Merge commit '06c20d3e32c33c4da6d9fbc43aebaeb38c45b859'
[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/libm.h"
23 #include "libavutil/log.h"
24 #include "libavutil/opt.h"
25 #include "avcodec.h"
26 #include "internal.h"
27 #include "snow_dwt.h"
28 #include "snow.h"
29
30 #include "rangecoder.h"
31 #include "mathops.h"
32
33 #include "mpegvideo.h"
34 #include "h263.h"
35
36 static av_cold int encode_init(AVCodecContext *avctx)
37 {
38     SnowContext *s = avctx->priv_data;
39     int plane_index, ret;
40     int i;
41
42 #if FF_API_PRIVATE_OPT
43 FF_DISABLE_DEPRECATION_WARNINGS
44     if (avctx->prediction_method)
45         s->pred = avctx->prediction_method;
46 FF_ENABLE_DEPRECATION_WARNINGS
47 #endif
48
49     if(s->pred == DWT_97
50        && (avctx->flags & AV_CODEC_FLAG_QSCALE)
51        && avctx->global_quality == 0){
52         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
53         return AVERROR(EINVAL);
54     }
55 #if FF_API_MOTION_EST
56 FF_DISABLE_DEPRECATION_WARNINGS
57     if (avctx->me_method == ME_ITER)
58         s->motion_est = FF_ME_ITER;
59 FF_ENABLE_DEPRECATION_WARNINGS
60 #endif
61
62     s->spatial_decomposition_type= s->pred; //FIXME add decorrelator type r transform_type
63
64     s->mv_scale       = (avctx->flags & AV_CODEC_FLAG_QPEL) ? 2 : 4;
65     s->block_max_depth= (avctx->flags & AV_CODEC_FLAG_4MV ) ? 1 : 0;
66
67     for(plane_index=0; plane_index<3; plane_index++){
68         s->plane[plane_index].diag_mc= 1;
69         s->plane[plane_index].htaps= 6;
70         s->plane[plane_index].hcoeff[0]=  40;
71         s->plane[plane_index].hcoeff[1]= -10;
72         s->plane[plane_index].hcoeff[2]=   2;
73         s->plane[plane_index].fast_mc= 1;
74     }
75
76     if ((ret = ff_snow_common_init(avctx)) < 0) {
77         return ret;
78     }
79     ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
80
81     ff_snow_alloc_blocks(s);
82
83     s->version=0;
84
85     s->m.avctx   = avctx;
86     s->m.bit_rate= avctx->bit_rate;
87     s->m.lmin    = avctx->mb_lmin;
88     s->m.lmax    = avctx->mb_lmax;
89
90     s->m.me.temp      =
91     s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
92     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
93     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
94     s->m.sc.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
95     if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.sc.obmc_scratchpad)
96         return AVERROR(ENOMEM);
97
98     ff_h263_encode_init(&s->m); //mv_penalty
99
100     s->max_ref_frames = av_clip(avctx->refs, 1, MAX_REF_FRAMES);
101
102     if(avctx->flags&AV_CODEC_FLAG_PASS1){
103         if(!avctx->stats_out)
104             avctx->stats_out = av_mallocz(256);
105
106         if (!avctx->stats_out)
107             return AVERROR(ENOMEM);
108     }
109     if((avctx->flags&AV_CODEC_FLAG_PASS2) || !(avctx->flags&AV_CODEC_FLAG_QSCALE)){
110         ret = ff_rate_control_init(&s->m);
111         if(ret < 0)
112             return ret;
113     }
114     s->pass1_rc= !(avctx->flags & (AV_CODEC_FLAG_QSCALE|AV_CODEC_FLAG_PASS2));
115
116     switch(avctx->pix_fmt){
117     case AV_PIX_FMT_YUV444P:
118 //    case AV_PIX_FMT_YUV422P:
119     case AV_PIX_FMT_YUV420P:
120 //    case AV_PIX_FMT_YUV411P:
121     case AV_PIX_FMT_YUV410P:
122         s->nb_planes = 3;
123         s->colorspace_type= 0;
124         break;
125     case AV_PIX_FMT_GRAY8:
126         s->nb_planes = 1;
127         s->colorspace_type = 1;
128         break;
129 /*    case AV_PIX_FMT_RGB32:
130         s->colorspace= 1;
131         break;*/
132     default:
133         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
134         return AVERROR_PATCHWELCOME;
135     }
136     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
137
138     ff_set_cmp(&s->mecc, s->mecc.me_cmp, s->avctx->me_cmp);
139     ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, s->avctx->me_sub_cmp);
140
141     s->input_picture = av_frame_alloc();
142     if (!s->input_picture)
143         return AVERROR(ENOMEM);
144
145     if ((ret = ff_snow_get_buffer(s, s->input_picture)) < 0)
146         return ret;
147
148     if(s->motion_est == FF_ME_ITER){
149         int size= s->b_width * s->b_height << 2*s->block_max_depth;
150         for(i=0; i<s->max_ref_frames; i++){
151             s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
152             s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
153             if (!s->ref_mvs[i] || !s->ref_scores[i])
154                 return AVERROR(ENOMEM);
155         }
156     }
157
158     return 0;
159 }
160
161 //near copy & paste from dsputil, FIXME
162 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
163 {
164     int s, i, j;
165
166     s = 0;
167     for (i = 0; i < h; i++) {
168         for (j = 0; j < w; j++) {
169             s += pix[0];
170             pix ++;
171         }
172         pix += line_size - w;
173     }
174     return s;
175 }
176
177 //near copy & paste from dsputil, FIXME
178 static int pix_norm1(uint8_t * pix, int line_size, int w)
179 {
180     int s, i, j;
181     uint32_t *sq = ff_square_tab + 256;
182
183     s = 0;
184     for (i = 0; i < w; i++) {
185         for (j = 0; j < w; j ++) {
186             s += sq[pix[0]];
187             pix ++;
188         }
189         pix += line_size - w;
190     }
191     return s;
192 }
193
194 static inline int get_penalty_factor(int lambda, int lambda2, int type){
195     switch(type&0xFF){
196     default:
197     case FF_CMP_SAD:
198         return lambda>>FF_LAMBDA_SHIFT;
199     case FF_CMP_DCT:
200         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
201     case FF_CMP_W53:
202         return (4*lambda)>>(FF_LAMBDA_SHIFT);
203     case FF_CMP_W97:
204         return (2*lambda)>>(FF_LAMBDA_SHIFT);
205     case FF_CMP_SATD:
206     case FF_CMP_DCT264:
207         return (2*lambda)>>FF_LAMBDA_SHIFT;
208     case FF_CMP_RD:
209     case FF_CMP_PSNR:
210     case FF_CMP_SSE:
211     case FF_CMP_NSSE:
212         return lambda2>>FF_LAMBDA_SHIFT;
213     case FF_CMP_BIT:
214         return 1;
215     }
216 }
217
218 //FIXME copy&paste
219 #define P_LEFT P[1]
220 #define P_TOP P[2]
221 #define P_TOPRIGHT P[3]
222 #define P_MEDIAN P[4]
223 #define P_MV1 P[9]
224 #define FLAG_QPEL   1 //must be 1
225
226 static int encode_q_branch(SnowContext *s, int level, int x, int y){
227     uint8_t p_buffer[1024];
228     uint8_t i_buffer[1024];
229     uint8_t p_state[sizeof(s->block_state)];
230     uint8_t i_state[sizeof(s->block_state)];
231     RangeCoder pc, ic;
232     uint8_t *pbbak= s->c.bytestream;
233     uint8_t *pbbak_start= s->c.bytestream_start;
234     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
235     const int w= s->b_width  << s->block_max_depth;
236     const int h= s->b_height << s->block_max_depth;
237     const int rem_depth= s->block_max_depth - level;
238     const int index= (x + y*w) << rem_depth;
239     const int block_w= 1<<(LOG2_MB_SIZE - level);
240     int trx= (x+1)<<rem_depth;
241     int try= (y+1)<<rem_depth;
242     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
243     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
244     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
245     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
246     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
247     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
248     int pl = left->color[0];
249     int pcb= left->color[1];
250     int pcr= left->color[2];
251     int pmx, pmy;
252     int mx=0, my=0;
253     int l,cr,cb;
254     const int stride= s->current_picture->linesize[0];
255     const int uvstride= s->current_picture->linesize[1];
256     uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y*  stride)*block_w,
257                                 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
258                                 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
259     int P[10][2];
260     int16_t last_mv[3][2];
261     int qpel= !!(s->avctx->flags & AV_CODEC_FLAG_QPEL); //unused
262     const int shift= 1+qpel;
263     MotionEstContext *c= &s->m.me;
264     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
265     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
266     int my_context= av_log2(2*FFABS(left->my - top->my));
267     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
268     int ref, best_ref, ref_score, ref_mx, ref_my;
269
270     av_assert0(sizeof(s->block_state) >= 256);
271     if(s->keyframe){
272         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
273         return 0;
274     }
275
276 //    clip predictors / edge ?
