Merge remote-tracking branch 'qatar/master'
[ffmpeg.git] / libavcodec / snowenc.c
1 /*
2  * Copyright (C) 2004 Michael Niedermayer <michaelni@gmx.at>
3  *
4  * This file is part of FFmpeg.
5  *
6  * FFmpeg is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * FFmpeg is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with FFmpeg; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20
21 #include "libavutil/intmath.h"
22 #include "libavutil/log.h"
23 #include "libavutil/opt.h"
24 #include "avcodec.h"
25 #include "internal.h"
26 #include "dsputil.h"
27 #include "dwt.h"
28 #include "snow.h"
29
30 #include "rangecoder.h"
31 #include "mathops.h"
32
33 #include "mpegvideo.h"
34 #include "h263.h"
35
36 #undef NDEBUG
37 #include <assert.h>
38
39 #define QUANTIZE2 0
40
41 #if QUANTIZE2==1
42 #define Q2_STEP 8
43
44 static void find_sse(SnowContext *s, Plane *p, int *score, int score_stride, IDWTELEM *r0, IDWTELEM *r1, int level, int orientation){
45     SubBand *b= &p->band[level][orientation];
46     int x, y;
47     int xo=0;
48     int yo=0;
49     int step= 1 << (s->spatial_decomposition_count - level);
50
51     if(orientation&1)
52         xo= step>>1;
53     if(orientation&2)
54         yo= step>>1;
55
56     //FIXME bias for nonzero ?
57     //FIXME optimize
58     memset(score, 0, sizeof(*score)*score_stride*((p->height + Q2_STEP-1)/Q2_STEP));
59     for(y=0; y<p->height; y++){
60         for(x=0; x<p->width; x++){
61             int sx= (x-xo + step/2) / step / Q2_STEP;
62             int sy= (y-yo + step/2) / step / Q2_STEP;
63             int v= r0[x + y*p->width] - r1[x + y*p->width];
64             assert(sx>=0 && sy>=0 && sx < score_stride);
65             v= ((v+8)>>4)<<4;
66             score[sx + sy*score_stride] += v*v;
67             assert(score[sx + sy*score_stride] >= 0);
68         }
69     }
70 }
71
72 static void dequantize_all(SnowContext *s, Plane *p, IDWTELEM *buffer, int width, int height){
73     int level, orientation;
74
75     for(level=0; level<s->spatial_decomposition_count; level++){
76         for(orientation=level ? 1 : 0; orientation<4; orientation++){
77             SubBand *b= &p->band[level][orientation];
78             IDWTELEM *dst= buffer + (b->ibuf - s->spatial_idwt_buffer);
79
80             dequantize(s, b, dst, b->stride);
81         }
82     }
83 }
84
85 static void dwt_quantize(SnowContext *s, Plane *p, DWTELEM *buffer, int width, int height, int stride, int type){
86     int level, orientation, ys, xs, x, y, pass;
87     IDWTELEM best_dequant[height * stride];
88     IDWTELEM idwt2_buffer[height * stride];
89     const int score_stride= (width + 10)/Q2_STEP;
90     int best_score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
91     int score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
92     int threshold= (s->m.lambda * s->m.lambda) >> 6;
93
94     //FIXME pass the copy cleanly ?
95
96 //    memcpy(dwt_buffer, buffer, height * stride * sizeof(DWTELEM));
97     ff_spatial_dwt(buffer, s->temp_dwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
98
99     for(level=0; level<s->spatial_decomposition_count; level++){
100         for(orientation=level ? 1 : 0; orientation<4; orientation++){
101             SubBand *b= &p->band[level][orientation];
102             IDWTELEM *dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
103              DWTELEM *src=       buffer + (b-> buf - s->spatial_dwt_buffer);
104             assert(src == b->buf); // code does not depend on this but it is true currently
105
106             quantize(s, b, dst, src, b->stride, s->qbias);
107         }
108     }
109     for(pass=0; pass<1; pass++){
110         if(s->qbias == 0) //keyframe
111             continue;
112         for(level=0; level<s->spatial_decomposition_count; level++){
113             for(orientation=level ? 1 : 0; orientation<4; orientation++){
114                 SubBand *b= &p->band[level][orientation];
115                 IDWTELEM *dst= idwt2_buffer + (b->ibuf - s->spatial_idwt_buffer);
116                 IDWTELEM *best_dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
117
118                 for(ys= 0; ys<Q2_STEP; ys++){
119                     for(xs= 0; xs<Q2_STEP; xs++){
120                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
121                         dequantize_all(s, p, idwt2_buffer, width, height);
122                         ff_spatial_idwt(idwt2_buffer, s->temp_idwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
123                         find_sse(s, p, best_score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
124                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
125                         for(y=ys; y<b->height; y+= Q2_STEP){
126                             for(x=xs; x<b->width; x+= Q2_STEP){
127                                 if(dst[x + y*b->stride]<0) dst[x + y*b->stride]++;
128                                 if(dst[x + y*b->stride]>0) dst[x + y*b->stride]--;
129                                 //FIXME try more than just --
130                             }
131                         }
132                         dequantize_all(s, p, idwt2_buffer, width, height);
133                         ff_spatial_idwt(idwt2_buffer, s->temp_idwt_buffer, width, height, stride, type, s->spatial_decomposition_count);
134                         find_sse(s, p, score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
135                         for(y=ys; y<b->height; y+= Q2_STEP){
136                             for(x=xs; x<b->width; x+= Q2_STEP){
137                                 int score_idx= x/Q2_STEP + (y/Q2_STEP)*score_stride;
138                                 if(score[score_idx] <= best_score[score_idx] + threshold){
139                                     best_score[score_idx]= score[score_idx];
140                                     if(best_dst[x + y*b->stride]<0) best_dst[x + y*b->stride]++;
141                                     if(best_dst[x + y*b->stride]>0) best_dst[x + y*b->stride]--;
142                                     //FIXME copy instead
143                                 }
144                             }
145                         }
146                     }
147                 }
148             }
149         }
150     }
151     memcpy(s->spatial_idwt_buffer, best_dequant, height * stride * sizeof(IDWTELEM)); //FIXME work with that directly instead of copy at the end
152 }
153
154 #endif /* QUANTIZE2==1 */
155
156 static av_cold int encode_init(AVCodecContext *avctx)
157 {
158     SnowContext *s = avctx->priv_data;
159     int plane_index, ret;
160
161     if(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL){
