Merge remote-tracking branch 'qatar/master'
[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 "dwt.h"
28 #include "snow.h"
29
30 #include "rangecoder.h"
31 #include "mathops.h"
32
33 #include "mpegvideo.h"
34 #include "h263.h"
35
36 #undef NDEBUG
37 #include <assert.h>
38
39 #define QUANTIZE2 0
40
41 #if QUANTIZE2==1
42 #define Q2_STEP 8
43
44 static void find_sse(SnowContext *s, Plane *p, int *score, int score_stride, IDWTELEM *r0, IDWTELEM *r1, int level, int orientation){
45     SubBand *b= &p->band[level][orientation];
46     int x, y;
47     int xo=0;
48     int yo=0;
49     int step= 1 << (s->spatial_decomposition_count - level);
50
51     if(orientation&1)
52         xo= step>>1;
53     if(orientation&2)
54         yo= step>>1;
55
56     //FIXME bias for nonzero ?
57     //FIXME optimize
58     memset(score, 0, sizeof(*score)*score_stride*((p->height + Q2_STEP-1)/Q2_STEP));
59     for(y=0; y<p->height; y++){
60         for(x=0; x<p->width; x++){
61             int sx= (x-xo + step/2) / step / Q2_STEP;
62             int sy= (y-yo + step/2) / step / Q2_STEP;
63             int v= r0[x + y*p->width] - r1[x + y*p->width];
64             assert(sx>=0 && sy>=0 && sx < score_stride);
65             v= ((v+8)>>4)<<4;
66             score[sx + sy*score_stride] += v*v;
67             assert(score[sx + sy*score_stride] >= 0);
68         }
69     }
70 }
71
72 static void dequantize_all(SnowContext *s, Plane *p, IDWTELEM *buffer, int width, int height){
73     int level, orientation;
74
75     for(level=0; level<s->spatial_decomposition_count; level++){
76         for(orientation=level ? 1 : 0; orientation<4; orientation++){
77             SubBand *b= &p->band[level][orientation];
78             IDWTELEM *dst= buffer + (b->ibuf - s->spatial_idwt_buffer);
79
80             dequantize(s, b, dst, b->stride);
81         }
82     }
83 }
84
85 static void dwt_quantize(SnowContext *s, Plane *p, DWTELEM *buffer, int width, int height, int stride, int type){
86     int level, orientation, ys, xs, x, y, pass;
87     IDWTELEM best_dequant[height * stride];
88     IDWTELEM idwt2_buffer[height * stride];
89     const int score_stride= (width + 10)/Q2_STEP;
90     int best_score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
91     int score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
92     int threshold= (s->m.lambda * s->m.lambda) >> 6;
93
94     //FIXME pass the copy cleanly ?
95
96 //    memcpy(dwt_buffer, buffer, height * stride * sizeof(DWTELEM));
97     ff_spatial_dwt(buffer, s->temp_dwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
98
99     for(level=0; level<s->spatial_decomposition_count; level++){
100         for(orientation=level ? 1 : 0; orientation<4; orientation++){
101             SubBand *b= &p->band[level][orientation];
102             IDWTELEM *dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
103              DWTELEM *src=       buffer + (b-> buf - s->spatial_dwt_buffer);
104             assert(src == b->buf); // code does not depend on this but it is true currently
105
106             quantize(s, b, dst, src, b->stride, s->qbias);
107         }
108     }
109     for(pass=0; pass<1; pass++){
110         if(s->qbias == 0) //keyframe
111             continue;
112         for(level=0; level<s->spatial_decomposition_count; level++){
113             for(orientation=level ? 1 : 0; orientation<4; orientation++){
114                 SubBand *b= &p->band[level][orientation];
115                 IDWTELEM *dst= idwt2_buffer + (b->ibuf - s->spatial_idwt_buffer);
116                 IDWTELEM *best_dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
117
118                 for(ys= 0; ys<Q2_STEP; ys++){
119                     for(xs= 0; xs<Q2_STEP; xs++){
120                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
121                         dequantize_all(s, p, idwt2_buffer, width, height);
122                         ff_spatial_idwt(idwt2_buffer, s->temp_idwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
123                         find_sse(s, p, best_score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
124                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
125                         for(y=ys; y<b->height; y+= Q2_STEP){
126                             for(x=xs; x<b->width; x+= Q2_STEP){
127                                 if(dst[x + y*b->stride]<0) dst[x + y*b->stride]++;
128                                 if(dst[x + y*b->stride]>0) dst[x + y*b->stride]--;
129                                 //FIXME try more than just --
130                             }
131                         }
132                         dequantize_all(s, p, idwt2_buffer, width, height);
133                         ff_spatial_idwt(idwt2_buffer, s->temp_idwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
134                         find_sse(s, p, score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
135                         for(y=ys; y<b->height; y+= Q2_STEP){
136                             for(x=xs; x<b->width; x+= Q2_STEP){
137                                 int score_idx= x/Q2_STEP + (y/Q2_STEP)*score_stride;
138                                 if(score[score_idx] <= best_score[score_idx] + threshold){
139                                     best_score[score_idx]= score[score_idx];
140                                     if(best_dst[x + y*b->stride]<0) best_dst[x + y*b->stride]++;
141                                     if(best_dst[x + y*b->stride]>0) best_dst[x + y*b->stride]--;
142                                     //FIXME copy instead
143                                 }
144                             }
145                         }
146                     }
147                 }
148             }
149         }
150     }
151     memcpy(s->spatial_idwt_buffer, best_dequant, height * stride * sizeof(IDWTELEM)); //FIXME work with that directly instead of copy at the end
152 }
153
154 #endif /* QUANTIZE2==1 */
155
156 static av_cold int encode_init(AVCodecContext *avctx)
157 {
158     SnowContext *s = avctx->priv_data;
159     int plane_index;
160
161     if(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL){
162         av_log(avctx, AV_LOG_ERROR, "This codec is under development, files encoded with it may not be decodable with future versions!!!\n"
163                "Use vstrict=-2 / -strict -2 to use it anyway.\n");
164         return -1;
165     }
166
167     if(avctx->prediction_method == DWT_97
168        && (avctx->flags & CODEC_FLAG_QSCALE)
169        && avctx->global_quality == 0){
170         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
171         return -1;
172     }
173
174     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
175
176     s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
177     s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
178
179     for(plane_index=0; plane_index<3; plane_index++){
180         s->plane[plane_index].diag_mc= 1;
181         s->plane[plane_index].htaps= 6;
182         s->plane[plane_index].hcoeff[0]=  40;
183         s->plane[plane_index].hcoeff[1]= -10;
184         s->plane[plane_index].hcoeff[2]=   2;
185         s->plane[plane_index].fast_mc= 1;
186     }
187
188     ff_snow_common_init(avctx);
189     ff_snow_alloc_blocks(s);
190
191     s->version=0;
192
193     s->m.avctx   = avctx;
194     s->m.flags   = avctx->flags;
195     s->m.bit_rate= avctx->bit_rate;
196
197     s->m.me.temp      =
198     s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
199     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
200     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
201     s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
202     ff_h263_encode_init(&s->m); //mv_penalty
203
204     s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
205
206     if(avctx->flags&CODEC_FLAG_PASS1){
207         if(!avctx->stats_out)
208             avctx->stats_out = av_mallocz(256);
209     }
210     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
211         if(ff_rate_control_init(&s->m) < 0)
212             return -1;
213     }
214     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
215
216     avctx->coded_frame= &s->current_picture;
217     switch(avctx->pix_fmt){
218     case PIX_FMT_YUV444P:
219 //    case PIX_FMT_YUV422P:
220     case PIX_FMT_YUV420P:
221 //     case PIX_FMT_GRAY8:
222 //    case PIX_FMT_YUV411P:
223     case PIX_FMT_YUV410P:
224         s->colorspace_type= 0;
225         break;
226 /*    case PIX_FMT_RGB32:
227         s->colorspace= 1;
228         break;*/
229     default:
230         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
231         return -1;
232     }
233     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
234
235     ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
236     ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
237
238     s->avctx->get_buffer(s->avctx, &s->input_picture);
239
240     if(s->avctx->me_method == ME_ITER){
241         int i;
242         int size= s->b_width * s->b_height << 2*s->block_max_depth;
243         for(i=0; i<s->max_ref_frames; i++){
244             s->ref_mvs[i]= av_mallocz(size*sizeof(int16_t[2]));
245             s->ref_scores[i]= av_mallocz(size*sizeof(uint32_t));
246         }
247     }
248
249     s->runs = av_malloc(avctx->width * avctx->height * sizeof(*s->runs));
250
251     return 0;
252 }
253
254 //near copy & paste from dsputil, FIXME
255 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
256 {
257     int s, i, j;
258
259     s = 0;
260     for (i = 0; i < h; i++) {
261         for (j = 0; j < w; j++) {
262             s += pix[0];
263             pix ++;
264         }
265         pix += line_size - w;
266     }
267     return s;
268 }
269
270 //near copy & paste from dsputil, FIXME
271 static int pix_norm1(uint8_t * pix, int line_size, int w)
272 {
273     int s, i, j;
274     uint32_t *sq = ff_squareTbl + 256;