277
278     P_LEFT[0]= left->mx;
279     P_LEFT[1]= left->my;
280     P_TOP [0]= top->mx;
281     P_TOP [1]= top->my;
282     P_TOPRIGHT[0]= tr->mx;
283     P_TOPRIGHT[1]= tr->my;
284
285     last_mv[0][0]= s->block[index].mx;
286     last_mv[0][1]= s->block[index].my;
287     last_mv[1][0]= right->mx;
288     last_mv[1][1]= right->my;
289     last_mv[2][0]= bottom->mx;
290     last_mv[2][1]= bottom->my;
291
292     s->m.mb_stride=2;
293     s->m.mb_x=
294     s->m.mb_y= 0;
295     c->skip= 0;
296
297     av_assert1(c->  stride ==   stride);
298     av_assert1(c->uvstride == uvstride);
299
300     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
301     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
302     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
303     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_DMV;
304
305     c->xmin = - x*block_w - 16+3;
306     c->ymin = - y*block_w - 16+3;
307     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
308     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
309
310     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
311     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
312     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
313     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
314     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
315     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
316     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
317
318     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
319     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
320
321     if (!y) {
322         c->pred_x= P_LEFT[0];
323         c->pred_y= P_LEFT[1];
324     } else {
325         c->pred_x = P_MEDIAN[0];
326         c->pred_y = P_MEDIAN[1];
327     }
328
329     score= INT_MAX;
330     best_ref= 0;
331     for(ref=0; ref<s->ref_frames; ref++){
332         init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
333
334         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
335                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
336
337         av_assert2(ref_mx >= c->xmin);
338         av_assert2(ref_mx <= c->xmax);
339         av_assert2(ref_my >= c->ymin);
340         av_assert2(ref_my <= c->ymax);
341
342         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
343         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
344         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
345         if(s->ref_mvs[ref]){
346             s->ref_mvs[ref][index][0]= ref_mx;
347             s->ref_mvs[ref][index][1]= ref_my;
348             s->ref_scores[ref][index]= ref_score;
349         }
350         if(score > ref_score){
351             score= ref_score;
352             best_ref= ref;
353             mx= ref_mx;
354             my= ref_my;
355         }
356     }
357     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
358
359   //  subpel search
360     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
361     pc= s->c;
362     pc.bytestream_start=
363     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
364     memcpy(p_state, s->block_state, sizeof(s->block_state));
365
366     if(level!=s->block_max_depth)
367         put_rac(&pc, &p_state[4 + s_context], 1);
368     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
369     if(s->ref_frames > 1)
370         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
371     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
372     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
373     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
374     p_len= pc.bytestream - pc.bytestream_start;
375     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
376
377     block_s= block_w*block_w;
378     sum = pix_sum(current_data[0], stride, block_w, block_w);
379     l= (sum + block_s/2)/block_s;
380     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
381
382     if (s->nb_planes > 2) {
383         block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
384         sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
385         cb= (sum + block_s/2)/block_s;
386     //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
387         sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
388         cr= (sum + block_s/2)/block_s;
389     //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
390     }else
391         cb = cr = 0;
392
393     ic= s->c;
394     ic.bytestream_start=
395     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
396     memcpy(i_state, s->block_state, sizeof(s->block_state));
397     if(level!=s->block_max_depth)
398         put_rac(&ic, &i_state[4 + s_context], 1);
399     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
400     put_symbol(&ic, &i_state[32],  l-pl , 1);
401     if (s->nb_planes > 2) {
402         put_symbol(&ic, &i_state[64], cb-pcb, 1);
403         put_symbol(&ic, &i_state[96], cr-pcr, 1);
404     }
405     i_len= ic.bytestream - ic.bytestream_start;
406     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
407
408     av_assert1(iscore < 255*255*256 + s->lambda2*10);
409     av_assert1(iscore >= 0);
410     av_assert1(l>=0 && l<=255);
411     av_assert1(pl>=0 && pl<=255);
412
413     if(level==0){
414         int varc= iscore >> 8;
415         int vard= score >> 8;
416         if (vard <= 64 || vard < varc)
417             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
418         else
419             c->scene_change_score+= s->m.qscale;
420     }
421
422     if(level!=s->block_max_depth){
423         put_rac(&s->c, &s->block_state[4 + s_context], 0);
424         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
425         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
426         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
427         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
428         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
429
430         if(score2 < score && score2 < iscore)
431             return score2;
432     }
433
434     if(iscore < score){
435         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
436         memcpy(pbbak, i_buffer, i_len);
437         s->c= ic;
438         s->c.bytestream_start= pbbak_start;
439         s->c.bytestream= pbbak + i_len;
440         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
441         memcpy(s->block_state, i_state, sizeof(s->block_state));
442         return iscore;
443     }else{
444         memcpy(pbbak, p_buffer, p_len);
445         s->c= pc;
446         s->c.bytestream_start= pbbak_start;
447         s->c.bytestream= pbbak + p_len;
448         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
449         memcpy(s->block_state, p_state, sizeof(s->block_state));
450         return score;
451     }
452 }
453
454 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
455     const int w= s->b_width  << s->block_max_depth;
456     const int rem_depth= s->block_max_depth - level;
457     const int index= (x + y*w) << rem_depth;
458     int trx= (x+1)<<rem_depth;
459     BlockNode *b= &s->block[index];
460     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
461     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
462     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
463     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
464     int pl = left->color[0];
465     int pcb= left->color[1];
466     int pcr= left->color[2];
467     int pmx, pmy;
468     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
469     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
470     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
471     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
472
473     if(s->keyframe){
474         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
475         return;
476     }
477
478     if(level!=s->block_max_depth){
479         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
480             put_rac(&s->c, &s->block_state[4 + s_context], 1);
481         }else{
482             put_rac(&s->c, &s->block_state[4 + s_context], 0);
483             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
484             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
485             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
486             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
487             return;
488         }
489     }
490     if(b->type & BLOCK_INTRA){
491         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
492         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
493         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
494         if (s->nb_planes > 2) {
495             put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
496             put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
497         }
498         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
499     }else{
500         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
501         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
502         if(s->ref_frames > 1)
503             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
504         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
505         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
506         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
507     }
508 }
509
510 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
511     int i, x2, y2;
512     Plane *p= &s->plane[plane_index];
513     const int block_size = MB_SIZE >> s->block_max_depth;
514     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
515     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
516     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
517     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
518     const int ref_stride= s->current_picture->linesize[plane_index];
519     uint8_t *src= s-> input_picture->data[plane_index];
