9c804acbd4e4566cedb16272b751919c890d79f3
[ffmpeg.git] / libavcodec / snowenc.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 "internal.h"
26 #include "snow_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 static av_cold int encode_init(AVCodecContext *avctx)
36 {
37     SnowContext *s = avctx->priv_data;
38     int plane_index, ret;
39     int i;
40
41     if(avctx->prediction_method == DWT_97
42        && (avctx->flags & CODEC_FLAG_QSCALE)
43        && avctx->global_quality == 0){
44         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
45         return -1;
46     }
47
48     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
49
50     s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
51     s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
52
53     for(plane_index=0; plane_index<3; plane_index++){
54         s->plane[plane_index].diag_mc= 1;
55         s->plane[plane_index].htaps= 6;
56         s->plane[plane_index].hcoeff[0]=  40;
57         s->plane[plane_index].hcoeff[1]= -10;
58         s->plane[plane_index].hcoeff[2]=   2;
59         s->plane[plane_index].fast_mc= 1;
60     }
61
62     if ((ret = ff_snow_common_init(avctx)) < 0) {
63         return ret;
64     }
65     ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
66
67     ff_snow_alloc_blocks(s);
68
69     s->version=0;
70
71     s->m.avctx   = avctx;
72     s->m.bit_rate= avctx->bit_rate;
73
74     s->m.me.temp      =
75     s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
76     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
77     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
78     s->m.sc.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
79     if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.sc.obmc_scratchpad)
80         return AVERROR(ENOMEM);
81
82     ff_h263_encode_init(&s->m); //mv_penalty
83
84     s->max_ref_frames = av_clip(avctx->refs, 1, MAX_REF_FRAMES);
85
86     if(avctx->flags&CODEC_FLAG_PASS1){
87         if(!avctx->stats_out)
88             avctx->stats_out = av_mallocz(256);
89
90         if (!avctx->stats_out)
91             return AVERROR(ENOMEM);
92     }
93     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
94         if(ff_rate_control_init(&s->m) < 0)
95             return -1;
96     }
97     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
98
99     switch(avctx->pix_fmt){
100     case AV_PIX_FMT_YUV444P:
101 //    case AV_PIX_FMT_YUV422P:
102     case AV_PIX_FMT_YUV420P:
103 //    case AV_PIX_FMT_YUV411P:
104     case AV_PIX_FMT_YUV410P:
105         s->nb_planes = 3;
106         s->colorspace_type= 0;
107         break;
108     case AV_PIX_FMT_GRAY8:
109         s->nb_planes = 1;
110         s->colorspace_type = 1;
111         break;
112 /*    case AV_PIX_FMT_RGB32:
113         s->colorspace= 1;
114         break;*/
115     default:
116         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
117         return -1;
118     }
119     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
120
121     ff_set_cmp(&s->mecc, s->mecc.me_cmp, s->avctx->me_cmp);
122     ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, s->avctx->me_sub_cmp);
123
124     s->input_picture = av_frame_alloc();
125     if (!s->input_picture)
126         return AVERROR(ENOMEM);
127
128     if ((ret = ff_snow_get_buffer(s, s->input_picture)) < 0)
129         return ret;
130
131     if(s->avctx->me_method == ME_ITER){
132         int size= s->b_width * s->b_height << 2*s->block_max_depth;
133         for(i=0; i<s->max_ref_frames; i++){
134             s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
135             s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
136             if (!s->ref_mvs[i] || !s->ref_scores[i])
137                 return AVERROR(ENOMEM);
138         }
139     }
140
141     return 0;
142 }
143
144 //near copy & paste from dsputil, FIXME
145 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
146 {
147     int s, i, j;
148
149     s = 0;
150     for (i = 0; i < h; i++) {
151         for (j = 0; j < w; j++) {
152             s += pix[0];
153             pix ++;
154         }
155         pix += line_size - w;
156     }
157     return s;
158 }
159
160 //near copy & paste from dsputil, FIXME
161 static int pix_norm1(uint8_t * pix, int line_size, int w)
162 {
163     int s, i, j;
164     uint32_t *sq = ff_square_tab + 256;
165
166     s = 0;
167     for (i = 0; i < w; i++) {
168         for (j = 0; j < w; j ++) {
169             s += sq[pix[0]];
170             pix ++;
171         }
172         pix += line_size - w;
173     }
174     return s;
175 }
176
177 static inline int get_penalty_factor(int lambda, int lambda2, int type){
178     switch(type&0xFF){
179     default:
180     case FF_CMP_SAD:
181         return lambda>>FF_LAMBDA_SHIFT;
182     case FF_CMP_DCT:
183         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
184     case FF_CMP_W53:
185         return (4*lambda)>>(FF_LAMBDA_SHIFT);
186     case FF_CMP_W97:
187         return (2*lambda)>>(FF_LAMBDA_SHIFT);
188     case FF_CMP_SATD:
189     case FF_CMP_DCT264:
190         return (2*lambda)>>FF_LAMBDA_SHIFT;
191     case FF_CMP_RD:
192     case FF_CMP_PSNR:
193     case FF_CMP_SSE:
194     case FF_CMP_NSSE:
195         return lambda2>>FF_LAMBDA_SHIFT;
196     case FF_CMP_BIT:
197         return 1;
198     }
199 }
200
201 //FIXME copy&paste
202 #define P_LEFT P[1]
203 #define P_TOP P[2]
204 #define P_TOPRIGHT P[3]
205 #define P_MEDIAN P[4]
206 #define P_MV1 P[9]
207 #define FLAG_QPEL   1 //must be 1
208
209 static int encode_q_branch(SnowContext *s, int level, int x, int y){
210     uint8_t p_buffer[1024];
211     uint8_t i_buffer[1024];
212     uint8_t p_state[sizeof(s->block_state)];
213     uint8_t i_state[sizeof(s->block_state)];
214     RangeCoder pc, ic;
215     uint8_t *pbbak= s->c.bytestream;
216     uint8_t *pbbak_start= s->c.bytestream_start;
217     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
218     const int w= s->b_width  << s->block_max_depth;
219     const int h= s->b_height << s->block_max_depth;
220     const int rem_depth= s->block_max_depth - level;
221     const int index= (x + y*w) << rem_depth;
222     const int block_w= 1<<(LOG2_MB_SIZE - level);
223     int trx= (x+1)<<rem_depth;
224     int try= (y+1)<<rem_depth;
225     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
226     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
227     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
228     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
229     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
230     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
231     int pl = left->color[0];
232     int pcb= left->color[1];
233     int pcr= left->color[2];
234     int pmx, pmy;
235     int mx=0, my=0;
236     int l,cr,cb;
237     const int stride= s->current_picture->linesize[0];
238     const int uvstride= s->current_picture->linesize[1];
239     uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y*  stride)*block_w,
240                                 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
241                                 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
242     int P[10][2];
243     int16_t last_mv[3][2];
244     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
245     const int shift= 1+qpel;
246     MotionEstContext *c= &s->m.me;
247     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
248     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
249     int my_context= av_log2(2*FFABS(left->my - top->my));
250     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
251     int ref, best_ref, ref_score, ref_mx, ref_my;
252
253     av_assert0(sizeof(s->block_state) >= 256);
254     if(s->keyframe){
255         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
256         return 0;
257     }
258
259 //    clip predictors / edge ?
