h264: fix loopfilter with threading at slice boundaries.
[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 "vdpau_internal.h"
40 #include "libavutil/avassert.h"
41
42 #include "cabac.h"
43
44 //#undef NDEBUG
45 #include <assert.h>
46
47 static const uint8_t rem6[QP_MAX_NUM+1]={
48 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,
49 };
50
51 static const uint8_t div6[QP_MAX_NUM+1]={
52 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,
53 };
54
55 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
56     PIX_FMT_DXVA2_VLD,
57     PIX_FMT_VAAPI_VLD,
58     PIX_FMT_YUVJ420P,
59     PIX_FMT_NONE
60 };
61
62 void ff_h264_write_back_intra_pred_mode(H264Context *h){
63     int8_t *mode= h->intra4x4_pred_mode + h->mb2br_xy[h->mb_xy];
64
65     AV_COPY32(mode, h->intra4x4_pred_mode_cache + 4 + 8*4);
66     mode[4]= h->intra4x4_pred_mode_cache[7+8*3];
67     mode[5]= h->intra4x4_pred_mode_cache[7+8*2];
68     mode[6]= h->intra4x4_pred_mode_cache[7+8*1];
69 }
70
71 /**
72  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
73  */
74 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
75     MpegEncContext * const s = &h->s;
76     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
77     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
78     int i;
79
80     if(!(h->top_samples_available&0x8000)){
81         for(i=0; i<4; i++){
82             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
83             if(status<0){
84                 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);
85                 return -1;
86             } else if(status){
87                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
88             }
89         }
90     }
91
92     if((h->left_samples_available&0x8888)!=0x8888){
93         static const int mask[4]={0x8000,0x2000,0x80,0x20};
94         for(i=0; i<4; i++){
95             if(!(h->left_samples_available&mask[i])){
96                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
97                 if(status<0){
98                     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);
99                     return -1;
100                 } else if(status){
101                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
102                 }
103             }
104         }
105     }
106
107     return 0;
108 } //FIXME cleanup like ff_h264_check_intra_pred_mode
109
110 /**
111  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
112  */
113 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
114     MpegEncContext * const s = &h->s;
115     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
116     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
117
118     if(mode > 6U) {
119         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);
120         return -1;
121     }
122
123     if(!(h->top_samples_available&0x8000)){
124         mode= top[ mode ];
125         if(mode<0){
126             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);
127             return -1;
128         }
129     }
130
131     if((h->left_samples_available&0x8080) != 0x8080){
132         mode= left[ mode ];
133         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
134             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
135         }
136         if(mode<0){
137             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);
138             return -1;
139         }
140     }
141
142     return mode;
143 }
144
145 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
146     int i, si, di;
147     uint8_t *dst;
148     int bufidx;
149
150 //    src[0]&0x80;                //forbidden bit
151     h->nal_ref_idc= src[0]>>5;
152     h->nal_unit_type= src[0]&0x1F;
153
154     src++; length--;
155
156 #if HAVE_FAST_UNALIGNED
157 # if HAVE_FAST_64BIT
158 #   define RS 7
159     for(i=0; i+1<length; i+=9){
160         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
161 # else
162 #   define RS 3
163     for(i=0; i+1<length; i+=5){
164         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
165 # endif
166             continue;
167         if(i>0 && !src[i]) i--;
168         while(src[i]) i++;
169 #else
170 #   define RS 0
171     for(i=0; i+1<length; i+=2){
172         if(src[i]) continue;
173         if(i>0 && src[i-1]==0) i--;
174 #endif
175         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
176             if(src[i+2]!=3){
177                 /* startcode, so we must be past the end */
178                 length=i;
179             }
180             break;
181         }
182         i-= RS;
183     }
184
185     if(i>=length-1){ //no escaped 0
186         *dst_length= length;
187         *consumed= length+1; //+1 for the header
188         return src;
189     }
190
191     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
192     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
193     dst= h->rbsp_buffer[bufidx];
194
195     if (dst == NULL){
196         return NULL;
197     }
198
199 //printf("decoding esc\n");
200     memcpy(dst, src, i);
201     si=di=i;
202     while(si+2<length){
203         //remove escapes (very rare 1:2^22)
204         if(src[si+2]>3){
205             dst[di++]= src[si++];
206             dst[di++]= src[si++];
207         }else if(src[si]==0 && src[si+1]==0){
208             if(src[si+2]==3){ //escape
209                 dst[di++]= 0;
210                 dst[di++]= 0;
211                 si+=3;
212                 continue;
213             }else //next start code
214                 goto nsc;
215         }
216
217         dst[di++]= src[si++];
218     }
219     while(si<length)
220         dst[di++]= src[si++];
221 nsc:
222
223     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
224
225     *dst_length= di;
226     *consumed= si + 1;//+1 for the header
227 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
228     return dst;
229 }
230
231 /**
232  * Identify the exact end of the bitstream
233  * @return the length of the trailing, or 0 if damaged
234  */
235 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
236     int v= *src;
237     int r;
238
239     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
240
241     for(r=1; r<9; r++){
242         if(v&1) return r;
243         v>>=1;
244     }
245     return 0;
246 }
247
248 #if 0
249 /**
250  * DCT transforms the 16 dc values.
251  * @param qp quantization parameter ??? FIXME
252  */
253 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
254 //    const int qmul= dequant_coeff[qp][0];
255     int i;
256     int temp[16]; //FIXME check if this is a good idea
257     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
258     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
259
260     for(i=0; i<4; i++){
261         const int offset= y_offset[i];
262         const int z0= block[offset+stride*0] + block[offset+stride*4];
263         const int z1= block[offset+stride*0] - block[offset+stride*4];
264         const int z2= block[offset+stride*1] - block[offset+stride*5];
265         const int z3= block[offset+stride*1] + block[offset+stride*5];
266
267         temp[4*i+0]= z0+z3;
268         temp[4*i+1]= z1+z2;
269         temp[4*i+2]= z1-z2;
270         temp[4*i+3]= z0-z3;
271     }
272
273     for(i=0; i<4; i++){
274         const int offset= x_offset[i];
275         const int z0= temp[4*0+i] + temp[4*2+i];
276         const int z1= temp[4*0+i] - temp[4*2+i];
277         const int z2= temp[4*1+i] - temp[4*3+i];
278         const int z3= temp[4*1+i] + temp[4*3+i];
279
280         block[stride*0 +offset]= (z0 + z3)>>1;
281         block[stride*2 +offset]= (z1 + z2)>>1;
282         block[stride*8 +offset]= (z1 - z2)>>1;
283         block[stride*10+offset]= (z0 - z3)>>1;
284     }
285 }
286 #endif
287
288 #undef xStride
289 #undef stride
290
291 #if 0
292 static void chroma_dc_dct_c(DCTELEM *block){
293     const int stride= 16*2;
294     const int xStride= 16;
295     int a,b,c,d,e;
296
297     a= block[stride*0 + xStride*0];
298     b= block[stride*0 + xStride*1];
299     c= block[stride*1 + xStride*0];
300     d= block[stride*1 + xStride*1];
301
302     e= a-b;
303     a= a+b;
304     b= c-d;
305     c= c+d;
306
307     block[stride*0 + xStride*0]= (a+c);
308     block[stride*0 + xStride*1]= (e+b);
309     block[stride*1 + xStride*0]= (a-c);
310     block[stride*1 + xStride*1]= (e-b);
311 }
312 #endif
313
314 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
315                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
316                            int src_x_offset, int src_y_offset,
317                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
318                            int pixel_shift){
319     MpegEncContext * const s = &h->s;
320     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
321     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
322     const int luma_xy= (mx&3) + ((my&3)<<2);
323     uint8_t * src_y = pic->data[0] + ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
324     uint8_t * src_cb, * src_cr;
325     int extra_width= h->emu_edge_width;
326     int extra_height= h->emu_edge_height;
327     int emu=0;
328     const int full_mx= mx>>2;
329     const int full_my= my>>2;
330     const int pic_width  = 16*s->mb_width;
331     const int pic_height = 16*s->mb_height >> MB_FIELD;
332
333     if(mx&7) extra_width -= 3;
334     if(my&7) extra_height -= 3;
335
336     if(   full_mx < 0-extra_width
337        || full_my < 0-extra_height
338        || full_mx + 16/*FIXME*/ > pic_width + extra_width
339        || full_my + 16/*FIXME*/ > pic_height + extra_height){
340         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);
341             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
342         emu=1;
343     }
344
345     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
346     if(!square){
347         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
348     }
349
350     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
351
352     if(MB_FIELD){
353         // chroma offset when predicting from a field of opposite parity
354         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
355         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
356     }
357     src_cb= pic->data[1] + ((mx>>3) << pixel_shift) + (my>>3)*h->mb_uvlinesize;
358     src_cr= pic->data[2] + ((mx>>3) << pixel_shift) + (my>>3)*h->mb_uvlinesize;
359
360     if(emu){
361         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);
362             src_cb= s->edge_emu_buffer;
363     }
364     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
365
366     if(emu){
367         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);
368             src_cr= s->edge_emu_buffer;
369     }
370     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
371 }
372
373 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
374                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
375                            int x_offset, int y_offset,
376                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
377                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
378                            int list0, int list1, int pixel_shift){
379     MpegEncContext * const s = &h->s;
380     qpel_mc_func *qpix_op=  qpix_put;
381     h264_chroma_mc_func chroma_op= chroma_put;
382
383     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->  mb_linesize;
384     dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
385     dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
386     x_offset += 8*s->mb_x;
387     y_offset += 8*(s->mb_y >> MB_FIELD);
388
389     if(list0){
390         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
391         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
392                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
393                            qpix_op, chroma_op, pixel_shift);
394
395         qpix_op=  qpix_avg;
396         chroma_op= chroma_avg;
397     }
398
399     if(list1){
400         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
401         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
402                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
403                            qpix_op, chroma_op, pixel_shift);
404     }
405 }
406
407 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
408                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
409                            int x_offset, int y_offset,
410                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
411                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
412                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
413                            int list0, int list1, int pixel_shift){
414     MpegEncContext * const s = &h->s;
415
416     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->  mb_linesize;
417     dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
418     dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
419     x_offset += 8*s->mb_x;
420     y_offset += 8*(s->mb_y >> MB_FIELD);
421
422     if(list0 && list1){
423         /* don't optimize for luma-only case, since B-frames usually
424          * use implicit weights => chroma too. */
425         uint8_t *tmp_cb = s->obmc_scratchpad;
426         uint8_t *tmp_cr = s->obmc_scratchpad + (8 << pixel_shift);
427         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
428         int refn0 = h->ref_cache[0][ scan8[n] ];
429         int refn1 = h->ref_cache[1][ scan8[n] ];
430
431         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
432                     dest_y, dest_cb, dest_cr,
433                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift);
434         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
435                     tmp_y, tmp_cb, tmp_cr,
436                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift);
437
438         if(h->use_weight == 2){
439             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
440             int weight1 = 64 - weight0;
441             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
442             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
443             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
444         }else{
445             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
446                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
447                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
448             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
449                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
450                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
451             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
452                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
453                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
454         }
455     }else{
456         int list = list1 ? 1 : 0;
457         int refn = h->ref_cache[list][ scan8[n] ];
458         Picture *ref= &h->ref_list[list][refn];
459         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
460                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
461                     qpix_put, chroma_put, pixel_shift);
462
463         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
464                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
465         if(h->use_weight_chroma){
466             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
467                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
468             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
469                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
470         }
471     }
472 }
473
474 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
475                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
476                            int x_offset, int y_offset,
477                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
478                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
479                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
480                            int list0, int list1, int pixel_shift){
481     if((h->use_weight==2 && list0 && list1
482         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
483        || h->use_weight==1)
484         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
485                          x_offset, y_offset, qpix_put, chroma_put,
486                          weight_op[0], weight_op[3], weight_avg[0],
487                          weight_avg[3], list0, list1, pixel_shift);
488     else
489         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
490                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
491                     chroma_avg, list0, list1, pixel_shift);
492 }
493
494 static inline void prefetch_motion(H264Context *h, int list, int pixel_shift){
495     /* fetch pixels for estimated mv 4 macroblocks ahead
496      * optimized for 64byte cache lines */
497     MpegEncContext * const s = &h->s;
498     const int refn = h->ref_cache[list][scan8[0]];
499     if(refn >= 0){
500         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
501         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
502         uint8_t **src= h->ref_list[list][refn].data;
503         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
504         s->dsp.prefetch(src[0]+off, s->linesize, 4);
505         off= ((mx>>1) << pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + (64 << pixel_shift);
506         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
507     }
508 }
509
510 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
511                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