162         av_log(avctx, AV_LOG_ERROR, "This codec is under development, files encoded with it may not be decodable with future versions!!!\n"
163                "Use vstrict=-2 / -strict -2 to use it anyway.\n");
164         return -1;
165     }
166
167     if(avctx->prediction_method == DWT_97
168        && (avctx->flags & CODEC_FLAG_QSCALE)
169        && avctx->global_quality == 0){
170         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
171         return -1;
172     }
173
174     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
175
176     s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
177     s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
178
179     for(plane_index=0; plane_index<3; plane_index++){
180         s->plane[plane_index].diag_mc= 1;
181         s->plane[plane_index].htaps= 6;
182         s->plane[plane_index].hcoeff[0]=  40;
183         s->plane[plane_index].hcoeff[1]= -10;
184         s->plane[plane_index].hcoeff[2]=   2;
185         s->plane[plane_index].fast_mc= 1;
186     }
187
188     if ((ret = ff_snow_common_init(avctx)) < 0) {
189         ff_snow_common_end(avctx->priv_data);
190         return ret;
191     }
192     ff_snow_alloc_blocks(s);
193
194     s->version=0;
195
196     s->m.avctx   = avctx;
197     s->m.flags   = avctx->flags;
198     s->m.bit_rate= avctx->bit_rate;
199
200     s->m.me.temp      =
201     s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
202     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
203     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
204     s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
205     ff_h263_encode_init(&s->m); //mv_penalty
206
207     s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
208
209     if(avctx->flags&CODEC_FLAG_PASS1){
210         if(!avctx->stats_out)
211             avctx->stats_out = av_mallocz(256);
212     }
213     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
214         if(ff_rate_control_init(&s->m) < 0)
215             return -1;
216     }
217     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
218
219     avctx->coded_frame= &s->current_picture;
220     switch(avctx->pix_fmt){
221     case AV_PIX_FMT_YUV444P:
222 //    case AV_PIX_FMT_YUV422P:
223     case AV_PIX_FMT_YUV420P:
224 //     case AV_PIX_FMT_GRAY8:
225 //    case AV_PIX_FMT_YUV411P:
226     case AV_PIX_FMT_YUV410P:
227         s->colorspace_type= 0;
228         break;
229 /*    case AV_PIX_FMT_RGB32:
230         s->colorspace= 1;
231         break;*/
232     default:
233         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
234         return -1;
235     }
236     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
237
238     ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
239     ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
240
241     s->avctx->get_buffer(s->avctx, &s->input_picture);
242
243     if(s->avctx->me_method == ME_ITER){
244         int i;
245         int size= s->b_width * s->b_height << 2*s->block_max_depth;
246         for(i=0; i<s->max_ref_frames; i++){
247             s->ref_mvs[i]= av_mallocz(size*sizeof(int16_t[2]));
248             s->ref_scores[i]= av_mallocz(size*sizeof(uint32_t));
249         }
250     }
251
252     return 0;
253 }
254
255 //near copy & paste from dsputil, FIXME
256 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
257 {
258     int s, i, j;
259
260     s = 0;
261     for (i = 0; i < h; i++) {
262         for (j = 0; j < w; j++) {
263             s += pix[0];
264             pix ++;
265         }
266         pix += line_size - w;
267     }
268     return s;
269 }
270
271 //near copy & paste from dsputil, FIXME
272 static int pix_norm1(uint8_t * pix, int line_size, int w)
273 {
274     int s, i, j;
275     uint32_t *sq = ff_squareTbl + 256;
276
277     s = 0;
278     for (i = 0; i < w; i++) {
279         for (j = 0; j < w; j ++) {
280             s += sq[pix[0]];
281             pix ++;
282         }
283         pix += line_size - w;
284     }
285     return s;
286 }
287
288 //FIXME copy&paste
289 #define P_LEFT P[1]
290 #define P_TOP P[2]
291 #define P_TOPRIGHT P[3]
292 #define P_MEDIAN P[4]
293 #define P_MV1 P[9]
294 #define FLAG_QPEL   1 //must be 1
295
296 static int encode_q_branch(SnowContext *s, int level, int x, int y){
297     uint8_t p_buffer[1024];
298     uint8_t i_buffer[1024];
299     uint8_t p_state[sizeof(s->block_state)];
300     uint8_t i_state[sizeof(s->block_state)];
301     RangeCoder pc, ic;
302     uint8_t *pbbak= s->c.bytestream;
303     uint8_t *pbbak_start= s->c.bytestream_start;
304     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
305     const int w= s->b_width  << s->block_max_depth;
306     const int h= s->b_height << s->block_max_depth;
307     const int rem_depth= s->block_max_depth - level;
308     const int index= (x + y*w) << rem_depth;
309     const int block_w= 1<<(LOG2_MB_SIZE - level);
310     int trx= (x+1)<<rem_depth;
311     int try= (y+1)<<rem_depth;
312     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
313     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
314     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
315     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
316     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
317     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
318     int pl = left->color[0];
319     int pcb= left->color[1];
320     int pcr= left->color[2];
321     int pmx, pmy;
322     int mx=0, my=0;
323     int l,cr,cb;
324     const int stride= s->current_picture.linesize[0];
325     const int uvstride= s->current_picture.linesize[1];
326     uint8_t *current_data[3]= { s->input_picture.data[0] + (x + y*  stride)*block_w,
327                                 s->input_picture.data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
328                                 s->input_picture.data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
329     int P[10][2];
330     int16_t last_mv[3][2];
331     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
332     const int shift= 1+qpel;
333     MotionEstContext *c= &s->m.me;
334     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
335     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
336     int my_context= av_log2(2*FFABS(left->my - top->my));
337     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
338     int ref, best_ref, ref_score, ref_mx, ref_my;
339
340     assert(sizeof(s->block_state) >= 256);
341     if(s->keyframe){
342         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
343         return 0;
344     }
345
346 //    clip predictors / edge ?