520     IDWTELEM *dst= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
521     const int b_stride = s->b_width << s->block_max_depth;
522     const int w= p->width;
523     const int h= p->height;
524     int index= mb_x + mb_y*b_stride;
525     BlockNode *b= &s->block[index];
526     BlockNode backup= *b;
527     int ab=0;
528     int aa=0;
529
530     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
531
532     b->type|= BLOCK_INTRA;
533     b->color[plane_index]= 0;
534     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
535
536     for(i=0; i<4; i++){
537         int mb_x2= mb_x + (i &1) - 1;
538         int mb_y2= mb_y + (i>>1) - 1;
539         int x= block_w*mb_x2 + block_w/2;
540         int y= block_h*mb_y2 + block_h/2;
541
542         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
543                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
544
545         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
546             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
547                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
548                 int obmc_v= obmc[index];
549                 int d;
550                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
551                 if(x<0) obmc_v += obmc[index + block_w];
552                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
553                 if(x+block_w>w) obmc_v += obmc[index - block_w];
554                 //FIXME precalculate this or simplify it somehow else
555
556                 d = -dst[index] + (1<<(FRAC_BITS-1));
557                 dst[index] = d;
558                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
559                 aa += obmc_v * obmc_v; //FIXME precalculate this
560             }
561         }
562     }
563     *b= backup;
564
565     return av_clip_uint8( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa) ); //FIXME we should not need clipping
566 }
567
568 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
569     const int b_stride = s->b_width << s->block_max_depth;
570     const int b_height = s->b_height<< s->block_max_depth;
571     int index= x + y*b_stride;
572     const BlockNode *b     = &s->block[index];
573     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
574     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
575     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
576     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
577     int dmx, dmy;
578 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
579 //  int my_context= av_log2(2*FFABS(left->my - top->my));
580
581     if(x<0 || x>=b_stride || y>=b_height)
582         return 0;
583 /*
584 1            0      0
585 01X          1-2    1
586 001XX        3-6    2-3
587 0001XXX      7-14   4-7
588 00001XXXX   15-30   8-15
589 */
590 //FIXME try accurate rate
591 //FIXME intra and inter predictors if surrounding blocks are not the same type
592     if(b->type & BLOCK_INTRA){
593         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
594                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
595                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
596     }else{
597         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
598         dmx-= b->mx;
599         dmy-= b->my;
600         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
601                     + av_log2(2*FFABS(dmy))
602                     + av_log2(2*b->ref));
603     }
604 }
605
606 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
607     Plane *p= &s->plane[plane_index];
608     const int block_size = MB_SIZE >> s->block_max_depth;
609     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
610     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
611     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
612     const int ref_stride= s->current_picture->linesize[plane_index];
613     uint8_t *dst= s->current_picture->data[plane_index];
614     uint8_t *src= s->  input_picture->data[plane_index];
615     IDWTELEM *pred= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4;
616     uint8_t *cur = s->scratchbuf;
617     uint8_t *tmp = s->emu_edge_buffer;
618     const int b_stride = s->b_width << s->block_max_depth;
619     const int b_height = s->b_height<< s->block_max_depth;
620     const int w= p->width;
621     const int h= p->height;
622     int distortion;
623     int rate= 0;
624     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
625     int sx= block_w*mb_x - block_w/2;
626     int sy= block_h*mb_y - block_h/2;
627     int x0= FFMAX(0,-sx);
628     int y0= FFMAX(0,-sy);
629     int x1= FFMIN(block_w*2, w-sx);
630     int y1= FFMIN(block_h*2, h-sy);
631     int i,x,y;
632
633     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
634
635     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);
636
637     for(y=y0; y<y1; y++){
638         const uint8_t *obmc1= obmc_edged[y];
639         const IDWTELEM *pred1 = pred + y*obmc_stride;
640         uint8_t *cur1 = cur + y*ref_stride;
641         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
642         for(x=x0; x<x1; x++){
643 #if FRAC_BITS >= LOG2_OBMC_MAX
644             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
645 #else
646             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
647 #endif
648             v = (v + pred1[x]) >> FRAC_BITS;
649             if(v&(~255)) v= ~(v>>31);
650             dst1[x] = v;
651         }
652     }
653
654     /* copy the regions where obmc[] = (uint8_t)256 */
655     if(LOG2_OBMC_MAX == 8
656         && (mb_x == 0 || mb_x == b_stride-1)
657         && (mb_y == 0 || mb_y == b_height-1)){
658         if(mb_x == 0)
659             x1 = block_w;
660         else
661             x0 = block_w;
662         if(mb_y == 0)
663             y1 = block_h;
664         else
665             y0 = block_h;
666         for(y=y0; y<y1; y++)
667             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
668     }
669
670     if(block_w==16){
671         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
672         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
673         /* FIXME cmps overlap but do not cover the wavelet's whole support.
674          * So improving the score of one block is not strictly guaranteed
675          * to improve the score of the whole frame, thus iterative motion
676          * estimation does not always converge. */
677         if(s->avctx->me_cmp == FF_CMP_W97)
678             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
679         else if(s->avctx->me_cmp == FF_CMP_W53)
680             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
681         else{
682             distortion = 0;
683             for(i=0; i<4; i++){
684                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
685                 distortion += s->mecc.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
686             }
687         }
688     }else{
689         av_assert2(block_w==8);
690         distortion = s->mecc.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
691     }
692
693     if(plane_index==0){
694         for(i=0; i<4; i++){
695 /* ..RRr
696  * .RXx.
697  * rxx..
698  */
699             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
700         }
701         if(mb_x == b_stride-2)
702             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
703     }
704     return distortion + rate*penalty_factor;
705 }
706
707 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
708     int i, y2;
709     Plane *p= &s->plane[plane_index];
710     const int block_size = MB_SIZE >> s->block_max_depth;
711     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
712     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
713     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
714     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
715     const int ref_stride= s->current_picture->linesize[plane_index];
716     uint8_t *dst= s->current_picture->data[plane_index];
717     uint8_t *src= s-> input_picture->data[plane_index];
718     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
719     // const has only been removed from zero_dst to suppress a warning
720     static IDWTELEM zero_dst[4096]; //FIXME
721     const int b_stride = s->b_width << s->block_max_depth;
722     const int w= p->width;
723     const int h= p->height;
724     int distortion= 0;
725     int rate= 0;
726     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
727
728     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
729
730     for(i=0; i<9; i++){
731         int mb_x2= mb_x + (i%3) - 1;
732         int mb_y2= mb_y + (i/3) - 1;
733         int x= block_w*mb_x2 + block_w/2;
734         int y= block_h*mb_y2 + block_h/2;
735
736         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
737                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
738
739         //FIXME find a cleaner/simpler way to skip the outside stuff
740         for(y2= y; y2<0; y2++)
741             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
742         for(y2= h; y2<y+block_h; y2++)
743             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
744         if(x<0){
745             for(y2= y; y2<y+block_h; y2++)
746                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
747         }
748         if(x+block_w > w){
749             for(y2= y; y2<y+block_h; y2++)
750                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
751         }
752
753         av_assert1(block_w== 8 || block_w==16);
754         distortion += s->mecc.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
755     }
756
757     if(plane_index==0){
758         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
759         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
760
761 /* ..RRRr
762  * .RXXx.
763  * .RXXx.
764  * rxxx.
