Merge commit '368f50359eb328b0b9d67451f56fda20b3255f9a'
[ffmpeg.git] / libavcodec / snowenc.c
1 /*
2  * Copyright (C) 2004 Michael Niedermayer <michaelni@gmx.at>
3  *
4  * This file is part of FFmpeg.
5  *
6  * FFmpeg is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * FFmpeg is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with FFmpeg; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20
21 #include "libavutil/intmath.h"
22 #include "libavutil/log.h"
23 #include "libavutil/opt.h"
24 #include "avcodec.h"
25 #include "dsputil.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
41     if(avctx->prediction_method == DWT_97
42        && (avctx->flags & CODEC_FLAG_QSCALE)
43        && avctx->global_quality == 0){
44         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
45         return -1;
46     }
47
48     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
49
50     s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
51     s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
52
53     for(plane_index=0; plane_index<3; plane_index++){
54         s->plane[plane_index].diag_mc= 1;
55         s->plane[plane_index].htaps= 6;
56         s->plane[plane_index].hcoeff[0]=  40;
57         s->plane[plane_index].hcoeff[1]= -10;
58         s->plane[plane_index].hcoeff[2]=   2;
59         s->plane[plane_index].fast_mc= 1;
60     }
61
62     if ((ret = ff_snow_common_init(avctx)) < 0) {
63         ff_snow_common_end(avctx->priv_data);
64         return ret;
65     }
66     ff_snow_alloc_blocks(s);
67
68     s->version=0;
69
70     s->m.avctx   = avctx;
71     s->m.flags   = avctx->flags;
72     s->m.bit_rate= avctx->bit_rate;
73
74     s->m.me.temp      =
75     s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
76     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
77     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
78     s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
79     if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.obmc_scratchpad)
80         return AVERROR(ENOMEM);
81
82     ff_h263_encode_init(&s->m); //mv_penalty
83
84     s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
85
86     if(avctx->flags&CODEC_FLAG_PASS1){
87         if(!avctx->stats_out)
88             avctx->stats_out = av_mallocz(256);
89
90         if (!avctx->stats_out)
91             return AVERROR(ENOMEM);
92     }
93     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
94         if(ff_rate_control_init(&s->m) < 0)
95             return -1;
96     }
97     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
98
99     switch(avctx->pix_fmt){
100     case AV_PIX_FMT_YUV444P:
101 //    case AV_PIX_FMT_YUV422P:
102     case AV_PIX_FMT_YUV420P:
103 //    case AV_PIX_FMT_YUV411P:
104     case AV_PIX_FMT_YUV410P:
105         s->nb_planes = 3;
106         s->colorspace_type= 0;
107         break;
108     case AV_PIX_FMT_GRAY8:
109         s->nb_planes = 1;
110         s->colorspace_type = 1;
111         break;
112 /*    case AV_PIX_FMT_RGB32:
113         s->colorspace= 1;
114         break;*/
115     default:
116         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
117         return -1;
118     }
119     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
120
121     ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
122     ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
123
124     s->input_picture = av_frame_alloc();
125     if (!s->input_picture)
126         return AVERROR(ENOMEM);
127     if ((ret = ff_get_buffer(s->avctx, s->input_picture, AV_GET_BUFFER_FLAG_REF)) < 0)
128         return ret;
129
130     if(s->avctx->me_method == ME_ITER){
131         int i;
132         int size= s->b_width * s->b_height << 2*s->block_max_depth;
133         for(i=0; i<s->max_ref_frames; i++){
134             s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
135             s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
136             if (!s->ref_mvs[i] || !s->ref_scores[i])
137                 return AVERROR(ENOMEM);
138         }
139     }
140
141     return 0;
142 }
143
144 //near copy & paste from dsputil, FIXME
145 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
146 {
147     int s, i, j;
148
149     s = 0;
150     for (i = 0; i < h; i++) {
151         for (j = 0; j < w; j++) {
152             s += pix[0];
153             pix ++;
154         }
155         pix += line_size - w;
156     }
157     return s;
158 }
159
160 //near copy & paste from dsputil, FIXME
161 static int pix_norm1(uint8_t * pix, int line_size, int w)
162 {
163     int s, i, j;
164     uint32_t *sq = ff_square_tab + 256;
165
166     s = 0;
167     for (i = 0; i < w; i++) {
168         for (j = 0; j < w; j ++) {
169             s += sq[pix[0]];
170             pix ++;
171         }
172         pix += line_size - w;
173     }
174     return s;
175 }
176
177 static inline int get_penalty_factor(int lambda, int lambda2, int type){
178     switch(type&0xFF){
179     default:
180     case FF_CMP_SAD:
181         return lambda>>FF_LAMBDA_SHIFT;
182     case FF_CMP_DCT:
183         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
184     case FF_CMP_W53:
185         return (4*lambda)>>(FF_LAMBDA_SHIFT);
186     case FF_CMP_W97:
187         return (2*lambda)>>(FF_LAMBDA_SHIFT);
188     case FF_CMP_SATD:
189     case FF_CMP_DCT264:
190         return (2*lambda)>>FF_LAMBDA_SHIFT;
191     case FF_CMP_RD:
192     case FF_CMP_PSNR:
193     case FF_CMP_SSE:
194     case FF_CMP_NSSE:
195         return lambda2>>FF_LAMBDA_SHIFT;
196     case FF_CMP_BIT:
197         return 1;
198     }
199 }
200
201 //FIXME copy&paste
202 #define P_LEFT P[1]
203 #define P_TOP P[2]
204 #define P_TOPRIGHT P[3]
205 #define P_MEDIAN P[4]
206 #define P_MV1 P[9]
207 #define FLAG_QPEL   1 //must be 1
208
209 static int encode_q_branch(SnowContext *s, int level, int x, int y){
210     uint8_t p_buffer[1024];
211     uint8_t i_buffer[1024];
212     uint8_t p_state[sizeof(s->block_state)];
213     uint8_t i_state[sizeof(s->block_state)];
214     RangeCoder pc, ic;
215     uint8_t *pbbak= s->c.bytestream;
216     uint8_t *pbbak_start= s->c.bytestream_start;
217     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
218     const int w= s->b_width  << s->block_max_depth;
219     const int h= s->b_height << s->block_max_depth;
220     const int rem_depth= s->block_max_depth - level;
221     const int index= (x + y*w) << rem_depth;
222     const int block_w= 1<<(LOG2_MB_SIZE - level);
223     int trx= (x+1)<<rem_depth;
224     int try= (y+1)<<rem_depth;
225     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
226     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
227     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
228     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
229     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
230     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
231     int pl = left->color[0];
232     int pcb= left->color[1];
233     int pcr= left->color[2];
234     int pmx, pmy;
235     int mx=0, my=0;
236     int l,cr,cb;
237     const int stride= s->current_picture->linesize[0];
238     const int uvstride= s->current_picture->linesize[1];
239     uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y*  stride)*block_w,
240                                 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
241                                 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
242     int P[10][2];
243     int16_t last_mv[3][2];
244     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
245     const int shift= 1+qpel;
246     MotionEstContext *c= &s->m.me;
247     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
248     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
249     int my_context= av_log2(2*FFABS(left->my - top->my));
250     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
251     int ref, best_ref, ref_score, ref_mx, ref_my;
252
253     av_assert0(sizeof(s->block_state) >= 256);
254     if(s->keyframe){
255         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
256         return 0;
257     }
258
259 //    clip predictors / edge ?