347
348     P_LEFT[0]= left->mx;
349     P_LEFT[1]= left->my;
350     P_TOP [0]= top->mx;
351     P_TOP [1]= top->my;
352     P_TOPRIGHT[0]= tr->mx;
353     P_TOPRIGHT[1]= tr->my;
354
355     last_mv[0][0]= s->block[index].mx;
356     last_mv[0][1]= s->block[index].my;
357     last_mv[1][0]= right->mx;
358     last_mv[1][1]= right->my;
359     last_mv[2][0]= bottom->mx;
360     last_mv[2][1]= bottom->my;
361
362     s->m.mb_stride=2;
363     s->m.mb_x=
364     s->m.mb_y= 0;
365     c->skip= 0;
366
367     assert(c->  stride ==   stride);
368     assert(c->uvstride == uvstride);
369
370     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
371     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
372     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
373     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
374
375     c->xmin = - x*block_w - 16+3;
376     c->ymin = - y*block_w - 16+3;
377     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
378     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
379
380     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
381     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
382     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
383     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
384     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
385     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
386     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
387
388     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
389     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
390
391     if (!y) {
392         c->pred_x= P_LEFT[0];
393         c->pred_y= P_LEFT[1];
394     } else {
395         c->pred_x = P_MEDIAN[0];
396         c->pred_y = P_MEDIAN[1];
397     }
398
399     score= INT_MAX;
400     best_ref= 0;
401     for(ref=0; ref<s->ref_frames; ref++){
402         init_ref(c, current_data, s->last_picture[ref].data, NULL, block_w*x, block_w*y, 0);
403
404         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
405                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
406
407         assert(ref_mx >= c->xmin);
408         assert(ref_mx <= c->xmax);
409         assert(ref_my >= c->ymin);
410         assert(ref_my <= c->ymax);
411
412         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
413         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
414         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
415         if(s->ref_mvs[ref]){
416             s->ref_mvs[ref][index][0]= ref_mx;
417             s->ref_mvs[ref][index][1]= ref_my;
418             s->ref_scores[ref][index]= ref_score;
419         }
420         if(score > ref_score){
421             score= ref_score;
422             best_ref= ref;
423             mx= ref_mx;
424             my= ref_my;
425         }
426     }
427     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
428
429   //  subpel search
430     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
431     pc= s->c;
432     pc.bytestream_start=
433     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
434     memcpy(p_state, s->block_state, sizeof(s->block_state));
435
436     if(level!=s->block_max_depth)
437         put_rac(&pc, &p_state[4 + s_context], 1);
438     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
439     if(s->ref_frames > 1)
440         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
441     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
442     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
443     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
444     p_len= pc.bytestream - pc.bytestream_start;
445     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
446
447     block_s= block_w*block_w;
448     sum = pix_sum(current_data[0], stride, block_w, block_w);
449     l= (sum + block_s/2)/block_s;
450     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
451
452     block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
453     sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
454     cb= (sum + block_s/2)/block_s;
455 //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
456     sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
457     cr= (sum + block_s/2)/block_s;
458 //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
459
460     ic= s->c;
461     ic.bytestream_start=
462     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
463     memcpy(i_state, s->block_state, sizeof(s->block_state));
464     if(level!=s->block_max_depth)
465         put_rac(&ic, &i_state[4 + s_context], 1);
466     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
467     put_symbol(&ic, &i_state[32],  l-pl , 1);
468     put_symbol(&ic, &i_state[64], cb-pcb, 1);
469     put_symbol(&ic, &i_state[96], cr-pcr, 1);
470     i_len= ic.bytestream - ic.bytestream_start;
471     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
472
473 //    assert(score==256*256*256*64-1);
474     assert(iscore < 255*255*256 + s->lambda2*10);
475     assert(iscore >= 0);
476     assert(l>=0 && l<=255);
477     assert(pl>=0 && pl<=255);
478
479     if(level==0){
480         int varc= iscore >> 8;
481         int vard= score >> 8;
482         if (vard <= 64 || vard < varc)
483             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
484         else
485             c->scene_change_score+= s->m.qscale;
486     }
487
488     if(level!=s->block_max_depth){
489         put_rac(&s->c, &s->block_state[4 + s_context], 0);
490         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
491         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
492         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
493         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
494         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
495
496         if(score2 < score && score2 < iscore)
497             return score2;
498     }
499
500     if(iscore < score){
501         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
502         memcpy(pbbak, i_buffer, i_len);
503         s->c= ic;
504         s->c.bytestream_start= pbbak_start;
505         s->c.bytestream= pbbak + i_len;
506         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
507         memcpy(s->block_state, i_state, sizeof(s->block_state));
508         return iscore;
509     }else{
510         memcpy(pbbak, p_buffer, p_len);
511         s->c= pc;
512         s->c.bytestream_start= pbbak_start;
513         s->c.bytestream= pbbak + p_len;
514         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
515         memcpy(s->block_state, p_state, sizeof(s->block_state));
516         return score;
517     }
518 }
519
520 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
521     const int w= s->b_width  << s->block_max_depth;
522     const int rem_depth= s->block_max_depth - level;
523     const int index= (x + y*w) << rem_depth;
524     int trx= (x+1)<<rem_depth;
525     BlockNode *b= &s->block[index];
526     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
527     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
528     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
529     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
530     int pl = left->color[0];
531     int pcb= left->color[1];
532     int pcr= left->color[2];
533     int pmx, pmy;
534     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
535     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
536     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
537     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
538
539     if(s->keyframe){
540         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
541         return;
542     }
543
544     if(level!=s->block_max_depth){
545         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
546             put_rac(&s->c, &s->block_state[4 + s_context], 1);
547         }else{
548             put_rac(&s->c, &s->block_state[4 + s_context], 0);
