jpegdec: be less picky on padding
[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, AV_GET_BUFFER_FLAG_REF)) < 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 static inline int get_penalty_factor(int lambda, int lambda2, int type){
291     switch(type&0xFF){
292     default:
293     case FF_CMP_SAD:
294         return lambda>>FF_LAMBDA_SHIFT;
295     case FF_CMP_DCT:
296         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
297     case FF_CMP_W53:
298         return (4*lambda)>>(FF_LAMBDA_SHIFT);
299     case FF_CMP_W97:
300         return (2*lambda)>>(FF_LAMBDA_SHIFT);
301     case FF_CMP_SATD:
302     case FF_CMP_DCT264:
303         return (2*lambda)>>FF_LAMBDA_SHIFT;
304     case FF_CMP_RD:
305     case FF_CMP_PSNR:
306     case FF_CMP_SSE:
307     case FF_CMP_NSSE:
308         return lambda2>>FF_LAMBDA_SHIFT;
309     case FF_CMP_BIT:
310         return 1;
311     }
312 }
313
314 //FIXME copy&paste
315 #define P_LEFT P[1]
316 #define P_TOP P[2]
317 #define P_TOPRIGHT P[3]
318 #define P_MEDIAN P[4]
319 #define P_MV1 P[9]
320 #define FLAG_QPEL   1 //must be 1
321
322 static int encode_q_branch(SnowContext *s, int level, int x, int y){
323     uint8_t p_buffer[1024];
324     uint8_t i_buffer[1024];
325     uint8_t p_state[sizeof(s->block_state)];
326     uint8_t i_state[sizeof(s->block_state)];
327     RangeCoder pc, ic;
328     uint8_t *pbbak= s->c.bytestream;
329     uint8_t *pbbak_start= s->c.bytestream_start;
330     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
331     const int w= s->b_width  << s->block_max_depth;
332     const int h= s->b_height << s->block_max_depth;
333     const int rem_depth= s->block_max_depth - level;
334     const int index= (x + y*w) << rem_depth;
335     const int block_w= 1<<(LOG2_MB_SIZE - level);
336     int trx= (x+1)<<rem_depth;
337     int try= (y+1)<<rem_depth;
338     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
339     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
340     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
341     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
342     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
343     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
344     int pl = left->color[0];
345     int pcb= left->color[1];
346     int pcr= left->color[2];
347     int pmx, pmy;
348     int mx=0, my=0;
349     int l,cr,cb;
350     const int stride= s->current_picture.linesize[0];
351     const int uvstride= s->current_picture.linesize[1];
352     uint8_t *current_data[3]= { s->input_picture.data[0] + (x + y*  stride)*block_w,
353                                 s->input_picture.data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
354                                 s->input_picture.data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
355     int P[10][2];
356     int16_t last_mv[3][2];
357     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
358     const int shift= 1+qpel;
359     MotionEstContext *c= &s->m.me;
360     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
361     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
362     int my_context= av_log2(2*FFABS(left->my - top->my));
363     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
364     int ref, best_ref, ref_score, ref_mx, ref_my;
365
366     assert(sizeof(s->block_state) >= 256);
367     if(s->keyframe){
368         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
369         return 0;
370     }
371
372 //    clip predictors / edge ?
373
374     P_LEFT[0]= left->mx;
375     P_LEFT[1]= left->my;
376     P_TOP [0]= top->mx;
377     P_TOP [1]= top->my;
378     P_TOPRIGHT[0]= tr->mx;
379     P_TOPRIGHT[1]= tr->my;
380
381     last_mv[0][0]= s->block[index].mx;
382     last_mv[0][1]= s->block[index].my;
383     last_mv[1][0]= right->mx;
384     last_mv[1][1]= right->my;
385     last_mv[2][0]= bottom->mx;
386     last_mv[2][1]= bottom->my;
387
388     s->m.mb_stride=2;
389     s->m.mb_x=
390     s->m.mb_y= 0;
391     c->skip= 0;
392
393     assert(c->  stride ==   stride);
394     assert(c->uvstride == uvstride);
395
396     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
397     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
398     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
399     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
400
401     c->xmin = - x*block_w - 16+3;
402     c->ymin = - y*block_w - 16+3;
403     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
404     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
405
406     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
407     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
408     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
409     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
410     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
411     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
412     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
413
414     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
415     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
416
417     if (!y) {
418         c->pred_x= P_LEFT[0];
419         c->pred_y= P_LEFT[1];
420     } else {
421         c->pred_x = P_MEDIAN[0];
422         c->pred_y = P_MEDIAN[1];
423     }
424
425     score= INT_MAX;
426     best_ref= 0;
427     for(ref=0; ref<s->ref_frames; ref++){
428         init_ref(c, current_data, s->last_picture[ref].data, NULL, block_w*x, block_w*y, 0);
429
430         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
431                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
432
433         assert(ref_mx >= c->xmin);
434         assert(ref_mx <= c->xmax);
435         assert(ref_my >= c->ymin);
436         assert(ref_my <= c->ymax);
437
438         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
439         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
440         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
441         if(s->ref_mvs[ref]){
442             s->ref_mvs[ref][index][0]= ref_mx;
443             s->ref_mvs[ref][index][1]= ref_my;
444             s->ref_scores[ref][index]= ref_score;
445         }
446         if(score > ref_score){
447             score= ref_score;
448             best_ref= ref;
449             mx= ref_mx;
450             my= ref_my;
451         }
452     }
453     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
454
455   //  subpel search
456     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
457     pc= s->c;
458     pc.bytestream_start=
459     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
460     memcpy(p_state, s->block_state, sizeof(s->block_state));
461
462     if(level!=s->block_max_depth)
463         put_rac(&pc, &p_state[4 + s_context], 1);
464     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
465     if(s->ref_frames > 1)
466         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
467     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
468     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
469     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
470     p_len= pc.bytestream - pc.bytestream_start;
471     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
472
473     block_s= block_w*block_w;
474     sum = pix_sum(current_data[0], stride, block_w, block_w);
475     l= (sum + block_s/2)/block_s;
476     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
477
478     block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
479     sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
480     cb= (sum + block_s/2)/block_s;
481 //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
482     sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
483     cr= (sum + block_s/2)/block_s;
484 //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
485
486     ic= s->c;
487     ic.bytestream_start=
488     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
489     memcpy(i_state, s->block_state, sizeof(s->block_state));
490     if(level!=s->block_max_depth)
491         put_rac(&ic, &i_state[4 + s_context], 1);
492     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
493     put_symbol(&ic, &i_state[32],  l-pl , 1);
494     put_symbol(&ic, &i_state[64], cb-pcb, 1);
495     put_symbol(&ic, &i_state[96], cr-pcr, 1);
496     i_len= ic.bytestream - ic.bytestream_start;
497     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
498
499 //    assert(score==256*256*256*64-1);
500     assert(iscore < 255*255*256 + s->lambda2*10);
501     assert(iscore >= 0);
502     assert(l>=0 && l<=255);
503     assert(pl>=0 && pl<=255);
504
505     if(level==0){
506         int varc= iscore >> 8;
507         int vard= score >> 8;
508         if (vard <= 64 || vard < varc)
509             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
510         else
511             c->scene_change_score+= s->m.qscale;
512     }
513
514     if(level!=s->block_max_depth){
515         put_rac(&s->c, &s->block_state[4 + s_context], 0);
516         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
517         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
518         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
519         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
520         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
521
522         if(score2 < score && score2 < iscore)
523             return score2;
524     }
525
526     if(iscore < score){
527         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
528         memcpy(pbbak, i_buffer, i_len);
529         s->c= ic;
530         s->c.bytestream_start= pbbak_start;
531         s->c.bytestream= pbbak + i_len;
532         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