512                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
513                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
514                       int pixel_shift){
515     MpegEncContext * const s = &h->s;
516     const int mb_xy= h->mb_xy;
517     const int mb_type= s->current_picture.mb_type[mb_xy];
518
519     assert(IS_INTER(mb_type));
520
521     prefetch_motion(h, 0, pixel_shift);
522
523     if(IS_16X16(mb_type)){
524         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
525                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
526                 weight_op, weight_avg,
527                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
528                 pixel_shift);
529     }else if(IS_16X8(mb_type)){
530         mc_part(h, 0, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
531                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
532                 &weight_op[1], &weight_avg[1],
533                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
534                 pixel_shift);
535         mc_part(h, 8, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
536                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
537                 &weight_op[1], &weight_avg[1],
538                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
539                 pixel_shift);
540     }else if(IS_8X16(mb_type)){
541         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
542                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
543                 &weight_op[2], &weight_avg[2],
544                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
545                 pixel_shift);
546         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
547                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
548                 &weight_op[2], &weight_avg[2],
549                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
550                 pixel_shift);
551     }else{
552         int i;
553
554         assert(IS_8X8(mb_type));
555
556         for(i=0; i<4; i++){
557             const int sub_mb_type= h->sub_mb_type[i];
558             const int n= 4*i;
559             int x_offset= (i&1)<<2;
560             int y_offset= (i&2)<<1;
561
562             if(IS_SUB_8X8(sub_mb_type)){
563                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
564                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
565                     &weight_op[3], &weight_avg[3],
566                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
567                     pixel_shift);
568             }else if(IS_SUB_8X4(sub_mb_type)){
569                 mc_part(h, n  , 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
570                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
571                     &weight_op[4], &weight_avg[4],
572                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
573                     pixel_shift);
574                 mc_part(h, n+2, 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
575                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
576                     &weight_op[4], &weight_avg[4],
577                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
578                     pixel_shift);
579             }else if(IS_SUB_4X8(sub_mb_type)){
580                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
581                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
582                     &weight_op[5], &weight_avg[5],
583                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
584                     pixel_shift);
585                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
586                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
587                     &weight_op[5], &weight_avg[5],
588                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
589                     pixel_shift);
590             }else{
591                 int j;
592                 assert(IS_SUB_4X4(sub_mb_type));
593                 for(j=0; j<4; j++){
594                     int sub_x_offset= x_offset + 2*(j&1);
595                     int sub_y_offset= y_offset +   (j&2);
596                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
597                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
598                         &weight_op[6], &weight_avg[6],
599                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
600                         pixel_shift);
601                 }
602             }
603         }
604     }
605
606     prefetch_motion(h, 1, pixel_shift);
607 }
608
609 #define hl_motion_fn(sh, bits) \
610 static av_always_inline void hl_motion_ ## bits(H264Context *h, \
611                                        uint8_t *dest_y, \
612                                        uint8_t *dest_cb, uint8_t *dest_cr, \
613                                        qpel_mc_func (*qpix_put)[16], \
614                                        h264_chroma_mc_func (*chroma_put), \
615                                        qpel_mc_func (*qpix_avg)[16], \
616                                        h264_chroma_mc_func (*chroma_avg), \
617                                        h264_weight_func *weight_op, \
618                                        h264_biweight_func *weight_avg) \
619 { \
620     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put, \
621               qpix_avg, chroma_avg, weight_op, weight_avg, sh); \
622 }
623 hl_motion_fn(0, 8);
624 hl_motion_fn(1, 16);
625
626 static void free_tables(H264Context *h, int free_rbsp){
627     int i;
628     H264Context *hx;
629     av_freep(&h->intra4x4_pred_mode);
630     av_freep(&h->chroma_pred_mode_table);
631     av_freep(&h->cbp_table);
632     av_freep(&h->mvd_table[0]);
633     av_freep(&h->mvd_table[1]);
634     av_freep(&h->direct_table);
635     av_freep(&h->non_zero_count);
636     av_freep(&h->slice_table_base);
637     h->slice_table= NULL;
638     av_freep(&h->list_counts);
639
640     av_freep(&h->mb2b_xy);
641     av_freep(&h->mb2br_xy);
642
643     for(i = 0; i < MAX_THREADS; i++) {
644         hx = h->thread_context[i];
645         if(!hx) continue;
646         av_freep(&hx->top_borders[1]);
647         av_freep(&hx->top_borders[0]);
648         av_freep(&hx->s.obmc_scratchpad);
649         if (free_rbsp){
650             av_freep(&hx->rbsp_buffer[1]);
651             av_freep(&hx->rbsp_buffer[0]);
652             hx->rbsp_buffer_size[0] = 0;
653             hx->rbsp_buffer_size[1] = 0;
654         }
655         if (i) av_freep(&h->thread_context[i]);
656     }
657 }
658
659 static void init_dequant8_coeff_table(H264Context *h){
660     int i,q,x;
661     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
662     h->dequant8_coeff[0] = h->dequant8_buffer[0];
663     h->dequant8_coeff[1] = h->dequant8_buffer[1];
664
665     for(i=0; i<2; i++ ){
666         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
667             h->dequant8_coeff[1] = h->dequant8_buffer[0];
668             break;
669         }
670
671         for(q=0; q<max_qp+1; q++){
672             int shift = div6[q];
673             int idx = rem6[q];
674             for(x=0; x<64; x++)
675                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
676                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
677                     h->pps.scaling_matrix8[i][x]) << shift;
678         }
679     }
680 }
681
682 static void init_dequant4_coeff_table(H264Context *h){
683     int i,j,q,x;
684     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
685     for(i=0; i<6; i++ ){
686         h->dequant4_coeff[i] = h->dequant4_buffer[i];
687         for(j=0; j<i; j++){
688             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
689                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
690                 break;
691             }
692         }
693         if(j<i)
694             continue;
695
696         for(q=0; q<max_qp+1; q++){
697             int shift = div6[q] + 2;
698             int idx = rem6[q];
699             for(x=0; x<16; x++)
700                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
701                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
702                     h->pps.scaling_matrix4[i][x]) << shift;
703         }
704     }
705 }
706
707 static void init_dequant_tables(H264Context *h){
708     int i,x;
709     init_dequant4_coeff_table(h);
710     if(h->pps.transform_8x8_mode)
711         init_dequant8_coeff_table(h);
712     if(h->sps.transform_bypass){
713         for(i=0; i<6; i++)
714             for(x=0; x<16; x++)
715                 h->dequant4_coeff[i][0][x] = 1<<6;
716         if(h->pps.transform_8x8_mode)
717             for(i=0; i<2; i++)
718                 for(x=0; x<64; x++)
719                     h->dequant8_coeff[i][0][x] = 1<<6;
720     }
721 }
722
723
724 int ff_h264_alloc_tables(H264Context *h){
725     MpegEncContext * const s = &h->s;
726     const int big_mb_num= s->mb_stride * (s->mb_height+1);
727     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
728     int x,y;
729
730     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
731
732     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 32 * sizeof(uint8_t), fail)
733     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
734     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
735
736     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
737     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
738     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
739     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
740     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
741
742     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
743     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
744
745     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
746     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
747     for(y=0; y<s->mb_height; y++){
748         for(x=0; x<s->mb_width; x++){
749             const int mb_xy= x + y*s->mb_stride;
750             const int b_xy = 4*x + 4*y*h->b_stride;
751
752             h->mb2b_xy [mb_xy]= b_xy;
753             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
754         }
755     }
756
757     s->obmc_scratchpad = NULL;
758
759     if(!h->dequant4_coeff[0])
760         init_dequant_tables(h);
761
762     return 0;
763 fail:
764     free_tables(h, 1);
765     return -1;
766 }
767
768 /**
769  * Mimic alloc_tables(), but for every context thread.
770  */
771 static void clone_tables(H264Context *dst, H264Context *src, int i){
772     MpegEncContext * const s = &src->s;
773     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
774     dst->non_zero_count           = src->non_zero_count;
775     dst->slice_table              = src->slice_table;
776     dst->cbp_table                = src->cbp_table;
777     dst->mb2b_xy                  = src->mb2b_xy;
778     dst->mb2br_xy                 = src->mb2br_xy;
779     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
780     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
781     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
782     dst->direct_table             = src->direct_table;
783     dst->list_counts              = src->list_counts;
784
785     dst->s.obmc_scratchpad = NULL;
786     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma);
787 }
788
789 /**
790  * Init context
791  * Allocate buffers which are not shared amongst multiple threads.
792  */
793 static int context_init(H264Context *h){
794     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t)*2, fail)
795     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t)*2, fail)
796
797     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
798     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
799
800     return 0;
801 fail:
802     return -1; // free_tables will clean up for us
803 }
804
805 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
806
807 static av_cold void common_init(H264Context *h){
808     MpegEncContext * const s = &h->s;
809
810     s->width = s->avctx->width;
811     s->height = s->avctx->height;
812     s->codec_id= s->avctx->codec->id;
813
814     ff_h264dsp_init(&h->h264dsp, 8);
815     ff_h264_pred_init(&h->hpc, s->codec_id, 8);
816
817     h->dequant_coeff_pps= -1;
818     s->unrestricted_mv=1;
819     s->decode=1; //FIXME
820
821     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
822
823     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
824     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
825 }
826
827 int ff_h264_decode_extradata(H264Context *h)
828 {
829     AVCodecContext *avctx = h->s.avctx;
830
831     if(*(char *)avctx->extradata == 1){
832         int i, cnt, nalsize;
833         unsigned char *p = avctx->extradata;
834
835         h->is_avc = 1;
836
837         if(avctx->extradata_size < 7) {
838             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
839             return -1;
840         }
841         /* sps and pps in the avcC always have length coded with 2 bytes,
842            so put a fake nal_length_size = 2 while parsing them */
843         h->nal_length_size = 2;
844         // Decode sps from avcC
845         cnt = *(p+5) & 0x1f; // Number of sps
846         p += 6;
847         for (i = 0; i < cnt; i++) {
848             nalsize = AV_RB16(p) + 2;
849             if(decode_nal_units(h, p, nalsize) < 0) {
850                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
851                 return -1;
852             }
853             p += nalsize;
854         }
855         // Decode pps from avcC
856         cnt = *(p++); // Number of pps
857         for (i = 0; i < cnt; i++) {
858             nalsize = AV_RB16(p) + 2;
859             if(decode_nal_units(h, p, nalsize)  != nalsize) {
860                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
861                 return -1;
862             }
863             p += nalsize;
864         }
865         // Now store right nal length size, that will be use to parse all other nals
866         h->nal_length_size = ((*(((char*)(avctx->extradata))+4))&0x03)+1;
867     } else {
868         h->is_avc = 0;
869         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
870             return -1;
871     }
872     return 0;
873 }
874
875 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
876     H264Context *h= avctx->priv_data;
877     MpegEncContext * const s = &h->s;
878
879     MPV_decode_defaults(s);
880
881     s->avctx = avctx;
882     common_init(h);
883
884     s->out_format = FMT_H264;
885     s->workaround_bugs= avctx->workaround_bugs;
886
887     // set defaults
888 //    s->decode_mb= ff_h263_decode_mb;
889     s->quarter_sample = 1;
890     if(!avctx->has_b_frames)
891     s->low_delay= 1;
892
893     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
894
895     ff_h264_decode_init_vlc();
896
897     h->pixel_shift = 0;
898     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
899
900     h->thread_context[0] = h;
901     h->outputed_poc = INT_MIN;
902     h->prev_poc_msb= 1<<16;
903     h->x264_build = -1;
904     ff_h264_reset_sei(h);
905     if(avctx->codec_id == CODEC_ID_H264){
906         if(avctx->ticks_per_frame == 1){
907             s->avctx->time_base.den *=2;
908         }
909         avctx->ticks_per_frame = 2;
910     }
911
912     if(avctx->extradata_size > 0 && avctx->extradata &&
913         ff_h264_decode_extradata(h))
914         return -1;
915
916     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
917         s->avctx->has_b_frames = h->sps.num_reorder_frames;
918         s->low_delay = 0;
919     }
920
921     return 0;
922 }
923
924 int ff_h264_frame_start(H264Context *h){
925     MpegEncContext * const s = &h->s;
926     int i;
927     const int pixel_shift = h->pixel_shift;
928
929     if(MPV_frame_start(s, s->avctx) < 0)
930         return -1;
931     ff_er_frame_start(s);
932     /*
933      * MPV_frame_start uses pict_type to derive key_frame.
934      * This is incorrect for H.264; IDR markings must be used.
935      * Zero here; IDR markings per slice in frame or fields are ORed in later.
936      * See decode_nal_units().
937      */
938     s->current_picture_ptr->key_frame= 0;
939     s->current_picture_ptr->mmco_reset= 0;
940
941     assert(s->linesize && s->uvlinesize);
942
943     for(i=0; i<16; i++){
944         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
945         h->block_offset[24+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
946     }
947     for(i=0; i<4; i++){
948         h->block_offset[16+i]=
949         h->block_offset[20+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
950         h->block_offset[24+16+i]=
951         h->block_offset[24+20+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
952     }
953
954     /* can't be in alloc_tables because linesize isn't known there.
955      * FIXME: redo bipred weight to not require extra buffer? */
956     for(i = 0; i < s->avctx->thread_count; i++)
957         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
958             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
959
960     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
961     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
962
963 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
964
965     // We mark the current picture as non-reference after allocating it, so
966     // that if we break out due to an error it can be released automatically
967     // in the next MPV_frame_start().
968     // SVQ3 as well as most other codecs have only last/next/current and thus
969     // get released even with set reference, besides SVQ3 and others do not
970     // mark frames as reference later "naturally".
