snow: remove a VLA used for edge emulation
[ffmpeg.git] / libavcodec / snowenc.c
1 /*
2  * Copyright (C) 2004 Michael Niedermayer <michaelni@gmx.at>
3  *
4  * This file is part of Libav.
5  *
6  * Libav 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  * Libav 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 Libav; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20
21 #include "libavutil/intmath.h"
22 #include "libavutil/log.h"
23 #include "libavutil/opt.h"
24 #include "avcodec.h"
25 #include "dsputil.h"
26 #include "dwt.h"
27 #include "snow.h"
28
29 #include "rangecoder.h"
30 #include "mathops.h"
31
32 #include "mpegvideo.h"
33 #include "h263.h"
34
35 #undef NDEBUG
36 #include <assert.h>
37
38 #define QUANTIZE2 0
39
40 #if QUANTIZE2==1
41 #define Q2_STEP 8
42
43 static void find_sse(SnowContext *s, Plane *p, int *score, int score_stride, IDWTELEM *r0, IDWTELEM *r1, int level, int orientation){
44     SubBand *b= &p->band[level][orientation];
45     int x, y;
46     int xo=0;
47     int yo=0;
48     int step= 1 << (s->spatial_decomposition_count - level);
49
50     if(orientation&1)
51         xo= step>>1;
52     if(orientation&2)
53         yo= step>>1;
54
55     //FIXME bias for nonzero ?
56     //FIXME optimize
57     memset(score, 0, sizeof(*score)*score_stride*((p->height + Q2_STEP-1)/Q2_STEP));
58     for(y=0; y<p->height; y++){
59         for(x=0; x<p->width; x++){
60             int sx= (x-xo + step/2) / step / Q2_STEP;
61             int sy= (y-yo + step/2) / step / Q2_STEP;
62             int v= r0[x + y*p->width] - r1[x + y*p->width];
63             assert(sx>=0 && sy>=0 && sx < score_stride);
64             v= ((v+8)>>4)<<4;
65             score[sx + sy*score_stride] += v*v;
66             assert(score[sx + sy*score_stride] >= 0);
67         }
68     }
69 }
70
71 static void dequantize_all(SnowContext *s, Plane *p, IDWTELEM *buffer, int width, int height){
72     int level, orientation;
73
74     for(level=0; level<s->spatial_decomposition_count; level++){
75         for(orientation=level ? 1 : 0; orientation<4; orientation++){
76             SubBand *b= &p->band[level][orientation];
77             IDWTELEM *dst= buffer + (b->ibuf - s->spatial_idwt_buffer);
78
79             dequantize(s, b, dst, b->stride);
80         }
81     }
82 }
83
84 static void dwt_quantize(SnowContext *s, Plane *p, DWTELEM *buffer, int width, int height, int stride, int type){
85     int level, orientation, ys, xs, x, y, pass;
86     IDWTELEM best_dequant[height * stride];
87     IDWTELEM idwt2_buffer[height * stride];
88     const int score_stride= (width + 10)/Q2_STEP;
89     int best_score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
90     int score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
91     int threshold= (s->m.lambda * s->m.lambda) >> 6;
92
93     //FIXME pass the copy cleanly ?
94
95 //    memcpy(dwt_buffer, buffer, height * stride * sizeof(DWTELEM));
96     ff_spatial_dwt(buffer, s->temp_dwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
97
98     for(level=0; level<s->spatial_decomposition_count; level++){
99         for(orientation=level ? 1 : 0; orientation<4; orientation++){
100             SubBand *b= &p->band[level][orientation];
101             IDWTELEM *dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
102              DWTELEM *src=       buffer + (b-> buf - s->spatial_dwt_buffer);
103             assert(src == b->buf); // code does not depend on this but it is true currently
104
105             quantize(s, b, dst, src, b->stride, s->qbias);
106         }
107     }
108     for(pass=0; pass<1; pass++){
109         if(s->qbias == 0) //keyframe
110             continue;
111         for(level=0; level<s->spatial_decomposition_count; level++){
112             for(orientation=level ? 1 : 0; orientation<4; orientation++){
113                 SubBand *b= &p->band[level][orientation];
114                 IDWTELEM *dst= idwt2_buffer + (b->ibuf - s->spatial_idwt_buffer);
115                 IDWTELEM *best_dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
116
117                 for(ys= 0; ys<Q2_STEP; ys++){
118                     for(xs= 0; xs<Q2_STEP; xs++){
119                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
120                         dequantize_all(s, p, idwt2_buffer, width, height);
121                         ff_spatial_idwt(idwt2_buffer, s->temp_idwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
122                         find_sse(s, p, best_score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
123                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
124                         for(y=ys; y<b->height; y+= Q2_STEP){
125                             for(x=xs; x<b->width; x+= Q2_STEP){
126                                 if(dst[x + y*b->stride]<0) dst[x + y*b->stride]++;
127                                 if(dst[x + y*b->stride]>0) dst[x + y*b->stride]--;
128                                 //FIXME try more than just --
129                             }
130                         }
131                         dequantize_all(s, p, idwt2_buffer, width, height);
132                         ff_spatial_idwt(idwt2_buffer, s->temp_idwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
133                         find_sse(s, p, score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
134                         for(y=ys; y<b->height; y+= Q2_STEP){
135                             for(x=xs; x<b->width; x+= Q2_STEP){
136                                 int score_idx= x/Q2_STEP + (y/Q2_STEP)*score_stride;
137                                 if(score[score_idx] <= best_score[score_idx] + threshold){
138                                     best_score[score_idx]= score[score_idx];
139                                     if(best_dst[x + y*b->stride]<0) best_dst[x + y*b->stride]++;
140                                     if(best_dst[x + y*b->stride]>0) best_dst[x + y*b->stride]--;
141                                     //FIXME copy instead
142                                 }
143                             }
144                         }
145                     }
146                 }
147             }
148         }
149     }
150     memcpy(s->spatial_idwt_buffer, best_dequant, height * stride * sizeof(IDWTELEM)); //FIXME work with that directly instead of copy at the end
151 }
152
153 #endif /* QUANTIZE2==1 */
154
155 static av_cold int encode_init(AVCodecContext *avctx)
156 {
157     SnowContext *s = avctx->priv_data;
158     int plane_index, ret;
159
160     if(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL){
161         av_log(avctx, AV_LOG_ERROR, "This codec is under development, files encoded with it may not be decodable with future versions!!!\n"
162                "Use vstrict=-2 / -strict -2 to use it anyway.\n");
163         return -1;
164     }
165
166     if(avctx->prediction_method == DWT_97
167        && (avctx->flags & CODEC_FLAG_QSCALE)
168        && avctx->global_quality == 0){
169         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
170         return -1;
171     }
172
173     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
174
175     s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
176     s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
177
178     for(plane_index=0; plane_index<3; plane_index++){
179         s->plane[plane_index].diag_mc= 1;
180         s->plane[plane_index].htaps= 6;
181         s->plane[plane_index].hcoeff[0]=  40;
182         s->plane[plane_index].hcoeff[1]= -10;
183         s->plane[plane_index].hcoeff[2]=   2;
184         s->plane[plane_index].fast_mc= 1;
185     }
186
187     if ((ret = ff_snow_common_init(avctx)) < 0) {
188         ff_snow_common_end(avctx->priv_data);
189         return ret;
190     }
191     ff_snow_alloc_blocks(s);
192
193     s->version=0;
194
195     s->m.avctx   = avctx;
196     s->m.flags   = avctx->flags;
197     s->m.bit_rate= avctx->bit_rate;
198
199     s->m.me.temp      =
200     s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
201     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
202     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
203     s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
204     ff_h263_encode_init(&s->m); //mv_penalty
205
206     s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
207
208     if(avctx->flags&CODEC_FLAG_PASS1){
209         if(!avctx->stats_out)
210             avctx->stats_out = av_mallocz(256);
211     }
212     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
213         if(ff_rate_control_init(&s->m) < 0)
214             return -1;
215     }
216     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
217
218     avctx->coded_frame= &s->current_picture;
219     switch(avctx->pix_fmt){
220 //    case PIX_FMT_YUV444P:
221 //    case PIX_FMT_YUV422P:
222     case PIX_FMT_YUV420P:
223     case PIX_FMT_GRAY8:
224 //    case PIX_FMT_YUV411P:
225 //    case PIX_FMT_YUV410P:
226         s->colorspace_type= 0;
227         break;
228 /*    case PIX_FMT_RGB32:
229         s->colorspace= 1;
230         break;*/
231     default:
232         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
233         return -1;
234     }
235 //    avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
236     s->chroma_h_shift= 1;
237     s->chroma_v_shift= 1;
238
239     ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
240     ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
241
242     s->avctx->get_buffer(s->avctx, &s->input_picture);
243
244     if(s->avctx->me_method == ME_ITER){
245         int i;
246         int size= s->b_width * s->b_height << 2*s->block_max_depth;
247         for(i=0; i<s->max_ref_frames; i++){
