]> git.sesse.net Git - ffmpeg/blob - libavcodec/snowdec.c
Merge remote-tracking branch 'shariman/wmall'
[ffmpeg] / libavcodec / snowdec.c
1 /*
2  * Copyright (C) 2004 Michael Niedermayer <michaelni@gmx.at>
3  *
4  * This file is part of FFmpeg.
5  *
6  * FFmpeg is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * FFmpeg is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with FFmpeg; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20
21 #include "libavutil/intmath.h"
22 #include "libavutil/log.h"
23 #include "libavutil/opt.h"
24 #include "avcodec.h"
25 #include "dsputil.h"
26 #include "dwt.h"
27 #include "snow.h"
28
29 #include "rangecoder.h"
30 #include "mathops.h"
31
32 #include "mpegvideo.h"
33 #include "h263.h"
34
35 #undef NDEBUG
36 #include <assert.h>
37
38 static av_always_inline void predict_slice_buffered(SnowContext *s, slice_buffer * sb, IDWTELEM * old_buffer, int plane_index, int add, int mb_y){
39     Plane *p= &s->plane[plane_index];
40     const int mb_w= s->b_width  << s->block_max_depth;
41     const int mb_h= s->b_height << s->block_max_depth;
42     int x, y, mb_x;
43     int block_size = MB_SIZE >> s->block_max_depth;
44     int block_w    = plane_index ? block_size/2 : block_size;
45     const uint8_t *obmc  = plane_index ? obmc_tab[s->block_max_depth+1] : obmc_tab[s->block_max_depth];
46     int obmc_stride= plane_index ? block_size : 2*block_size;
47     int ref_stride= s->current_picture.linesize[plane_index];
48     uint8_t *dst8= s->current_picture.data[plane_index];
49     int w= p->width;
50     int h= p->height;
51
52     if(s->keyframe || (s->avctx->debug&512)){
53         if(mb_y==mb_h)
54             return;
55
56         if(add){
57             for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++){
58 //                DWTELEM * line = slice_buffer_get_line(sb, y);
59                 IDWTELEM * line = sb->line[y];
60                 for(x=0; x<w; x++){
61 //                    int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
62                     int v= line[x] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
63                     v >>= FRAC_BITS;
64                     if(v&(~255)) v= ~(v>>31);
65                     dst8[x + y*ref_stride]= v;
66                 }
67             }
68         }else{
69             for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++){
70 //                DWTELEM * line = slice_buffer_get_line(sb, y);
71                 IDWTELEM * line = sb->line[y];
72                 for(x=0; x<w; x++){
73                     line[x] -= 128 << FRAC_BITS;
74 //                    buf[x + y*w]-= 128<<FRAC_BITS;
75                 }
76             }
77         }
78
79         return;
80     }
81
82     for(mb_x=0; mb_x<=mb_w; mb_x++){
83         add_yblock(s, 1, sb, old_buffer, dst8, obmc,
84                    block_w*mb_x - block_w/2,
85                    block_w*mb_y - block_w/2,
86                    block_w, block_w,
87                    w, h,
88                    w, ref_stride, obmc_stride,
89                    mb_x - 1, mb_y - 1,
90                    add, 0, plane_index);
91     }
92 }
93
94 static inline void decode_subband_slice_buffered(SnowContext *s, SubBand *b, slice_buffer * sb, int start_y, int h, int save_state[1]){
95     const int w= b->width;
96     int y;
97     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
98     int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
99     int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
100     int new_index = 0;
101
102     if(b->ibuf == s->spatial_idwt_buffer || s->qlog == LOSSLESS_QLOG){
103         qadd= 0;
104         qmul= 1<<QEXPSHIFT;
105     }
106
107     /* If we are on the second or later slice, restore our index. */
108     if (start_y != 0)
109         new_index = save_state[0];
110
111
112     for(y=start_y; y<h; y++){
113         int x = 0;
114         int v;
115         IDWTELEM * line = slice_buffer_get_line(sb, y * b->stride_line + b->buf_y_offset) + b->buf_x_offset;
116         memset(line, 0, b->width*sizeof(IDWTELEM));
117         v = b->x_coeff[new_index].coeff;
118         x = b->x_coeff[new_index++].x;
119         while(x < w){
120             register int t= ( (v>>1)*qmul + qadd)>>QEXPSHIFT;
121             register int u= -(v&1);
122             line[x] = (t^u) - u;
123
124             v = b->x_coeff[new_index].coeff;
125             x = b->x_coeff[new_index++].x;
126         }
127     }
128
