avcodec/snowenc: Set mb_num to avoid ratecontrol floating point divisions by 0.0
[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 "libavutil/pixdesc.h"
26 #include "avcodec.h"
27 #include "internal.h"
28 #include "snow_dwt.h"
29 #include "snow.h"
30
31 #include "rangecoder.h"
32 #include "mathops.h"
33
34 #include "mpegvideo.h"
35 #include "h263.h"
36
37 static av_cold int encode_init(AVCodecContext *avctx)
38 {
39     SnowContext *s = avctx->priv_data;
40     int plane_index, ret;
41     int i;
42
43 #if FF_API_PRIVATE_OPT
44 FF_DISABLE_DEPRECATION_WARNINGS
45     if (avctx->prediction_method)
46         s->pred = avctx->prediction_method;
47 FF_ENABLE_DEPRECATION_WARNINGS
48 #endif
49
50     if(s->pred == DWT_97
51        && (avctx->flags & AV_CODEC_FLAG_QSCALE)
52        && avctx->global_quality == 0){
53         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
54         return AVERROR(EINVAL);
55     }
56
57     s->spatial_decomposition_type= s->pred; //FIXME add decorrelator type r transform_type
58
59     s->mv_scale       = (avctx->flags & AV_CODEC_FLAG_QPEL) ? 2 : 4;
60     s->block_max_depth= (avctx->flags & AV_CODEC_FLAG_4MV ) ? 1 : 0;
61
62     for(plane_index=0; plane_index<3; plane_index++){
63         s->plane[plane_index].diag_mc= 1;
64         s->plane[plane_index].htaps= 6;
65         s->plane[plane_index].hcoeff[0]=  40;
66         s->plane[plane_index].hcoeff[1]= -10;
67         s->plane[plane_index].hcoeff[2]=   2;
68         s->plane[plane_index].fast_mc= 1;
69     }
70
71     if ((ret = ff_snow_common_init(avctx)) < 0) {
72         return ret;
73     }
74     ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
75
76     ff_snow_alloc_blocks(s);
77
78     s->version=0;
79
80     s->m.avctx   = avctx;
81     s->m.bit_rate= avctx->bit_rate;
82     s->m.lmin    = avctx->mb_lmin;
83     s->m.lmax    = avctx->mb_lmax;
84     s->m.mb_num  = (avctx->width * avctx->height + 255) / 256; // For ratecontrol
85
86     s->m.me.temp      =
87     s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
88     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
89     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
90     s->m.sc.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
91     if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.sc.obmc_scratchpad)
92         return AVERROR(ENOMEM);
93
94     ff_h263_encode_init(&s->m); //mv_penalty
95
96     s->max_ref_frames = av_clip(avctx->refs, 1, MAX_REF_FRAMES);
97
98     if(avctx->flags&AV_CODEC_FLAG_PASS1){
99         if(!avctx->stats_out)
100             avctx->stats_out = av_mallocz(256);
101
102         if (!avctx->stats_out)
103             return AVERROR(ENOMEM);
104     }
105     if((avctx->flags&AV_CODEC_FLAG_PASS2) || !(avctx->flags&AV_CODEC_FLAG_QSCALE)){
106         ret = ff_rate_control_init(&s->m);
107         if(ret < 0)
108             return ret;
109     }
110     s->pass1_rc= !(avctx->flags & (AV_CODEC_FLAG_QSCALE|AV_CODEC_FLAG_PASS2));
111
112     switch(avctx->pix_fmt){
113     case AV_PIX_FMT_YUV444P:
114 //    case AV_PIX_FMT_YUV422P:
115     case AV_PIX_FMT_YUV420P:
116 //    case AV_PIX_FMT_YUV411P:
117     case AV_PIX_FMT_YUV410P:
118         s->nb_planes = 3;
119         s->colorspace_type= 0;
120         break;
121     case AV_PIX_FMT_GRAY8:
122         s->nb_planes = 1;
123         s->colorspace_type = 1;
124         break;
125 /*    case AV_PIX_FMT_RGB32:
126         s->colorspace= 1;
127         break;*/
128     default:
129         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
130         return AVERROR_PATCHWELCOME;
131     }
132
133     ret = av_pix_fmt_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift,
134                                            &s->chroma_v_shift);
135     if (ret) {
136         av_log(avctx, AV_LOG_ERROR, "pixel format invalid or unknown\n");
137         return ret;
138     }
139
140     ff_set_cmp(&s->mecc, s->mecc.me_cmp, s->avctx->me_cmp);
141     ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, s->avctx->me_sub_cmp);
142
143     s->input_picture = av_frame_alloc();
144     if (!s->input_picture)
145         return AVERROR(ENOMEM);
146
147     if ((ret = ff_snow_get_buffer(s, s->input_picture)) < 0)
148         return ret;
149
150     if(s->motion_est == FF_ME_ITER){
151         int size= s->b_width * s->b_height << 2*s->block_max_depth;
152         for(i=0; i<s->max_ref_frames; i++){
153             s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
154             s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
155             if (!s->ref_mvs[i] || !s->ref_scores[i])
156                 return AVERROR(ENOMEM);
157         }
158     }
159
160     return 0;
161 }
162
163 //near copy & paste from dsputil, FIXME
164 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
165 {
166     int s, i, j;
167
168     s = 0;
169     for (i = 0; i < h; i++) {
170         for (j = 0; j < w; j++) {
171             s += pix[0];
172             pix ++;
173         }
174         pix += line_size - w;
175     }
176     return s;
177 }
178
179 //near copy & paste from dsputil, FIXME
180 static int pix_norm1(uint8_t * pix, int line_size, int w)
181 {
182     int s, i, j;
183     const uint32_t *sq = ff_square_tab + 256;
184
185     s = 0;
186     for (i = 0; i < w; i++) {
187         for (j = 0; j < w; j ++) {
188             s += sq[pix[0]];
189             pix ++;
190         }
191         pix += line_size - w;
192     }
193     return s;
194 }
195
196 static inline int get_penalty_factor(int lambda, int lambda2, int type){
197     switch(type&0xFF){
198     default:
199     case FF_CMP_SAD:
200         return lambda>>FF_LAMBDA_SHIFT;
201     case FF_CMP_DCT:
202         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
203     case FF_CMP_W53:
204         return (4*lambda)>>(FF_LAMBDA_SHIFT);
205     case FF_CMP_W97:
206         return (2*lambda)>>(FF_LAMBDA_SHIFT);
207     case FF_CMP_SATD:
208     case FF_CMP_DCT264:
209         return (2*lambda)>>FF_LAMBDA_SHIFT;
210     case FF_CMP_RD:
211     case FF_CMP_PSNR:
212     case FF_CMP_SSE:
213     case FF_CMP_NSSE:
214         return lambda2>>FF_LAMBDA_SHIFT;
215     case FF_CMP_BIT:
216         return 1;
217     }
218 }
219
220 //FIXME copy&paste
221 #define P_LEFT P[1]
222 #define P_TOP P[2]
223 #define P_TOPRIGHT P[3]
224 #define P_MEDIAN P[4]
225 #define P_MV1 P[9]
226 #define FLAG_QPEL   1 //must be 1
227
228 static int encode_q_branch(SnowContext *s, int level, int x, int y){
229     uint8_t p_buffer[1024];
230     uint8_t i_buffer[1024];
231     uint8_t p_state[sizeof(s->block_state)];
232     uint8_t i_state[sizeof(s->block_state)];
233     RangeCoder pc, ic;
234     uint8_t *pbbak= s->c.bytestream;
235     uint8_t *pbbak_start= s->c.bytestream_start;
236     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
237     const int w= s->b_width  << s->block_max_depth;
238     const int h= s->b_height << s->block_max_depth;
239     const int rem_depth= s->block_max_depth - level;
240     const int index= (x + y*w) << rem_depth;
241     const int block_w= 1<<(LOG2_MB_SIZE - level);
242     int trx= (x+1)<<rem_depth;
243     int try= (y+1)<<rem_depth;
244     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
245     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
246     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
247     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
248     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
249     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
250     int pl = left->color[0];
251     int pcb= left->color[1];
252     int pcr= left->color[2];
253     int pmx, pmy;
254     int mx=0, my=0;
255     int l,cr,cb;
256     const int stride= s->current_picture->linesize[0];
257     const int uvstride= s->current_picture->linesize[1];
258     uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y*  stride)*block_w,
259                                 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
260                                 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
261     int P[10][2];
262     int16_t last_mv[3][2];
263     int qpel= !!(s->avctx->flags & AV_CODEC_FLAG_QPEL); //unused
264     const int shift= 1+qpel;
265     MotionEstContext *c= &s->m.me;
266     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
267     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
268     int my_context= av_log2(2*FFABS(left->my - top->my));
269     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
270     int ref, best_ref, ref_score, ref_mx, ref_my;
271
272     av_assert0(sizeof(s->block_state) >= 256);
273     if(s->keyframe){
274         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
275         return 0;
276     }
277
278 //    clip predictors / edge ?