260
261     P_LEFT[0]= left->mx;
262     P_LEFT[1]= left->my;
263     P_TOP [0]= top->mx;
264     P_TOP [1]= top->my;
265     P_TOPRIGHT[0]= tr->mx;
266     P_TOPRIGHT[1]= tr->my;
267
268     last_mv[0][0]= s->block[index].mx;
269     last_mv[0][1]= s->block[index].my;
270     last_mv[1][0]= right->mx;
271     last_mv[1][1]= right->my;
272     last_mv[2][0]= bottom->mx;
273     last_mv[2][1]= bottom->my;
274
275     s->m.mb_stride=2;
276     s->m.mb_x=
277     s->m.mb_y= 0;
278     c->skip= 0;
279
280     av_assert1(c->  stride ==   stride);
281     av_assert1(c->uvstride == uvstride);
282
283     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
284     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
285     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
286     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
287
288     c->xmin = - x*block_w - 16+3;
289     c->ymin = - y*block_w - 16+3;
290     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
291     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
292
293     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
294     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
295     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
296     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
297     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
298     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
299     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
300
301     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
302     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
303
304     if (!y) {
305         c->pred_x= P_LEFT[0];
306         c->pred_y= P_LEFT[1];
307     } else {
308         c->pred_x = P_MEDIAN[0];
309         c->pred_y = P_MEDIAN[1];
310     }
311
312     score= INT_MAX;
313     best_ref= 0;
314     for(ref=0; ref<s->ref_frames; ref++){
315         init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
316
317         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
318                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
319
320         av_assert2(ref_mx >= c->xmin);
321         av_assert2(ref_mx <= c->xmax);
322         av_assert2(ref_my >= c->ymin);
323         av_assert2(ref_my <= c->ymax);
324
325         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
326         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
327         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
328         if(s->ref_mvs[ref]){
329             s->ref_mvs[ref][index][0]= ref_mx;
330             s->ref_mvs[ref][index][1]= ref_my;
331             s->ref_scores[ref][index]= ref_score;
332         }
333         if(score > ref_score){
334             score= ref_score;
335             best_ref= ref;
336             mx= ref_mx;
337             my= ref_my;
338         }
339     }
340     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
341
342   //  subpel search
343     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
344     pc= s->c;
345     pc.bytestream_start=
346     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
347     memcpy(p_state, s->block_state, sizeof(s->block_state));
348
349     if(level!=s->block_max_depth)
350         put_rac(&pc, &p_state[4 + s_context], 1);
351     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
352     if(s->ref_frames > 1)
353         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
354     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
355     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
356     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
357     p_len= pc.bytestream - pc.bytestream_start;
358     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
359
360     block_s= block_w*block_w;
361     sum = pix_sum(current_data[0], stride, block_w, block_w);
362     l= (sum + block_s/2)/block_s;
363     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
364
365     if (s->nb_planes > 2) {
366         block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
367         sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
368         cb= (sum + block_s/2)/block_s;
369     //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
370         sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
371         cr= (sum + block_s/2)/block_s;
372     //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
373     }else
374         cb = cr = 0;
375
376     ic= s->c;
377     ic.bytestream_start=
378     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
379     memcpy(i_state, s->block_state, sizeof(s->block_state));
380     if(level!=s->block_max_depth)
381         put_rac(&ic, &i_state[4 + s_context], 1);
382     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
383     put_symbol(&ic, &i_state[32],  l-pl , 1);
384     if (s->nb_planes > 2) {
385         put_symbol(&ic, &i_state[64], cb-pcb, 1);
386         put_symbol(&ic, &i_state[96], cr-pcr, 1);
387     }
388     i_len= ic.bytestream - ic.bytestream_start;
389     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
390
391     av_assert1(iscore < 255*255*256 + s->lambda2*10);
392     av_assert1(iscore >= 0);
393     av_assert1(l>=0 && l<=255);
394     av_assert1(pl>=0 && pl<=255);
395
396     if(level==0){
397         int varc= iscore >> 8;
398         int vard= score >> 8;
399         if (vard <= 64 || vard < varc)
400             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
401         else
402             c->scene_change_score+= s->m.qscale;
403     }
404
405     if(level!=s->block_max_depth){
406         put_rac(&s->c, &s->block_state[4 + s_context], 0);
407         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
408         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
409         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
410         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
411         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
412
413         if(score2 < score && score2 < iscore)
414             return score2;
415     }
416
417     if(iscore < score){
418         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
419         memcpy(pbbak, i_buffer, i_len);
420         s->c= ic;
421         s->c.bytestream_start= pbbak_start;
422         s->c.bytestream= pbbak + i_len;
423         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
424         memcpy(s->block_state, i_state, sizeof(s->block_state));
425         return iscore;
426     }else{
427         memcpy(pbbak, p_buffer, p_len);
428         s->c= pc;
429         s->c.bytestream_start= pbbak_start;
430         s->c.bytestream= pbbak + p_len;
431         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
432         memcpy(s->block_state, p_state, sizeof(s->block_state));
433         return score;
434     }
435 }
436
437 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
438     const int w= s->b_width  << s->block_max_depth;
439     const int rem_depth= s->block_max_depth - level;
440     const int index= (x + y*w) << rem_depth;
441     int trx= (x+1)<<rem_depth;
442     BlockNode *b= &s->block[index];
443     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
444     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
445     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
446     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
447     int pl = left->color[0];
448     int pcb= left->color[1];
449     int pcr= left->color[2];
450     int pmx, pmy;
451     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
452     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
453     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
454     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
455
456     if(s->keyframe){
457         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
458         return;
459     }
460
461     if(level!=s->block_max_depth){
462         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
463             put_rac(&s->c, &s->block_state[4 + s_context], 1);
464         }else{
465             put_rac(&s->c, &s->block_state[4 + s_context], 0);
466             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
467             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
468             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
469             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
470             return;
471         }
472     }
473     if(b->type & BLOCK_INTRA){
474         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
475         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
476         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
477         if (s->nb_planes > 2) {
478             put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
479             put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
480         }
481         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
482     }else{
483         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
484         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
485         if(s->ref_frames > 1)
486             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
487         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
488         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
489         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
490     }
491 }
492
493 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
494     int i, x2, y2;
495     Plane *p= &s->plane[plane_index];
496     const int block_size = MB_SIZE >> s->block_max_depth;
497     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
498     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
499     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
500     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
501     const int ref_stride= s->current_picture->linesize[plane_index];
502     uint8_t *src= s-> input_picture->data[plane_index];