248             s->ref_mvs[i]= av_mallocz(size*sizeof(int16_t[2]));
249             s->ref_scores[i]= av_mallocz(size*sizeof(uint32_t));
250         }
251     }
252
253     return 0;
254 }
255
256 //near copy & paste from dsputil, FIXME
257 static int pix_sum(uint8_t * pix, int line_size, int w)
258 {
259     int s, i, j;
260
261     s = 0;
262     for (i = 0; i < w; i++) {
263         for (j = 0; j < w; j++) {
264             s += pix[0];
265             pix ++;
266         }
267         pix += line_size - w;
268     }
269     return s;
270 }
271
272 //near copy & paste from dsputil, FIXME
273 static int pix_norm1(uint8_t * pix, int line_size, int w)
274 {
275     int s, i, j;
276     uint32_t *sq = ff_squareTbl + 256;
277
278     s = 0;
279     for (i = 0; i < w; i++) {
280         for (j = 0; j < w; j ++) {
281             s += sq[pix[0]];
282             pix ++;
283         }
284         pix += line_size - w;
285     }
286     return s;
287 }
288
289 //FIXME copy&paste
290 #define P_LEFT P[1]
291 #define P_TOP P[2]
292 #define P_TOPRIGHT P[3]
293 #define P_MEDIAN P[4]
294 #define P_MV1 P[9]
295 #define FLAG_QPEL   1 //must be 1
296
297 static int encode_q_branch(SnowContext *s, int level, int x, int y){
298     uint8_t p_buffer[1024];
299     uint8_t i_buffer[1024];
300     uint8_t p_state[sizeof(s->block_state)];
301     uint8_t i_state[sizeof(s->block_state)];
302     RangeCoder pc, ic;
303     uint8_t *pbbak= s->c.bytestream;
304     uint8_t *pbbak_start= s->c.bytestream_start;
305     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
306     const int w= s->b_width  << s->block_max_depth;
307     const int h= s->b_height << s->block_max_depth;
308     const int rem_depth= s->block_max_depth - level;
309     const int index= (x + y*w) << rem_depth;
310     const int block_w= 1<<(LOG2_MB_SIZE - level);
311     int trx= (x+1)<<rem_depth;
312     int try= (y+1)<<rem_depth;
313     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
314     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
315     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
316     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
317     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
318     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
319     int pl = left->color[0];
320     int pcb= left->color[1];
321     int pcr= left->color[2];
322     int pmx, pmy;
323     int mx=0, my=0;
324     int l,cr,cb;
325     const int stride= s->current_picture.linesize[0];
326     const int uvstride= s->current_picture.linesize[1];
327     uint8_t *current_data[3]= { s->input_picture.data[0] + (x + y*  stride)*block_w,
328                                 s->input_picture.data[1] + (x + y*uvstride)*block_w/2,
329                                 s->input_picture.data[2] + (x + y*uvstride)*block_w/2};
330     int P[10][2];
331     int16_t last_mv[3][2];
332     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
333     const int shift= 1+qpel;
334     MotionEstContext *c= &s->m.me;
335     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
336     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
337     int my_context= av_log2(2*FFABS(left->my - top->my));
338     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
339     int ref, best_ref, ref_score, ref_mx, ref_my;
340
341     assert(sizeof(s->block_state) >= 256);
342     if(s->keyframe){
343         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
344         return 0;
345     }
346
347 //    clip predictors / edge ?
348
349     P_LEFT[0]= left->mx;
350     P_LEFT[1]= left->my;
351     P_TOP [0]= top->mx;
352     P_TOP [1]= top->my;
353     P_TOPRIGHT[0]= tr->mx;
354     P_TOPRIGHT[1]= tr->my;
355
356     last_mv[0][0]= s->block[index].mx;
357     last_mv[0][1]= s->block[index].my;
358     last_mv[1][0]= right->mx;
359     last_mv[1][1]= right->my;
360     last_mv[2][0]= bottom->mx;
361     last_mv[2][1]= bottom->my;
362
363     s->m.mb_stride=2;
364     s->m.mb_x=
365     s->m.mb_y= 0;
366     c->skip= 0;
367
368     assert(c->  stride ==   stride);
369     assert(c->uvstride == uvstride);
370
371     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
372     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
373     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
374     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
375
376     c->xmin = - x*block_w - 16+3;
377     c->ymin = - y*block_w - 16+3;
378     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
379     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
380
381     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
382     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
383     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
384     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
385     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
386     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
387     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
388
389     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
390     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
391
392     if (!y) {
393         c->pred_x= P_LEFT[0];
394         c->pred_y= P_LEFT[1];
395     } else {
396         c->pred_x = P_MEDIAN[0];
397         c->pred_y = P_MEDIAN[1];
398     }
399
400     score= INT_MAX;
401     best_ref= 0;
402     for(ref=0; ref<s->ref_frames; ref++){
403         init_ref(c, current_data, s->last_picture[ref].data, NULL, block_w*x, block_w*y, 0);
404
405         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
406                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
407
408         assert(ref_mx >= c->xmin);
409         assert(ref_mx <= c->xmax);
410         assert(ref_my >= c->ymin);
411         assert(ref_my <= c->ymax);
412
413         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
414         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
415         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
416         if(s->ref_mvs[ref]){
417             s->ref_mvs[ref][index][0]= ref_mx;
418             s->ref_mvs[ref][index][1]= ref_my;
419             s->ref_scores[ref][index]= ref_score;
420         }
421         if(score > ref_score){
422             score= ref_score;
423             best_ref= ref;
424             mx= ref_mx;
425             my= ref_my;
426         }
427     }
428     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
429
430   //  subpel search
431     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
432     pc= s->c;
433     pc.bytestream_start=
434     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
435     memcpy(p_state, s->block_state, sizeof(s->block_state));
436
437     if(level!=s->block_max_depth)
438         put_rac(&pc, &p_state[4 + s_context], 1);
439     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
440     if(s->ref_frames > 1)
441         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
442     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
443     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
444     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
445     p_len= pc.bytestream - pc.bytestream_start;
446     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
447
448     block_s= block_w*block_w;
449     sum = pix_sum(current_data[0], stride, block_w);
450     l= (sum + block_s/2)/block_s;
451     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
452
453     block_s= block_w*block_w>>2;
454     sum = pix_sum(current_data[1], uvstride, block_w>>1);
455     cb= (sum + block_s/2)/block_s;
456 //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
457     sum = pix_sum(current_data[2], uvstride, block_w>>1);
458     cr= (sum + block_s/2)/block_s;
459 //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
460
461     ic= s->c;
462     ic.bytestream_start=
463     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
464     memcpy(i_state, s->block_state, sizeof(s->block_state));
465     if(level!=s->block_max_depth)
466         put_rac(&ic, &i_state[4 + s_context], 1);
467     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
468     put_symbol(&ic, &i_state[32],  l-pl , 1);
469     put_symbol(&ic, &i_state[64], cb-pcb, 1);
470     put_symbol(&ic, &i_state[96], cr-pcr, 1);
471     i_len= ic.bytestream - ic.bytestream_start;
472     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
473
474 //    assert(score==256*256*256*64-1);
475     assert(iscore < 255*255*256 + s->lambda2*10);
476     assert(iscore >= 0);
477     assert(l>=0 && l<=255);
478     assert(pl>=0 && pl<=255);
479
480     if(level==0){
481         int varc= iscore >> 8;
482         int vard= score >> 8;
483         if (vard <= 64 || vard < varc)
484             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
485         else
486             c->scene_change_score+= s->m.qscale;
487     }
488
489     if(level!=s->block_max_depth){
490         put_rac(&s->c, &s->block_state[4 + s_context], 0);
491         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
492         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
493         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
494         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
495         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
496
497         if(score2 < score && score2 < iscore)
498             return score2;
499     }
500
501     if(iscore < score){
502         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
503         memcpy(pbbak, i_buffer, i_len);
504         s->c= ic;
505         s->c.bytestream_start= pbbak_start;
506         s->c.bytestream= pbbak + i_len;
507         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