279
280     P_LEFT[0]= left->mx;
281     P_LEFT[1]= left->my;
282     P_TOP [0]= top->mx;
283     P_TOP [1]= top->my;
284     P_TOPRIGHT[0]= tr->mx;
285     P_TOPRIGHT[1]= tr->my;
286
287     last_mv[0][0]= s->block[index].mx;
288     last_mv[0][1]= s->block[index].my;
289     last_mv[1][0]= right->mx;
290     last_mv[1][1]= right->my;
291     last_mv[2][0]= bottom->mx;
292     last_mv[2][1]= bottom->my;
293
294     s->m.mb_stride=2;
295     s->m.mb_x=
296     s->m.mb_y= 0;
297     c->skip= 0;
298
299     av_assert1(c->  stride ==   stride);
300     av_assert1(c->uvstride == uvstride);
301
302     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
303     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
304     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
305     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_DMV;
306
307     c->xmin = - x*block_w - 16+3;
308     c->ymin = - y*block_w - 16+3;
309     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
310     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
311
312     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
313     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
314     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
315     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
316     if(P_TOPRIGHT[0] < (c->xmin * (1<<shift))) P_TOPRIGHT[0]= (c->xmin * (1<<shift));
317     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
318     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
319
320     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
321     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
322
323     if (!y) {
324         c->pred_x= P_LEFT[0];
325         c->pred_y= P_LEFT[1];
326     } else {
327         c->pred_x = P_MEDIAN[0];
328         c->pred_y = P_MEDIAN[1];
329     }
330
331     score= INT_MAX;
332     best_ref= 0;
333     for(ref=0; ref<s->ref_frames; ref++){
334         init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
335
336         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
337                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
338
339         av_assert2(ref_mx >= c->xmin);
340         av_assert2(ref_mx <= c->xmax);
341         av_assert2(ref_my >= c->ymin);
342         av_assert2(ref_my <= c->ymax);
343
344         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
345         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
346         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
347         if(s->ref_mvs[ref]){
348             s->ref_mvs[ref][index][0]= ref_mx;
349             s->ref_mvs[ref][index][1]= ref_my;
350             s->ref_scores[ref][index]= ref_score;
351         }
352         if(score > ref_score){
353             score= ref_score;
354             best_ref= ref;
355             mx= ref_mx;
356             my= ref_my;
357         }
358     }
359     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
360
361   //  subpel search
362     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
363     pc= s->c;
364     pc.bytestream_start=
365     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
366     memcpy(p_state, s->block_state, sizeof(s->block_state));
367
368     if(level!=s->block_max_depth)
369         put_rac(&pc, &p_state[4 + s_context], 1);
370     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
371     if(s->ref_frames > 1)
372         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
373     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
374     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
375     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
376     p_len= pc.bytestream - pc.bytestream_start;
377     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
378
379     block_s= block_w*block_w;
380     sum = pix_sum(current_data[0], stride, block_w, block_w);
381     l= (sum + block_s/2)/block_s;
382     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
383
384     if (s->nb_planes > 2) {
385         block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
386         sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
387         cb= (sum + block_s/2)/block_s;
388     //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
389         sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
390         cr= (sum + block_s/2)/block_s;
391     //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
392     }else
393         cb = cr = 0;
394
395     ic= s->c;
396     ic.bytestream_start=
397     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
398     memcpy(i_state, s->block_state, sizeof(s->block_state));
399     if(level!=s->block_max_depth)
400         put_rac(&ic, &i_state[4 + s_context], 1);
401     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
402     put_symbol(&ic, &i_state[32],  l-pl , 1);
403     if (s->nb_planes > 2) {
404         put_symbol(&ic, &i_state[64], cb-pcb, 1);
405         put_symbol(&ic, &i_state[96], cr-pcr, 1);
406     }
407     i_len= ic.bytestream - ic.bytestream_start;
408     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
409
410     av_assert1(iscore < 255*255*256 + s->lambda2*10);
411     av_assert1(iscore >= 0);
412     av_assert1(l>=0 && l<=255);
413     av_assert1(pl>=0 && pl<=255);
414
415     if(level==0){
416         int varc= iscore >> 8;
417         int vard= score >> 8;
418         if (vard <= 64 || vard < varc)
419             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
420         else
421             c->scene_change_score+= s->m.qscale;
422     }
423
424     if(level!=s->block_max_depth){
425         put_rac(&s->c, &s->block_state[4 + s_context], 0);
426         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
427         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
428         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
429         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
430         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
431
432         if(score2 < score && score2 < iscore)
433             return score2;
434     }
435
436     if(iscore < score){
437         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
438         memcpy(pbbak, i_buffer, i_len);
439         s->c= ic;
440         s->c.bytestream_start= pbbak_start;
441         s->c.bytestream= pbbak + i_len;
442         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
443         memcpy(s->block_state, i_state, sizeof(s->block_state));
444         return iscore;
445     }else{
446         memcpy(pbbak, p_buffer, p_len);
447         s->c= pc;
448         s->c.bytestream_start= pbbak_start;
449         s->c.bytestream= pbbak + p_len;
450         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
451         memcpy(s->block_state, p_state, sizeof(s->block_state));
452         return score;
453     }
454 }
455
456 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
457     const int w= s->b_width  << s->block_max_depth;
458     const int rem_depth= s->block_max_depth - level;
459     const int index= (x + y*w) << rem_depth;
460     int trx= (x+1)<<rem_depth;
461     BlockNode *b= &s->block[index];
462     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
463     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
464     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
465     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
466     int pl = left->color[0];
467     int pcb= left->color[1];
468     int pcr= left->color[2];
469     int pmx, pmy;
470     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
471     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
472     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
473     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
474
475     if(s->keyframe){
476         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
477         return;
478     }
479
480     if(level!=s->block_max_depth){
481         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
482             put_rac(&s->c, &s->block_state[4 + s_context], 1);
483         }else{
484             put_rac(&s->c, &s->block_state[4 + s_context], 0);
485             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
486             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
487             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
488             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
489             return;
490         }
491     }
492     if(b->type & BLOCK_INTRA){
493         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
494         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
495         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
496         if (s->nb_planes > 2) {
497             put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
498             put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
499         }
500         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
501     }else{
502         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
503         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
504         if(s->ref_frames > 1)
505             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
506         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
507         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
508         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
509     }
510 }
511
512 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
513     int i, x2, y2;
514     Plane *p= &s->plane[plane_index];
515     const int block_size = MB_SIZE >> s->block_max_depth;
516     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
517     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
518     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
519     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
520     const int ref_stride= s->current_picture->linesize[plane_index];