765  */
766         if(merged)
767             rate = get_block_bits(s, mb_x, mb_y, 2);
768         for(i=merged?4:0; i<9; i++){
769             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
770             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
771         }
772     }
773     return distortion + rate*penalty_factor;
774 }
775
776 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
777     const int w= b->width;
778     const int h= b->height;
779     int x, y;
780
781     if(1){
782         int run=0;
783         int *runs = s->run_buffer;
784         int run_index=0;
785         int max_index;
786
787         for(y=0; y<h; y++){
788             for(x=0; x<w; x++){
789                 int v, p=0;
790                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
791                 v= src[x + y*stride];
792
793                 if(y){
794                     t= src[x + (y-1)*stride];
795                     if(x){
796                         lt= src[x - 1 + (y-1)*stride];
797                     }
798                     if(x + 1 < w){
799                         rt= src[x + 1 + (y-1)*stride];
800                     }
801                 }
802                 if(x){
803                     l= src[x - 1 + y*stride];
804                     /*if(x > 1){
805                         if(orientation==1) ll= src[y + (x-2)*stride];
806                         else               ll= src[x - 2 + y*stride];
807                     }*/
808                 }
809                 if(parent){
810                     int px= x>>1;
811                     int py= y>>1;
812                     if(px<b->parent->width && py<b->parent->height)
813                         p= parent[px + py*2*stride];
814                 }
815                 if(!(/*ll|*/l|lt|t|rt|p)){
816                     if(v){
817                         runs[run_index++]= run;
818                         run=0;
819                     }else{
820                         run++;
821                     }
822                 }
823             }
824         }
825         max_index= run_index;
826         runs[run_index++]= run;
827         run_index=0;
828         run= runs[run_index++];
829
830         put_symbol2(&s->c, b->state[30], max_index, 0);
831         if(run_index <= max_index)
832             put_symbol2(&s->c, b->state[1], run, 3);
833
834         for(y=0; y<h; y++){
835             if(s->c.bytestream_end - s->c.bytestream < w*40){
836                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
837                 return AVERROR(ENOMEM);
838             }
839             for(x=0; x<w; x++){
840                 int v, p=0;
841                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
842                 v= src[x + y*stride];
843
844                 if(y){
845                     t= src[x + (y-1)*stride];
846                     if(x){
847                         lt= src[x - 1 + (y-1)*stride];
848                     }
849                     if(x + 1 < w){
850                         rt= src[x + 1 + (y-1)*stride];
851                     }
852                 }
853                 if(x){
854                     l= src[x - 1 + y*stride];
855                     /*if(x > 1){
856                         if(orientation==1) ll= src[y + (x-2)*stride];
857                         else               ll= src[x - 2 + y*stride];
858                     }*/
859                 }
860                 if(parent){
861                     int px= x>>1;
862                     int py= y>>1;
863                     if(px<b->parent->width && py<b->parent->height)
864                         p= parent[px + py*2*stride];
865                 }
866                 if(/*ll|*/l|lt|t|rt|p){
867                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
868
869                     put_rac(&s->c, &b->state[0][context], !!v);
870                 }else{
871                     if(!run){
872                         run= runs[run_index++];
873
874                         if(run_index <= max_index)
875                             put_symbol2(&s->c, b->state[1], run, 3);
876                         av_assert2(v);
877                     }else{
878                         run--;
879                         av_assert2(!v);
880                     }
881                 }
882                 if(v){
883                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
884                     int l2= 2*FFABS(l) + (l<0);
885                     int t2= 2*FFABS(t) + (t<0);
886
887                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
888                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
889                 }
890             }
891         }
892     }
893     return 0;
894 }
895
896 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
897 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
898 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
899     return encode_subband_c0run(s, b, src, parent, stride, orientation);
900 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
901 }
902
903 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){
904     const int b_stride= s->b_width << s->block_max_depth;
905     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
906     BlockNode backup= *block;
907     unsigned value;
908     int rd, index;
909
910     av_assert2(mb_x>=0 && mb_y>=0);
911     av_assert2(mb_x<b_stride);
912
913     if(intra){
914         block->color[0] = p[0];
915         block->color[1] = p[1];
916         block->color[2] = p[2];
917         block->type |= BLOCK_INTRA;
918     }else{
919         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
920         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
921         if(s->me_cache[index] == value)
922             return 0;
923         s->me_cache[index]= value;
924
925         block->mx= p[0];
926         block->my= p[1];
927         block->type &= ~BLOCK_INTRA;
928     }
929
930     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged) + s->intra_penalty * !!intra;
931
932 //FIXME chroma
933     if(rd < *best_rd){
934         *best_rd= rd;
935         return 1;
936     }else{
937         *block= backup;
938         return 0;
939     }
940 }
941
942 /* special case for int[2] args we discard afterwards,
943  * fixes compilation problem with gcc 2.95 */
944 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){
945     int p[2] = {p0, p1};
946     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
947 }
948
949 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){
950     const int b_stride= s->b_width << s->block_max_depth;
951     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
952     BlockNode backup[4];
953     unsigned value;
954     int rd, index;
955
956     /* We don't initialize backup[] during variable declaration, because
957      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
958      * 'int16_t'". */
959     backup[0] = block[0];
960     backup[1] = block[1];
961     backup[2] = block[b_stride];
962     backup[3] = block[b_stride + 1];
963
964     av_assert2(mb_x>=0 && mb_y>=0);
965     av_assert2(mb_x<b_stride);
966     av_assert2(((mb_x|mb_y)&1) == 0);
967
968     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
969     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
970     if(s->me_cache[index] == value)
971         return 0;
972     s->me_cache[index]= value;
973
974     block->mx= p0;
975     block->my= p1;
976     block->ref= ref;
977     block->type &= ~BLOCK_INTRA;
978     block[1]= block[b_stride]= block[b_stride+1]= *block;
979
980     rd= get_4block_rd(s, mb_x, mb_y, 0);
981
982 //FIXME chroma
983     if(rd < *best_rd){
984         *best_rd= rd;
985         return 1;
986     }else{
987         block[0]= backup[0];
988         block[1]= backup[1];
989         block[b_stride]= backup[2];
990         block[b_stride+1]= backup[3];
991         return 0;
992     }
993 }
994
995 static void iterative_me(SnowContext *s){
996     int pass, mb_x, mb_y;
997     const int b_width = s->b_width  << s->block_max_depth;
998     const int b_height= s->b_height << s->block_max_depth;
999     const int b_stride= b_width;
1000     int color[3];
1001
1002     {
1003         RangeCoder r = s->c;
1004         uint8_t state[sizeof(s->block_state)];
1005         memcpy(state, s->block_state, sizeof(s->block_state));
1006         for(mb_y= 0; mb_y<s->b_height; mb_y++)
1007             for(mb_x= 0; mb_x<s->b_width; mb_x++)
1008                 encode_q_branch(s, 0, mb_x, mb_y);
1009         s->c = r;
1010         memcpy(s->block_state, state, sizeof(s->block_state));
1011     }
1012
1013     for(pass=0; pass<25; pass++){
1014         int change= 0;
1015
1016         for(mb_y= 0; mb_y<b_height; mb_y++){
1017             for(mb_x= 0; mb_x<b_width; mb_x++){
1018                 int dia_change, i, j, ref;
1019                 int best_rd= INT_MAX, ref_rd;
1020                 BlockNode backup, ref_b;
1021                 const int index= mb_x + mb_y * b_stride;
1022                 BlockNode *block= &s->block[index];
1023                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1024                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1025                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1026                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1027                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1028                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1029                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1030                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1031                 const int b_w= (MB_SIZE >> s->block_max_depth);
1032                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1033
1034                 if(pass && (block->type & BLOCK_OPT))
