lavc: split snow and dirac DWTs
[ffmpeg.git] / libavcodec / snowenc.c
1 /*
2  * Copyright (C) 2004 Michael Niedermayer <michaelni@gmx.at>
3  *
4  * This file is part of FFmpeg.
5  *
6  * FFmpeg is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * FFmpeg is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with FFmpeg; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20
21 #include "libavutil/intmath.h"
22 #include "libavutil/log.h"
23 #include "libavutil/opt.h"
24 #include "avcodec.h"
25 #include "internal.h"
26 #include "dsputil.h"
27 #include "internal.h"
28 #include "snow_dwt.h"
29 #include "snow.h"
30
31 #include "rangecoder.h"
32 #include "mathops.h"
33
34 #include "mpegvideo.h"
35 #include "h263.h"
36
37 #undef NDEBUG
38 #include <assert.h>
39
40 #define QUANTIZE2 0
41
42 #if QUANTIZE2==1
43 #define Q2_STEP 8
44
45 static void find_sse(SnowContext *s, Plane *p, int *score, int score_stride, IDWTELEM *r0, IDWTELEM *r1, int level, int orientation){
46     SubBand *b= &p->band[level][orientation];
47     int x, y;
48     int xo=0;
49     int yo=0;
50     int step= 1 << (s->spatial_decomposition_count - level);
51
52     if(orientation&1)
53         xo= step>>1;
54     if(orientation&2)
55         yo= step>>1;
56
57     //FIXME bias for nonzero ?
58     //FIXME optimize
59     memset(score, 0, sizeof(*score)*score_stride*((p->height + Q2_STEP-1)/Q2_STEP));
60     for(y=0; y<p->height; y++){
61         for(x=0; x<p->width; x++){
62             int sx= (x-xo + step/2) / step / Q2_STEP;
63             int sy= (y-yo + step/2) / step / Q2_STEP;
64             int v= r0[x + y*p->width] - r1[x + y*p->width];
65             assert(sx>=0 && sy>=0 && sx < score_stride);
66             v= ((v+8)>>4)<<4;
67             score[sx + sy*score_stride] += v*v;
68             assert(score[sx + sy*score_stride] >= 0);
69         }
70     }
71 }
72
73 static void dequantize_all(SnowContext *s, Plane *p, IDWTELEM *buffer, int width, int height){
74     int level, orientation;
75
76     for(level=0; level<s->spatial_decomposition_count; level++){
77         for(orientation=level ? 1 : 0; orientation<4; orientation++){
78             SubBand *b= &p->band[level][orientation];
79             IDWTELEM *dst= buffer + (b->ibuf - s->spatial_idwt_buffer);
80
81             dequantize(s, b, dst, b->stride);
82         }
83     }
84 }
85
86 static void dwt_quantize(SnowContext *s, Plane *p, DWTELEM *buffer, int width, int height, int stride, int type){
87     int level, orientation, ys, xs, x, y, pass;
88     IDWTELEM best_dequant[height * stride];
89     IDWTELEM idwt2_buffer[height * stride];
90     const int score_stride= (width + 10)/Q2_STEP;
91     int best_score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
92     int score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
93     int threshold= (s->m.lambda * s->m.lambda) >> 6;
94
95     //FIXME pass the copy cleanly ?
96
97 //    memcpy(dwt_buffer, buffer, height * stride * sizeof(DWTELEM));
98     ff_spatial_dwt(buffer, s->temp_dwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
99
100     for(level=0; level<s->spatial_decomposition_count; level++){
101         for(orientation=level ? 1 : 0; orientation<4; orientation++){
102             SubBand *b= &p->band[level][orientation];
103             IDWTELEM *dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
104              DWTELEM *src=       buffer + (b-> buf - s->spatial_dwt_buffer);
105             assert(src == b->buf); // code does not depend on this but it is true currently
106
107             quantize(s, b, dst, src, b->stride, s->qbias);
108         }
109     }
110     for(pass=0; pass<1; pass++){
111         if(s->qbias == 0) //keyframe
112             continue;
113         for(level=0; level<s->spatial_decomposition_count; level++){
114             for(orientation=level ? 1 : 0; orientation<4; orientation++){
115                 SubBand *b= &p->band[level][orientation];
116                 IDWTELEM *dst= idwt2_buffer + (b->ibuf - s->spatial_idwt_buffer);
117                 IDWTELEM *best_dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
118
119                 for(ys= 0; ys<Q2_STEP; ys++){
120                     for(xs= 0; xs<Q2_STEP; xs++){
121                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
122                         dequantize_all(s, p, idwt2_buffer, width, height);
123                         ff_spatial_idwt(idwt2_buffer, s->temp_idwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
124                         find_sse(s, p, best_score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
125                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
126                         for(y=ys; y<b->height; y+= Q2_STEP){
127                             for(x=xs; x<b->width; x+= Q2_STEP){
128                                 if(dst[x + y*b->stride]<0) dst[x + y*b->stride]++;
129                                 if(dst[x + y*b->stride]>0) dst[x + y*b->stride]--;
130                                 //FIXME try more than just --
131                             }
132                         }
133                         dequantize_all(s, p, idwt2_buffer, width, height);
134                         ff_spatial_idwt(idwt2_buffer, s->temp_idwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
135                         find_sse(s, p, score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
136                         for(y=ys; y<b->height; y+= Q2_STEP){
137                             for(x=xs; x<b->width; x+= Q2_STEP){
138                                 int score_idx= x/Q2_STEP + (y/Q2_STEP)*score_stride;
139                                 if(score[score_idx] <= best_score[score_idx] + threshold){
140                                     best_score[score_idx]= score[score_idx];
141                                     if(best_dst[x + y*b->stride]<0) best_dst[x + y*b->stride]++;
142                                     if(best_dst[x + y*b->stride]>0) best_dst[x + y*b->stride]--;
143                                     //FIXME copy instead
144                                 }
145                             }
146                         }
147                     }
148                 }
149             }
150         }
151     }
152     memcpy(s->spatial_idwt_buffer, best_dequant, height * stride * sizeof(IDWTELEM)); //FIXME work with that directly instead of copy at the end
153 }
154
155 #endif /* QUANTIZE2==1 */
156
157 static av_cold int encode_init(AVCodecContext *avctx)
158 {
159     SnowContext *s = avctx->priv_data;
160     int plane_index, ret;
161
162     if(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL){
163         av_log(avctx, AV_LOG_ERROR, "This codec is under development, files encoded with it may not be decodable with future versions!!!\n"
164                "Use vstrict=-2 / -strict -2 to use it anyway.\n");
165         return -1;
166     }
167
168     if(avctx->prediction_method == DWT_97
169        && (avctx->flags & CODEC_FLAG_QSCALE)
170        && avctx->global_quality == 0){
171         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
172         return -1;
173     }
174
175     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
176
177     s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
178     s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
179
180     for(plane_index=0; plane_index<3; plane_index++){
181         s->plane[plane_index].diag_mc= 1;
182         s->plane[plane_index].htaps= 6;
183         s->plane[plane_index].hcoeff[0]=  40;
184         s->plane[plane_index].hcoeff[1]= -10;
185         s->plane[plane_index].hcoeff[2]=   2;
186         s->plane[plane_index].fast_mc= 1;
187     }
188
189     if ((ret = ff_snow_common_init(avctx)) < 0) {
190         ff_snow_common_end(avctx->priv_data);
191         return ret;
192     }
193     ff_snow_alloc_blocks(s);
194
195     s->version=0;
196
197     s->m.avctx   = avctx;
198     s->m.flags   = avctx->flags;
199     s->m.bit_rate= avctx->bit_rate;
200
201     s->m.me.temp      =
202     s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
203     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
204     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
205     s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
206     ff_h263_encode_init(&s->m); //mv_penalty
207
208     s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
209
210     if(avctx->flags&CODEC_FLAG_PASS1){
211         if(!avctx->stats_out)
212             avctx->stats_out = av_mallocz(256);
213     }
214     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
215         if(ff_rate_control_init(&s->m) < 0)
216             return -1;
217     }
218     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
219
220     avctx->coded_frame= &s->current_picture;
221     switch(avctx->pix_fmt){
222     case AV_PIX_FMT_YUV444P:
223 //    case AV_PIX_FMT_YUV422P:
224     case AV_PIX_FMT_YUV420P:
225 //     case AV_PIX_FMT_GRAY8:
226 //    case AV_PIX_FMT_YUV411P:
227     case AV_PIX_FMT_YUV410P:
228         s->colorspace_type= 0;
229         break;
230 /*    case AV_PIX_FMT_RGB32:
231         s->colorspace= 1;
232         break;*/
233     default:
234         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
235         return -1;
236     }
237     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
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     if ((ret = ff_get_buffer(s->avctx, &s->input_picture)) < 0)
243         return ret;
244
245     if(s->avctx->me_method == ME_ITER){
246         int i;
247         int size= s->b_width * s->b_height << 2*s->block_max_depth;
248         for(i=0; i<s->max_ref_frames; i++){
249             s->ref_mvs[i]= av_mallocz(size*sizeof(int16_t[2]));
250             s->ref_scores[i]= av_mallocz(size*sizeof(uint32_t));
251         }
252     }
253
254     return 0;
255 }
256
257 //near copy & paste from dsputil, FIXME
258 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
259 {
260     int s, i, j;
261
262     s = 0;
263     for (i = 0; i < h; i++) {