521     uint8_t *src= s-> input_picture->data[plane_index];
522     IDWTELEM *dst= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
523     const int b_stride = s->b_width << s->block_max_depth;
524     const int w= p->width;
525     const int h= p->height;
526     int index= mb_x + mb_y*b_stride;
527     BlockNode *b= &s->block[index];
528     BlockNode backup= *b;
529     int ab=0;
530     int aa=0;
531
532     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
533
534     b->type|= BLOCK_INTRA;
535     b->color[plane_index]= 0;
536     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
537
538     for(i=0; i<4; i++){
539         int mb_x2= mb_x + (i &1) - 1;
540         int mb_y2= mb_y + (i>>1) - 1;
541         int x= block_w*mb_x2 + block_w/2;
542         int y= block_h*mb_y2 + block_h/2;
543
544         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
545                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
546
547         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
548             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
549                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
550                 int obmc_v= obmc[index];
551                 int d;
552                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
553                 if(x<0) obmc_v += obmc[index + block_w];
554                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
555                 if(x+block_w>w) obmc_v += obmc[index - block_w];
556                 //FIXME precalculate this or simplify it somehow else
557
558                 d = -dst[index] + (1<<(FRAC_BITS-1));
559                 dst[index] = d;
560                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
561                 aa += obmc_v * obmc_v; //FIXME precalculate this
562             }
563         }
564     }
565     *b= backup;
566
567     return av_clip_uint8( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa) ); //FIXME we should not need clipping
568 }
569
570 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
571     const int b_stride = s->b_width << s->block_max_depth;
572     const int b_height = s->b_height<< s->block_max_depth;
573     int index= x + y*b_stride;
574     const BlockNode *b     = &s->block[index];
575     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
576     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
577     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
578     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
579     int dmx, dmy;
580 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
581 //  int my_context= av_log2(2*FFABS(left->my - top->my));
582
583     if(x<0 || x>=b_stride || y>=b_height)
584         return 0;
585 /*
586 1            0      0
587 01X          1-2    1
588 001XX        3-6    2-3
589 0001XXX      7-14   4-7
590 00001XXXX   15-30   8-15
591 */
592 //FIXME try accurate rate
593 //FIXME intra and inter predictors if surrounding blocks are not the same type
594     if(b->type & BLOCK_INTRA){
595         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
596                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
597                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
598     }else{
599         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
600         dmx-= b->mx;
601         dmy-= b->my;
602         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
603                     + av_log2(2*FFABS(dmy))
604                     + av_log2(2*b->ref));
605     }
606 }
607
608 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
609     Plane *p= &s->plane[plane_index];
610     const int block_size = MB_SIZE >> s->block_max_depth;
611     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
612     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
613     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
614     const int ref_stride= s->current_picture->linesize[plane_index];
615     uint8_t *dst= s->current_picture->data[plane_index];
616     uint8_t *src= s->  input_picture->data[plane_index];
617     IDWTELEM *pred= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4;
618     uint8_t *cur = s->scratchbuf;
619     uint8_t *tmp = s->emu_edge_buffer;
620     const int b_stride = s->b_width << s->block_max_depth;
621     const int b_height = s->b_height<< s->block_max_depth;
622     const int w= p->width;
623     const int h= p->height;
624     int distortion;
625     int rate= 0;
626     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
627     int sx= block_w*mb_x - block_w/2;
628     int sy= block_h*mb_y - block_h/2;
629     int x0= FFMAX(0,-sx);
630     int y0= FFMAX(0,-sy);
631     int x1= FFMIN(block_w*2, w-sx);
632     int y1= FFMIN(block_h*2, h-sy);
633     int i,x,y;
634
635     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
636
637     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);
638
639     for(y=y0; y<y1; y++){
640         const uint8_t *obmc1= obmc_edged[y];
641         const IDWTELEM *pred1 = pred + y*obmc_stride;
642         uint8_t *cur1 = cur + y*ref_stride;
643         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
644         for(x=x0; x<x1; x++){
645 #if FRAC_BITS >= LOG2_OBMC_MAX
646             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
647 #else
648             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
649 #endif
650             v = (v + pred1[x]) >> FRAC_BITS;
651             if(v&(~255)) v= ~(v>>31);
652             dst1[x] = v;
653         }
654     }
655
656     /* copy the regions where obmc[] = (uint8_t)256 */
657     if(LOG2_OBMC_MAX == 8
658         && (mb_x == 0 || mb_x == b_stride-1)
659         && (mb_y == 0 || mb_y == b_height-1)){
660         if(mb_x == 0)
661             x1 = block_w;
662         else
663             x0 = block_w;
664         if(mb_y == 0)
665             y1 = block_h;
666         else
667             y0 = block_h;
668         for(y=y0; y<y1; y++)
669             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
670     }
671
672     if(block_w==16){
673         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
674         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
675         /* FIXME cmps overlap but do not cover the wavelet's whole support.
676          * So improving the score of one block is not strictly guaranteed
677          * to improve the score of the whole frame, thus iterative motion
678          * estimation does not always converge. */
679         if(s->avctx->me_cmp == FF_CMP_W97)
680             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
681         else if(s->avctx->me_cmp == FF_CMP_W53)
682             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
683         else{
684             distortion = 0;
685             for(i=0; i<4; i++){
686                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
687                 distortion += s->mecc.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
688             }
689         }
690     }else{
691         av_assert2(block_w==8);
692         distortion = s->mecc.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
693     }
694
695     if(plane_index==0){
696         for(i=0; i<4; i++){
697 /* ..RRr
698  * .RXx.
699  * rxx..
700  */
701             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
702         }
703         if(mb_x == b_stride-2)
704             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
705     }
706     return distortion + rate*penalty_factor;
707 }
708
709 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
710     int i, y2;
711     Plane *p= &s->plane[plane_index];
712     const int block_size = MB_SIZE >> s->block_max_depth;
713     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
714     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
715     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
716     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
717     const int ref_stride= s->current_picture->linesize[plane_index];
718     uint8_t *dst= s->current_picture->data[plane_index];
719     uint8_t *src= s-> input_picture->data[plane_index];
720     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
721     // const has only been removed from zero_dst to suppress a warning
722     static IDWTELEM zero_dst[4096]; //FIXME
723     const int b_stride = s->b_width << s->block_max_depth;
724     const int w= p->width;
725     const int h= p->height;
726     int distortion= 0;
727     int rate= 0;
728     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
729
730     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
731
732     for(i=0; i<9; i++){
733         int mb_x2= mb_x + (i%3) - 1;
734         int mb_y2= mb_y + (i/3) - 1;
735         int x= block_w*mb_x2 + block_w/2;
736         int y= block_h*mb_y2 + block_h/2;
737
738         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
739                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
740
741         //FIXME find a cleaner/simpler way to skip the outside stuff
742         for(y2= y; y2<0; y2++)
743             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
744         for(y2= h; y2<y+block_h; y2++)
745             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
746         if(x<0){
747             for(y2= y; y2<y+block_h; y2++)
748                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
749         }
750         if(x+block_w > w){
751             for(y2= y; y2<y+block_h; y2++)
752                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
753         }
754
755         av_assert1(block_w== 8 || block_w==16);
756         distortion += s->mecc.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
757     }
758
759     if(plane_index==0){
760         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
761         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
762
763 /* ..RRRr
764  * .RXXx.
765  * .RXXx.
766  * rxxx.