549             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
550             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
551             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
552             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
553             return;
554         }
555     }
556     if(b->type & BLOCK_INTRA){
557         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
558         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
559         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
560         put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
561         put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
562         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
563     }else{
564         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
565         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
566         if(s->ref_frames > 1)
567             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
568         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
569         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
570         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
571     }
572 }
573
574 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
575     int i, x2, y2;
576     Plane *p= &s->plane[plane_index];
577     const int block_size = MB_SIZE >> s->block_max_depth;
578     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
579     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
580     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
581     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
582     const int ref_stride= s->current_picture.linesize[plane_index];
583     uint8_t *src= s-> input_picture.data[plane_index];
584     IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
585     const int b_stride = s->b_width << s->block_max_depth;
586     const int w= p->width;
587     const int h= p->height;
588     int index= mb_x + mb_y*b_stride;
589     BlockNode *b= &s->block[index];
590     BlockNode backup= *b;
591     int ab=0;
592     int aa=0;
593
594     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
595
596     b->type|= BLOCK_INTRA;
597     b->color[plane_index]= 0;
598     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
599
600     for(i=0; i<4; i++){
601         int mb_x2= mb_x + (i &1) - 1;
602         int mb_y2= mb_y + (i>>1) - 1;
603         int x= block_w*mb_x2 + block_w/2;
604         int y= block_h*mb_y2 + block_h/2;
605
606         add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
607                     x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
608
609         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
610             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
611                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
612                 int obmc_v= obmc[index];
613                 int d;
614                 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
615                 if(x<0) obmc_v += obmc[index + block_w];
616                 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
617                 if(x+block_w>w) obmc_v += obmc[index - block_w];
618                 //FIXME precalculate this or simplify it somehow else
619
620                 d = -dst[index] + (1<<(FRAC_BITS-1));
621                 dst[index] = d;
622                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
623                 aa += obmc_v * obmc_v; //FIXME precalculate this
624             }
625         }
626     }
627     *b= backup;
628
629     return av_clip(((ab<<LOG2_OBMC_MAX) + aa/2)/aa, 0, 255); //FIXME we should not need clipping
630 }
631
632 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
633     const int b_stride = s->b_width << s->block_max_depth;
634     const int b_height = s->b_height<< s->block_max_depth;
635     int index= x + y*b_stride;
636     const BlockNode *b     = &s->block[index];
637     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
638     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
639     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
640     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
641     int dmx, dmy;
642 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
643 //  int my_context= av_log2(2*FFABS(left->my - top->my));
644
645     if(x<0 || x>=b_stride || y>=b_height)
646         return 0;
647 /*
648 1            0      0
649 01X          1-2    1
650 001XX        3-6    2-3
651 0001XXX      7-14   4-7
652 00001XXXX   15-30   8-15
653 */
654 //FIXME try accurate rate
655 //FIXME intra and inter predictors if surrounding blocks are not the same type
656     if(b->type & BLOCK_INTRA){
657         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
658                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
659                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
660     }else{
661         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
662         dmx-= b->mx;
663         dmy-= b->my;
664         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
665                     + av_log2(2*FFABS(dmy))
666                     + av_log2(2*b->ref));
667     }
668 }
669
670 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
671     Plane *p= &s->plane[plane_index];
672     const int block_size = MB_SIZE >> s->block_max_depth;
673     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
674     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
675     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
676     const int ref_stride= s->current_picture.linesize[plane_index];
677     uint8_t *dst= s->current_picture.data[plane_index];
678     uint8_t *src= s->  input_picture.data[plane_index];
679     IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
680     uint8_t *cur = s->scratchbuf;
681     uint8_t *tmp = s->emu_edge_buffer;
682     const int b_stride = s->b_width << s->block_max_depth;
683     const int b_height = s->b_height<< s->block_max_depth;
684     const int w= p->width;
685     const int h= p->height;
686     int distortion;
687     int rate= 0;
688     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
689     int sx= block_w*mb_x - block_w/2;
690     int sy= block_h*mb_y - block_h/2;
691     int x0= FFMAX(0,-sx);
692     int y0= FFMAX(0,-sy);
693     int x1= FFMIN(block_w*2, w-sx);
694     int y1= FFMIN(block_h*2, h-sy);
695     int i,x,y;
696
697     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
698
699     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);
700
701     for(y=y0; y<y1; y++){
702         const uint8_t *obmc1= obmc_edged[y];
703         const IDWTELEM *pred1 = pred + y*obmc_stride;
704         uint8_t *cur1 = cur + y*ref_stride;
705         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
706         for(x=x0; x<x1; x++){
707 #if FRAC_BITS >= LOG2_OBMC_MAX
708             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
709 #else
710             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
711 #endif
712             v = (v + pred1[x]) >> FRAC_BITS;
713             if(v&(~255)) v= ~(v>>31);
714             dst1[x] = v;
715         }
716     }
717
718     /* copy the regions where obmc[] = (uint8_t)256 */
719     if(LOG2_OBMC_MAX == 8
720         && (mb_x == 0 || mb_x == b_stride-1)
721         && (mb_y == 0 || mb_y == b_height-1)){
722         if(mb_x == 0)
723             x1 = block_w;
724         else
725             x0 = block_w;
726         if(mb_y == 0)
727             y1 = block_h;
728         else
729             y0 = block_h;
730         for(y=y0; y<y1; y++)
731             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
732     }
733
734     if(block_w==16){
735         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
736         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
737         /* FIXME cmps overlap but do not cover the wavelet's whole support.