971     if(s->codec_id != CODEC_ID_SVQ3)
972         s->current_picture_ptr->reference= 0;
973
974     s->current_picture_ptr->field_poc[0]=
975     s->current_picture_ptr->field_poc[1]= INT_MAX;
976     assert(s->current_picture_ptr->long_ref==0);
977
978     return 0;
979 }
980
981 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 simple){
982     MpegEncContext * const s = &h->s;
983     uint8_t *top_border;
984     int top_idx = 1;
985     const int pixel_shift = h->pixel_shift;
986
987     src_y  -=   linesize;
988     src_cb -= uvlinesize;
989     src_cr -= uvlinesize;
990
991     if(!simple && FRAME_MBAFF){
992         if(s->mb_y&1){
993             if(!MB_MBAFF){
994                 top_border = h->top_borders[0][s->mb_x];
995                 AV_COPY128(top_border, src_y + 15*linesize);
996                 if (pixel_shift)
997                     AV_COPY128(top_border+16, src_y+15*linesize+16);
998                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
999                     if (pixel_shift) {
1000                         AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1001                         AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1002                     } else {
1003                     AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1004                     AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1005                     }
1006                 }
1007             }
1008         }else if(MB_MBAFF){
1009             top_idx = 0;
1010         }else
1011             return;
1012     }
1013
1014     top_border = h->top_borders[top_idx][s->mb_x];
1015     // There are two lines saved, the line above the the top macroblock of a pair,
1016     // and the line above the bottom macroblock
1017     AV_COPY128(top_border, src_y + 16*linesize);
1018     if (pixel_shift)
1019         AV_COPY128(top_border+16, src_y+16*linesize+16);
1020
1021     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1022         if (pixel_shift) {
1023             AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1024             AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1025         } else {
1026         AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1027         AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1028         }
1029     }
1030 }
1031
1032 static inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1033                                   uint8_t *src_cb, uint8_t *src_cr,
1034                                   int linesize, int uvlinesize,
1035                                   int xchg, int simple, int pixel_shift){
1036     MpegEncContext * const s = &h->s;
1037     int deblock_left;
1038     int deblock_top;
1039     int top_idx = 1;
1040     uint8_t *top_border_m1;
1041     uint8_t *top_border;
1042
1043     if(!simple && FRAME_MBAFF){
1044         if(s->mb_y&1){
1045             if(!MB_MBAFF)
1046                 return;
1047         }else{
1048             top_idx = MB_MBAFF ? 0 : 1;
1049         }
1050     }
1051
1052     if(h->deblocking_filter == 2) {
1053         deblock_left = h->left_type[0];
1054         deblock_top  = h->top_type;
1055     } else {
1056         deblock_left = (s->mb_x > 0);
1057         deblock_top =  (s->mb_y > !!MB_FIELD);
1058     }
1059
1060     src_y  -=   linesize + 1 + pixel_shift;
1061     src_cb -= uvlinesize + 1 + pixel_shift;
1062     src_cr -= uvlinesize + 1 + pixel_shift;
1063
1064     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1065     top_border    = h->top_borders[top_idx][s->mb_x];
1066
1067 #define XCHG(a,b,xchg)\
1068     if (pixel_shift) {\
1069         if (xchg) {\
1070             AV_SWAP64(b+0,a+0);\
1071             AV_SWAP64(b+8,a+8);\
1072         } else {\
1073             AV_COPY128(b,a); \
1074         }\
1075     } else \
1076 if (xchg) AV_SWAP64(b,a);\
1077 else      AV_COPY64(b,a);
1078
1079     if(deblock_top){
1080         if(deblock_left){
1081             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1082         }
1083         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1084         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1085         if(s->mb_x+1 < s->mb_width){
1086             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1087         }
1088     }
1089     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1090         if(deblock_top){
1091             if(deblock_left){
1092                 XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1093                 XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1094             }
1095             XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1096             XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1097         }
1098     }
1099 }
1100
1101 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1102     if (high_bit_depth) {
1103         return AV_RN32A(((int32_t*)mb) + index);
1104     } else
1105         return AV_RN16A(mb + index);
1106 }
1107
1108 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1109     if (high_bit_depth) {
1110         AV_WN32A(((int32_t*)mb) + index, value);
1111     } else
1112         AV_WN16A(mb + index, value);
1113 }
1114
1115 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1116     MpegEncContext * const s = &h->s;
1117     const int mb_x= s->mb_x;
1118     const int mb_y= s->mb_y;
1119     const int mb_xy= h->mb_xy;
1120     const int mb_type= s->current_picture.mb_type[mb_xy];
1121     uint8_t  *dest_y, *dest_cb, *dest_cr;
1122     int linesize, uvlinesize /*dct_offset*/;
1123     int i;
1124     int *block_offset = &h->block_offset[0];
1125     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1126     /* is_h264 should always be true if SVQ3 is disabled. */
1127     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1128     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1129     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1130
1131     dest_y  = s->current_picture.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1132     dest_cb = s->current_picture.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
1133     dest_cr = s->current_picture.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
1134
1135     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1136     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1137
1138     h->list_counts[mb_xy]= h->list_count;
1139
1140     if (!simple && MB_FIELD) {
1141         linesize   = h->mb_linesize   = s->linesize * 2;
1142         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1143         block_offset = &h->block_offset[24];
1144         if(mb_y&1){ //FIXME move out of this function?
1145             dest_y -= s->linesize*15;
1146             dest_cb-= s->uvlinesize*7;
1147             dest_cr-= s->uvlinesize*7;
1148         }
1149         if(FRAME_MBAFF) {
1150             int list;
1151             for(list=0; list<h->list_count; list++){
1152                 if(!USES_LIST(mb_type, list))
1153                     continue;
1154                 if(IS_16X16(mb_type)){
1155                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1156                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1157                 }else{
1158                     for(i=0; i<16; i+=4){
1159                         int ref = h->ref_cache[list][scan8[i]];
1160                         if(ref >= 0)
1161                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1162                     }
1163                 }
1164             }
1165         }
1166     } else {
1167         linesize   = h->mb_linesize   = s->linesize;
1168         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1169 //        dct_offset = s->linesize * 16;
1170     }
1171
1172     if (!simple && IS_INTRA_PCM(mb_type)) {
1173         if (pixel_shift) {
1174             const int bit_depth = h->sps.bit_depth_luma;
1175             int j;
1176             GetBitContext gb;
1177             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1178
1179             for (i = 0; i < 16; i++) {
1180                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1181                 for (j = 0; j < 16; j++)
1182                     tmp_y[j] = get_bits(&gb, bit_depth);
1183             }
1184             for (i = 0; i < 8; i++) {
1185                 uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1186                 for (j = 0; j < 8; j++)
1187                     tmp_cb[j] = get_bits(&gb, bit_depth);
1188             }
1189             for (i = 0; i < 8; i++) {
1190                 uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1191                 for (j = 0; j < 8; j++)
1192                     tmp_cr[j] = get_bits(&gb, bit_depth);
1193             }
1194         } else {
1195         for (i=0; i<16; i++) {
1196             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1197         }
1198         for (i=0; i<8; i++) {
1199             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
1200             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
1201         }
1202         }
1203     } else {
1204         if(IS_INTRA(mb_type)){
1205             if(h->deblocking_filter)
1206                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple, pixel_shift);
1207
1208             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1209                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1210                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1211             }
1212
1213             if(IS_INTRA4x4(mb_type)){
1214                 if(simple || !s->encoding){
1215                     if(IS_8x8DCT(mb_type)){
1216                         if(transform_bypass){
1217                             idct_dc_add =
1218                             idct_add    = s->dsp.add_pixels8;
1219                         }else{
1220                             idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1221                             idct_add    = h->h264dsp.h264_idct8_add;
1222                         }
1223                         for(i=0; i<16; i+=4){
1224                             uint8_t * const ptr= dest_y + block_offset[i];
1225                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1226                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1227                                 h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16 << pixel_shift), linesize);
1228                             }else{
1229                                 const int nnz = h->non_zero_count_cache[ scan8[i] ];
1230                                 h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1231                                                             (h->topright_samples_available<<i)&0x4000, linesize);
1232                                 if(nnz){
1233                                     if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16))
1234                                         idct_dc_add(ptr, h->mb + (i*16 << pixel_shift), linesize);
1235                                     else
1236                                         idct_add   (ptr, h->mb + (i*16 << pixel_shift), linesize);
1237                                 }
1238                             }
1239                         }
1240                     }else{
1241                         if(transform_bypass){
1242                             idct_dc_add =
1243                             idct_add    = s->dsp.add_pixels4;
1244                         }else{
1245                             idct_dc_add = h->h264dsp.h264_idct_dc_add;
1246                             idct_add    = h->h264dsp.h264_idct_add;
1247                         }
1248                         for(i=0; i<16; i++){
1249                             uint8_t * const ptr= dest_y + block_offset[i];
1250                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1251
1252                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1253                                 h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16 << pixel_shift), linesize);
1254                             }else{
1255                                 uint8_t *topright;
1256                                 int nnz, tr;
1257                                 uint64_t tr_high;
1258                                 if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1259                                     const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1260                                     assert(mb_y || linesize <= block_offset[i]);
1261                                     if(!topright_avail){
1262                                         if (pixel_shift) {
1263                                             tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1264                                             topright= (uint8_t*) &tr_high;
1265                                         } else {
1266                                         tr= ptr[3 - linesize]*0x01010101;
1267                                         topright= (uint8_t*) &tr;
1268                                         }
1269                                     }else
1270                                         topright= ptr + (4 << pixel_shift) - linesize;
1271                                 }else
1272                                     topright= NULL;
1273
1274                                 h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1275                                 nnz = h->non_zero_count_cache[ scan8[i] ];
1276                                 if(nnz){
1277                                     if(is_h264){
1278                                         if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16))
1279                                             idct_dc_add(ptr, h->mb + (i*16 << pixel_shift), linesize);
1280                                         else
1281                                             idct_add   (ptr, h->mb + (i*16 << pixel_shift), linesize);
1282                                     }else
1283                                         ff_svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
1284                                 }
1285                             }
1286                         }
1287                     }
1288                 }
1289             }else{
1290                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1291                 if(is_h264){
1292                     if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX] ]){
1293                         if(!transform_bypass)
1294                             h->h264dsp.h264_luma_dc_dequant_idct(h->mb, h->mb_luma_dc, h->dequant4_coeff[0][s->qscale][0]);
1295                         else{
1296                             static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1297                                                                     8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1298                             for(i = 0; i < 16; i++)
1299                                 dctcoef_set(h->mb, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc, pixel_shift, i));
1300                         }
1301                     }
1302                 }else
1303                     ff_svq3_luma_dc_dequant_idct_c(h->mb, h->mb_luma_dc, s->qscale);
1304             }
1305             if(h->deblocking_filter)
1306                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple, pixel_shift);
1307         }else if(is_h264){
1308             if (pixel_shift) {
1309                 hl_motion_16(h, dest_y, dest_cb, dest_cr,
1310                              s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1311                              s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1312                              h->h264dsp.weight_h264_pixels_tab,
1313                              h->h264dsp.biweight_h264_pixels_tab);
1314             } else
1315                 hl_motion_8(h, dest_y, dest_cb, dest_cr,
1316                             s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1317                             s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1318                             h->h264dsp.weight_h264_pixels_tab,
1319                             h->h264dsp.biweight_h264_pixels_tab);
1320         }
1321
1322
1323         if(!IS_INTRA4x4(mb_type)){
1324             if(is_h264){
1325                 if(IS_INTRA16x16(mb_type)){
1326                     if(transform_bypass){
1327                         if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1328                             h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb, linesize);
1329                         }else{
1330                             for(i=0; i<16; i++){
1331                                 if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1332                                     s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16 << pixel_shift), linesize);
1333                             }
1334                         }
1335                     }else{
1336                          h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
1337                     }
1338                 }else if(h->cbp&15){
1339                     if(transform_bypass){
1340                         const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1341                         idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1342                         for(i=0; i<16; i+=di){
1343                             if(h->non_zero_count_cache[ scan8[i] ]){
1344                                 idct_add(dest_y + block_offset[i], h->mb + (i*16 << pixel_shift), linesize);
1345                             }
1346                         }
1347                     }else{
1348                         if(IS_8x8DCT(mb_type)){
1349                             h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
1350                         }else{
1351                             h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
1352                         }
1353                     }
1354                 }
1355             }else{
1356                 for(i=0; i<16; i++){
1357                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
1358                         uint8_t * const ptr= dest_y + block_offset[i];
1359                         ff_svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1360                     }
1361                 }
1362             }
1363         }
1364
1365         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1366             uint8_t *dest[2] = {dest_cb, dest_cr};
1367             if(transform_bypass){
1368                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1369                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16 << pixel_shift), uvlinesize);
1370                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 20, h->mb + (20*16 << pixel_shift), uvlinesize);
1371                 }else{
1372                     idct_add = s->dsp.add_pixels4;
1373                     for(i=16; i<16+8; i++){
1374                         if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1375                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1376                     }
1377                 }
1378             }else{
1379                 if(is_h264){
1380                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1381                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16 << pixel_shift)       , h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1382                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1383                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + ((16*16+4*16) << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
1384                     h->h264dsp.h264_idct_add8(dest, block_offset,
1385                                               h->mb, uvlinesize,
1386                                               h->non_zero_count_cache);
1387                 }else{
1388                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16*16     , h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1389                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16*16+4*16, h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
1390                     for(i=16; i<16+8; i++){
1391                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1392                             uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
1393                             ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1394                         }
1395                     }
1396                 }
1397             }
1398         }
1399     }
1400     if(h->cbp || IS_INTRA(mb_type))
1401         s->dsp.clear_blocks(h->mb);
1402 }
1403
1404 /**
1405  * Process a macroblock; this case avoids checks for expensive uncommon cases.
1406  */
1407 #define hl_decode_mb_simple(sh, bits) \
1408 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
1409     hl_decode_mb_internal(h, 1, sh); \
1410 }
1411 hl_decode_mb_simple(0, 8);
1412 hl_decode_mb_simple(1, 16);
1413
1414 /**
1415  * Process a macroblock; this handles edge cases, such as interlacing.