264         for (j = 0; j < w; j++) {
265             s += pix[0];
266             pix ++;
267         }
268         pix += line_size - w;
269     }
270     return s;
271 }
272
273 //near copy & paste from dsputil, FIXME
274 static int pix_norm1(uint8_t * pix, int line_size, int w)
275 {
276     int s, i, j;
277     uint32_t *sq = ff_squareTbl + 256;
278
279     s = 0;
280     for (i = 0; i < w; i++) {
281         for (j = 0; j < w; j ++) {
282             s += sq[pix[0]];
283             pix ++;
284         }
285         pix += line_size - w;
286     }
287     return s;
288 }
289
290 //FIXME copy&paste
291 #define P_LEFT P[1]
292 #define P_TOP P[2]
293 #define P_TOPRIGHT P[3]
294 #define P_MEDIAN P[4]
295 #define P_MV1 P[9]
296 #define FLAG_QPEL   1 //must be 1
297
298 static int encode_q_branch(SnowContext *s, int level, int x, int y){
299     uint8_t p_buffer[1024];
300     uint8_t i_buffer[1024];
301     uint8_t p_state[sizeof(s->block_state)];
302     uint8_t i_state[sizeof(s->block_state)];
303     RangeCoder pc, ic;
304     uint8_t *pbbak= s->c.bytestream;
305     uint8_t *pbbak_start= s->c.bytestream_start;
306     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
307     const int w= s->b_width  << s->block_max_depth;
308     const int h= s->b_height << s->block_max_depth;
309     const int rem_depth= s->block_max_depth - level;
310     const int index= (x + y*w) << rem_depth;
311     const int block_w= 1<<(LOG2_MB_SIZE - level);
312     int trx= (x+1)<<rem_depth;
313     int try= (y+1)<<rem_depth;
314     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
315     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
316     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
317     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
318     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
319     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
320     int pl = left->color[0];
321     int pcb= left->color[1];
322     int pcr= left->color[2];
323     int pmx, pmy;
324     int mx=0, my=0;
325     int l,cr,cb;
326     const int stride= s->current_picture.linesize[0];
327     const int uvstride= s->current_picture.linesize[1];
328     uint8_t *current_data[3]= { s->input_picture.data[0] + (x + y*  stride)*block_w,
329                                 s->input_picture.data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
330                                 s->input_picture.data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
331     int P[10][2];
332     int16_t last_mv[3][2];
333     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
334     const int shift= 1+qpel;
335     MotionEstContext *c= &s->m.me;
336     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
337     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
338     int my_context= av_log2(2*FFABS(left->my - top->my));
339     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
340     int ref, best_ref, ref_score, ref_mx, ref_my;
341
342     assert(sizeof(s->block_state) >= 256);
343     if(s->keyframe){
344         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
345         return 0;
346     }
347
348 //    clip predictors / edge ?
349
350     P_LEFT[0]= left->mx;
351     P_LEFT[1]= left->my;
352     P_TOP [0]= top->mx;
353     P_TOP [1]= top->my;
354     P_TOPRIGHT[0]= tr->mx;
355     P_TOPRIGHT[1]= tr->my;
356
357     last_mv[0][0]= s->block[index].mx;
358     last_mv[0][1]= s->block[index].my;
359     last_mv[1][0]= right->mx;
360     last_mv[1][1]= right->my;
361     last_mv[2][0]= bottom->mx;
362     last_mv[2][1]= bottom->my;
363
364     s->m.mb_stride=2;
365     s->m.mb_x=
366     s->m.mb_y= 0;
367     c->skip= 0;
368
369     assert(c->  stride ==   stride);
370     assert(c->uvstride == uvstride);
371
372     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
373     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
374     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
375     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
376
377     c->xmin = - x*block_w - 16+3;
378     c->ymin = - y*block_w - 16+3;
379     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
380     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
381
382     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
383     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
384     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
385     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
386     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
387     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
388     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
389
390     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
391     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
392
393     if (!y) {
394         c->pred_x= P_LEFT[0];
395         c->pred_y= P_LEFT[1];
396     } else {
397         c->pred_x = P_MEDIAN[0];
398         c->pred_y = P_MEDIAN[1];
399     }
400
401     score= INT_MAX;
402     best_ref= 0;
403     for(ref=0; ref<s->ref_frames; ref++){
404         init_ref(c, current_data, s->last_picture[ref].data, NULL, block_w*x, block_w*y, 0);
405
406         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
407                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
408
409         assert(ref_mx >= c->xmin);
410         assert(ref_mx <= c->xmax);
411         assert(ref_my >= c->ymin);
412         assert(ref_my <= c->ymax);
413
414         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
415         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
416         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
417         if(s->ref_mvs[ref]){
418             s->ref_mvs[ref][index][0]= ref_mx;
419             s->ref_mvs[ref][index][1]= ref_my;
420             s->ref_scores[ref][index]= ref_score;
421         }
422         if(score > ref_score){
423             score= ref_score;
424             best_ref= ref;
425             mx= ref_mx;
426             my= ref_my;
427         }
428     }
429     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
430
431   //  subpel search
432     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
433     pc= s->c;
434     pc.bytestream_start=
435     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
436     memcpy(p_state, s->block_state, sizeof(s->block_state));
437
438     if(level!=s->block_max_depth)
439         put_rac(&pc, &p_state[4 + s_context], 1);
440     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
441     if(s->ref_frames > 1)
442         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
443     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
444     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
445     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
446     p_len= pc.bytestream - pc.bytestream_start;
447     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
448
449     block_s= block_w*block_w;
450     sum = pix_sum(current_data[0], stride, block_w, block_w);
451     l= (sum + block_s/2)/block_s;
452     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
453
454     block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
455     sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
456     cb= (sum + block_s/2)/block_s;
457 //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
458     sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
459     cr= (sum + block_s/2)/block_s;
460 //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
461
462     ic= s->c;
463     ic.bytestream_start=
464     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
465     memcpy(i_state, s->block_state, sizeof(s->block_state));
466     if(level!=s->block_max_depth)
467         put_rac(&ic, &i_state[4 + s_context], 1);
468     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
469     put_symbol(&ic, &i_state[32],  l-pl , 1);
470     put_symbol(&ic, &i_state[64], cb-pcb, 1);
471     put_symbol(&ic, &i_state[96], cr-pcr, 1);
472     i_len= ic.bytestream - ic.bytestream_start;
473     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
474
475 //    assert(score==256*256*256*64-1);
476     assert(iscore < 255*255*256 + s->lambda2*10);
477     assert(iscore >= 0);
478     assert(l>=0 && l<=255);
479     assert(pl>=0 && pl<=255);
480
481     if(level==0){
482         int varc= iscore >> 8;
483         int vard= score >> 8;
484         if (vard <= 64 || vard < varc)
485             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
486         else
487             c->scene_change_score+= s->m.qscale;
488     }
489
490     if(level!=s->block_max_depth){
491         put_rac(&s->c, &s->block_state[4 + s_context], 0);
492         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
493         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
494         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
495         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
496         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
497
498         if(score2 < score && score2 < iscore)
499             return score2;
500     }
501
502     if(iscore < score){
503         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
504         memcpy(pbbak, i_buffer, i_len);
505         s->c= ic;
506         s->c.bytestream_start= pbbak_start;
507         s->c.bytestream= pbbak + i_len;
508         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
509         memcpy(s->block_state, i_state, sizeof(s->block_state));
510         return iscore;
511     }else{
512         memcpy(pbbak, p_buffer, p_len);
513         s->c= pc;
514         s->c.bytestream_start= pbbak_start;
515         s->c.bytestream= pbbak + p_len;
516         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
517         memcpy(s->block_state, p_state, sizeof(s->block_state));
518         return score;
519     }
520 }
521
522 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
523     const int w= s->b_width  << s->block_max_depth;
524     const int rem_depth= s->block_max_depth - level;