503     IDWTELEM *dst= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
504     const int b_stride = s->b_width << s->block_max_depth;
505     const int w= p->width;
506     const int h= p->height;
507     int index= mb_x + mb_y*b_stride;
508     BlockNode *b= &s->block[index];
509     BlockNode backup= *b;
510     int ab=0;
511     int aa=0;
512
513     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
514
515     b->type|= BLOCK_INTRA;
516     b->color[plane_index]= 0;
517     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
518
519     for(i=0; i<4; i++){
520         int mb_x2= mb_x + (i &1) - 1;
521         int mb_y2= mb_y + (i>>1) - 1;
522         int x= block_w*mb_x2 + block_w/2;
523         int y= block_h*mb_y2 + block_h/2;
524
525         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
526                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
527
528         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
529             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
530                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
531                 int obmc_v= obmc[index];
532                 int d;
533                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
534                 if(x<0) obmc_v += obmc[index + block_w];
535                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
536                 if(x+block_w>w) obmc_v += obmc[index - block_w];
537                 //FIXME precalculate this or simplify it somehow else
538
539                 d = -dst[index] + (1<<(FRAC_BITS-1));
540                 dst[index] = d;
541                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
542                 aa += obmc_v * obmc_v; //FIXME precalculate this
543             }
544         }
545     }
546     *b= backup;
547
548     return av_clip_uint8( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa) ); //FIXME we should not need clipping
549 }
550
551 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
552     const int b_stride = s->b_width << s->block_max_depth;
553     const int b_height = s->b_height<< s->block_max_depth;
554     int index= x + y*b_stride;
555     const BlockNode *b     = &s->block[index];
556     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
557     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
558     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
559     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
560     int dmx, dmy;
561 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
562 //  int my_context= av_log2(2*FFABS(left->my - top->my));
563
564     if(x<0 || x>=b_stride || y>=b_height)
565         return 0;
566 /*
567 1            0      0
568 01X          1-2    1
569 001XX        3-6    2-3
570 0001XXX      7-14   4-7
571 00001XXXX   15-30   8-15
572 */
573 //FIXME try accurate rate
574 //FIXME intra and inter predictors if surrounding blocks are not the same type
575     if(b->type & BLOCK_INTRA){
576         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
577                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
578                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
579     }else{
580         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
581         dmx-= b->mx;
582         dmy-= b->my;
583         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
584                     + av_log2(2*FFABS(dmy))
585                     + av_log2(2*b->ref));
586     }
587 }
588
589 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
590     Plane *p= &s->plane[plane_index];
591     const int block_size = MB_SIZE >> s->block_max_depth;
592     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
593     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
594     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
595     const int ref_stride= s->current_picture->linesize[plane_index];
596     uint8_t *dst= s->current_picture->data[plane_index];
597     uint8_t *src= s->  input_picture->data[plane_index];
598     IDWTELEM *pred= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4;
599     uint8_t *cur = s->scratchbuf;
600     uint8_t *tmp = s->emu_edge_buffer;
601     const int b_stride = s->b_width << s->block_max_depth;
602     const int b_height = s->b_height<< s->block_max_depth;
603     const int w= p->width;
604     const int h= p->height;
605     int distortion;
606     int rate= 0;
607     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
608     int sx= block_w*mb_x - block_w/2;
609     int sy= block_h*mb_y - block_h/2;
610     int x0= FFMAX(0,-sx);
611     int y0= FFMAX(0,-sy);
612     int x1= FFMIN(block_w*2, w-sx);
613     int y1= FFMIN(block_h*2, h-sy);
614     int i,x,y;
615
616     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
617
618     ff_snow_pred_block(s, cur, tmp, ref_stride, sx, sy, block_w*2, block_h*2, &s->block[mb_x + mb_y*b_stride], plane_index, w, h);
619
620     for(y=y0; y<y1; y++){
621         const uint8_t *obmc1= obmc_edged[y];
622         const IDWTELEM *pred1 = pred + y*obmc_stride;
623         uint8_t *cur1 = cur + y*ref_stride;
624         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
625         for(x=x0; x<x1; x++){
626 #if FRAC_BITS >= LOG2_OBMC_MAX
627             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
628 #else
629             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
630 #endif
631             v = (v + pred1[x]) >> FRAC_BITS;
632             if(v&(~255)) v= ~(v>>31);
633             dst1[x] = v;
634         }
635     }
636
637     /* copy the regions where obmc[] = (uint8_t)256 */
638     if(LOG2_OBMC_MAX == 8
639         && (mb_x == 0 || mb_x == b_stride-1)
640         && (mb_y == 0 || mb_y == b_height-1)){
641         if(mb_x == 0)
642             x1 = block_w;
643         else
644             x0 = block_w;
645         if(mb_y == 0)
646             y1 = block_h;
647         else
648             y0 = block_h;
649         for(y=y0; y<y1; y++)
650             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
651     }
652
653     if(block_w==16){
654         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
655         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
656         /* FIXME cmps overlap but do not cover the wavelet's whole support.
657          * So improving the score of one block is not strictly guaranteed
658          * to improve the score of the whole frame, thus iterative motion
659          * estimation does not always converge. */
660         if(s->avctx->me_cmp == FF_CMP_W97)
661             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
662         else if(s->avctx->me_cmp == FF_CMP_W53)
663             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
664         else{
665             distortion = 0;
666             for(i=0; i<4; i++){
667                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
668                 distortion += s->mecc.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
669             }
670         }
671     }else{
672         av_assert2(block_w==8);
673         distortion = s->mecc.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
674     }
675
676     if(plane_index==0){
677         for(i=0; i<4; i++){
678 /* ..RRr
679  * .RXx.
680  * rxx..
681  */
682             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
683         }
684         if(mb_x == b_stride-2)
685             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
686     }
687     return distortion + rate*penalty_factor;
688 }
689
690 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
691     int i, y2;
692     Plane *p= &s->plane[plane_index];
693     const int block_size = MB_SIZE >> s->block_max_depth;
694     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
695     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
696     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
697     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
698     const int ref_stride= s->current_picture->linesize[plane_index];
699     uint8_t *dst= s->current_picture->data[plane_index];
700     uint8_t *src= s-> input_picture->data[plane_index];
701     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
702     // const has only been removed from zero_dst to suppress a warning
703     static IDWTELEM zero_dst[4096]; //FIXME
704     const int b_stride = s->b_width << s->block_max_depth;
705     const int w= p->width;
706     const int h= p->height;
707     int distortion= 0;
708     int rate= 0;
709     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
710
711     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
712
713     for(i=0; i<9; i++){
714         int mb_x2= mb_x + (i%3) - 1;
715         int mb_y2= mb_y + (i/3) - 1;
716         int x= block_w*mb_x2 + block_w/2;
717         int y= block_h*mb_y2 + block_h/2;
718
719         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
720                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
721
722         //FIXME find a cleaner/simpler way to skip the outside stuff
723         for(y2= y; y2<0; y2++)
724             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
725         for(y2= h; y2<y+block_h; y2++)
726             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
727         if(x<0){
728             for(y2= y; y2<y+block_h; y2++)
729                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
730         }
731         if(x+block_w > w){
732             for(y2= y; y2<y+block_h; y2++)
733                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
734         }
735
736         av_assert1(block_w== 8 || block_w==16);
737         distortion += s->mecc.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
738     }
739
740     if(plane_index==0){
741         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
742         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
743
744 /* ..RRRr
745  * .RXXx.
746  * .RXXx.
747  * rxxx.