1416  */
1417 static void av_noinline hl_decode_mb_complex(H264Context *h){
1418     hl_decode_mb_internal(h, 0, h->pixel_shift);
1419 }
1420
1421 void ff_h264_hl_decode_mb(H264Context *h){
1422     MpegEncContext * const s = &h->s;
1423     const int mb_xy= h->mb_xy;
1424     const int mb_type= s->current_picture.mb_type[mb_xy];
1425     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
1426
1427     if (is_complex) {
1428         hl_decode_mb_complex(h);
1429     } else if (h->pixel_shift) {
1430         hl_decode_mb_simple_16(h);
1431     } else
1432         hl_decode_mb_simple_8(h);
1433 }
1434
1435 static int pred_weight_table(H264Context *h){
1436     MpegEncContext * const s = &h->s;
1437     int list, i;
1438     int luma_def, chroma_def;
1439
1440     h->use_weight= 0;
1441     h->use_weight_chroma= 0;
1442     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
1443     if(CHROMA)
1444         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
1445     luma_def = 1<<h->luma_log2_weight_denom;
1446     chroma_def = 1<<h->chroma_log2_weight_denom;
1447
1448     for(list=0; list<2; list++){
1449         h->luma_weight_flag[list]   = 0;
1450         h->chroma_weight_flag[list] = 0;
1451         for(i=0; i<h->ref_count[list]; i++){
1452             int luma_weight_flag, chroma_weight_flag;
1453
1454             luma_weight_flag= get_bits1(&s->gb);
1455             if(luma_weight_flag){
1456                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
1457                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
1458                 if(   h->luma_weight[i][list][0] != luma_def
1459                    || h->luma_weight[i][list][1] != 0) {
1460                     h->use_weight= 1;
1461                     h->luma_weight_flag[list]= 1;
1462                 }
1463             }else{
1464                 h->luma_weight[i][list][0]= luma_def;
1465                 h->luma_weight[i][list][1]= 0;
1466             }
1467
1468             if(CHROMA){
1469                 chroma_weight_flag= get_bits1(&s->gb);
1470                 if(chroma_weight_flag){
1471                     int j;
1472                     for(j=0; j<2; j++){
1473                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
1474                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
1475                         if(   h->chroma_weight[i][list][j][0] != chroma_def
1476                            || h->chroma_weight[i][list][j][1] != 0) {
1477                             h->use_weight_chroma= 1;
1478                             h->chroma_weight_flag[list]= 1;
1479                         }
1480                     }
1481                 }else{
1482                     int j;
1483                     for(j=0; j<2; j++){
1484                         h->chroma_weight[i][list][j][0]= chroma_def;
1485                         h->chroma_weight[i][list][j][1]= 0;
1486                     }
1487                 }
1488             }
1489         }
1490         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
1491     }
1492     h->use_weight= h->use_weight || h->use_weight_chroma;
1493     return 0;
1494 }
1495
1496 /**
1497  * Initialize implicit_weight table.
1498  * @param field  0/1 initialize the weight for interlaced MBAFF
1499  *                -1 initializes the rest
1500  */
1501 static void implicit_weight_table(H264Context *h, int field){
1502     MpegEncContext * const s = &h->s;
1503     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
1504
1505     for (i = 0; i < 2; i++) {
1506         h->luma_weight_flag[i]   = 0;
1507         h->chroma_weight_flag[i] = 0;
1508     }
1509
1510     if(field < 0){
1511         cur_poc = s->current_picture_ptr->poc;
1512     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
1513        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
1514         h->use_weight= 0;
1515         h->use_weight_chroma= 0;
1516         return;
1517     }
1518         ref_start= 0;
1519         ref_count0= h->ref_count[0];
1520         ref_count1= h->ref_count[1];
1521     }else{
1522         cur_poc = s->current_picture_ptr->field_poc[field];
1523         ref_start= 16;
1524         ref_count0= 16+2*h->ref_count[0];
1525         ref_count1= 16+2*h->ref_count[1];
1526     }
1527
1528     h->use_weight= 2;
1529     h->use_weight_chroma= 2;
1530     h->luma_log2_weight_denom= 5;
1531     h->chroma_log2_weight_denom= 5;
1532
1533     for(ref0=ref_start; ref0 < ref_count0; ref0++){
1534         int poc0 = h->ref_list[0][ref0].poc;
1535         for(ref1=ref_start; ref1 < ref_count1; ref1++){
1536             int poc1 = h->ref_list[1][ref1].poc;
1537             int td = av_clip(poc1 - poc0, -128, 127);
1538             int w= 32;
1539             if(td){
1540                 int tb = av_clip(cur_poc - poc0, -128, 127);
1541                 int tx = (16384 + (FFABS(td) >> 1)) / td;
1542                 int dist_scale_factor = (tb*tx + 32) >> 8;
1543                 if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
1544                     w = 64 - dist_scale_factor;
1545             }
1546             if(field<0){
1547                 h->implicit_weight[ref0][ref1][0]=
1548                 h->implicit_weight[ref0][ref1][1]= w;
1549             }else{
1550                 h->implicit_weight[ref0][ref1][field]=w;
1551             }
1552         }
1553     }
1554 }
1555
1556 /**
1557  * instantaneous decoder refresh.
1558  */
1559 static void idr(H264Context *h){
1560     ff_h264_remove_all_refs(h);
1561     h->prev_frame_num= 0;
1562     h->prev_frame_num_offset= 0;
1563     h->prev_poc_msb=
1564     h->prev_poc_lsb= 0;
1565 }
1566
1567 /* forget old pics after a seek */
1568 static void flush_dpb(AVCodecContext *avctx){
1569     H264Context *h= avctx->priv_data;
1570     int i;
1571     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
1572         if(h->delayed_pic[i])
1573             h->delayed_pic[i]->reference= 0;
1574         h->delayed_pic[i]= NULL;
1575     }
1576     h->outputed_poc= INT_MIN;
1577     h->prev_interlaced_frame = 1;
1578     idr(h);
1579     if(h->s.current_picture_ptr)
1580         h->s.current_picture_ptr->reference= 0;
1581     h->s.first_field= 0;
1582     ff_h264_reset_sei(h);
1583     ff_mpeg_flush(avctx);
1584 }
1585
1586 static int init_poc(H264Context *h){
1587     MpegEncContext * const s = &h->s;
1588     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
1589     int field_poc[2];
1590     Picture *cur = s->current_picture_ptr;
1591
1592     h->frame_num_offset= h->prev_frame_num_offset;
1593     if(h->frame_num < h->prev_frame_num)
1594         h->frame_num_offset += max_frame_num;
1595
1596     if(h->sps.poc_type==0){
1597         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
1598
1599         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
1600             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
1601         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
1602             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
1603         else
1604             h->poc_msb = h->prev_poc_msb;
1605 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
1606         field_poc[0] =
1607         field_poc[1] = h->poc_msb + h->poc_lsb;
1608         if(s->picture_structure == PICT_FRAME)
1609             field_poc[1] += h->delta_poc_bottom;
1610     }else if(h->sps.poc_type==1){
1611         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
1612         int i;
1613
1614         if(h->sps.poc_cycle_length != 0)
1615             abs_frame_num = h->frame_num_offset + h->frame_num;
1616         else
1617             abs_frame_num = 0;
1618
1619         if(h->nal_ref_idc==0 && abs_frame_num > 0)
1620             abs_frame_num--;
1621
1622         expected_delta_per_poc_cycle = 0;
1623         for(i=0; i < h->sps.poc_cycle_length; i++)
1624             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
1625
1626         if(abs_frame_num > 0){
1627             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
1628             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
1629
1630             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
1631             for(i = 0; i <= frame_num_in_poc_cycle; i++)
1632                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
1633         } else
1634             expectedpoc = 0;
1635
1636         if(h->nal_ref_idc == 0)
1637             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
1638
1639         field_poc[0] = expectedpoc + h->delta_poc[0];
1640         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
1641
1642         if(s->picture_structure == PICT_FRAME)
1643             field_poc[1] += h->delta_poc[1];
1644     }else{
1645         int poc= 2*(h->frame_num_offset + h->frame_num);
1646
1647         if(!h->nal_ref_idc)
1648             poc--;
1649
1650         field_poc[0]= poc;
1651         field_poc[1]= poc;
1652     }
1653
1654     if(s->picture_structure != PICT_BOTTOM_FIELD)
1655         s->current_picture_ptr->field_poc[0]= field_poc[0];
1656     if(s->picture_structure != PICT_TOP_FIELD)
1657         s->current_picture_ptr->field_poc[1]= field_poc[1];
1658     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
1659
1660     return 0;
1661 }
1662
1663
1664 /**
1665  * initialize scan tables
1666  */
1667 static void init_scan_tables(H264Context *h){
1668     int i;
1669     for(i=0; i<16; i++){
1670 #define T(x) (x>>2) | ((x<<2) & 0xF)
1671         h->zigzag_scan[i] = T(zigzag_scan[i]);
1672         h-> field_scan[i] = T( field_scan[i]);
1673 #undef T
1674     }
1675     for(i=0; i<64; i++){
1676 #define T(x) (x>>3) | ((x&7)<<3)
1677         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
1678         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
1679         h->field_scan8x8[i]        = T(field_scan8x8[i]);
1680         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
1681 #undef T
1682     }
1683     if(h->sps.transform_bypass){ //FIXME same ugly
1684         h->zigzag_scan_q0          = zigzag_scan;
1685         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
1686         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
1687         h->field_scan_q0           = field_scan;
1688         h->field_scan8x8_q0        = field_scan8x8;
1689         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
1690     }else{
1691         h->zigzag_scan_q0          = h->zigzag_scan;
1692         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
1693         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
1694         h->field_scan_q0           = h->field_scan;
1695         h->field_scan8x8_q0        = h->field_scan8x8;
1696         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
1697     }
1698 }
1699
1700 static void field_end(H264Context *h){
1701     MpegEncContext * const s = &h->s;
1702     AVCodecContext * const avctx= s->avctx;
1703     s->mb_y= 0;
1704
1705     s->current_picture_ptr->qscale_type= FF_QSCALE_TYPE_H264;
1706     s->current_picture_ptr->pict_type= s->pict_type;
1707
1708     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
1709         ff_vdpau_h264_set_reference_frames(s);
1710
1711     if(!s->dropable) {
1712         ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1713         h->prev_poc_msb= h->poc_msb;
1714         h->prev_poc_lsb= h->poc_lsb;
1715     }
1716     h->prev_frame_num_offset= h->frame_num_offset;
1717     h->prev_frame_num= h->frame_num;
1718
1719     if (avctx->hwaccel) {
1720         if (avctx->hwaccel->end_frame(avctx) < 0)
1721             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
1722     }
1723
1724     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
1725         ff_vdpau_h264_picture_complete(s);
1726
1727     /*
1728      * FIXME: Error handling code does not seem to support interlaced
1729      * when slices span multiple rows
1730      * The ff_er_add_slice calls don't work right for bottom
1731      * fields; they cause massive erroneous error concealing
1732      * Error marking covers both fields (top and bottom).
1733      * This causes a mismatched s->error_count
1734      * and a bad error table. Further, the error count goes to
1735      * INT_MAX when called for bottom field, because mb_y is
1736      * past end by one (callers fault) and resync_mb_y != 0
1737      * causes problems for the first MB line, too.
1738      */
1739     if (!FIELD_PICTURE)
1740         ff_er_frame_end(s);
1741
1742     MPV_frame_end(s);
1743
1744     h->current_slice=0;
1745 }
1746
1747 /**
1748  * Replicate H264 "master" context to thread contexts.
1749  */
1750 static void clone_slice(H264Context *dst, H264Context *src)
1751 {
1752     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
1753     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
1754     dst->s.current_picture      = src->s.current_picture;
1755     dst->s.linesize             = src->s.linesize;
1756     dst->s.uvlinesize           = src->s.uvlinesize;
1757     dst->s.first_field          = src->s.first_field;
1758
1759     dst->prev_poc_msb           = src->prev_poc_msb;
1760     dst->prev_poc_lsb           = src->prev_poc_lsb;
1761     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
1762     dst->prev_frame_num         = src->prev_frame_num;
1763     dst->short_ref_count        = src->short_ref_count;
1764
1765     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
1766     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
1767     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
1768     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
1769
1770     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
1771     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
1772 }
1773
1774 /**
1775  * computes profile from profile_idc and constraint_set?_flags
1776  *
1777  * @param sps SPS
1778  *
1779  * @return profile as defined by FF_PROFILE_H264_*
1780  */
1781 int ff_h264_get_profile(SPS *sps)
1782 {
1783     int profile = sps->profile_idc;
1784
1785     switch(sps->profile_idc) {
1786     case FF_PROFILE_H264_BASELINE:
1787         // constraint_set1_flag set to 1
1788         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
1789         break;
1790     case FF_PROFILE_H264_HIGH_10:
1791     case FF_PROFILE_H264_HIGH_422:
1792     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
1793         // constraint_set3_flag set to 1
1794         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
1795         break;
1796     }
1797
1798     return profile;
1799 }
1800
1801 /**
1802  * decodes a slice header.
1803  * This will also call MPV_common_init() and frame_start() as needed.