260
261     P_LEFT[0]= left->mx;
262     P_LEFT[1]= left->my;
263     P_TOP [0]= top->mx;
264     P_TOP [1]= top->my;
265     P_TOPRIGHT[0]= tr->mx;
266     P_TOPRIGHT[1]= tr->my;
267
268     last_mv[0][0]= s->block[index].mx;
269     last_mv[0][1]= s->block[index].my;
270     last_mv[1][0]= right->mx;
271     last_mv[1][1]= right->my;
272     last_mv[2][0]= bottom->mx;
273     last_mv[2][1]= bottom->my;
274
275     s->m.mb_stride=2;
276     s->m.mb_x=
277     s->m.mb_y= 0;
278     c->skip= 0;
279
280     av_assert1(c->  stride ==   stride);
281     av_assert1(c->uvstride == uvstride);
282
283     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
284     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
285     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
286     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
287
288     c->xmin = - x*block_w - 16+3;
289     c->ymin = - y*block_w - 16+3;
290     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
291     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
292
293     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
294     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
295     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
296     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
297     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
298     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
299     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
300
301     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
302     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
303
304     if (!y) {
305         c->pred_x= P_LEFT[0];
306         c->pred_y= P_LEFT[1];
307     } else {
308         c->pred_x = P_MEDIAN[0];
309         c->pred_y = P_MEDIAN[1];
310     }
311
312     score= INT_MAX;
313     best_ref= 0;
314     for(ref=0; ref<s->ref_frames; ref++){
315         init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
316
317         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
318                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
319
320         av_assert2(ref_mx >= c->xmin);
321         av_assert2(ref_mx <= c->xmax);
322         av_assert2(ref_my >= c->ymin);
323         av_assert2(ref_my <= c->ymax);
324
325         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
326         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
327         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
328         if(s->ref_mvs[ref]){
329             s->ref_mvs[ref][index][0]= ref_mx;
330             s->ref_mvs[ref][index][1]= ref_my;
331             s->ref_scores[ref][index]= ref_score;
332         }
333         if(score > ref_score){
334             score= ref_score;
335             best_ref= ref;
336             mx= ref_mx;
337             my= ref_my;
338         }
339     }
340     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
341
342   //  subpel search
343     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
344     pc= s->c;
345     pc.bytestream_start=
346     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
347     memcpy(p_state, s->block_state, sizeof(s->block_state));
348
349     if(level!=s->block_max_depth)
350         put_rac(&pc, &p_state[4 + s_context], 1);
351     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
352     if(s->ref_frames > 1)
353         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
354     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
355     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
356     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
357     p_len= pc.bytestream - pc.bytestream_start;
358     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
359
360     block_s= block_w*block_w;
361     sum = pix_sum(current_data[0], stride, block_w, block_w);
362     l= (sum + block_s/2)/block_s;
363     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
364
365     if (s->nb_planes > 2) {
366         block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
367         sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
368         cb= (sum + block_s/2)/block_s;
369     //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
370         sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
371         cr= (sum + block_s/2)/block_s;
372     //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
373     }else
374         cb = cr = 0;
375
376     ic= s->c;
377     ic.bytestream_start=
378     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
379     memcpy(i_state, s->block_state, sizeof(s->block_state));
380     if(level!=s->block_max_depth)
381         put_rac(&ic, &i_state[4 + s_context], 1);
382     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
383     put_symbol(&ic, &i_state[32],  l-pl , 1);
384     if (s->nb_planes > 2) {
385         put_symbol(&ic, &i_state[64], cb-pcb, 1);
386         put_symbol(&ic, &i_state[96], cr-pcr, 1);
387     }
388     i_len= ic.bytestream - ic.bytestream_start;
389     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
390
391 //    assert(score==256*256*256*64-1);
392     av_assert1(iscore < 255*255*256 + s->lambda2*10);
393     av_assert1(iscore >= 0);
394     av_assert1(l>=0 && l<=255);
395     av_assert1(pl>=0 && pl<=255);
396
397     if(level==0){
398         int varc= iscore >> 8;
399         int vard= score >> 8;
400         if (vard <= 64 || vard < varc)
401             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
402         else
403             c->scene_change_score+= s->m.qscale;
404     }
405
406     if(level!=s->block_max_depth){
407         put_rac(&s->c, &s->block_state[4 + s_context], 0);
408         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
409         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
410         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
411         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
412         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
413
414         if(score2 < score && score2 < iscore)
415             return score2;
416     }
417
418     if(iscore < score){
419         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
420         memcpy(pbbak, i_buffer, i_len);
421         s->c= ic;
422         s->c.bytestream_start= pbbak_start;
423         s->c.bytestream= pbbak + i_len;
424         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
425         memcpy(s->block_state, i_state, sizeof(s->block_state));
426         return iscore;
427     }else{
428         memcpy(pbbak, p_buffer, p_len);
429         s->c= pc;
430         s->c.bytestream_start= pbbak_start;
431         s->c.bytestream= pbbak + p_len;
432         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
433         memcpy(s->block_state, p_state, sizeof(s->block_state));
434         return score;
435     }
436 }
437
438 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
439     const int w= s->b_width  << s->block_max_depth;
440     const int rem_depth= s->block_max_depth - level;
441     const int index= (x + y*w) << rem_depth;
442     int trx= (x+1)<<rem_depth;
443     BlockNode *b= &s->block[index];
444     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
445     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
446     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
447     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
448     int pl = left->color[0];
449     int pcb= left->color[1];
450     int pcr= left->color[2];
451     int pmx, pmy;
452     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
453     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
454     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
455     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
456
457     if(s->keyframe){
458         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
459         return;
460     }
461
462     if(level!=s->block_max_depth){
463         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
464             put_rac(&s->c, &s->block_state[4 + s_context], 1);
465         }else{
466             put_rac(&s->c, &s->block_state[4 + s_context], 0);
467             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
468             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
469             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
470             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
471             return;
472         }
473     }
474     if(b->type & BLOCK_INTRA){
475         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
476         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
477         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
478         if (s->nb_planes > 2) {
479             put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
480             put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
481         }
482         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
483     }else{
484         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
485         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
486         if(s->ref_frames > 1)
487             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
488         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
489         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
490         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
491     }
492 }
493
494 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
495     int i, x2, y2;
496     Plane *p= &s->plane[plane_index];
497     const int block_size = MB_SIZE >> s->block_max_depth;
498     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
499     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
500     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
501     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
502     const int ref_stride= s->current_picture->linesize[plane_index];