748  */
749         if(merged)
750             rate = get_block_bits(s, mb_x, mb_y, 2);
751         for(i=merged?4:0; i<9; i++){
752             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
753             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
754         }
755     }
756     return distortion + rate*penalty_factor;
757 }
758
759 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
760     const int w= b->width;
761     const int h= b->height;
762     int x, y;
763
764     if(1){
765         int run=0;
766         int *runs = s->run_buffer;
767         int run_index=0;
768         int max_index;
769
770         for(y=0; y<h; y++){
771             for(x=0; x<w; x++){
772                 int v, p=0;
773                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
774                 v= src[x + y*stride];
775
776                 if(y){
777                     t= src[x + (y-1)*stride];
778                     if(x){
779                         lt= src[x - 1 + (y-1)*stride];
780                     }
781                     if(x + 1 < w){
782                         rt= src[x + 1 + (y-1)*stride];
783                     }
784                 }
785                 if(x){
786                     l= src[x - 1 + y*stride];
787                     /*if(x > 1){
788                         if(orientation==1) ll= src[y + (x-2)*stride];
789                         else               ll= src[x - 2 + y*stride];
790                     }*/
791                 }
792                 if(parent){
793                     int px= x>>1;
794                     int py= y>>1;
795                     if(px<b->parent->width && py<b->parent->height)
796                         p= parent[px + py*2*stride];
797                 }
798                 if(!(/*ll|*/l|lt|t|rt|p)){
799                     if(v){
800                         runs[run_index++]= run;
801                         run=0;
802                     }else{
803                         run++;
804                     }
805                 }
806             }
807         }
808         max_index= run_index;
809         runs[run_index++]= run;
810         run_index=0;
811         run= runs[run_index++];
812
813         put_symbol2(&s->c, b->state[30], max_index, 0);
814         if(run_index <= max_index)
815             put_symbol2(&s->c, b->state[1], run, 3);
816
817         for(y=0; y<h; y++){
818             if(s->c.bytestream_end - s->c.bytestream < w*40){
819                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
820                 return -1;
821             }
822             for(x=0; x<w; x++){
823                 int v, p=0;
824                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
825                 v= src[x + y*stride];
826
827                 if(y){
828                     t= src[x + (y-1)*stride];
829                     if(x){
830                         lt= src[x - 1 + (y-1)*stride];
831                     }
832                     if(x + 1 < w){
833                         rt= src[x + 1 + (y-1)*stride];
834                     }
835                 }
836                 if(x){
837                     l= src[x - 1 + y*stride];
838                     /*if(x > 1){
839                         if(orientation==1) ll= src[y + (x-2)*stride];
840                         else               ll= src[x - 2 + y*stride];
841                     }*/
842                 }
843                 if(parent){
844                     int px= x>>1;
845                     int py= y>>1;
846                     if(px<b->parent->width && py<b->parent->height)
847                         p= parent[px + py*2*stride];
848                 }
849                 if(/*ll|*/l|lt|t|rt|p){
850                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
851
852                     put_rac(&s->c, &b->state[0][context], !!v);
853                 }else{
854                     if(!run){
855                         run= runs[run_index++];
856
857                         if(run_index <= max_index)
858                             put_symbol2(&s->c, b->state[1], run, 3);
859                         av_assert2(v);
860                     }else{
861                         run--;
862                         av_assert2(!v);
863                     }
864                 }
865                 if(v){
866                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
867                     int l2= 2*FFABS(l) + (l<0);
868                     int t2= 2*FFABS(t) + (t<0);
869
870                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
871                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
872                 }
873             }
874         }
875     }
876     return 0;
877 }
878
879 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
880 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
881 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
882     return encode_subband_c0run(s, b, src, parent, stride, orientation);
883 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
884 }
885
886 static av_always_inline int check_block(SnowContext *s, int mb_x, int mb_y, int p[3], int intra, uint8_t (*obmc_edged)[MB_SIZE * 2], int *best_rd){
887     const int b_stride= s->b_width << s->block_max_depth;
888     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
889     BlockNode backup= *block;
890     unsigned value;
891     int rd, index;
892
893     av_assert2(mb_x>=0 && mb_y>=0);
894     av_assert2(mb_x<b_stride);
895
896     if(intra){
897         block->color[0] = p[0];
898         block->color[1] = p[1];
899         block->color[2] = p[2];
900         block->type |= BLOCK_INTRA;
901     }else{
902         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
903         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
904         if(s->me_cache[index] == value)
905             return 0;
906         s->me_cache[index]= value;
907
908         block->mx= p[0];
909         block->my= p[1];
910         block->type &= ~BLOCK_INTRA;
911     }
912
913     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged) + s->intra_penalty * !!intra;
914
915 //FIXME chroma
916     if(rd < *best_rd){
917         *best_rd= rd;
918         return 1;
919     }else{
920         *block= backup;
921         return 0;
922     }
923 }
924
925 /* special case for int[2] args we discard afterwards,
926  * fixes compilation problem with gcc 2.95 */
927 static av_always_inline int check_block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, uint8_t (*obmc_edged)[MB_SIZE * 2], int *best_rd){
928     int p[2] = {p0, p1};
929     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
930 }
931
932 static av_always_inline int check_4block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, int ref, int *best_rd){
933     const int b_stride= s->b_width << s->block_max_depth;
934     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
935     BlockNode backup[4];
936     unsigned value;
937     int rd, index;
938
939     /* We don't initialize backup[] during variable declaration, because
940      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
941      * 'int16_t'". */
942     backup[0] = block[0];
943     backup[1] = block[1];
944     backup[2] = block[b_stride];
945     backup[3] = block[b_stride + 1];
946
947     av_assert2(mb_x>=0 && mb_y>=0);
948     av_assert2(mb_x<b_stride);
949     av_assert2(((mb_x|mb_y)&1) == 0);
950
951     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
952     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
953     if(s->me_cache[index] == value)
954         return 0;
955     s->me_cache[index]= value;
956
957     block->mx= p0;
958     block->my= p1;
959     block->ref= ref;
960     block->type &= ~BLOCK_INTRA;
961     block[1]= block[b_stride]= block[b_stride+1]= *block;
962
963     rd= get_4block_rd(s, mb_x, mb_y, 0);
964
965 //FIXME chroma
966     if(rd < *best_rd){
967         *best_rd= rd;
968         return 1;
969     }else{
970         block[0]= backup[0];
971         block[1]= backup[1];
972         block[b_stride]= backup[2];
973         block[b_stride+1]= backup[3];
974         return 0;
975     }
976 }
977
978 static void iterative_me(SnowContext *s){
979     int pass, mb_x, mb_y;
980     const int b_width = s->b_width  << s->block_max_depth;
981     const int b_height= s->b_height << s->block_max_depth;
982     const int b_stride= b_width;
983     int color[3];
984
985     {
986         RangeCoder r = s->c;
987         uint8_t state[sizeof(s->block_state)];
988         memcpy(state, s->block_state, sizeof(s->block_state));
989         for(mb_y= 0; mb_y<s->b_height; mb_y++)
990             for(mb_x= 0; mb_x<s->b_width; mb_x++)
991                 encode_q_branch(s, 0, mb_x, mb_y);
992         s->c = r;
993         memcpy(s->block_state, state, sizeof(s->block_state));
994     }
995
996     for(pass=0; pass<25; pass++){
997         int change= 0;
998
999         for(mb_y= 0; mb_y<b_height; mb_y++){
1000             for(mb_x= 0; mb_x<b_width; mb_x++){
1001                 int dia_change, i, j, ref;
1002                 int best_rd= INT_MAX, ref_rd;
1003                 BlockNode backup, ref_b;
1004                 const int index= mb_x + mb_y * b_stride;
1005                 BlockNode *block= &s->block[index];
1006                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1007                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1008                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1009                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1010                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1011                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1012                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1013                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1014                 const int b_w= (MB_SIZE >> s->block_max_depth);
1015                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1016
1017                 if(pass && (block->type & BLOCK_OPT))
1018                     continue;
1019                 block->type |= BLOCK_OPT;
1020
1021                 backup= *block;
1022
1023                 if(!s->me_cache_generation)
1024                     memset(s->me_cache, 0, sizeof(s->me_cache));