1035                     continue;
1036                 block->type |= BLOCK_OPT;
1037
1038                 backup= *block;
1039
1040                 if(!s->me_cache_generation)
1041                     memset(s->me_cache, 0, sizeof(s->me_cache));
1042                 s->me_cache_generation += 1<<22;
1043
1044                 //FIXME precalculate
1045                 {
1046                     int x, y;
1047                     for (y = 0; y < b_w * 2; y++)
1048                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1049                     if(mb_x==0)
1050                         for(y=0; y<b_w*2; y++)
1051                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1052                     if(mb_x==b_stride-1)
1053                         for(y=0; y<b_w*2; y++)
1054                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1055                     if(mb_y==0){
1056                         for(x=0; x<b_w*2; x++)
1057                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1058                         for(y=1; y<b_w; y++)
1059                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1060                     }
1061                     if(mb_y==b_height-1){
1062                         for(x=0; x<b_w*2; x++)
1063                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1064                         for(y=b_w; y<b_w*2-1; y++)
1065                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1066                     }
1067                 }
1068
1069                 //skip stuff outside the picture
1070                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1071                     uint8_t *src= s->  input_picture->data[0];
1072                     uint8_t *dst= s->current_picture->data[0];
1073                     const int stride= s->current_picture->linesize[0];
1074                     const int block_w= MB_SIZE >> s->block_max_depth;
1075                     const int block_h= MB_SIZE >> s->block_max_depth;
1076                     const int sx= block_w*mb_x - block_w/2;
1077                     const int sy= block_h*mb_y - block_h/2;
1078                     const int w= s->plane[0].width;
1079                     const int h= s->plane[0].height;
1080                     int y;
1081
1082                     for(y=sy; y<0; y++)
1083                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1084                     for(y=h; y<sy+block_h*2; y++)
1085                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1086                     if(sx<0){
1087                         for(y=sy; y<sy+block_h*2; y++)
1088                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1089                     }
1090                     if(sx+block_w*2 > w){
1091                         for(y=sy; y<sy+block_h*2; y++)
1092                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1093                     }
1094                 }
1095
1096                 // intra(black) = neighbors' contribution to the current block
1097                 for(i=0; i < s->nb_planes; i++)
1098                     color[i]= get_dc(s, mb_x, mb_y, i);
1099
1100                 // get previous score (cannot be cached due to OBMC)
1101                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1102                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1103                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1104                 }else
1105                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1106
1107                 ref_b= *block;
1108                 ref_rd= best_rd;
1109                 for(ref=0; ref < s->ref_frames; ref++){
1110                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1111                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1112                         continue;
1113                     block->ref= ref;
1114                     best_rd= INT_MAX;
1115
1116                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1117                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1118                     if(tb)
1119                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1120                     if(lb)
1121                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1122                     if(rb)
1123                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1124                     if(bb)
1125                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1126
1127                     /* fullpel ME */
1128                     //FIXME avoid subpel interpolation / round to nearest integer
1129                     do{
1130                         int newx = block->mx;
1131                         int newy = block->my;
1132                         int dia_size = s->iterative_dia_size ? s->iterative_dia_size : FFMAX(s->avctx->dia_size, 1);
1133                         dia_change=0;
1134                         for(i=0; i < dia_size; i++){
1135                             for(j=0; j<i; j++){
1136                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+4*(i-j), newy+(4*j), obmc_edged, &best_rd);
1137                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-4*(i-j), newy-(4*j), obmc_edged, &best_rd);
1138                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-(4*j), newy+4*(i-j), obmc_edged, &best_rd);
1139                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+(4*j), newy-4*(i-j), obmc_edged, &best_rd);
1140                             }
1141                         }
1142                     }while(dia_change);
1143                     /* subpel ME */
1144                     do{
1145                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1146                         dia_change=0;
1147                         for(i=0; i<8; i++)
1148                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1149                     }while(dia_change);
1150                     //FIXME or try the standard 2 pass qpel or similar
1151
1152                     mvr[0][0]= block->mx;
1153                     mvr[0][1]= block->my;
1154                     if(ref_rd > best_rd){
1155                         ref_rd= best_rd;
1156                         ref_b= *block;
1157                     }
1158                 }
1159                 best_rd= ref_rd;
1160                 *block= ref_b;
1161                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1162                 //FIXME RD style color selection
1163                 if(!same_block(block, &backup)){
1164                     if(tb ) tb ->type &= ~BLOCK_OPT;
1165                     if(lb ) lb ->type &= ~BLOCK_OPT;
1166                     if(rb ) rb ->type &= ~BLOCK_OPT;
1167                     if(bb ) bb ->type &= ~BLOCK_OPT;
1168                     if(tlb) tlb->type &= ~BLOCK_OPT;
1169                     if(trb) trb->type &= ~BLOCK_OPT;
1170                     if(blb) blb->type &= ~BLOCK_OPT;
1171                     if(brb) brb->type &= ~BLOCK_OPT;
1172                     change ++;
1173                 }
1174             }
1175         }
1176         av_log(s->avctx, AV_LOG_DEBUG, "pass:%d changed:%d\n", pass, change);
1177         if(!change)
1178             break;
1179     }
1180
1181     if(s->block_max_depth == 1){
1182         int change= 0;
1183         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1184             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1185                 int i;
1186                 int best_rd, init_rd;
1187                 const int index= mb_x + mb_y * b_stride;
1188                 BlockNode *b[4];
1189
1190                 b[0]= &s->block[index];
1191                 b[1]= b[0]+1;
1192                 b[2]= b[0]+b_stride;
1193                 b[3]= b[2]+1;
1194                 if(same_block(b[0], b[1]) &&
1195                    same_block(b[0], b[2]) &&
1196                    same_block(b[0], b[3]))
1197                     continue;
1198
1199                 if(!s->me_cache_generation)
1200                     memset(s->me_cache, 0, sizeof(s->me_cache));
1201                 s->me_cache_generation += 1<<22;
1202
1203                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1204
1205                 //FIXME more multiref search?
1206                 check_4block_inter(s, mb_x, mb_y,
1207                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1208                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1209
1210                 for(i=0; i<4; i++)
1211                     if(!(b[i]->type&BLOCK_INTRA))
1212                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1213
1214                 if(init_rd != best_rd)
1215                     change++;
1216             }
1217         }
1218         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1219     }
1220 }
1221
1222 static void encode_blocks(SnowContext *s, int search){
1223     int x, y;
1224     int w= s->b_width;
1225     int h= s->b_height;
1226
1227     if(s->motion_est == FF_ME_ITER && !s->keyframe && search)
1228         iterative_me(s);
1229
1230     for(y=0; y<h; y++){
1231         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1232             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1233             return;
1234         }
1235         for(x=0; x<w; x++){
1236             if(s->motion_est == FF_ME_ITER || !search)
1237                 encode_q_branch2(s, 0, x, y);
1238             else
1239                 encode_q_branch (s, 0, x, y);
1240         }
1241     }
1242 }
1243
1244 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1245     const int w= b->width;
1246     const int h= b->height;
1247     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1248     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1249     int x,y, thres1, thres2;
