2747         } else {
2748             /* Frame or first field in a potentially complementary pair */
2749             assert(!s0->current_picture_ptr);
2750             s0->first_field = FIELD_PICTURE;
2751         }
2752
2753         if(!FIELD_PICTURE || s0->first_field) {
2754             if (ff_h264_frame_start(h) < 0) {
2755                 s0->first_field = 0;
2756                 return -1;
2757             }
2758         } else {
2759             ff_release_unused_pictures(s, 0);
2760         }
2761     }
2762     if(h != h0)
2763         clone_slice(h, h0);
2764
2765     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2766
2767     assert(s->mb_num == s->mb_width * s->mb_height);
2768     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2769        first_mb_in_slice                    >= s->mb_num){
2770         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2771         return -1;
2772     }
2773     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2774     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2775     if (s->picture_structure == PICT_BOTTOM_FIELD)
2776         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2777     assert(s->mb_y < s->mb_height);
2778
2779     if(s->picture_structure==PICT_FRAME){
2780         h->curr_pic_num=   h->frame_num;
2781         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2782     }else{
2783         h->curr_pic_num= 2*h->frame_num + 1;
2784         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2785     }
2786
2787     if(h->nal_unit_type == NAL_IDR_SLICE){
2788         get_ue_golomb(&s->gb); /* idr_pic_id */
2789     }
2790
2791     if(h->sps.poc_type==0){
2792         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2793
2794         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2795             h->delta_poc_bottom= get_se_golomb(&s->gb);
2796         }
2797     }
2798
2799     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2800         h->delta_poc[0]= get_se_golomb(&s->gb);
2801
2802         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2803             h->delta_poc[1]= get_se_golomb(&s->gb);
2804     }
2805
2806     init_poc(h);
2807
2808     if(h->pps.redundant_pic_cnt_present){
2809         h->redundant_pic_count= get_ue_golomb(&s->gb);
2810     }
2811
2812     //set defaults, might be overridden a few lines later
2813     h->ref_count[0]= h->pps.ref_count[0];
2814     h->ref_count[1]= h->pps.ref_count[1];
2815
2816     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2817         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2818             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2819         }
2820         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2821
2822         if(num_ref_idx_active_override_flag){
2823             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2824             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2825                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2826
2827             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2828                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2829                 h->ref_count[0]= h->ref_count[1]= 1;
2830                 return -1;
2831             }
2832         }
2833         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2834             h->list_count= 2;
2835         else
2836             h->list_count= 1;
2837     }else
2838         h->list_count= 0;
2839
2840     if(!default_ref_list_done){
2841         ff_h264_fill_default_ref_list(h);
2842     }
2843
2844     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0)
2845         return -1;
2846
2847     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2848         s->last_picture_ptr= &h->ref_list[0][0];
2849         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2850     }
2851     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2852         s->next_picture_ptr= &h->ref_list[1][0];
2853         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2854     }
2855
2856     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2857        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2858         pred_weight_table(h);
2859     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2860         implicit_weight_table(h, -1);
2861     }else {
2862         h->use_weight = 0;
2863         for (i = 0; i < 2; i++) {
2864             h->luma_weight_flag[i]   = 0;
2865             h->chroma_weight_flag[i] = 0;
2866         }
2867     }
2868
2869     if(h->nal_ref_idc)
2870         ff_h264_decode_ref_pic_marking(h0, &s->gb);
2871
2872     if(FRAME_MBAFF){
2873         ff_h264_fill_mbaff_ref_list(h);
2874
2875         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2876             implicit_weight_table(h, 0);
2877             implicit_weight_table(h, 1);
2878         }
2879     }
2880
2881     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2882         ff_h264_direct_dist_scale_factor(h);
2883     ff_h264_direct_ref_list_init(h);
2884
2885     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2886         tmp = get_ue_golomb_31(&s->gb);
2887         if(tmp > 2){
2888             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2889             return -1;
2890         }
2891         h->cabac_init_idc= tmp;
2892     }
2893
2894     h->last_qscale_diff = 0;
2895     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2896     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
2897         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2898         return -1;
2899     }
2900     s->qscale= tmp;
2901     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2902     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2903     //FIXME qscale / qp ... stuff