275
276     s = 0;
277     for (i = 0; i < w; i++) {
278         for (j = 0; j < w; j ++) {
279             s += sq[pix[0]];
280             pix ++;
281         }
282         pix += line_size - w;
283     }
284     return s;
285 }
286
287 //FIXME copy&paste
288 #define P_LEFT P[1]
289 #define P_TOP P[2]
290 #define P_TOPRIGHT P[3]
291 #define P_MEDIAN P[4]
292 #define P_MV1 P[9]
293 #define FLAG_QPEL   1 //must be 1
294
295 static int encode_q_branch(SnowContext *s, int level, int x, int y){
296     uint8_t p_buffer[1024];
297     uint8_t i_buffer[1024];
298     uint8_t p_state[sizeof(s->block_state)];
299     uint8_t i_state[sizeof(s->block_state)];
300     RangeCoder pc, ic;
301     uint8_t *pbbak= s->c.bytestream;
302     uint8_t *pbbak_start= s->c.bytestream_start;
303     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
304     const int w= s->b_width  << s->block_max_depth;
305     const int h= s->b_height << s->block_max_depth;
306     const int rem_depth= s->block_max_depth - level;
307     const int index= (x + y*w) << rem_depth;
308     const int block_w= 1<<(LOG2_MB_SIZE - level);
309     int trx= (x+1)<<rem_depth;
310     int try= (y+1)<<rem_depth;
311     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
312     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
313     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
314     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
315     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
316     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
317     int pl = left->color[0];
318     int pcb= left->color[1];
319     int pcr= left->color[2];
320     int pmx, pmy;
321     int mx=0, my=0;
322     int l,cr,cb;
323     const int stride= s->current_picture.linesize[0];
324     const int uvstride= s->current_picture.linesize[1];
325     uint8_t *current_data[3]= { s->input_picture.data[0] + (x + y*  stride)*block_w,
326                                 s->input_picture.data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
327                                 s->input_picture.data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
328     int P[10][2];
329     int16_t last_mv[3][2];
330     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
331     const int shift= 1+qpel;
332     MotionEstContext *c= &s->m.me;
333     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
334     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
335     int my_context= av_log2(2*FFABS(left->my - top->my));
336     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
337     int ref, best_ref, ref_score, ref_mx, ref_my;
338
339     assert(sizeof(s->block_state) >= 256);
340     if(s->keyframe){
341         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
342         return 0;
343     }
344
345 //    clip predictors / edge ?
346
347     P_LEFT[0]= left->mx;
348     P_LEFT[1]= left->my;
349     P_TOP [0]= top->mx;
350     P_TOP [1]= top->my;
351     P_TOPRIGHT[0]= tr->mx;
352     P_TOPRIGHT[1]= tr->my;
353
354     last_mv[0][0]= s->block[index].mx;
355     last_mv[0][1]= s->block[index].my;
356     last_mv[1][0]= right->mx;
357     last_mv[1][1]= right->my;
358     last_mv[2][0]= bottom->mx;
359     last_mv[2][1]= bottom->my;
360
361     s->m.mb_stride=2;
362     s->m.mb_x=
363     s->m.mb_y= 0;
364     c->skip= 0;
365
366     assert(c->  stride ==   stride);
367     assert(c->uvstride == uvstride);
368
369     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
370     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
371     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
372     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
373
374     c->xmin = - x*block_w - 16+3;
375     c->ymin = - y*block_w - 16+3;
376     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
377     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
378
379     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
380     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
381     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
382     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
383     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
384     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
385     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
386
387     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
388     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
389
390     if (!y) {
391         c->pred_x= P_LEFT[0];
392         c->pred_y= P_LEFT[1];
393     } else {
394         c->pred_x = P_MEDIAN[0];
395         c->pred_y = P_MEDIAN[1];
396     }
397
398     score= INT_MAX;
399     best_ref= 0;
400     for(ref=0; ref<s->ref_frames; ref++){
401         init_ref(c, current_data, s->last_picture[ref].data, NULL, block_w*x, block_w*y, 0);
402
403         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
404                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
405
406         assert(ref_mx >= c->xmin);
407         assert(ref_mx <= c->xmax);
408         assert(ref_my >= c->ymin);
409         assert(ref_my <= c->ymax);
410
411         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
412         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
413         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
414         if(s->ref_mvs[ref]){
415             s->ref_mvs[ref][index][0]= ref_mx;
416             s->ref_mvs[ref][index][1]= ref_my;
417             s->ref_scores[ref][index]= ref_score;
418         }
419         if(score > ref_score){
420             score= ref_score;
421             best_ref= ref;
422             mx= ref_mx;
423             my= ref_my;
424         }
425     }
426     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
427
428   //  subpel search
429     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
430     pc= s->c;
431     pc.bytestream_start=
432     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
433     memcpy(p_state, s->block_state, sizeof(s->block_state));
434
435     if(level!=s->block_max_depth)
436         put_rac(&pc, &p_state[4 + s_context], 1);
437     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
438     if(s->ref_frames > 1)
439         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
440     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
441     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
442     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
443     p_len= pc.bytestream - pc.bytestream_start;
444     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
445
446     block_s= block_w*block_w;
447     sum = pix_sum(current_data[0], stride, block_w, block_w);
448     l= (sum + block_s/2)/block_s;
449     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
450
451     block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
452     sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
453     cb= (sum + block_s/2)/block_s;
454 //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
455     sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
456     cr= (sum + block_s/2)/block_s;
457 //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
458
459     ic= s->c;
460     ic.bytestream_start=
461     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
462     memcpy(i_state, s->block_state, sizeof(s->block_state));
463     if(level!=s->block_max_depth)
464         put_rac(&ic, &i_state[4 + s_context], 1);
465     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
466     put_symbol(&ic, &i_state[32],  l-pl , 1);
467     put_symbol(&ic, &i_state[64], cb-pcb, 1);
468     put_symbol(&ic, &i_state[96], cr-pcr, 1);
469     i_len= ic.bytestream - ic.bytestream_start;
470     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
471
472 //    assert(score==256*256*256*64-1);
473     assert(iscore < 255*255*256 + s->lambda2*10);
474     assert(iscore >= 0);
475     assert(l>=0 && l<=255);
476     assert(pl>=0 && pl<=255);
477
478     if(level==0){
479         int varc= iscore >> 8;
480         int vard= score >> 8;
481         if (vard <= 64 || vard < varc)
482             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
483         else
484             c->scene_change_score+= s->m.qscale;
485     }
486
487     if(level!=s->block_max_depth){
488         put_rac(&s->c, &s->block_state[4 + s_context], 0);
489         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
490         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
491         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
492         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
493         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
494
495         if(score2 < score && score2 < iscore)
496             return score2;
497     }
498
499     if(iscore < score){
500         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
501         memcpy(pbbak, i_buffer, i_len);
502         s->c= ic;
503         s->c.bytestream_start= pbbak_start;
504         s->c.bytestream= pbbak + i_len;
505         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
506         memcpy(s->block_state, i_state, sizeof(s->block_state));
507         return iscore;
508     }else{
509         memcpy(pbbak, p_buffer, p_len);
510         s->c= pc;
511         s->c.bytestream_start= pbbak_start;
512         s->c.bytestream= pbbak + p_len;
513         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
514         memcpy(s->block_state, p_state, sizeof(s->block_state));
515         return score;
516     }
517 }
518
519 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
520     const int w= s->b_width  << s->block_max_depth;
521     const int rem_depth= s->block_max_depth - level;
522     const int index= (x + y*w) << rem_depth;
523     int trx= (x+1)<<rem_depth;
524     BlockNode *b= &s->block[index];
525     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