767  */
768         if(merged)
769             rate = get_block_bits(s, mb_x, mb_y, 2);
770         for(i=merged?4:0; i<9; i++){
771             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
772             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
773         }
774     }
775     return distortion + rate*penalty_factor;
776 }
777
778 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
779     const int w= b->width;
780     const int h= b->height;
781     int x, y;
782
783     if(1){
784         int run=0;
785         int *runs = s->run_buffer;
786         int run_index=0;
787         int max_index;
788
789         for(y=0; y<h; y++){
790             for(x=0; x<w; x++){
791                 int v, p=0;
792                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
793                 v= src[x + y*stride];
794
795                 if(y){
796                     t= src[x + (y-1)*stride];
797                     if(x){
798                         lt= src[x - 1 + (y-1)*stride];
799                     }
800                     if(x + 1 < w){
801                         rt= src[x + 1 + (y-1)*stride];
802                     }
803                 }
804                 if(x){
805                     l= src[x - 1 + y*stride];
806                     /*if(x > 1){
807                         if(orientation==1) ll= src[y + (x-2)*stride];
808                         else               ll= src[x - 2 + y*stride];
809                     }*/
810                 }
811                 if(parent){
812                     int px= x>>1;
813                     int py= y>>1;
814                     if(px<b->parent->width && py<b->parent->height)
815                         p= parent[px + py*2*stride];
816                 }
817                 if(!(/*ll|*/l|lt|t|rt|p)){
818                     if(v){
819                         runs[run_index++]= run;
820                         run=0;
821                     }else{
822                         run++;
823                     }
824                 }
825             }
826         }
827         max_index= run_index;
828         runs[run_index++]= run;
829         run_index=0;
830         run= runs[run_index++];
831
832         put_symbol2(&s->c, b->state[30], max_index, 0);
833         if(run_index <= max_index)
834             put_symbol2(&s->c, b->state[1], run, 3);
835
836         for(y=0; y<h; y++){
837             if(s->c.bytestream_end - s->c.bytestream < w*40){
838                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
839                 return AVERROR(ENOMEM);
840             }
841             for(x=0; x<w; x++){
842                 int v, p=0;
843                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
844                 v= src[x + y*stride];
845
846                 if(y){
847                     t= src[x + (y-1)*stride];
848                     if(x){
849                         lt= src[x - 1 + (y-1)*stride];
850                     }
851                     if(x + 1 < w){
852                         rt= src[x + 1 + (y-1)*stride];
853                     }
854                 }
855                 if(x){
856                     l= src[x - 1 + y*stride];
857                     /*if(x > 1){
858                         if(orientation==1) ll= src[y + (x-2)*stride];
859                         else               ll= src[x - 2 + y*stride];
860                     }*/
861                 }
862                 if(parent){
863                     int px= x>>1;
864                     int py= y>>1;
865                     if(px<b->parent->width && py<b->parent->height)
866                         p= parent[px + py*2*stride];
867                 }
868                 if(/*ll|*/l|lt|t|rt|p){
869                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
870
871                     put_rac(&s->c, &b->state[0][context], !!v);
872                 }else{
873                     if(!run){
874                         run= runs[run_index++];
875
876                         if(run_index <= max_index)
877                             put_symbol2(&s->c, b->state[1], run, 3);
878                         av_assert2(v);
879                     }else{
880                         run--;
881                         av_assert2(!v);
882                     }
883                 }
884                 if(v){
885                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
886                     int l2= 2*FFABS(l) + (l<0);
887                     int t2= 2*FFABS(t) + (t<0);
888
889                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
890                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
891                 }
892             }
893         }
894     }
895     return 0;
896 }
897
898 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
899 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
900 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
901     return encode_subband_c0run(s, b, src, parent, stride, orientation);
902 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
903 }
904
905 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){
906     const int b_stride= s->b_width << s->block_max_depth;
907     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
908     BlockNode backup= *block;
909     unsigned value;
910     int rd, index;
911
912     av_assert2(mb_x>=0 && mb_y>=0);
913     av_assert2(mb_x<b_stride);
914
915     if(intra){
916         block->color[0] = p[0];
917         block->color[1] = p[1];
918         block->color[2] = p[2];
919         block->type |= BLOCK_INTRA;
920     }else{
921         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
922         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
923         if(s->me_cache[index] == value)
924             return 0;
925         s->me_cache[index]= value;
926
927         block->mx= p[0];
928         block->my= p[1];
929         block->type &= ~BLOCK_INTRA;
930     }
931
932     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged) + s->intra_penalty * !!intra;
933
934 //FIXME chroma
935     if(rd < *best_rd){
936         *best_rd= rd;
937         return 1;
938     }else{
939         *block= backup;
940         return 0;
941     }
942 }
943
944 /* special case for int[2] args we discard afterwards,
945  * fixes compilation problem with gcc 2.95 */
946 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){
947     int p[2] = {p0, p1};
948     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
949 }
950
951 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){
952     const int b_stride= s->b_width << s->block_max_depth;
953     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
954     BlockNode backup[4];
955     unsigned value;
956     int rd, index;
957
958     /* We don't initialize backup[] during variable declaration, because
959      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
960      * 'int16_t'". */
961     backup[0] = block[0];
962     backup[1] = block[1];
963     backup[2] = block[b_stride];
964     backup[3] = block[b_stride + 1];
965
966     av_assert2(mb_x>=0 && mb_y>=0);
967     av_assert2(mb_x<b_stride);
968     av_assert2(((mb_x|mb_y)&1) == 0);
969
970     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
971     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
972     if(s->me_cache[index] == value)
973         return 0;
974     s->me_cache[index]= value;
975
976     block->mx= p0;
977     block->my= p1;
978     block->ref= ref;
979     block->type &= ~BLOCK_INTRA;
980     block[1]= block[b_stride]= block[b_stride+1]= *block;
981
982     rd= get_4block_rd(s, mb_x, mb_y, 0);
983
984 //FIXME chroma
985     if(rd < *best_rd){
986         *best_rd= rd;
987         return 1;
988     }else{
989         block[0]= backup[0];
990         block[1]= backup[1];
991         block[b_stride]= backup[2];
992         block[b_stride+1]= backup[3];
993         return 0;
994     }
995 }
996
997 static void iterative_me(SnowContext *s){
998     int pass, mb_x, mb_y;
999     const int b_width = s->b_width  << s->block_max_depth;
1000     const int b_height= s->b_height << s->block_max_depth;
1001     const int b_stride= b_width;
1002     int color[3];
1003
1004     {
1005         RangeCoder r = s->c;
1006         uint8_t state[sizeof(s->block_state)];
1007         memcpy(state, s->block_state, sizeof(s->block_state));
1008         for(mb_y= 0; mb_y<s->b_height; mb_y++)
1009             for(mb_x= 0; mb_x<s->b_width; mb_x++)
1010                 encode_q_branch(s, 0, mb_x, mb_y);
1011         s->c = r;
1012         memcpy(s->block_state, state, sizeof(s->block_state));
1013     }
1014
1015     for(pass=0; pass<25; pass++){
1016         int change= 0;
1017
1018         for(mb_y= 0; mb_y<b_height; mb_y++){
1019             for(mb_x= 0; mb_x<b_width; mb_x++){
1020                 int dia_change, i, j, ref;
1021                 int best_rd= INT_MAX, ref_rd;
1022                 BlockNode backup, ref_b;
1023                 const int index= mb_x + mb_y * b_stride;
1024                 BlockNode *block= &s->block[index];