503     uint8_t *src= s-> input_picture->data[plane_index];
504     IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
505     const int b_stride = s->b_width << s->block_max_depth;
506     const int w= p->width;
507     const int h= p->height;
508     int index= mb_x + mb_y*b_stride;
509     BlockNode *b= &s->block[index];
510     BlockNode backup= *b;
511     int ab=0;
512     int aa=0;
513
514     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
515
516     b->type|= BLOCK_INTRA;
517     b->color[plane_index]= 0;
518     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
519
520     for(i=0; i<4; i++){
521         int mb_x2= mb_x + (i &1) - 1;
522         int mb_y2= mb_y + (i>>1) - 1;
523         int x= block_w*mb_x2 + block_w/2;
524         int y= block_h*mb_y2 + block_h/2;
525
526         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
527                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
528
529         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
530             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
531                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
532                 int obmc_v= obmc[index];
533                 int d;
534                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
535                 if(x<0) obmc_v += obmc[index + block_w];
536                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
537                 if(x+block_w>w) obmc_v += obmc[index - block_w];
538                 //FIXME precalculate this or simplify it somehow else
539
540                 d = -dst[index] + (1<<(FRAC_BITS-1));
541                 dst[index] = d;
542                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
543                 aa += obmc_v * obmc_v; //FIXME precalculate this
544             }
545         }
546     }
547     *b= backup;
548
549     return av_clip( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa), 0, 255); //FIXME we should not need clipping
550 }
551
552 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
553     const int b_stride = s->b_width << s->block_max_depth;
554     const int b_height = s->b_height<< s->block_max_depth;
555     int index= x + y*b_stride;
556     const BlockNode *b     = &s->block[index];
557     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
558     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
559     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
560     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
561     int dmx, dmy;
562 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
563 //  int my_context= av_log2(2*FFABS(left->my - top->my));
564
565     if(x<0 || x>=b_stride || y>=b_height)
566         return 0;
567 /*
568 1            0      0
569 01X          1-2    1
570 001XX        3-6    2-3
571 0001XXX      7-14   4-7
572 00001XXXX   15-30   8-15
573 */
574 //FIXME try accurate rate
575 //FIXME intra and inter predictors if surrounding blocks are not the same type
576     if(b->type & BLOCK_INTRA){
577         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
578                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
579                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
580     }else{
581         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
582         dmx-= b->mx;
583         dmy-= b->my;
584         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
585                     + av_log2(2*FFABS(dmy))
586                     + av_log2(2*b->ref));
587     }
588 }
589
590 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
591     Plane *p= &s->plane[plane_index];
592     const int block_size = MB_SIZE >> s->block_max_depth;
593     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
594     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
595     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
596     const int ref_stride= s->current_picture->linesize[plane_index];
597     uint8_t *dst= s->current_picture->data[plane_index];
598     uint8_t *src= s->  input_picture->data[plane_index];
599     IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
600     uint8_t *cur = s->scratchbuf;
601     uint8_t *tmp = s->emu_edge_buffer;
602     const int b_stride = s->b_width << s->block_max_depth;
603     const int b_height = s->b_height<< s->block_max_depth;
604     const int w= p->width;
605     const int h= p->height;
606     int distortion;
607     int rate= 0;
608     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
609     int sx= block_w*mb_x - block_w/2;
610     int sy= block_h*mb_y - block_h/2;
611     int x0= FFMAX(0,-sx);
612     int y0= FFMAX(0,-sy);
613     int x1= FFMIN(block_w*2, w-sx);
614     int y1= FFMIN(block_h*2, h-sy);
615     int i,x,y;
616
617     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
618
619     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);
620
621     for(y=y0; y<y1; y++){
622         const uint8_t *obmc1= obmc_edged[y];
623         const IDWTELEM *pred1 = pred + y*obmc_stride;
624         uint8_t *cur1 = cur + y*ref_stride;
625         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
626         for(x=x0; x<x1; x++){
627 #if FRAC_BITS >= LOG2_OBMC_MAX
628             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
629 #else
630             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
631 #endif
632             v = (v + pred1[x]) >> FRAC_BITS;
633             if(v&(~255)) v= ~(v>>31);
634             dst1[x] = v;
635         }
636     }
637
638     /* copy the regions where obmc[] = (uint8_t)256 */
639     if(LOG2_OBMC_MAX == 8
640         && (mb_x == 0 || mb_x == b_stride-1)
641         && (mb_y == 0 || mb_y == b_height-1)){
642         if(mb_x == 0)
643             x1 = block_w;
644         else
645             x0 = block_w;
646         if(mb_y == 0)
647             y1 = block_h;
648         else
649             y0 = block_h;
650         for(y=y0; y<y1; y++)
651             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
652     }
653
654     if(block_w==16){
655         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
656         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
657         /* FIXME cmps overlap but do not cover the wavelet's whole support.
658          * So improving the score of one block is not strictly guaranteed
659          * to improve the score of the whole frame, thus iterative motion
660          * estimation does not always converge. */
661         if(s->avctx->me_cmp == FF_CMP_W97)
662             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
663         else if(s->avctx->me_cmp == FF_CMP_W53)
664             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
665         else{
666             distortion = 0;
667             for(i=0; i<4; i++){
668                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
669                 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
670             }
671         }
672     }else{
673         av_assert2(block_w==8);
674         distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
675     }
676
677     if(plane_index==0){
678         for(i=0; i<4; i++){
679 /* ..RRr
680  * .RXx.
681  * rxx..
682  */
683             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
684         }
685         if(mb_x == b_stride-2)
686             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
687     }
688     return distortion + rate*penalty_factor;
689 }
690
691 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
692     int i, y2;
693     Plane *p= &s->plane[plane_index];
694     const int block_size = MB_SIZE >> s->block_max_depth;
695     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
696     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
697     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
698     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
699     const int ref_stride= s->current_picture->linesize[plane_index];
700     uint8_t *dst= s->current_picture->data[plane_index];
701     uint8_t *src= s-> input_picture->data[plane_index];
702     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
703     // const has only been removed from zero_dst to suppress a warning
704     static IDWTELEM zero_dst[4096]; //FIXME
705     const int b_stride = s->b_width << s->block_max_depth;
706     const int w= p->width;
707     const int h= p->height;
708     int distortion= 0;
709     int rate= 0;
710     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
711
712     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
713
714     for(i=0; i<9; i++){
715         int mb_x2= mb_x + (i%3) - 1;
716         int mb_y2= mb_y + (i/3) - 1;
717         int x= block_w*mb_x2 + block_w/2;
718         int y= block_h*mb_y2 + block_h/2;
719
720         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
721                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
722
723         //FIXME find a cleaner/simpler way to skip the outside stuff
724         for(y2= y; y2<0; y2++)
725             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
726         for(y2= h; y2<y+block_h; y2++)
727             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
728         if(x<0){
729             for(y2= y; y2<y+block_h; y2++)
730                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
731         }
732         if(x+block_w > w){
733             for(y2= y; y2<y+block_h; y2++)
734                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
735         }
736
737         av_assert1(block_w== 8 || block_w==16);
738         distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
739     }
740
741     if(plane_index==0){
742         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
743         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
744
745 /* ..RRRr
746  * .RXXx.
747  * .RXXx.
748  * rxxx.