525     const int index= (x + y*w) << rem_depth;
526     int trx= (x+1)<<rem_depth;
527     BlockNode *b= &s->block[index];
528     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
529     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
530     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
531     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
532     int pl = left->color[0];
533     int pcb= left->color[1];
534     int pcr= left->color[2];
535     int pmx, pmy;
536     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
537     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
538     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
539     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
540
541     if(s->keyframe){
542         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
543         return;
544     }
545
546     if(level!=s->block_max_depth){
547         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
548             put_rac(&s->c, &s->block_state[4 + s_context], 1);
549         }else{
550             put_rac(&s->c, &s->block_state[4 + s_context], 0);
551             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
552             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
553             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
554             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
555             return;
556         }
557     }
558     if(b->type & BLOCK_INTRA){
559         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
560         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
561         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
562         put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
563         put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
564         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
565     }else{
566         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
567         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
568         if(s->ref_frames > 1)
569             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
570         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
571         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
572         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
573     }
574 }
575
576 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
577     int i, x2, y2;
578     Plane *p= &s->plane[plane_index];
579     const int block_size = MB_SIZE >> s->block_max_depth;
580     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
581     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
582     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
583     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
584     const int ref_stride= s->current_picture.linesize[plane_index];
585     uint8_t *src= s-> input_picture.data[plane_index];
586     IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
587     const int b_stride = s->b_width << s->block_max_depth;
588     const int w= p->width;
589     const int h= p->height;
590     int index= mb_x + mb_y*b_stride;
591     BlockNode *b= &s->block[index];
592     BlockNode backup= *b;
593     int ab=0;
594     int aa=0;
595
596     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
597
598     b->type|= BLOCK_INTRA;
599     b->color[plane_index]= 0;
600     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
601
602     for(i=0; i<4; i++){
603         int mb_x2= mb_x + (i &1) - 1;
604         int mb_y2= mb_y + (i>>1) - 1;
605         int x= block_w*mb_x2 + block_w/2;
606         int y= block_h*mb_y2 + block_h/2;
607
608         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
609                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
610
611         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
612             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
613                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
614                 int obmc_v= obmc[index];
615                 int d;
616                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
617                 if(x<0) obmc_v += obmc[index + block_w];
618                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
619                 if(x+block_w>w) obmc_v += obmc[index - block_w];
620                 //FIXME precalculate this or simplify it somehow else
621
622                 d = -dst[index] + (1<<(FRAC_BITS-1));
623                 dst[index] = d;
624                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
625                 aa += obmc_v * obmc_v; //FIXME precalculate this
626             }
627         }
628     }
629     *b= backup;
630
631     return av_clip( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa), 0, 255); //FIXME we should not need clipping
632 }
633
634 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
635     const int b_stride = s->b_width << s->block_max_depth;
636     const int b_height = s->b_height<< s->block_max_depth;
637     int index= x + y*b_stride;
638     const BlockNode *b     = &s->block[index];
639     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
640     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
641     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
642     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
643     int dmx, dmy;
644 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
645 //  int my_context= av_log2(2*FFABS(left->my - top->my));
646
647     if(x<0 || x>=b_stride || y>=b_height)
648         return 0;
649 /*
650 1            0      0
651 01X          1-2    1
652 001XX        3-6    2-3
653 0001XXX      7-14   4-7
654 00001XXXX   15-30   8-15
655 */
656 //FIXME try accurate rate
657 //FIXME intra and inter predictors if surrounding blocks are not the same type
658     if(b->type & BLOCK_INTRA){
659         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
660                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
661                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
662     }else{
663         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
664         dmx-= b->mx;
665         dmy-= b->my;
666         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
667                     + av_log2(2*FFABS(dmy))
668                     + av_log2(2*b->ref));
669     }
670 }
671
672 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
673     Plane *p= &s->plane[plane_index];
674     const int block_size = MB_SIZE >> s->block_max_depth;
675     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
676     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
677     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
678     const int ref_stride= s->current_picture.linesize[plane_index];
679     uint8_t *dst= s->current_picture.data[plane_index];
680     uint8_t *src= s->  input_picture.data[plane_index];
681     IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
682     uint8_t *cur = s->scratchbuf;
683     uint8_t *tmp = s->emu_edge_buffer;
684     const int b_stride = s->b_width << s->block_max_depth;
685     const int b_height = s->b_height<< s->block_max_depth;
686     const int w= p->width;
687     const int h= p->height;
688     int distortion;
689     int rate= 0;
690     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
691     int sx= block_w*mb_x - block_w/2;
692     int sy= block_h*mb_y - block_h/2;
693     int x0= FFMAX(0,-sx);
694     int y0= FFMAX(0,-sy);
695     int x1= FFMIN(block_w*2, w-sx);
696     int y1= FFMIN(block_h*2, h-sy);
697     int i,x,y;
698
699     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
700
701     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);
702
703     for(y=y0; y<y1; y++){
704         const uint8_t *obmc1= obmc_edged[y];
705         const IDWTELEM *pred1 = pred + y*obmc_stride;
706         uint8_t *cur1 = cur + y*ref_stride;
707         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
708         for(x=x0; x<x1; x++){
709 #if FRAC_BITS >= LOG2_OBMC_MAX
710             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
711 #else
712             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
713 #endif
714             v = (v + pred1[x]) >> FRAC_BITS;
715             if(v&(~255)) v= ~(v>>31);
716             dst1[x] = v;
717         }
718     }
719
720     /* copy the regions where obmc[] = (uint8_t)256 */
721     if(LOG2_OBMC_MAX == 8
722         && (mb_x == 0 || mb_x == b_stride-1)
723         && (mb_y == 0 || mb_y == b_height-1)){
724         if(mb_x == 0)
725             x1 = block_w;
726         else
727             x0 = block_w;
728         if(mb_y == 0)
729             y1 = block_h;
730         else
731             y0 = block_h;
732         for(y=y0; y<y1; y++)
733             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
734     }
735
736     if(block_w==16){
737         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
738         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
739         /* FIXME cmps overlap but do not cover the wavelet's whole support.
740          * So improving the score of one block is not strictly guaranteed
741          * to improve the score of the whole frame, thus iterative motion
742          * estimation does not always converge. */
743         if(s->avctx->me_cmp == FF_CMP_W97)
744             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
745         else if(s->avctx->me_cmp == FF_CMP_W53)
746             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
747         else{
748             distortion = 0;
749             for(i=0; i<4; i++){
750                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
751                 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
752             }
753         }
754     }else{
755         assert(block_w==8);
756         distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
757     }
758
759     if(plane_index==0){
760         for(i=0; i<4; i++){
761 /* ..RRr
762  * .RXx.
763  * rxx..