1025                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1026                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1027                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1028                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1029                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1030                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1031                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1032                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1033                 const int b_w= (MB_SIZE >> s->block_max_depth);
1034                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1035
1036                 if(pass && (block->type & BLOCK_OPT))
1037                     continue;
1038                 block->type |= BLOCK_OPT;
1039
1040                 backup= *block;
1041
1042                 if(!s->me_cache_generation)
1043                     memset(s->me_cache, 0, sizeof(s->me_cache));
1044                 s->me_cache_generation += 1<<22;
1045
1046                 //FIXME precalculate
1047                 {
1048                     int x, y;
1049                     for (y = 0; y < b_w * 2; y++)
1050                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1051                     if(mb_x==0)
1052                         for(y=0; y<b_w*2; y++)
1053                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1054                     if(mb_x==b_stride-1)
1055                         for(y=0; y<b_w*2; y++)
1056                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1057                     if(mb_y==0){
1058                         for(x=0; x<b_w*2; x++)
1059                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1060                         for(y=1; y<b_w; y++)
1061                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1062                     }
1063                     if(mb_y==b_height-1){
1064                         for(x=0; x<b_w*2; x++)
1065                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1066                         for(y=b_w; y<b_w*2-1; y++)
1067                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1068                     }
1069                 }
1070
1071                 //skip stuff outside the picture
1072                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1073                     uint8_t *src= s->  input_picture->data[0];
1074                     uint8_t *dst= s->current_picture->data[0];
1075                     const int stride= s->current_picture->linesize[0];
1076                     const int block_w= MB_SIZE >> s->block_max_depth;
1077                     const int block_h= MB_SIZE >> s->block_max_depth;
1078                     const int sx= block_w*mb_x - block_w/2;
1079                     const int sy= block_h*mb_y - block_h/2;
1080                     const int w= s->plane[0].width;
1081                     const int h= s->plane[0].height;
1082                     int y;
1083
1084                     for(y=sy; y<0; y++)
1085                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1086                     for(y=h; y<sy+block_h*2; y++)
1087                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1088                     if(sx<0){
1089                         for(y=sy; y<sy+block_h*2; y++)
1090                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1091                     }
1092                     if(sx+block_w*2 > w){
1093                         for(y=sy; y<sy+block_h*2; y++)
1094                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1095                     }
1096                 }
1097
1098                 // intra(black) = neighbors' contribution to the current block
1099                 for(i=0; i < s->nb_planes; i++)
1100                     color[i]= get_dc(s, mb_x, mb_y, i);
1101
1102                 // get previous score (cannot be cached due to OBMC)
1103                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1104                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1105                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1106                 }else
1107                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1108
1109                 ref_b= *block;
1110                 ref_rd= best_rd;
1111                 for(ref=0; ref < s->ref_frames; ref++){
1112                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1113                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1114                         continue;
1115                     block->ref= ref;
1116                     best_rd= INT_MAX;
1117
1118                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1119                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1120                     if(tb)
1121                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1122                     if(lb)
1123                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1124                     if(rb)
1125                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1126                     if(bb)
1127                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1128
1129                     /* fullpel ME */
1130                     //FIXME avoid subpel interpolation / round to nearest integer
1131                     do{
1132                         int newx = block->mx;
1133                         int newy = block->my;
1134                         int dia_size = s->iterative_dia_size ? s->iterative_dia_size : FFMAX(s->avctx->dia_size, 1);
1135                         dia_change=0;
1136                         for(i=0; i < dia_size; i++){
1137                             for(j=0; j<i; j++){
1138                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+4*(i-j), newy+(4*j), obmc_edged, &best_rd);
1139                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-4*(i-j), newy-(4*j), obmc_edged, &best_rd);
1140                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx-(4*j), newy+4*(i-j), obmc_edged, &best_rd);
1141                                 dia_change |= check_block_inter(s, mb_x, mb_y, newx+(4*j), newy-4*(i-j), obmc_edged, &best_rd);
1142                             }
1143                         }
1144                     }while(dia_change);
1145                     /* subpel ME */
1146                     do{
1147                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1148                         dia_change=0;
1149                         for(i=0; i<8; i++)
1150                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1151                     }while(dia_change);
1152                     //FIXME or try the standard 2 pass qpel or similar
1153
1154                     mvr[0][0]= block->mx;
1155                     mvr[0][1]= block->my;
1156                     if(ref_rd > best_rd){
1157                         ref_rd= best_rd;
1158                         ref_b= *block;
1159                     }
1160                 }
1161                 best_rd= ref_rd;
1162                 *block= ref_b;
1163                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1164                 //FIXME RD style color selection
1165                 if(!same_block(block, &backup)){
1166                     if(tb ) tb ->type &= ~BLOCK_OPT;
1167                     if(lb ) lb ->type &= ~BLOCK_OPT;
1168                     if(rb ) rb ->type &= ~BLOCK_OPT;
1169                     if(bb ) bb ->type &= ~BLOCK_OPT;
1170                     if(tlb) tlb->type &= ~BLOCK_OPT;
1171                     if(trb) trb->type &= ~BLOCK_OPT;
1172                     if(blb) blb->type &= ~BLOCK_OPT;
1173                     if(brb) brb->type &= ~BLOCK_OPT;
1174                     change ++;
1175                 }
1176             }
1177         }
1178         av_log(s->avctx, AV_LOG_DEBUG, "pass:%d changed:%d\n", pass, change);
1179         if(!change)
1180             break;
1181     }
1182
1183     if(s->block_max_depth == 1){
1184         int change= 0;
1185         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1186             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1187                 int i;
1188                 int best_rd, init_rd;
1189                 const int index= mb_x + mb_y * b_stride;
1190                 BlockNode *b[4];
1191
1192                 b[0]= &s->block[index];
1193                 b[1]= b[0]+1;
1194                 b[2]= b[0]+b_stride;
1195                 b[3]= b[2]+1;
1196                 if(same_block(b[0], b[1]) &&
1197                    same_block(b[0], b[2]) &&
1198                    same_block(b[0], b[3]))
1199                     continue;
1200
1201                 if(!s->me_cache_generation)
1202                     memset(s->me_cache, 0, sizeof(s->me_cache));
1203                 s->me_cache_generation += 1<<22;
1204
1205                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1206
1207                 //FIXME more multiref search?