749  */
750         if(merged)
751             rate = get_block_bits(s, mb_x, mb_y, 2);
752         for(i=merged?4:0; i<9; i++){
753             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
754             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
755         }
756     }
757     return distortion + rate*penalty_factor;
758 }
759
760 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
761     const int w= b->width;
762     const int h= b->height;
763     int x, y;
764
765     if(1){
766         int run=0;
767         int *runs = s->run_buffer;
768         int run_index=0;
769         int max_index;
770
771         for(y=0; y<h; y++){
772             for(x=0; x<w; x++){
773                 int v, p=0;
774                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
775                 v= src[x + y*stride];
776
777                 if(y){
778                     t= src[x + (y-1)*stride];
779                     if(x){
780                         lt= src[x - 1 + (y-1)*stride];
781                     }
782                     if(x + 1 < w){
783                         rt= src[x + 1 + (y-1)*stride];
784                     }
785                 }
786                 if(x){
787                     l= src[x - 1 + y*stride];
788                     /*if(x > 1){
789                         if(orientation==1) ll= src[y + (x-2)*stride];
790                         else               ll= src[x - 2 + y*stride];
791                     }*/
792                 }
793                 if(parent){
794                     int px= x>>1;
795                     int py= y>>1;
796                     if(px<b->parent->width && py<b->parent->height)
797                         p= parent[px + py*2*stride];
798                 }
799                 if(!(/*ll|*/l|lt|t|rt|p)){
800                     if(v){
801                         runs[run_index++]= run;
802                         run=0;
803                     }else{
804                         run++;
805                     }
806                 }
807             }
808         }
809         max_index= run_index;
810         runs[run_index++]= run;
811         run_index=0;
812         run= runs[run_index++];
813
814         put_symbol2(&s->c, b->state[30], max_index, 0);
815         if(run_index <= max_index)
816             put_symbol2(&s->c, b->state[1], run, 3);
817
818         for(y=0; y<h; y++){
819             if(s->c.bytestream_end - s->c.bytestream < w*40){
820                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
821                 return -1;
822             }
823             for(x=0; x<w; x++){
824                 int v, p=0;
825                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
826                 v= src[x + y*stride];
827
828                 if(y){
829                     t= src[x + (y-1)*stride];
830                     if(x){
831                         lt= src[x - 1 + (y-1)*stride];
832                     }
833                     if(x + 1 < w){
834                         rt= src[x + 1 + (y-1)*stride];
835                     }
836                 }
837                 if(x){
838                     l= src[x - 1 + y*stride];
839                     /*if(x > 1){
840                         if(orientation==1) ll= src[y + (x-2)*stride];
841                         else               ll= src[x - 2 + y*stride];
842                     }*/
843                 }
844                 if(parent){
845                     int px= x>>1;
846                     int py= y>>1;
847                     if(px<b->parent->width && py<b->parent->height)
848                         p= parent[px + py*2*stride];
849                 }
850                 if(/*ll|*/l|lt|t|rt|p){
851                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
852
853                     put_rac(&s->c, &b->state[0][context], !!v);
854                 }else{
855                     if(!run){
856                         run= runs[run_index++];
857
858                         if(run_index <= max_index)
859                             put_symbol2(&s->c, b->state[1], run, 3);
860                         av_assert2(v);
861                     }else{
862                         run--;
863                         av_assert2(!v);
864                     }
865                 }
866                 if(v){
867                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
868                     int l2= 2*FFABS(l) + (l<0);
869                     int t2= 2*FFABS(t) + (t<0);
870
871                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
872                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
873                 }
874             }
875         }
876     }
877     return 0;
878 }
879
880 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
881 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
882 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
883     return encode_subband_c0run(s, b, src, parent, stride, orientation);
884 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
885 }
886
887 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){
888     const int b_stride= s->b_width << s->block_max_depth;
889     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
890     BlockNode backup= *block;
891     unsigned value;
892     int rd, index;
893
894     av_assert2(mb_x>=0 && mb_y>=0);
895     av_assert2(mb_x<b_stride);
896
897     if(intra){
898         block->color[0] = p[0];
899         block->color[1] = p[1];
900         block->color[2] = p[2];
901         block->type |= BLOCK_INTRA;
902     }else{
903         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
904         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
905         if(s->me_cache[index] == value)
906             return 0;
907         s->me_cache[index]= value;
908
909         block->mx= p[0];
910         block->my= p[1];
911         block->type &= ~BLOCK_INTRA;
912     }
913
914     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
915
916 //FIXME chroma
917     if(rd < *best_rd){
918         *best_rd= rd;
919         return 1;
920     }else{
921         *block= backup;
922         return 0;
923     }
924 }
925
926 /* special case for int[2] args we discard afterwards,
927  * fixes compilation problem with gcc 2.95 */
928 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){
929     int p[2] = {p0, p1};
930     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
931 }
932
933 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){
934     const int b_stride= s->b_width << s->block_max_depth;
935     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
936     BlockNode backup[4];
937     unsigned value;
938     int rd, index;
939
940     /* We don't initialize backup[] during variable declaration, because
941      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
942      * 'int16_t'". */
943     backup[0] = block[0];
944     backup[1] = block[1];
945     backup[2] = block[b_stride];
946     backup[3] = block[b_stride + 1];
947
948     av_assert2(mb_x>=0 && mb_y>=0);
949     av_assert2(mb_x<b_stride);
950     av_assert2(((mb_x|mb_y)&1) == 0);
951
952     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
953     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
954     if(s->me_cache[index] == value)
955         return 0;
956     s->me_cache[index]= value;
957
958     block->mx= p0;
959     block->my= p1;
960     block->ref= ref;
961     block->type &= ~BLOCK_INTRA;
962     block[1]= block[b_stride]= block[b_stride+1]= *block;
963
964     rd= get_4block_rd(s, mb_x, mb_y, 0);
965
966 //FIXME chroma
967     if(rd < *best_rd){
968         *best_rd= rd;
969         return 1;
970     }else{
971         block[0]= backup[0];
972         block[1]= backup[1];
973         block[b_stride]= backup[2];
974         block[b_stride+1]= backup[3];
975         return 0;
976     }
977 }
978
979 static void iterative_me(SnowContext *s){
980     int pass, mb_x, mb_y;
981     const int b_width = s->b_width  << s->block_max_depth;
982     const int b_height= s->b_height << s->block_max_depth;
983     const int b_stride= b_width;
984     int color[3];
985
986     {
987         RangeCoder r = s->c;
988         uint8_t state[sizeof(s->block_state)];
989         memcpy(state, s->block_state, sizeof(s->block_state));
990         for(mb_y= 0; mb_y<s->b_height; mb_y++)
991             for(mb_x= 0; mb_x<s->b_width; mb_x++)
992                 encode_q_branch(s, 0, mb_x, mb_y);
993         s->c = r;
994         memcpy(s->block_state, state, sizeof(s->block_state));
995     }
996
997     for(pass=0; pass<25; pass++){
998         int change= 0;
999
1000         for(mb_y= 0; mb_y<b_height; mb_y++){
1001             for(mb_x= 0; mb_x<b_width; mb_x++){
1002                 int dia_change, i, j, ref;
1003                 int best_rd= INT_MAX, ref_rd;
1004                 BlockNode backup, ref_b;
1005                 const int index= mb_x + mb_y * b_stride;
1006                 BlockNode *block= &s->block[index];
1007                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1008                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1009                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1010                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1011                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1012                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1013                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1014                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1015                 const int b_w= (MB_SIZE >> s->block_max_depth);
1016                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1017
1018                 if(pass && (block->type & BLOCK_OPT))
1019                     continue;
1020                 block->type |= BLOCK_OPT;