764  */
765             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
766         }
767         if(mb_x == b_stride-2)
768             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
769     }
770     return distortion + rate*penalty_factor;
771 }
772
773 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
774     int i, y2;
775     Plane *p= &s->plane[plane_index];
776     const int block_size = MB_SIZE >> s->block_max_depth;
777     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
778     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
779     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
780     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
781     const int ref_stride= s->current_picture.linesize[plane_index];
782     uint8_t *dst= s->current_picture.data[plane_index];
783     uint8_t *src= s-> input_picture.data[plane_index];
784     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
785     // const has only been removed from zero_dst to suppress a warning
786     static IDWTELEM zero_dst[4096]; //FIXME
787     const int b_stride = s->b_width << s->block_max_depth;
788     const int w= p->width;
789     const int h= p->height;
790     int distortion= 0;
791     int rate= 0;
792     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
793
794     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
795
796     for(i=0; i<9; i++){
797         int mb_x2= mb_x + (i%3) - 1;
798         int mb_y2= mb_y + (i/3) - 1;
799         int x= block_w*mb_x2 + block_w/2;
800         int y= block_h*mb_y2 + block_h/2;
801
802         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
803                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
804
805         //FIXME find a cleaner/simpler way to skip the outside stuff
806         for(y2= y; y2<0; y2++)
807             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
808         for(y2= h; y2<y+block_h; y2++)
809             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
810         if(x<0){
811             for(y2= y; y2<y+block_h; y2++)
812                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
813         }
814         if(x+block_w > w){
815             for(y2= y; y2<y+block_h; y2++)
816                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
817         }
818
819         assert(block_w== 8 || block_w==16);
820         distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
821     }
822
823     if(plane_index==0){
824         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
825         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
826
827 /* ..RRRr
828  * .RXXx.
829  * .RXXx.
830  * rxxx.
831  */
832         if(merged)
833             rate = get_block_bits(s, mb_x, mb_y, 2);
834         for(i=merged?4:0; i<9; i++){
835             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
836             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
837         }
838     }
839     return distortion + rate*penalty_factor;
840 }
841
842 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
843     const int w= b->width;
844     const int h= b->height;
845     int x, y;
846
847     if(1){
848         int run=0;
849         int *runs = s->run_buffer;
850         int run_index=0;
851         int max_index;
852
853         for(y=0; y<h; y++){
854             for(x=0; x<w; x++){
855                 int v, p=0;
856                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
857                 v= src[x + y*stride];
858
859                 if(y){
860                     t= src[x + (y-1)*stride];
861                     if(x){
862                         lt= src[x - 1 + (y-1)*stride];
863                     }
864                     if(x + 1 < w){
865                         rt= src[x + 1 + (y-1)*stride];
866                     }
867                 }
868                 if(x){
869                     l= src[x - 1 + y*stride];
870                     /*if(x > 1){
871                         if(orientation==1) ll= src[y + (x-2)*stride];
872                         else               ll= src[x - 2 + y*stride];
873                     }*/
874                 }
875                 if(parent){
876                     int px= x>>1;
877                     int py= y>>1;
878                     if(px<b->parent->width && py<b->parent->height)
879                         p= parent[px + py*2*stride];
880                 }
881                 if(!(/*ll|*/l|lt|t|rt|p)){
882                     if(v){
883                         runs[run_index++]= run;
884                         run=0;
885                     }else{
886                         run++;
887                     }
888                 }
889             }
890         }
891         max_index= run_index;
892         runs[run_index++]= run;
893         run_index=0;
894         run= runs[run_index++];
895
896         put_symbol2(&s->c, b->state[30], max_index, 0);
897         if(run_index <= max_index)
898             put_symbol2(&s->c, b->state[1], run, 3);
899
900         for(y=0; y<h; y++){
901             if(s->c.bytestream_end - s->c.bytestream < w*40){
902                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
903                 return -1;
904             }
905             for(x=0; x<w; x++){
906                 int v, p=0;
907                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
908                 v= src[x + y*stride];
909
910                 if(y){
911                     t= src[x + (y-1)*stride];
912                     if(x){
913                         lt= src[x - 1 + (y-1)*stride];
914                     }
915                     if(x + 1 < w){
916                         rt= src[x + 1 + (y-1)*stride];
917                     }
918                 }
919                 if(x){
920                     l= src[x - 1 + y*stride];
921                     /*if(x > 1){
922                         if(orientation==1) ll= src[y + (x-2)*stride];
923                         else               ll= src[x - 2 + y*stride];
924                     }*/
925                 }
926                 if(parent){
927                     int px= x>>1;
928                     int py= y>>1;
929                     if(px<b->parent->width && py<b->parent->height)
930                         p= parent[px + py*2*stride];
931                 }
932                 if(/*ll|*/l|lt|t|rt|p){
933                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
934
935                     put_rac(&s->c, &b->state[0][context], !!v);
936                 }else{
937                     if(!run){
938                         run= runs[run_index++];
939
940                         if(run_index <= max_index)
941                             put_symbol2(&s->c, b->state[1], run, 3);
942                         assert(v);
943                     }else{
944                         run--;
945                         assert(!v);
946                     }
947                 }
948                 if(v){
949                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
950                     int l2= 2*FFABS(l) + (l<0);
951                     int t2= 2*FFABS(t) + (t<0);
952
953                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
954                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
955                 }
956             }
957         }
958     }
959     return 0;
960 }
961
962 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
963 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
964 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
965     return encode_subband_c0run(s, b, src, parent, stride, orientation);
966 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
967 }
968
969 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){
970     const int b_stride= s->b_width << s->block_max_depth;
971     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
972     BlockNode backup= *block;
973     unsigned value;
974     int rd, index;
975
976     assert(mb_x>=0 && mb_y>=0);
977     assert(mb_x<b_stride);
978
979     if(intra){
980         block->color[0] = p[0];
981         block->color[1] = p[1];
982         block->color[2] = p[2];
983         block->type |= BLOCK_INTRA;
984     }else{
985         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
986         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
987         if(s->me_cache[index] == value)
988             return 0;
989         s->me_cache[index]= value;
990
991         block->mx= p[0];
992         block->my= p[1];
993         block->type &= ~BLOCK_INTRA;
994     }
995
996     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
997
998 //FIXME chroma
999     if(rd < *best_rd){
1000         *best_rd= rd;
1001         return 1;
1002     }else{
1003         *block= backup;
1004         return 0;
1005     }
1006 }
1007
1008 /* special case for int[2] args we discard afterwards,
1009  * fixes compilation problem with gcc 2.95 */
1010 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){
1011     int p[2] = {p0, p1};
1012     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
1013 }
1014
1015 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){
1016     const int b_stride= s->b_width << s->block_max_depth;
1017     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
1018     BlockNode backup[4];
1019     unsigned value;
1020     int rd, index;
1021
1022     /* We don't initialize backup[] during variable declaration, because
1023      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
1024      * 'int16_t'". */
1025     backup[0] = block[0];
1026     backup[1] = block[1];
1027     backup[2] = block[b_stride];
1028     backup[3] = block[b_stride + 1];
1029
1030     assert(mb_x>=0 && mb_y>=0);
1031     assert(mb_x<b_stride);
1032     assert(((mb_x|mb_y)&1) == 0);
1033
1034     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
1035     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
1036     if(s->me_cache[index] == value)
1037         return 0;
1038     s->me_cache[index]= value;
1039
1040     block->mx= p0;
1041     block->my= p1;
1042     block->ref= ref;
1043     block->type &= ~BLOCK_INTRA;
1044     block[1]= block[b_stride]= block[b_stride+1]= *block;