526     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
527     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
528     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
529     int pl = left->color[0];
530     int pcb= left->color[1];
531     int pcr= left->color[2];
532     int pmx, pmy;
533     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
534     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
535     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
536     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
537
538     if(s->keyframe){
539         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
540         return;
541     }
542
543     if(level!=s->block_max_depth){
544         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
545             put_rac(&s->c, &s->block_state[4 + s_context], 1);
546         }else{
547             put_rac(&s->c, &s->block_state[4 + s_context], 0);
548             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
549             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
550             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
551             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
552             return;
553         }
554     }
555     if(b->type & BLOCK_INTRA){
556         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
557         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
558         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
559         put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
560         put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
561         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
562     }else{
563         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
564         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
565         if(s->ref_frames > 1)
566             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
567         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
568         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
569         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
570     }
571 }
572
573 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
574     int i, x2, y2;
575     Plane *p= &s->plane[plane_index];
576     const int block_size = MB_SIZE >> s->block_max_depth;
577     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
578     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
579     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
580     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
581     const int ref_stride= s->current_picture.linesize[plane_index];
582     uint8_t *src= s-> input_picture.data[plane_index];
583     IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
584     const int b_stride = s->b_width << s->block_max_depth;
585     const int w= p->width;
586     const int h= p->height;
587     int index= mb_x + mb_y*b_stride;
588     BlockNode *b= &s->block[index];
589     BlockNode backup= *b;
590     int ab=0;
591     int aa=0;
592
593     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
594
595     b->type|= BLOCK_INTRA;
596     b->color[plane_index]= 0;
597     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
598
599     for(i=0; i<4; i++){
600         int mb_x2= mb_x + (i &1) - 1;
601         int mb_y2= mb_y + (i>>1) - 1;
602         int x= block_w*mb_x2 + block_w/2;
603         int y= block_h*mb_y2 + block_h/2;
604
605         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
606                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
607
608         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
609             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
610                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
611                 int obmc_v= obmc[index];
612                 int d;
613                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
614                 if(x<0) obmc_v += obmc[index + block_w];
615                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
616                 if(x+block_w>w) obmc_v += obmc[index - block_w];
617                 //FIXME precalculate this or simplify it somehow else
618
619                 d = -dst[index] + (1<<(FRAC_BITS-1));
620                 dst[index] = d;
621                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
622                 aa += obmc_v * obmc_v; //FIXME precalculate this
623             }
624         }
625     }
626     *b= backup;
627
628     return av_clip(((ab<<LOG2_OBMC_MAX) + aa/2)/aa, 0, 255); //FIXME we should not need clipping
629 }
630
631 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
632     const int b_stride = s->b_width << s->block_max_depth;
633     const int b_height = s->b_height<< s->block_max_depth;
634     int index= x + y*b_stride;
635     const BlockNode *b     = &s->block[index];
636     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
637     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
638     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
639     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
640     int dmx, dmy;
641 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
642 //  int my_context= av_log2(2*FFABS(left->my - top->my));
643
644     if(x<0 || x>=b_stride || y>=b_height)
645         return 0;
646 /*
647 1            0      0
648 01X          1-2    1
649 001XX        3-6    2-3
650 0001XXX      7-14   4-7
651 00001XXXX   15-30   8-15
652 */
653 //FIXME try accurate rate
654 //FIXME intra and inter predictors if surrounding blocks are not the same type
655     if(b->type & BLOCK_INTRA){
656         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
657                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
658                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
659     }else{
660         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
661         dmx-= b->mx;
662         dmy-= b->my;
663         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
664                     + av_log2(2*FFABS(dmy))
665                     + av_log2(2*b->ref));
666     }
667 }
668
669 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
670     Plane *p= &s->plane[plane_index];
671     const int block_size = MB_SIZE >> s->block_max_depth;
672     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
673     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
674     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
675     const int ref_stride= s->current_picture.linesize[plane_index];
676     uint8_t *dst= s->current_picture.data[plane_index];
677     uint8_t *src= s->  input_picture.data[plane_index];
678     IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
679     uint8_t *cur = s->scratchbuf;
680     uint8_t tmp[ref_stride*(2*MB_SIZE+HTAPS_MAX-1)];
681     const int b_stride = s->b_width << s->block_max_depth;
682     const int b_height = s->b_height<< s->block_max_depth;
683     const int w= p->width;
684     const int h= p->height;
685     int distortion;
686     int rate= 0;
687     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
688     int sx= block_w*mb_x - block_w/2;
689     int sy= block_h*mb_y - block_h/2;
690     int x0= FFMAX(0,-sx);
691     int y0= FFMAX(0,-sy);
692     int x1= FFMIN(block_w*2, w-sx);
693     int y1= FFMIN(block_h*2, h-sy);
694     int i,x,y;
695
696     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
697
698     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);
699
700     for(y=y0; y<y1; y++){
701         const uint8_t *obmc1= obmc_edged[y];
702         const IDWTELEM *pred1 = pred + y*obmc_stride;
703         uint8_t *cur1 = cur + y*ref_stride;
704         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
705         for(x=x0; x<x1; x++){
706 #if FRAC_BITS >= LOG2_OBMC_MAX
707             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
708 #else
709             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
710 #endif
711             v = (v + pred1[x]) >> FRAC_BITS;
712             if(v&(~255)) v= ~(v>>31);
713             dst1[x] = v;
714         }
715     }
716
717     /* copy the regions where obmc[] = (uint8_t)256 */
718     if(LOG2_OBMC_MAX == 8
719         && (mb_x == 0 || mb_x == b_stride-1)
720         && (mb_y == 0 || mb_y == b_height-1)){
721         if(mb_x == 0)
722             x1 = block_w;
723         else
724             x0 = block_w;
725         if(mb_y == 0)
726             y1 = block_h;
727         else
728             y0 = block_h;
729         for(y=y0; y<y1; y++)
730             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
731     }
732
733     if(block_w==16){
734         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
735         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
736         /* FIXME cmps overlap but do not cover the wavelet's whole support.
737          * So improving the score of one block is not strictly guaranteed
738          * to improve the score of the whole frame, thus iterative motion
739          * estimation does not always converge. */
740         if(s->avctx->me_cmp == FF_CMP_W97)
741             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
742         else if(s->avctx->me_cmp == FF_CMP_W53)
743             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
744         else{
745             distortion = 0;
746             for(i=0; i<4; i++){
747                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
748                 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
749             }
750         }
751     }else{
752         assert(block_w==8);
753         distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
754     }
755
756     if(plane_index==0){
757         for(i=0; i<4; i++){
758 /* ..RRr
759  * .RXx.
760  * rxx..