1804  *
1805  * @param h h264context
1806  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
1807  *
1808  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
1809  */
1810 static int decode_slice_header(H264Context *h, H264Context *h0){
1811     MpegEncContext * const s = &h->s;
1812     MpegEncContext * const s0 = &h0->s;
1813     unsigned int first_mb_in_slice;
1814     unsigned int pps_id;
1815     int num_ref_idx_active_override_flag;
1816     unsigned int slice_type, tmp, i, j;
1817     int default_ref_list_done = 0;
1818     int last_pic_structure;
1819
1820     s->dropable= h->nal_ref_idc == 0;
1821
1822     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc){
1823         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
1824         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
1825     }else{
1826         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
1827         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
1828     }
1829
1830     first_mb_in_slice= get_ue_golomb(&s->gb);
1831
1832     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
1833         if(h0->current_slice && FIELD_PICTURE){
1834             field_end(h);
1835         }
1836
1837         h0->current_slice = 0;
1838         if (!s0->first_field)
1839             s->current_picture_ptr= NULL;
1840     }
1841
1842     slice_type= get_ue_golomb_31(&s->gb);
1843     if(slice_type > 9){
1844         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);
1845         return -1;
1846     }
1847     if(slice_type > 4){
1848         slice_type -= 5;
1849         h->slice_type_fixed=1;
1850     }else
1851         h->slice_type_fixed=0;
1852
1853     slice_type= golomb_to_pict_type[ slice_type ];
1854     if (slice_type == AV_PICTURE_TYPE_I
1855         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
1856         default_ref_list_done = 1;
1857     }
1858     h->slice_type= slice_type;
1859     h->slice_type_nos= slice_type & 3;
1860
1861     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
1862
1863     pps_id= get_ue_golomb(&s->gb);
1864     if(pps_id>=MAX_PPS_COUNT){
1865         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
1866         return -1;
1867     }
1868     if(!h0->pps_buffers[pps_id]) {
1869         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
1870         return -1;
1871     }
1872     h->pps= *h0->pps_buffers[pps_id];
1873
1874     if(!h0->sps_buffers[h->pps.sps_id]) {
1875         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
1876         return -1;
1877     }
1878     h->sps = *h0->sps_buffers[h->pps.sps_id];
1879
1880     s->avctx->profile = ff_h264_get_profile(&h->sps);
1881     s->avctx->level   = h->sps.level_idc;
1882     s->avctx->refs    = h->sps.ref_frame_count;
1883
1884     if(h == h0 && h->dequant_coeff_pps != pps_id){
1885         h->dequant_coeff_pps = pps_id;
1886         init_dequant_tables(h);
1887     }
1888
1889     s->mb_width= h->sps.mb_width;
1890     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
1891
1892     h->b_stride=  s->mb_width*4;
1893
1894     s->width = 16*s->mb_width - 2*FFMIN(h->sps.crop_right, 7);
1895     if(h->sps.frame_mbs_only_flag)
1896         s->height= 16*s->mb_height - 2*FFMIN(h->sps.crop_bottom, 7);
1897     else
1898         s->height= 16*s->mb_height - 4*FFMIN(h->sps.crop_bottom, 7);
1899
1900     if (s->context_initialized
1901         && (   s->width != s->avctx->width || s->height != s->avctx->height
1902             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
1903         if(h != h0)
1904             return -1;   // width / height changed during parallelized decoding
1905         free_tables(h, 0);
1906         flush_dpb(s->avctx);
1907         MPV_common_end(s);
1908     }
1909     if (!s->context_initialized) {
1910         if(h != h0)
1911             return -1;  // we cant (re-)initialize context during parallel decoding
1912
1913         avcodec_set_dimensions(s->avctx, s->width, s->height);
1914         s->avctx->sample_aspect_ratio= h->sps.sar;
1915         av_assert0(s->avctx->sample_aspect_ratio.den);
1916
1917         if(h->sps.video_signal_type_present_flag){
1918             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
1919             if(h->sps.colour_description_present_flag){
1920                 s->avctx->color_primaries = h->sps.color_primaries;
1921                 s->avctx->color_trc       = h->sps.color_trc;
1922                 s->avctx->colorspace      = h->sps.colorspace;
1923             }
1924         }
1925
1926         if(h->sps.timing_info_present_flag){
1927             int64_t den= h->sps.time_scale;
1928             if(h->x264_build < 44U)
1929                 den *= 2;
1930             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
1931                       h->sps.num_units_in_tick, den, 1<<30);
1932         }
1933
1934         switch (h->sps.bit_depth_luma) {
1935             case 9 :
1936                 s->avctx->pix_fmt = PIX_FMT_YUV420P9;
1937                 break;
1938             case 10 :
1939                 s->avctx->pix_fmt = PIX_FMT_YUV420P10;
1940                 break;
1941             default:
1942         s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
1943                                                  s->avctx->codec->pix_fmts ?
1944                                                  s->avctx->codec->pix_fmts :
1945                                                  s->avctx->color_range == AVCOL_RANGE_JPEG ?
1946                                                  hwaccel_pixfmt_list_h264_jpeg_420 :
1947                                                  ff_hwaccel_pixfmt_list_420);
1948         }
1949
1950         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
1951
1952         if (MPV_common_init(s) < 0)
1953             return -1;
1954         s->first_field = 0;
1955         h->prev_interlaced_frame = 1;
1956
1957         init_scan_tables(h);
1958         ff_h264_alloc_tables(h);
1959
1960         for(i = 1; i < s->avctx->thread_count; i++) {
1961             H264Context *c;
1962             c = h->thread_context[i] = av_malloc(sizeof(H264Context));
1963             memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
1964             memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
1965             c->h264dsp = h->h264dsp;
1966             c->sps = h->sps;
1967             c->pps = h->pps;
1968             c->pixel_shift = h->pixel_shift;
1969             init_scan_tables(c);
1970             clone_tables(c, h, i);
1971         }
1972
1973         for(i = 0; i < s->avctx->thread_count; i++)
1974             if(context_init(h->thread_context[i]) < 0)
1975                 return -1;
1976     }
1977
1978     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
1979
1980     h->mb_mbaff = 0;
1981     h->mb_aff_frame = 0;
1982     last_pic_structure = s0->picture_structure;
1983     if(h->sps.frame_mbs_only_flag){
1984         s->picture_structure= PICT_FRAME;
1985     }else{
1986         if(get_bits1(&s->gb)) { //field_pic_flag
1987             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
1988         } else {
1989             s->picture_structure= PICT_FRAME;
1990             h->mb_aff_frame = h->sps.mb_aff;
1991         }
1992     }
1993     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
1994
1995     if(h0->current_slice == 0){
1996         while(h->frame_num !=  h->prev_frame_num &&
1997               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
1998             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
1999             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2000             if (ff_h264_frame_start(h) < 0)
2001                 return -1;
2002             h->prev_frame_num++;
2003             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2004             s->current_picture_ptr->frame_num= h->prev_frame_num;
2005             ff_generate_sliding_window_mmcos(h);
2006             ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2007             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2008              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2009              * about there being no actual duplicates.
2010              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2011              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2012              * be fixed. */
2013             if (h->short_ref_count) {
2014                 if (prev) {
2015                     av_image_copy(h->short_ref[0]->data, h->short_ref[0]->linesize,
2016                                   (const uint8_t**)prev->data, prev->linesize,
2017                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2018                     h->short_ref[0]->poc = prev->poc+2;
2019                 }
2020                 h->short_ref[0]->frame_num = h->prev_frame_num;
2021             }
2022         }
2023
2024         /* See if we have a decoded first field looking for a pair... */
2025         if (s0->first_field) {
2026             assert(s0->current_picture_ptr);
2027             assert(s0->current_picture_ptr->data[0]);
2028             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
2029
2030             /* figure out if we have a complementary field pair */
2031             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2032                 /*
2033                  * Previous field is unmatched. Don't display it, but let it
2034                  * remain for reference if marked as such.
2035                  */
2036                 s0->current_picture_ptr = NULL;
2037                 s0->first_field = FIELD_PICTURE;
2038
2039             } else {
2040                 if (h->nal_ref_idc &&
2041                         s0->current_picture_ptr->reference &&
2042                         s0->current_picture_ptr->frame_num != h->frame_num) {
2043                     /*
2044                      * This and previous field were reference, but had
2045                      * different frame_nums. Consider this field first in
2046                      * pair. Throw away previous field except for reference
2047                      * purposes.
2048                      */
2049                     s0->first_field = 1;
2050                     s0->current_picture_ptr = NULL;
2051
2052                 } else {
2053                     /* Second field in complementary pair */
2054                     s0->first_field = 0;
2055                 }
2056             }
2057
2058         } else {
2059             /* Frame or first field in a potentially complementary pair */
2060             assert(!s0->current_picture_ptr);
2061             s0->first_field = FIELD_PICTURE;
2062         }
2063
2064         if((!FIELD_PICTURE || s0->first_field) && ff_h264_frame_start(h) < 0) {
2065             s0->first_field = 0;
2066             return -1;
2067         }
2068     }
2069     if(h != h0)
2070         clone_slice(h, h0);
2071
2072     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2073
2074     assert(s->mb_num == s->mb_width * s->mb_height);
2075     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2076        first_mb_in_slice                    >= s->mb_num){
2077         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2078         return -1;
2079     }
2080     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2081     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2082     if (s->picture_structure == PICT_BOTTOM_FIELD)
2083         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2084     assert(s->mb_y < s->mb_height);
2085
2086     if(s->picture_structure==PICT_FRAME){
2087         h->curr_pic_num=   h->frame_num;
2088         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2089     }else{
2090         h->curr_pic_num= 2*h->frame_num + 1;
2091         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2092     }
2093
2094     if(h->nal_unit_type == NAL_IDR_SLICE){
2095         get_ue_golomb(&s->gb); /* idr_pic_id */
2096     }
2097
2098     if(h->sps.poc_type==0){
2099         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2100
2101         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2102             h->delta_poc_bottom= get_se_golomb(&s->gb);
2103         }
2104     }
2105
2106     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2107         h->delta_poc[0]= get_se_golomb(&s->gb);
2108
2109         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2110             h->delta_poc[1]= get_se_golomb(&s->gb);
2111     }
2112
2113     init_poc(h);
2114
2115     if(h->pps.redundant_pic_cnt_present){
2116         h->redundant_pic_count= get_ue_golomb(&s->gb);
2117     }
2118
2119     //set defaults, might be overridden a few lines later
2120     h->ref_count[0]= h->pps.ref_count[0];
2121     h->ref_count[1]= h->pps.ref_count[1];
2122
2123     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2124         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2125             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2126         }
2127         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2128
2129         if(num_ref_idx_active_override_flag){
2130             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2131             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2132                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2133
2134             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2135                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2136                 h->ref_count[0]= h->ref_count[1]= 1;
2137                 return -1;
2138             }
2139         }
2140         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2141             h->list_count= 2;
2142         else
2143             h->list_count= 1;
2144     }else
2145         h->list_count= 0;
2146
2147     if(!default_ref_list_done){
2148         ff_h264_fill_default_ref_list(h);
2149     }
2150
2151     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0)
2152         return -1;
2153
2154     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2155         s->last_picture_ptr= &h->ref_list[0][0];
2156         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2157     }
2158     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2159         s->next_picture_ptr= &h->ref_list[1][0];
2160         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2161     }
2162
2163     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2164        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2165         pred_weight_table(h);
2166     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2167         implicit_weight_table(h, -1);
2168     }else {
2169         h->use_weight = 0;
2170         for (i = 0; i < 2; i++) {
2171             h->luma_weight_flag[i]   = 0;
2172             h->chroma_weight_flag[i] = 0;
2173         }
2174     }
2175
2176     if(h->nal_ref_idc)
2177         ff_h264_decode_ref_pic_marking(h0, &s->gb);
2178
2179     if(FRAME_MBAFF){
2180         ff_h264_fill_mbaff_ref_list(h);
2181
2182         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2183             implicit_weight_table(h, 0);
2184             implicit_weight_table(h, 1);
2185         }
2186     }
2187
2188     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2189         ff_h264_direct_dist_scale_factor(h);
2190     ff_h264_direct_ref_list_init(h);
2191
2192     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2193         tmp = get_ue_golomb_31(&s->gb);
2194         if(tmp > 2){
2195             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2196             return -1;
2197         }
2198         h->cabac_init_idc= tmp;
2199     }
2200
2201     h->last_qscale_diff = 0;
2202     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2203     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
2204         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2205         return -1;
2206     }
2207     s->qscale= tmp;
2208     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2209     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2210     //FIXME qscale / qp ... stuff
2211     if(h->slice_type == AV_PICTURE_TYPE_SP){
2212         get_bits1(&s->gb); /* sp_for_switch_flag */
2213     }
2214     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
2215         get_se_golomb(&s->gb); /* slice_qs_delta */
2216     }
2217
2218     h->deblocking_filter = 1;
2219     h->slice_alpha_c0_offset = 52;
2220     h->slice_beta_offset = 52;
2221     if( h->pps.deblocking_filter_parameters_present ) {
2222         tmp= get_ue_golomb_31(&s->gb);
2223         if(tmp > 2){
2224             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2225             return -1;
2226         }
2227         h->deblocking_filter= tmp;
2228         if(h->deblocking_filter < 2)
2229             h->deblocking_filter^= 1; // 1<->0
2230
2231         if( h->deblocking_filter ) {
2232             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2233             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2234             if(   h->slice_alpha_c0_offset > 104U
2235                || h->slice_beta_offset     > 104U){
2236                 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);
2237                 return -1;
2238             }
2239         }
2240     }
2241
2242     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2243        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
2244        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
2245        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2246         h->deblocking_filter= 0;
2247
2248     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
2249         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
2250             /* Cheat slightly for speed:
2251                Do not bother to deblock across slices. */
2252             h->deblocking_filter = 2;
2253         } else {
2254             h0->max_contexts = 1;
2255             if(!h0->single_decode_warning) {
2256                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
2257                 h0->single_decode_warning = 1;
2258             }
2259             if(h != h0)
2260                 return 1; // deblocking switched inside frame
2261         }
2262     }
2263     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]);
2264
2265 #if 0 //FMO
2266     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
2267         slice_group_change_cycle= get_bits(&s->gb, ?);
2268 #endif
2269
2270     h0->last_slice_type = slice_type;
2271     h->slice_num = ++h0->current_slice;
2272     if(h->slice_num >= MAX_SLICES){
2273         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
2274     }
2275
2276     for(j=0; j<2; j++){
2277         int id_list[16];
2278         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
2279         for(i=0; i<16; i++){
2280             id_list[i]= 60;
2281             if(h->ref_list[j][i].data[0]){
2282                 int k;
2283                 uint8_t *base= h->ref_list[j][i].base[0];
2284                 for(k=0; k<h->short_ref_count; k++)
2285                     if(h->short_ref[k]->base[0] == base){
2286                         id_list[i]= k;
2287                         break;
2288                     }
2289                 for(k=0; k<h->long_ref_count; k++)
2290                     if(h->long_ref[k] && h->long_ref[k]->base[0] == base){
2291                         id_list[i]= h->short_ref_count + k;
2292                         break;
2293                     }
2294             }
2295         }
2296
2297         ref2frm[0]=
2298         ref2frm[1]= -1;
2299         for(i=0; i<16; i++)
2300             ref2frm[i+2]= 4*id_list[i]
2301                           +(h->ref_list[j][i].reference&3);
2302         ref2frm[18+0]=
2303         ref2frm[18+1]= -1;
2304         for(i=16; i<48; i++)
2305             ref2frm[i+4]= 4*id_list[(i-16)>>1]
2306                           +(h->ref_list[j][i].reference&3);
2307     }
2308
2309     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE) ? 0 : 16;
2310     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
2311
2312     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
2313         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",
2314                h->slice_num,
2315                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
2316                first_mb_in_slice,
2317                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
2318                pps_id, h->frame_num,
2319                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
2320                h->ref_count[0], h->ref_count[1],
2321                s->qscale,
2322                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
2323                h->use_weight,
2324                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
2325                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
2326                );
2327     }
2328
2329     return 0;
2330 }
2331
2332 int ff_h264_get_slice_type(const H264Context *h)
2333 {
2334     switch (h->slice_type) {
2335     case AV_PICTURE_TYPE_P:  return 0;
2336     case AV_PICTURE_TYPE_B:  return 1;
2337     case AV_PICTURE_TYPE_I:  return 2;
2338     case AV_PICTURE_TYPE_SP: return 3;
2339     case AV_PICTURE_TYPE_SI: return 4;
2340     default:         return -1;
2341     }
2342 }
2343
2344 /**
2345  *
2346  * @return non zero if the loop filter can be skiped
2347  */
2348 static int fill_filter_caches(H264Context *h, int mb_type){
2349     MpegEncContext * const s = &h->s;
2350     const int mb_xy= h->mb_xy;
2351     int top_xy, left_xy[2];
2352     int top_type, left_type[2];
2353
2354     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
2355
2356     //FIXME deblocking could skip the intra and nnz parts.