1045
1046     rd= get_4block_rd(s, mb_x, mb_y, 0);
1047
1048 //FIXME chroma
1049     if(rd < *best_rd){
1050         *best_rd= rd;
1051         return 1;
1052     }else{
1053         block[0]= backup[0];
1054         block[1]= backup[1];
1055         block[b_stride]= backup[2];
1056         block[b_stride+1]= backup[3];
1057         return 0;
1058     }
1059 }
1060
1061 static void iterative_me(SnowContext *s){
1062     int pass, mb_x, mb_y;
1063     const int b_width = s->b_width  << s->block_max_depth;
1064     const int b_height= s->b_height << s->block_max_depth;
1065     const int b_stride= b_width;
1066     int color[3];
1067
1068     {
1069         RangeCoder r = s->c;
1070         uint8_t state[sizeof(s->block_state)];
1071         memcpy(state, s->block_state, sizeof(s->block_state));
1072         for(mb_y= 0; mb_y<s->b_height; mb_y++)
1073             for(mb_x= 0; mb_x<s->b_width; mb_x++)
1074                 encode_q_branch(s, 0, mb_x, mb_y);
1075         s->c = r;
1076         memcpy(s->block_state, state, sizeof(s->block_state));
1077     }
1078
1079     for(pass=0; pass<25; pass++){
1080         int change= 0;
1081
1082         for(mb_y= 0; mb_y<b_height; mb_y++){
1083             for(mb_x= 0; mb_x<b_width; mb_x++){
1084                 int dia_change, i, j, ref;
1085                 int best_rd= INT_MAX, ref_rd;
1086                 BlockNode backup, ref_b;
1087                 const int index= mb_x + mb_y * b_stride;
1088                 BlockNode *block= &s->block[index];
1089                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1090                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1091                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1092                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1093                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1094                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1095                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1096                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1097                 const int b_w= (MB_SIZE >> s->block_max_depth);
1098                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1099
1100                 if(pass && (block->type & BLOCK_OPT))
1101                     continue;
1102                 block->type |= BLOCK_OPT;
1103
1104                 backup= *block;
1105
1106                 if(!s->me_cache_generation)
1107                     memset(s->me_cache, 0, sizeof(s->me_cache));
1108                 s->me_cache_generation += 1<<22;
1109
1110                 //FIXME precalculate
1111                 {
1112                     int x, y;
1113                     for (y = 0; y < b_w * 2; y++)
1114                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1115                     if(mb_x==0)
1116                         for(y=0; y<b_w*2; y++)
1117                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1118                     if(mb_x==b_stride-1)
1119                         for(y=0; y<b_w*2; y++)
1120                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1121                     if(mb_y==0){
1122                         for(x=0; x<b_w*2; x++)
1123                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1124                         for(y=1; y<b_w; y++)
1125                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1126                     }
1127                     if(mb_y==b_height-1){
1128                         for(x=0; x<b_w*2; x++)
1129                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1130                         for(y=b_w; y<b_w*2-1; y++)
1131                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1132                     }
1133                 }
1134
1135                 //skip stuff outside the picture
1136                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1137                     uint8_t *src= s->  input_picture.data[0];
1138                     uint8_t *dst= s->current_picture.data[0];
1139                     const int stride= s->current_picture.linesize[0];
1140                     const int block_w= MB_SIZE >> s->block_max_depth;
1141                     const int block_h= MB_SIZE >> s->block_max_depth;
1142                     const int sx= block_w*mb_x - block_w/2;
1143                     const int sy= block_h*mb_y - block_h/2;
1144                     const int w= s->plane[0].width;
1145                     const int h= s->plane[0].height;
1146                     int y;
1147
1148                     for(y=sy; y<0; y++)
1149                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1150                     for(y=h; y<sy+block_h*2; y++)
1151                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1152                     if(sx<0){
1153                         for(y=sy; y<sy+block_h*2; y++)
1154                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1155                     }
1156                     if(sx+block_w*2 > w){
1157                         for(y=sy; y<sy+block_h*2; y++)
1158                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1159                     }
1160                 }
1161
1162                 // intra(black) = neighbors' contribution to the current block
1163                 for(i=0; i<3; i++)
1164                     color[i]= get_dc(s, mb_x, mb_y, i);
1165
1166                 // get previous score (cannot be cached due to OBMC)
1167                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1168                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1169                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1170                 }else
1171                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1172
1173                 ref_b= *block;
1174                 ref_rd= best_rd;
1175                 for(ref=0; ref < s->ref_frames; ref++){
1176                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1177                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1178                         continue;
1179                     block->ref= ref;
1180                     best_rd= INT_MAX;
1181
1182                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1183                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1184                     if(tb)
1185                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1186                     if(lb)
1187                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1188                     if(rb)
1189                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1190                     if(bb)
1191                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1192
1193                     /* fullpel ME */
1194                     //FIXME avoid subpel interpolation / round to nearest integer
1195                     do{
1196                         dia_change=0;
1197                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1198                             for(j=0; j<i; j++){
1199                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1200                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1201                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1202                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1203                             }
1204                         }
1205                     }while(dia_change);
1206                     /* subpel ME */
1207                     do{
1208                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1209                         dia_change=0;
1210                         for(i=0; i<8; i++)
1211                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1212                     }while(dia_change);
1213                     //FIXME or try the standard 2 pass qpel or similar
1214
1215                     mvr[0][0]= block->mx;
1216                     mvr[0][1]= block->my;
1217                     if(ref_rd > best_rd){
1218                         ref_rd= best_rd;
1219                         ref_b= *block;
1220                     }
1221                 }
1222                 best_rd= ref_rd;
1223                 *block= ref_b;
1224                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1225                 //FIXME RD style color selection
1226                 if(!same_block(block, &backup)){
1227                     if(tb ) tb ->type &= ~BLOCK_OPT;
1228                     if(lb ) lb ->type &= ~BLOCK_OPT;
1229                     if(rb ) rb ->type &= ~BLOCK_OPT;
1230                     if(bb ) bb ->type &= ~BLOCK_OPT;
1231                     if(tlb) tlb->type &= ~BLOCK_OPT;
1232                     if(trb) trb->type &= ~BLOCK_OPT;
1233                     if(blb) blb->type &= ~BLOCK_OPT;
1234                     if(brb) brb->type &= ~BLOCK_OPT;
1235                     change ++;
1236                 }
1237             }
1238         }
1239         av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1240         if(!change)
1241             break;
1242     }
1243
1244     if(s->block_max_depth == 1){
1245         int change= 0;
1246         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1247             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1248                 int i;
1249                 int best_rd, init_rd;
1250                 const int index= mb_x + mb_y * b_stride;
1251                 BlockNode *b[4];
1252
1253                 b[0]= &s->block[index];
1254                 b[1]= b[0]+1;
1255                 b[2]= b[0]+b_stride;
1256                 b[3]= b[2]+1;
1257                 if(same_block(b[0], b[1]) &&
1258                    same_block(b[0], b[2]) &&
1259                    same_block(b[0], b[3]))
1260                     continue;
1261
1262                 if(!s->me_cache_generation)
1263                     memset(s->me_cache, 0, sizeof(s->me_cache));
1264                 s->me_cache_generation += 1<<22;
1265
1266                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1267
1268                 //FIXME more multiref search?