533         memcpy(s->block_state, i_state, sizeof(s->block_state));
534         return iscore;
535     }else{
536         memcpy(pbbak, p_buffer, p_len);
537         s->c= pc;
538         s->c.bytestream_start= pbbak_start;
539         s->c.bytestream= pbbak + p_len;
540         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
541         memcpy(s->block_state, p_state, sizeof(s->block_state));
542         return score;
543     }
544 }
545
546 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
547     const int w= s->b_width  << s->block_max_depth;
548     const int rem_depth= s->block_max_depth - level;
549     const int index= (x + y*w) << rem_depth;
550     int trx= (x+1)<<rem_depth;
551     BlockNode *b= &s->block[index];
552     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
553     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
554     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
555     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
556     int pl = left->color[0];
557     int pcb= left->color[1];
558     int pcr= left->color[2];
559     int pmx, pmy;
560     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
561     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
562     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
563     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
564
565     if(s->keyframe){
566         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
567         return;
568     }
569
570     if(level!=s->block_max_depth){
571         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
572             put_rac(&s->c, &s->block_state[4 + s_context], 1);
573         }else{
574             put_rac(&s->c, &s->block_state[4 + s_context], 0);
575             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
576             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
577             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
578             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
579             return;
580         }
581     }
582     if(b->type & BLOCK_INTRA){
583         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
584         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
585         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
586         put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
587         put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
588         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
589     }else{
590         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
591         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
592         if(s->ref_frames > 1)
593             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
594         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
595         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
596         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
597     }
598 }
599
600 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
601     int i, x2, y2;
602     Plane *p= &s->plane[plane_index];
603     const int block_size = MB_SIZE >> s->block_max_depth;
604     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
605     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
606     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
607     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
608     const int ref_stride= s->current_picture.linesize[plane_index];
609     uint8_t *src= s-> input_picture.data[plane_index];
610     IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
611     const int b_stride = s->b_width << s->block_max_depth;
612     const int w= p->width;
613     const int h= p->height;
614     int index= mb_x + mb_y*b_stride;
615     BlockNode *b= &s->block[index];
616     BlockNode backup= *b;
617     int ab=0;
618     int aa=0;
619
620     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
621
622     b->type|= BLOCK_INTRA;
623     b->color[plane_index]= 0;
624     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
625
626     for(i=0; i<4; i++){
627         int mb_x2= mb_x + (i &1) - 1;
628         int mb_y2= mb_y + (i>>1) - 1;
629         int x= block_w*mb_x2 + block_w/2;
630         int y= block_h*mb_y2 + block_h/2;
631
632         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
633                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
634
635         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
636             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
637                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
638                 int obmc_v= obmc[index];
639                 int d;
640                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
641                 if(x<0) obmc_v += obmc[index + block_w];
642                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
643                 if(x+block_w>w) obmc_v += obmc[index - block_w];
644                 //FIXME precalculate this or simplify it somehow else
645
646                 d = -dst[index] + (1<<(FRAC_BITS-1));
647                 dst[index] = d;
648                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
649                 aa += obmc_v * obmc_v; //FIXME precalculate this
650             }
651         }
652     }
653     *b= backup;
654
655     return av_clip( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa), 0, 255); //FIXME we should not need clipping
656 }
657
658 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
659     const int b_stride = s->b_width << s->block_max_depth;
660     const int b_height = s->b_height<< s->block_max_depth;
661     int index= x + y*b_stride;
662     const BlockNode *b     = &s->block[index];
663     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
664     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
665     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
666     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
667     int dmx, dmy;
668 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
669 //  int my_context= av_log2(2*FFABS(left->my - top->my));
670
671     if(x<0 || x>=b_stride || y>=b_height)
672         return 0;
673 /*
674 1            0      0
675 01X          1-2    1
676 001XX        3-6    2-3
677 0001XXX      7-14   4-7
678 00001XXXX   15-30   8-15
679 */
680 //FIXME try accurate rate
681 //FIXME intra and inter predictors if surrounding blocks are not the same type
682     if(b->type & BLOCK_INTRA){
683         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
684                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
685                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
686     }else{
687         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
688         dmx-= b->mx;
689         dmy-= b->my;
690         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
691                     + av_log2(2*FFABS(dmy))
692                     + av_log2(2*b->ref));
693     }
694 }
695
696 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
697     Plane *p= &s->plane[plane_index];
698     const int block_size = MB_SIZE >> s->block_max_depth;
699     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
700     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
701     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
702     const int ref_stride= s->current_picture.linesize[plane_index];
703     uint8_t *dst= s->current_picture.data[plane_index];
704     uint8_t *src= s->  input_picture.data[plane_index];
705     IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
706     uint8_t *cur = s->scratchbuf;
707     uint8_t *tmp = s->emu_edge_buffer;
708     const int b_stride = s->b_width << s->block_max_depth;
709     const int b_height = s->b_height<< s->block_max_depth;
710     const int w= p->width;
711     const int h= p->height;
712     int distortion;
713     int rate= 0;
714     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
715     int sx= block_w*mb_x - block_w/2;
716     int sy= block_h*mb_y - block_h/2;
717     int x0= FFMAX(0,-sx);
718     int y0= FFMAX(0,-sy);
719     int x1= FFMIN(block_w*2, w-sx);
720     int y1= FFMIN(block_h*2, h-sy);
721     int i,x,y;
722
723     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
724
725     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);
726
727     for(y=y0; y<y1; y++){
728         const uint8_t *obmc1= obmc_edged[y];
729         const IDWTELEM *pred1 = pred + y*obmc_stride;
730         uint8_t *cur1 = cur + y*ref_stride;
731         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
732         for(x=x0; x<x1; x++){
733 #if FRAC_BITS >= LOG2_OBMC_MAX
734             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
735 #else
736             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
737 #endif
738             v = (v + pred1[x]) >> FRAC_BITS;
739             if(v&(~255)) v= ~(v>>31);
740             dst1[x] = v;
741         }
742     }
743
744     /* copy the regions where obmc[] = (uint8_t)256 */
745     if(LOG2_OBMC_MAX == 8
746         && (mb_x == 0 || mb_x == b_stride-1)
747         && (mb_y == 0 || mb_y == b_height-1)){
748         if(mb_x == 0)
749             x1 = block_w;
750         else
751             x0 = block_w;
752         if(mb_y == 0)
753             y1 = block_h;
754         else
755             y0 = block_h;
756         for(y=y0; y<y1; y++)
757             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
758     }
759
760     if(block_w==16){
761         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
762         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
763         /* FIXME cmps overlap but do not cover the wavelet's whole support.