2904     if(h->slice_type == AV_PICTURE_TYPE_SP){
2905         get_bits1(&s->gb); /* sp_for_switch_flag */
2906     }
2907     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
2908         get_se_golomb(&s->gb); /* slice_qs_delta */
2909     }
2910
2911     h->deblocking_filter = 1;
2912     h->slice_alpha_c0_offset = 52;
2913     h->slice_beta_offset = 52;
2914     if( h->pps.deblocking_filter_parameters_present ) {
2915         tmp= get_ue_golomb_31(&s->gb);
2916         if(tmp > 2){
2917             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2918             return -1;
2919         }
2920         h->deblocking_filter= tmp;
2921         if(h->deblocking_filter < 2)
2922             h->deblocking_filter^= 1; // 1<->0
2923
2924         if( h->deblocking_filter ) {
2925             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2926             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2927             if(   h->slice_alpha_c0_offset > 104U
2928                || h->slice_beta_offset     > 104U){
2929                 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);
2930                 return -1;
2931             }
2932         }
2933     }
2934
2935     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2936        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
2937        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
2938        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2939         h->deblocking_filter= 0;
2940
2941     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
2942         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
2943             /* Cheat slightly for speed:
2944                Do not bother to deblock across slices. */
2945             h->deblocking_filter = 2;
2946         } else {
2947             h0->max_contexts = 1;
2948             if(!h0->single_decode_warning) {
2949                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
2950                 h0->single_decode_warning = 1;
2951             }
2952             if (h != h0) {
2953                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
2954                 return 1;
2955             }
2956         }
2957     }
2958     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]);
2959
2960 #if 0 //FMO
2961     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
2962         slice_group_change_cycle= get_bits(&s->gb, ?);
2963 #endif
2964
2965     h0->last_slice_type = slice_type;
2966     h->slice_num = ++h0->current_slice;
2967     if(h->slice_num >= MAX_SLICES){
2968         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
2969     }
2970
2971     for(j=0; j<2; j++){
2972         int id_list[16];
2973         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
2974         for(i=0; i<16; i++){
2975             id_list[i]= 60;
2976             if(h->ref_list[j][i].data[0]){
2977                 int k;
2978                 uint8_t *base= h->ref_list[j][i].base[0];
2979                 for(k=0; k<h->short_ref_count; k++)
2980                     if(h->short_ref[k]->base[0] == base){
2981                         id_list[i]= k;
2982                         break;
2983                     }
2984                 for(k=0; k<h->long_ref_count; k++)
2985                     if(h->long_ref[k] && h->long_ref[k]->base[0] == base){
2986                         id_list[i]= h->short_ref_count + k;
2987                         break;
2988                     }
2989             }
2990         }
2991
2992         ref2frm[0]=
2993         ref2frm[1]= -1;
2994         for(i=0; i<16; i++)
2995             ref2frm[i+2]= 4*id_list[i]
2996                           +(h->ref_list[j][i].reference&3);
2997         ref2frm[18+0]=
2998         ref2frm[18+1]= -1;
2999         for(i=16; i<48; i++)
3000             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3001                           +(h->ref_list[j][i].reference&3);
3002     }
3003
3004     //FIXME: fix draw_edges+PAFF+frame threads
3005     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3006     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3007
3008     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3009         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",
3010                h->slice_num,
3011                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3012                first_mb_in_slice,
3013                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3014                pps_id, h->frame_num,
3015                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3016                h->ref_count[0], h->ref_count[1],
3017                s->qscale,
3018                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3019                h->use_weight,
3020                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3021                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3022                );
3023     }
3024
3025     return 0;
3026 }
3027
3028 int ff_h264_get_slice_type(const H264Context *h)
3029 {
3030     switch (h->slice_type) {
3031     case AV_PICTURE_TYPE_P:  return 0;
3032     case AV_PICTURE_TYPE_B:  return 1;
3033     case AV_PICTURE_TYPE_I:  return 2;
3034     case AV_PICTURE_TYPE_SP: return 3;
3035     case AV_PICTURE_TYPE_SI: return 4;
3036     default:         return -1;
3037     }
3038 }
3039
3040 /**
3041  *
3042  * @return non zero if the loop filter can be skiped
3043  */
3044 static int fill_filter_caches(H264Context *h, int mb_type){
3045     MpegEncContext * const s = &h->s;
3046     const int mb_xy= h->mb_xy;
3047     int top_xy, left_xy[2];
3048     int top_type, left_type[2];
3049
3050     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3051
3052     //FIXME deblocking could skip the intra and nnz parts.