508         memcpy(s->block_state, i_state, sizeof(s->block_state));
509         return iscore;
510     }else{
511         memcpy(pbbak, p_buffer, p_len);
512         s->c= pc;
513         s->c.bytestream_start= pbbak_start;
514         s->c.bytestream= pbbak + p_len;
515         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
516         memcpy(s->block_state, p_state, sizeof(s->block_state));
517         return score;
518     }
519 }
520
521 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
522     const int w= s->b_width  << s->block_max_depth;
523     const int rem_depth= s->block_max_depth - level;
524     const int index= (x + y*w) << rem_depth;
525     int trx= (x+1)<<rem_depth;
526     BlockNode *b= &s->block[index];
527     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
528     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
529     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
530     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
531     int pl = left->color[0];
532     int pcb= left->color[1];
533     int pcr= left->color[2];
534     int pmx, pmy;
535     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
536     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
537     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
538     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
539
540     if(s->keyframe){
541         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
542         return;
543     }
544
545     if(level!=s->block_max_depth){
546         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
547             put_rac(&s->c, &s->block_state[4 + s_context], 1);
548         }else{
549             put_rac(&s->c, &s->block_state[4 + s_context], 0);
550             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
551             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
552             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
553             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
554             return;
555         }
556     }
557     if(b->type & BLOCK_INTRA){
558         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
559         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
560         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
561         put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
562         put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
563         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
564     }else{
565         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
566         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
567         if(s->ref_frames > 1)
568             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
569         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
570         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
571         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
572     }
573 }
574
575 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
576     int i, x2, y2;
577     Plane *p= &s->plane[plane_index];
578     const int block_size = MB_SIZE >> s->block_max_depth;
579     const int block_w    = plane_index ? block_size/2 : block_size;
580     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+1] : ff_obmc_tab[s->block_max_depth];
581     const int obmc_stride= plane_index ? block_size : 2*block_size;
582     const int ref_stride= s->current_picture.linesize[plane_index];
583     uint8_t *src= s-> input_picture.data[plane_index];
584     IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
585     const int b_stride = s->b_width << s->block_max_depth;
586     const int w= p->width;
587     const int h= p->height;
588     int index= mb_x + mb_y*b_stride;
589     BlockNode *b= &s->block[index];
590     BlockNode backup= *b;
591     int ab=0;
592     int aa=0;
593
594     b->type|= BLOCK_INTRA;
595     b->color[plane_index]= 0;
596     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
597
598     for(i=0; i<4; i++){
599         int mb_x2= mb_x + (i &1) - 1;
600         int mb_y2= mb_y + (i>>1) - 1;
601         int x= block_w*mb_x2 + block_w/2;
602         int y= block_w*mb_y2 + block_w/2;
603
604         add_yblock(s, 0, NULL, dst + ((i&1)+(i>>1)*obmc_stride)*block_w, NULL, obmc,
605                     x, y, block_w, block_w, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
606
607         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_w); y2++){
608             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
609                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_w*mb_y - block_w/2))*obmc_stride;
610                 int obmc_v= obmc[index];
611                 int d;
612                 if(y<0) obmc_v += obmc[index + block_w*obmc_stride];
613                 if(x<0) obmc_v += obmc[index + block_w];
614                 if(y+block_w>h) obmc_v += obmc[index - block_w*obmc_stride];
615                 if(x+block_w>w) obmc_v += obmc[index - block_w];
616                 //FIXME precalculate this or simplify it somehow else
617
618                 d = -dst[index] + (1<<(FRAC_BITS-1));
619                 dst[index] = d;
620                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
621                 aa += obmc_v * obmc_v; //FIXME precalculate this
622             }
623         }
624     }
625     *b= backup;
626
627     return av_clip(((ab<<LOG2_OBMC_MAX) + aa/2)/aa, 0, 255); //FIXME we should not need clipping
628 }
629
630 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
631     const int b_stride = s->b_width << s->block_max_depth;
632     const int b_height = s->b_height<< s->block_max_depth;
633     int index= x + y*b_stride;
634     const BlockNode *b     = &s->block[index];
635     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
636     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
637     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
638     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
639     int dmx, dmy;
640 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
641 //  int my_context= av_log2(2*FFABS(left->my - top->my));
642
643     if(x<0 || x>=b_stride || y>=b_height)
644         return 0;
645 /*
646 1            0      0
647 01X          1-2    1
648 001XX        3-6    2-3
649 0001XXX      7-14   4-7
650 00001XXXX   15-30   8-15
651 */
652 //FIXME try accurate rate
653 //FIXME intra and inter predictors if surrounding blocks are not the same type
654     if(b->type & BLOCK_INTRA){
655         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
656                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
657                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
658     }else{
659         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
660         dmx-= b->mx;
661         dmy-= b->my;
662         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
663                     + av_log2(2*FFABS(dmy))
664                     + av_log2(2*b->ref));
665     }
666 }
667
668 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
669     Plane *p= &s->plane[plane_index];
670     const int block_size = MB_SIZE >> s->block_max_depth;
671     const int block_w    = plane_index ? block_size/2 : block_size;
672     const int obmc_stride= plane_index ? block_size : 2*block_size;
673     const int ref_stride= s->current_picture.linesize[plane_index];
674     uint8_t *dst= s->current_picture.data[plane_index];
675     uint8_t *src= s->  input_picture.data[plane_index];
676     IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
677     uint8_t *cur = s->scratchbuf;
678     uint8_t *tmp = s->emu_edge_buffer;
679     const int b_stride = s->b_width << s->block_max_depth;
680     const int b_height = s->b_height<< s->block_max_depth;
681     const int w= p->width;
682     const int h= p->height;
683     int distortion;
684     int rate= 0;
685     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
686     int sx= block_w*mb_x - block_w/2;
687     int sy= block_w*mb_y - block_w/2;
688     int x0= FFMAX(0,-sx);
689     int y0= FFMAX(0,-sy);
690     int x1= FFMIN(block_w*2, w-sx);
691     int y1= FFMIN(block_w*2, h-sy);
692     int i,x,y;
693
694     ff_snow_pred_block(s, cur, tmp, ref_stride, sx, sy, block_w*2, block_w*2, &s->block[mb_x + mb_y*b_stride], plane_index, w, h);
695
696     for(y=y0; y<y1; y++){
697         const uint8_t *obmc1= obmc_edged[y];
698         const IDWTELEM *pred1 = pred + y*obmc_stride;
699         uint8_t *cur1 = cur + y*ref_stride;
700         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
701         for(x=x0; x<x1; x++){
702 #if FRAC_BITS >= LOG2_OBMC_MAX
703             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
704 #else
705             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
706 #endif
707             v = (v + pred1[x]) >> FRAC_BITS;
708             if(v&(~255)) v= ~(v>>31);
709             dst1[x] = v;
710         }
711     }
712
713     /* copy the regions where obmc[] = (uint8_t)256 */
714     if(LOG2_OBMC_MAX == 8
715         && (mb_x == 0 || mb_x == b_stride-1)
716         && (mb_y == 0 || mb_y == b_height-1)){
717         if(mb_x == 0)
718             x1 = block_w;
719         else
720             x0 = block_w;
721         if(mb_y == 0)
722             y1 = block_w;
723         else
724             y0 = block_w;
725         for(y=y0; y<y1; y++)
726             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
727     }
728
729     if(block_w==16){
730         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
731         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
732         /* FIXME cmps overlap but do not cover the wavelet's whole support.