764          * So improving the score of one block is not strictly guaranteed
765          * to improve the score of the whole frame, thus iterative motion
766          * estimation does not always converge. */
767         if(s->avctx->me_cmp == FF_CMP_W97)
768             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
769         else if(s->avctx->me_cmp == FF_CMP_W53)
770             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
771         else{
772             distortion = 0;
773             for(i=0; i<4; i++){
774                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
775                 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
776             }
777         }
778     }else{
779         assert(block_w==8);
780         distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
781     }
782
783     if(plane_index==0){
784         for(i=0; i<4; i++){
785 /* ..RRr
786  * .RXx.
787  * rxx..
788  */
789             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
790         }
791         if(mb_x == b_stride-2)
792             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
793     }
794     return distortion + rate*penalty_factor;
795 }
796
797 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
798     int i, y2;
799     Plane *p= &s->plane[plane_index];
800     const int block_size = MB_SIZE >> s->block_max_depth;
801     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
802     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
803     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
804     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
805     const int ref_stride= s->current_picture.linesize[plane_index];
806     uint8_t *dst= s->current_picture.data[plane_index];
807     uint8_t *src= s-> input_picture.data[plane_index];
808     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
809     // const has only been removed from zero_dst to suppress a warning
810     static IDWTELEM zero_dst[4096]; //FIXME
811     const int b_stride = s->b_width << s->block_max_depth;
812     const int w= p->width;
813     const int h= p->height;
814     int distortion= 0;
815     int rate= 0;
816     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
817
818     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
819
820     for(i=0; i<9; i++){
821         int mb_x2= mb_x + (i%3) - 1;
822         int mb_y2= mb_y + (i/3) - 1;
823         int x= block_w*mb_x2 + block_w/2;
824         int y= block_h*mb_y2 + block_h/2;
825
826         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
827                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
828
829         //FIXME find a cleaner/simpler way to skip the outside stuff
830         for(y2= y; y2<0; y2++)
831             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
832         for(y2= h; y2<y+block_h; y2++)
833             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
834         if(x<0){
835             for(y2= y; y2<y+block_h; y2++)
836                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
837         }
838         if(x+block_w > w){
839             for(y2= y; y2<y+block_h; y2++)
840                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
841         }
842
843         assert(block_w== 8 || block_w==16);
844         distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
845     }
846
847     if(plane_index==0){
848         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
849         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
850
851 /* ..RRRr
852  * .RXXx.
853  * .RXXx.
854  * rxxx.
855  */
856         if(merged)
857             rate = get_block_bits(s, mb_x, mb_y, 2);
858         for(i=merged?4:0; i<9; i++){
859             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
860             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
861         }
862     }
863     return distortion + rate*penalty_factor;
864 }
865
866 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
867     const int w= b->width;
868     const int h= b->height;
869     int x, y;
870
871     if(1){
872         int run=0;
873         int *runs = s->run_buffer;
874         int run_index=0;
875         int max_index;
876
877         for(y=0; y<h; y++){
878             for(x=0; x<w; x++){
879                 int v, p=0;
880                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
881                 v= src[x + y*stride];
882
883                 if(y){
884                     t= src[x + (y-1)*stride];
885                     if(x){
886                         lt= src[x - 1 + (y-1)*stride];
887                     }
888                     if(x + 1 < w){
889                         rt= src[x + 1 + (y-1)*stride];
890                     }
891                 }
892                 if(x){
893                     l= src[x - 1 + y*stride];
894                     /*if(x > 1){
895                         if(orientation==1) ll= src[y + (x-2)*stride];
896                         else               ll= src[x - 2 + y*stride];
897                     }*/
898                 }
899                 if(parent){
900                     int px= x>>1;
901                     int py= y>>1;
902                     if(px<b->parent->width && py<b->parent->height)
903                         p= parent[px + py*2*stride];
904                 }
905                 if(!(/*ll|*/l|lt|t|rt|p)){
906                     if(v){
907                         runs[run_index++]= run;
908                         run=0;
909                     }else{
910                         run++;
911                     }
912                 }
913             }
914         }
915         max_index= run_index;
916         runs[run_index++]= run;
917         run_index=0;
918         run= runs[run_index++];
919
920         put_symbol2(&s->c, b->state[30], max_index, 0);
921         if(run_index <= max_index)
922             put_symbol2(&s->c, b->state[1], run, 3);
923
924         for(y=0; y<h; y++){
925             if(s->c.bytestream_end - s->c.bytestream < w*40){
926                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
927                 return -1;
928             }
929             for(x=0; x<w; x++){
930                 int v, p=0;
931                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
932                 v= src[x + y*stride];
933
934                 if(y){
935                     t= src[x + (y-1)*stride];
936                     if(x){
937                         lt= src[x - 1 + (y-1)*stride];
938                     }
939                     if(x + 1 < w){
940                         rt= src[x + 1 + (y-1)*stride];
941                     }
942                 }
943                 if(x){
944                     l= src[x - 1 + y*stride];
945                     /*if(x > 1){
946                         if(orientation==1) ll= src[y + (x-2)*stride];
947                         else               ll= src[x - 2 + y*stride];
948                     }*/
949                 }
950                 if(parent){
951                     int px= x>>1;
952                     int py= y>>1;
953                     if(px<b->parent->width && py<b->parent->height)
954                         p= parent[px + py*2*stride];
955                 }
956                 if(/*ll|*/l|lt|t|rt|p){
957                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
958
959                     put_rac(&s->c, &b->state[0][context], !!v);
960                 }else{
961                     if(!run){
962                         run= runs[run_index++];
963
964                         if(run_index <= max_index)
965                             put_symbol2(&s->c, b->state[1], run, 3);
966                         assert(v);
967                     }else{
968                         run--;
969                         assert(!v);
970                     }
971                 }
972                 if(v){
973                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
974                     int l2= 2*FFABS(l) + (l<0);
975                     int t2= 2*FFABS(t) + (t<0);
976
977                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
978                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
979                 }
980             }
981         }
982     }
983     return 0;
984 }
985
986 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
987 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
988 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
989     return encode_subband_c0run(s, b, src, parent, stride, orientation);
990 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
991 }
992
993 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){
994     const int b_stride= s->b_width << s->block_max_depth;