1269                 check_4block_inter(s, mb_x, mb_y,
1270                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1271                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1272
1273                 for(i=0; i<4; i++)
1274                     if(!(b[i]->type&BLOCK_INTRA))
1275                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1276
1277                 if(init_rd != best_rd)
1278                     change++;
1279             }
1280         }
1281         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1282     }
1283 }
1284
1285 static void encode_blocks(SnowContext *s, int search){
1286     int x, y;
1287     int w= s->b_width;
1288     int h= s->b_height;
1289
1290     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1291         iterative_me(s);
1292
1293     for(y=0; y<h; y++){
1294         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1295             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1296             return;
1297         }
1298         for(x=0; x<w; x++){
1299             if(s->avctx->me_method == ME_ITER || !search)
1300                 encode_q_branch2(s, 0, x, y);
1301             else
1302                 encode_q_branch (s, 0, x, y);
1303         }
1304     }
1305 }
1306
1307 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1308     const int w= b->width;
1309     const int h= b->height;
1310     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1311     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1312     int x,y, thres1, thres2;
1313
1314     if(s->qlog == LOSSLESS_QLOG){
1315         for(y=0; y<h; y++)
1316             for(x=0; x<w; x++)
1317                 dst[x + y*stride]= src[x + y*stride];
1318         return;
1319     }
1320
1321     bias= bias ? 0 : (3*qmul)>>3;
1322     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1323     thres2= 2*thres1;
1324
1325     if(!bias){
1326         for(y=0; y<h; y++){
1327             for(x=0; x<w; x++){
1328                 int i= src[x + y*stride];
1329
1330                 if((unsigned)(i+thres1) > thres2){
1331                     if(i>=0){
1332                         i<<= QEXPSHIFT;
1333                         i/= qmul; //FIXME optimize
1334                         dst[x + y*stride]=  i;
1335                     }else{
1336                         i= -i;
1337                         i<<= QEXPSHIFT;
1338                         i/= qmul; //FIXME optimize
1339                         dst[x + y*stride]= -i;
1340                     }
1341                 }else
1342                     dst[x + y*stride]= 0;
1343             }
1344         }
1345     }else{
1346         for(y=0; y<h; y++){
1347             for(x=0; x<w; x++){
1348                 int i= src[x + y*stride];
1349
1350                 if((unsigned)(i+thres1) > thres2){
1351                     if(i>=0){
1352                         i<<= QEXPSHIFT;
1353                         i= (i + bias) / qmul; //FIXME optimize
1354                         dst[x + y*stride]=  i;
1355                     }else{
1356                         i= -i;
1357                         i<<= QEXPSHIFT;
1358                         i= (i + bias) / qmul; //FIXME optimize
1359                         dst[x + y*stride]= -i;
1360                     }
1361                 }else
1362                     dst[x + y*stride]= 0;
1363             }
1364         }
1365     }
1366 }
1367
1368 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1369     const int w= b->width;
1370     const int h= b->height;
1371     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1372     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1373     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1374     int x,y;
1375
1376     if(s->qlog == LOSSLESS_QLOG) return;
1377
1378     for(y=0; y<h; y++){
1379         for(x=0; x<w; x++){
1380             int i= src[x + y*stride];
1381             if(i<0){
1382                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1383             }else if(i>0){
1384                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1385             }
1386         }
1387     }
1388 }
1389
1390 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1391     const int w= b->width;
1392     const int h= b->height;
1393     int x,y;
1394
1395     for(y=h-1; y>=0; y--){
1396         for(x=w-1; x>=0; x--){
1397             int i= x + y*stride;
1398
1399             if(x){
1400                 if(use_median){
1401                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1402                     else  src[i] -= src[i - 1];
1403                 }else{
1404                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1405                     else  src[i] -= src[i - 1];
1406                 }
1407             }else{
1408                 if(y) src[i] -= src[i - stride];
1409             }
1410         }
1411     }
1412 }
1413
1414 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1415     const int w= b->width;
1416     const int h= b->height;
1417     int x,y;
1418
1419     for(y=0; y<h; y++){
1420         for(x=0; x<w; x++){
1421             int i= x + y*stride;
1422
1423             if(x){
1424                 if(use_median){
1425                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1426                     else  src[i] += src[i - 1];
1427                 }else{
1428                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1429                     else  src[i] += src[i - 1];
1430                 }
1431             }else{
1432                 if(y) src[i] += src[i - stride];
1433             }
1434         }
1435     }
1436 }
1437
1438 static void encode_qlogs(SnowContext *s){
1439     int plane_index, level, orientation;
1440
1441     for(plane_index=0; plane_index<2; plane_index++){
1442         for(level=0; level<s->spatial_decomposition_count; level++){
1443             for(orientation=level ? 1:0; orientation<4; orientation++){
1444                 if(orientation==2) continue;
1445                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1446             }
1447         }
1448     }
1449 }
1450
1451 static void encode_header(SnowContext *s){
1452     int plane_index, i;
1453     uint8_t kstate[32];
1454
1455     memset(kstate, MID_STATE, sizeof(kstate));
1456
1457     put_rac(&s->c, kstate, s->keyframe);
1458     if(s->keyframe || s->always_reset){
1459         ff_snow_reset_contexts(s);
1460         s->last_spatial_decomposition_type=
1461         s->last_qlog=
1462         s->last_qbias=
1463         s->last_mv_scale=
1464         s->last_block_max_depth= 0;
1465         for(plane_index=0; plane_index<2; plane_index++){
1466             Plane *p= &s->plane[plane_index];
1467             p->last_htaps=0;
1468             p->last_diag_mc=0;
1469             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1470         }
1471     }
1472     if(s->keyframe){
1473         put_symbol(&s->c, s->header_state, s->version, 0);
1474         put_rac(&s->c, s->header_state, s->always_reset);
1475         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1476         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1477         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1478         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1479         put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1480         put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1481         put_rac(&s->c, s->header_state, s->spatial_scalability);
1482 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1483         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1484
1485         encode_qlogs(s);
1486     }
1487
1488     if(!s->keyframe){
1489         int update_mc=0;
1490         for(plane_index=0; plane_index<2; plane_index++){
1491             Plane *p= &s->plane[plane_index];
1492             update_mc |= p->last_htaps   != p->htaps;
1493             update_mc |= p->last_diag_mc != p->diag_mc;
1494             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1495         }
1496         put_rac(&s->c, s->header_state, update_mc);
1497         if(update_mc){
1498             for(plane_index=0; plane_index<2; plane_index++){
1499                 Plane *p= &s->plane[plane_index];
1500                 put_rac(&s->c, s->header_state, p->diag_mc);
1501                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1502                 for(i= p->htaps/2; i; i--)
1503                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1504             }
1505         }
1506         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1507             put_rac(&s->c, s->header_state, 1);
1508             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1509             encode_qlogs(s);
1510         }else
1511             put_rac(&s->c, s->header_state, 0);
1512     }
1513
1514     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1515     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1516     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1517     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1518     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1519
1520 }
1521
1522 static void update_last_header_values(SnowContext *s){
1523     int plane_index;
1524
1525     if(!s->keyframe){
1526         for(plane_index=0; plane_index<2; plane_index++){
1527             Plane *p= &s->plane[plane_index];
1528             p->last_diag_mc= p->diag_mc;
1529             p->last_htaps  = p->htaps;
1530             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1531         }
1532     }
1533
1534     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1535     s->last_qlog                        = s->qlog;
1536     s->last_qbias                       = s->qbias;
1537     s->last_mv_scale                    = s->mv_scale;
1538     s->last_block_max_depth             = s->block_max_depth;
1539     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1540 }
1541
1542 static int qscale2qlog(int qscale){
1543     return rint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1544            + 61*QROOT/8; ///< 64 > 60
1545 }
1546
1547 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1548 {
1549     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1550      * FIXME we know exact mv bits at this point,
1551      * but ratecontrol isn't set up to include them. */
1552     uint32_t coef_sum= 0;
1553     int level, orientation, delta_qlog;