1021
1022                 backup= *block;
1023
1024                 if(!s->me_cache_generation)
1025                     memset(s->me_cache, 0, sizeof(s->me_cache));
1026                 s->me_cache_generation += 1<<22;
1027
1028                 //FIXME precalculate
1029                 {
1030                     int x, y;
1031                     for (y = 0; y < b_w * 2; y++)
1032                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1033                     if(mb_x==0)
1034                         for(y=0; y<b_w*2; y++)
1035                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1036                     if(mb_x==b_stride-1)
1037                         for(y=0; y<b_w*2; y++)
1038                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1039                     if(mb_y==0){
1040                         for(x=0; x<b_w*2; x++)
1041                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1042                         for(y=1; y<b_w; y++)
1043                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1044                     }
1045                     if(mb_y==b_height-1){
1046                         for(x=0; x<b_w*2; x++)
1047                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1048                         for(y=b_w; y<b_w*2-1; y++)
1049                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1050                     }
1051                 }
1052
1053                 //skip stuff outside the picture
1054                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1055                     uint8_t *src= s->  input_picture->data[0];
1056                     uint8_t *dst= s->current_picture->data[0];
1057                     const int stride= s->current_picture->linesize[0];
1058                     const int block_w= MB_SIZE >> s->block_max_depth;
1059                     const int block_h= MB_SIZE >> s->block_max_depth;
1060                     const int sx= block_w*mb_x - block_w/2;
1061                     const int sy= block_h*mb_y - block_h/2;
1062                     const int w= s->plane[0].width;
1063                     const int h= s->plane[0].height;
1064                     int y;
1065
1066                     for(y=sy; y<0; y++)
1067                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1068                     for(y=h; y<sy+block_h*2; y++)
1069                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1070                     if(sx<0){
1071                         for(y=sy; y<sy+block_h*2; y++)
1072                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1073                     }
1074                     if(sx+block_w*2 > w){
1075                         for(y=sy; y<sy+block_h*2; y++)
1076                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1077                     }
1078                 }
1079
1080                 // intra(black) = neighbors' contribution to the current block
1081                 for(i=0; i < s->nb_planes; i++)
1082                     color[i]= get_dc(s, mb_x, mb_y, i);
1083
1084                 // get previous score (cannot be cached due to OBMC)
1085                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1086                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1087                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1088                 }else
1089                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1090
1091                 ref_b= *block;
1092                 ref_rd= best_rd;
1093                 for(ref=0; ref < s->ref_frames; ref++){
1094                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1095                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1096                         continue;
1097                     block->ref= ref;
1098                     best_rd= INT_MAX;
1099
1100                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1101                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1102                     if(tb)
1103                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1104                     if(lb)
1105                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1106                     if(rb)
1107                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1108                     if(bb)
1109                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1110
1111                     /* fullpel ME */
1112                     //FIXME avoid subpel interpolation / round to nearest integer
1113                     do{
1114                         dia_change=0;
1115                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1116                             for(j=0; j<i; j++){
1117                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1118                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1119                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1120                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1121                             }
1122                         }
1123                     }while(dia_change);
1124                     /* subpel ME */
1125                     do{
1126                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1127                         dia_change=0;
1128                         for(i=0; i<8; i++)
1129                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1130                     }while(dia_change);
1131                     //FIXME or try the standard 2 pass qpel or similar
1132
1133                     mvr[0][0]= block->mx;
1134                     mvr[0][1]= block->my;
1135                     if(ref_rd > best_rd){
1136                         ref_rd= best_rd;
1137                         ref_b= *block;
1138                     }
1139                 }
1140                 best_rd= ref_rd;
1141                 *block= ref_b;
1142                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1143                 //FIXME RD style color selection
1144                 if(!same_block(block, &backup)){
1145                     if(tb ) tb ->type &= ~BLOCK_OPT;
1146                     if(lb ) lb ->type &= ~BLOCK_OPT;
1147                     if(rb ) rb ->type &= ~BLOCK_OPT;
1148                     if(bb ) bb ->type &= ~BLOCK_OPT;
1149                     if(tlb) tlb->type &= ~BLOCK_OPT;
1150                     if(trb) trb->type &= ~BLOCK_OPT;
1151                     if(blb) blb->type &= ~BLOCK_OPT;
1152                     if(brb) brb->type &= ~BLOCK_OPT;
1153                     change ++;
1154                 }
1155             }
1156         }
1157         av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1158         if(!change)
1159             break;
1160     }
1161
1162     if(s->block_max_depth == 1){
1163         int change= 0;
1164         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1165             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1166                 int i;
1167                 int best_rd, init_rd;
1168                 const int index= mb_x + mb_y * b_stride;
1169                 BlockNode *b[4];
1170
1171                 b[0]= &s->block[index];
1172                 b[1]= b[0]+1;
1173                 b[2]= b[0]+b_stride;
1174                 b[3]= b[2]+1;
1175                 if(same_block(b[0], b[1]) &&
1176                    same_block(b[0], b[2]) &&
1177                    same_block(b[0], b[3]))
1178                     continue;
1179
1180                 if(!s->me_cache_generation)
1181                     memset(s->me_cache, 0, sizeof(s->me_cache));
1182                 s->me_cache_generation += 1<<22;
1183
1184                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1185
1186                 //FIXME more multiref search?
1187                 check_4block_inter(s, mb_x, mb_y,
1188                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1189                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1190
1191                 for(i=0; i<4; i++)
1192                     if(!(b[i]->type&BLOCK_INTRA))
1193                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1194
1195                 if(init_rd != best_rd)
1196                     change++;
1197             }
1198         }
1199         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1200     }
1201 }
1202
1203 static void encode_blocks(SnowContext *s, int search){
1204     int x, y;
1205     int w= s->b_width;
1206     int h= s->b_height;
1207
1208     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1209         iterative_me(s);
1210
1211     for(y=0; y<h; y++){
1212         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1213             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1214             return;
1215         }
1216         for(x=0; x<w; x++){
1217             if(s->avctx->me_method == ME_ITER || !search)
1218                 encode_q_branch2(s, 0, x, y);
1219             else
1220                 encode_q_branch (s, 0, x, y);
1221         }
1222     }
1223 }
1224
1225 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1226     const int w= b->width;
1227     const int h= b->height;
1228     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1229     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1230     int x,y, thres1, thres2;
1231
1232     if(s->qlog == LOSSLESS_QLOG){
1233         for(y=0; y<h; y++)
1234             for(x=0; x<w; x++)
1235                 dst[x + y*stride]= src[x + y*stride];
1236         return;
1237     }
1238
1239     bias= bias ? 0 : (3*qmul)>>3;
1240     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1241     thres2= 2*thres1;
1242
1243     if(!bias){
1244         for(y=0; y<h; y++){
1245             for(x=0; x<w; x++){
1246                 int i= src[x + y*stride];