995     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
996     BlockNode backup= *block;
997     unsigned value;
998     int rd, index;
999
1000     assert(mb_x>=0 && mb_y>=0);
1001     assert(mb_x<b_stride);
1002
1003     if(intra){
1004         block->color[0] = p[0];
1005         block->color[1] = p[1];
1006         block->color[2] = p[2];
1007         block->type |= BLOCK_INTRA;
1008     }else{
1009         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
1010         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
1011         if(s->me_cache[index] == value)
1012             return 0;
1013         s->me_cache[index]= value;
1014
1015         block->mx= p[0];
1016         block->my= p[1];
1017         block->type &= ~BLOCK_INTRA;
1018     }
1019
1020     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
1021
1022 //FIXME chroma
1023     if(rd < *best_rd){
1024         *best_rd= rd;
1025         return 1;
1026     }else{
1027         *block= backup;
1028         return 0;
1029     }
1030 }
1031
1032 /* special case for int[2] args we discard afterwards,
1033  * fixes compilation problem with gcc 2.95 */
1034 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){
1035     int p[2] = {p0, p1};
1036     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
1037 }
1038
1039 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){
1040     const int b_stride= s->b_width << s->block_max_depth;
1041     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
1042     BlockNode backup[4];
1043     unsigned value;
1044     int rd, index;
1045
1046     /* We don't initialize backup[] during variable declaration, because
1047      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
1048      * 'int16_t'". */
1049     backup[0] = block[0];
1050     backup[1] = block[1];
1051     backup[2] = block[b_stride];
1052     backup[3] = block[b_stride + 1];
1053
1054     assert(mb_x>=0 && mb_y>=0);
1055     assert(mb_x<b_stride);
1056     assert(((mb_x|mb_y)&1) == 0);
1057
1058     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
1059     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
1060     if(s->me_cache[index] == value)
1061         return 0;
1062     s->me_cache[index]= value;
1063
1064     block->mx= p0;
1065     block->my= p1;
1066     block->ref= ref;
1067     block->type &= ~BLOCK_INTRA;
1068     block[1]= block[b_stride]= block[b_stride+1]= *block;
1069
1070     rd= get_4block_rd(s, mb_x, mb_y, 0);
1071
1072 //FIXME chroma
1073     if(rd < *best_rd){
1074         *best_rd= rd;
1075         return 1;
1076     }else{
1077         block[0]= backup[0];
1078         block[1]= backup[1];
1079         block[b_stride]= backup[2];
1080         block[b_stride+1]= backup[3];
1081         return 0;
1082     }
1083 }
1084
1085 static void iterative_me(SnowContext *s){
1086     int pass, mb_x, mb_y;
1087     const int b_width = s->b_width  << s->block_max_depth;
1088     const int b_height= s->b_height << s->block_max_depth;
1089     const int b_stride= b_width;
1090     int color[3];
1091
1092     {
1093         RangeCoder r = s->c;
1094         uint8_t state[sizeof(s->block_state)];
1095         memcpy(state, s->block_state, sizeof(s->block_state));
1096         for(mb_y= 0; mb_y<s->b_height; mb_y++)
1097             for(mb_x= 0; mb_x<s->b_width; mb_x++)
1098                 encode_q_branch(s, 0, mb_x, mb_y);
1099         s->c = r;
1100         memcpy(s->block_state, state, sizeof(s->block_state));
1101     }
1102
1103     for(pass=0; pass<25; pass++){
1104         int change= 0;
1105
1106         for(mb_y= 0; mb_y<b_height; mb_y++){
1107             for(mb_x= 0; mb_x<b_width; mb_x++){
1108                 int dia_change, i, j, ref;
1109                 int best_rd= INT_MAX, ref_rd;
1110                 BlockNode backup, ref_b;
1111                 const int index= mb_x + mb_y * b_stride;
1112                 BlockNode *block= &s->block[index];
1113                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1114                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1115                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1116                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1117                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1118                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1119                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1120                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1121                 const int b_w= (MB_SIZE >> s->block_max_depth);
1122                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1123
1124                 if(pass && (block->type & BLOCK_OPT))
1125                     continue;
1126                 block->type |= BLOCK_OPT;
1127
1128                 backup= *block;
1129
1130                 if(!s->me_cache_generation)
1131                     memset(s->me_cache, 0, sizeof(s->me_cache));
1132                 s->me_cache_generation += 1<<22;
1133
1134                 //FIXME precalculate
1135                 {
1136                     int x, y;
1137                     for (y = 0; y < b_w * 2; y++)
1138                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1139                     if(mb_x==0)
1140                         for(y=0; y<b_w*2; y++)
1141                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1142                     if(mb_x==b_stride-1)
1143                         for(y=0; y<b_w*2; y++)
1144                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1145                     if(mb_y==0){
1146                         for(x=0; x<b_w*2; x++)
1147                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1148                         for(y=1; y<b_w; y++)
1149                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1150                     }
1151                     if(mb_y==b_height-1){
1152                         for(x=0; x<b_w*2; x++)
1153                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1154                         for(y=b_w; y<b_w*2-1; y++)
1155                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1156                     }
1157                 }
1158
1159                 //skip stuff outside the picture
1160                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1161                     uint8_t *src= s->  input_picture.data[0];
1162                     uint8_t *dst= s->current_picture.data[0];
1163                     const int stride= s->current_picture.linesize[0];
1164                     const int block_w= MB_SIZE >> s->block_max_depth;
1165                     const int block_h= MB_SIZE >> s->block_max_depth;
1166                     const int sx= block_w*mb_x - block_w/2;
1167                     const int sy= block_h*mb_y - block_h/2;
1168                     const int w= s->plane[0].width;
1169                     const int h= s->plane[0].height;
1170                     int y;
1171
1172                     for(y=sy; y<0; y++)
1173                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1174                     for(y=h; y<sy+block_h*2; y++)
1175                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1176                     if(sx<0){
1177                         for(y=sy; y<sy+block_h*2; y++)
1178                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1179                     }
1180                     if(sx+block_w*2 > w){
1181                         for(y=sy; y<sy+block_h*2; y++)
1182                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1183                     }
1184                 }
1185
1186                 // intra(black) = neighbors' contribution to the current block
1187                 for(i=0; i<3; i++)
1188                     color[i]= get_dc(s, mb_x, mb_y, i);
1189
1190                 // get previous score (cannot be cached due to OBMC)
1191                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1192                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1193                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1194                 }else
1195                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1196
1197                 ref_b= *block;
1198                 ref_rd= best_rd;
1199                 for(ref=0; ref < s->ref_frames; ref++){
1200                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1201                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1202                         continue;
1203                     block->ref= ref;
1204                     best_rd= INT_MAX;
1205