129     /* Save our variables for the next slice. */
130     save_state[0] = new_index;
131
132     return;
133 }
134
135 static void decode_q_branch(SnowContext *s, int level, int x, int y){
136     const int w= s->b_width << s->block_max_depth;
137     const int rem_depth= s->block_max_depth - level;
138     const int index= (x + y*w) << rem_depth;
139     int trx= (x+1)<<rem_depth;
140     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
141     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
142     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
143     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
144     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
145
146     if(s->keyframe){
147         set_blocks(s, level, x, y, null_block.color[0], null_block.color[1], null_block.color[2], null_block.mx, null_block.my, null_block.ref, BLOCK_INTRA);
148         return;
149     }
150
151     if(level==s->block_max_depth || get_rac(&s->c, &s->block_state[4 + s_context])){
152         int type, mx, my;
153         int l = left->color[0];
154         int cb= left->color[1];
155         int cr= left->color[2];
156         int ref = 0;
157         int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
158         int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 0*av_log2(2*FFABS(tr->mx - top->mx));
159         int my_context= av_log2(2*FFABS(left->my - top->my)) + 0*av_log2(2*FFABS(tr->my - top->my));
160
161         type= get_rac(&s->c, &s->block_state[1 + left->type + top->type]) ? BLOCK_INTRA : 0;
162
163         if(type){
164             pred_mv(s, &mx, &my, 0, left, top, tr);
165             l += get_symbol(&s->c, &s->block_state[32], 1);
166             cb+= get_symbol(&s->c, &s->block_state[64], 1);
167             cr+= get_symbol(&s->c, &s->block_state[96], 1);
168         }else{
169             if(s->ref_frames > 1)
170                 ref= get_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], 0);
171             pred_mv(s, &mx, &my, ref, left, top, tr);
172             mx+= get_symbol(&s->c, &s->block_state[128 + 32*(mx_context + 16*!!ref)], 1);
173             my+= get_symbol(&s->c, &s->block_state[128 + 32*(my_context + 16*!!ref)], 1);
174         }
175         set_blocks(s, level, x, y, l, cb, cr, mx, my, ref, type);
176     }else{
177         decode_q_branch(s, level+1, 2*x+0, 2*y+0);
178         decode_q_branch(s, level+1, 2*x+1, 2*y+0);
179         decode_q_branch(s, level+1, 2*x+0, 2*y+1);
180         decode_q_branch(s, level+1, 2*x+1, 2*y+1);
181     }
182 }
183
184 static void dequantize_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand *b, IDWTELEM *src, int stride, int start_y, int end_y){
185     const int w= b->width;
186     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
187     const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
188     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
189     int x,y;
190
191     if(s->qlog == LOSSLESS_QLOG) return;
192
193     for(y=start_y; y<end_y; y++){
194 //        DWTELEM * line = slice_buffer_get_line_from_address(sb, src + (y * stride));
195         IDWTELEM * line = slice_buffer_get_line(sb, (y * b->stride_line) + b->buf_y_offset) + b->buf_x_offset;
196         for(x=0; x<w; x++){
197             int i= line[x];
198             if(i<0){
199                 line[x]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
200             }else if(i>0){
201                 line[x]=  (( i*qmul + qadd)>>(QEXPSHIFT));
202             }
203         }
204     }
205 }
206
207 static void correlate_slice_buffered(SnowContext *s, slice_buffer * sb, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median, int start_y, int end_y){
208     const int w= b->width;
209     int x,y;
210
211     IDWTELEM * line=0; // silence silly "could be used without having been initialized" warning
212     IDWTELEM * prev;
213
214     if (start_y != 0)
215         line = slice_buffer_get_line(sb, ((start_y - 1) * b->stride_line) + b->buf_y_offset) + b->buf_x_offset;
216
217     for(y=start_y; y<end_y; y++){
218         prev = line;
219 //        line = slice_buffer_get_line_from_address(sb, src + (y * stride));
220         line = slice_buffer_get_line(sb, (y * b->stride_line) + b->buf_y_offset) + b->buf_x_offset;
221         for(x=0; x<w; x++){
222             if(x){
223                 if(use_median){
224                     if(y && x+1<w) line[x] += mid_pred(line[x - 1], prev[x], prev[x + 1]);