738          * So improving the score of one block is not strictly guaranteed
739          * to improve the score of the whole frame, thus iterative motion
740          * estimation does not always converge. */
741         if(s->avctx->me_cmp == FF_CMP_W97)
742             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
743         else if(s->avctx->me_cmp == FF_CMP_W53)
744             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
745         else{
746             distortion = 0;
747             for(i=0; i<4; i++){
748                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
749                 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
750             }
751         }
752     }else{
753         assert(block_w==8);
754         distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
755     }
756
757     if(plane_index==0){
758         for(i=0; i<4; i++){
759 /* ..RRr
760  * .RXx.
761  * rxx..
762  */
763             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
764         }
765         if(mb_x == b_stride-2)
766             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
767     }
768     return distortion + rate*penalty_factor;
769 }
770
771 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
772     int i, y2;
773     Plane *p= &s->plane[plane_index];
774     const int block_size = MB_SIZE >> s->block_max_depth;
775     const int block_w    = plane_index ? block_size>>s->chroma_h_shift : block_size;
776     const int block_h    = plane_index ? block_size>>s->chroma_v_shift : block_size;
777     const uint8_t *obmc  = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
778     const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
779     const int ref_stride= s->current_picture.linesize[plane_index];
780     uint8_t *dst= s->current_picture.data[plane_index];
781     uint8_t *src= s-> input_picture.data[plane_index];
782     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
783     // const has only been removed from zero_dst to suppress a warning
784     static IDWTELEM zero_dst[4096]; //FIXME
785     const int b_stride = s->b_width << s->block_max_depth;
786     const int w= p->width;
787     const int h= p->height;
788     int distortion= 0;
789     int rate= 0;
790     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
791
792     av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
793
794     for(i=0; i<9; i++){
795         int mb_x2= mb_x + (i%3) - 1;
796         int mb_y2= mb_y + (i/3) - 1;
797         int x= block_w*mb_x2 + block_w/2;
798         int y= block_h*mb_y2 + block_h/2;
799
800         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
801                    x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
802
803         //FIXME find a cleaner/simpler way to skip the outside stuff
804         for(y2= y; y2<0; y2++)
805             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
806         for(y2= h; y2<y+block_h; y2++)
807             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
808         if(x<0){
809             for(y2= y; y2<y+block_h; y2++)
810                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
811         }
812         if(x+block_w > w){
813             for(y2= y; y2<y+block_h; y2++)
814                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
815         }
816
817         assert(block_w== 8 || block_w==16);
818         distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
819     }
820
821     if(plane_index==0){
822         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
823         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
824
825 /* ..RRRr
826  * .RXXx.
827  * .RXXx.
828  * rxxx.