3053
3054     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3055      * stuff, I can't imagine that these complex rules are worth it. */
3056
3057     left_xy[1] = left_xy[0] = mb_xy-1;
3058     if(FRAME_MBAFF){
3059         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[mb_xy-1]);
3060         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3061         if(s->mb_y&1){
3062             if (left_mb_field_flag != curr_mb_field_flag) {
3063                 left_xy[0] -= s->mb_stride;
3064             }
3065         }else{
3066             if(curr_mb_field_flag){
3067                 top_xy      += s->mb_stride & (((s->current_picture.mb_type[top_xy    ]>>7)&1)-1);
3068             }
3069             if (left_mb_field_flag != curr_mb_field_flag) {
3070                 left_xy[1] += s->mb_stride;
3071             }
3072         }
3073     }
3074
3075     h->top_mb_xy = top_xy;
3076     h->left_mb_xy[0] = left_xy[0];
3077     h->left_mb_xy[1] = left_xy[1];
3078     {
3079         //for sufficiently low qp, filtering wouldn't do anything
3080         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3081         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3082         int qp = s->current_picture.qscale_table[mb_xy];
3083         if(qp <= qp_thresh
3084            && (left_xy[0]<0 || ((qp + s->current_picture.qscale_table[left_xy[0]] + 1)>>1) <= qp_thresh)
3085            && (top_xy   < 0 || ((qp + s->current_picture.qscale_table[top_xy    ] + 1)>>1) <= qp_thresh)){
3086             if(!FRAME_MBAFF)
3087                 return 1;
3088             if(   (left_xy[0]< 0            || ((qp + s->current_picture.qscale_table[left_xy[1]             ] + 1)>>1) <= qp_thresh)
3089                && (top_xy    < s->mb_stride || ((qp + s->current_picture.qscale_table[top_xy    -s->mb_stride] + 1)>>1) <= qp_thresh))
3090                 return 1;
3091         }
3092     }
3093
3094     top_type     = s->current_picture.mb_type[top_xy]    ;
3095     left_type[0] = s->current_picture.mb_type[left_xy[0]];
3096     left_type[1] = s->current_picture.mb_type[left_xy[1]];
3097     if(h->deblocking_filter == 2){
3098         if(h->slice_table[top_xy     ] != h->slice_num) top_type= 0;
3099         if(h->slice_table[left_xy[0] ] != h->slice_num) left_type[0]= left_type[1]= 0;
3100     }else{
3101         if(h->slice_table[top_xy     ] == 0xFFFF) top_type= 0;
3102         if(h->slice_table[left_xy[0] ] == 0xFFFF) left_type[0]= left_type[1] =0;
3103     }
3104     h->top_type    = top_type    ;
3105     h->left_type[0]= left_type[0];
3106     h->left_type[1]= left_type[1];
3107
3108     if(IS_INTRA(mb_type))
3109         return 0;
3110
3111     AV_COPY32(&h->non_zero_count_cache[4+8* 1], &h->non_zero_count[mb_xy][ 0]);
3112     AV_COPY32(&h->non_zero_count_cache[4+8* 2], &h->non_zero_count[mb_xy][ 4]);
3113     AV_COPY32(&h->non_zero_count_cache[4+8* 3], &h->non_zero_count[mb_xy][ 8]);
3114     AV_COPY32(&h->non_zero_count_cache[4+8* 4], &h->non_zero_count[mb_xy][12]);
3115
3116     h->cbp= h->cbp_table[mb_xy];
3117
3118     {
3119         int list;
3120         for(list=0; list<h->list_count; list++){
3121             int8_t *ref;
3122             int y, b_stride;
3123             int16_t (*mv_dst)[2];
3124             int16_t (*mv_src)[2];
3125
3126             if(!USES_LIST(mb_type, list)){
3127                 fill_rectangle(  h->mv_cache[list][scan8[0]], 4, 4, 8, pack16to32(0,0), 4);
3128                 AV_WN32A(&h->ref_cache[list][scan8[ 0]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3129                 AV_WN32A(&h->ref_cache[list][scan8[ 2]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3130                 AV_WN32A(&h->ref_cache[list][scan8[ 8]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3131                 AV_WN32A(&h->ref_cache[list][scan8[10]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3132                 continue;
3133             }
3134
3135             ref = &s->current_picture.ref_index[list][4*mb_xy];
3136             {
3137                 int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3138                 AV_WN32A(&h->ref_cache[list][scan8[ 0]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
3139                 AV_WN32A(&h->ref_cache[list][scan8[ 2]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
3140                 ref += 2;
3141                 AV_WN32A(&h->ref_cache[list][scan8[ 8]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
3142                 AV_WN32A(&h->ref_cache[list][scan8[10]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
3143             }
3144
3145             b_stride = h->b_stride;
3146             mv_dst   = &h->mv_cache[list][scan8[0]];
3147             mv_src   = &s->current_picture.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3148             for(y=0; y<4; y++){
3149                 AV_COPY128(mv_dst + 8*y, mv_src + y*b_stride);
3150             }
3151
3152         }
3153     }
3154
3155
3156 /*
3157 0 . T T. T T T T
3158 1 L . .L . . . .
3159 2 L . .L . . . .
3160 3 . T TL . . . .
3161 4 L . .L . . . .
3162 5 L . .. . . . .