1247
1248                 if((unsigned)(i+thres1) > thres2){
1249                     if(i>=0){
1250                         i<<= QEXPSHIFT;
1251                         i/= qmul; //FIXME optimize
1252                         dst[x + y*stride]=  i;
1253                     }else{
1254                         i= -i;
1255                         i<<= QEXPSHIFT;
1256                         i/= qmul; //FIXME optimize
1257                         dst[x + y*stride]= -i;
1258                     }
1259                 }else
1260                     dst[x + y*stride]= 0;
1261             }
1262         }
1263     }else{
1264         for(y=0; y<h; y++){
1265             for(x=0; x<w; x++){
1266                 int i= src[x + y*stride];
1267
1268                 if((unsigned)(i+thres1) > thres2){
1269                     if(i>=0){
1270                         i<<= QEXPSHIFT;
1271                         i= (i + bias) / qmul; //FIXME optimize
1272                         dst[x + y*stride]=  i;
1273                     }else{
1274                         i= -i;
1275                         i<<= QEXPSHIFT;
1276                         i= (i + bias) / qmul; //FIXME optimize
1277                         dst[x + y*stride]= -i;
1278                     }
1279                 }else
1280                     dst[x + y*stride]= 0;
1281             }
1282         }
1283     }
1284 }
1285
1286 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1287     const int w= b->width;
1288     const int h= b->height;
1289     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1290     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1291     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1292     int x,y;
1293
1294     if(s->qlog == LOSSLESS_QLOG) return;
1295
1296     for(y=0; y<h; y++){
1297         for(x=0; x<w; x++){
1298             int i= src[x + y*stride];
1299             if(i<0){
1300                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1301             }else if(i>0){
1302                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1303             }
1304         }
1305     }
1306 }
1307
1308 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1309     const int w= b->width;
1310     const int h= b->height;
1311     int x,y;
1312
1313     for(y=h-1; y>=0; y--){
1314         for(x=w-1; x>=0; x--){
1315             int i= x + y*stride;
1316
1317             if(x){
1318                 if(use_median){
1319                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1320                     else  src[i] -= src[i - 1];
1321                 }else{
1322                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1323                     else  src[i] -= src[i - 1];
1324                 }
1325             }else{
1326                 if(y) src[i] -= src[i - stride];
1327             }
1328         }
1329     }
1330 }
1331
1332 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1333     const int w= b->width;
1334     const int h= b->height;
1335     int x,y;
1336
1337     for(y=0; y<h; y++){
1338         for(x=0; x<w; x++){
1339             int i= x + y*stride;
1340
1341             if(x){
1342                 if(use_median){
1343                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1344                     else  src[i] += src[i - 1];
1345                 }else{
1346                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1347                     else  src[i] += src[i - 1];
1348                 }
1349             }else{
1350                 if(y) src[i] += src[i - stride];
1351             }
1352         }
1353     }
1354 }
1355
1356 static void encode_qlogs(SnowContext *s){
1357     int plane_index, level, orientation;
1358
1359     for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1360         for(level=0; level<s->spatial_decomposition_count; level++){
1361             for(orientation=level ? 1:0; orientation<4; orientation++){
1362                 if(orientation==2) continue;
1363                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1364             }
1365         }
1366     }
1367 }
1368
1369 static void encode_header(SnowContext *s){
1370     int plane_index, i;
1371     uint8_t kstate[32];
1372
1373     memset(kstate, MID_STATE, sizeof(kstate));
1374
1375     put_rac(&s->c, kstate, s->keyframe);
1376     if(s->keyframe || s->always_reset){
1377         ff_snow_reset_contexts(s);
1378         s->last_spatial_decomposition_type=
1379         s->last_qlog=
1380         s->last_qbias=
1381         s->last_mv_scale=
1382         s->last_block_max_depth= 0;
1383         for(plane_index=0; plane_index<2; plane_index++){
1384             Plane *p= &s->plane[plane_index];
1385             p->last_htaps=0;
1386             p->last_diag_mc=0;
1387             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1388         }
1389     }
1390     if(s->keyframe){
1391         put_symbol(&s->c, s->header_state, s->version, 0);
1392         put_rac(&s->c, s->header_state, s->always_reset);
1393         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1394         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1395         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1396         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1397         if (s->nb_planes > 2) {
1398             put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1399             put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1400         }
1401         put_rac(&s->c, s->header_state, s->spatial_scalability);
1402 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1403         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1404
1405         encode_qlogs(s);
1406     }
1407
1408     if(!s->keyframe){
1409         int update_mc=0;
1410         for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1411             Plane *p= &s->plane[plane_index];
1412             update_mc |= p->last_htaps   != p->htaps;
1413             update_mc |= p->last_diag_mc != p->diag_mc;
1414             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1415         }
1416         put_rac(&s->c, s->header_state, update_mc);
1417         if(update_mc){
1418             for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1419                 Plane *p= &s->plane[plane_index];
1420                 put_rac(&s->c, s->header_state, p->diag_mc);
1421                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1422                 for(i= p->htaps/2; i; i--)
1423                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1424             }
1425         }
1426         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1427             put_rac(&s->c, s->header_state, 1);
1428             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1429             encode_qlogs(s);
1430         }else
1431             put_rac(&s->c, s->header_state, 0);
1432     }
1433
1434     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1435     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1436     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1437     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1438     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1439
1440 }
1441
1442 static void update_last_header_values(SnowContext *s){
1443     int plane_index;
1444
1445     if(!s->keyframe){
1446         for(plane_index=0; plane_index<2; plane_index++){
1447             Plane *p= &s->plane[plane_index];
1448             p->last_diag_mc= p->diag_mc;
1449             p->last_htaps  = p->htaps;
1450             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1451         }
1452     }
1453
1454     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1455     s->last_qlog                        = s->qlog;
1456     s->last_qbias                       = s->qbias;
1457     s->last_mv_scale                    = s->mv_scale;
1458     s->last_block_max_depth             = s->block_max_depth;
1459     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1460 }
1461
1462 static int qscale2qlog(int qscale){
1463     return rint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1464            + 61*QROOT/8; ///< 64 > 60
1465 }
1466
1467 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1468 {
1469     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1470      * FIXME we know exact mv bits at this point,
1471      * but ratecontrol isn't set up to include them. */
1472     uint32_t coef_sum= 0;
1473     int level, orientation, delta_qlog;
1474
1475     for(level=0; level<s->spatial_decomposition_count; level++){
1476         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1477             SubBand *b= &s->plane[0].band[level][orientation];
1478             IDWTELEM *buf= b->ibuf;
1479             const int w= b->width;
1480             const int h= b->height;
1481             const int stride= b->stride;
1482             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1483             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1484             const int qdiv= (1<<16)/qmul;
1485             int x, y;
1486             //FIXME this is ugly
1487             for(y=0; y<h; y++)
1488                 for(x=0; x<w; x++)
1489                     buf[x+y*stride]= b->buf[x+y*stride];
1490             if(orientation==0)
1491                 decorrelate(s, b, buf, stride, 1, 0);
1492             for(y=0; y<h; y++)
1493                 for(x=0; x<w; x++)
1494                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1495         }
1496     }
1497
1498     /* ugly, ratecontrol just takes a sqrt again */
1499     av_assert0(coef_sum < INT_MAX);
1500     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1501
1502     if(pict->pict_type == AV_PICTURE_TYPE_I){
1503         s->m.current_picture.mb_var_sum= coef_sum;
1504         s->m.current_picture.mc_mb_var_sum= 0;