1554
1555     for(level=0; level<s->spatial_decomposition_count; level++){
1556         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1557             SubBand *b= &s->plane[0].band[level][orientation];
1558             IDWTELEM *buf= b->ibuf;
1559             const int w= b->width;
1560             const int h= b->height;
1561             const int stride= b->stride;
1562             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1563             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1564             const int qdiv= (1<<16)/qmul;
1565             int x, y;
1566             //FIXME this is ugly
1567             for(y=0; y<h; y++)
1568                 for(x=0; x<w; x++)
1569                     buf[x+y*stride]= b->buf[x+y*stride];
1570             if(orientation==0)
1571                 decorrelate(s, b, buf, stride, 1, 0);
1572             for(y=0; y<h; y++)
1573                 for(x=0; x<w; x++)
1574                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1575         }
1576     }
1577
1578     /* ugly, ratecontrol just takes a sqrt again */
1579     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1580     assert(coef_sum < INT_MAX);
1581
1582     if(pict->pict_type == AV_PICTURE_TYPE_I){
1583         s->m.current_picture.mb_var_sum= coef_sum;
1584         s->m.current_picture.mc_mb_var_sum= 0;
1585     }else{
1586         s->m.current_picture.mc_mb_var_sum= coef_sum;
1587         s->m.current_picture.mb_var_sum= 0;
1588     }
1589
1590     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1591     if (pict->quality < 0)
1592         return INT_MIN;
1593     s->lambda= pict->quality * 3/2;
1594     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1595     s->qlog+= delta_qlog;
1596     return delta_qlog;
1597 }
1598
1599 static void calculate_visual_weight(SnowContext *s, Plane *p){
1600     int width = p->width;
1601     int height= p->height;
1602     int level, orientation, x, y;
1603
1604     for(level=0; level<s->spatial_decomposition_count; level++){
1605         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1606             SubBand *b= &p->band[level][orientation];
1607             IDWTELEM *ibuf= b->ibuf;
1608             int64_t error=0;
1609
1610             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1611             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1612             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1613             for(y=0; y<height; y++){
1614                 for(x=0; x<width; x++){
1615                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1616                     error += d*d;
1617                 }
1618             }
1619
1620             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1621         }
1622     }
1623 }
1624
1625 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1626                         const AVFrame *pict, int *got_packet)
1627 {
1628     SnowContext *s = avctx->priv_data;
1629     RangeCoder * const c= &s->c;
1630     AVFrame *pic = &s->new_picture;
1631     const int width= s->avctx->width;
1632     const int height= s->avctx->height;
1633     int level, orientation, plane_index, i, y, ret;
1634     uint8_t rc_header_bak[sizeof(s->header_state)];
1635     uint8_t rc_block_bak[sizeof(s->block_state)];
1636
1637     if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0)
1638         return ret;
1639
1640     ff_init_range_encoder(c, pkt->data, pkt->size);
1641     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1642
1643     for(i=0; i<3; i++){
1644         int hshift= i ? s->chroma_h_shift : 0;
1645         int vshift= i ? s->chroma_v_shift : 0;
1646         for(y=0; y<(height>>vshift); y++)
1647             memcpy(&s->input_picture.data[i][y * s->input_picture.linesize[i]],
1648                    &pict->data[i][y * pict->linesize[i]],
1649                    width>>hshift);
1650     }
1651     s->new_picture = *pict;
1652
1653     s->m.picture_number= avctx->frame_number;
1654     if(avctx->flags&CODEC_FLAG_PASS2){
1655         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1656         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1657         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1658             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1659             if (pic->quality < 0)
1660                 return -1;
1661         }
1662     }else{
1663         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1664         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1665     }
1666
1667     if(s->pass1_rc && avctx->frame_number == 0)
1668         pic->quality = 2*FF_QP2LAMBDA;
1669     if (pic->quality) {
1670         s->qlog   = qscale2qlog(pic->quality);
1671         s->lambda = pic->quality * 3/2;
1672     }
1673     if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1674         s->qlog= LOSSLESS_QLOG;
1675         s->lambda = 0;
1676     }//else keep previous frame's qlog until after motion estimation
1677
1678     ff_snow_frame_start(s);
1679
1680     s->m.current_picture_ptr= &s->m.current_picture;
1681     s->m.last_picture.f.pts = s->m.current_picture.f.pts;
1682     s->m.current_picture.f.pts = pict->pts;
1683     if(pic->pict_type == AV_PICTURE_TYPE_P){
1684         int block_width = (width +15)>>4;
1685         int block_height= (height+15)>>4;
1686         int stride= s->current_picture.linesize[0];
1687
1688         assert(s->current_picture.data[0]);
1689         assert(s->last_picture[0].data[0]);
1690
1691         s->m.avctx= s->avctx;
1692         s->m.current_picture.f.data[0] = s->current_picture.data[0];
1693         s->m.   last_picture.f.data[0] = s->last_picture[0].data[0];
1694         s->m.    new_picture.f.data[0] = s->  input_picture.data[0];
1695         s->m.   last_picture_ptr= &s->m.   last_picture;
1696         s->m.linesize=
1697         s->m.   last_picture.f.linesize[0] =
1698         s->m.    new_picture.f.linesize[0] =
1699         s->m.current_picture.f.linesize[0] = stride;
1700         s->m.uvlinesize= s->current_picture.linesize[1];
1701         s->m.width = width;
1702         s->m.height= height;
1703         s->m.mb_width = block_width;
1704         s->m.mb_height= block_height;
1705         s->m.mb_stride=   s->m.mb_width+1;
1706         s->m.b8_stride= 2*s->m.mb_width+1;
1707         s->m.f_code=1;
1708         s->m.pict_type = pic->pict_type;
1709         s->m.me_method= s->avctx->me_method;
1710         s->m.me.scene_change_score=0;
1711         s->m.flags= s->avctx->flags;
1712         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1713         s->m.out_format= FMT_H263;
1714         s->m.unrestricted_mv= 1;
1715
1716         s->m.lambda = s->lambda;
1717         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1718         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1719
1720         s->m.dsp= s->dsp; //move
1721         ff_init_me(&s->m);
1722         s->dsp= s->m.dsp;
1723     }
1724
1725     if(s->pass1_rc){
1726         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1727         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1728     }
1729
1730 redo_frame:
1731
1732     if (pic->pict_type == AV_PICTURE_TYPE_I)
1733         s->spatial_decomposition_count= 5;
1734     else
1735         s->spatial_decomposition_count= 5;
1736
1737     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1738           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1739         s->spatial_decomposition_count--;
1740
1741     s->m.pict_type = pic->pict_type;
1742     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1743
1744     ff_snow_common_init_after_header(avctx);
1745
1746     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1747         for(plane_index=0; plane_index<3; plane_index++){
1748             calculate_visual_weight(s, &s->plane[plane_index]);
1749         }
1750     }
1751
1752     encode_header(s);
1753     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1754     encode_blocks(s, 1);
1755     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1756
1757     for(plane_index=0; plane_index<3; plane_index++){
1758         Plane *p= &s->plane[plane_index];
1759         int w= p->width;
1760         int h= p->height;
1761         int x, y;
1762 //        int bits= put_bits_count(&s->c.pb);
1763
1764         if (!s->memc_only) {
1765             //FIXME optimize
1766             if(pict->data[plane_index]) //FIXME gray hack
1767                 for(y=0; y<h; y++){
1768                     for(x=0; x<w; x++){
1769                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1770                     }
1771                 }
1772             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1773
1774             if(   plane_index==0
1775                && pic->pict_type == AV_PICTURE_TYPE_P
1776                && !(avctx->flags&CODEC_FLAG_PASS2)
1777                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1778                 ff_init_range_encoder(c, pkt->data, pkt->size);
1779                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1780                 pic->pict_type= AV_PICTURE_TYPE_I;
1781                 s->keyframe=1;
1782                 s->current_picture.key_frame=1;
1783                 goto redo_frame;
1784             }
1785
1786             if(s->qlog == LOSSLESS_QLOG){
1787                 for(y=0; y<h; y++){
1788                     for(x=0; x<w; x++){
1789                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1790                     }
1791                 }
1792             }else{
1793                 for(y=0; y<h; y++){
1794                     for(x=0; x<w; x++){
1795                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1796                     }
1797                 }
1798             }
1799
1800             /*  if(QUANTIZE2)
1801                 dwt_quantize(s, p, s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type);
1802             else*/
1803                 ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1804
1805             if(s->pass1_rc && plane_index==0){
1806                 int delta_qlog = ratecontrol_1pass(s, pic);
1807                 if (delta_qlog <= INT_MIN)
1808                     return -1;
1809                 if(delta_qlog){
1810                     //reordering qlog in the bitstream would eliminate this reset
1811                     ff_init_range_encoder(c, pkt->data, pkt->size);