1208                 check_4block_inter(s, mb_x, mb_y,
1209                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1210                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1211
1212                 for(i=0; i<4; i++)
1213                     if(!(b[i]->type&BLOCK_INTRA))
1214                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1215
1216                 if(init_rd != best_rd)
1217                     change++;
1218             }
1219         }
1220         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1221     }
1222 }
1223
1224 static void encode_blocks(SnowContext *s, int search){
1225     int x, y;
1226     int w= s->b_width;
1227     int h= s->b_height;
1228
1229     if(s->motion_est == FF_ME_ITER && !s->keyframe && search)
1230         iterative_me(s);
1231
1232     for(y=0; y<h; y++){
1233         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1234             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1235             return;
1236         }
1237         for(x=0; x<w; x++){
1238             if(s->motion_est == FF_ME_ITER || !search)
1239                 encode_q_branch2(s, 0, x, y);
1240             else
1241                 encode_q_branch (s, 0, x, y);
1242         }
1243     }
1244 }
1245
1246 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1247     const int w= b->width;
1248     const int h= b->height;
1249     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1250     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1251     int x,y, thres1, thres2;
1252
1253     if(s->qlog == LOSSLESS_QLOG){
1254         for(y=0; y<h; y++)
1255             for(x=0; x<w; x++)
1256                 dst[x + y*stride]= src[x + y*stride];
1257         return;
1258     }
1259
1260     bias= bias ? 0 : (3*qmul)>>3;
1261     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1262     thres2= 2*thres1;
1263
1264     if(!bias){
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/= qmul; //FIXME optimize
1273                         dst[x + y*stride]=  i;
1274                     }else{
1275                         i= -i;
1276                         i<<= QEXPSHIFT;
1277                         i/= qmul; //FIXME optimize
1278                         dst[x + y*stride]= -i;
1279                     }
1280                 }else
1281                     dst[x + y*stride]= 0;
1282             }
1283         }
1284     }else{
1285         for(y=0; y<h; y++){
1286             for(x=0; x<w; x++){
1287                 int i= src[x + y*stride];
1288
1289                 if((unsigned)(i+thres1) > thres2){
1290                     if(i>=0){
1291                         i<<= QEXPSHIFT;
1292                         i= (i + bias) / qmul; //FIXME optimize
1293                         dst[x + y*stride]=  i;
1294                     }else{
1295                         i= -i;
1296                         i<<= QEXPSHIFT;
1297                         i= (i + bias) / qmul; //FIXME optimize
1298                         dst[x + y*stride]= -i;
1299                     }
1300                 }else
1301                     dst[x + y*stride]= 0;
1302             }
1303         }
1304     }
1305 }
1306
1307 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1308     const int w= b->width;
1309     const int h= b->height;
1310     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1311     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1312     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1313     int x,y;
1314
1315     if(s->qlog == LOSSLESS_QLOG) return;
1316
1317     for(y=0; y<h; y++){
1318         for(x=0; x<w; x++){
1319             int i= src[x + y*stride];
1320             if(i<0){
1321                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1322             }else if(i>0){
1323                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1324             }
1325         }
1326     }
1327 }
1328
1329 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1330     const int w= b->width;
1331     const int h= b->height;
1332     int x,y;
1333
1334     for(y=h-1; y>=0; y--){
1335         for(x=w-1; x>=0; x--){
1336             int i= x + y*stride;
1337
1338             if(x){
1339                 if(use_median){
1340                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1341                     else  src[i] -= src[i - 1];
1342                 }else{
1343                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1344                     else  src[i] -= src[i - 1];
1345                 }
1346             }else{
1347                 if(y) src[i] -= src[i - stride];
1348             }
1349         }
1350     }
1351 }
1352
1353 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1354     const int w= b->width;
1355     const int h= b->height;
1356     int x,y;
1357
1358     for(y=0; y<h; y++){
1359         for(x=0; x<w; x++){
1360             int i= x + y*stride;
1361
1362             if(x){
1363                 if(use_median){
1364                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1365                     else  src[i] += src[i - 1];
1366                 }else{
1367                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1368                     else  src[i] += src[i - 1];
1369                 }
1370             }else{
1371                 if(y) src[i] += src[i - stride];
1372             }
1373         }
1374     }
1375 }
1376
1377 static void encode_qlogs(SnowContext *s){
1378     int plane_index, level, orientation;
1379
1380     for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1381         for(level=0; level<s->spatial_decomposition_count; level++){
1382             for(orientation=level ? 1:0; orientation<4; orientation++){
1383                 if(orientation==2) continue;
1384                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1385             }
1386         }
1387     }
1388 }
1389
1390 static void encode_header(SnowContext *s){
1391     int plane_index, i;
1392     uint8_t kstate[32];
1393
1394     memset(kstate, MID_STATE, sizeof(kstate));
1395
1396     put_rac(&s->c, kstate, s->keyframe);
1397     if(s->keyframe || s->always_reset){
1398         ff_snow_reset_contexts(s);
1399         s->last_spatial_decomposition_type=
1400         s->last_qlog=
1401         s->last_qbias=
1402         s->last_mv_scale=
1403         s->last_block_max_depth= 0;
1404         for(plane_index=0; plane_index<2; plane_index++){
1405             Plane *p= &s->plane[plane_index];
1406             p->last_htaps=0;
1407             p->last_diag_mc=0;
1408             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1409         }
1410     }
1411     if(s->keyframe){
1412         put_symbol(&s->c, s->header_state, s->version, 0);
1413         put_rac(&s->c, s->header_state, s->always_reset);
1414         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1415         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1416         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1417         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1418         if (s->nb_planes > 2) {
1419             put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1420             put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1421         }
1422         put_rac(&s->c, s->header_state, s->spatial_scalability);
1423 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1424         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1425
1426         encode_qlogs(s);
1427     }
1428
1429     if(!s->keyframe){
1430         int update_mc=0;
1431         for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1432             Plane *p= &s->plane[plane_index];
1433             update_mc |= p->last_htaps   != p->htaps;
1434             update_mc |= p->last_diag_mc != p->diag_mc;
1435             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1436         }
1437         put_rac(&s->c, s->header_state, update_mc);
1438         if(update_mc){
1439             for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1440                 Plane *p= &s->plane[plane_index];
1441                 put_rac(&s->c, s->header_state, p->diag_mc);
1442                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1443                 for(i= p->htaps/2; i; i--)
1444                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1445             }
1446         }
1447         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1448             put_rac(&s->c, s->header_state, 1);
1449             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1450             encode_qlogs(s);
1451         }else
1452             put_rac(&s->c, s->header_state, 0);
1453     }
1454
1455     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1456     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1457     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1458     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1459     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1460
1461 }
1462
1463 static void update_last_header_values(SnowContext *s){
1464     int plane_index;
1465
1466     if(!s->keyframe){
1467         for(plane_index=0; plane_index<2; plane_index++){
1468             Plane *p= &s->plane[plane_index];
1469             p->last_diag_mc= p->diag_mc;
1470             p->last_htaps  = p->htaps;
1471             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1472         }
1473     }
1474
1475     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1476     s->last_qlog                        = s->qlog;
1477     s->last_qbias                       = s->qbias;
1478     s->last_mv_scale                    = s->mv_scale;
1479     s->last_block_max_depth             = s->block_max_depth;
1480     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1481 }
1482
1483 static int qscale2qlog(int qscale){
1484     return lrint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1485            + 61*QROOT/8; ///< 64 > 60
1486 }
1487
1488 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1489 {
1490     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1491      * FIXME we know exact mv bits at this point,