2357
2358     /* Wow, what a mess, why didn't they simplify the interlacing & intra
2359      * stuff, I can't imagine that these complex rules are worth it. */
2360
2361     left_xy[1] = left_xy[0] = mb_xy-1;
2362     if(FRAME_MBAFF){
2363         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[mb_xy-1]);
2364         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
2365         if(s->mb_y&1){
2366             if (left_mb_field_flag != curr_mb_field_flag) {
2367                 left_xy[0] -= s->mb_stride;
2368             }
2369         }else{
2370             if(curr_mb_field_flag){
2371                 top_xy      += s->mb_stride & (((s->current_picture.mb_type[top_xy    ]>>7)&1)-1);
2372             }
2373             if (left_mb_field_flag != curr_mb_field_flag) {
2374                 left_xy[1] += s->mb_stride;
2375             }
2376         }
2377     }
2378
2379     h->top_mb_xy = top_xy;
2380     h->left_mb_xy[0] = left_xy[0];
2381     h->left_mb_xy[1] = left_xy[1];
2382     {
2383         //for sufficiently low qp, filtering wouldn't do anything
2384         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
2385         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
2386         int qp = s->current_picture.qscale_table[mb_xy];
2387         if(qp <= qp_thresh
2388            && (left_xy[0]<0 || ((qp + s->current_picture.qscale_table[left_xy[0]] + 1)>>1) <= qp_thresh)
2389            && (top_xy   < 0 || ((qp + s->current_picture.qscale_table[top_xy    ] + 1)>>1) <= qp_thresh)){
2390             if(!FRAME_MBAFF)
2391                 return 1;
2392             if(   (left_xy[0]< 0            || ((qp + s->current_picture.qscale_table[left_xy[1]             ] + 1)>>1) <= qp_thresh)
2393                && (top_xy    < s->mb_stride || ((qp + s->current_picture.qscale_table[top_xy    -s->mb_stride] + 1)>>1) <= qp_thresh))
2394                 return 1;
2395         }
2396     }
2397
2398     top_type     = s->current_picture.mb_type[top_xy]    ;
2399     left_type[0] = s->current_picture.mb_type[left_xy[0]];
2400     left_type[1] = s->current_picture.mb_type[left_xy[1]];
2401     if(h->deblocking_filter == 2){
2402         if(h->slice_table[top_xy     ] != h->slice_num) top_type= 0;
2403         if(h->slice_table[left_xy[0] ] != h->slice_num) left_type[0]= left_type[1]= 0;
2404     }else{
2405         if(h->slice_table[top_xy     ] == 0xFFFF) top_type= 0;
2406         if(h->slice_table[left_xy[0] ] == 0xFFFF) left_type[0]= left_type[1] =0;
2407     }
2408     h->top_type    = top_type    ;
2409     h->left_type[0]= left_type[0];
2410     h->left_type[1]= left_type[1];
2411
2412     if(IS_INTRA(mb_type))
2413         return 0;
2414
2415     AV_COPY64(&h->non_zero_count_cache[0+8*1], &h->non_zero_count[mb_xy][ 0]);
2416     AV_COPY64(&h->non_zero_count_cache[0+8*2], &h->non_zero_count[mb_xy][ 8]);
2417     AV_COPY32(&h->non_zero_count_cache[0+8*5], &h->non_zero_count[mb_xy][16]);
2418     AV_COPY32(&h->non_zero_count_cache[4+8*3], &h->non_zero_count[mb_xy][20]);
2419     AV_COPY64(&h->non_zero_count_cache[0+8*4], &h->non_zero_count[mb_xy][24]);
2420
2421     h->cbp= h->cbp_table[mb_xy];
2422
2423     {
2424         int list;
2425         for(list=0; list<h->list_count; list++){
2426             int8_t *ref;
2427             int y, b_stride;
2428             int16_t (*mv_dst)[2];
2429             int16_t (*mv_src)[2];
2430
2431             if(!USES_LIST(mb_type, list)){
2432                 fill_rectangle(  h->mv_cache[list][scan8[0]], 4, 4, 8, pack16to32(0,0), 4);
2433                 AV_WN32A(&h->ref_cache[list][scan8[ 0]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2434                 AV_WN32A(&h->ref_cache[list][scan8[ 2]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2435                 AV_WN32A(&h->ref_cache[list][scan8[ 8]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2436                 AV_WN32A(&h->ref_cache[list][scan8[10]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2437                 continue;
2438             }
2439
2440             ref = &s->current_picture.ref_index[list][4*mb_xy];
2441             {
2442                 int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
2443                 AV_WN32A(&h->ref_cache[list][scan8[ 0]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2444                 AV_WN32A(&h->ref_cache[list][scan8[ 2]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2445                 ref += 2;
2446                 AV_WN32A(&h->ref_cache[list][scan8[ 8]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2447                 AV_WN32A(&h->ref_cache[list][scan8[10]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2448             }
2449
2450             b_stride = h->b_stride;
2451             mv_dst   = &h->mv_cache[list][scan8[0]];
2452             mv_src   = &s->current_picture.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
2453             for(y=0; y<4; y++){
2454                 AV_COPY128(mv_dst + 8*y, mv_src + y*b_stride);
2455             }
2456
2457         }
2458     }
2459
2460
2461 /*
2462 0 . T T. T T T T
2463 1 L . .L . . . .
2464 2 L . .L . . . .
2465 3 . T TL . . . .
2466 4 L . .L . . . .
2467 5 L . .. . . . .
2468 */
2469 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
2470     if(top_type){
2471         AV_COPY32(&h->non_zero_count_cache[4+8*0], &h->non_zero_count[top_xy][4+3*8]);
2472     }
2473
2474     if(left_type[0]){
2475         h->non_zero_count_cache[3+8*1]= h->non_zero_count[left_xy[0]][7+0*8];
2476         h->non_zero_count_cache[3+8*2]= h->non_zero_count[left_xy[0]][7+1*8];
2477         h->non_zero_count_cache[3+8*3]= h->non_zero_count[left_xy[0]][7+2*8];
2478         h->non_zero_count_cache[3+8*4]= h->non_zero_count[left_xy[0]][7+3*8];
2479     }
2480
2481     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
2482     if(!CABAC && h->pps.transform_8x8_mode){
2483         if(IS_8x8DCT(top_type)){
2484             h->non_zero_count_cache[4+8*0]=
2485             h->non_zero_count_cache[5+8*0]= h->cbp_table[top_xy] & 4;
2486             h->non_zero_count_cache[6+8*0]=
2487             h->non_zero_count_cache[7+8*0]= h->cbp_table[top_xy] & 8;
2488         }
2489         if(IS_8x8DCT(left_type[0])){
2490             h->non_zero_count_cache[3+8*1]=
2491             h->non_zero_count_cache[3+8*2]= h->cbp_table[left_xy[0]]&2; //FIXME check MBAFF
2492         }
2493         if(IS_8x8DCT(left_type[1])){
2494             h->non_zero_count_cache[3+8*3]=
2495             h->non_zero_count_cache[3+8*4]= h->cbp_table[left_xy[1]]&8; //FIXME check MBAFF
2496         }
2497
2498         if(IS_8x8DCT(mb_type)){
2499             h->non_zero_count_cache[scan8[0   ]]= h->non_zero_count_cache[scan8[1   ]]=
2500             h->non_zero_count_cache[scan8[2   ]]= h->non_zero_count_cache[scan8[3   ]]= h->cbp & 1;
2501
2502             h->non_zero_count_cache[scan8[0+ 4]]= h->non_zero_count_cache[scan8[1+ 4]]=
2503             h->non_zero_count_cache[scan8[2+ 4]]= h->non_zero_count_cache[scan8[3+ 4]]= h->cbp & 2;
2504
2505             h->non_zero_count_cache[scan8[0+ 8]]= h->non_zero_count_cache[scan8[1+ 8]]=
2506             h->non_zero_count_cache[scan8[2+ 8]]= h->non_zero_count_cache[scan8[3+ 8]]= h->cbp & 4;
2507
2508             h->non_zero_count_cache[scan8[0+12]]= h->non_zero_count_cache[scan8[1+12]]=
2509             h->non_zero_count_cache[scan8[2+12]]= h->non_zero_count_cache[scan8[3+12]]= h->cbp & 8;
2510         }
2511     }
2512
2513     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
2514         int list;
2515         for(list=0; list<h->list_count; list++){
2516             if(USES_LIST(top_type, list)){
2517                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
2518                 const int b8_xy= 4*top_xy + 2;
2519                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
2520                 AV_COPY128(h->mv_cache[list][scan8[0] + 0 - 1*8], s->current_picture.motion_val[list][b_xy + 0]);
2521                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
2522                 h->ref_cache[list][scan8[0] + 1 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 0]];
2523                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
2524                 h->ref_cache[list][scan8[0] + 3 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 1]];
2525             }else{
2526                 AV_ZERO128(h->mv_cache[list][scan8[0] + 0 - 1*8]);
2527                 AV_WN32A(&h->ref_cache[list][scan8[0] + 0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2528             }
2529
2530             if(!IS_INTERLACED(mb_type^left_type[0])){
2531                 if(USES_LIST(left_type[0], list)){
2532                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
2533                     const int b8_xy= 4*left_xy[0] + 1;
2534                     int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[0]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
2535                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 0 ], s->current_picture.motion_val[list][b_xy + h->b_stride*0]);
2536                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 8 ], s->current_picture.motion_val[list][b_xy + h->b_stride*1]);
2537                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 +16 ], s->current_picture.motion_val[list][b_xy + h->b_stride*2]);
2538                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 +24 ], s->current_picture.motion_val[list][b_xy + h->b_stride*3]);
2539                     h->ref_cache[list][scan8[0] - 1 + 0 ]=
2540                     h->ref_cache[list][scan8[0] - 1 + 8 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*0]];
2541                     h->ref_cache[list][scan8[0] - 1 +16 ]=
2542                     h->ref_cache[list][scan8[0] - 1 +24 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*1]];
2543                 }else{
2544                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 0 ]);
2545                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 8 ]);
2546                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 +16 ]);
2547                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 +24 ]);
2548                     h->ref_cache[list][scan8[0] - 1 + 0  ]=
2549                     h->ref_cache[list][scan8[0] - 1 + 8  ]=
2550                     h->ref_cache[list][scan8[0] - 1 + 16 ]=
2551                     h->ref_cache[list][scan8[0] - 1 + 24 ]= LIST_NOT_USED;
2552                 }
2553             }
2554         }
2555     }
2556
2557     return 0;
2558 }
2559
2560 static void loop_filter(H264Context *h, int start_x, int end_x){
2561     MpegEncContext * const s = &h->s;
2562     uint8_t  *dest_y, *dest_cb, *dest_cr;
2563     int linesize, uvlinesize, mb_x, mb_y;
2564     const int end_mb_y= s->mb_y + FRAME_MBAFF;
2565     const int old_slice_type= h->slice_type;
2566     const int pixel_shift = h->pixel_shift;
2567
2568     if(h->deblocking_filter) {
2569         for(mb_x= start_x; mb_x<end_x; mb_x++){
2570             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
2571                 int mb_xy, mb_type;
2572                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
2573                 h->slice_num= h->slice_table[mb_xy];
2574                 mb_type= s->current_picture.mb_type[mb_xy];
2575                 h->list_count= h->list_counts[mb_xy];
2576
2577                 if(FRAME_MBAFF)
2578                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
2579
2580                 s->mb_x= mb_x;
2581                 s->mb_y= mb_y;
2582                 dest_y  = s->current_picture.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
2583                 dest_cb = s->current_picture.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
2584                 dest_cr = s->current_picture.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
2585                     //FIXME simplify above
2586
2587                 if (MB_FIELD) {
2588                     linesize   = h->mb_linesize   = s->linesize * 2;
2589                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2590                     if(mb_y&1){ //FIXME move out of this function?