1812                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1813                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1814                     encode_header(s);
1815                     encode_blocks(s, 0);
1816                 }
1817             }
1818
1819             for(level=0; level<s->spatial_decomposition_count; level++){
1820                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1821                     SubBand *b= &p->band[level][orientation];
1822
1823                     if(!QUANTIZE2)
1824                         quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1825                     if(orientation==0)
1826                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1827                     if (!s->no_bitstream)
1828                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1829                     assert(b->parent==NULL || b->parent->stride == b->stride*2);
1830                     if(orientation==0)
1831                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1832                 }
1833             }
1834
1835             for(level=0; level<s->spatial_decomposition_count; level++){
1836                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1837                     SubBand *b= &p->band[level][orientation];
1838
1839                     dequantize(s, b, b->ibuf, b->stride);
1840                 }
1841             }
1842
1843             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1844             if(s->qlog == LOSSLESS_QLOG){
1845                 for(y=0; y<h; y++){
1846                     for(x=0; x<w; x++){
1847                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1848                     }
1849                 }
1850             }
1851             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1852         }else{
1853             //ME/MC only
1854             if(pic->pict_type == AV_PICTURE_TYPE_I){
1855                 for(y=0; y<h; y++){
1856                     for(x=0; x<w; x++){
1857                         s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]=
1858                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1859                     }
1860                 }
1861             }else{
1862                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1863                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1864             }
1865         }
1866         if(s->avctx->flags&CODEC_FLAG_PSNR){
1867             int64_t error= 0;
1868
1869             if(pict->data[plane_index]) //FIXME gray hack
1870                 for(y=0; y<h; y++){
1871                     for(x=0; x<w; x++){
1872                         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];
1873                         error += d*d;
1874                     }
1875                 }
1876             s->avctx->error[plane_index] += error;
1877             s->current_picture.error[plane_index] = error;
1878         }
1879
1880     }
1881
1882     update_last_header_values(s);
1883
1884     ff_snow_release_buffer(avctx);
1885
1886     s->current_picture.coded_picture_number = avctx->frame_number;
1887     s->current_picture.pict_type = pict->pict_type;
1888     s->current_picture.quality = pict->quality;
1889     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1890     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1891     s->m.current_picture.f.display_picture_number =
1892     s->m.current_picture.f.coded_picture_number   = avctx->frame_number;
1893     s->m.current_picture.f.quality                = pic->quality;
1894     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1895     if(s->pass1_rc)
1896         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1897             return -1;
1898     if(avctx->flags&CODEC_FLAG_PASS1)
1899         ff_write_pass1_stats(&s->m);
1900     s->m.last_pict_type = s->m.pict_type;
1901     avctx->frame_bits = s->m.frame_bits;
1902     avctx->mv_bits = s->m.mv_bits;
1903     avctx->misc_bits = s->m.misc_bits;
1904     avctx->p_tex_bits = s->m.p_tex_bits;
1905
1906     emms_c();
1907
1908     pkt->size = ff_rac_terminate(c);
1909     if (avctx->coded_frame->key_frame)
1910         pkt->flags |= AV_PKT_FLAG_KEY;
1911     *got_packet = 1;
1912
1913     return 0;
1914 }
1915
1916 static av_cold int encode_end(AVCodecContext *avctx)
1917 {
1918     SnowContext *s = avctx->priv_data;
1919
1920     ff_snow_common_end(s);
1921     if (s->input_picture.data[0])
1922         avctx->release_buffer(avctx, &s->input_picture);
1923     av_free(avctx->stats_out);
1924
1925     return 0;
1926 }
1927
1928 #define OFFSET(x) offsetof(SnowContext, x)
1929 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1930 static const AVOption options[] = {
1931     { "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 },
1932     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1933     { NULL },
1934 };
1935
1936 static const AVClass snowenc_class = {
1937     .class_name = "snow encoder",
1938     .item_name  = av_default_item_name,
1939     .option     = options,
1940     .version    = LIBAVUTIL_VERSION_INT,
1941 };
1942
1943 AVCodec ff_snow_encoder = {
1944     .name           = "snow",
1945     .type           = AVMEDIA_TYPE_VIDEO,
1946     .id             = AV_CODEC_ID_SNOW,
1947     .priv_data_size = sizeof(SnowContext),
1948     .init           = encode_init,
1949     .encode2        = encode_frame,
1950     .close          = encode_end,
1951     .pix_fmts       = (const enum AVPixelFormat[]){
1952         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1953         AV_PIX_FMT_NONE
1954     },
1955     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1956     .priv_class     = &snowenc_class,
1957 };
1958
1959
1960 #ifdef TEST
1961 #undef malloc
1962 #undef free
1963 #undef printf
1964
1965 #include "libavutil/lfg.h"
1966 #include "libavutil/mathematics.h"
1967
1968 int main(void){
1969 #define width  256
1970 #define height 256
1971     int buffer[2][width*height];
1972     SnowContext s;
1973     int i;
1974     AVLFG prng;
1975     s.spatial_decomposition_count=6;
1976     s.spatial_decomposition_type=1;
1977
1978     s.temp_dwt_buffer  = av_mallocz(width * sizeof(DWTELEM));
1979     s.temp_idwt_buffer = av_mallocz(width * sizeof(IDWTELEM));
1980
1981     av_lfg_init(&prng, 1);
1982
1983     printf("testing 5/3 DWT\n");
1984     for(i=0; i<width*height; i++)
1985         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1986
1987     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1988     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1989
1990     for(i=0; i<width*height; i++)
1991         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1992
1993     printf("testing 9/7 DWT\n");
1994     s.spatial_decomposition_type=0;
1995     for(i=0; i<width*height; i++)
1996         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1997
1998     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1999     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2000
2001     for(i=0; i<width*height; i++)
2002         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
2003
2004     {
2005     int level, orientation, x, y;
2006     int64_t errors[8][4];
2007     int64_t g=0;
2008
2009         memset(errors, 0, sizeof(errors));
2010         s.spatial_decomposition_count=3;
2011         s.spatial_decomposition_type=0;
2012         for(level=0; level<s.spatial_decomposition_count; level++){
2013             for(orientation=level ? 1 : 0; orientation<4; orientation++){
2014                 int w= width  >> (s.spatial_decomposition_count-level);
2015                 int h= height >> (s.spatial_decomposition_count-level);
2016                 int stride= width  << (s.spatial_decomposition_count-level);
2017                 DWTELEM *buf= buffer[0];
2018                 int64_t error=0;
2019
2020                 if(orientation&1) buf+=w;
2021                 if(orientation>1) buf+=stride>>1;
2022
2023                 memset(buffer[0], 0, sizeof(int)*width*height);
2024                 buf[w/2 + h/2*stride]= 256*256;
2025                 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2026                 for(y=0; y<height; y++){
2027                     for(x=0; x<width; x++){
2028                         int64_t d= buffer[0][x + y*width];
2029                         error += d*d;
2030                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
2031                     }
2032                     if(FFABS(height/2-y)<9 && level==2) printf("\n");
2033                 }
2034                 error= (int)(sqrt(error)+0.5);
2035                 errors[level][orientation]= error;
2036                 if(g) g=av_gcd(g, error);
2037                 else g= error;
2038             }
2039         }
2040         printf("static int const visual_weight[][4]={\n");
2041         for(level=0; level<s.spatial_decomposition_count; level++){
2042             printf("  {");
2043             for(orientation=0; orientation<4; orientation++){
2044                 printf("%8"PRId64",", errors[level][orientation]/g);
2045             }
2046             printf("},\n");
2047         }
2048         printf("};\n");
2049         {
2050             int level=2;
2051             int w= width  >> (s.spatial_decomposition_count-level);
2052             //int h= height >> (s.spatial_decomposition_count-level);
2053             int stride= width  << (s.spatial_decomposition_count-level);
2054             DWTELEM *buf= buffer[0];
2055             int64_t error=0;
2056
2057             buf+=w;
2058             buf+=stride>>1;
2059
2060             memset(buffer[0], 0, sizeof(int)*width*height);
2061             for(y=0; y<height; y++){
2062                 for(x=0; x<width; x++){
2063                     int tab[4]={0,2,3,1};
2064                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2065                 }
2066             }
2067             ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2068             for(y=0; y<height; y++){
2069                 for(x=0; x<width; x++){
2070                     int64_t d= buffer[0][x + y*width];
2071                     error += d*d;
2072                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2073                 }
2074                 if(FFABS(height/2-y)<9) printf("\n");
2075             }
2076         }
2077
2078     }
2079     return 0;
2080 }
2081 #endif /* TEST */