761  */
762             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
763         }
764         if(mb_x == b_stride-2)
765             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
766     }
767     return distortion + rate*penalty_factor;
768 }
769
770 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
771     int i, y2;
772     Plane *p= &s->plane[plane_index];
773     const int block_size = MB_SIZE >> s->block_max_depth;
774     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
775     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
776     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
777     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
778     const int ref_stride= s->current_picture.linesize[plane_index];
779     uint8_t *dst= s->current_picture.data[plane_index];
780     uint8_t *src= s-> input_picture.data[plane_index];
781     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
782     // const has only been removed from zero_dst to suppress a warning
783     static IDWTELEM zero_dst[4096]; //FIXME
784     const int b_stride = s->b_width << s->block_max_depth;
785     const int w= p->width;
786     const int h= p->height;
787     int distortion= 0;
788     int rate= 0;
789     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
790
791     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
792
793     for(i=0; i<9; i++){
794         int mb_x2= mb_x + (i%3) - 1;
795         int mb_y2= mb_y + (i/3) - 1;
796         int x= block_w*mb_x2 + block_w/2;
797         int y= block_h*mb_y2 + block_h/2;
798
799         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
800                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
801
802         //FIXME find a cleaner/simpler way to skip the outside stuff
803         for(y2= y; y2<0; y2++)
804             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
805         for(y2= h; y2<y+block_h; y2++)
806             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
807         if(x<0){
808             for(y2= y; y2<y+block_h; y2++)
809                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
810         }
811         if(x+block_w > w){
812             for(y2= y; y2<y+block_h; y2++)
813                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
814         }
815
816         assert(block_w== 8 || block_w==16);
817         distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
818     }
819
820     if(plane_index==0){
821         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
822         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
823
824 /* ..RRRr
825  * .RXXx.
826  * .RXXx.
827  * rxxx.
828  */
829         if(merged)
830             rate = get_block_bits(s, mb_x, mb_y, 2);
831         for(i=merged?4:0; i<9; i++){
832             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
833             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
834         }
835     }
836     return distortion + rate*penalty_factor;
837 }
838
839 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
840     const int w= b->width;
841     const int h= b->height;
842     int x, y;
843
844     if(1){
845         int run=0;
846         int run_index=0;
847         int max_index;
848
849         for(y=0; y<h; y++){
850             for(x=0; x<w; x++){
851                 int v, p=0;
852                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
853                 v= src[x + y*stride];
854
855                 if(y){
856                     t= src[x + (y-1)*stride];
857                     if(x){
858                         lt= src[x - 1 + (y-1)*stride];
859                     }
860                     if(x + 1 < w){
861                         rt= src[x + 1 + (y-1)*stride];
862                     }
863                 }
864                 if(x){
865                     l= src[x - 1 + y*stride];
866                     /*if(x > 1){
867                         if(orientation==1) ll= src[y + (x-2)*stride];
868                         else               ll= src[x - 2 + y*stride];
869                     }*/
870                 }
871                 if(parent){
872                     int px= x>>1;
873                     int py= y>>1;
874                     if(px<b->parent->width && py<b->parent->height)
875                         p= parent[px + py*2*stride];
876                 }
877                 if(!(/*ll|*/l|lt|t|rt|p)){
878                     if(v){
879                         s->runs[run_index++]= run;
880                         run=0;
881                     }else{
882                         run++;
883                     }
884                 }
885             }
886         }
887         max_index= run_index;
888         s->runs[run_index++]= run;
889         run_index=0;
890         run= s->runs[run_index++];
891
892         put_symbol2(&s->c, b->state[30], max_index, 0);
893         if(run_index <= max_index)
894             put_symbol2(&s->c, b->state[1], run, 3);
895
896         for(y=0; y<h; y++){
897             if(s->c.bytestream_end - s->c.bytestream < w*40){
898                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
899                 return -1;
900             }
901             for(x=0; x<w; x++){
902                 int v, p=0;
903                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
904                 v= src[x + y*stride];
905
906                 if(y){
907                     t= src[x + (y-1)*stride];
908                     if(x){
909                         lt= src[x - 1 + (y-1)*stride];
910                     }
911                     if(x + 1 < w){
912                         rt= src[x + 1 + (y-1)*stride];
913                     }
914                 }
915                 if(x){
916                     l= src[x - 1 + y*stride];
917                     /*if(x > 1){
918                         if(orientation==1) ll= src[y + (x-2)*stride];
919                         else               ll= src[x - 2 + y*stride];
920                     }*/
921                 }
922                 if(parent){
923                     int px= x>>1;
924                     int py= y>>1;
925                     if(px<b->parent->width && py<b->parent->height)
926                         p= parent[px + py*2*stride];
927                 }
928                 if(/*ll|*/l|lt|t|rt|p){
929                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
930
931                     put_rac(&s->c, &b->state[0][context], !!v);
932                 }else{
933                     if(!run){
934                         run= s->runs[run_index++];
935
936                         if(run_index <= max_index)
937                             put_symbol2(&s->c, b->state[1], run, 3);
938                         assert(v);
939                     }else{
940                         run--;
941                         assert(!v);
942                     }
943                 }
944                 if(v){
945                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
946                     int l2= 2*FFABS(l) + (l<0);
947                     int t2= 2*FFABS(t) + (t<0);
948
949                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
950                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
951                 }
952             }
953         }
954     }
955     return 0;
956 }
957
958 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
959 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
960 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
961     return encode_subband_c0run(s, b, src, parent, stride, orientation);
962 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
963 }
964
965 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){
966     const int b_stride= s->b_width << s->block_max_depth;
967     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
968     BlockNode backup= *block;
969     unsigned value;
970     int rd, index;
971
972     assert(mb_x>=0 && mb_y>=0);
973     assert(mb_x<b_stride);
974
975     if(intra){
976         block->color[0] = p[0];
977         block->color[1] = p[1];
978         block->color[2] = p[2];
979         block->type |= BLOCK_INTRA;
980     }else{
981         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
982         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
983         if(s->me_cache[index] == value)
984             return 0;
985         s->me_cache[index]= value;
986
987         block->mx= p[0];
988         block->my= p[1];
989         block->type &= ~BLOCK_INTRA;
990     }
991
992     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
993
994 //FIXME chroma
995     if(rd < *best_rd){
996         *best_rd= rd;
997         return 1;
998     }else{
999         *block= backup;
1000         return 0;
1001     }
1002 }
1003
1004 /* special case for int[2] args we discard afterwards,
1005  * fixes compilation problem with gcc 2.95 */
1006 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){
1007     int p[2] = {p0, p1};
1008     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
1009 }
1010
1011 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){
1012     const int b_stride= s->b_width << s->block_max_depth;
1013     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
1014     BlockNode backup[4]= {block[0], block[1], block[b_stride], block[b_stride+1]};
1015     unsigned value;
1016     int rd, index;
1017
1018     assert(mb_x>=0 && mb_y>=0);
1019     assert(mb_x<b_stride);
1020     assert(((mb_x|mb_y)&1) == 0);
1021
1022     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
1023     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
1024     if(s->me_cache[index] == value)
1025         return 0;
1026     s->me_cache[index]= value;
1027
1028     block->mx= p0;
1029     block->my= p1;
1030     block->ref= ref;
1031     block->type &= ~BLOCK_INTRA;
1032     block[1]= block[b_stride]= block[b_stride+1]= *block;
1033
1034     rd= get_4block_rd(s, mb_x, mb_y, 0);
1035
1036 //FIXME chroma
1037     if(rd < *best_rd){
1038         *best_rd= rd;