1025                 s->me_cache_generation += 1<<22;
1026
1027                 //FIXME precalculate
1028                 {
1029                     int x, y;
1030                     for (y = 0; y < b_w * 2; y++)
1031                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1032                     if(mb_x==0)
1033                         for(y=0; y<b_w*2; y++)
1034                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1035                     if(mb_x==b_stride-1)
1036                         for(y=0; y<b_w*2; y++)
1037                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1038                     if(mb_y==0){
1039                         for(x=0; x<b_w*2; x++)
1040                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1041                         for(y=1; y<b_w; y++)
1042                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1043                     }
1044                     if(mb_y==b_height-1){
1045                         for(x=0; x<b_w*2; x++)
1046                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1047                         for(y=b_w; y<b_w*2-1; y++)
1048                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1049                     }
1050                 }
1051
1052                 //skip stuff outside the picture
1053                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1054                     uint8_t *src= s->  input_picture->data[0];
1055                     uint8_t *dst= s->current_picture->data[0];
1056                     const int stride= s->current_picture->linesize[0];
1057                     const int block_w= MB_SIZE >> s->block_max_depth;
1058                     const int block_h= MB_SIZE >> s->block_max_depth;
1059                     const int sx= block_w*mb_x - block_w/2;
1060                     const int sy= block_h*mb_y - block_h/2;
1061                     const int w= s->plane[0].width;
1062                     const int h= s->plane[0].height;
1063                     int y;
1064
1065                     for(y=sy; y<0; y++)
1066                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1067                     for(y=h; y<sy+block_h*2; y++)
1068                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1069                     if(sx<0){
1070                         for(y=sy; y<sy+block_h*2; y++)
1071                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1072                     }
1073                     if(sx+block_w*2 > w){
1074                         for(y=sy; y<sy+block_h*2; y++)
1075                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1076                     }
1077                 }
1078
1079                 // intra(black) = neighbors' contribution to the current block
1080                 for(i=0; i < s->nb_planes; i++)
1081                     color[i]= get_dc(s, mb_x, mb_y, i);
1082
1083                 // get previous score (cannot be cached due to OBMC)
1084                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1085                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1086                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1087                 }else
1088                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1089
1090                 ref_b= *block;
1091                 ref_rd= best_rd;
1092                 for(ref=0; ref < s->ref_frames; ref++){
1093                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1094                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1095                         continue;
1096                     block->ref= ref;
1097                     best_rd= INT_MAX;
1098
1099                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1100                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1101                     if(tb)
1102                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1103                     if(lb)
1104                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1105                     if(rb)
1106                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1107                     if(bb)
1108                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1109
1110                     /* fullpel ME */
1111                     //FIXME avoid subpel interpolation / round to nearest integer
1112                     do{
1113                         int newx = block->mx;
1114                         int newy = block->my;
1115                         dia_change=0;
1116                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1117                             for(j=0; j<i; j++){
1118                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+4*(i-j), newy+(4*j), obmc_edged, &best_rd);
1119                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-4*(i-j), newy-(4*j), obmc_edged, &best_rd);
1120                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-(4*j), newy+4*(i-j), obmc_edged, &best_rd);
1121                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+(4*j), newy-4*(i-j), obmc_edged, &best_rd);
1122                             }
1123                         }
1124                     }while(dia_change);
1125                     /* subpel ME */
1126                     do{
1127                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1128                         dia_change=0;
1129                         for(i=0; i<8; i++)
1130                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1131                     }while(dia_change);
1132                     //FIXME or try the standard 2 pass qpel or similar
1133
1134                     mvr[0][0]= block->mx;
1135                     mvr[0][1]= block->my;
1136                     if(ref_rd > best_rd){
1137                         ref_rd= best_rd;
1138                         ref_b= *block;
1139                     }
1140                 }
1141                 best_rd= ref_rd;
1142                 *block= ref_b;
1143                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1144                 //FIXME RD style color selection
1145                 if(!same_block(block, &backup)){
1146                     if(tb ) tb ->type &= ~BLOCK_OPT;
1147                     if(lb ) lb ->type &= ~BLOCK_OPT;
1148                     if(rb ) rb ->type &= ~BLOCK_OPT;
1149                     if(bb ) bb ->type &= ~BLOCK_OPT;
1150                     if(tlb) tlb->type &= ~BLOCK_OPT;
1151                     if(trb) trb->type &= ~BLOCK_OPT;
1152                     if(blb) blb->type &= ~BLOCK_OPT;
1153                     if(brb) brb->type &= ~BLOCK_OPT;
1154                     change ++;
1155                 }
1156             }
1157         }
1158         av_log(s->avctx, AV_LOG_DEBUG, "pass:%d changed:%d\n", pass, change);
1159         if(!change)
1160             break;
1161     }
1162
1163     if(s->block_max_depth == 1){
1164         int change= 0;
1165         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1166             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1167                 int i;
1168                 int best_rd, init_rd;
1169                 const int index= mb_x + mb_y * b_stride;
1170                 BlockNode *b[4];
1171
1172                 b[0]= &s->block[index];
1173                 b[1]= b[0]+1;
1174                 b[2]= b[0]+b_stride;
1175                 b[3]= b[2]+1;
1176                 if(same_block(b[0], b[1]) &&
1177                    same_block(b[0], b[2]) &&
1178                    same_block(b[0], b[3]))
1179                     continue;
1180
1181                 if(!s->me_cache_generation)
1182                     memset(s->me_cache, 0, sizeof(s->me_cache));
1183                 s->me_cache_generation += 1<<22;
1184
1185                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1186
1187                 //FIXME more multiref search?
1188                 check_4block_inter(s, mb_x, mb_y,
1189                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1190                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1191
1192                 for(i=0; i<4; i++)
1193                     if(!(b[i]->type&BLOCK_INTRA))
1194                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1195
1196                 if(init_rd != best_rd)
1197                     change++;
1198             }
1199         }
1200         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1201     }
1202 }
1203
1204 static void encode_blocks(SnowContext *s, int search){
1205     int x, y;
1206     int w= s->b_width;
1207     int h= s->b_height;
1208
1209     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1210         iterative_me(s);
1211
1212     for(y=0; y<h; y++){
1213         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1214             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1215             return;
1216         }
1217         for(x=0; x<w; x++){
1218             if(s->avctx->me_method == ME_ITER || !search)
1219                 encode_q_branch2(s, 0, x, y);
1220             else
1221                 encode_q_branch (s, 0, x, y);
1222         }
1223     }
1224 }
1225
1226 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1227     const int w= b->width;
1228     const int h= b->height;
1229     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1230     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1231     int x,y, thres1, thres2;
1232
1233     if(s->qlog == LOSSLESS_QLOG){
1234         for(y=0; y<h; y++)
1235             for(x=0; x<w; x++)
1236                 dst[x + y*stride]= src[x + y*stride];
1237         return;
1238     }
1239
1240     bias= bias ? 0 : (3*qmul)>>3;
1241     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1242     thres2= 2*thres1;
1243
1244     if(!bias){
1245         for(y=0; y<h; y++){
1246             for(x=0; x<w; x++){
1247                 int i= src[x + y*stride];
1248
1249                 if((unsigned)(i+thres1) > thres2){
1250                     if(i>=0){
1251                         i<<= QEXPSHIFT;
1252                         i/= qmul; //FIXME optimize
1253                         dst[x + y*stride]=  i;
1254                     }else{
1255                         i= -i;
1256                         i<<= QEXPSHIFT;