1206                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1207                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1208                     if(tb)
1209                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1210                     if(lb)
1211                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1212                     if(rb)
1213                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1214                     if(bb)
1215                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1216
1217                     /* fullpel ME */
1218                     //FIXME avoid subpel interpolation / round to nearest integer
1219                     do{
1220                         dia_change=0;
1221                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1222                             for(j=0; j<i; j++){
1223                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1224                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1225                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1226                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1227                             }
1228                         }
1229                     }while(dia_change);
1230                     /* subpel ME */
1231                     do{
1232                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1233                         dia_change=0;
1234                         for(i=0; i<8; i++)
1235                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1236                     }while(dia_change);
1237                     //FIXME or try the standard 2 pass qpel or similar
1238
1239                     mvr[0][0]= block->mx;
1240                     mvr[0][1]= block->my;
1241                     if(ref_rd > best_rd){
1242                         ref_rd= best_rd;
1243                         ref_b= *block;
1244                     }
1245                 }
1246                 best_rd= ref_rd;
1247                 *block= ref_b;
1248                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1249                 //FIXME RD style color selection
1250                 if(!same_block(block, &backup)){
1251                     if(tb ) tb ->type &= ~BLOCK_OPT;
1252                     if(lb ) lb ->type &= ~BLOCK_OPT;
1253                     if(rb ) rb ->type &= ~BLOCK_OPT;
1254                     if(bb ) bb ->type &= ~BLOCK_OPT;
1255                     if(tlb) tlb->type &= ~BLOCK_OPT;
1256                     if(trb) trb->type &= ~BLOCK_OPT;
1257                     if(blb) blb->type &= ~BLOCK_OPT;
1258                     if(brb) brb->type &= ~BLOCK_OPT;
1259                     change ++;
1260                 }
1261             }
1262         }
1263         av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1264         if(!change)
1265             break;
1266     }
1267
1268     if(s->block_max_depth == 1){
1269         int change= 0;
1270         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1271             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1272                 int i;
1273                 int best_rd, init_rd;
1274                 const int index= mb_x + mb_y * b_stride;
1275                 BlockNode *b[4];
1276
1277                 b[0]= &s->block[index];
1278                 b[1]= b[0]+1;
1279                 b[2]= b[0]+b_stride;
1280                 b[3]= b[2]+1;
1281                 if(same_block(b[0], b[1]) &&
1282                    same_block(b[0], b[2]) &&
1283                    same_block(b[0], b[3]))
1284                     continue;
1285
1286                 if(!s->me_cache_generation)
1287                     memset(s->me_cache, 0, sizeof(s->me_cache));
1288                 s->me_cache_generation += 1<<22;
1289
1290                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1291
1292                 //FIXME more multiref search?
1293                 check_4block_inter(s, mb_x, mb_y,
1294                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1295                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1296
1297                 for(i=0; i<4; i++)
1298                     if(!(b[i]->type&BLOCK_INTRA))
1299                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1300
1301                 if(init_rd != best_rd)
1302                     change++;
1303             }
1304         }
1305         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1306     }
1307 }
1308
1309 static void encode_blocks(SnowContext *s, int search){
1310     int x, y;
1311     int w= s->b_width;
1312     int h= s->b_height;
1313
1314     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1315         iterative_me(s);
1316
1317     for(y=0; y<h; y++){
1318         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1319             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1320             return;
1321         }
1322         for(x=0; x<w; x++){
1323             if(s->avctx->me_method == ME_ITER || !search)
1324                 encode_q_branch2(s, 0, x, y);
1325             else
1326                 encode_q_branch (s, 0, x, y);
1327         }
1328     }
1329 }
1330
1331 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1332     const int w= b->width;
1333     const int h= b->height;
1334     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1335     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1336     int x,y, thres1, thres2;
1337
1338     if(s->qlog == LOSSLESS_QLOG){
1339         for(y=0; y<h; y++)
1340             for(x=0; x<w; x++)
1341                 dst[x + y*stride]= src[x + y*stride];
1342         return;
1343     }
1344
1345     bias= bias ? 0 : (3*qmul)>>3;
1346     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1347     thres2= 2*thres1;
1348
1349     if(!bias){
1350         for(y=0; y<h; y++){
1351             for(x=0; x<w; x++){
1352                 int i= src[x + y*stride];
1353
1354                 if((unsigned)(i+thres1) > thres2){
1355                     if(i>=0){
1356                         i<<= QEXPSHIFT;
1357                         i/= qmul; //FIXME optimize
1358                         dst[x + y*stride]=  i;
1359                     }else{
1360                         i= -i;
1361                         i<<= QEXPSHIFT;
1362                         i/= qmul; //FIXME optimize
1363                         dst[x + y*stride]= -i;
1364                     }
1365                 }else
1366                     dst[x + y*stride]= 0;
1367             }
1368         }
1369     }else{
1370         for(y=0; y<h; y++){
1371             for(x=0; x<w; x++){
1372                 int i= src[x + y*stride];
1373
1374                 if((unsigned)(i+thres1) > thres2){
1375                     if(i>=0){
1376                         i<<= QEXPSHIFT;
1377                         i= (i + bias) / qmul; //FIXME optimize
1378                         dst[x + y*stride]=  i;
1379                     }else{
1380                         i= -i;
1381                         i<<= QEXPSHIFT;
1382                         i= (i + bias) / qmul; //FIXME optimize
1383                         dst[x + y*stride]= -i;
1384                     }
1385                 }else
1386                     dst[x + y*stride]= 0;
1387             }
1388         }
1389     }
1390 }
1391
1392 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1393     const int w= b->width;
1394     const int h= b->height;
1395     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1396     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1397     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1398     int x,y;
1399
1400     if(s->qlog == LOSSLESS_QLOG) return;
1401
1402     for(y=0; y<h; y++){
1403         for(x=0; x<w; x++){
1404             int i= src[x + y*stride];
1405             if(i<0){
1406                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1407             }else if(i>0){
1408                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1409             }
1410         }
1411     }
1412 }
1413
1414 static void decorrelate(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=h-1; y>=0; y--){
1420         for(x=w-1; x>=0; 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 correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1439     const int w= b->width;
1440     const int h= b->height;
1441     int x,y;
1442
1443     for(y=0; y<h; y++){
1444         for(x=0; x<w; x++){
1445             int i= x + y*stride;
1446
1447             if(x){
1448                 if(use_median){
1449                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1450                     else  src[i] += src[i - 1];
1451                 }else{
1452                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1453                     else  src[i] += src[i - 1];