1039         return 1;
1040     }else{
1041         block[0]= backup[0];
1042         block[1]= backup[1];
1043         block[b_stride]= backup[2];
1044         block[b_stride+1]= backup[3];
1045         return 0;
1046     }
1047 }
1048
1049 static void iterative_me(SnowContext *s){
1050     int pass, mb_x, mb_y;
1051     const int b_width = s->b_width  << s->block_max_depth;
1052     const int b_height= s->b_height << s->block_max_depth;
1053     const int b_stride= b_width;
1054     int color[3];
1055
1056     {
1057         RangeCoder r = s->c;
1058         uint8_t state[sizeof(s->block_state)];
1059         memcpy(state, s->block_state, sizeof(s->block_state));
1060         for(mb_y= 0; mb_y<s->b_height; mb_y++)
1061             for(mb_x= 0; mb_x<s->b_width; mb_x++)
1062                 encode_q_branch(s, 0, mb_x, mb_y);
1063         s->c = r;
1064         memcpy(s->block_state, state, sizeof(s->block_state));
1065     }
1066
1067     for(pass=0; pass<25; pass++){
1068         int change= 0;
1069
1070         for(mb_y= 0; mb_y<b_height; mb_y++){
1071             for(mb_x= 0; mb_x<b_width; mb_x++){
1072                 int dia_change, i, j, ref;
1073                 int best_rd= INT_MAX, ref_rd;
1074                 BlockNode backup, ref_b;
1075                 const int index= mb_x + mb_y * b_stride;
1076                 BlockNode *block= &s->block[index];
1077                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1078                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1079                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1080                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1081                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1082                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1083                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1084                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1085                 const int b_w= (MB_SIZE >> s->block_max_depth);
1086                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1087
1088                 if(pass && (block->type & BLOCK_OPT))
1089                     continue;
1090                 block->type |= BLOCK_OPT;
1091
1092                 backup= *block;
1093
1094                 if(!s->me_cache_generation)
1095                     memset(s->me_cache, 0, sizeof(s->me_cache));
1096                 s->me_cache_generation += 1<<22;
1097
1098                 //FIXME precalculate
1099                 {
1100                     int x, y;
1101                     for (y = 0; y < b_w * 2; y++)
1102                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1103                     if(mb_x==0)
1104                         for(y=0; y<b_w*2; y++)
1105                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1106                     if(mb_x==b_stride-1)
1107                         for(y=0; y<b_w*2; y++)
1108                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1109                     if(mb_y==0){
1110                         for(x=0; x<b_w*2; x++)
1111                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1112                         for(y=1; y<b_w; y++)
1113                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1114                     }
1115                     if(mb_y==b_height-1){
1116                         for(x=0; x<b_w*2; x++)
1117                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1118                         for(y=b_w; y<b_w*2-1; y++)
1119                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1120                     }
1121                 }
1122
1123                 //skip stuff outside the picture
1124                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1125                     uint8_t *src= s->  input_picture.data[0];
1126                     uint8_t *dst= s->current_picture.data[0];
1127                     const int stride= s->current_picture.linesize[0];
1128                     const int block_w= MB_SIZE >> s->block_max_depth;
1129                     const int block_h= MB_SIZE >> s->block_max_depth;
1130                     const int sx= block_w*mb_x - block_w/2;
1131                     const int sy= block_h*mb_y - block_h/2;
1132                     const int w= s->plane[0].width;
1133                     const int h= s->plane[0].height;
1134                     int y;
1135
1136                     for(y=sy; y<0; y++)
1137                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1138                     for(y=h; y<sy+block_h*2; y++)
1139                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1140                     if(sx<0){
1141                         for(y=sy; y<sy+block_h*2; y++)
1142                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1143                     }
1144                     if(sx+block_w*2 > w){
1145                         for(y=sy; y<sy+block_h*2; y++)
1146                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1147                     }
1148                 }
1149
1150                 // intra(black) = neighbors' contribution to the current block
1151                 for(i=0; i<3; i++)
1152                     color[i]= get_dc(s, mb_x, mb_y, i);
1153
1154                 // get previous score (cannot be cached due to OBMC)
1155                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1156                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1157                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1158                 }else
1159                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1160
1161                 ref_b= *block;
1162                 ref_rd= best_rd;
1163                 for(ref=0; ref < s->ref_frames; ref++){
1164                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1165                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1166                         continue;
1167                     block->ref= ref;
1168                     best_rd= INT_MAX;
1169
1170                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1171                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1172                     if(tb)
1173                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1174                     if(lb)
1175                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1176                     if(rb)
1177                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1178                     if(bb)
1179                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1180
1181                     /* fullpel ME */
1182                     //FIXME avoid subpel interpolation / round to nearest integer
1183                     do{
1184                         dia_change=0;
1185                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1186                             for(j=0; j<i; j++){
1187                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1188                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1189                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1190                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1191                             }
1192                         }
1193                     }while(dia_change);
1194                     /* subpel ME */
1195                     do{
1196                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1197                         dia_change=0;
1198                         for(i=0; i<8; i++)
1199                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1200                     }while(dia_change);
1201                     //FIXME or try the standard 2 pass qpel or similar
1202
1203                     mvr[0][0]= block->mx;
1204                     mvr[0][1]= block->my;
1205                     if(ref_rd > best_rd){
1206                         ref_rd= best_rd;
1207                         ref_b= *block;
1208                     }
1209                 }
1210                 best_rd= ref_rd;
1211                 *block= ref_b;
1212                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1213                 //FIXME RD style color selection
1214                 if(!same_block(block, &backup)){
1215                     if(tb ) tb ->type &= ~BLOCK_OPT;
1216                     if(lb ) lb ->type &= ~BLOCK_OPT;
1217                     if(rb ) rb ->type &= ~BLOCK_OPT;
1218                     if(bb ) bb ->type &= ~BLOCK_OPT;
1219                     if(tlb) tlb->type &= ~BLOCK_OPT;
1220                     if(trb) trb->type &= ~BLOCK_OPT;
1221                     if(blb) blb->type &= ~BLOCK_OPT;
1222                     if(brb) brb->type &= ~BLOCK_OPT;
1223                     change ++;
1224                 }
1225             }
1226         }
1227         av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1228         if(!change)
1229             break;
1230     }
1231
1232     if(s->block_max_depth == 1){
1233         int change= 0;
1234         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1235             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1236                 int i;
1237                 int best_rd, init_rd;
1238                 const int index= mb_x + mb_y * b_stride;
1239                 BlockNode *b[4];
1240
1241                 b[0]= &s->block[index];
1242                 b[1]= b[0]+1;
1243                 b[2]= b[0]+b_stride;
1244                 b[3]= b[2]+1;
1245                 if(same_block(b[0], b[1]) &&
1246                    same_block(b[0], b[2]) &&
1247                    same_block(b[0], b[3]))
1248                     continue;
1249
1250                 if(!s->me_cache_generation)
1251                     memset(s->me_cache, 0, sizeof(s->me_cache));
1252                 s->me_cache_generation += 1<<22;
1253
1254                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1255
1256                 //FIXME more multiref search?