225                     else  line[x] += line[x - 1];
226                 }else{
227                     if(y) line[x] += mid_pred(line[x - 1], prev[x], line[x - 1] + prev[x] - prev[x - 1]);
228                     else  line[x] += line[x - 1];
229                 }
230             }else{
231                 if(y) line[x] += prev[x];
232             }
233         }
234     }
235 }
236
237 static void decode_qlogs(SnowContext *s){
238     int plane_index, level, orientation;
239
240     for(plane_index=0; plane_index<3; plane_index++){
241         for(level=0; level<s->spatial_decomposition_count; level++){
242             for(orientation=level ? 1:0; orientation<4; orientation++){
243                 int q;
244                 if     (plane_index==2) q= s->plane[1].band[level][orientation].qlog;
245                 else if(orientation==2) q= s->plane[plane_index].band[level][1].qlog;
246                 else                    q= get_symbol(&s->c, s->header_state, 1);
247                 s->plane[plane_index].band[level][orientation].qlog= q;
248             }
249         }
250     }
251 }
252
253 #define GET_S(dst, check) \
254     tmp= get_symbol(&s->c, s->header_state, 0);\
255     if(!(check)){\
256         av_log(s->avctx, AV_LOG_ERROR, "Error " #dst " is %d\n", tmp);\
257         return -1;\
258     }\
259     dst= tmp;
260
261 static int decode_header(SnowContext *s){
262     int plane_index, tmp;
263     uint8_t kstate[32];
264
265     memset(kstate, MID_STATE, sizeof(kstate));
266
267     s->keyframe= get_rac(&s->c, kstate);
268     if(s->keyframe || s->always_reset){
269         ff_snow_reset_contexts(s);
270         s->spatial_decomposition_type=
271         s->qlog=
272         s->qbias=
273         s->mv_scale=
274         s->block_max_depth= 0;
275     }
276     if(s->keyframe){
277         GET_S(s->version, tmp <= 0U)
278         s->always_reset= get_rac(&s->c, s->header_state);
279         s->temporal_decomposition_type= get_symbol(&s->c, s->header_state, 0);
280         s->temporal_decomposition_count= get_symbol(&s->c, s->header_state, 0);
281         GET_S(s->spatial_decomposition_count, 0 < tmp && tmp <= MAX_DECOMPOSITIONS)
282         s->colorspace_type= get_symbol(&s->c, s->header_state, 0);
283         s->chroma_h_shift= get_symbol(&s->c, s->header_state, 0);
284         s->chroma_v_shift= get_symbol(&s->c, s->header_state, 0);
285         s->spatial_scalability= get_rac(&s->c, s->header_state);
286 //        s->rate_scalability= get_rac(&s->c, s->header_state);
287         GET_S(s->max_ref_frames, tmp < (unsigned)MAX_REF_FRAMES)
288         s->max_ref_frames++;
289
290         decode_qlogs(s);
291     }
292
293     if(!s->keyframe){
294         if(get_rac(&s->c, s->header_state)){
295             for(plane_index=0; plane_index<2; plane_index++){
296                 int htaps, i, sum=0;
297                 Plane *p= &s->plane[plane_index];
298                 p->diag_mc= get_rac(&s->c, s->header_state);
299                 htaps= get_symbol(&s->c, s->header_state, 0)*2 + 2;
300                 if((unsigned)htaps > HTAPS_MAX || htaps==0)
301                     return -1;
302                 p->htaps= htaps;
303                 for(i= htaps/2; i; i--){
304                     p->hcoeff[i]= get_symbol(&s->c, s->header_state, 0) * (1-2*(i&1));
305                     sum += p->hcoeff[i];
306                 }
307                 p->hcoeff[0]= 32-sum;
308             }
309             s->plane[2].diag_mc= s->plane[1].diag_mc;
310             s->plane[2].htaps  = s->plane[1].htaps;
311             memcpy(s->plane[2].hcoeff, s->plane[1].hcoeff, sizeof(s->plane[1].hcoeff));
312         }
313         if(get_rac(&s->c, s->header_state)){
314             GET_S(s->spatial_decomposition_count, 0 < tmp && tmp <= MAX_DECOMPOSITIONS)
315             decode_qlogs(s);
316         }
317     }
318
319     s->spatial_decomposition_type+= get_symbol(&s->c, s->header_state, 1);
320     if(s->spatial_decomposition_type > 1U){
321         av_log(s->avctx, AV_LOG_ERROR, "spatial_decomposition_type %d not supported", s->spatial_decomposition_type);
322         return -1;
323     }
324     if(FFMIN(s->avctx-> width>>s->chroma_h_shift,
325              s->avctx->height>>s->chroma_v_shift) >> (s->spatial_decomposition_count-1) <= 0){
326         av_log(s->avctx, AV_LOG_ERROR, "spatial_decomposition_count %d too large for size", s->spatial_decomposition_count);