733          * So improving the score of one block is not strictly guaranteed
734          * to improve the score of the whole frame, thus iterative motion
735          * estimation does not always converge. */
736         if(s->avctx->me_cmp == FF_CMP_W97)
737             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
738         else if(s->avctx->me_cmp == FF_CMP_W53)
739             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
740         else{
741             distortion = 0;
742             for(i=0; i<4; i++){
743                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
744                 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
745             }
746         }
747     }else{
748         assert(block_w==8);
749         distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
750     }
751
752     if(plane_index==0){
753         for(i=0; i<4; i++){
754 /* ..RRr
755  * .RXx.
756  * rxx..
757  */
758             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
759         }
760         if(mb_x == b_stride-2)
761             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
762     }
763     return distortion + rate*penalty_factor;
764 }
765
766 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
767     int i, y2;
768     Plane *p= &s->plane[plane_index];
769     const int block_size = MB_SIZE >> s->block_max_depth;
770     const int block_w    = plane_index ? block_size/2 : block_size;
771     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+1] : ff_obmc_tab[s->block_max_depth];
772     const int obmc_stride= plane_index ? block_size : 2*block_size;
773     const int ref_stride= s->current_picture.linesize[plane_index];
774     uint8_t *dst= s->current_picture.data[plane_index];
775     uint8_t *src= s-> input_picture.data[plane_index];
776     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
777     // const has only been removed from zero_dst to suppress a warning
778     static IDWTELEM zero_dst[4096]; //FIXME
779     const int b_stride = s->b_width << s->block_max_depth;
780     const int w= p->width;
781     const int h= p->height;
782     int distortion= 0;
783     int rate= 0;
784     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
785
786     for(i=0; i<9; i++){
787         int mb_x2= mb_x + (i%3) - 1;
788         int mb_y2= mb_y + (i/3) - 1;
789         int x= block_w*mb_x2 + block_w/2;
790         int y= block_w*mb_y2 + block_w/2;
791
792         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
793                    x, y, block_w, block_w, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
794
795         //FIXME find a cleaner/simpler way to skip the outside stuff
796         for(y2= y; y2<0; y2++)
797             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
798         for(y2= h; y2<y+block_w; y2++)
799             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
800         if(x<0){
801             for(y2= y; y2<y+block_w; y2++)
802                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
803         }
804         if(x+block_w > w){
805             for(y2= y; y2<y+block_w; y2++)
806                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
807         }
808
809         assert(block_w== 8 || block_w==16);
810         distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_w);
811     }
812
813     if(plane_index==0){
814         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
815         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
816
817 /* ..RRRr
818  * .RXXx.
819  * .RXXx.
820  * rxxx.
821  */
822         if(merged)
823             rate = get_block_bits(s, mb_x, mb_y, 2);
824         for(i=merged?4:0; i<9; i++){
825             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
826             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
827         }
828     }
829     return distortion + rate*penalty_factor;
830 }
831
832 static int encode_subband_c0run(SnowContext *s, SubBand *b, IDWTELEM *src, IDWTELEM *parent, int stride, int orientation){
833     const int w= b->width;
834     const int h= b->height;
835     int x, y;
836
837     if(1){
838         int run=0;
839         int *runs = s->run_buffer;
840         int run_index=0;
841         int max_index;
842
843         for(y=0; y<h; y++){
844             for(x=0; x<w; x++){
845                 int v, p=0;
846                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
847                 v= src[x + y*stride];
848
849                 if(y){
850                     t= src[x + (y-1)*stride];
851                     if(x){
852                         lt= src[x - 1 + (y-1)*stride];
853                     }
854                     if(x + 1 < w){
855                         rt= src[x + 1 + (y-1)*stride];
856                     }
857                 }
858                 if(x){
859                     l= src[x - 1 + y*stride];
860                     /*if(x > 1){
861                         if(orientation==1) ll= src[y + (x-2)*stride];
862                         else               ll= src[x - 2 + y*stride];
863                     }*/
864                 }
865                 if(parent){
866                     int px= x>>1;
867                     int py= y>>1;
868                     if(px<b->parent->width && py<b->parent->height)
869                         p= parent[px + py*2*stride];
870                 }
871                 if(!(/*ll|*/l|lt|t|rt|p)){
872                     if(v){
873                         runs[run_index++]= run;
874                         run=0;
875                     }else{
876                         run++;
877                     }
878                 }
879             }
880         }
881         max_index= run_index;
882         runs[run_index++]= run;
883         run_index=0;
884         run= runs[run_index++];
885
886         put_symbol2(&s->c, b->state[30], max_index, 0);
887         if(run_index <= max_index)
888             put_symbol2(&s->c, b->state[1], run, 3);
889
890         for(y=0; y<h; y++){
891             if(s->c.bytestream_end - s->c.bytestream < w*40){
892                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
893                 return -1;
894             }
895             for(x=0; x<w; x++){
896                 int v, p=0;
897                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
898                 v= src[x + y*stride];
899
900                 if(y){
901                     t= src[x + (y-1)*stride];
902                     if(x){
903                         lt= src[x - 1 + (y-1)*stride];
904                     }
905                     if(x + 1 < w){
906                         rt= src[x + 1 + (y-1)*stride];
907                     }
908                 }
909                 if(x){
910                     l= src[x - 1 + y*stride];
911                     /*if(x > 1){
912                         if(orientation==1) ll= src[y + (x-2)*stride];
913                         else               ll= src[x - 2 + y*stride];
914                     }*/
915                 }
916                 if(parent){
917                     int px= x>>1;
918                     int py= y>>1;
919                     if(px<b->parent->width && py<b->parent->height)
920                         p= parent[px + py*2*stride];
921                 }
922                 if(/*ll|*/l|lt|t|rt|p){
923                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
924
925                     put_rac(&s->c, &b->state[0][context], !!v);
926                 }else{
927                     if(!run){
928                         run= runs[run_index++];
929
930                         if(run_index <= max_index)
931                             put_symbol2(&s->c, b->state[1], run, 3);
932                         assert(v);
933                     }else{
934                         run--;
935                         assert(!v);
936                     }
937                 }
938                 if(v){
939                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
940                     int l2= 2*FFABS(l) + (l<0);
941                     int t2= 2*FFABS(t) + (t<0);
942
943                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
944                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
945                 }
946             }
947         }
948     }
949     return 0;
950 }
951
952 static int encode_subband(SnowContext *s, SubBand *b, IDWTELEM *src, IDWTELEM *parent, int stride, int orientation){
953 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
954 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
955     return encode_subband_c0run(s, b, src, parent, stride, orientation);
956 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
957 }
958
959 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){
960     const int b_stride= s->b_width << s->block_max_depth;
961     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
962     BlockNode backup= *block;
963     unsigned value;
964     int rd, index;
965
966     assert(mb_x>=0 && mb_y>=0);
967     assert(mb_x<b_stride);
968
969     if(intra){
970         block->color[0] = p[0];
971         block->color[1] = p[1];
972         block->color[2] = p[2];
973         block->type |= BLOCK_INTRA;
974     }else{
975         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
976         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
977         if(s->me_cache[index] == value)
978             return 0;
979         s->me_cache[index]= value;
980
981         block->mx= p[0];
982         block->my= p[1];
983         block->type &= ~BLOCK_INTRA;
984     }
985
986     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
987
988 //FIXME chroma
989     if(rd < *best_rd){
990         *best_rd= rd;
991         return 1;
992     }else{
993         *block= backup;
994         return 0;
995     }
996 }
997
998 /* special case for int[2] args we discard afterwards,
999  * fixes compilation problem with gcc 2.95 */
1000 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){
1001     int p[2] = {p0, p1};
1002     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
1003 }
1004
1005 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){
1006     const int b_stride= s->b_width << s->block_max_depth;
1007     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
1008     BlockNode backup[4]= {block[0], block[1], block[b_stride], block[b_stride+1]};
1009     unsigned value;
1010     int rd, index;
1011
1012     assert(mb_x>=0 && mb_y>=0);
1013     assert(mb_x<b_stride);
1014     assert(((mb_x|mb_y)&1) == 0);