1257                 check_4block_inter(s, mb_x, mb_y,
1258                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1259                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1260
1261                 for(i=0; i<4; i++)
1262                     if(!(b[i]->type&BLOCK_INTRA))
1263                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1264
1265                 if(init_rd != best_rd)
1266                     change++;
1267             }
1268         }
1269         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1270     }
1271 }
1272
1273 static void encode_blocks(SnowContext *s, int search){
1274     int x, y;
1275     int w= s->b_width;
1276     int h= s->b_height;
1277
1278     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1279         iterative_me(s);
1280
1281     for(y=0; y<h; y++){
1282         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1283             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1284             return;
1285         }
1286         for(x=0; x<w; x++){
1287             if(s->avctx->me_method == ME_ITER || !search)
1288                 encode_q_branch2(s, 0, x, y);
1289             else
1290                 encode_q_branch (s, 0, x, y);
1291         }
1292     }
1293 }
1294
1295 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1296     const int w= b->width;
1297     const int h= b->height;
1298     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1299     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1300     int x,y, thres1, thres2;
1301
1302     if(s->qlog == LOSSLESS_QLOG){
1303         for(y=0; y<h; y++)
1304             for(x=0; x<w; x++)
1305                 dst[x + y*stride]= src[x + y*stride];
1306         return;
1307     }
1308
1309     bias= bias ? 0 : (3*qmul)>>3;
1310     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1311     thres2= 2*thres1;
1312
1313     if(!bias){
1314         for(y=0; y<h; y++){
1315             for(x=0; x<w; x++){
1316                 int i= src[x + y*stride];
1317
1318                 if((unsigned)(i+thres1) > thres2){
1319                     if(i>=0){
1320                         i<<= QEXPSHIFT;
1321                         i/= qmul; //FIXME optimize
1322                         dst[x + y*stride]=  i;
1323                     }else{
1324                         i= -i;
1325                         i<<= QEXPSHIFT;
1326                         i/= qmul; //FIXME optimize
1327                         dst[x + y*stride]= -i;
1328                     }
1329                 }else
1330                     dst[x + y*stride]= 0;
1331             }
1332         }
1333     }else{
1334         for(y=0; y<h; y++){
1335             for(x=0; x<w; x++){
1336                 int i= src[x + y*stride];
1337
1338                 if((unsigned)(i+thres1) > thres2){
1339                     if(i>=0){
1340                         i<<= QEXPSHIFT;
1341                         i= (i + bias) / qmul; //FIXME optimize
1342                         dst[x + y*stride]=  i;
1343                     }else{
1344                         i= -i;
1345                         i<<= QEXPSHIFT;
1346                         i= (i + bias) / qmul; //FIXME optimize
1347                         dst[x + y*stride]= -i;
1348                     }
1349                 }else
1350                     dst[x + y*stride]= 0;
1351             }
1352         }
1353     }
1354 }
1355
1356 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1357     const int w= b->width;
1358     const int h= b->height;
1359     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1360     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1361     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1362     int x,y;
1363
1364     if(s->qlog == LOSSLESS_QLOG) return;
1365
1366     for(y=0; y<h; y++){
1367         for(x=0; x<w; x++){
1368             int i= src[x + y*stride];
1369             if(i<0){
1370                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1371             }else if(i>0){
1372                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1373             }
1374         }
1375     }
1376 }
1377
1378 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1379     const int w= b->width;
1380     const int h= b->height;
1381     int x,y;
1382
1383     for(y=h-1; y>=0; y--){
1384         for(x=w-1; x>=0; x--){
1385             int i= x + y*stride;
1386
1387             if(x){
1388                 if(use_median){
1389                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1390                     else  src[i] -= src[i - 1];
1391                 }else{
1392                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1393                     else  src[i] -= src[i - 1];
1394                 }
1395             }else{
1396                 if(y) src[i] -= src[i - stride];
1397             }
1398         }
1399     }
1400 }
1401
1402 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1403     const int w= b->width;
1404     const int h= b->height;
1405     int x,y;
1406
1407     for(y=0; y<h; y++){
1408         for(x=0; x<w; x++){
1409             int i= x + y*stride;
1410
1411             if(x){
1412                 if(use_median){
1413                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1414                     else  src[i] += src[i - 1];
1415                 }else{
1416                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1417                     else  src[i] += src[i - 1];
1418                 }
1419             }else{
1420                 if(y) src[i] += src[i - stride];
1421             }
1422         }
1423     }
1424 }
1425
1426 static void encode_qlogs(SnowContext *s){
1427     int plane_index, level, orientation;
1428
1429     for(plane_index=0; plane_index<2; plane_index++){
1430         for(level=0; level<s->spatial_decomposition_count; level++){
1431             for(orientation=level ? 1:0; orientation<4; orientation++){
1432                 if(orientation==2) continue;
1433                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1434             }
1435         }
1436     }
1437 }
1438
1439 static void encode_header(SnowContext *s){
1440     int plane_index, i;
1441     uint8_t kstate[32];
1442
1443     memset(kstate, MID_STATE, sizeof(kstate));
1444
1445     put_rac(&s->c, kstate, s->keyframe);
1446     if(s->keyframe || s->always_reset){
1447         ff_snow_reset_contexts(s);
1448         s->last_spatial_decomposition_type=
1449         s->last_qlog=
1450         s->last_qbias=
1451         s->last_mv_scale=
1452         s->last_block_max_depth= 0;
1453         for(plane_index=0; plane_index<2; plane_index++){
1454             Plane *p= &s->plane[plane_index];
1455             p->last_htaps=0;
1456             p->last_diag_mc=0;
1457             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1458         }
1459     }
1460     if(s->keyframe){
1461         put_symbol(&s->c, s->header_state, s->version, 0);
1462         put_rac(&s->c, s->header_state, s->always_reset);
1463         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1464         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1465         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1466         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1467         put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1468         put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1469         put_rac(&s->c, s->header_state, s->spatial_scalability);
1470 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1471         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1472
1473         encode_qlogs(s);
1474     }
1475
1476     if(!s->keyframe){
1477         int update_mc=0;
1478         for(plane_index=0; plane_index<2; plane_index++){
1479             Plane *p= &s->plane[plane_index];
1480             update_mc |= p->last_htaps   != p->htaps;
1481             update_mc |= p->last_diag_mc != p->diag_mc;
1482             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1483         }
1484         put_rac(&s->c, s->header_state, update_mc);
1485         if(update_mc){
1486             for(plane_index=0; plane_index<2; plane_index++){
1487                 Plane *p= &s->plane[plane_index];
1488                 put_rac(&s->c, s->header_state, p->diag_mc);
1489                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1490                 for(i= p->htaps/2; i; i--)
1491                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1492             }
1493         }
1494         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1495             put_rac(&s->c, s->header_state, 1);
1496             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1497             encode_qlogs(s);
1498         }else
1499             put_rac(&s->c, s->header_state, 0);
1500     }
1501
1502     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1503     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1504     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1505     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1506     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1507
1508 }
1509
1510 static void update_last_header_values(SnowContext *s){
1511     int plane_index;
1512
1513     if(!s->keyframe){
1514         for(plane_index=0; plane_index<2; plane_index++){
1515             Plane *p= &s->plane[plane_index];
1516             p->last_diag_mc= p->diag_mc;
1517             p->last_htaps  = p->htaps;
1518             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1519         }
1520     }
1521
1522     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1523     s->last_qlog                        = s->qlog;
1524     s->last_qbias                       = s->qbias;
1525     s->last_mv_scale                    = s->mv_scale;
1526     s->last_block_max_depth             = s->block_max_depth;
1527     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1528 }
1529
1530 static int qscale2qlog(int qscale){
1531     return rint(QROOT*log(qscale / (float)FF_QP2LAMBDA)/log(2))
1532            + 61*QROOT/8; ///< 64 > 60
1533 }
1534
1535 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1536 {
1537     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1538      * FIXME we know exact mv bits at this point,
1539      * but ratecontrol isn't set up to include them. */
1540     uint32_t coef_sum= 0;