327         return -1;
328     }
329
330     s->qlog           += get_symbol(&s->c, s->header_state, 1);
331     s->mv_scale       += get_symbol(&s->c, s->header_state, 1);
332     s->qbias          += get_symbol(&s->c, s->header_state, 1);
333     s->block_max_depth+= get_symbol(&s->c, s->header_state, 1);
334     if(s->block_max_depth > 1 || s->block_max_depth < 0){
335         av_log(s->avctx, AV_LOG_ERROR, "block_max_depth= %d is too large", s->block_max_depth);
336         s->block_max_depth= 0;
337         return -1;
338     }
339
340     return 0;
341 }
342
343 static av_cold int decode_init(AVCodecContext *avctx)
344 {
345     avctx->pix_fmt= PIX_FMT_YUV420P;
346
347     ff_snow_common_init(avctx);
348
349     return 0;
350 }
351
352 static void decode_blocks(SnowContext *s){
353     int x, y;
354     int w= s->b_width;
355     int h= s->b_height;
356
357     for(y=0; y<h; y++){
358         for(x=0; x<w; x++){
359             decode_q_branch(s, 0, x, y);
360         }
361     }
362 }
363
364 static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, AVPacket *avpkt){
365     const uint8_t *buf = avpkt->data;
366     int buf_size = avpkt->size;
367     SnowContext *s = avctx->priv_data;
368     RangeCoder * const c= &s->c;
369     int bytes_read;
370     AVFrame *picture = data;
371     int level, orientation, plane_index;
372
373     ff_init_range_decoder(c, buf, buf_size);
374     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
375
376     s->current_picture.pict_type= AV_PICTURE_TYPE_I; //FIXME I vs. P
377     if(decode_header(s)<0)
378         return -1;
379     ff_snow_common_init_after_header(avctx);
380
381     // realloc slice buffer for the case that spatial_decomposition_count changed
382     ff_slice_buffer_destroy(&s->sb);
383     ff_slice_buffer_init(&s->sb, s->plane[0].height, (MB_SIZE >> s->block_max_depth) + s->spatial_decomposition_count * 8 + 1, s->plane[0].width, s->spatial_idwt_buffer);
384
385     for(plane_index=0; plane_index<3; plane_index++){
386         Plane *p= &s->plane[plane_index];
387         p->fast_mc= p->diag_mc && p->htaps==6 && p->hcoeff[0]==40
388                                               && p->hcoeff[1]==-10
389                                               && p->hcoeff[2]==2;
390     }
391
392     ff_snow_alloc_blocks(s);
393
394     if(ff_snow_frame_start(s) < 0)
395         return -1;
396     //keyframe flag duplication mess FIXME
397     if(avctx->debug&FF_DEBUG_PICT_INFO)
398         av_log(avctx, AV_LOG_ERROR, "keyframe:%d qlog:%d\n", s->keyframe, s->qlog);
399
400     decode_blocks(s);
401
402     for(plane_index=0; plane_index<3; plane_index++){
403         Plane *p= &s->plane[plane_index];
404         int w= p->width;
405         int h= p->height;
406         int x, y;
407         int decode_state[MAX_DECOMPOSITIONS][4][1]; /* Stored state info for unpack_coeffs. 1 variable per instance. */
408
409         if(s->avctx->debug&2048){
410             memset(s->spatial_dwt_buffer, 0, sizeof(DWTELEM)*w*h);
411             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
412
413             for(y=0; y<h; y++){
414                 for(x=0; x<w; x++){
415                     int v= s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x];
416                     s->mconly_picture.data[plane_index][y*s->mconly_picture.linesize[plane_index] + x]= v;
417                 }
418             }
419         }
420
421         {
422         for(level=0; level<s->spatial_decomposition_count; level++){
423             for(orientation=level ? 1 : 0; orientation<4; orientation++){
424                 SubBand *b= &p->band[level][orientation];
425                 unpack_coeffs(s, b, b->parent, orientation);
426             }
427         }
428         }
429
430         {
431         const int mb_h= s->b_height << s->block_max_depth;
432         const int block_size = MB_SIZE >> s->block_max_depth;
433         const int block_w    = plane_index ? block_size/2 : block_size;
434         int mb_y;
435         DWTCompose cs[MAX_DECOMPOSITIONS];
436         int yd=0, yq=0;
437         int y;
438         int end_y;
439
440         ff_spatial_idwt_buffered_init(cs, &s->sb, w, h, 1, s->spatial_decomposition_type, s->spatial_decomposition_count);
441         for(mb_y=0; mb_y<=mb_h; mb_y++){
442
443             int slice_starty = block_w*mb_y;
444             int slice_h = block_w*(mb_y+1);
445             if (!(s->keyframe || s->avctx->debug&512)){