3163 */
3164 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
3165     if(top_type){
3166         AV_COPY32(&h->non_zero_count_cache[4+8*0], &h->non_zero_count[top_xy][3*4]);
3167     }
3168
3169     if(left_type[0]){
3170         h->non_zero_count_cache[3+8*1]= h->non_zero_count[left_xy[0]][3+0*4];
3171         h->non_zero_count_cache[3+8*2]= h->non_zero_count[left_xy[0]][3+1*4];
3172         h->non_zero_count_cache[3+8*3]= h->non_zero_count[left_xy[0]][3+2*4];
3173         h->non_zero_count_cache[3+8*4]= h->non_zero_count[left_xy[0]][3+3*4];
3174     }
3175
3176     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3177     if(!CABAC && h->pps.transform_8x8_mode){
3178         if(IS_8x8DCT(top_type)){
3179             h->non_zero_count_cache[4+8*0]=
3180             h->non_zero_count_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3181             h->non_zero_count_cache[6+8*0]=
3182             h->non_zero_count_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3183         }
3184         if(IS_8x8DCT(left_type[0])){
3185             h->non_zero_count_cache[3+8*1]=
3186             h->non_zero_count_cache[3+8*2]= (h->cbp_table[left_xy[0]]&0x2000) >> 12; //FIXME check MBAFF
3187         }
3188         if(IS_8x8DCT(left_type[1])){
3189             h->non_zero_count_cache[3+8*3]=
3190             h->non_zero_count_cache[3+8*4]= (h->cbp_table[left_xy[1]]&0x8000) >> 12; //FIXME check MBAFF
3191         }
3192
3193         if(IS_8x8DCT(mb_type)){
3194             h->non_zero_count_cache[scan8[0   ]]= h->non_zero_count_cache[scan8[1   ]]=
3195             h->non_zero_count_cache[scan8[2   ]]= h->non_zero_count_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3196
3197             h->non_zero_count_cache[scan8[0+ 4]]= h->non_zero_count_cache[scan8[1+ 4]]=
3198             h->non_zero_count_cache[scan8[2+ 4]]= h->non_zero_count_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3199
3200             h->non_zero_count_cache[scan8[0+ 8]]= h->non_zero_count_cache[scan8[1+ 8]]=
3201             h->non_zero_count_cache[scan8[2+ 8]]= h->non_zero_count_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3202
3203             h->non_zero_count_cache[scan8[0+12]]= h->non_zero_count_cache[scan8[1+12]]=
3204             h->non_zero_count_cache[scan8[2+12]]= h->non_zero_count_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3205         }
3206     }
3207
3208     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3209         int list;
3210         for(list=0; list<h->list_count; list++){
3211             if(USES_LIST(top_type, list)){
3212                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
3213                 const int b8_xy= 4*top_xy + 2;
3214                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3215                 AV_COPY128(h->mv_cache[list][scan8[0] + 0 - 1*8], s->current_picture.motion_val[list][b_xy + 0]);
3216                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
3217                 h->ref_cache[list][scan8[0] + 1 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 0]];
3218                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
3219                 h->ref_cache[list][scan8[0] + 3 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 1]];
3220             }else{
3221                 AV_ZERO128(h->mv_cache[list][scan8[0] + 0 - 1*8]);
3222                 AV_WN32A(&h->ref_cache[list][scan8[0] + 0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3223             }
3224
3225             if(!IS_INTERLACED(mb_type^left_type[0])){
3226                 if(USES_LIST(left_type[0], list)){
3227                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
3228                     const int b8_xy= 4*left_xy[0] + 1;
3229                     int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[0]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3230                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 0 ], s->current_picture.motion_val[list][b_xy + h->b_stride*0]);
3231                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 8 ], s->current_picture.motion_val[list][b_xy + h->b_stride*1]);
3232                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 +16 ], s->current_picture.motion_val[list][b_xy + h->b_stride*2]);
3233                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 +24 ], s->current_picture.motion_val[list][b_xy + h->b_stride*3]);
3234                     h->ref_cache[list][scan8[0] - 1 + 0 ]=
3235                     h->ref_cache[list][scan8[0] - 1 + 8 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*0]];
3236                     h->ref_cache[list][scan8[0] - 1 +16 ]=
3237                     h->ref_cache[list][scan8[0] - 1 +24 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*1]];
3238                 }else{
3239                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 0 ]);
3240                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 8 ]);
3241                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 +16 ]);
3242                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 +24 ]);
3243                     h->ref_cache[list][scan8[0] - 1 + 0  ]=
3244                     h->ref_cache[list][scan8[0] - 1 + 8  ]=
3245                     h->ref_cache[list][scan8[0] - 1 + 16 ]=
3246                     h->ref_cache[list][scan8[0] - 1 + 24 ]= LIST_NOT_USED;
3247                 }
3248             }
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*7;
3288                         dest_cr-= s->uvlinesize*7;
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);
3467
3468                         return 0;
3469                     }else{