2591                         dest_y -= s->linesize*15;
2592                         dest_cb-= s->uvlinesize*7;
2593                         dest_cr-= s->uvlinesize*7;
2594                     }
2595                 } else {
2596                     linesize   = h->mb_linesize   = s->linesize;
2597                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2598                 }
2599                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
2600                 if(fill_filter_caches(h, mb_type))
2601                     continue;
2602                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2603                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2604
2605                 if (FRAME_MBAFF) {
2606                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2607                 } else {
2608                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2609                 }
2610             }
2611         }
2612     }
2613     h->slice_type= old_slice_type;
2614     s->mb_x= 0;
2615     s->mb_y= end_mb_y - FRAME_MBAFF;
2616     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2617     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2618 }
2619
2620 static void predict_field_decoding_flag(H264Context *h){
2621     MpegEncContext * const s = &h->s;
2622     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
2623     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
2624                 ? s->current_picture.mb_type[mb_xy-1]
2625                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
2626                 ? s->current_picture.mb_type[mb_xy-s->mb_stride]
2627                 : 0;
2628     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
2629 }
2630
2631 static int decode_slice(struct AVCodecContext *avctx, void *arg){
2632     H264Context *h = *(void**)arg;
2633     MpegEncContext * const s = &h->s;
2634     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
2635     int lf_x_start = s->mb_x;
2636
2637     s->mb_skip_run= -1;
2638
2639     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
2640                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
2641
2642     if( h->pps.cabac ) {
2643         /* realign */
2644         align_get_bits( &s->gb );
2645
2646         /* init cabac */
2647         ff_init_cabac_states( &h->cabac);
2648         ff_init_cabac_decoder( &h->cabac,
2649                                s->gb.buffer + get_bits_count(&s->gb)/8,
2650                                (get_bits_left(&s->gb) + 7)/8);
2651
2652         ff_h264_init_cabac_states(h);
2653
2654         for(;;){
2655 //START_TIMER
2656             int ret = ff_h264_decode_mb_cabac(h);
2657             int eos;
2658 //STOP_TIMER("decode_mb_cabac")
2659
2660             if(ret>=0) ff_h264_hl_decode_mb(h);
2661
2662             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
2663                 s->mb_y++;
2664
2665                 ret = ff_h264_decode_mb_cabac(h);
2666
2667                 if(ret>=0) ff_h264_hl_decode_mb(h);
2668                 s->mb_y--;
2669             }
2670             eos = get_cabac_terminate( &h->cabac );
2671
2672             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
2673                 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);
2674                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
2675                 return 0;
2676             }
2677             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
2678                 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);
2679                 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);
2680                 return -1;
2681             }
2682
2683             if( ++s->mb_x >= s->mb_width ) {
2684                 loop_filter(h, lf_x_start, s->mb_x);
2685                 s->mb_x = lf_x_start = 0;
2686                 ff_draw_horiz_band(s, 16*s->mb_y, 16);
2687                 ++s->mb_y;
2688                 if(FIELD_OR_MBAFF_PICTURE) {
2689                     ++s->mb_y;
2690                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
2691                         predict_field_decoding_flag(h);
2692                 }
2693             }
2694
2695             if( eos || s->mb_y >= s->mb_height ) {
2696                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
2697                 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);
2698                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
2699                 return 0;
2700             }
2701         }
2702
2703     } else {
2704         for(;;){
2705             int ret = ff_h264_decode_mb_cavlc(h);
2706
2707             if(ret>=0) ff_h264_hl_decode_mb(h);
2708
2709             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
2710                 s->mb_y++;
2711                 ret = ff_h264_decode_mb_cavlc(h);
2712
2713                 if(ret>=0) ff_h264_hl_decode_mb(h);
2714                 s->mb_y--;
2715             }
2716
2717             if(ret<0){
2718                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
2719                 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);
2720                 return -1;
2721             }
2722
2723             if(++s->mb_x >= s->mb_width){
2724                 loop_filter(h, lf_x_start, s->mb_x);
2725                 s->mb_x = lf_x_start = 0;
2726                 ff_draw_horiz_band(s, 16*s->mb_y, 16);
2727                 ++s->mb_y;
2728                 if(FIELD_OR_MBAFF_PICTURE) {
2729                     ++s->mb_y;
2730                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
2731                         predict_field_decoding_flag(h);
2732                 }
2733                 if(s->mb_y >= s->mb_height){
2734                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
2735
2736                     if(get_bits_count(&s->gb) == s->gb.size_in_bits ) {
2737                         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);
2738
2739                         return 0;
2740                     }else{
2741                         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);
2742
2743                         return -1;
2744                     }
2745                 }
2746             }
2747
2748             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
2749                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
2750                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
2751                     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);
2752                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
2753
2754                     return 0;
2755                 }else{
2756                     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);
2757
2758                     return -1;
2759                 }
2760             }
2761         }
2762     }
2763
2764 #if 0
2765     for(;s->mb_y < s->mb_height; s->mb_y++){
2766         for(;s->mb_x < s->mb_width; s->mb_x++){
2767             int ret= decode_mb(h);
2768
2769             ff_h264_hl_decode_mb(h);
2770
2771             if(ret<0){
2772                 av_log(s->avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
2773                 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);
2774
2775                 return -1;
2776             }
2777
2778             if(++s->mb_x >= s->mb_width){
2779                 s->mb_x=0;
2780                 if(++s->mb_y >= s->mb_height){
2781                     if(get_bits_count(s->gb) == s->gb.size_in_bits){
2782                         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);
2783
2784                         return 0;
2785                     }else{
2786                         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);
2787
2788                         return -1;
2789                     }
2790                 }
2791             }
2792
2793             if(get_bits_count(s->?gb) >= s->gb?.size_in_bits){
2794                 if(get_bits_count(s->gb) == s->gb.size_in_bits){
2795                     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);
2796
2797                     return 0;
2798                 }else{
2799                     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);
2800
2801                     return -1;
2802                 }
2803             }
2804         }
2805         s->mb_x=0;
2806         ff_draw_horiz_band(s, 16*s->mb_y, 16);
2807     }
2808 #endif
2809     return -1; //not reached
2810 }
2811
2812 /**
2813  * Call decode_slice() for each context.
2814  *
2815  * @param h h264 master context
2816  * @param context_count number of contexts to execute
2817  */
2818 static void execute_decode_slices(H264Context *h, int context_count){
2819     MpegEncContext * const s = &h->s;
2820     AVCodecContext * const avctx= s->avctx;
2821     H264Context *hx;
2822     int i;
2823
2824     if (s->avctx->hwaccel)
2825         return;
2826     if(s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2827         return;
2828     if(context_count == 1) {
2829         decode_slice(avctx, &h);
2830     } else {
2831         for(i = 1; i < context_count; i++) {
2832             hx = h->thread_context[i];
2833             hx->s.error_recognition = avctx->error_recognition;
2834             hx->s.error_count = 0;
2835         }
2836
2837         avctx->execute(avctx, (void *)decode_slice,
2838                        h->thread_context, NULL, context_count, sizeof(void*));
2839
2840         /* pull back stuff from slices to master context */
2841         hx = h->thread_context[context_count - 1];
2842         s->mb_x = hx->s.mb_x;
2843         s->mb_y = hx->s.mb_y;
2844         s->dropable = hx->s.dropable;
2845         s->picture_structure = hx->s.picture_structure;
2846         for(i = 1; i < context_count; i++)
2847             h->s.error_count += h->thread_context[i]->s.error_count;
2848     }
2849 }
2850
2851
2852 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
2853     MpegEncContext * const s = &h->s;
2854     AVCodecContext * const avctx= s->avctx;
2855     int buf_index=0;
2856     H264Context *hx; ///< thread context
2857     int context_count = 0;
2858     int next_avc= h->is_avc ? 0 : buf_size;
2859
2860     h->max_contexts = avctx->thread_count;
2861     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
2862         h->current_slice = 0;
2863         if (!s->first_field)
2864             s->current_picture_ptr= NULL;
2865         ff_h264_reset_sei(h);
2866     }
2867
2868     for(;;){
2869         int consumed;
2870         int dst_length;
2871         int bit_length;
2872         const uint8_t *ptr;
2873         int i, nalsize = 0;
2874         int err;
2875
2876         if(buf_index >= next_avc) {
2877             if(buf_index >= buf_size) break;
2878             nalsize = 0;
2879             for(i = 0; i < h->nal_length_size; i++)
2880                 nalsize = (nalsize << 8) | buf[buf_index++];
2881             if(nalsize <= 0 || nalsize > buf_size - buf_index){
2882                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
2883                 break;
2884             }
2885             next_avc= buf_index + nalsize;
2886         } else {
2887             // start code prefix search
2888             for(; buf_index + 3 < next_avc; buf_index++){
2889                 // This should always succeed in the first iteration.
2890                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
2891                     break;
2892             }
2893
2894             if(buf_index+3 >= buf_size) break;
2895
2896             buf_index+=3;
2897             if(buf_index >= next_avc) continue;
2898         }
2899
2900         hx = h->thread_context[context_count];
2901
2902         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
2903         if (ptr==NULL || dst_length < 0){
2904             return -1;
2905         }
2906         i= buf_index + consumed;
2907         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
2908            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
2909             s->workaround_bugs |= FF_BUG_TRUNCATED;
2910
2911         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
2912         while(ptr[dst_length - 1] == 0 && dst_length > 0)
2913             dst_length--;
2914         }
2915         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
2916
2917         if(s->avctx->debug&FF_DEBUG_STARTCODE){
2918             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);
2919         }
2920
2921         if (h->is_avc && (nalsize != consumed) && nalsize){
2922             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
2923         }
2924
2925         buf_index += consumed;
2926
2927         //FIXME do not discard SEI id
2928         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
2929             continue;
2930
2931       again:
2932         err = 0;
2933         switch(hx->nal_unit_type){
2934         case NAL_IDR_SLICE:
2935             if (h->nal_unit_type != NAL_IDR_SLICE) {
2936                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
2937                 return -1;
2938             }
2939             idr(h); //FIXME ensure we don't loose some frames if there is reordering
2940         case NAL_SLICE:
2941             init_get_bits(&hx->s.gb, ptr, bit_length);
2942             hx->intra_gb_ptr=
2943             hx->inter_gb_ptr= &hx->s.gb;
2944             hx->s.data_partitioning = 0;
2945
2946             if((err = decode_slice_header(hx, h)))
2947                break;
2948
2949             if (h->current_slice == 1) {
2950                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
2951                     return -1;
2952                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2953                     ff_vdpau_h264_picture_start(s);
2954             }
2955
2956             s->current_picture_ptr->key_frame |=
2957                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
2958                     (h->sei_recovery_frame_cnt >= 0);
2959             if(hx->redundant_pic_count==0
2960                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
2961                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
2962                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
2963                && avctx->skip_frame < AVDISCARD_ALL){
2964                 if(avctx->hwaccel) {
2965                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
2966                         return -1;
2967                 }else
2968                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
2969                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
2970                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
2971                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
2972                 }else
2973                     context_count++;
2974             }
2975             break;
2976         case NAL_DPA:
2977             init_get_bits(&hx->s.gb, ptr, bit_length);
2978             hx->intra_gb_ptr=
2979             hx->inter_gb_ptr= NULL;
2980
2981             if ((err = decode_slice_header(hx, h)) < 0)
2982                 break;
2983
2984             hx->s.data_partitioning = 1;
2985
2986             break;
2987         case NAL_DPB:
2988             init_get_bits(&hx->intra_gb, ptr, bit_length);
2989             hx->intra_gb_ptr= &hx->intra_gb;
2990             break;
2991         case NAL_DPC:
2992             init_get_bits(&hx->inter_gb, ptr, bit_length);
2993             hx->inter_gb_ptr= &hx->inter_gb;
2994
2995             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
2996                && s->context_initialized
2997                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
2998                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
2999                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3000                && avctx->skip_frame < AVDISCARD_ALL)
3001                 context_count++;
3002             break;
3003         case NAL_SEI:
3004             init_get_bits(&s->gb, ptr, bit_length);
3005             ff_h264_decode_sei(h);
3006             break;
3007         case NAL_SPS:
3008             init_get_bits(&s->gb, ptr, bit_length);
3009             ff_h264_decode_seq_parameter_set(h);
3010
3011             if(s->flags& CODEC_FLAG_LOW_DELAY)
3012                 s->low_delay=1;
3013
3014             if(avctx->has_b_frames < 2)
3015                 avctx->has_b_frames= !s->low_delay;
3016
3017             if (avctx->bits_per_raw_sample != h->sps.bit_depth_luma) {
3018                 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
3019                     avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3020                     h->pixel_shift = h->sps.bit_depth_luma > 8;
3021
3022                     ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma);
3023                     ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma);
3024                     dsputil_init(&s->dsp, s->avctx);
3025                 } else {
3026                     av_log(avctx, AV_LOG_DEBUG, "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3027                     return -1;
3028                 }
3029             }
3030             break;
3031         case NAL_PPS:
3032             init_get_bits(&s->gb, ptr, bit_length);
3033
3034             ff_h264_decode_picture_parameter_set(h, bit_length);
3035
3036             break;
3037         case NAL_AUD:
3038         case NAL_END_SEQUENCE:
3039         case NAL_END_STREAM:
3040         case NAL_FILLER_DATA:
3041         case NAL_SPS_EXT:
3042         case NAL_AUXILIARY_SLICE:
3043             break;
3044         default:
3045             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
3046         }
3047
3048         if(context_count == h->max_contexts) {
3049             execute_decode_slices(h, context_count);
3050             context_count = 0;
3051         }
3052
3053         if (err < 0)
3054             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3055         else if(err == 1) {
3056             /* Slice could not be decoded in parallel mode, copy down
3057              * NAL unit stuff to context 0 and restart. Note that
3058              * rbsp_buffer is not transferred, but since we no longer
3059              * run in parallel mode this should not be an issue. */
3060             h->nal_unit_type = hx->nal_unit_type;
3061             h->nal_ref_idc   = hx->nal_ref_idc;
3062             hx = h;
3063             goto again;
3064         }
3065     }
3066     if(context_count)
3067         execute_decode_slices(h, context_count);
3068     return buf_index;
3069 }
3070
3071 /**
3072  * returns the number of bytes consumed for building the current frame
3073  */
3074 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
3075         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
3076         if(pos+10>buf_size) pos=buf_size; // oops ;)
3077
3078         return pos;
3079 }
3080
3081 static int decode_frame(AVCodecContext *avctx,
3082                              void *data, int *data_size,
3083                              AVPacket *avpkt)
3084 {
3085     const uint8_t *buf = avpkt->data;
3086     int buf_size = avpkt->size;
3087     H264Context *h = avctx->priv_data;
3088     MpegEncContext *s = &h->s;
3089     AVFrame *pict = data;
3090     int buf_index;
3091
3092     s->flags= avctx->flags;
3093     s->flags2= avctx->flags2;
3094
3095    /* end of stream, output what is still in the buffers */
3096  out:
3097     if (buf_size == 0) {
3098         Picture *out;
3099         int i, out_idx;
3100
3101 //FIXME factorize this with the output code below
3102         out = h->delayed_pic[0];
3103         out_idx = 0;
3104         for(i=1; h->delayed_pic[i] && !h->delayed_pic[i]->key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3105             if(h->delayed_pic[i]->poc < out->poc){
3106                 out = h->delayed_pic[i];
3107                 out_idx = i;
3108             }
3109
3110         for(i=out_idx; h->delayed_pic[i]; i++)
3111             h->delayed_pic[i] = h->delayed_pic[i+1];
3112
3113         if(out){
3114             *data_size = sizeof(AVFrame);
3115             *pict= *(AVFrame*)out;
3116         }
3117
3118         return 0;
3119     }
3120
3121     buf_index=decode_nal_units(h, buf, buf_size);
3122     if(buf_index < 0)
3123         return -1;
3124
3125     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
3126         buf_size = 0;
3127         goto out;
3128     }
3129
3130     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
3131         if (avctx->skip_frame >= AVDISCARD_NONREF)
3132             return 0;
3133         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
3134         return -1;
3135     }
3136
3137     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
3138         Picture *out = s->current_picture_ptr;
3139         Picture *cur = s->current_picture_ptr;
3140         int i, pics, out_of_order, out_idx;
3141
3142         field_end(h);
3143
3144         if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
3145             /* Wait for second field. */
3146             *data_size = 0;
3147
3148         } else {
3149             cur->interlaced_frame = 0;
3150             cur->repeat_pict = 0;
3151
3152             /* Signal interlacing information externally. */
3153             /* Prioritize picture timing SEI information over used decoding process if it exists. */
3154
3155             if(h->sps.pic_struct_present_flag){
3156                 switch (h->sei_pic_struct)
3157                 {
3158                 case SEI_PIC_STRUCT_FRAME:
3159                     break;
3160                 case SEI_PIC_STRUCT_TOP_FIELD:
3161                 case SEI_PIC_STRUCT_BOTTOM_FIELD:
3162                     cur->interlaced_frame = 1;
3163                     break;
3164                 case SEI_PIC_STRUCT_TOP_BOTTOM:
3165                 case SEI_PIC_STRUCT_BOTTOM_TOP:
3166                     if (FIELD_OR_MBAFF_PICTURE)
3167                         cur->interlaced_frame = 1;
3168                     else
3169                         // try to flag soft telecine progressive
3170                         cur->interlaced_frame = h->prev_interlaced_frame;
3171                     break;
3172                 case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
3173                 case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
3174                     // Signal the possibility of telecined film externally (pic_struct 5,6)
3175                     // From these hints, let the applications decide if they apply deinterlacing.