1015
1016     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
1017     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
1018     if(s->me_cache[index] == value)
1019         return 0;
1020     s->me_cache[index]= value;
1021
1022     block->mx= p0;
1023     block->my= p1;
1024     block->ref= ref;
1025     block->type &= ~BLOCK_INTRA;
1026     block[1]= block[b_stride]= block[b_stride+1]= *block;
1027
1028     rd= get_4block_rd(s, mb_x, mb_y, 0);
1029
1030 //FIXME chroma
1031     if(rd < *best_rd){
1032         *best_rd= rd;
1033         return 1;
1034     }else{
1035         block[0]= backup[0];
1036         block[1]= backup[1];
1037         block[b_stride]= backup[2];
1038         block[b_stride+1]= backup[3];
1039         return 0;
1040     }
1041 }
1042
1043 static void iterative_me(SnowContext *s){
1044     int pass, mb_x, mb_y;
1045     const int b_width = s->b_width  << s->block_max_depth;
1046     const int b_height= s->b_height << s->block_max_depth;
1047     const int b_stride= b_width;
1048     int color[3];
1049
1050     {
1051         RangeCoder r = s->c;
1052         uint8_t state[sizeof(s->block_state)];
1053         memcpy(state, s->block_state, sizeof(s->block_state));
1054         for(mb_y= 0; mb_y<s->b_height; mb_y++)
1055             for(mb_x= 0; mb_x<s->b_width; mb_x++)
1056                 encode_q_branch(s, 0, mb_x, mb_y);
1057         s->c = r;
1058         memcpy(s->block_state, state, sizeof(s->block_state));
1059     }
1060
1061     for(pass=0; pass<25; pass++){
1062         int change= 0;
1063
1064         for(mb_y= 0; mb_y<b_height; mb_y++){
1065             for(mb_x= 0; mb_x<b_width; mb_x++){
1066                 int dia_change, i, j, ref;
1067                 int best_rd= INT_MAX, ref_rd;
1068                 BlockNode backup, ref_b;
1069                 const int index= mb_x + mb_y * b_stride;
1070                 BlockNode *block= &s->block[index];
1071                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1072                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1073                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1074                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1075                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1076                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1077                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1078                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1079                 const int b_w= (MB_SIZE >> s->block_max_depth);
1080                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1081
1082                 if(pass && (block->type & BLOCK_OPT))
1083                     continue;
1084                 block->type |= BLOCK_OPT;
1085
1086                 backup= *block;
1087
1088                 if(!s->me_cache_generation)
1089                     memset(s->me_cache, 0, sizeof(s->me_cache));
1090                 s->me_cache_generation += 1<<22;
1091
1092                 //FIXME precalculate
1093                 {
1094                     int x, y;
1095                     for (y = 0; y < b_w * 2; y++)
1096                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1097                     if(mb_x==0)
1098                         for(y=0; y<b_w*2; y++)
1099                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1100                     if(mb_x==b_stride-1)
1101                         for(y=0; y<b_w*2; y++)
1102                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1103                     if(mb_y==0){
1104                         for(x=0; x<b_w*2; x++)
1105                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1106                         for(y=1; y<b_w; y++)
1107                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1108                     }
1109                     if(mb_y==b_height-1){
1110                         for(x=0; x<b_w*2; x++)
1111                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1112                         for(y=b_w; y<b_w*2-1; y++)
1113                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1114                     }
1115                 }
1116
1117                 //skip stuff outside the picture
1118                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1119                     uint8_t *src= s->  input_picture.data[0];
1120                     uint8_t *dst= s->current_picture.data[0];
1121                     const int stride= s->current_picture.linesize[0];
1122                     const int block_w= MB_SIZE >> s->block_max_depth;
1123                     const int sx= block_w*mb_x - block_w/2;
1124                     const int sy= block_w*mb_y - block_w/2;
1125                     const int w= s->plane[0].width;
1126                     const int h= s->plane[0].height;
1127                     int y;
1128
1129                     for(y=sy; y<0; y++)
1130                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1131                     for(y=h; y<sy+block_w*2; y++)
1132                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1133                     if(sx<0){
1134                         for(y=sy; y<sy+block_w*2; y++)
1135                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1136                     }
1137                     if(sx+block_w*2 > w){
1138                         for(y=sy; y<sy+block_w*2; y++)
1139                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1140                     }
1141                 }
1142
1143                 // intra(black) = neighbors' contribution to the current block
1144                 for(i=0; i<3; i++)
1145                     color[i]= get_dc(s, mb_x, mb_y, i);
1146
1147                 // get previous score (cannot be cached due to OBMC)
1148                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1149                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1150                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1151                 }else
1152                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1153
1154                 ref_b= *block;
1155                 ref_rd= best_rd;
1156                 for(ref=0; ref < s->ref_frames; ref++){
1157                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1158                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1159                         continue;
1160                     block->ref= ref;
1161                     best_rd= INT_MAX;
1162
1163                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1164                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1165                     if(tb)
1166                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1167                     if(lb)
1168                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1169                     if(rb)
1170                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1171                     if(bb)
1172                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1173
1174                     /* fullpel ME */
1175                     //FIXME avoid subpel interpolation / round to nearest integer
1176                     do{
1177                         dia_change=0;
1178                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1179                             for(j=0; j<i; j++){
1180                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1181                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1182                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1183                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1184                             }
1185                         }
1186                     }while(dia_change);
1187                     /* subpel ME */
1188                     do{
1189                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1190                         dia_change=0;
1191                         for(i=0; i<8; i++)
1192                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1193                     }while(dia_change);
1194                     //FIXME or try the standard 2 pass qpel or similar
1195
1196                     mvr[0][0]= block->mx;
1197                     mvr[0][1]= block->my;
1198                     if(ref_rd > best_rd){
1199                         ref_rd= best_rd;
1200                         ref_b= *block;
1201                     }
1202                 }
1203                 best_rd= ref_rd;
1204                 *block= ref_b;
1205                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1206                 //FIXME RD style color selection
1207                 if(!same_block(block, &backup)){
1208                     if(tb ) tb ->type &= ~BLOCK_OPT;
1209                     if(lb ) lb ->type &= ~BLOCK_OPT;
1210                     if(rb ) rb ->type &= ~BLOCK_OPT;
1211                     if(bb ) bb ->type &= ~BLOCK_OPT;
1212                     if(tlb) tlb->type &= ~BLOCK_OPT;
1213                     if(trb) trb->type &= ~BLOCK_OPT;
1214                     if(blb) blb->type &= ~BLOCK_OPT;
1215                     if(brb) brb->type &= ~BLOCK_OPT;
1216                     change ++;
1217                 }
1218             }
1219         }
1220         av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1221         if(!change)
1222             break;
1223     }
1224
1225     if(s->block_max_depth == 1){
1226         int change= 0;
1227         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1228             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1229                 int i;
1230                 int best_rd, init_rd;
1231                 const int index= mb_x + mb_y * b_stride;
1232                 BlockNode *b[4];
1233
1234                 b[0]= &s->block[index];
1235                 b[1]= b[0]+1;
1236                 b[2]= b[0]+b_stride;
1237                 b[3]= b[2]+1;
1238                 if(same_block(b[0], b[1]) &&
1239                    same_block(b[0], b[2]) &&
1240                    same_block(b[0], b[3]))
1241                     continue;
1242
1243                 if(!s->me_cache_generation)
1244                     memset(s->me_cache, 0, sizeof(s->me_cache));
1245                 s->me_cache_generation += 1<<22;
1246
1247                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1248
1249                 //FIXME more multiref search?