3470                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3471
3472                         return -1;
3473                     }
3474                 }
3475             }
3476
3477             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
3478                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3479                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
3480                     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);
3481                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3482
3483                     return 0;
3484                 }else{
3485                     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);
3486
3487                     return -1;
3488                 }
3489             }
3490         }
3491     }
3492
3493 #if 0
3494     for(;s->mb_y < s->mb_height; s->mb_y++){
3495         for(;s->mb_x < s->mb_width; s->mb_x++){
3496             int ret= decode_mb(h);
3497
3498             ff_h264_hl_decode_mb(h);
3499
3500             if(ret<0){
3501                 av_log(s->avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3502                 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);
3503
3504                 return -1;
3505             }
3506
3507             if(++s->mb_x >= s->mb_width){
3508                 s->mb_x=0;
3509                 if(++s->mb_y >= s->mb_height){
3510                     if(get_bits_count(s->gb) == s->gb.size_in_bits){
3511                         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);
3512
3513                         return 0;
3514                     }else{
3515                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3516
3517                         return -1;
3518                     }
3519                 }
3520             }
3521
3522             if(get_bits_count(s->?gb) >= s->gb?.size_in_bits){
3523                 if(get_bits_count(s->gb) == s->gb.size_in_bits){
3524                     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);
3525
3526                     return 0;
3527                 }else{
3528                     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);
3529
3530                     return -1;
3531                 }
3532             }
3533         }
3534         s->mb_x=0;
3535         ff_draw_horiz_band(s, 16*s->mb_y, 16);
3536     }
3537 #endif
3538     return -1; //not reached
3539 }
3540
3541 /**
3542  * Call decode_slice() for each context.
3543  *
3544  * @param h h264 master context
3545  * @param context_count number of contexts to execute
3546  */
3547 static void execute_decode_slices(H264Context *h, int context_count){
3548     MpegEncContext * const s = &h->s;
3549     AVCodecContext * const avctx= s->avctx;
3550     H264Context *hx;
3551     int i;
3552
3553     if (s->avctx->hwaccel)
3554         return;
3555     if(s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3556         return;
3557     if(context_count == 1) {
3558         decode_slice(avctx, &h);
3559     } else {
3560         for(i = 1; i < context_count; i++) {
3561             hx = h->thread_context[i];
3562             hx->s.error_recognition = avctx->error_recognition;
3563             hx->s.error_count = 0;
3564         }
3565
3566         avctx->execute(avctx, (void *)decode_slice,
3567                        h->thread_context, NULL, context_count, sizeof(void*));
3568
3569         /* pull back stuff from slices to master context */
3570         hx = h->thread_context[context_count - 1];
3571         s->mb_x = hx->s.mb_x;
3572         s->mb_y = hx->s.mb_y;
3573         s->dropable = hx->s.dropable;
3574         s->picture_structure = hx->s.picture_structure;
3575         for(i = 1; i < context_count; i++)
3576             h->s.error_count += h->thread_context[i]->s.error_count;
3577     }
3578 }
3579
3580
3581 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
3582     MpegEncContext * const s = &h->s;
3583     AVCodecContext * const avctx= s->avctx;
3584     H264Context *hx; ///< thread context
3585     int buf_index;
3586     int context_count;
3587     int next_avc;
3588     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3589     int nals_needed=0; ///< number of NALs that need decoding before the next frame thread starts
3590     int nal_index;
3591
3592     h->max_contexts = (HAVE_THREADS && (s->avctx->active_thread_type&FF_THREAD_SLICE)) ? avctx->thread_count : 1;
3593     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
3594         h->current_slice = 0;
3595         if (!s->first_field)
3596             s->current_picture_ptr= NULL;
3597         ff_h264_reset_sei(h);
3598     }
3599
3600     for(;pass <= 1;pass++){
3601         buf_index = 0;
3602         context_count = 0;
3603         next_avc = h->is_avc ? 0 : buf_size;
3604         nal_index = 0;
3605     for(;;){
3606         int consumed;
3607         int dst_length;
3608         int bit_length;
3609         const uint8_t *ptr;
3610         int i, nalsize = 0;
3611         int err;
3612
3613         if(buf_index >= next_avc) {
3614             if(buf_index >= buf_size) break;
3615             nalsize = 0;
3616             for(i = 0; i < h->nal_length_size; i++)
3617                 nalsize = (nalsize << 8) | buf[buf_index++];
3618             if(nalsize <= 0 || nalsize > buf_size - buf_index){
3619                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
3620                 break;
3621             }
3622             next_avc= buf_index + nalsize;
3623         } else {
3624             // start code prefix search
3625             for(; buf_index + 3 < next_avc; buf_index++){
3626                 // This should always succeed in the first iteration.