1505     }else{
1506         s->m.current_picture.mc_mb_var_sum= coef_sum;
1507         s->m.current_picture.mb_var_sum= 0;
1508     }
1509
1510     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1511     if (pict->quality < 0)
1512         return INT_MIN;
1513     s->lambda= pict->quality * 3/2;
1514     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1515     s->qlog+= delta_qlog;
1516     return delta_qlog;
1517 }
1518
1519 static void calculate_visual_weight(SnowContext *s, Plane *p){
1520     int width = p->width;
1521     int height= p->height;
1522     int level, orientation, x, y;
1523
1524     for(level=0; level<s->spatial_decomposition_count; level++){
1525         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1526             SubBand *b= &p->band[level][orientation];
1527             IDWTELEM *ibuf= b->ibuf;
1528             int64_t error=0;
1529
1530             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1531             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1532             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1533             for(y=0; y<height; y++){
1534                 for(x=0; x<width; x++){
1535                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1536                     error += d*d;
1537                 }
1538             }
1539
1540             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1541         }
1542     }
1543 }
1544
1545 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1546                         const AVFrame *pict, int *got_packet)
1547 {
1548     SnowContext *s = avctx->priv_data;
1549     RangeCoder * const c= &s->c;
1550     AVFrame *pic = pict;
1551     const int width= s->avctx->width;
1552     const int height= s->avctx->height;
1553     int level, orientation, plane_index, i, y, ret;
1554     uint8_t rc_header_bak[sizeof(s->header_state)];
1555     uint8_t rc_block_bak[sizeof(s->block_state)];
1556
1557     if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0)
1558         return ret;
1559
1560     ff_init_range_encoder(c, pkt->data, pkt->size);
1561     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1562
1563     for(i=0; i < s->nb_planes; i++){
1564         int hshift= i ? s->chroma_h_shift : 0;
1565         int vshift= i ? s->chroma_v_shift : 0;
1566         for(y=0; y<(height>>vshift); y++)
1567             memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1568                    &pict->data[i][y * pict->linesize[i]],
1569                    width>>hshift);
1570         s->dsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1571                             width >> hshift, height >> vshift,
1572                             EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1573                             EDGE_TOP | EDGE_BOTTOM);
1574
1575     }
1576     emms_c();
1577     s->new_picture = pict;
1578
1579     s->m.picture_number= avctx->frame_number;
1580     if(avctx->flags&CODEC_FLAG_PASS2){
1581         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1582         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1583         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1584             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1585             if (pic->quality < 0)
1586                 return -1;
1587         }
1588     }else{
1589         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1590         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1591     }
1592
1593     if(s->pass1_rc && avctx->frame_number == 0)
1594         pic->quality = 2*FF_QP2LAMBDA;
1595     if (pic->quality) {
1596         s->qlog   = qscale2qlog(pic->quality);
1597         s->lambda = pic->quality * 3/2;
1598     }
1599     if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1600         s->qlog= LOSSLESS_QLOG;
1601         s->lambda = 0;
1602     }//else keep previous frame's qlog until after motion estimation
1603
1604     ff_snow_frame_start(s);
1605     avctx->coded_frame= s->current_picture;
1606
1607     s->m.current_picture_ptr= &s->m.current_picture;
1608     s->m.current_picture.f = s->current_picture;
1609     s->m.current_picture.f->pts = pict->pts;
1610     if(pic->pict_type == AV_PICTURE_TYPE_P){
1611         int block_width = (width +15)>>4;
1612         int block_height= (height+15)>>4;
1613         int stride= s->current_picture->linesize[0];
1614
1615         av_assert0(s->current_picture->data[0]);
1616         av_assert0(s->last_picture[0]->data[0]);
1617
1618         s->m.avctx= s->avctx;
1619         s->m.   last_picture.f = s->last_picture[0];
1620         s->m.    new_picture.f = s->input_picture;
1621         s->m.   last_picture_ptr= &s->m.   last_picture;
1622         s->m.linesize = stride;
1623         s->m.uvlinesize= s->current_picture->linesize[1];
1624         s->m.width = width;
1625         s->m.height= height;
1626         s->m.mb_width = block_width;
1627         s->m.mb_height= block_height;
1628         s->m.mb_stride=   s->m.mb_width+1;
1629         s->m.b8_stride= 2*s->m.mb_width+1;
1630         s->m.f_code=1;
1631         s->m.pict_type = pic->pict_type;
1632         s->m.me_method= s->avctx->me_method;
1633         s->m.me.scene_change_score=0;
1634         s->m.flags= s->avctx->flags;
1635         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1636         s->m.out_format= FMT_H263;
1637         s->m.unrestricted_mv= 1;
1638
1639         s->m.lambda = s->lambda;
1640         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1641         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1642
1643         s->m.dsp= s->dsp; //move
1644         s->m.qdsp= s->qdsp; //move
1645         s->m.hdsp = s->hdsp;
1646         ff_init_me(&s->m);
1647         s->hdsp = s->m.hdsp;
1648         s->dsp= s->m.dsp;
1649     }
1650
1651     if(s->pass1_rc){
1652         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1653         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1654     }
1655
1656 redo_frame:
1657
1658     s->spatial_decomposition_count= 5;
1659
1660     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1661           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1662         s->spatial_decomposition_count--;
1663
1664     if (s->spatial_decomposition_count <= 0) {
1665         av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1666         return AVERROR(EINVAL);
1667     }
1668
1669     s->m.pict_type = pic->pict_type;
1670     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1671
1672     ff_snow_common_init_after_header(avctx);
1673
1674     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1675         for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1676             calculate_visual_weight(s, &s->plane[plane_index]);
1677         }
1678     }
1679
1680     encode_header(s);
1681     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1682     encode_blocks(s, 1);
1683     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1684
1685     for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1686         Plane *p= &s->plane[plane_index];
1687         int w= p->width;
1688         int h= p->height;
1689         int x, y;
1690 //        int bits= put_bits_count(&s->c.pb);
1691
1692         if (!s->memc_only) {
1693             //FIXME optimize
1694             if(pict->data[plane_index]) //FIXME gray hack
1695                 for(y=0; y<h; y++){
1696                     for(x=0; x<w; x++){
1697                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1698                     }
1699                 }
1700             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1701
1702             if(   plane_index==0
1703                && pic->pict_type == AV_PICTURE_TYPE_P
1704                && !(avctx->flags&CODEC_FLAG_PASS2)
1705                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1706                 ff_init_range_encoder(c, pkt->data, pkt->size);
1707                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1708                 pic->pict_type= AV_PICTURE_TYPE_I;
1709                 s->keyframe=1;
1710                 s->current_picture->key_frame=1;
1711                 goto redo_frame;
1712             }
1713
1714             if(s->qlog == LOSSLESS_QLOG){
1715                 for(y=0; y<h; y++){
1716                     for(x=0; x<w; x++){
1717                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1718                     }
1719                 }
1720             }else{
1721                 for(y=0; y<h; y++){
1722                     for(x=0; x<w; x++){
1723                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1724                     }
1725                 }
1726             }
1727
1728             ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1729
1730             if(s->pass1_rc && plane_index==0){
1731                 int delta_qlog = ratecontrol_1pass(s, pic);
1732                 if (delta_qlog <= INT_MIN)
1733                     return -1;
1734                 if(delta_qlog){
1735                     //reordering qlog in the bitstream would eliminate this reset
1736                     ff_init_range_encoder(c, pkt->data, pkt->size);
1737                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1738                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1739                     encode_header(s);
1740                     encode_blocks(s, 0);
1741                 }
1742             }
1743
1744             for(level=0; level<s->spatial_decomposition_count; level++){
1745                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1746                     SubBand *b= &p->band[level][orientation];
1747
1748                     quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1749                     if(orientation==0)