829  */
830         if(merged)
831             rate = get_block_bits(s, mb_x, mb_y, 2);
832         for(i=merged?4:0; i<9; i++){
833             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
834             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
835         }
836     }
837     return distortion + rate*penalty_factor;
838 }
839
840 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
841     const int w= b->width;
842     const int h= b->height;
843     int x, y;
844
845     if(1){
846         int run=0;
847         int *runs = s->run_buffer;
848         int run_index=0;
849         int max_index;
850
851         for(y=0; y<h; y++){
852             for(x=0; x<w; x++){
853                 int v, p=0;
854                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
855                 v= src[x + y*stride];
856
857                 if(y){
858                     t= src[x + (y-1)*stride];
859                     if(x){
860                         lt= src[x - 1 + (y-1)*stride];
861                     }
862                     if(x + 1 < w){
863                         rt= src[x + 1 + (y-1)*stride];
864                     }
865                 }
866                 if(x){
867                     l= src[x - 1 + y*stride];
868                     /*if(x > 1){
869                         if(orientation==1) ll= src[y + (x-2)*stride];
870                         else               ll= src[x - 2 + y*stride];
871                     }*/
872                 }
873                 if(parent){
874                     int px= x>>1;
875                     int py= y>>1;
876                     if(px<b->parent->width && py<b->parent->height)
877                         p= parent[px + py*2*stride];
878                 }
879                 if(!(/*ll|*/l|lt|t|rt|p)){
880                     if(v){
881                         runs[run_index++]= run;
882                         run=0;
883                     }else{
884                         run++;
885                     }
886                 }
887             }
888         }
889         max_index= run_index;
890         runs[run_index++]= run;
891         run_index=0;
892         run= runs[run_index++];
893
894         put_symbol2(&s->c, b->state[30], max_index, 0);
895         if(run_index <= max_index)
896             put_symbol2(&s->c, b->state[1], run, 3);
897
898         for(y=0; y<h; y++){
899             if(s->c.bytestream_end - s->c.bytestream < w*40){
900                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
901                 return -1;
902             }
903             for(x=0; x<w; x++){
904                 int v, p=0;
905                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
906                 v= src[x + y*stride];
907
908                 if(y){
909                     t= src[x + (y-1)*stride];
910                     if(x){
911                         lt= src[x - 1 + (y-1)*stride];
912                     }
913                     if(x + 1 < w){
914                         rt= src[x + 1 + (y-1)*stride];
915                     }
916                 }
917                 if(x){
918                     l= src[x - 1 + y*stride];
919                     /*if(x > 1){
920                         if(orientation==1) ll= src[y + (x-2)*stride];
921                         else               ll= src[x - 2 + y*stride];
922                     }*/
923                 }
924                 if(parent){
925                     int px= x>>1;
926                     int py= y>>1;
927                     if(px<b->parent->width && py<b->parent->height)
928                         p= parent[px + py*2*stride];
929                 }
930                 if(/*ll|*/l|lt|t|rt|p){
931                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
932
933                     put_rac(&s->c, &b->state[0][context], !!v);
934                 }else{
935                     if(!run){
936                         run= runs[run_index++];
937
938                         if(run_index <= max_index)
939                             put_symbol2(&s->c, b->state[1], run, 3);
940                         assert(v);
941                     }else{
942                         run--;
943                         assert(!v);
944                     }
945                 }
946                 if(v){
947                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
948                     int l2= 2*FFABS(l) + (l<0);
949                     int t2= 2*FFABS(t) + (t<0);
950
951                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
952                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
953                 }
954             }
955         }
956     }
957     return 0;
958 }
959
960 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
961 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
962 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
963     return encode_subband_c0run(s, b, src, parent, stride, orientation);
964 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
965 }
966
967 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){
968     const int b_stride= s->b_width << s->block_max_depth;
969     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
970     BlockNode backup= *block;
971     unsigned value;
972     int rd, index;
973
974     assert(mb_x>=0 && mb_y>=0);
975     assert(mb_x<b_stride);
976
977     if(intra){
978         block->color[0] = p[0];
979         block->color[1] = p[1];
980         block->color[2] = p[2];
981         block->type |= BLOCK_INTRA;
982     }else{
983         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
984         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
985         if(s->me_cache[index] == value)
986             return 0;
987         s->me_cache[index]= value;
988
989         block->mx= p[0];
990         block->my= p[1];
991         block->type &= ~BLOCK_INTRA;
992     }
993
994     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
995
996 //FIXME chroma
997     if(rd < *best_rd){
998         *best_rd= rd;
999         return 1;
1000     }else{
1001         *block= backup;
1002         return 0;
1003     }
1004 }
1005
1006 /* special case for int[2] args we discard afterwards,
1007  * fixes compilation problem with gcc 2.95 */
1008 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){
1009     int p[2] = {p0, p1};
1010     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
1011 }
1012
1013 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){
1014     const int b_stride= s->b_width << s->block_max_depth;
1015     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
1016     BlockNode backup[4];
1017     unsigned value;
1018     int rd, index;
1019
1020     /* We don't initialize backup[] during variable declaration, because
1021      * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
1022      * 'int16_t'". */
1023     backup[0] = block[0];
1024     backup[1] = block[1];
1025     backup[2] = block[b_stride];
1026     backup[3] = block[b_stride + 1];
1027
1028     assert(mb_x>=0 && mb_y>=0);
1029     assert(mb_x<b_stride);
1030     assert(((mb_x|mb_y)&1) == 0);
1031
1032     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
1033     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
1034     if(s->me_cache[index] == value)
1035         return 0;
1036     s->me_cache[index]= value;
1037
1038     block->mx= p0;
1039     block->my= p1;
1040     block->ref= ref;
1041     block->type &= ~BLOCK_INTRA;
1042     block[1]= block[b_stride]= block[b_stride+1]= *block;
1043
1044     rd= get_4block_rd(s, mb_x, mb_y, 0);
1045
1046 //FIXME chroma
1047     if(rd < *best_rd){
1048         *best_rd= rd;
1049         return 1;
1050     }else{
1051         block[0]= backup[0];
1052         block[1]= backup[1];
1053         block[b_stride]= backup[2];
1054         block[b_stride+1]= backup[3];
1055         return 0;
1056     }
1057 }
1058
1059 static void iterative_me(SnowContext *s){
1060     int pass, mb_x, mb_y;
1061     const int b_width = s->b_width  << s->block_max_depth;
1062     const int b_height= s->b_height << s->block_max_depth;
1063     const int b_stride= b_width;
1064     int color[3];
1065
1066     {
1067         RangeCoder r = s->c;
1068         uint8_t state[sizeof(s->block_state)];
1069         memcpy(state, s->block_state, sizeof(s->block_state));
1070         for(mb_y= 0; mb_y<s->b_height; mb_y++)
1071             for(mb_x= 0; mb_x<s->b_width; mb_x++)
1072                 encode_q_branch(s, 0, mb_x, mb_y);
1073         s->c = r;
1074         memcpy(s->block_state, state, sizeof(s->block_state));
1075     }
1076
1077     for(pass=0; pass<25; pass++){
1078         int change= 0;
1079
1080         for(mb_y= 0; mb_y<b_height; mb_y++){
1081             for(mb_x= 0; mb_x<b_width; mb_x++){
1082                 int dia_change, i, j, ref;
1083                 int best_rd= INT_MAX, ref_rd;