3627                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
3628                     break;
3629             }
3630
3631             if(buf_index+3 >= buf_size) break;
3632
3633             buf_index+=3;
3634             if(buf_index >= next_avc) continue;
3635         }
3636
3637         hx = h->thread_context[context_count];
3638
3639         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
3640         if (ptr==NULL || dst_length < 0){
3641             return -1;
3642         }
3643         i= buf_index + consumed;
3644         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
3645            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
3646             s->workaround_bugs |= FF_BUG_TRUNCATED;
3647
3648         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
3649         while(ptr[dst_length - 1] == 0 && dst_length > 0)
3650             dst_length--;
3651         }
3652         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
3653
3654         if(s->avctx->debug&FF_DEBUG_STARTCODE){
3655             av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d at %d/%d length %d\n", hx->nal_unit_type, buf_index, buf_size, dst_length);
3656         }
3657
3658         if (h->is_avc && (nalsize != consumed) && nalsize){
3659             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
3660         }
3661
3662         buf_index += consumed;
3663         nal_index++;
3664
3665         if(pass == 0) {
3666             // packets can sometimes contain multiple PPS/SPS
3667             // e.g. two PAFF field pictures in one packet, or a demuxer which splits NALs strangely
3668             // if so, when frame threading we can't start the next thread until we've read all of them
3669             switch (hx->nal_unit_type) {
3670                 case NAL_SPS:
3671                 case NAL_PPS:
3672                     nals_needed = nal_index;
3673             }
3674             continue;
3675         }
3676
3677         //FIXME do not discard SEI id
3678         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
3679             continue;
3680
3681       again:
3682         err = 0;
3683         switch(hx->nal_unit_type){
3684         case NAL_IDR_SLICE:
3685             if (h->nal_unit_type != NAL_IDR_SLICE) {
3686                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
3687                 return -1;
3688             }
3689             idr(h); //FIXME ensure we don't loose some frames if there is reordering
3690         case NAL_SLICE:
3691             init_get_bits(&hx->s.gb, ptr, bit_length);
3692             hx->intra_gb_ptr=
3693             hx->inter_gb_ptr= &hx->s.gb;
3694             hx->s.data_partitioning = 0;
3695
3696             if((err = decode_slice_header(hx, h)))
3697                break;
3698
3699             s->current_picture_ptr->key_frame |=
3700                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
3701                     (h->sei_recovery_frame_cnt >= 0);
3702
3703             if (h->current_slice == 1) {
3704                 if(!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3705                     decode_postinit(h, nal_index >= nals_needed);
3706                 }
3707
3708                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3709                     return -1;
3710                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3711                     ff_vdpau_h264_picture_start(s);
3712             }
3713
3714             if(hx->redundant_pic_count==0
3715                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3716                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3717                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3718                && avctx->skip_frame < AVDISCARD_ALL){
3719                 if(avctx->hwaccel) {
3720                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
3721                         return -1;
3722                 }else
3723                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
3724                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
3725                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
3726                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
3727                 }else
3728                     context_count++;
3729             }
3730             break;
3731         case NAL_DPA:
3732             init_get_bits(&hx->s.gb, ptr, bit_length);
3733             hx->intra_gb_ptr=
3734             hx->inter_gb_ptr= NULL;
3735
3736             if ((err = decode_slice_header(hx, h)) < 0)
3737                 break;
3738
3739             hx->s.data_partitioning = 1;
3740
3741             break;
3742         case NAL_DPB:
3743             init_get_bits(&hx->intra_gb, ptr, bit_length);
3744             hx->intra_gb_ptr= &hx->intra_gb;
3745             break;
3746         case NAL_DPC:
3747             init_get_bits(&hx->inter_gb, ptr, bit_length);
3748             hx->inter_gb_ptr= &hx->inter_gb;
3749
3750             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