3176                     cur->repeat_pict = 1;
3177                     break;
3178                 case SEI_PIC_STRUCT_FRAME_DOUBLING:
3179                     // Force progressive here, as doubling interlaced frame is a bad idea.
3180                     cur->repeat_pict = 2;
3181                     break;
3182                 case SEI_PIC_STRUCT_FRAME_TRIPLING:
3183                     cur->repeat_pict = 4;
3184                     break;
3185                 }
3186
3187                 if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
3188                     cur->interlaced_frame = (h->sei_ct_type & (1<<1)) != 0;
3189             }else{
3190                 /* Derive interlacing flag from used decoding process. */
3191                 cur->interlaced_frame = FIELD_OR_MBAFF_PICTURE;
3192             }
3193             h->prev_interlaced_frame = cur->interlaced_frame;
3194
3195             if (cur->field_poc[0] != cur->field_poc[1]){
3196                 /* Derive top_field_first from field pocs. */
3197                 cur->top_field_first = cur->field_poc[0] < cur->field_poc[1];
3198             }else{
3199                 if(cur->interlaced_frame || h->sps.pic_struct_present_flag){
3200                     /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
3201                     if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
3202                       || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
3203                         cur->top_field_first = 1;
3204                     else
3205                         cur->top_field_first = 0;
3206                 }else{
3207                     /* Most likely progressive */
3208                     cur->top_field_first = 0;
3209                 }
3210             }
3211
3212         //FIXME do something with unavailable reference frames
3213
3214             /* Sort B-frames into display order */
3215
3216             if(h->sps.bitstream_restriction_flag
3217                && s->avctx->has_b_frames < h->sps.num_reorder_frames){
3218                 s->avctx->has_b_frames = h->sps.num_reorder_frames;
3219                 s->low_delay = 0;
3220             }
3221
3222             if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
3223                && !h->sps.bitstream_restriction_flag){
3224                 s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
3225                 s->low_delay= 0;
3226             }
3227
3228             pics = 0;
3229             while(h->delayed_pic[pics]) pics++;
3230
3231             assert(pics <= MAX_DELAYED_PIC_COUNT);
3232
3233             h->delayed_pic[pics++] = cur;
3234             if(cur->reference == 0)
3235                 cur->reference = DELAYED_PIC_REF;
3236
3237             out = h->delayed_pic[0];
3238             out_idx = 0;
3239             for(i=1; h->delayed_pic[i] && !h->delayed_pic[i]->key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3240                 if(h->delayed_pic[i]->poc < out->poc){
3241                     out = h->delayed_pic[i];
3242                     out_idx = i;
3243                 }
3244             if(s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset))
3245                 h->outputed_poc= INT_MIN;
3246             out_of_order = out->poc < h->outputed_poc;
3247
3248             if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
3249                 { }
3250             else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
3251                || (s->low_delay &&
3252                 ((h->outputed_poc != INT_MIN && out->poc > h->outputed_poc + 2)
3253                  || cur->pict_type == AV_PICTURE_TYPE_B)))
3254             {
3255                 s->low_delay = 0;
3256                 s->avctx->has_b_frames++;
3257             }
3258
3259             if(out_of_order || pics > s->avctx->has_b_frames){
3260                 out->reference &= ~DELAYED_PIC_REF;
3261                 for(i=out_idx; h->delayed_pic[i]; i++)
3262                     h->delayed_pic[i] = h->delayed_pic[i+1];
3263             }
3264             if(!out_of_order && pics > s->avctx->has_b_frames){
3265                 *data_size = sizeof(AVFrame);
3266
3267                 if(out_idx==0 && h->delayed_pic[0] && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset)) {
3268                     h->outputed_poc = INT_MIN;
3269                 } else
3270                     h->outputed_poc = out->poc;
3271                 *pict= *(AVFrame*)out;
3272             }else{
3273                 av_log(avctx, AV_LOG_DEBUG, "no picture\n");
3274             }
3275         }
3276     }
3277
3278     assert(pict->data[0] || !*data_size);
3279     ff_print_debug_info(s, pict);
3280 //printf("out %d\n", (int)pict->data[0]);
3281
3282     return get_consumed_bytes(s, buf_index, buf_size);
3283 }
3284 #if 0
3285 static inline void fill_mb_avail(H264Context *h){
3286     MpegEncContext * const s = &h->s;
3287     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3288
3289     if(s->mb_y){
3290         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
3291         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
3292         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
3293     }else{
3294         h->mb_avail[0]=
3295         h->mb_avail[1]=
3296         h->mb_avail[2]= 0;
3297     }
3298     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
3299     h->mb_avail[4]= 1; //FIXME move out
3300     h->mb_avail[5]= 0; //FIXME move out
3301 }
3302 #endif
3303
3304 #ifdef TEST
3305 #undef printf
3306 #undef random
3307 #define COUNT 8000
3308 #define SIZE (COUNT*40)
3309 int main(void){
3310     int i;
3311     uint8_t temp[SIZE];
3312     PutBitContext pb;
3313     GetBitContext gb;
3314 //    int int_temp[10000];
3315     DSPContext dsp;
3316     AVCodecContext avctx;
3317
3318     dsputil_init(&dsp, &avctx);
3319
3320     init_put_bits(&pb, temp, SIZE);
3321     printf("testing unsigned exp golomb\n");
3322     for(i=0; i<COUNT; i++){
3323         START_TIMER
3324         set_ue_golomb(&pb, i);
3325         STOP_TIMER("set_ue_golomb");
3326     }
3327     flush_put_bits(&pb);
3328
3329     init_get_bits(&gb, temp, 8*SIZE);
3330     for(i=0; i<COUNT; i++){
3331         int j, s;
3332
3333         s= show_bits(&gb, 24);
3334
3335         START_TIMER
3336         j= get_ue_golomb(&gb);
3337         if(j != i){
3338             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3339 //            return -1;
3340         }
3341         STOP_TIMER("get_ue_golomb");
3342     }
3343
3344
3345     init_put_bits(&pb, temp, SIZE);
3346     printf("testing signed exp golomb\n");
3347     for(i=0; i<COUNT; i++){
3348         START_TIMER
3349         set_se_golomb(&pb, i - COUNT/2);
3350         STOP_TIMER("set_se_golomb");
3351     }
3352     flush_put_bits(&pb);
3353
3354     init_get_bits(&gb, temp, 8*SIZE);
3355     for(i=0; i<COUNT; i++){
3356         int j, s;
3357
3358         s= show_bits(&gb, 24);
3359
3360         START_TIMER
3361         j= get_se_golomb(&gb);
3362         if(j != i - COUNT/2){
3363             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3364 //            return -1;
3365         }
3366         STOP_TIMER("get_se_golomb");
3367     }
3368
3369 #if 0
3370     printf("testing 4x4 (I)DCT\n");
3371
3372     DCTELEM block[16];
3373     uint8_t src[16], ref[16];
3374     uint64_t error= 0, max_error=0;
3375
3376     for(i=0; i<COUNT; i++){
3377         int j;
3378 //        printf("%d %d %d\n", r1, r2, (r2-r1)*16);
3379         for(j=0; j<16; j++){
3380             ref[j]= random()%255;
3381             src[j]= random()%255;
3382         }
3383
3384         h264_diff_dct_c(block, src, ref, 4);
3385
3386         //normalize
3387         for(j=0; j<16; j++){
3388 //            printf("%d ", block[j]);
3389             block[j]= block[j]*4;
3390             if(j&1) block[j]= (block[j]*4 + 2)/5;
3391             if(j&4) block[j]= (block[j]*4 + 2)/5;
3392         }
3393 //        printf("\n");
3394
3395         h->h264dsp.h264_idct_add(ref, block, 4);
3396 /*        for(j=0; j<16; j++){
3397             printf("%d ", ref[j]);
3398         }
3399         printf("\n");*/
3400
3401         for(j=0; j<16; j++){
3402             int diff= FFABS(src[j] - ref[j]);
3403
3404             error+= diff*diff;
3405             max_error= FFMAX(max_error, diff);
3406         }
3407     }
3408     printf("error=%f max_error=%d\n", ((float)error)/COUNT/16, (int)max_error );
3409     printf("testing quantizer\n");
3410     for(qp=0; qp<52; qp++){
3411         for(i=0; i<16; i++)
3412             src1_block[i]= src2_block[i]= random()%255;
3413
3414     }
3415     printf("Testing NAL layer\n");
3416
3417     uint8_t bitstream[COUNT];
3418     uint8_t nal[COUNT*2];
3419     H264Context h;
3420     memset(&h, 0, sizeof(H264Context));
3421
3422     for(i=0; i<COUNT; i++){
3423         int zeros= i;
3424         int nal_length;
3425         int consumed;
3426         int out_length;
3427         uint8_t *out;
3428         int j;
3429
3430         for(j=0; j<COUNT; j++){
3431             bitstream[j]= (random() % 255) + 1;
3432         }
3433
3434         for(j=0; j<zeros; j++){
3435             int pos= random() % COUNT;
3436             while(bitstream[pos] == 0){
3437                 pos++;
3438                 pos %= COUNT;
3439             }
3440             bitstream[pos]=0;
3441         }
3442
3443         START_TIMER
3444
3445         nal_length= encode_nal(&h, nal, bitstream, COUNT, COUNT*2);
3446         if(nal_length<0){
3447             printf("encoding failed\n");
3448             return -1;
3449         }
3450
3451         out= ff_h264_decode_nal(&h, nal, &out_length, &consumed, nal_length);
3452
3453         STOP_TIMER("NAL")
3454
3455         if(out_length != COUNT){
3456             printf("incorrect length %d %d\n", out_length, COUNT);
3457             return -1;
3458         }
3459
3460         if(consumed != nal_length){
3461             printf("incorrect consumed length %d %d\n", nal_length, consumed);
3462             return -1;
3463         }
3464
3465         if(memcmp(bitstream, out, COUNT)){
3466             printf("mismatch\n");