1084                 BlockNode backup, ref_b;
1085                 const int index= mb_x + mb_y * b_stride;
1086                 BlockNode *block= &s->block[index];
1087                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1088                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1089                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1090                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1091                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1092                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1093                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1094                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1095                 const int b_w= (MB_SIZE >> s->block_max_depth);
1096                 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1097
1098                 if(pass && (block->type & BLOCK_OPT))
1099                     continue;
1100                 block->type |= BLOCK_OPT;
1101
1102                 backup= *block;
1103
1104                 if(!s->me_cache_generation)
1105                     memset(s->me_cache, 0, sizeof(s->me_cache));
1106                 s->me_cache_generation += 1<<22;
1107
1108                 //FIXME precalculate
1109                 {
1110                     int x, y;
1111                     for (y = 0; y < b_w * 2; y++)
1112                         memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1113                     if(mb_x==0)
1114                         for(y=0; y<b_w*2; y++)
1115                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1116                     if(mb_x==b_stride-1)
1117                         for(y=0; y<b_w*2; y++)
1118                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1119                     if(mb_y==0){
1120                         for(x=0; x<b_w*2; x++)
1121                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1122                         for(y=1; y<b_w; y++)
1123                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1124                     }
1125                     if(mb_y==b_height-1){
1126                         for(x=0; x<b_w*2; x++)
1127                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1128                         for(y=b_w; y<b_w*2-1; y++)
1129                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1130                     }
1131                 }
1132
1133                 //skip stuff outside the picture
1134                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1135                     uint8_t *src= s->  input_picture.data[0];
1136                     uint8_t *dst= s->current_picture.data[0];
1137                     const int stride= s->current_picture.linesize[0];
1138                     const int block_w= MB_SIZE >> s->block_max_depth;
1139                     const int block_h= MB_SIZE >> s->block_max_depth;
1140                     const int sx= block_w*mb_x - block_w/2;
1141                     const int sy= block_h*mb_y - block_h/2;
1142                     const int w= s->plane[0].width;
1143                     const int h= s->plane[0].height;
1144                     int y;
1145
1146                     for(y=sy; y<0; y++)
1147                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1148                     for(y=h; y<sy+block_h*2; y++)
1149                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1150                     if(sx<0){
1151                         for(y=sy; y<sy+block_h*2; y++)
1152                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1153                     }
1154                     if(sx+block_w*2 > w){
1155                         for(y=sy; y<sy+block_h*2; y++)
1156                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1157                     }
1158                 }
1159
1160                 // intra(black) = neighbors' contribution to the current block
1161                 for(i=0; i<3; i++)
1162                     color[i]= get_dc(s, mb_x, mb_y, i);
1163
1164                 // get previous score (cannot be cached due to OBMC)
1165                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1166                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1167                     check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1168                 }else
1169                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1170
1171                 ref_b= *block;
1172                 ref_rd= best_rd;
1173                 for(ref=0; ref < s->ref_frames; ref++){
1174                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1175                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1176                         continue;
1177                     block->ref= ref;
1178                     best_rd= INT_MAX;
1179
1180                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1181                     check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1182                     if(tb)
1183                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1184                     if(lb)
1185                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1186                     if(rb)
1187                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1188                     if(bb)
1189                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1190
1191                     /* fullpel ME */
1192                     //FIXME avoid subpel interpolation / round to nearest integer
1193                     do{
1194                         dia_change=0;
1195                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1196                             for(j=0; j<i; j++){
1197                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1198                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1199                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1200                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1201                             }
1202                         }
1203                     }while(dia_change);
1204                     /* subpel ME */
1205                     do{
1206                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1207                         dia_change=0;
1208                         for(i=0; i<8; i++)
1209                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1210                     }while(dia_change);
1211                     //FIXME or try the standard 2 pass qpel or similar
1212
1213                     mvr[0][0]= block->mx;
1214                     mvr[0][1]= block->my;
1215                     if(ref_rd > best_rd){
1216                         ref_rd= best_rd;
1217                         ref_b= *block;
1218                     }
1219                 }
1220                 best_rd= ref_rd;
1221                 *block= ref_b;
1222                 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1223                 //FIXME RD style color selection
1224                 if(!same_block(block, &backup)){
1225                     if(tb ) tb ->type &= ~BLOCK_OPT;
1226                     if(lb ) lb ->type &= ~BLOCK_OPT;
1227                     if(rb ) rb ->type &= ~BLOCK_OPT;
1228                     if(bb ) bb ->type &= ~BLOCK_OPT;
1229                     if(tlb) tlb->type &= ~BLOCK_OPT;
1230                     if(trb) trb->type &= ~BLOCK_OPT;
1231                     if(blb) blb->type &= ~BLOCK_OPT;
1232                     if(brb) brb->type &= ~BLOCK_OPT;
1233                     change ++;
1234                 }
1235             }
1236         }
1237         av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1238         if(!change)
1239             break;
1240     }
1241
1242     if(s->block_max_depth == 1){
1243         int change= 0;
1244         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1245             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1246                 int i;
1247                 int best_rd, init_rd;
1248                 const int index= mb_x + mb_y * b_stride;
1249                 BlockNode *b[4];
1250
1251                 b[0]= &s->block[index];
1252                 b[1]= b[0]+1;
1253                 b[2]= b[0]+b_stride;
1254                 b[3]= b[2]+1;
1255                 if(same_block(b[0], b[1]) &&
1256                    same_block(b[0], b[2]) &&
1257                    same_block(b[0], b[3]))
1258                     continue;
1259
1260                 if(!s->me_cache_generation)
1261                     memset(s->me_cache, 0, sizeof(s->me_cache));
1262                 s->me_cache_generation += 1<<22;
1263
1264                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1265
1266                 //FIXME more multiref search?