1250                 check_4block_inter(s, mb_x, mb_y,
1251                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1252                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1253
1254                 for(i=0; i<4; i++)
1255                     if(!(b[i]->type&BLOCK_INTRA))
1256                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1257
1258                 if(init_rd != best_rd)
1259                     change++;
1260             }
1261         }
1262         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1263     }
1264 }
1265
1266 static void encode_blocks(SnowContext *s, int search){
1267     int x, y;
1268     int w= s->b_width;
1269     int h= s->b_height;
1270
1271     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1272         iterative_me(s);
1273
1274     for(y=0; y<h; y++){
1275         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1276             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1277             return;
1278         }
1279         for(x=0; x<w; x++){
1280             if(s->avctx->me_method == ME_ITER || !search)
1281                 encode_q_branch2(s, 0, x, y);
1282             else
1283                 encode_q_branch (s, 0, x, y);
1284         }
1285     }
1286 }
1287
1288 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1289     const int w= b->width;
1290     const int h= b->height;
1291     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1292     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1293     int x,y, thres1, thres2;
1294
1295     if(s->qlog == LOSSLESS_QLOG){
1296         for(y=0; y<h; y++)
1297             for(x=0; x<w; x++)
1298                 dst[x + y*stride]= src[x + y*stride];
1299         return;
1300     }
1301
1302     bias= bias ? 0 : (3*qmul)>>3;
1303     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1304     thres2= 2*thres1;
1305
1306     if(!bias){
1307         for(y=0; y<h; y++){
1308             for(x=0; x<w; x++){
1309                 int i= src[x + y*stride];
1310
1311                 if((unsigned)(i+thres1) > thres2){
1312                     if(i>=0){
1313                         i<<= QEXPSHIFT;
1314                         i/= qmul; //FIXME optimize
1315                         dst[x + y*stride]=  i;
1316                     }else{
1317                         i= -i;
1318                         i<<= QEXPSHIFT;
1319                         i/= qmul; //FIXME optimize
1320                         dst[x + y*stride]= -i;
1321                     }
1322                 }else
1323                     dst[x + y*stride]= 0;
1324             }
1325         }
1326     }else{
1327         for(y=0; y<h; y++){
1328             for(x=0; x<w; x++){
1329                 int i= src[x + y*stride];
1330
1331                 if((unsigned)(i+thres1) > thres2){
1332                     if(i>=0){
1333                         i<<= QEXPSHIFT;
1334                         i= (i + bias) / qmul; //FIXME optimize
1335                         dst[x + y*stride]=  i;
1336                     }else{
1337                         i= -i;
1338                         i<<= QEXPSHIFT;
1339                         i= (i + bias) / qmul; //FIXME optimize
1340                         dst[x + y*stride]= -i;
1341                     }
1342                 }else
1343                     dst[x + y*stride]= 0;
1344             }
1345         }
1346     }
1347 }
1348
1349 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1350     const int w= b->width;
1351     const int h= b->height;
1352     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1353     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1354     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1355     int x,y;
1356
1357     if(s->qlog == LOSSLESS_QLOG) return;
1358
1359     for(y=0; y<h; y++){
1360         for(x=0; x<w; x++){
1361             int i= src[x + y*stride];
1362             if(i<0){
1363                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1364             }else if(i>0){
1365                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1366             }
1367         }
1368     }
1369 }
1370
1371 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1372     const int w= b->width;
1373     const int h= b->height;
1374     int x,y;
1375
1376     for(y=h-1; y>=0; y--){
1377         for(x=w-1; x>=0; x--){
1378             int i= x + y*stride;
1379
1380             if(x){
1381                 if(use_median){
1382                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1383                     else  src[i] -= src[i - 1];
1384                 }else{
1385                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1386                     else  src[i] -= src[i - 1];
1387                 }
1388             }else{
1389                 if(y) src[i] -= src[i - stride];
1390             }
1391         }
1392     }
1393 }
1394
1395 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1396     const int w= b->width;
1397     const int h= b->height;
1398     int x,y;
1399
1400     for(y=0; y<h; y++){
1401         for(x=0; x<w; x++){
1402             int i= x + y*stride;
1403
1404             if(x){
1405                 if(use_median){
1406                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1407                     else  src[i] += src[i - 1];
1408                 }else{
1409                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1410                     else  src[i] += src[i - 1];
1411                 }
1412             }else{
1413                 if(y) src[i] += src[i - stride];
1414             }
1415         }
1416     }
1417 }
1418
1419 static void encode_qlogs(SnowContext *s){
1420     int plane_index, level, orientation;
1421
1422     for(plane_index=0; plane_index<2; plane_index++){
1423         for(level=0; level<s->spatial_decomposition_count; level++){
1424             for(orientation=level ? 1:0; orientation<4; orientation++){
1425                 if(orientation==2) continue;
1426                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1427             }
1428         }
1429     }
1430 }
1431
1432 static void encode_header(SnowContext *s){
1433     int plane_index, i;
1434     uint8_t kstate[32];
1435
1436     memset(kstate, MID_STATE, sizeof(kstate));
1437
1438     put_rac(&s->c, kstate, s->keyframe);
1439     if(s->keyframe || s->always_reset){
1440         ff_snow_reset_contexts(s);
1441         s->last_spatial_decomposition_type=
1442         s->last_qlog=
1443         s->last_qbias=
1444         s->last_mv_scale=
1445         s->last_block_max_depth= 0;
1446         for(plane_index=0; plane_index<2; plane_index++){
1447             Plane *p= &s->plane[plane_index];
1448             p->last_htaps=0;
1449             p->last_diag_mc=0;
1450             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1451         }
1452     }
1453     if(s->keyframe){
1454         put_symbol(&s->c, s->header_state, s->version, 0);
1455         put_rac(&s->c, s->header_state, s->always_reset);
1456         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1457         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1458         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1459         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1460         put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1461         put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1462         put_rac(&s->c, s->header_state, s->spatial_scalability);
1463 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1464         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1465
1466         encode_qlogs(s);
1467     }
1468
1469     if(!s->keyframe){
1470         int update_mc=0;
1471         for(plane_index=0; plane_index<2; plane_index++){
1472             Plane *p= &s->plane[plane_index];
1473             update_mc |= p->last_htaps   != p->htaps;
1474             update_mc |= p->last_diag_mc != p->diag_mc;
1475             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1476         }
1477         put_rac(&s->c, s->header_state, update_mc);
1478         if(update_mc){
1479             for(plane_index=0; plane_index<2; plane_index++){
1480                 Plane *p= &s->plane[plane_index];
1481                 put_rac(&s->c, s->header_state, p->diag_mc);
1482                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1483                 for(i= p->htaps/2; i; i--)
1484                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1485             }
1486         }
1487         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1488             put_rac(&s->c, s->header_state, 1);
1489             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1490             encode_qlogs(s);