1454                 }
1455             }else{
1456                 if(y) src[i] += src[i - stride];
1457             }
1458         }
1459     }
1460 }
1461
1462 static void encode_qlogs(SnowContext *s){
1463     int plane_index, level, orientation;
1464
1465     for(plane_index=0; plane_index<2; plane_index++){
1466         for(level=0; level<s->spatial_decomposition_count; level++){
1467             for(orientation=level ? 1:0; orientation<4; orientation++){
1468                 if(orientation==2) continue;
1469                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1470             }
1471         }
1472     }
1473 }
1474
1475 static void encode_header(SnowContext *s){
1476     int plane_index, i;
1477     uint8_t kstate[32];
1478
1479     memset(kstate, MID_STATE, sizeof(kstate));
1480
1481     put_rac(&s->c, kstate, s->keyframe);
1482     if(s->keyframe || s->always_reset){
1483         ff_snow_reset_contexts(s);
1484         s->last_spatial_decomposition_type=
1485         s->last_qlog=
1486         s->last_qbias=
1487         s->last_mv_scale=
1488         s->last_block_max_depth= 0;
1489         for(plane_index=0; plane_index<2; plane_index++){
1490             Plane *p= &s->plane[plane_index];
1491             p->last_htaps=0;
1492             p->last_diag_mc=0;
1493             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1494         }
1495     }
1496     if(s->keyframe){
1497         put_symbol(&s->c, s->header_state, s->version, 0);
1498         put_rac(&s->c, s->header_state, s->always_reset);
1499         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1500         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1501         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1502         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1503         put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1504         put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1505         put_rac(&s->c, s->header_state, s->spatial_scalability);
1506 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1507         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1508
1509         encode_qlogs(s);
1510     }
1511
1512     if(!s->keyframe){
1513         int update_mc=0;
1514         for(plane_index=0; plane_index<2; plane_index++){
1515             Plane *p= &s->plane[plane_index];
1516             update_mc |= p->last_htaps   != p->htaps;
1517             update_mc |= p->last_diag_mc != p->diag_mc;
1518             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1519         }
1520         put_rac(&s->c, s->header_state, update_mc);
1521         if(update_mc){
1522             for(plane_index=0; plane_index<2; plane_index++){
1523                 Plane *p= &s->plane[plane_index];
1524                 put_rac(&s->c, s->header_state, p->diag_mc);
1525                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1526                 for(i= p->htaps/2; i; i--)
1527                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1528             }
1529         }
1530         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1531             put_rac(&s->c, s->header_state, 1);
1532             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1533             encode_qlogs(s);
1534         }else
1535             put_rac(&s->c, s->header_state, 0);
1536     }
1537
1538     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1539     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1540     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1541     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1542     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1543
1544 }
1545
1546 static void update_last_header_values(SnowContext *s){
1547     int plane_index;
1548
1549     if(!s->keyframe){
1550         for(plane_index=0; plane_index<2; plane_index++){
1551             Plane *p= &s->plane[plane_index];
1552             p->last_diag_mc= p->diag_mc;
1553             p->last_htaps  = p->htaps;
1554             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1555         }
1556     }
1557
1558     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1559     s->last_qlog                        = s->qlog;
1560     s->last_qbias                       = s->qbias;
1561     s->last_mv_scale                    = s->mv_scale;
1562     s->last_block_max_depth             = s->block_max_depth;
1563     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1564 }
1565
1566 static int qscale2qlog(int qscale){
1567     return rint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1568            + 61*QROOT/8; ///< 64 > 60
1569 }
1570
1571 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1572 {
1573     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1574      * FIXME we know exact mv bits at this point,
1575      * but ratecontrol isn't set up to include them. */
1576     uint32_t coef_sum= 0;
1577     int level, orientation, delta_qlog;
1578
1579     for(level=0; level<s->spatial_decomposition_count; level++){
1580         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1581             SubBand *b= &s->plane[0].band[level][orientation];
1582             IDWTELEM *buf= b->ibuf;
1583             const int w= b->width;
1584             const int h= b->height;
1585             const int stride= b->stride;
1586             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1587             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1588             const int qdiv= (1<<16)/qmul;
1589             int x, y;
1590             //FIXME this is ugly
1591             for(y=0; y<h; y++)
1592                 for(x=0; x<w; x++)
1593                     buf[x+y*stride]= b->buf[x+y*stride];
1594             if(orientation==0)
1595                 decorrelate(s, b, buf, stride, 1, 0);
1596             for(y=0; y<h; y++)
1597                 for(x=0; x<w; x++)
1598                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1599         }
1600     }
1601
1602     /* ugly, ratecontrol just takes a sqrt again */
1603     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1604     assert(coef_sum < INT_MAX);
1605
1606     if(pict->pict_type == AV_PICTURE_TYPE_I){
1607         s->m.current_picture.mb_var_sum= coef_sum;
1608         s->m.current_picture.mc_mb_var_sum= 0;
1609     }else{
1610         s->m.current_picture.mc_mb_var_sum= coef_sum;
1611         s->m.current_picture.mb_var_sum= 0;
1612     }
1613
1614     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1615     if (pict->quality < 0)
1616         return INT_MIN;
1617     s->lambda= pict->quality * 3/2;
1618     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1619     s->qlog+= delta_qlog;
1620     return delta_qlog;
1621 }
1622
1623 static void calculate_visual_weight(SnowContext *s, Plane *p){
1624     int width = p->width;
1625     int height= p->height;
1626     int level, orientation, x, y;
1627
1628     for(level=0; level<s->spatial_decomposition_count; level++){
1629         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1630             SubBand *b= &p->band[level][orientation];
1631             IDWTELEM *ibuf= b->ibuf;
1632             int64_t error=0;
1633
1634             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1635             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1636             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1637             for(y=0; y<height; y++){
1638                 for(x=0; x<width; x++){
1639                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1640                     error += d*d;
1641                 }
1642             }
1643
1644             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1645         }
1646     }
1647 }
1648
1649 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1650                         const AVFrame *pict, int *got_packet)
1651 {
1652     SnowContext *s = avctx->priv_data;
1653     RangeCoder * const c= &s->c;
1654     AVFrame *pic = &s->new_picture;
1655     const int width= s->avctx->width;
1656     const int height= s->avctx->height;
1657     int level, orientation, plane_index, i, y, ret;
1658     uint8_t rc_header_bak[sizeof(s->header_state)];
1659     uint8_t rc_block_bak[sizeof(s->block_state)];
1660
1661     if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0)
1662         return ret;
1663
1664     ff_init_range_encoder(c, pkt->data, pkt->size);
1665     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1666
1667     for(i=0; i<3; i++){
1668         int hshift= i ? s->chroma_h_shift : 0;
1669         int vshift= i ? s->chroma_v_shift : 0;
1670         for(y=0; y<(height>>vshift); y++)