1250
1251     if(s->qlog == LOSSLESS_QLOG){
1252         for(y=0; y<h; y++)
1253             for(x=0; x<w; x++)
1254                 dst[x + y*stride]= src[x + y*stride];
1255         return;
1256     }
1257
1258     bias= bias ? 0 : (3*qmul)>>3;
1259     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1260     thres2= 2*thres1;
1261
1262     if(!bias){
1263         for(y=0; y<h; y++){
1264             for(x=0; x<w; x++){
1265                 int i= src[x + y*stride];
1266
1267                 if((unsigned)(i+thres1) > thres2){
1268                     if(i>=0){
1269                         i<<= QEXPSHIFT;
1270                         i/= qmul; //FIXME optimize
1271                         dst[x + y*stride]=  i;
1272                     }else{
1273                         i= -i;
1274                         i<<= QEXPSHIFT;
1275                         i/= qmul; //FIXME optimize
1276                         dst[x + y*stride]= -i;
1277                     }
1278                 }else
1279                     dst[x + y*stride]= 0;
1280             }
1281         }
1282     }else{
1283         for(y=0; y<h; y++){
1284             for(x=0; x<w; x++){
1285                 int i= src[x + y*stride];
1286
1287                 if((unsigned)(i+thres1) > thres2){
1288                     if(i>=0){
1289                         i<<= QEXPSHIFT;
1290                         i= (i + bias) / qmul; //FIXME optimize
1291                         dst[x + y*stride]=  i;
1292                     }else{
1293                         i= -i;
1294                         i<<= QEXPSHIFT;
1295                         i= (i + bias) / qmul; //FIXME optimize
1296                         dst[x + y*stride]= -i;
1297                     }
1298                 }else
1299                     dst[x + y*stride]= 0;
1300             }
1301         }
1302     }
1303 }
1304
1305 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1306     const int w= b->width;
1307     const int h= b->height;
1308     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1309     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1310     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1311     int x,y;
1312
1313     if(s->qlog == LOSSLESS_QLOG) return;
1314
1315     for(y=0; y<h; y++){
1316         for(x=0; x<w; x++){
1317             int i= src[x + y*stride];
1318             if(i<0){
1319                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1320             }else if(i>0){
1321                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1322             }
1323         }
1324     }
1325 }
1326
1327 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1328     const int w= b->width;
1329     const int h= b->height;
1330     int x,y;
1331
1332     for(y=h-1; y>=0; y--){
1333         for(x=w-1; x>=0; x--){
1334             int i= x + y*stride;
1335
1336             if(x){
1337                 if(use_median){
1338                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1339                     else  src[i] -= src[i - 1];
1340                 }else{
1341                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1342                     else  src[i] -= src[i - 1];
1343                 }
1344             }else{
1345                 if(y) src[i] -= src[i - stride];
1346             }
1347         }
1348     }
1349 }
1350
1351 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1352     const int w= b->width;
1353     const int h= b->height;
1354     int x,y;
1355
1356     for(y=0; y<h; y++){
1357         for(x=0; x<w; x++){
1358             int i= x + y*stride;
1359
1360             if(x){
1361                 if(use_median){
1362                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1363                     else  src[i] += src[i - 1];
1364                 }else{
1365                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1366                     else  src[i] += src[i - 1];
1367                 }
1368             }else{
1369                 if(y) src[i] += src[i - stride];
1370             }
1371         }
1372     }
1373 }
1374
1375 static void encode_qlogs(SnowContext *s){
1376     int plane_index, level, orientation;
1377
1378     for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1379         for(level=0; level<s->spatial_decomposition_count; level++){
1380             for(orientation=level ? 1:0; orientation<4; orientation++){
1381                 if(orientation==2) continue;
1382                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1383             }
1384         }
1385     }
1386 }
1387
1388 static void encode_header(SnowContext *s){
1389     int plane_index, i;
1390     uint8_t kstate[32];
1391
1392     memset(kstate, MID_STATE, sizeof(kstate));
1393
1394     put_rac(&s->c, kstate, s->keyframe);
1395     if(s->keyframe || s->always_reset){
1396         ff_snow_reset_contexts(s);
1397         s->last_spatial_decomposition_type=
1398         s->last_qlog=
1399         s->last_qbias=
1400         s->last_mv_scale=
1401         s->last_block_max_depth= 0;
1402         for(plane_index=0; plane_index<2; plane_index++){
1403             Plane *p= &s->plane[plane_index];
1404             p->last_htaps=0;
1405             p->last_diag_mc=0;
1406             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1407         }
1408     }
1409     if(s->keyframe){
1410         put_symbol(&s->c, s->header_state, s->version, 0);
1411         put_rac(&s->c, s->header_state, s->always_reset);
1412         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1413         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1414         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1415         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1416         if (s->nb_planes > 2) {
1417             put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1418             put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1419         }
1420         put_rac(&s->c, s->header_state, s->spatial_scalability);
1421 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1422         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1423
1424         encode_qlogs(s);
1425     }
1426
1427     if(!s->keyframe){
1428         int update_mc=0;
1429         for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1430             Plane *p= &s->plane[plane_index];
1431             update_mc |= p->last_htaps   != p->htaps;
1432             update_mc |= p->last_diag_mc != p->diag_mc;
1433             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1434         }
1435         put_rac(&s->c, s->header_state, update_mc);
1436         if(update_mc){
1437             for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1438                 Plane *p= &s->plane[plane_index];
1439                 put_rac(&s->c, s->header_state, p->diag_mc);
1440                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1441                 for(i= p->htaps/2; i; i--)
1442                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1443             }
1444         }
1445         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1446             put_rac(&s->c, s->header_state, 1);
1447             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1448             encode_qlogs(s);
1449         }else
1450             put_rac(&s->c, s->header_state, 0);
1451     }
1452
1453     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1454     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1455     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1456     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1457     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1458
1459 }
1460
1461 static void update_last_header_values(SnowContext *s){
1462     int plane_index;
1463
1464     if(!s->keyframe){
1465         for(plane_index=0; plane_index<2; plane_index++){
1466             Plane *p= &s->plane[plane_index];
1467             p->last_diag_mc= p->diag_mc;
1468             p->last_htaps  = p->htaps;
1469             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1470         }
1471     }
1472
1473     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1474     s->last_qlog                        = s->qlog;
1475     s->last_qbias                       = s->qbias;
1476     s->last_mv_scale                    = s->mv_scale;
1477     s->last_block_max_depth             = s->block_max_depth;
1478     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1479 }
1480
1481 static int qscale2qlog(int qscale){
1482     return lrint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1483            + 61*QROOT/8; ///< 64 > 60
1484 }
1485
1486 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1487 {
1488     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1489      * FIXME we know exact mv bits at this point,
1490      * but ratecontrol isn't set up to include them. */
1491     uint32_t coef_sum= 0;
1492     int level, orientation, delta_qlog;
1493
1494     for(level=0; level<s->spatial_decomposition_count; level++){
1495         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1496             SubBand *b= &s->plane[0].band[level][orientation];
1497             IDWTELEM *buf= b->ibuf;
1498             const int w= b->width;
1499             const int h= b->height;
1500             const int stride= b->stride;
1501             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1502             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1503             const int qdiv= (1<<16)/qmul;
1504             int x, y;
1505             //FIXME this is ugly
1506             for(y=0; y<h; y++)
1507                 for(x=0; x<w; x++)
1508                     buf[x+y*stride]= b->buf[x+y*stride];
1509             if(orientation==0)