1491         }else
1492             put_rac(&s->c, s->header_state, 0);
1493     }
1494
1495     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1496     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1497     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1498     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1499     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1500
1501 }
1502
1503 static void update_last_header_values(SnowContext *s){
1504     int plane_index;
1505
1506     if(!s->keyframe){
1507         for(plane_index=0; plane_index<2; plane_index++){
1508             Plane *p= &s->plane[plane_index];
1509             p->last_diag_mc= p->diag_mc;
1510             p->last_htaps  = p->htaps;
1511             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1512         }
1513     }
1514
1515     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1516     s->last_qlog                        = s->qlog;
1517     s->last_qbias                       = s->qbias;
1518     s->last_mv_scale                    = s->mv_scale;
1519     s->last_block_max_depth             = s->block_max_depth;
1520     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1521 }
1522
1523 static int qscale2qlog(int qscale){
1524     return rint(QROOT*log(qscale / (float)FF_QP2LAMBDA)/log(2))
1525            + 61*QROOT/8; ///< 64 > 60
1526 }
1527
1528 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1529 {
1530     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1531      * FIXME we know exact mv bits at this point,
1532      * but ratecontrol isn't set up to include them. */
1533     uint32_t coef_sum= 0;
1534     int level, orientation, delta_qlog;
1535
1536     for(level=0; level<s->spatial_decomposition_count; level++){
1537         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1538             SubBand *b= &s->plane[0].band[level][orientation];
1539             IDWTELEM *buf= b->ibuf;
1540             const int w= b->width;
1541             const int h= b->height;
1542             const int stride= b->stride;
1543             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1544             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1545             const int qdiv= (1<<16)/qmul;
1546             int x, y;
1547             //FIXME this is ugly
1548             for(y=0; y<h; y++)
1549                 for(x=0; x<w; x++)
1550                     buf[x+y*stride]= b->buf[x+y*stride];
1551             if(orientation==0)
1552                 decorrelate(s, b, buf, stride, 1, 0);
1553             for(y=0; y<h; y++)
1554                 for(x=0; x<w; x++)
1555                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1556         }
1557     }
1558
1559     /* ugly, ratecontrol just takes a sqrt again */
1560     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1561     assert(coef_sum < INT_MAX);
1562
1563     if(pict->pict_type == AV_PICTURE_TYPE_I){
1564         s->m.current_picture.mb_var_sum= coef_sum;
1565         s->m.current_picture.mc_mb_var_sum= 0;
1566     }else{
1567         s->m.current_picture.mc_mb_var_sum= coef_sum;
1568         s->m.current_picture.mb_var_sum= 0;
1569     }
1570
1571     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1572     if (pict->quality < 0)
1573         return INT_MIN;
1574     s->lambda= pict->quality * 3/2;
1575     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1576     s->qlog+= delta_qlog;
1577     return delta_qlog;
1578 }
1579
1580 static void calculate_visual_weight(SnowContext *s, Plane *p){
1581     int width = p->width;
1582     int height= p->height;
1583     int level, orientation, x, y;
1584
1585     for(level=0; level<s->spatial_decomposition_count; level++){
1586         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1587             SubBand *b= &p->band[level][orientation];
1588             IDWTELEM *ibuf= b->ibuf;
1589             int64_t error=0;
1590
1591             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1592             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1593             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1594             for(y=0; y<height; y++){
1595                 for(x=0; x<width; x++){
1596                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1597                     error += d*d;
1598                 }
1599             }
1600
1601             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1602         }
1603     }
1604 }
1605
1606 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1607                         const AVFrame *pict, int *got_packet)
1608 {
1609     SnowContext *s = avctx->priv_data;
1610     RangeCoder * const c= &s->c;
1611     AVFrame *pic = &s->new_picture;
1612     const int width= s->avctx->width;
1613     const int height= s->avctx->height;
1614     int level, orientation, plane_index, i, y, ret;
1615     uint8_t rc_header_bak[sizeof(s->header_state)];
1616     uint8_t rc_block_bak[sizeof(s->block_state)];
1617
1618     if (!pkt->data &&
1619         (ret = av_new_packet(pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0) {
1620         av_log(avctx, AV_LOG_ERROR, "Error getting output packet.\n");
1621         return ret;
1622     }
1623
1624     ff_init_range_encoder(c, pkt->data, pkt->size);
1625     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1626
1627     for(i=0; i<3; i++){
1628         int shift= !!i;
1629         for(y=0; y<(height>>shift); y++)
1630             memcpy(&s->input_picture.data[i][y * s->input_picture.linesize[i]],
1631                    &pict->data[i][y * pict->linesize[i]],
1632                    width>>shift);
1633     }
1634     s->new_picture = *pict;
1635
1636     s->m.picture_number= avctx->frame_number;
1637     if(avctx->flags&CODEC_FLAG_PASS2){
1638         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1639         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1640         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1641             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1642             if (pic->quality < 0)
1643                 return -1;
1644         }
1645     }else{
1646         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1647         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1648     }
1649
1650     if(s->pass1_rc && avctx->frame_number == 0)
1651         pic->quality = 2*FF_QP2LAMBDA;
1652     if (pic->quality) {
1653         s->qlog   = qscale2qlog(pic->quality);
1654         s->lambda = pic->quality * 3/2;
1655     }
1656     if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1657         s->qlog= LOSSLESS_QLOG;
1658         s->lambda = 0;
1659     }//else keep previous frame's qlog until after motion estimation
1660
1661     ff_snow_frame_start(s);
1662
1663     s->m.current_picture_ptr= &s->m.current_picture;
1664     s->m.last_picture.f.pts = s->m.current_picture.f.pts;
1665     s->m.current_picture.f.pts = pict->pts;
1666     if(pic->pict_type == AV_PICTURE_TYPE_P){
1667         int block_width = (width +15)>>4;
1668         int block_height= (height+15)>>4;
1669         int stride= s->current_picture.linesize[0];
1670
1671         assert(s->current_picture.data[0]);
1672         assert(s->last_picture[0].data[0]);
1673
1674         s->m.avctx= s->avctx;
1675         s->m.current_picture.f.data[0] = s->current_picture.data[0];
1676         s->m.   last_picture.f.data[0] = s->last_picture[0].data[0];
1677         s->m.    new_picture.f.data[0] = s->  input_picture.data[0];
1678         s->m.   last_picture_ptr= &s->m.   last_picture;
1679         s->m.linesize=
1680         s->m.   last_picture.f.linesize[0] =
1681         s->m.    new_picture.f.linesize[0] =
1682         s->m.current_picture.f.linesize[0] = stride;
1683         s->m.uvlinesize= s->current_picture.linesize[1];
1684         s->m.width = width;
1685         s->m.height= height;
1686         s->m.mb_width = block_width;
1687         s->m.mb_height= block_height;
1688         s->m.mb_stride=   s->m.mb_width+1;
1689         s->m.b8_stride= 2*s->m.mb_width+1;
1690         s->m.f_code=1;
1691         s->m.pict_type = pic->pict_type;
1692         s->m.me_method= s->avctx->me_method;
1693         s->m.me.scene_change_score=0;
1694         s->m.flags= s->avctx->flags;
1695         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1696         s->m.out_format= FMT_H263;
1697         s->m.unrestricted_mv= 1;
1698
1699         s->m.lambda = s->lambda;
1700         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1701         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1702
1703         s->m.dsp= s->dsp; //move
1704         ff_init_me(&s->m);
1705         s->dsp= s->m.dsp;