1257                         i/= qmul; //FIXME optimize
1258                         dst[x + y*stride]= -i;
1259                     }
1260                 }else
1261                     dst[x + y*stride]= 0;
1262             }
1263         }
1264     }else{
1265         for(y=0; y<h; y++){
1266             for(x=0; x<w; x++){
1267                 int i= src[x + y*stride];
1268
1269                 if((unsigned)(i+thres1) > thres2){
1270                     if(i>=0){
1271                         i<<= QEXPSHIFT;
1272                         i= (i + bias) / qmul; //FIXME optimize
1273                         dst[x + y*stride]=  i;
1274                     }else{
1275                         i= -i;
1276                         i<<= QEXPSHIFT;
1277                         i= (i + bias) / qmul; //FIXME optimize
1278                         dst[x + y*stride]= -i;
1279                     }
1280                 }else
1281                     dst[x + y*stride]= 0;
1282             }
1283         }
1284     }
1285 }
1286
1287 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1288     const int w= b->width;
1289     const int h= b->height;
1290     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1291     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1292     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1293     int x,y;
1294
1295     if(s->qlog == LOSSLESS_QLOG) return;
1296
1297     for(y=0; y<h; y++){
1298         for(x=0; x<w; x++){
1299             int i= src[x + y*stride];
1300             if(i<0){
1301                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1302             }else if(i>0){
1303                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1304             }
1305         }
1306     }
1307 }
1308
1309 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1310     const int w= b->width;
1311     const int h= b->height;
1312     int x,y;
1313
1314     for(y=h-1; y>=0; y--){
1315         for(x=w-1; x>=0; x--){
1316             int i= x + y*stride;
1317
1318             if(x){
1319                 if(use_median){
1320                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1321                     else  src[i] -= src[i - 1];
1322                 }else{
1323                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1324                     else  src[i] -= src[i - 1];
1325                 }
1326             }else{
1327                 if(y) src[i] -= src[i - stride];
1328             }
1329         }
1330     }
1331 }
1332
1333 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1334     const int w= b->width;
1335     const int h= b->height;
1336     int x,y;
1337
1338     for(y=0; y<h; y++){
1339         for(x=0; x<w; x++){
1340             int i= x + y*stride;
1341
1342             if(x){
1343                 if(use_median){
1344                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1345                     else  src[i] += src[i - 1];
1346                 }else{
1347                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1348                     else  src[i] += src[i - 1];
1349                 }
1350             }else{
1351                 if(y) src[i] += src[i - stride];
1352             }
1353         }
1354     }
1355 }
1356
1357 static void encode_qlogs(SnowContext *s){
1358     int plane_index, level, orientation;
1359
1360     for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1361         for(level=0; level<s->spatial_decomposition_count; level++){
1362             for(orientation=level ? 1:0; orientation<4; orientation++){
1363                 if(orientation==2) continue;
1364                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1365             }
1366         }
1367     }
1368 }
1369
1370 static void encode_header(SnowContext *s){
1371     int plane_index, i;
1372     uint8_t kstate[32];
1373
1374     memset(kstate, MID_STATE, sizeof(kstate));
1375
1376     put_rac(&s->c, kstate, s->keyframe);
1377     if(s->keyframe || s->always_reset){
1378         ff_snow_reset_contexts(s);
1379         s->last_spatial_decomposition_type=
1380         s->last_qlog=
1381         s->last_qbias=
1382         s->last_mv_scale=
1383         s->last_block_max_depth= 0;
1384         for(plane_index=0; plane_index<2; plane_index++){
1385             Plane *p= &s->plane[plane_index];
1386             p->last_htaps=0;
1387             p->last_diag_mc=0;
1388             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1389         }
1390     }
1391     if(s->keyframe){
1392         put_symbol(&s->c, s->header_state, s->version, 0);
1393         put_rac(&s->c, s->header_state, s->always_reset);
1394         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1395         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1396         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1397         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1398         if (s->nb_planes > 2) {
1399             put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1400             put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1401         }
1402         put_rac(&s->c, s->header_state, s->spatial_scalability);
1403 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1404         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1405
1406         encode_qlogs(s);
1407     }
1408
1409     if(!s->keyframe){
1410         int update_mc=0;
1411         for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1412             Plane *p= &s->plane[plane_index];
1413             update_mc |= p->last_htaps   != p->htaps;
1414             update_mc |= p->last_diag_mc != p->diag_mc;
1415             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1416         }
1417         put_rac(&s->c, s->header_state, update_mc);
1418         if(update_mc){
1419             for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1420                 Plane *p= &s->plane[plane_index];
1421                 put_rac(&s->c, s->header_state, p->diag_mc);
1422                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1423                 for(i= p->htaps/2; i; i--)
1424                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1425             }
1426         }
1427         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1428             put_rac(&s->c, s->header_state, 1);
1429             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1430             encode_qlogs(s);
1431         }else
1432             put_rac(&s->c, s->header_state, 0);
1433     }
1434
1435     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1436     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1437     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1438     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1439     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1440
1441 }
1442
1443 static void update_last_header_values(SnowContext *s){
1444     int plane_index;
1445
1446     if(!s->keyframe){
1447         for(plane_index=0; plane_index<2; plane_index++){
1448             Plane *p= &s->plane[plane_index];
1449             p->last_diag_mc= p->diag_mc;
1450             p->last_htaps  = p->htaps;
1451             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1452         }
1453     }
1454
1455     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1456     s->last_qlog                        = s->qlog;
1457     s->last_qbias                       = s->qbias;
1458     s->last_mv_scale                    = s->mv_scale;
1459     s->last_block_max_depth             = s->block_max_depth;
1460     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1461 }
1462
1463 static int qscale2qlog(int qscale){
1464     return rint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1465            + 61*QROOT/8; ///< 64 > 60
1466 }
1467
1468 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1469 {
1470     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1471      * FIXME we know exact mv bits at this point,
1472      * but ratecontrol isn't set up to include them. */
1473     uint32_t coef_sum= 0;
1474     int level, orientation, delta_qlog;
1475
1476     for(level=0; level<s->spatial_decomposition_count; level++){
1477         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1478             SubBand *b= &s->plane[0].band[level][orientation];
1479             IDWTELEM *buf= b->ibuf;
1480             const int w= b->width;
1481             const int h= b->height;
1482             const int stride= b->stride;
1483             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1484             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1485             const int qdiv= (1<<16)/qmul;
1486             int x, y;
1487             //FIXME this is ugly
1488             for(y=0; y<h; y++)
1489                 for(x=0; x<w; x++)
1490                     buf[x+y*stride]= b->buf[x+y*stride];
1491             if(orientation==0)
1492                 decorrelate(s, b, buf, stride, 1, 0);
1493             for(y=0; y<h; y++)
1494                 for(x=0; x<w; x++)
1495                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1496         }
1497     }
1498
1499     /* ugly, ratecontrol just takes a sqrt again */
1500     av_assert0(coef_sum < INT_MAX);
1501     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1502
1503     if(pict->pict_type == AV_PICTURE_TYPE_I){
1504         s->m.current_picture.mb_var_sum= coef_sum;
1505         s->m.current_picture.mc_mb_var_sum= 0;
1506     }else{
1507         s->m.current_picture.mc_mb_var_sum= coef_sum;
1508         s->m.current_picture.mb_var_sum= 0;
1509     }
1510
1511     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1512     if (pict->quality < 0)
1513         return INT_MIN;
1514     s->lambda= pict->quality * 3/2;
1515     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1516     s->qlog+= delta_qlog;
1517     return delta_qlog;
1518 }
1519
1520 static void calculate_visual_weight(SnowContext *s, Plane *p){
1521     int width = p->width;
1522     int height= p->height;
1523     int level, orientation, x, y;
1524
1525     for(level=0; level<s->spatial_decomposition_count; level++){
1526         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1527             SubBand *b= &p->band[level][orientation];