3751                && s->context_initialized
3752                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3753                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3754                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3755                && avctx->skip_frame < AVDISCARD_ALL)
3756                 context_count++;
3757             break;
3758         case NAL_SEI:
3759             init_get_bits(&s->gb, ptr, bit_length);
3760             ff_h264_decode_sei(h);
3761             break;
3762         case NAL_SPS:
3763             init_get_bits(&s->gb, ptr, bit_length);
3764             ff_h264_decode_seq_parameter_set(h);
3765
3766             if(s->flags& CODEC_FLAG_LOW_DELAY)
3767                 s->low_delay=1;
3768
3769             if(avctx->has_b_frames < 2)
3770                 avctx->has_b_frames= !s->low_delay;
3771
3772             if (avctx->bits_per_raw_sample != h->sps.bit_depth_luma) {
3773                 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
3774                     avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3775                     h->pixel_shift = h->sps.bit_depth_luma > 8;
3776
3777                     ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma);
3778                     ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma);
3779                     dsputil_init(&s->dsp, s->avctx);
3780                 } else {
3781                     av_log(avctx, AV_LOG_DEBUG, "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3782                     return -1;
3783                 }
3784             }
3785             break;
3786         case NAL_PPS:
3787             init_get_bits(&s->gb, ptr, bit_length);
3788
3789             ff_h264_decode_picture_parameter_set(h, bit_length);
3790
3791             break;
3792         case NAL_AUD:
3793         case NAL_END_SEQUENCE:
3794         case NAL_END_STREAM:
3795         case NAL_FILLER_DATA:
3796         case NAL_SPS_EXT:
3797         case NAL_AUXILIARY_SLICE:
3798             break;
3799         default:
3800             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
3801         }
3802
3803         if(context_count == h->max_contexts) {
3804             execute_decode_slices(h, context_count);
3805             context_count = 0;
3806         }
3807
3808         if (err < 0)
3809             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3810         else if(err == 1) {
3811             /* Slice could not be decoded in parallel mode, copy down
3812              * NAL unit stuff to context 0 and restart. Note that
3813              * rbsp_buffer is not transferred, but since we no longer
3814              * run in parallel mode this should not be an issue. */
3815             h->nal_unit_type = hx->nal_unit_type;
3816             h->nal_ref_idc   = hx->nal_ref_idc;
3817             hx = h;
3818             goto again;
3819         }
3820     }
3821     }
3822     if(context_count)
3823         execute_decode_slices(h, context_count);
3824     return buf_index;
3825 }
3826
3827 /**
3828  * returns the number of bytes consumed for building the current frame
3829  */
3830 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
3831         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
3832         if(pos+10>buf_size) pos=buf_size; // oops ;)
3833
3834         return pos;
3835 }
3836
3837 static int decode_frame(AVCodecContext *avctx,
3838                              void *data, int *data_size,
3839                              AVPacket *avpkt)
3840 {
3841     const uint8_t *buf = avpkt->data;
3842     int buf_size = avpkt->size;
3843     H264Context *h = avctx->priv_data;
3844     MpegEncContext *s = &h->s;
3845     AVFrame *pict = data;
3846     int buf_index;
3847
3848     s->flags= avctx->flags;
3849     s->flags2= avctx->flags2;
3850
3851    /* end of stream, output what is still in the buffers */
3852  out:
3853     if (buf_size == 0) {
3854         Picture *out;
3855         int i, out_idx;
3856
3857         s->current_picture_ptr = NULL;
3858
3859 //FIXME factorize this with the output code below
3860         out = h->delayed_pic[0];
3861         out_idx = 0;
3862         for(i=1; h->delayed_pic[i] && !h->delayed_pic[i]->key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3863             if(h->delayed_pic[i]->poc < out->poc){
3864                 out = h->delayed_pic[i];
3865                 out_idx = i;
3866             }
3867
3868         for(i=out_idx; h->delayed_pic[i]; i++)
3869             h->delayed_pic[i] = h->delayed_pic[i+1];
3870
3871         if(out){
3872             *data_size = sizeof(AVFrame);
3873             *pict= *(AVFrame*)out;
3874         }
3875
3876         return 0;
3877     }
3878
3879     buf_index=decode_nal_units(h, buf, buf_size);
3880     if(buf_index < 0)
3881         return -1;
3882
3883     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
3884         buf_size = 0;
3885         goto out;
3886     }
3887
3888     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
3889         if (avctx->skip_frame >= AVDISCARD_NONREF)
3890             return 0;