1671             memcpy(&s->input_picture.data[i][y * s->input_picture.linesize[i]],
1672                    &pict->data[i][y * pict->linesize[i]],
1673                    width>>hshift);
1674         s->dsp.draw_edges(s->input_picture.data[i], s->input_picture.linesize[i],
1675                             width >> hshift, height >> vshift,
1676                             EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1677                             EDGE_TOP | EDGE_BOTTOM);
1678
1679     }
1680     emms_c();
1681     s->new_picture = *pict;
1682
1683     s->m.picture_number= avctx->frame_number;
1684     if(avctx->flags&CODEC_FLAG_PASS2){
1685         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1686         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1687         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1688             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1689             if (pic->quality < 0)
1690                 return -1;
1691         }
1692     }else{
1693         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1694         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1695     }
1696
1697     if(s->pass1_rc && avctx->frame_number == 0)
1698         pic->quality = 2*FF_QP2LAMBDA;
1699     if (pic->quality) {
1700         s->qlog   = qscale2qlog(pic->quality);
1701         s->lambda = pic->quality * 3/2;
1702     }
1703     if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1704         s->qlog= LOSSLESS_QLOG;
1705         s->lambda = 0;
1706     }//else keep previous frame's qlog until after motion estimation
1707
1708     ff_snow_frame_start(s);
1709
1710     s->m.current_picture_ptr= &s->m.current_picture;
1711     s->m.last_picture.f.pts = s->m.current_picture.f.pts;
1712     s->m.current_picture.f.pts = pict->pts;
1713     if(pic->pict_type == AV_PICTURE_TYPE_P){
1714         int block_width = (width +15)>>4;
1715         int block_height= (height+15)>>4;
1716         int stride= s->current_picture.linesize[0];
1717
1718         assert(s->current_picture.data[0]);
1719         assert(s->last_picture[0].data[0]);
1720
1721         s->m.avctx= s->avctx;
1722         s->m.current_picture.f.data[0] = s->current_picture.data[0];
1723         s->m.   last_picture.f.data[0] = s->last_picture[0].data[0];
1724         s->m.    new_picture.f.data[0] = s->  input_picture.data[0];
1725         s->m.   last_picture_ptr= &s->m.   last_picture;
1726         s->m.linesize=
1727         s->m.   last_picture.f.linesize[0] =
1728         s->m.    new_picture.f.linesize[0] =
1729         s->m.current_picture.f.linesize[0] = stride;
1730         s->m.uvlinesize= s->current_picture.linesize[1];
1731         s->m.width = width;
1732         s->m.height= height;
1733         s->m.mb_width = block_width;
1734         s->m.mb_height= block_height;
1735         s->m.mb_stride=   s->m.mb_width+1;
1736         s->m.b8_stride= 2*s->m.mb_width+1;
1737         s->m.f_code=1;
1738         s->m.pict_type = pic->pict_type;
1739         s->m.me_method= s->avctx->me_method;
1740         s->m.me.scene_change_score=0;
1741         s->m.flags= s->avctx->flags;
1742         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1743         s->m.out_format= FMT_H263;
1744         s->m.unrestricted_mv= 1;
1745
1746         s->m.lambda = s->lambda;
1747         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1748         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1749
1750         s->m.dsp= s->dsp; //move
1751         s->m.hdsp = s->hdsp;
1752         ff_init_me(&s->m);
1753         s->hdsp = s->m.hdsp;
1754         s->dsp= s->m.dsp;
1755     }
1756
1757     if(s->pass1_rc){
1758         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1759         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1760     }
1761
1762 redo_frame:
1763
1764     if (pic->pict_type == AV_PICTURE_TYPE_I)
1765         s->spatial_decomposition_count= 5;
1766     else
1767         s->spatial_decomposition_count= 5;
1768
1769     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1770           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1771         s->spatial_decomposition_count--;
1772
1773     s->m.pict_type = pic->pict_type;
1774     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1775
1776     ff_snow_common_init_after_header(avctx);
1777
1778     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1779         for(plane_index=0; plane_index<3; plane_index++){
1780             calculate_visual_weight(s, &s->plane[plane_index]);
1781         }
1782     }
1783
1784     encode_header(s);
1785     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1786     encode_blocks(s, 1);
1787     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1788
1789     for(plane_index=0; plane_index<3; plane_index++){
1790         Plane *p= &s->plane[plane_index];
1791         int w= p->width;
1792         int h= p->height;
1793         int x, y;
1794 //        int bits= put_bits_count(&s->c.pb);
1795
1796         if (!s->memc_only) {
1797             //FIXME optimize
1798             if(pict->data[plane_index]) //FIXME gray hack
1799                 for(y=0; y<h; y++){
1800                     for(x=0; x<w; x++){
1801                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1802                     }
1803                 }
1804             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1805
1806             if(   plane_index==0
1807                && pic->pict_type == AV_PICTURE_TYPE_P
1808                && !(avctx->flags&CODEC_FLAG_PASS2)
1809                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1810                 ff_init_range_encoder(c, pkt->data, pkt->size);
1811                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1812                 pic->pict_type= AV_PICTURE_TYPE_I;
1813                 s->keyframe=1;
1814                 s->current_picture.key_frame=1;
1815                 goto redo_frame;
1816             }
1817
1818             if(s->qlog == LOSSLESS_QLOG){
1819                 for(y=0; y<h; y++){
1820                     for(x=0; x<w; x++){
1821                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1822                     }
1823                 }
1824             }else{
1825                 for(y=0; y<h; y++){
1826                     for(x=0; x<w; x++){
1827                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1828                     }
1829                 }
1830             }
1831
1832             /*  if(QUANTIZE2)
1833                 dwt_quantize(s, p, s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type);
1834             else*/
1835                 ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1836
1837             if(s->pass1_rc && plane_index==0){
1838                 int delta_qlog = ratecontrol_1pass(s, pic);
1839                 if (delta_qlog <= INT_MIN)
1840                     return -1;
1841                 if(delta_qlog){
1842                     //reordering qlog in the bitstream would eliminate this reset
1843                     ff_init_range_encoder(c, pkt->data, pkt->size);
1844                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1845                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1846                     encode_header(s);
1847                     encode_blocks(s, 0);
1848                 }
1849             }
1850
1851             for(level=0; level<s->spatial_decomposition_count; level++){
1852                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1853                     SubBand *b= &p->band[level][orientation];
1854
1855                     if(!QUANTIZE2)
1856                         quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1857                     if(orientation==0)
1858                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1859                     if (!s->no_bitstream)
1860                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1861                     assert(b->parent==NULL || b->parent->stride == b->stride*2);
1862                     if(orientation==0)
1863                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1864                 }
1865             }
1866
1867             for(level=0; level<s->spatial_decomposition_count; level++){
1868                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1869                     SubBand *b= &p->band[level][orientation];
1870
1871                     dequantize(s, b, b->ibuf, b->stride);
1872                 }
1873             }
1874
1875             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1876             if(s->qlog == LOSSLESS_QLOG){
1877                 for(y=0; y<h; y++){
1878                     for(x=0; x<w; x++){