1492      * but ratecontrol isn't set up to include them. */
1493     uint32_t coef_sum= 0;
1494     int level, orientation, delta_qlog;
1495
1496     for(level=0; level<s->spatial_decomposition_count; level++){
1497         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1498             SubBand *b= &s->plane[0].band[level][orientation];
1499             IDWTELEM *buf= b->ibuf;
1500             const int w= b->width;
1501             const int h= b->height;
1502             const int stride= b->stride;
1503             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1504             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1505             const int qdiv= (1<<16)/qmul;
1506             int x, y;
1507             //FIXME this is ugly
1508             for(y=0; y<h; y++)
1509                 for(x=0; x<w; x++)
1510                     buf[x+y*stride]= b->buf[x+y*stride];
1511             if(orientation==0)
1512                 decorrelate(s, b, buf, stride, 1, 0);
1513             for(y=0; y<h; y++)
1514                 for(x=0; x<w; x++)
1515                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1516         }
1517     }
1518
1519     /* ugly, ratecontrol just takes a sqrt again */
1520     av_assert0(coef_sum < INT_MAX);
1521     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1522
1523     if(pict->pict_type == AV_PICTURE_TYPE_I){
1524         s->m.current_picture.mb_var_sum= coef_sum;
1525         s->m.current_picture.mc_mb_var_sum= 0;
1526     }else{
1527         s->m.current_picture.mc_mb_var_sum= coef_sum;
1528         s->m.current_picture.mb_var_sum= 0;
1529     }
1530
1531     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1532     if (pict->quality < 0)
1533         return INT_MIN;
1534     s->lambda= pict->quality * 3/2;
1535     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1536     s->qlog+= delta_qlog;
1537     return delta_qlog;
1538 }
1539
1540 static void calculate_visual_weight(SnowContext *s, Plane *p){
1541     int width = p->width;
1542     int height= p->height;
1543     int level, orientation, x, y;
1544
1545     for(level=0; level<s->spatial_decomposition_count; level++){
1546         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1547             SubBand *b= &p->band[level][orientation];
1548             IDWTELEM *ibuf= b->ibuf;
1549             int64_t error=0;
1550
1551             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1552             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1553             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1554             for(y=0; y<height; y++){
1555                 for(x=0; x<width; x++){
1556                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1557                     error += d*d;
1558                 }
1559             }
1560
1561             b->qlog= (int)(QROOT * log2(352256.0/sqrt(error)) + 0.5);
1562         }
1563     }
1564 }
1565
1566 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1567                         const AVFrame *pict, int *got_packet)
1568 {
1569     SnowContext *s = avctx->priv_data;
1570     RangeCoder * const c= &s->c;
1571     AVFrame *pic;
1572     const int width= s->avctx->width;
1573     const int height= s->avctx->height;
1574     int level, orientation, plane_index, i, y, ret;
1575     uint8_t rc_header_bak[sizeof(s->header_state)];
1576     uint8_t rc_block_bak[sizeof(s->block_state)];
1577
1578     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)
1579         return ret;
1580
1581     ff_init_range_encoder(c, pkt->data, pkt->size);
1582     ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1583
1584     for(i=0; i < s->nb_planes; i++){
1585         int hshift= i ? s->chroma_h_shift : 0;
1586         int vshift= i ? s->chroma_v_shift : 0;
1587         for(y=0; y<AV_CEIL_RSHIFT(height, vshift); y++)
1588             memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1589                    &pict->data[i][y * pict->linesize[i]],
1590                    AV_CEIL_RSHIFT(width, hshift));
1591         s->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1592                                 AV_CEIL_RSHIFT(width, hshift), AV_CEIL_RSHIFT(height, vshift),
1593                                 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1594                                 EDGE_TOP | EDGE_BOTTOM);
1595
1596     }
1597     emms_c();
1598     pic = s->input_picture;
1599     pic->pict_type = pict->pict_type;
1600     pic->quality = pict->quality;
1601
1602     s->m.picture_number= avctx->frame_number;
1603     if(avctx->flags&AV_CODEC_FLAG_PASS2){
1604         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1605         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1606         if(!(avctx->flags&AV_CODEC_FLAG_QSCALE)) {
1607             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1608             if (pic->quality < 0)
1609                 return -1;
1610         }
1611     }else{
1612         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1613         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1614     }
1615
1616     if(s->pass1_rc && avctx->frame_number == 0)
1617         pic->quality = 2*FF_QP2LAMBDA;
1618     if (pic->quality) {
1619         s->qlog   = qscale2qlog(pic->quality);
1620         s->lambda = pic->quality * 3/2;
1621     }
1622     if (s->qlog < 0 || (!pic->quality && (avctx->flags & AV_CODEC_FLAG_QSCALE))) {
1623         s->qlog= LOSSLESS_QLOG;
1624         s->lambda = 0;
1625     }//else keep previous frame's qlog until after motion estimation
1626
1627     if (s->current_picture->data[0]) {
1628         int w = s->avctx->width;
1629         int h = s->avctx->height;
1630
1631         s->mpvencdsp.draw_edges(s->current_picture->data[0],
1632                                 s->current_picture->linesize[0], w   , h   ,
1633                                 EDGE_WIDTH  , EDGE_WIDTH  , EDGE_TOP | EDGE_BOTTOM);
1634         if (s->current_picture->data[2]) {
1635             s->mpvencdsp.draw_edges(s->current_picture->data[1],
1636                                     s->current_picture->linesize[1], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1637                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1638             s->mpvencdsp.draw_edges(s->current_picture->data[2],
1639                                     s->current_picture->linesize[2], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1640                                     EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1641         }
1642         emms_c();
1643     }
1644
1645     ff_snow_frame_start(s);
1646 #if FF_API_CODED_FRAME
1647 FF_DISABLE_DEPRECATION_WARNINGS
1648     av_frame_unref(avctx->coded_frame);
1649     ret = av_frame_ref(avctx->coded_frame, s->current_picture);
1650 FF_ENABLE_DEPRECATION_WARNINGS
1651 #endif
1652     if (ret < 0)
1653         return ret;
1654
1655     s->m.current_picture_ptr= &s->m.current_picture;
1656     s->m.current_picture.f = s->current_picture;
1657     s->m.current_picture.f->pts = pict->pts;
1658     if(pic->pict_type == AV_PICTURE_TYPE_P){
1659         int block_width = (width +15)>>4;
1660         int block_height= (height+15)>>4;
1661         int stride= s->current_picture->linesize[0];
1662
1663         av_assert0(s->current_picture->data[0]);
1664         av_assert0(s->last_picture[0]->data[0]);
1665
1666         s->m.avctx= s->avctx;
1667         s->m.   last_picture.f = s->last_picture[0];
1668         s->m.    new_picture.f = s->input_picture;
1669         s->m.   last_picture_ptr= &s->m.   last_picture;
1670         s->m.linesize = stride;
1671         s->m.uvlinesize= s->current_picture->linesize[1];
1672         s->m.width = width;
1673         s->m.height= height;
1674         s->m.mb_width = block_width;
1675         s->m.mb_height= block_height;
1676         s->m.mb_stride=   s->m.mb_width+1;
1677         s->m.b8_stride= 2*s->m.mb_width+1;
1678         s->m.f_code=1;
1679         s->m.pict_type = pic->pict_type;
1680         s->m.motion_est= s->motion_est;
1681         s->m.me.scene_change_score=0;
1682         s->m.me.dia_size = avctx->dia_size;
1683         s->m.quarter_sample= (s->avctx->flags & AV_CODEC_FLAG_QPEL)!=0;
1684         s->m.out_format= FMT_H263;
1685         s->m.unrestricted_mv= 1;
1686
1687         s->m.lambda = s->lambda;
1688         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1689         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1690
1691         s->m.mecc= s->mecc; //move
1692         s->m.qdsp= s->qdsp; //move
1693         s->m.hdsp = s->hdsp;
1694         ff_init_me(&s->m);
1695         s->hdsp = s->m.hdsp;
1696         s->mecc= s->m.mecc;
1697     }
1698
1699     if(s->pass1_rc){
1700         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1701         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1702     }
1703
1704 redo_frame:
1705
1706     s->spatial_decomposition_count= 5;
1707
1708     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1709           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1710         s->spatial_decomposition_count--;
1711
1712     if (s->spatial_decomposition_count <= 0) {
1713         av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1714         return AVERROR(EINVAL);
1715     }
1716
1717     s->m.pict_type = pic->pict_type;
1718     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1719
1720     ff_snow_common_init_after_header(avctx);
1721
1722     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1723         for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1724             calculate_visual_weight(s, &s->plane[plane_index]);
1725         }
1726     }
1727
1728     encode_header(s);
1729     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1730     encode_blocks(s, 1);
1731     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1732
1733     for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1734         Plane *p= &s->plane[plane_index];
1735         int w= p->width;
1736         int h= p->height;
1737         int x, y;