1528             IDWTELEM *ibuf= b->ibuf;
1529             int64_t error=0;
1530
1531             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1532             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1533             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1534             for(y=0; y<height; y++){
1535                 for(x=0; x<width; x++){
1536                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1537                     error += d*d;
1538                 }
1539             }
1540
1541             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1542         }
1543     }
1544 }
1545
1546 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1547                         const AVFrame *pict, int *got_packet)
1548 {
1549     SnowContext *s = avctx->priv_data;
1550     RangeCoder * const c= &s->c;
1551     AVFrame *pic = pict;
1552     const int width= s->avctx->width;
1553     const int height= s->avctx->height;
1554     int level, orientation, plane_index, i, y, ret;
1555     uint8_t rc_header_bak[sizeof(s->header_state)];
1556     uint8_t rc_block_bak[sizeof(s->block_state)];
1557
1558     if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0)
1559         return ret;
1560
1561     ff_init_range_encoder(c, pkt->data, pkt->size);
1562     ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1563
1564     for(i=0; i < s->nb_planes; i++){
1565         int hshift= i ? s->chroma_h_shift : 0;
1566         int vshift= i ? s->chroma_v_shift : 0;
1567         for(y=0; y<FF_CEIL_RSHIFT(height, vshift); y++)
1568             memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1569                    &pict->data[i][y * pict->linesize[i]],
1570                    FF_CEIL_RSHIFT(width, hshift));
1571         s->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1572                                 FF_CEIL_RSHIFT(width, hshift), FF_CEIL_RSHIFT(height, vshift),
1573                                 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1574                                 EDGE_TOP | EDGE_BOTTOM);
1575
1576     }
1577     emms_c();
1578     s->new_picture = pict;
1579
1580     s->m.picture_number= avctx->frame_number;
1581     if(avctx->flags&CODEC_FLAG_PASS2){
1582         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1583         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1584         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1585             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1586             if (pic->quality < 0)
1587                 return -1;
1588         }
1589     }else{
1590         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1591         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1592     }
1593
1594     if(s->pass1_rc && avctx->frame_number == 0)
1595         pic->quality = 2*FF_QP2LAMBDA;
1596     if (pic->quality) {
1597         s->qlog   = qscale2qlog(pic->quality);
1598         s->lambda = pic->quality * 3/2;
1599     }
1600     if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1601         s->qlog= LOSSLESS_QLOG;
1602         s->lambda = 0;
1603     }//else keep previous frame's qlog until after motion estimation
1604
1605     if (s->current_picture->data[0] && !(s->avctx->flags&CODEC_FLAG_EMU_EDGE)) {
1606         int w = s->avctx->width;
1607         int h = s->avctx->height;
1608
1609         s->mpvencdsp.draw_edges(s->current_picture->data[0],
1610                                 s->current_picture->linesize[0], w   , h   ,
1611                                 EDGE_WIDTH  , EDGE_WIDTH  , EDGE_TOP | EDGE_BOTTOM);
1612         if (s->current_picture->data[2]) {
1613             s->mpvencdsp.draw_edges(s->current_picture->data[1],
1614                                     s->current_picture->linesize[1], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1615                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1616             s->mpvencdsp.draw_edges(s->current_picture->data[2],
1617                                     s->current_picture->linesize[2], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1618                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1619         }
1620     }
1621
1622     ff_snow_frame_start(s);
1623     av_frame_unref(avctx->coded_frame);
1624     ret = av_frame_ref(avctx->coded_frame, s->current_picture);
1625     if (ret < 0)
1626         return ret;
1627
1628     s->m.current_picture_ptr= &s->m.current_picture;
1629     s->m.current_picture.f = s->current_picture;
1630     s->m.current_picture.f->pts = pict->pts;
1631     if(pic->pict_type == AV_PICTURE_TYPE_P){
1632         int block_width = (width +15)>>4;
1633         int block_height= (height+15)>>4;
1634         int stride= s->current_picture->linesize[0];
1635
1636         av_assert0(s->current_picture->data[0]);
1637         av_assert0(s->last_picture[0]->data[0]);
1638
1639         s->m.avctx= s->avctx;
1640         s->m.   last_picture.f = s->last_picture[0];
1641         s->m.    new_picture.f = s->input_picture;
1642         s->m.   last_picture_ptr= &s->m.   last_picture;
1643         s->m.linesize = stride;
1644         s->m.uvlinesize= s->current_picture->linesize[1];
1645         s->m.width = width;
1646         s->m.height= height;
1647         s->m.mb_width = block_width;
1648         s->m.mb_height= block_height;
1649         s->m.mb_stride=   s->m.mb_width+1;
1650         s->m.b8_stride= 2*s->m.mb_width+1;
1651         s->m.f_code=1;
1652         s->m.pict_type = pic->pict_type;
1653         s->m.me_method= s->avctx->me_method;
1654         s->m.me.scene_change_score=0;
1655         s->m.me.dia_size = avctx->dia_size;
1656         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1657         s->m.out_format= FMT_H263;
1658         s->m.unrestricted_mv= 1;
1659
1660         s->m.lambda = s->lambda;
1661         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1662         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1663
1664         s->m.mecc= s->mecc; //move
1665         s->m.qdsp= s->qdsp; //move
1666         s->m.hdsp = s->hdsp;
1667         ff_init_me(&s->m);
1668         s->hdsp = s->m.hdsp;
1669         s->mecc= s->m.mecc;
1670     }
1671
1672     if(s->pass1_rc){
1673         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1674         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1675     }
1676
1677 redo_frame:
1678
1679     s->spatial_decomposition_count= 5;
1680
1681     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1682           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1683         s->spatial_decomposition_count--;
1684
1685     if (s->spatial_decomposition_count <= 0) {
1686         av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1687         return AVERROR(EINVAL);
1688     }
1689
1690     s->m.pict_type = pic->pict_type;
1691     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1692
1693     ff_snow_common_init_after_header(avctx);
1694
1695     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1696         for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1697             calculate_visual_weight(s, &s->plane[plane_index]);
1698         }
1699     }
1700
1701     encode_header(s);
1702     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1703     encode_blocks(s, 1);
1704     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1705
1706     for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1707         Plane *p= &s->plane[plane_index];
1708         int w= p->width;
1709         int h= p->height;
1710         int x, y;
1711 //        int bits= put_bits_count(&s->c.pb);
1712
1713         if (!s->memc_only) {
1714             //FIXME optimize
1715             if(pict->data[plane_index]) //FIXME gray hack
1716                 for(y=0; y<h; y++){
1717                     for(x=0; x<w; x++){
1718                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1719                     }
1720                 }
1721             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1722
1723             if(   plane_index==0
1724                && pic->pict_type == AV_PICTURE_TYPE_P
1725                && !(avctx->flags&CODEC_FLAG_PASS2)
1726                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1727                 ff_init_range_encoder(c, pkt->data, pkt->size);
1728                 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1729                 pic->pict_type= AV_PICTURE_TYPE_I;
1730                 s->keyframe=1;
1731                 s->current_picture->key_frame=1;
1732                 goto redo_frame;
1733             }
1734
1735             if(s->qlog == LOSSLESS_QLOG){
1736                 for(y=0; y<h; y++){
1737                     for(x=0; x<w; x++){
1738                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1739                     }
1740                 }
1741             }else{
1742                 for(y=0; y<h; y++){
1743                     for(x=0; x<w; x++){
1744                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1745                     }
1746                 }
1747             }
1748
1749             ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1750
1751             if(s->pass1_rc && plane_index==0){
1752                 int delta_qlog = ratecontrol_1pass(s, pic);
1753                 if (delta_qlog <= INT_MIN)
1754                     return -1;
1755                 if(delta_qlog){
1756                     //reordering qlog in the bitstream would eliminate this reset
1757                     ff_init_range_encoder(c, pkt->data, pkt->size);
1758                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1759                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1760                     encode_header(s);
1761                     encode_blocks(s, 0);
1762                 }
1763             }
1764
1765             for(level=0; level<s->spatial_decomposition_count; level++){
1766                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1767                     SubBand *b= &p->band[level][orientation];
1768
1769                     quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1770                     if(orientation==0)
1771                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1772                     if (!s->no_bitstream)