1541     int level, orientation, delta_qlog;
1542
1543     for(level=0; level<s->spatial_decomposition_count; level++){
1544         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1545             SubBand *b= &s->plane[0].band[level][orientation];
1546             IDWTELEM *buf= b->ibuf;
1547             const int w= b->width;
1548             const int h= b->height;
1549             const int stride= b->stride;
1550             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1551             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1552             const int qdiv= (1<<16)/qmul;
1553             int x, y;
1554             //FIXME this is ugly
1555             for(y=0; y<h; y++)
1556                 for(x=0; x<w; x++)
1557                     buf[x+y*stride]= b->buf[x+y*stride];
1558             if(orientation==0)
1559                 decorrelate(s, b, buf, stride, 1, 0);
1560             for(y=0; y<h; y++)
1561                 for(x=0; x<w; x++)
1562                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1563         }
1564     }
1565
1566     /* ugly, ratecontrol just takes a sqrt again */
1567     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1568     assert(coef_sum < INT_MAX);
1569
1570     if(pict->pict_type == AV_PICTURE_TYPE_I){
1571         s->m.current_picture.mb_var_sum= coef_sum;
1572         s->m.current_picture.mc_mb_var_sum= 0;
1573     }else{
1574         s->m.current_picture.mc_mb_var_sum= coef_sum;
1575         s->m.current_picture.mb_var_sum= 0;
1576     }
1577
1578     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1579     if (pict->quality < 0)
1580         return INT_MIN;
1581     s->lambda= pict->quality * 3/2;
1582     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1583     s->qlog+= delta_qlog;
1584     return delta_qlog;
1585 }
1586
1587 static void calculate_visual_weight(SnowContext *s, Plane *p){
1588     int width = p->width;
1589     int height= p->height;
1590     int level, orientation, x, y;
1591
1592     for(level=0; level<s->spatial_decomposition_count; level++){
1593         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1594             SubBand *b= &p->band[level][orientation];
1595             IDWTELEM *ibuf= b->ibuf;
1596             int64_t error=0;
1597
1598             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1599             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1600             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1601             for(y=0; y<height; y++){
1602                 for(x=0; x<width; x++){
1603                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1604                     error += d*d;
1605                 }
1606             }
1607
1608             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1609         }
1610     }
1611 }
1612
1613 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1614                         const AVFrame *pict, int *got_packet)
1615 {
1616     SnowContext *s = avctx->priv_data;
1617     RangeCoder * const c= &s->c;
1618     AVFrame *pic = &s->new_picture;
1619     const int width= s->avctx->width;
1620     const int height= s->avctx->height;
1621     int level, orientation, plane_index, i, y, ret;
1622     uint8_t rc_header_bak[sizeof(s->header_state)];
1623     uint8_t rc_block_bak[sizeof(s->block_state)];
1624
1625     if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0)
1626         return ret;
1627
1628     ff_init_range_encoder(c, pkt->data, pkt->size);
1629     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1630
1631     for(i=0; i<3; i++){
1632         int hshift= i ? s->chroma_h_shift : 0;
1633         int vshift= i ? s->chroma_v_shift : 0;
1634         for(y=0; y<(height>>vshift); y++)
1635             memcpy(&s->input_picture.data[i][y * s->input_picture.linesize[i]],
1636                    &pict->data[i][y * pict->linesize[i]],
1637                    width>>hshift);
1638     }
1639     s->new_picture = *pict;
1640
1641     s->m.picture_number= avctx->frame_number;
1642     if(avctx->flags&CODEC_FLAG_PASS2){
1643         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1644         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1645         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1646             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1647             if (pic->quality < 0)
1648                 return -1;
1649         }
1650     }else{
1651         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1652         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1653     }
1654
1655     if(s->pass1_rc && avctx->frame_number == 0)
1656         pic->quality = 2*FF_QP2LAMBDA;
1657     if (pic->quality) {
1658         s->qlog   = qscale2qlog(pic->quality);
1659         s->lambda = pic->quality * 3/2;
1660     }
1661     if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1662         s->qlog= LOSSLESS_QLOG;
1663         s->lambda = 0;
1664     }//else keep previous frame's qlog until after motion estimation
1665
1666     ff_snow_frame_start(s);
1667
1668     s->m.current_picture_ptr= &s->m.current_picture;
1669     s->m.last_picture.f.pts = s->m.current_picture.f.pts;
1670     s->m.current_picture.f.pts = pict->pts;
1671     if(pic->pict_type == AV_PICTURE_TYPE_P){
1672         int block_width = (width +15)>>4;
1673         int block_height= (height+15)>>4;
1674         int stride= s->current_picture.linesize[0];
1675
1676         assert(s->current_picture.data[0]);
1677         assert(s->last_picture[0].data[0]);
1678
1679         s->m.avctx= s->avctx;
1680         s->m.current_picture.f.data[0] = s->current_picture.data[0];
1681         s->m.   last_picture.f.data[0] = s->last_picture[0].data[0];
1682         s->m.    new_picture.f.data[0] = s->  input_picture.data[0];
1683         s->m.   last_picture_ptr= &s->m.   last_picture;
1684         s->m.linesize=
1685         s->m.   last_picture.f.linesize[0] =
1686         s->m.    new_picture.f.linesize[0] =
1687         s->m.current_picture.f.linesize[0] = stride;
1688         s->m.uvlinesize= s->current_picture.linesize[1];
1689         s->m.width = width;
1690         s->m.height= height;
1691         s->m.mb_width = block_width;
1692         s->m.mb_height= block_height;
1693         s->m.mb_stride=   s->m.mb_width+1;
1694         s->m.b8_stride= 2*s->m.mb_width+1;
1695         s->m.f_code=1;
1696         s->m.pict_type = pic->pict_type;
1697         s->m.me_method= s->avctx->me_method;
1698         s->m.me.scene_change_score=0;
1699         s->m.flags= s->avctx->flags;
1700         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1701         s->m.out_format= FMT_H263;
1702         s->m.unrestricted_mv= 1;
1703
1704         s->m.lambda = s->lambda;
1705         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1706         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1707
1708         s->m.dsp= s->dsp; //move
1709         ff_init_me(&s->m);
1710         s->dsp= s->m.dsp;
1711     }
1712
1713     if(s->pass1_rc){
1714         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1715         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1716     }
1717
1718 redo_frame:
1719
1720     if (pic->pict_type == AV_PICTURE_TYPE_I)
1721         s->spatial_decomposition_count= 5;
1722     else
1723         s->spatial_decomposition_count= 5;
1724
1725     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1726           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1727         s->spatial_decomposition_count--;
1728
1729     s->m.pict_type = pic->pict_type;
1730     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1731
1732     ff_snow_common_init_after_header(avctx);
1733
1734     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1735         for(plane_index=0; plane_index<3; plane_index++){
1736             calculate_visual_weight(s, &s->plane[plane_index]);
1737         }
1738     }
1739
1740     encode_header(s);
1741     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1742     encode_blocks(s, 1);
1743     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1744
1745     for(plane_index=0; plane_index<3; plane_index++){
1746         Plane *p= &s->plane[plane_index];
1747         int w= p->width;
1748         int h= p->height;
1749         int x, y;
1750 //        int bits= put_bits_count(&s->c.pb);
1751
1752         if (!s->memc_only) {
1753             //FIXME optimize
1754             if(pict->data[plane_index]) //FIXME gray hack
1755                 for(y=0; y<h; y++){
1756                     for(x=0; x<w; x++){
1757                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1758                     }
1759                 }
1760             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1761
1762             if(   plane_index==0
1763                && pic->pict_type == AV_PICTURE_TYPE_P
1764                && !(avctx->flags&CODEC_FLAG_PASS2)
1765                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1766                 ff_init_range_encoder(c, pkt->data, pkt->size);
1767                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1768                 pic->pict_type= AV_PICTURE_TYPE_I;
1769                 s->keyframe=1;
1770                 s->current_picture.key_frame=1;
1771                 goto redo_frame;
1772             }
1773
1774             if(s->qlog == LOSSLESS_QLOG){
1775                 for(y=0; y<h; y++){
1776                     for(x=0; x<w; x++){
1777                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1778                     }
1779                 }
1780             }else{
1781                 for(y=0; y<h; y++){
1782                     for(x=0; x<w; x++){
1783                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1784                     }
1785                 }
1786             }
1787
1788             /*  if(QUANTIZE2)
1789                 dwt_quantize(s, p, s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type);
1790             else*/
1791                 ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1792
1793             if(s->pass1_rc && plane_index==0){
1794                 int delta_qlog = ratecontrol_1pass(s, pic);
1795                 if (delta_qlog <= INT_MIN)
1796                     return -1;
1797                 if(delta_qlog){
1798                     //reordering qlog in the bitstream would eliminate this reset
1799                     ff_init_range_encoder(c, pkt->data, pkt->size);