1706     }
1707
1708     if(s->pass1_rc){
1709         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1710         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1711     }
1712
1713 redo_frame:
1714
1715     if (pic->pict_type == AV_PICTURE_TYPE_I)
1716         s->spatial_decomposition_count= 5;
1717     else
1718         s->spatial_decomposition_count= 5;
1719
1720     s->m.pict_type = pic->pict_type;
1721     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1722
1723     ff_snow_common_init_after_header(avctx);
1724
1725     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1726         for(plane_index=0; plane_index<3; plane_index++){
1727             calculate_visual_weight(s, &s->plane[plane_index]);
1728         }
1729     }
1730
1731     encode_header(s);
1732     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1733     encode_blocks(s, 1);
1734     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1735
1736     for(plane_index=0; plane_index<3; plane_index++){
1737         Plane *p= &s->plane[plane_index];
1738         int w= p->width;
1739         int h= p->height;
1740         int x, y;
1741 //        int bits= put_bits_count(&s->c.pb);
1742
1743         if (!s->memc_only) {
1744             //FIXME optimize
1745             if(pict->data[plane_index]) //FIXME gray hack
1746                 for(y=0; y<h; y++){
1747                     for(x=0; x<w; x++){
1748                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1749                     }
1750                 }
1751             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1752
1753             if(   plane_index==0
1754                && pic->pict_type == AV_PICTURE_TYPE_P
1755                && !(avctx->flags&CODEC_FLAG_PASS2)
1756                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1757                 ff_init_range_encoder(c, pkt->data, pkt->size);
1758                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1759                 pic->pict_type= AV_PICTURE_TYPE_I;
1760                 s->keyframe=1;
1761                 s->current_picture.key_frame=1;
1762                 goto redo_frame;
1763             }
1764
1765             if(s->qlog == LOSSLESS_QLOG){
1766                 for(y=0; y<h; y++){
1767                     for(x=0; x<w; x++){
1768                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1769                     }
1770                 }
1771             }else{
1772                 for(y=0; y<h; y++){
1773                     for(x=0; x<w; x++){
1774                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1775                     }
1776                 }
1777             }
1778
1779             /*  if(QUANTIZE2)
1780                 dwt_quantize(s, p, s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type);
1781             else*/
1782                 ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1783
1784             if(s->pass1_rc && plane_index==0){
1785                 int delta_qlog = ratecontrol_1pass(s, pic);
1786                 if (delta_qlog <= INT_MIN)
1787                     return -1;
1788                 if(delta_qlog){
1789                     //reordering qlog in the bitstream would eliminate this reset
1790                     ff_init_range_encoder(c, pkt->data, pkt->size);
1791                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1792                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1793                     encode_header(s);
1794                     encode_blocks(s, 0);
1795                 }
1796             }
1797
1798             for(level=0; level<s->spatial_decomposition_count; level++){
1799                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1800                     SubBand *b= &p->band[level][orientation];
1801
1802                     if(!QUANTIZE2)
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                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1807                     assert(b->parent==NULL || b->parent->stride == b->stride*2);
1808                     if(orientation==0)
1809                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1810                 }
1811             }
1812
1813             for(level=0; level<s->spatial_decomposition_count; level++){
1814                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1815                     SubBand *b= &p->band[level][orientation];
1816
1817                     dequantize(s, b, b->ibuf, b->stride);
1818                 }
1819             }
1820
1821             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1822             if(s->qlog == LOSSLESS_QLOG){
1823                 for(y=0; y<h; y++){
1824                     for(x=0; x<w; x++){
1825                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1826                     }
1827                 }
1828             }
1829             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1830         }else{
1831             //ME/MC only
1832             if(pic->pict_type == AV_PICTURE_TYPE_I){
1833                 for(y=0; y<h; y++){
1834                     for(x=0; x<w; x++){
1835                         s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]=
1836                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1837                     }
1838                 }
1839             }else{
1840                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1841                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1842             }
1843         }
1844         if(s->avctx->flags&CODEC_FLAG_PSNR){
1845             int64_t error= 0;
1846
1847             if(pict->data[plane_index]) //FIXME gray hack
1848                 for(y=0; y<h; y++){
1849                     for(x=0; x<w; x++){
1850                         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];
1851                         error += d*d;
1852                     }
1853                 }
1854             s->avctx->error[plane_index] += error;
1855             s->current_picture.error[plane_index] = error;
1856         }
1857
1858     }
1859
1860     update_last_header_values(s);
1861
1862     ff_snow_release_buffer(avctx);
1863
1864     s->current_picture.coded_picture_number = avctx->frame_number;
1865     s->current_picture.pict_type = pict->pict_type;
1866     s->current_picture.quality = pict->quality;
1867     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1868     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1869     s->m.current_picture.f.display_picture_number =
1870     s->m.current_picture.f.coded_picture_number   = avctx->frame_number;
1871     s->m.current_picture.f.quality                = pic->quality;
1872     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1873     if(s->pass1_rc)
1874         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1875             return -1;
1876     if(avctx->flags&CODEC_FLAG_PASS1)
1877         ff_write_pass1_stats(&s->m);
1878     s->m.last_pict_type = s->m.pict_type;
1879     avctx->frame_bits = s->m.frame_bits;
1880     avctx->mv_bits = s->m.mv_bits;
1881     avctx->misc_bits = s->m.misc_bits;
1882     avctx->p_tex_bits = s->m.p_tex_bits;
1883
1884     emms_c();
1885
1886     pkt->size = ff_rac_terminate(c);
1887     if (avctx->coded_frame->key_frame)
1888         pkt->flags |= AV_PKT_FLAG_KEY;
1889     *got_packet = 1;
1890
1891     return 0;
1892 }
1893
1894 static av_cold int encode_end(AVCodecContext *avctx)
1895 {
1896     SnowContext *s = avctx->priv_data;
1897
1898     ff_snow_common_end(s);
1899     if (s->input_picture.data[0])
1900         avctx->release_buffer(avctx, &s->input_picture);
1901     av_free(avctx->stats_out);
1902
1903     return 0;
1904 }
1905
1906 #define OFFSET(x) offsetof(SnowContext, x)
1907 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1908 static const AVOption options[] = {
1909     { "memc_only",      "Only do ME/MC (I frames -> ref, P frame -> ME+MC).",   OFFSET(memc_only), AV_OPT_TYPE_INT, { 0 }, 0, 1, VE },
1910     { NULL },
1911 };
1912
1913 static const AVClass snowenc_class = {
1914     .class_name = "snow encoder",
1915     .item_name  = av_default_item_name,
1916     .option     = options,
1917     .version    = LIBAVUTIL_VERSION_INT,
1918 };
1919
1920 AVCodec ff_snow_encoder = {
1921     .name           = "snow",
1922     .type           = AVMEDIA_TYPE_VIDEO,
1923     .id             = CODEC_ID_SNOW,
1924     .priv_data_size = sizeof(SnowContext),
1925     .init           = encode_init,
1926     .encode2        = encode_frame,
1927     .close          = encode_end,
1928     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1929     .priv_class     = &snowenc_class,
1930 };