1510                 decorrelate(s, b, buf, stride, 1, 0);
1511             for(y=0; y<h; y++)
1512                 for(x=0; x<w; x++)
1513                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1514         }
1515     }
1516
1517     /* ugly, ratecontrol just takes a sqrt again */
1518     av_assert0(coef_sum < INT_MAX);
1519     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1520
1521     if(pict->pict_type == AV_PICTURE_TYPE_I){
1522         s->m.current_picture.mb_var_sum= coef_sum;
1523         s->m.current_picture.mc_mb_var_sum= 0;
1524     }else{
1525         s->m.current_picture.mc_mb_var_sum= coef_sum;
1526         s->m.current_picture.mb_var_sum= 0;
1527     }
1528
1529     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1530     if (pict->quality < 0)
1531         return INT_MIN;
1532     s->lambda= pict->quality * 3/2;
1533     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1534     s->qlog+= delta_qlog;
1535     return delta_qlog;
1536 }
1537
1538 static void calculate_visual_weight(SnowContext *s, Plane *p){
1539     int width = p->width;
1540     int height= p->height;
1541     int level, orientation, x, y;
1542
1543     for(level=0; level<s->spatial_decomposition_count; level++){
1544         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1545             SubBand *b= &p->band[level][orientation];
1546             IDWTELEM *ibuf= b->ibuf;
1547             int64_t error=0;
1548
1549             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1550             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1551             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1552             for(y=0; y<height; y++){
1553                 for(x=0; x<width; x++){
1554                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1555                     error += d*d;
1556                 }
1557             }
1558
1559             b->qlog= (int)(QROOT * log2(352256.0/sqrt(error)) + 0.5);
1560         }
1561     }
1562 }
1563
1564 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1565                         const AVFrame *pict, int *got_packet)
1566 {
1567     SnowContext *s = avctx->priv_data;
1568     RangeCoder * const c= &s->c;
1569     AVFrame *pic;
1570     const int width= s->avctx->width;
1571     const int height= s->avctx->height;
1572     int level, orientation, plane_index, i, y, ret;
1573     uint8_t rc_header_bak[sizeof(s->header_state)];
1574     uint8_t rc_block_bak[sizeof(s->block_state)];
1575
1576     if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + AV_INPUT_BUFFER_MIN_SIZE, 0)) < 0)
1577         return ret;
1578
1579     ff_init_range_encoder(c, pkt->data, pkt->size);
1580     ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1581
1582     for(i=0; i < s->nb_planes; i++){
1583         int hshift= i ? s->chroma_h_shift : 0;
1584         int vshift= i ? s->chroma_v_shift : 0;
1585         for(y=0; y<AV_CEIL_RSHIFT(height, vshift); y++)
1586             memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1587                    &pict->data[i][y * pict->linesize[i]],
1588                    AV_CEIL_RSHIFT(width, hshift));
1589         s->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1590                                 AV_CEIL_RSHIFT(width, hshift), AV_CEIL_RSHIFT(height, vshift),
1591                                 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1592                                 EDGE_TOP | EDGE_BOTTOM);
1593
1594     }
1595     emms_c();
1596     pic = s->input_picture;
1597     pic->pict_type = pict->pict_type;
1598     pic->quality = pict->quality;
1599
1600     s->m.picture_number= avctx->frame_number;
1601     if(avctx->flags&AV_CODEC_FLAG_PASS2){
1602         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1603         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1604         if(!(avctx->flags&AV_CODEC_FLAG_QSCALE)) {
1605             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1606             if (pic->quality < 0)
1607                 return -1;
1608         }
1609     }else{
1610         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1611         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1612     }
1613
1614     if(s->pass1_rc && avctx->frame_number == 0)
1615         pic->quality = 2*FF_QP2LAMBDA;
1616     if (pic->quality) {
1617         s->qlog   = qscale2qlog(pic->quality);
1618         s->lambda = pic->quality * 3/2;
1619     }
1620     if (s->qlog < 0 || (!pic->quality && (avctx->flags & AV_CODEC_FLAG_QSCALE))) {
1621         s->qlog= LOSSLESS_QLOG;
1622         s->lambda = 0;
1623     }//else keep previous frame's qlog until after motion estimation
1624
1625     if (s->current_picture->data[0]
1626 #if FF_API_EMU_EDGE
1627         && !(s->avctx->flags&CODEC_FLAG_EMU_EDGE)
1628 #endif
1629         ) {
1630         int w = s->avctx->width;
1631         int h = s->avctx->height;
1632
1633         s->mpvencdsp.draw_edges(s->current_picture->data[0],
1634                                 s->current_picture->linesize[0], w   , h   ,
1635                                 EDGE_WIDTH  , EDGE_WIDTH  , EDGE_TOP | EDGE_BOTTOM);
1636         if (s->current_picture->data[2]) {
1637             s->mpvencdsp.draw_edges(s->current_picture->data[1],
1638                                     s->current_picture->linesize[1], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1639                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1640             s->mpvencdsp.draw_edges(s->current_picture->data[2],
1641                                     s->current_picture->linesize[2], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1642                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1643         }
1644         emms_c();
1645     }
1646
1647     ff_snow_frame_start(s);
1648 #if FF_API_CODED_FRAME
1649 FF_DISABLE_DEPRECATION_WARNINGS
1650     av_frame_unref(avctx->coded_frame);
1651     ret = av_frame_ref(avctx->coded_frame, s->current_picture);
1652 FF_ENABLE_DEPRECATION_WARNINGS
1653 #endif
1654     if (ret < 0)
1655         return ret;
1656
1657     s->m.current_picture_ptr= &s->m.current_picture;
1658     s->m.current_picture.f = s->current_picture;
1659     s->m.current_picture.f->pts = pict->pts;
1660     if(pic->pict_type == AV_PICTURE_TYPE_P){
1661         int block_width = (width +15)>>4;
1662         int block_height= (height+15)>>4;
1663         int stride= s->current_picture->linesize[0];
1664
1665         av_assert0(s->current_picture->data[0]);
1666         av_assert0(s->last_picture[0]->data[0]);
1667
1668         s->m.avctx= s->avctx;
1669         s->m.   last_picture.f = s->last_picture[0];
1670         s->m.    new_picture.f = s->input_picture;
1671         s->m.   last_picture_ptr= &s->m.   last_picture;
1672         s->m.linesize = stride;
1673         s->m.uvlinesize= s->current_picture->linesize[1];
1674         s->m.width = width;
1675         s->m.height= height;
1676         s->m.mb_width = block_width;
1677         s->m.mb_height= block_height;
1678         s->m.mb_stride=   s->m.mb_width+1;
1679         s->m.b8_stride= 2*s->m.mb_width+1;
1680         s->m.f_code=1;
1681         s->m.pict_type = pic->pict_type;
1682 #if FF_API_MOTION_EST
1683 FF_DISABLE_DEPRECATION_WARNINGS
1684         s->m.me_method= s->avctx->me_method;
1685 FF_ENABLE_DEPRECATION_WARNINGS
1686 #endif
1687         s->m.motion_est= s->motion_est;
1688         s->m.me.scene_change_score=0;
1689         s->m.me.dia_size = avctx->dia_size;
1690         s->m.quarter_sample= (s->avctx->flags & AV_CODEC_FLAG_QPEL)!=0;
1691         s->m.out_format= FMT_H263;
1692         s->m.unrestricted_mv= 1;
1693
1694         s->m.lambda = s->lambda;
1695         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1696         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1697
1698         s->m.mecc= s->mecc; //move
1699         s->m.qdsp= s->qdsp; //move
1700         s->m.hdsp = s->hdsp;
1701         ff_init_me(&s->m);
1702         s->hdsp = s->m.hdsp;
1703         s->mecc= s->m.mecc;
1704     }
1705
1706     if(s->pass1_rc){
1707         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1708         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1709     }
1710
1711 redo_frame:
1712
1713     s->spatial_decomposition_count= 5;
1714
1715     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1716           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1717         s->spatial_decomposition_count--;
1718
1719     if (s->spatial_decomposition_count <= 0) {
1720         av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1721         return AVERROR(EINVAL);
1722     }
1723
1724     s->m.pict_type = pic->pict_type;
1725     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1726
1727     ff_snow_common_init_after_header(avctx);
1728
1729     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1730         for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1731             calculate_visual_weight(s, &s->plane[plane_index]);
1732         }
1733     }
1734
1735     encode_header(s);
1736     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1737     encode_blocks(s, 1);
1738     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1739
1740     for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1741         Plane *p= &s->plane[plane_index];
1742         int w= p->width;
1743         int h= p->height;
1744         int x, y;
1745 //        int bits= put_bits_count(&s->c.pb);
1746
1747         if (!s->memc_only) {
1748             //FIXME optimize
1749             if(pict->data[plane_index]) //FIXME gray hack
1750                 for(y=0; y<h; y++){
1751                     for(x=0; x<w; x++){
1752                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1753                     }
1754                 }