1879                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1880                     }
1881                 }
1882             }
1883             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1884         }else{
1885             //ME/MC only
1886             if(pic->pict_type == AV_PICTURE_TYPE_I){
1887                 for(y=0; y<h; y++){
1888                     for(x=0; x<w; x++){
1889                         s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]=
1890                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1891                     }
1892                 }
1893             }else{
1894                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1895                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1896             }
1897         }
1898         if(s->avctx->flags&CODEC_FLAG_PSNR){
1899             int64_t error= 0;
1900
1901             if(pict->data[plane_index]) //FIXME gray hack
1902                 for(y=0; y<h; y++){
1903                     for(x=0; x<w; x++){
1904                         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];
1905                         error += d*d;
1906                     }
1907                 }
1908             s->avctx->error[plane_index] += error;
1909             s->current_picture.error[plane_index] = error;
1910         }
1911
1912     }
1913
1914     update_last_header_values(s);
1915
1916     ff_snow_release_buffer(avctx);
1917
1918     s->current_picture.coded_picture_number = avctx->frame_number;
1919     s->current_picture.pict_type = pict->pict_type;
1920     s->current_picture.quality = pict->quality;
1921     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1922     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1923     s->m.current_picture.f.display_picture_number =
1924     s->m.current_picture.f.coded_picture_number   = avctx->frame_number;
1925     s->m.current_picture.f.quality                = pic->quality;
1926     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1927     if(s->pass1_rc)
1928         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1929             return -1;
1930     if(avctx->flags&CODEC_FLAG_PASS1)
1931         ff_write_pass1_stats(&s->m);
1932     s->m.last_pict_type = s->m.pict_type;
1933     avctx->frame_bits = s->m.frame_bits;
1934     avctx->mv_bits = s->m.mv_bits;
1935     avctx->misc_bits = s->m.misc_bits;
1936     avctx->p_tex_bits = s->m.p_tex_bits;
1937
1938     emms_c();
1939
1940     pkt->size = ff_rac_terminate(c);
1941     if (avctx->coded_frame->key_frame)
1942         pkt->flags |= AV_PKT_FLAG_KEY;
1943     *got_packet = 1;
1944
1945     return 0;
1946 }
1947
1948 static av_cold int encode_end(AVCodecContext *avctx)
1949 {
1950     SnowContext *s = avctx->priv_data;
1951
1952     ff_snow_common_end(s);
1953     av_frame_unref(&s->input_picture);
1954     av_free(avctx->stats_out);
1955
1956     return 0;
1957 }
1958
1959 #define OFFSET(x) offsetof(SnowContext, x)
1960 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1961 static const AVOption options[] = {
1962     { "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 },
1963     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1964     { NULL },
1965 };
1966
1967 static const AVClass snowenc_class = {
1968     .class_name = "snow encoder",
1969     .item_name  = av_default_item_name,
1970     .option     = options,
1971     .version    = LIBAVUTIL_VERSION_INT,
1972 };
1973
1974 AVCodec ff_snow_encoder = {
1975     .name           = "snow",
1976     .type           = AVMEDIA_TYPE_VIDEO,
1977     .id             = AV_CODEC_ID_SNOW,
1978     .priv_data_size = sizeof(SnowContext),
1979     .init           = encode_init,
1980     .encode2        = encode_frame,
1981     .close          = encode_end,
1982     .pix_fmts       = (const enum AVPixelFormat[]){
1983         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1984         AV_PIX_FMT_NONE
1985     },
1986     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1987     .priv_class     = &snowenc_class,
1988 };
1989
1990
1991 #ifdef TEST
1992 #undef malloc
1993 #undef free
1994 #undef printf
1995
1996 #include "libavutil/lfg.h"
1997 #include "libavutil/mathematics.h"
1998
1999 int main(void){
2000 #define width  256
2001 #define height 256
2002     int buffer[2][width*height];
2003     SnowContext s;
2004     int i;
2005     AVLFG prng;
2006     s.spatial_decomposition_count=6;
2007     s.spatial_decomposition_type=1;
2008
2009     s.temp_dwt_buffer  = av_mallocz(width * sizeof(DWTELEM));
2010     s.temp_idwt_buffer = av_mallocz(width * sizeof(IDWTELEM));
2011
2012     av_lfg_init(&prng, 1);
2013
2014     printf("testing 5/3 DWT\n");
2015     for(i=0; i<width*height; i++)
2016         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
2017
2018     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2019     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2020
2021     for(i=0; i<width*height; i++)
2022         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
2023
2024     printf("testing 9/7 DWT\n");
2025     s.spatial_decomposition_type=0;
2026     for(i=0; i<width*height; i++)
2027         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
2028
2029     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2030     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2031
2032     for(i=0; i<width*height; i++)
2033         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
2034
2035     {
2036     int level, orientation, x, y;
2037     int64_t errors[8][4];
2038     int64_t g=0;
2039
2040         memset(errors, 0, sizeof(errors));
2041         s.spatial_decomposition_count=3;
2042         s.spatial_decomposition_type=0;
2043         for(level=0; level<s.spatial_decomposition_count; level++){
2044             for(orientation=level ? 1 : 0; orientation<4; orientation++){
2045                 int w= width  >> (s.spatial_decomposition_count-level);
2046                 int h= height >> (s.spatial_decomposition_count-level);
2047                 int stride= width  << (s.spatial_decomposition_count-level);
2048                 DWTELEM *buf= buffer[0];
2049                 int64_t error=0;
2050
2051                 if(orientation&1) buf+=w;
2052                 if(orientation>1) buf+=stride>>1;
2053
2054                 memset(buffer[0], 0, sizeof(int)*width*height);
2055                 buf[w/2 + h/2*stride]= 256*256;
2056                 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2057                 for(y=0; y<height; y++){
2058                     for(x=0; x<width; x++){
2059                         int64_t d= buffer[0][x + y*width];
2060                         error += d*d;
2061                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
2062                     }
2063                     if(FFABS(height/2-y)<9 && level==2) printf("\n");
2064                 }
2065                 error= (int)(sqrt(error)+0.5);
2066                 errors[level][orientation]= error;
2067                 if(g) g=av_gcd(g, error);
2068                 else g= error;
2069             }
2070         }
2071         printf("static int const visual_weight[][4]={\n");
2072         for(level=0; level<s.spatial_decomposition_count; level++){
2073             printf("  {");
2074             for(orientation=0; orientation<4; orientation++){
2075                 printf("%8"PRId64",", errors[level][orientation]/g);
2076             }
2077             printf("},\n");
2078         }
2079         printf("};\n");
2080         {
2081             int level=2;
2082             int w= width  >> (s.spatial_decomposition_count-level);
2083             //int h= height >> (s.spatial_decomposition_count-level);
2084             int stride= width  << (s.spatial_decomposition_count-level);
2085             DWTELEM *buf= buffer[0];
2086             int64_t error=0;
2087
2088             buf+=w;
2089             buf+=stride>>1;
2090
2091             memset(buffer[0], 0, sizeof(int)*width*height);
2092             for(y=0; y<height; y++){
2093                 for(x=0; x<width; x++){
2094                     int tab[4]={0,2,3,1};
2095                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2096                 }
2097             }
2098             ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2099             for(y=0; y<height; y++){
2100                 for(x=0; x<width; x++){
2101                     int64_t d= buffer[0][x + y*width];
2102                     error += d*d;
2103                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2104                 }
2105                 if(FFABS(height/2-y)<9) printf("\n");
2106             }
2107         }
2108
2109     }
2110     return 0;
2111 }
2112 #endif /* TEST */