1738 //        int bits= put_bits_count(&s->c.pb);
1739
1740         if (!s->memc_only) {
1741             //FIXME optimize
1742             if(pict->data[plane_index]) //FIXME gray hack
1743                 for(y=0; y<h; y++){
1744                     for(x=0; x<w; x++){
1745                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1746                     }
1747                 }
1748             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1749
1750 #if FF_API_PRIVATE_OPT
1751 FF_DISABLE_DEPRECATION_WARNINGS
1752             if(s->avctx->scenechange_threshold)
1753                 s->scenechange_threshold = s->avctx->scenechange_threshold;
1754 FF_ENABLE_DEPRECATION_WARNINGS
1755 #endif
1756
1757             if(   plane_index==0
1758                && pic->pict_type == AV_PICTURE_TYPE_P
1759                && !(avctx->flags&AV_CODEC_FLAG_PASS2)
1760                && s->m.me.scene_change_score > s->scenechange_threshold){
1761                 ff_init_range_encoder(c, pkt->data, pkt->size);
1762                 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1763                 pic->pict_type= AV_PICTURE_TYPE_I;
1764                 s->keyframe=1;
1765                 s->current_picture->key_frame=1;
1766                 goto redo_frame;
1767             }
1768
1769             if(s->qlog == LOSSLESS_QLOG){
1770                 for(y=0; y<h; y++){
1771                     for(x=0; x<w; x++){
1772                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1773                     }
1774                 }
1775             }else{
1776                 for(y=0; y<h; y++){
1777                     for(x=0; x<w; x++){
1778                         s->spatial_dwt_buffer[y*w + x]= s->spatial_idwt_buffer[y*w + x] * (1 << ENCODER_EXTRA_BITS);
1779                     }
1780                 }
1781             }
1782
1783             ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1784
1785             if(s->pass1_rc && plane_index==0){
1786                 int delta_qlog = ratecontrol_1pass(s, pic);
1787                 if (delta_qlog <= INT_MIN)
1788                     return -1;
1789                 if(delta_qlog){
1790                     //reordering qlog in the bitstream would eliminate this reset
1791                     ff_init_range_encoder(c, pkt->data, pkt->size);
1792                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1793                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1794                     encode_header(s);
1795                     encode_blocks(s, 0);
1796                 }
1797             }
1798
1799             for(level=0; level<s->spatial_decomposition_count; level++){
1800                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1801                     SubBand *b= &p->band[level][orientation];
1802
1803                     quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1804                     if(orientation==0)
1805                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1806                     if (!s->no_bitstream)
1807                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1808                     av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1809                     if(orientation==0)
1810                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1811                 }
1812             }
1813
1814             for(level=0; level<s->spatial_decomposition_count; level++){
1815                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1816                     SubBand *b= &p->band[level][orientation];
1817
1818                     dequantize(s, b, b->ibuf, b->stride);
1819                 }
1820             }
1821
1822             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1823             if(s->qlog == LOSSLESS_QLOG){
1824                 for(y=0; y<h; y++){
1825                     for(x=0; x<w; x++){
1826                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1827                     }
1828                 }
1829             }
1830             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1831         }else{
1832             //ME/MC only
1833             if(pic->pict_type == AV_PICTURE_TYPE_I){
1834                 for(y=0; y<h; y++){
1835                     for(x=0; x<w; x++){
1836                         s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1837                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1838                     }
1839                 }
1840             }else{
1841                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1842                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1843             }
1844         }
1845         if(s->avctx->flags&AV_CODEC_FLAG_PSNR){
1846             int64_t error= 0;
1847
1848             if(pict->data[plane_index]) //FIXME gray hack
1849                 for(y=0; y<h; y++){
1850                     for(x=0; x<w; x++){
1851                         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];
1852                         error += d*d;
1853                     }
1854                 }
1855             s->avctx->error[plane_index] += error;
1856             s->encoding_error[plane_index] = error;
1857         }
1858
1859     }
1860     emms_c();
1861
1862     update_last_header_values(s);
1863
1864     ff_snow_release_buffer(avctx);
1865
1866     s->current_picture->coded_picture_number = avctx->frame_number;
1867     s->current_picture->pict_type = pic->pict_type;
1868     s->current_picture->quality = pic->quality;
1869     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1870     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1871     s->m.current_picture.f->display_picture_number =
1872     s->m.current_picture.f->coded_picture_number   = avctx->frame_number;
1873     s->m.current_picture.f->quality                = pic->quality;
1874     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1875     if(s->pass1_rc)
1876         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1877             return -1;
1878     if(avctx->flags&AV_CODEC_FLAG_PASS1)
1879         ff_write_pass1_stats(&s->m);
1880     s->m.last_pict_type = s->m.pict_type;
1881 #if FF_API_STAT_BITS
1882 FF_DISABLE_DEPRECATION_WARNINGS
1883     avctx->frame_bits = s->m.frame_bits;
1884     avctx->mv_bits = s->m.mv_bits;
1885     avctx->misc_bits = s->m.misc_bits;
1886     avctx->p_tex_bits = s->m.p_tex_bits;
1887 FF_ENABLE_DEPRECATION_WARNINGS
1888 #endif
1889
1890     emms_c();
1891
1892     ff_side_data_set_encoder_stats(pkt, s->current_picture->quality,
1893                                    s->encoding_error,
1894                                    (s->avctx->flags&AV_CODEC_FLAG_PSNR) ? 4 : 0,
1895                                    s->current_picture->pict_type);
1896
1897 #if FF_API_ERROR_FRAME
1898 FF_DISABLE_DEPRECATION_WARNINGS
1899     memcpy(s->current_picture->error, s->encoding_error, sizeof(s->encoding_error));
1900 FF_ENABLE_DEPRECATION_WARNINGS
1901 #endif
1902
1903     pkt->size = ff_rac_terminate(c, 0);
1904     if (s->current_picture->key_frame)
1905         pkt->flags |= AV_PKT_FLAG_KEY;
1906     *got_packet = 1;
1907
1908     return 0;
1909 }
1910
1911 static av_cold int encode_end(AVCodecContext *avctx)
1912 {
1913     SnowContext *s = avctx->priv_data;
1914
1915     ff_snow_common_end(s);
1916     ff_rate_control_uninit(&s->m);
1917     av_frame_free(&s->input_picture);
1918     av_freep(&avctx->stats_out);
1919
1920     return 0;
1921 }
1922
1923 #define OFFSET(x) offsetof(SnowContext, x)
1924 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1925 static const AVOption options[] = {
1926     {"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" },
1927     { "zero", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ZERO }, 0, 0, VE, "motion_est" },
1928     { "epzs", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_EPZS }, 0, 0, VE, "motion_est" },
1929     { "xone", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_XONE }, 0, 0, VE, "motion_est" },
1930     { "iter", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ITER }, 0, 0, VE, "motion_est" },
1931     { "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 },
1932     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
1933     { "intra_penalty",  "Penalty for intra blocks in block decission",      OFFSET(intra_penalty), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1934     { "iterative_dia_size",  "Dia size for the iterative ME",          OFFSET(iterative_dia_size), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1935     { "sc_threshold",   "Scene change threshold",                   OFFSET(scenechange_threshold), AV_OPT_TYPE_INT, { .i64 = 0 }, INT_MIN, INT_MAX, VE },
1936     { "pred",           "Spatial decomposition type",                                OFFSET(pred), AV_OPT_TYPE_INT, { .i64 = 0 }, DWT_97, DWT_53, VE, "pred" },
1937         { "dwt97", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, INT_MIN, INT_MAX, VE, "pred" },
1938         { "dwt53", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN, INT_MAX, VE, "pred" },
1939     { NULL },
1940 };
1941
1942 static const AVClass snowenc_class = {
1943     .class_name = "snow encoder",
1944     .item_name  = av_default_item_name,
1945     .option     = options,
1946     .version    = LIBAVUTIL_VERSION_INT,
1947 };
1948
1949 AVCodec ff_snow_encoder = {
1950     .name           = "snow",
1951     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1952     .type           = AVMEDIA_TYPE_VIDEO,
1953     .id             = AV_CODEC_ID_SNOW,
1954     .priv_data_size = sizeof(SnowContext),
1955     .init           = encode_init,
1956     .encode2        = encode_frame,
1957     .close          = encode_end,
1958     .pix_fmts       = (const enum AVPixelFormat[]){
1959         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1960         AV_PIX_FMT_GRAY8,
1961         AV_PIX_FMT_NONE
1962     },
1963     .priv_class     = &snowenc_class,
1964     .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE |
1965                       FF_CODEC_CAP_INIT_CLEANUP,
1966 };