1750                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1751                     if (!s->no_bitstream)
1752                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1753                     av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1754                     if(orientation==0)
1755                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1756                 }
1757             }
1758
1759             for(level=0; level<s->spatial_decomposition_count; level++){
1760                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1761                     SubBand *b= &p->band[level][orientation];
1762
1763                     dequantize(s, b, b->ibuf, b->stride);
1764                 }
1765             }
1766
1767             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1768             if(s->qlog == LOSSLESS_QLOG){
1769                 for(y=0; y<h; y++){
1770                     for(x=0; x<w; x++){
1771                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1772                     }
1773                 }
1774             }
1775             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1776         }else{
1777             //ME/MC only
1778             if(pic->pict_type == AV_PICTURE_TYPE_I){
1779                 for(y=0; y<h; y++){
1780                     for(x=0; x<w; x++){
1781                         s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1782                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1783                     }
1784                 }
1785             }else{
1786                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1787                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1788             }
1789         }
1790         if(s->avctx->flags&CODEC_FLAG_PSNR){
1791             int64_t error= 0;
1792
1793             if(pict->data[plane_index]) //FIXME gray hack
1794                 for(y=0; y<h; y++){
1795                     for(x=0; x<w; x++){
1796                         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];
1797                         error += d*d;
1798                     }
1799                 }
1800             s->avctx->error[plane_index] += error;
1801             s->current_picture->error[plane_index] = error;
1802         }
1803
1804     }
1805
1806     update_last_header_values(s);
1807
1808     ff_snow_release_buffer(avctx);
1809
1810     s->current_picture->coded_picture_number = avctx->frame_number;
1811     s->current_picture->pict_type = pict->pict_type;
1812     s->current_picture->quality = pict->quality;
1813     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1814     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1815     s->m.current_picture.f->display_picture_number =
1816     s->m.current_picture.f->coded_picture_number   = avctx->frame_number;
1817     s->m.current_picture.f->quality                = pic->quality;
1818     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1819     if(s->pass1_rc)
1820         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1821             return -1;
1822     if(avctx->flags&CODEC_FLAG_PASS1)
1823         ff_write_pass1_stats(&s->m);
1824     s->m.last_pict_type = s->m.pict_type;
1825     avctx->frame_bits = s->m.frame_bits;
1826     avctx->mv_bits = s->m.mv_bits;
1827     avctx->misc_bits = s->m.misc_bits;
1828     avctx->p_tex_bits = s->m.p_tex_bits;
1829
1830     emms_c();
1831
1832     pkt->size = ff_rac_terminate(c);
1833     if (avctx->coded_frame->key_frame)
1834         pkt->flags |= AV_PKT_FLAG_KEY;
1835     *got_packet = 1;
1836
1837     return 0;
1838 }
1839
1840 static av_cold int encode_end(AVCodecContext *avctx)
1841 {
1842     SnowContext *s = avctx->priv_data;
1843
1844     ff_snow_common_end(s);
1845     ff_rate_control_uninit(&s->m);
1846     av_frame_free(&s->input_picture);
1847     av_free(avctx->stats_out);
1848
1849     return 0;
1850 }
1851
1852 #define OFFSET(x) offsetof(SnowContext, x)
1853 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1854 static const AVOption options[] = {
1855     { "memc_only",      "Only do ME/MC (I frames -> ref, P frame -> ME+MC).",   OFFSET(memc_only), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1856     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1857     { NULL },
1858 };
1859
1860 static const AVClass snowenc_class = {
1861     .class_name = "snow encoder",
1862     .item_name  = av_default_item_name,
1863     .option     = options,
1864     .version    = LIBAVUTIL_VERSION_INT,
1865 };
1866
1867 AVCodec ff_snow_encoder = {
1868     .name           = "snow",
1869     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1870     .type           = AVMEDIA_TYPE_VIDEO,
1871     .id             = AV_CODEC_ID_SNOW,
1872     .priv_data_size = sizeof(SnowContext),
1873     .init           = encode_init,
1874     .encode2        = encode_frame,
1875     .close          = encode_end,
1876     .pix_fmts       = (const enum AVPixelFormat[]){
1877         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1878         AV_PIX_FMT_GRAY8,
1879         AV_PIX_FMT_NONE
1880     },
1881     .priv_class     = &snowenc_class,
1882 };
1883
1884
1885 #ifdef TEST
1886 #undef malloc
1887 #undef free
1888 #undef printf
1889
1890 #include "libavutil/lfg.h"
1891 #include "libavutil/mathematics.h"
1892
1893 int main(void){
1894 #define width  256
1895 #define height 256
1896     int buffer[2][width*height];
1897     SnowContext s;
1898     int i;
1899     AVLFG prng;
1900     s.spatial_decomposition_count=6;
1901     s.spatial_decomposition_type=1;
1902
1903     s.temp_dwt_buffer  = av_mallocz(width * sizeof(DWTELEM));
1904     s.temp_idwt_buffer = av_mallocz(width * sizeof(IDWTELEM));
1905
1906     av_lfg_init(&prng, 1);
1907
1908     printf("testing 5/3 DWT\n");
1909     for(i=0; i<width*height; i++)
1910         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1911
1912     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1913     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1914
1915     for(i=0; i<width*height; i++)
1916         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1917
1918     printf("testing 9/7 DWT\n");
1919     s.spatial_decomposition_type=0;
1920     for(i=0; i<width*height; i++)
1921         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1922
1923     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1924     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1925
1926     for(i=0; i<width*height; i++)
1927         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1928
1929     {
1930     int level, orientation, x, y;
1931     int64_t errors[8][4];
1932     int64_t g=0;
1933
1934         memset(errors, 0, sizeof(errors));
1935         s.spatial_decomposition_count=3;
1936         s.spatial_decomposition_type=0;
1937         for(level=0; level<s.spatial_decomposition_count; level++){
1938             for(orientation=level ? 1 : 0; orientation<4; orientation++){
1939                 int w= width  >> (s.spatial_decomposition_count-level);
1940                 int h= height >> (s.spatial_decomposition_count-level);
1941                 int stride= width  << (s.spatial_decomposition_count-level);
1942                 DWTELEM *buf= buffer[0];
1943                 int64_t error=0;
1944
1945                 if(orientation&1) buf+=w;
1946                 if(orientation>1) buf+=stride>>1;
1947
1948                 memset(buffer[0], 0, sizeof(int)*width*height);
1949                 buf[w/2 + h/2*stride]= 256*256;
1950                 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1951                 for(y=0; y<height; y++){
1952                     for(x=0; x<width; x++){
1953                         int64_t d= buffer[0][x + y*width];
1954                         error += d*d;
1955                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
1956                     }
1957                     if(FFABS(height/2-y)<9 && level==2) printf("\n");
1958                 }
1959                 error= (int)(sqrt(error)+0.5);
1960                 errors[level][orientation]= error;
1961                 if(g) g=av_gcd(g, error);
1962                 else g= error;
1963             }
1964         }
1965         printf("static int const visual_weight[][4]={\n");
1966         for(level=0; level<s.spatial_decomposition_count; level++){
1967             printf("  {");
1968             for(orientation=0; orientation<4; orientation++){
1969                 printf("%8"PRId64",", errors[level][orientation]/g);
1970             }
1971             printf("},\n");
1972         }
1973         printf("};\n");
1974         {
1975             int level=2;
1976             int w= width  >> (s.spatial_decomposition_count-level);
1977             //int h= height >> (s.spatial_decomposition_count-level);
1978             int stride= width  << (s.spatial_decomposition_count-level);
1979             DWTELEM *buf= buffer[0];
1980             int64_t error=0;
1981
1982             buf+=w;
1983             buf+=stride>>1;
1984
1985             memset(buffer[0], 0, sizeof(int)*width*height);
1986             for(y=0; y<height; y++){
1987                 for(x=0; x<width; x++){
1988                     int tab[4]={0,2,3,1};
1989                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
1990                 }
1991             }
1992             ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1993             for(y=0; y<height; y++){
1994                 for(x=0; x<width; x++){
1995                     int64_t d= buffer[0][x + y*width];
1996                     error += d*d;
1997                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
1998                 }
1999                 if(FFABS(height/2-y)<9) printf("\n");
2000             }
2001         }
2002
2003     }
2004     return 0;
2005 }
2006 #endif /* TEST */