1773                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1774                     av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1775                     if(orientation==0)
1776                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1777                 }
1778             }
1779
1780             for(level=0; level<s->spatial_decomposition_count; level++){
1781                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1782                     SubBand *b= &p->band[level][orientation];
1783
1784                     dequantize(s, b, b->ibuf, b->stride);
1785                 }
1786             }
1787
1788             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1789             if(s->qlog == LOSSLESS_QLOG){
1790                 for(y=0; y<h; y++){
1791                     for(x=0; x<w; x++){
1792                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1793                     }
1794                 }
1795             }
1796             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1797         }else{
1798             //ME/MC only
1799             if(pic->pict_type == AV_PICTURE_TYPE_I){
1800                 for(y=0; y<h; y++){
1801                     for(x=0; x<w; x++){
1802                         s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1803                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1804                     }
1805                 }
1806             }else{
1807                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1808                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1809             }
1810         }
1811         if(s->avctx->flags&CODEC_FLAG_PSNR){
1812             int64_t error= 0;
1813
1814             if(pict->data[plane_index]) //FIXME gray hack
1815                 for(y=0; y<h; y++){
1816                     for(x=0; x<w; x++){
1817                         int d= s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x] - pict->data[plane_index][y*pict->linesize[plane_index] + x];
1818                         error += d*d;
1819                     }
1820                 }
1821             s->avctx->error[plane_index] += error;
1822             s->current_picture->error[plane_index] = error;
1823         }
1824
1825     }
1826
1827     update_last_header_values(s);
1828
1829     ff_snow_release_buffer(avctx);
1830
1831     s->current_picture->coded_picture_number = avctx->frame_number;
1832     s->current_picture->pict_type = pict->pict_type;
1833     s->current_picture->quality = pict->quality;
1834     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1835     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1836     s->m.current_picture.f->display_picture_number =
1837     s->m.current_picture.f->coded_picture_number   = avctx->frame_number;
1838     s->m.current_picture.f->quality                = pic->quality;
1839     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1840     if(s->pass1_rc)
1841         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1842             return -1;
1843     if(avctx->flags&CODEC_FLAG_PASS1)
1844         ff_write_pass1_stats(&s->m);
1845     s->m.last_pict_type = s->m.pict_type;
1846     avctx->frame_bits = s->m.frame_bits;
1847     avctx->mv_bits = s->m.mv_bits;
1848     avctx->misc_bits = s->m.misc_bits;
1849     avctx->p_tex_bits = s->m.p_tex_bits;
1850
1851     emms_c();
1852
1853     pkt->size = ff_rac_terminate(c);
1854     if (s->current_picture->key_frame)
1855         pkt->flags |= AV_PKT_FLAG_KEY;
1856     *got_packet = 1;
1857
1858     return 0;
1859 }
1860
1861 static av_cold int encode_end(AVCodecContext *avctx)
1862 {
1863     SnowContext *s = avctx->priv_data;
1864
1865     ff_snow_common_end(s);
1866     ff_rate_control_uninit(&s->m);
1867     av_frame_free(&s->input_picture);
1868     av_freep(&avctx->stats_out);
1869
1870     return 0;
1871 }
1872
1873 #define OFFSET(x) offsetof(SnowContext, x)
1874 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1875 static const AVOption options[] = {
1876     FF_MPV_COMMON_OPTS
1877     { "memc_only",      "Only do ME/MC (I frames -> ref, P frame -> ME+MC).",   OFFSET(memc_only), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1878     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1879     { "intra_penalty",  "Penalty for intra blocks in block decission",      OFFSET(intra_penalty), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1880     { NULL },
1881 };
1882
1883 static const AVClass snowenc_class = {
1884     .class_name = "snow encoder",
1885     .item_name  = av_default_item_name,
1886     .option     = options,
1887     .version    = LIBAVUTIL_VERSION_INT,
1888 };
1889
1890 AVCodec ff_snow_encoder = {
1891     .name           = "snow",
1892     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1893     .type           = AVMEDIA_TYPE_VIDEO,
1894     .id             = AV_CODEC_ID_SNOW,
1895     .priv_data_size = sizeof(SnowContext),
1896     .init           = encode_init,
1897     .encode2        = encode_frame,
1898     .close          = encode_end,
1899     .pix_fmts       = (const enum AVPixelFormat[]){
1900         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1901         AV_PIX_FMT_GRAY8,
1902         AV_PIX_FMT_NONE
1903     },
1904     .priv_class     = &snowenc_class,
1905     .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE |
1906                       FF_CODEC_CAP_INIT_CLEANUP,
1907 };
1908
1909
1910 #ifdef TEST
1911 #undef malloc
1912 #undef free
1913 #undef printf
1914
1915 #include "libavutil/lfg.h"
1916 #include "libavutil/mathematics.h"
1917
1918 int main(void){
1919 #define width  256
1920 #define height 256
1921     int buffer[2][width*height];
1922     SnowContext s;
1923     int i;
1924     AVLFG prng;
1925     s.spatial_decomposition_count=6;
1926     s.spatial_decomposition_type=1;
1927
1928     s.temp_dwt_buffer  = av_mallocz_array(width, sizeof(DWTELEM));
1929     s.temp_idwt_buffer = av_mallocz_array(width, sizeof(IDWTELEM));
1930
1931     if (!s.temp_dwt_buffer || !s.temp_idwt_buffer) {
1932         fprintf(stderr, "Failed to allocate memory\n");
1933         return 1;
1934     }
1935
1936     av_lfg_init(&prng, 1);
1937
1938     printf("testing 5/3 DWT\n");
1939     for(i=0; i<width*height; i++)
1940         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1941
1942     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1943     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1944
1945     for(i=0; i<width*height; i++)
1946         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1947
1948     printf("testing 9/7 DWT\n");
1949     s.spatial_decomposition_type=0;
1950     for(i=0; i<width*height; i++)
1951         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1952
1953     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1954     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1955
1956     for(i=0; i<width*height; i++)
1957         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1958
1959     {
1960     int level, orientation, x, y;
1961     int64_t errors[8][4];
1962     int64_t g=0;
1963
1964         memset(errors, 0, sizeof(errors));
1965         s.spatial_decomposition_count=3;
1966         s.spatial_decomposition_type=0;
1967         for(level=0; level<s.spatial_decomposition_count; level++){
1968             for(orientation=level ? 1 : 0; orientation<4; orientation++){
1969                 int w= width  >> (s.spatial_decomposition_count-level);
1970                 int h= height >> (s.spatial_decomposition_count-level);
1971                 int stride= width  << (s.spatial_decomposition_count-level);
1972                 DWTELEM *buf= buffer[0];
1973                 int64_t error=0;
1974
1975                 if(orientation&1) buf+=w;
1976                 if(orientation>1) buf+=stride>>1;
1977
1978                 memset(buffer[0], 0, sizeof(int)*width*height);
1979                 buf[w/2 + h/2*stride]= 256*256;
1980                 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1981                 for(y=0; y<height; y++){
1982                     for(x=0; x<width; x++){
1983                         int64_t d= buffer[0][x + y*width];
1984                         error += d*d;
1985                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
1986                     }
1987                     if(FFABS(height/2-y)<9 && level==2) printf("\n");
1988                 }
1989                 error= (int)(sqrt(error)+0.5);
1990                 errors[level][orientation]= error;
1991                 if(g) g=av_gcd(g, error);
1992                 else g= error;
1993             }
1994         }
1995         printf("static int const visual_weight[][4]={\n");
1996         for(level=0; level<s.spatial_decomposition_count; level++){
1997             printf("  {");
1998             for(orientation=0; orientation<4; orientation++){
1999                 printf("%8"PRId64",", errors[level][orientation]/g);
2000             }
2001             printf("},\n");
2002         }
2003         printf("};\n");
2004         {
2005             int level=2;
2006             int w= width  >> (s.spatial_decomposition_count-level);
2007             //int h= height >> (s.spatial_decomposition_count-level);
2008             int stride= width  << (s.spatial_decomposition_count-level);
2009             DWTELEM *buf= buffer[0];
2010             int64_t error=0;
2011
2012             buf+=w;
2013             buf+=stride>>1;
2014
2015             memset(buffer[0], 0, sizeof(int)*width*height);
2016             for(y=0; y<height; y++){
2017                 for(x=0; x<width; x++){
2018                     int tab[4]={0,2,3,1};
2019                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2020                 }
2021             }
2022             ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2023             for(y=0; y<height; y++){
2024                 for(x=0; x<width; x++){
2025                     int64_t d= buffer[0][x + y*width];
2026                     error += d*d;
2027                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2028                 }
2029                 if(FFABS(height/2-y)<9) printf("\n");
2030             }
2031         }
2032
2033     }
2034     return 0;
2035 }
2036 #endif /* TEST */