1755             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1756
1757 #if FF_API_PRIVATE_OPT
1758 FF_DISABLE_DEPRECATION_WARNINGS
1759             if(s->avctx->scenechange_threshold)
1760                 s->scenechange_threshold = s->avctx->scenechange_threshold;
1761 FF_ENABLE_DEPRECATION_WARNINGS
1762 #endif
1763
1764             if(   plane_index==0
1765                && pic->pict_type == AV_PICTURE_TYPE_P
1766                && !(avctx->flags&AV_CODEC_FLAG_PASS2)
1767                && s->m.me.scene_change_score > s->scenechange_threshold){
1768                 ff_init_range_encoder(c, pkt->data, pkt->size);
1769                 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1770                 pic->pict_type= AV_PICTURE_TYPE_I;
1771                 s->keyframe=1;
1772                 s->current_picture->key_frame=1;
1773                 goto redo_frame;
1774             }
1775
1776             if(s->qlog == LOSSLESS_QLOG){
1777                 for(y=0; y<h; y++){
1778                     for(x=0; x<w; x++){
1779                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1780                     }
1781                 }
1782             }else{
1783                 for(y=0; y<h; y++){
1784                     for(x=0; x<w; x++){
1785                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1786                     }
1787                 }
1788             }
1789
1790             ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1791
1792             if(s->pass1_rc && plane_index==0){
1793                 int delta_qlog = ratecontrol_1pass(s, pic);
1794                 if (delta_qlog <= INT_MIN)
1795                     return -1;
1796                 if(delta_qlog){
1797                     //reordering qlog in the bitstream would eliminate this reset
1798                     ff_init_range_encoder(c, pkt->data, pkt->size);
1799                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1800                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1801                     encode_header(s);
1802                     encode_blocks(s, 0);
1803                 }
1804             }
1805
1806             for(level=0; level<s->spatial_decomposition_count; level++){
1807                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1808                     SubBand *b= &p->band[level][orientation];
1809
1810                     quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1811                     if(orientation==0)
1812                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1813                     if (!s->no_bitstream)
1814                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1815                     av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1816                     if(orientation==0)
1817                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1818                 }
1819             }
1820
1821             for(level=0; level<s->spatial_decomposition_count; level++){
1822                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1823                     SubBand *b= &p->band[level][orientation];
1824
1825                     dequantize(s, b, b->ibuf, b->stride);
1826                 }
1827             }
1828
1829             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1830             if(s->qlog == LOSSLESS_QLOG){
1831                 for(y=0; y<h; y++){
1832                     for(x=0; x<w; x++){
1833                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1834                     }
1835                 }
1836             }
1837             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1838         }else{
1839             //ME/MC only
1840             if(pic->pict_type == AV_PICTURE_TYPE_I){
1841                 for(y=0; y<h; y++){
1842                     for(x=0; x<w; x++){
1843                         s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1844                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1845                     }
1846                 }
1847             }else{
1848                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1849                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1850             }
1851         }
1852         if(s->avctx->flags&AV_CODEC_FLAG_PSNR){
1853             int64_t error= 0;
1854
1855             if(pict->data[plane_index]) //FIXME gray hack
1856                 for(y=0; y<h; y++){
1857                     for(x=0; x<w; x++){
1858                         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];
1859                         error += d*d;
1860                     }
1861                 }
1862             s->avctx->error[plane_index] += error;
1863             s->encoding_error[plane_index] = error;
1864         }
1865
1866     }
1867     emms_c();
1868
1869     update_last_header_values(s);
1870
1871     ff_snow_release_buffer(avctx);
1872
1873     s->current_picture->coded_picture_number = avctx->frame_number;
1874     s->current_picture->pict_type = pic->pict_type;
1875     s->current_picture->quality = pic->quality;
1876     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1877     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1878     s->m.current_picture.f->display_picture_number =
1879     s->m.current_picture.f->coded_picture_number   = avctx->frame_number;
1880     s->m.current_picture.f->quality                = pic->quality;
1881     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1882     if(s->pass1_rc)
1883         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1884             return -1;
1885     if(avctx->flags&AV_CODEC_FLAG_PASS1)
1886         ff_write_pass1_stats(&s->m);
1887     s->m.last_pict_type = s->m.pict_type;
1888 #if FF_API_STAT_BITS
1889 FF_DISABLE_DEPRECATION_WARNINGS
1890     avctx->frame_bits = s->m.frame_bits;
1891     avctx->mv_bits = s->m.mv_bits;
1892     avctx->misc_bits = s->m.misc_bits;
1893     avctx->p_tex_bits = s->m.p_tex_bits;
1894 FF_ENABLE_DEPRECATION_WARNINGS
1895 #endif
1896
1897     emms_c();
1898
1899     ff_side_data_set_encoder_stats(pkt, s->current_picture->quality,
1900                                    s->encoding_error,
1901                                    (s->avctx->flags&AV_CODEC_FLAG_PSNR) ? 4 : 0,
1902                                    s->current_picture->pict_type);
1903
1904 #if FF_API_ERROR_FRAME
1905 FF_DISABLE_DEPRECATION_WARNINGS
1906     memcpy(s->current_picture->error, s->encoding_error, sizeof(s->encoding_error));
1907 FF_ENABLE_DEPRECATION_WARNINGS
1908 #endif
1909
1910     pkt->size = ff_rac_terminate(c);
1911     if (s->current_picture->key_frame)
1912         pkt->flags |= AV_PKT_FLAG_KEY;
1913     *got_packet = 1;
1914
1915     return 0;
1916 }
1917
1918 static av_cold int encode_end(AVCodecContext *avctx)
1919 {
1920     SnowContext *s = avctx->priv_data;
1921
1922     ff_snow_common_end(s);
1923     ff_rate_control_uninit(&s->m);
1924     av_frame_free(&s->input_picture);
1925     av_freep(&avctx->stats_out);
1926
1927     return 0;
1928 }
1929
1930 #define OFFSET(x) offsetof(SnowContext, x)
1931 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1932 static const AVOption options[] = {
1933     {"motion_est", "motion estimation algorithm", OFFSET(motion_est), AV_OPT_TYPE_INT, {.i64 = FF_ME_EPZS }, FF_ME_ZERO, FF_ME_ITER, VE, "motion_est" },
1934     { "zero", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ZERO }, 0, 0, VE, "motion_est" },
1935     { "epzs", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_EPZS }, 0, 0, VE, "motion_est" },
1936     { "xone", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_XONE }, 0, 0, VE, "motion_est" },
1937     { "iter", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ITER }, 0, 0, VE, "motion_est" },
1938     { "memc_only",      "Only do ME/MC (I frames -> ref, P frame -> ME+MC).",   OFFSET(memc_only), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
1939     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
1940     { "intra_penalty",  "Penalty for intra blocks in block decission",      OFFSET(intra_penalty), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1941     { "iterative_dia_size",  "Dia size for the iterative ME",          OFFSET(iterative_dia_size), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1942     { "sc_threshold",   "Scene change threshold",                   OFFSET(scenechange_threshold), AV_OPT_TYPE_INT, { .i64 = 0 }, INT_MIN, INT_MAX, VE },
1943     { "pred",           "Spatial decomposition type",                                OFFSET(pred), AV_OPT_TYPE_INT, { .i64 = 0 }, DWT_97, DWT_53, VE, "pred" },
1944         { "dwt97", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, INT_MIN, INT_MAX, VE, "pred" },
1945         { "dwt53", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN, INT_MAX, VE, "pred" },
1946     { NULL },
1947 };
1948
1949 static const AVClass snowenc_class = {
1950     .class_name = "snow encoder",
1951     .item_name  = av_default_item_name,
1952     .option     = options,
1953     .version    = LIBAVUTIL_VERSION_INT,
1954 };
1955
1956 AVCodec ff_snow_encoder = {
1957     .name           = "snow",
1958     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1959     .type           = AVMEDIA_TYPE_VIDEO,
1960     .id             = AV_CODEC_ID_SNOW,
1961     .priv_data_size = sizeof(SnowContext),
1962     .init           = encode_init,
1963     .encode2        = encode_frame,
1964     .close          = encode_end,
1965     .pix_fmts       = (const enum AVPixelFormat[]){
1966         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1967         AV_PIX_FMT_GRAY8,
1968         AV_PIX_FMT_NONE
1969     },
1970     .priv_class     = &snowenc_class,
1971     .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE |
1972                       FF_CODEC_CAP_INIT_CLEANUP,
1973 };