446                 slice_starty = FFMAX(0, slice_starty - (block_w >> 1));
447                 slice_h -= (block_w >> 1);
448             }
449
450             for(level=0; level<s->spatial_decomposition_count; level++){
451                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
452                     SubBand *b= &p->band[level][orientation];
453                     int start_y;
454                     int end_y;
455                     int our_mb_start = mb_y;
456                     int our_mb_end = (mb_y + 1);
457                     const int extra= 3;
458                     start_y = (mb_y ? ((block_w * our_mb_start) >> (s->spatial_decomposition_count - level)) + s->spatial_decomposition_count - level + extra: 0);
459                     end_y = (((block_w * our_mb_end) >> (s->spatial_decomposition_count - level)) + s->spatial_decomposition_count - level + extra);
460                     if (!(s->keyframe || s->avctx->debug&512)){
461                         start_y = FFMAX(0, start_y - (block_w >> (1+s->spatial_decomposition_count - level)));
462                         end_y = FFMAX(0, end_y - (block_w >> (1+s->spatial_decomposition_count - level)));
463                     }
464                     start_y = FFMIN(b->height, start_y);
465                     end_y = FFMIN(b->height, end_y);
466
467                     if (start_y != end_y){
468                         if (orientation == 0){
469                             SubBand * correlate_band = &p->band[0][0];
470                             int correlate_end_y = FFMIN(b->height, end_y + 1);
471                             int correlate_start_y = FFMIN(b->height, (start_y ? start_y + 1 : 0));
472                             decode_subband_slice_buffered(s, correlate_band, &s->sb, correlate_start_y, correlate_end_y, decode_state[0][0]);
473                             correlate_slice_buffered(s, &s->sb, correlate_band, correlate_band->ibuf, correlate_band->stride, 1, 0, correlate_start_y, correlate_end_y);
474                             dequantize_slice_buffered(s, &s->sb, correlate_band, correlate_band->ibuf, correlate_band->stride, start_y, end_y);
475                         }
476                         else
477                             decode_subband_slice_buffered(s, b, &s->sb, start_y, end_y, decode_state[level][orientation]);
478                     }
479                 }
480             }
481
482             for(; yd<slice_h; yd+=4){
483                 ff_spatial_idwt_buffered_slice(&s->dwt, cs, &s->sb, w, h, 1, s->spatial_decomposition_type, s->spatial_decomposition_count, yd);
484             }
485
486             if(s->qlog == LOSSLESS_QLOG){
487                 for(; yq<slice_h && yq<h; yq++){
488                     IDWTELEM * line = slice_buffer_get_line(&s->sb, yq);
489                     for(x=0; x<w; x++){
490                         line[x] <<= FRAC_BITS;
491                     }
492                 }
493             }
494
495             predict_slice_buffered(s, &s->sb, s->spatial_idwt_buffer, plane_index, 1, mb_y);
496
497             y = FFMIN(p->height, slice_starty);
498             end_y = FFMIN(p->height, slice_h);
499             while(y < end_y)
500                 ff_slice_buffer_release(&s->sb, y++);
501         }
502
503         ff_slice_buffer_flush(&s->sb);
504         }
505
506     }
507
508     emms_c();
509
510     ff_snow_release_buffer(avctx);
511
512     if(!(s->avctx->debug&2048))
513         *picture= s->current_picture;
514     else
515         *picture= s->mconly_picture;
516
517     *data_size = sizeof(AVFrame);
518
519     bytes_read= c->bytestream - c->bytestream_start;
520     if(bytes_read ==0) av_log(s->avctx, AV_LOG_ERROR, "error at end of frame\n"); //FIXME
521
522     return bytes_read;
523 }
524
525 static av_cold int decode_end(AVCodecContext *avctx)
526 {
527     SnowContext *s = avctx->priv_data;
528
529     ff_slice_buffer_destroy(&s->sb);
530
531     ff_snow_common_end(s);
532
533     return 0;
534 }
535
536 AVCodec ff_snow_decoder = {
537     .name           = "snow",
538     .type           = AVMEDIA_TYPE_VIDEO,
539     .id             = CODEC_ID_SNOW,
540     .priv_data_size = sizeof(SnowContext),
541     .init           = decode_init,
542     .close          = decode_end,
543     .decode         = decode_frame,
544     .capabilities   = CODEC_CAP_DR1 /*| CODEC_CAP_DRAW_HORIZ_BAND*/,
545     .long_name = NULL_IF_CONFIG_SMALL("Snow"),
546 };