3891         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
3892         return -1;
3893     }
3894
3895     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
3896
3897         if(s->flags2 & CODEC_FLAG2_CHUNKS) decode_postinit(h, 1);
3898
3899         field_end(h, 0);
3900
3901         if (!h->next_output_pic) {
3902             /* Wait for second field. */
3903             *data_size = 0;
3904
3905         } else {
3906             *data_size = sizeof(AVFrame);
3907             *pict = *(AVFrame*)h->next_output_pic;
3908         }
3909     }
3910
3911     assert(pict->data[0] || !*data_size);
3912     ff_print_debug_info(s, pict);
3913 //printf("out %d\n", (int)pict->data[0]);
3914
3915     return get_consumed_bytes(s, buf_index, buf_size);
3916 }
3917 #if 0
3918 static inline void fill_mb_avail(H264Context *h){
3919     MpegEncContext * const s = &h->s;
3920     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3921
3922     if(s->mb_y){
3923         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
3924         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
3925         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
3926     }else{
3927         h->mb_avail[0]=
3928         h->mb_avail[1]=
3929         h->mb_avail[2]= 0;
3930     }
3931     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
3932     h->mb_avail[4]= 1; //FIXME move out
3933     h->mb_avail[5]= 0; //FIXME move out
3934 }
3935 #endif
3936
3937 #ifdef TEST
3938 #undef printf
3939 #undef random
3940 #define COUNT 8000
3941 #define SIZE (COUNT*40)
3942 int main(void){
3943     int i;
3944     uint8_t temp[SIZE];
3945     PutBitContext pb;
3946     GetBitContext gb;
3947 //    int int_temp[10000];
3948     DSPContext dsp;
3949     AVCodecContext avctx;
3950
3951     dsputil_init(&dsp, &avctx);
3952
3953     init_put_bits(&pb, temp, SIZE);
3954     printf("testing unsigned exp golomb\n");
3955     for(i=0; i<COUNT; i++){
3956         START_TIMER
3957         set_ue_golomb(&pb, i);
3958         STOP_TIMER("set_ue_golomb");
3959     }
3960     flush_put_bits(&pb);
3961
3962     init_get_bits(&gb, temp, 8*SIZE);
3963     for(i=0; i<COUNT; i++){
3964         int j, s;
3965
3966         s= show_bits(&gb, 24);
3967
3968         START_TIMER
3969         j= get_ue_golomb(&gb);
3970         if(j != i){
3971             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3972 //            return -1;
3973         }
3974         STOP_TIMER("get_ue_golomb");
3975     }
3976
3977
3978     init_put_bits(&pb, temp, SIZE);
3979     printf("testing signed exp golomb\n");
3980     for(i=0; i<COUNT; i++){
3981         START_TIMER
3982         set_se_golomb(&pb, i - COUNT/2);
3983         STOP_TIMER("set_se_golomb");
3984     }
3985     flush_put_bits(&pb);
3986
3987     init_get_bits(&gb, temp, 8*SIZE);
3988     for(i=0; i<COUNT; i++){
3989         int j, s;
3990
3991         s= show_bits(&gb, 24);
3992
3993         START_TIMER
3994         j= get_se_golomb(&gb);
3995         if(j != i - COUNT/2){
3996             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3997 //            return -1;
3998         }
3999         STOP_TIMER("get_se_golomb");
4000     }
4001
4002 #if 0
4003     printf("testing 4x4 (I)DCT\n");
4004
4005     DCTELEM block[16];
4006     uint8_t src[16], ref[16];
4007     uint64_t error= 0, max_error=0;
4008
4009     for(i=0; i<COUNT; i++){
4010         int j;
4011 //        printf("%d %d %d\n", r1, r2, (r2-r1)*16);
4012         for(j=0; j<16; j++){
4013             ref[j]= random()%255;
4014             src[j]= random()%255;
4015         }
4016
4017         h264_diff_dct_c(block, src, ref, 4);
4018
4019         //normalize
4020         for(j=0; j<16; j++){
4021 //            printf("%d ", block[j]);
4022             block[j]= block[j]*4;
4023             if(j&1) block[j]= (block[j]*4 + 2)/5;
4024             if(j&4) block[j]= (block[j]*4 + 2)/5;
4025         }
4026 //        printf("\n");
4027
4028         h->h264dsp.h264_idct_add(ref, block, 4);
4029 /*        for(j=0; j<16; j++){
4030             printf("%d ", ref[j]);
4031         }
4032         printf("\n");*/
4033
4034         for(j=0; j<16; j++){
4035             int diff= FFABS(src[j] - ref[j]);
4036
4037             error+= diff*diff;
4038             max_error= FFMAX(max_error, diff);
4039         }
4040     }
4041     printf("error=%f max_error=%d\n", ((float)error)/COUNT/16, (int)max_error );
4042     printf("testing quantizer\n");
4043     for(qp=0; qp<52; qp++){
4044         for(i=0; i<16; i++)
4045             src1_block[i]= src2_block[i]= random()%255;
4046
4047     }
4048     printf("Testing NAL layer\n");
4049
4050     uint8_t bitstream[COUNT];
4051     uint8_t nal[COUNT*2];
4052     H264Context h;
4053     memset(&h, 0, sizeof(H264Context));
4054
4055     for(i=0; i<COUNT; i++){
4056         int zeros= i;
4057         int nal_length;
4058         int consumed;
4059         int out_length;
4060         uint8_t *out;
4061         int j;
4062
4063         for(j=0; j<COUNT; j++){
4064             bitstream[j]= (random() % 255) + 1;
4065         }
4066
4067         for(j=0; j<zeros; j++){