1800                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1801                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1802                     encode_header(s);
1803                     encode_blocks(s, 0);
1804                 }
1805             }
1806
1807             for(level=0; level<s->spatial_decomposition_count; level++){
1808                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1809                     SubBand *b= &p->band[level][orientation];
1810
1811                     if(!QUANTIZE2)
1812                         quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1813                     if(orientation==0)
1814                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1815                     if (!s->no_bitstream)
1816                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1817                     assert(b->parent==NULL || b->parent->stride == b->stride*2);
1818                     if(orientation==0)
1819                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1820                 }
1821             }
1822
1823             for(level=0; level<s->spatial_decomposition_count; level++){
1824                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1825                     SubBand *b= &p->band[level][orientation];
1826
1827                     dequantize(s, b, b->ibuf, b->stride);
1828                 }
1829             }
1830
1831             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1832             if(s->qlog == LOSSLESS_QLOG){
1833                 for(y=0; y<h; y++){
1834                     for(x=0; x<w; x++){
1835                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1836                     }
1837                 }
1838             }
1839             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1840         }else{
1841             //ME/MC only
1842             if(pic->pict_type == AV_PICTURE_TYPE_I){
1843                 for(y=0; y<h; y++){
1844                     for(x=0; x<w; x++){
1845                         s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]=
1846                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1847                     }
1848                 }
1849             }else{
1850                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1851                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1852             }
1853         }
1854         if(s->avctx->flags&CODEC_FLAG_PSNR){
1855             int64_t error= 0;
1856
1857             if(pict->data[plane_index]) //FIXME gray hack
1858                 for(y=0; y<h; y++){
1859                     for(x=0; x<w; x++){
1860                         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];
1861                         error += d*d;
1862                     }
1863                 }
1864             s->avctx->error[plane_index] += error;
1865             s->current_picture.error[plane_index] = error;
1866         }
1867
1868     }
1869
1870     update_last_header_values(s);
1871
1872     ff_snow_release_buffer(avctx);
1873
1874     s->current_picture.coded_picture_number = avctx->frame_number;
1875     s->current_picture.pict_type = pict->pict_type;
1876     s->current_picture.quality = pict->quality;
1877     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1878     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1879     s->m.current_picture.f.display_picture_number =
1880     s->m.current_picture.f.coded_picture_number   = avctx->frame_number;
1881     s->m.current_picture.f.quality                = pic->quality;
1882     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1883     if(s->pass1_rc)
1884         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1885             return -1;
1886     if(avctx->flags&CODEC_FLAG_PASS1)
1887         ff_write_pass1_stats(&s->m);
1888     s->m.last_pict_type = s->m.pict_type;
1889     avctx->frame_bits = s->m.frame_bits;
1890     avctx->mv_bits = s->m.mv_bits;
1891     avctx->misc_bits = s->m.misc_bits;
1892     avctx->p_tex_bits = s->m.p_tex_bits;
1893
1894     emms_c();
1895
1896     pkt->size = ff_rac_terminate(c);
1897     if (avctx->coded_frame->key_frame)
1898         pkt->flags |= AV_PKT_FLAG_KEY;
1899     *got_packet = 1;
1900
1901     return 0;
1902 }
1903
1904 static av_cold int encode_end(AVCodecContext *avctx)
1905 {
1906     SnowContext *s = avctx->priv_data;
1907
1908     ff_snow_common_end(s);
1909     if (s->input_picture.data[0])
1910         avctx->release_buffer(avctx, &s->input_picture);
1911     av_free(avctx->stats_out);
1912     av_freep(&s->runs);
1913
1914     return 0;
1915 }
1916
1917 #define OFFSET(x) offsetof(SnowContext, x)
1918 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1919 static const AVOption options[] = {
1920     { "memc_only",      "Only do ME/MC (I frames -> ref, P frame -> ME+MC).",   OFFSET(memc_only), AV_OPT_TYPE_INT, { 0 }, 0, 1, VE },
1921     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_INT, { 0 }, 0, 1, VE },
1922     { NULL },
1923 };
1924
1925 static const AVClass snowenc_class = {
1926     .class_name = "snow encoder",
1927     .item_name  = av_default_item_name,
1928     .option     = options,
1929     .version    = LIBAVUTIL_VERSION_INT,
1930 };
1931
1932 AVCodec ff_snow_encoder = {
1933     .name           = "snow",
1934     .type           = AVMEDIA_TYPE_VIDEO,
1935     .id             = CODEC_ID_SNOW,
1936     .priv_data_size = sizeof(SnowContext),
1937     .init           = encode_init,
1938     .encode2        = encode_frame,
1939     .close          = encode_end,
1940     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1941     .priv_class     = &snowenc_class,
1942 };
1943
1944
1945 #ifdef TEST
1946 #undef malloc
1947 #undef free
1948 #undef printf
1949
1950 #include "libavutil/lfg.h"
1951 #include "libavutil/mathematics.h"
1952
1953 int main(void){
1954     int width=256;
1955     int height=256;
1956     int buffer[2][width*height];
1957     SnowContext s;
1958     int i;
1959     AVLFG prng;
1960     s.spatial_decomposition_count=6;
1961     s.spatial_decomposition_type=1;
1962
1963     s.temp_dwt_buffer  = av_mallocz(width * sizeof(DWTELEM));
1964     s.temp_idwt_buffer = av_mallocz(width * sizeof(IDWTELEM));
1965
1966     av_lfg_init(&prng, 1);
1967
1968     printf("testing 5/3 DWT\n");
1969     for(i=0; i<width*height; i++)
1970         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1971
1972     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1973     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1974
1975     for(i=0; i<width*height; i++)
1976         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1977
1978     printf("testing 9/7 DWT\n");
1979     s.spatial_decomposition_type=0;
1980     for(i=0; i<width*height; i++)
1981         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1982
1983     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1984     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1985
1986     for(i=0; i<width*height; i++)
1987         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1988
1989     {
1990     int level, orientation, x, y;
1991     int64_t errors[8][4];
1992     int64_t g=0;
1993
1994         memset(errors, 0, sizeof(errors));
1995         s.spatial_decomposition_count=3;
1996         s.spatial_decomposition_type=0;
1997         for(level=0; level<s.spatial_decomposition_count; level++){
1998             for(orientation=level ? 1 : 0; orientation<4; orientation++){
1999                 int w= width  >> (s.spatial_decomposition_count-level);
2000                 int h= height >> (s.spatial_decomposition_count-level);
2001                 int stride= width  << (s.spatial_decomposition_count-level);
2002                 DWTELEM *buf= buffer[0];
2003                 int64_t error=0;
2004
2005                 if(orientation&1) buf+=w;
2006                 if(orientation>1) buf+=stride>>1;
2007
2008                 memset(buffer[0], 0, sizeof(int)*width*height);
2009                 buf[w/2 + h/2*stride]= 256*256;
2010                 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2011                 for(y=0; y<height; y++){
2012                     for(x=0; x<width; x++){
2013                         int64_t d= buffer[0][x + y*width];
2014                         error += d*d;
2015                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
2016                     }
2017                     if(FFABS(height/2-y)<9 && level==2) printf("\n");
2018                 }
2019                 error= (int)(sqrt(error)+0.5);
2020                 errors[level][orientation]= error;
2021                 if(g) g=av_gcd(g, error);
2022                 else g= error;
2023             }
2024         }
2025         printf("static int const visual_weight[][4]={\n");
2026         for(level=0; level<s.spatial_decomposition_count; level++){
2027             printf("  {");
2028             for(orientation=0; orientation<4; orientation++){
2029                 printf("%8"PRId64",", errors[level][orientation]/g);
2030             }
2031             printf("},\n");
2032         }
2033         printf("};\n");
2034         {
2035             int level=2;
2036             int w= width  >> (s.spatial_decomposition_count-level);
2037             //int h= height >> (s.spatial_decomposition_count-level);
2038             int stride= width  << (s.spatial_decomposition_count-level);
2039             DWTELEM *buf= buffer[0];
2040             int64_t error=0;
2041
2042             buf+=w;
2043             buf+=stride>>1;
2044
2045             memset(buffer[0], 0, sizeof(int)*width*height);
2046             for(y=0; y<height; y++){
2047                 for(x=0; x<width; x++){
2048                     int tab[4]={0,2,3,1};
2049                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2050                 }
2051             }
2052             ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2053             for(y=0; y<height; y++){
2054                 for(x=0; x<width; x++){
2055                     int64_t d= buffer[0][x + y*width];
2056                     error += d*d;
2057                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2058                 }
2059                 if(FFABS(height/2-y)<9) printf("\n");
2060             }
2061         }
2062
2063     }
2064     return 0;
2065 }
2066 #endif /* TEST */