1267                 check_4block_inter(s, mb_x, mb_y,
1268                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1269                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1270
1271                 for(i=0; i<4; i++)
1272                     if(!(b[i]->type&BLOCK_INTRA))
1273                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1274
1275                 if(init_rd != best_rd)
1276                     change++;
1277             }
1278         }
1279         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1280     }
1281 }
1282
1283 static void encode_blocks(SnowContext *s, int search){
1284     int x, y;
1285     int w= s->b_width;
1286     int h= s->b_height;
1287
1288     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1289         iterative_me(s);
1290
1291     for(y=0; y<h; y++){
1292         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1293             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1294             return;
1295         }
1296         for(x=0; x<w; x++){
1297             if(s->avctx->me_method == ME_ITER || !search)
1298                 encode_q_branch2(s, 0, x, y);
1299             else
1300                 encode_q_branch (s, 0, x, y);
1301         }
1302     }
1303 }
1304
1305 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1306     const int w= b->width;
1307     const int h= b->height;
1308     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1309     const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1310     int x,y, thres1, thres2;
1311
1312     if(s->qlog == LOSSLESS_QLOG){
1313         for(y=0; y<h; y++)
1314             for(x=0; x<w; x++)
1315                 dst[x + y*stride]= src[x + y*stride];
1316         return;
1317     }
1318
1319     bias= bias ? 0 : (3*qmul)>>3;
1320     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1321     thres2= 2*thres1;
1322
1323     if(!bias){
1324         for(y=0; y<h; y++){
1325             for(x=0; x<w; x++){
1326                 int i= src[x + y*stride];
1327
1328                 if((unsigned)(i+thres1) > thres2){
1329                     if(i>=0){
1330                         i<<= QEXPSHIFT;
1331                         i/= qmul; //FIXME optimize
1332                         dst[x + y*stride]=  i;
1333                     }else{
1334                         i= -i;
1335                         i<<= QEXPSHIFT;
1336                         i/= qmul; //FIXME optimize
1337                         dst[x + y*stride]= -i;
1338                     }
1339                 }else
1340                     dst[x + y*stride]= 0;
1341             }
1342         }
1343     }else{
1344         for(y=0; y<h; y++){
1345             for(x=0; x<w; x++){
1346                 int i= src[x + y*stride];
1347
1348                 if((unsigned)(i+thres1) > thres2){
1349                     if(i>=0){
1350                         i<<= QEXPSHIFT;
1351                         i= (i + bias) / qmul; //FIXME optimize
1352                         dst[x + y*stride]=  i;
1353                     }else{
1354                         i= -i;
1355                         i<<= QEXPSHIFT;
1356                         i= (i + bias) / qmul; //FIXME optimize
1357                         dst[x + y*stride]= -i;
1358                     }
1359                 }else
1360                     dst[x + y*stride]= 0;
1361             }
1362         }
1363     }
1364 }
1365
1366 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1367     const int w= b->width;
1368     const int h= b->height;
1369     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1370     const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1371     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1372     int x,y;
1373
1374     if(s->qlog == LOSSLESS_QLOG) return;
1375
1376     for(y=0; y<h; y++){
1377         for(x=0; x<w; x++){
1378             int i= src[x + y*stride];
1379             if(i<0){
1380                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1381             }else if(i>0){
1382                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1383             }
1384         }
1385     }
1386 }
1387
1388 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1389     const int w= b->width;
1390     const int h= b->height;
1391     int x,y;
1392
1393     for(y=h-1; y>=0; y--){
1394         for(x=w-1; x>=0; x--){
1395             int i= x + y*stride;
1396
1397             if(x){
1398                 if(use_median){
1399                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1400                     else  src[i] -= src[i - 1];
1401                 }else{
1402                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1403                     else  src[i] -= src[i - 1];
1404                 }
1405             }else{
1406                 if(y) src[i] -= src[i - stride];
1407             }
1408         }
1409     }
1410 }
1411
1412 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1413     const int w= b->width;
1414     const int h= b->height;
1415     int x,y;
1416
1417     for(y=0; y<h; y++){
1418         for(x=0; x<w; x++){
1419             int i= x + y*stride;
1420
1421             if(x){
1422                 if(use_median){
1423                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1424                     else  src[i] += src[i - 1];
1425                 }else{
1426                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1427                     else  src[i] += src[i - 1];
1428                 }
1429             }else{
1430                 if(y) src[i] += src[i - stride];
1431             }
1432         }
1433     }
1434 }
1435
1436 static void encode_qlogs(SnowContext *s){
1437     int plane_index, level, orientation;
1438
1439     for(plane_index=0; plane_index<2; plane_index++){
1440         for(level=0; level<s->spatial_decomposition_count; level++){
1441             for(orientation=level ? 1:0; orientation<4; orientation++){
1442                 if(orientation==2) continue;
1443                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1444             }
1445         }
1446     }
1447 }
1448
1449 static void encode_header(SnowContext *s){
1450     int plane_index, i;
1451     uint8_t kstate[32];
1452
1453     memset(kstate, MID_STATE, sizeof(kstate));
1454
1455     put_rac(&s->c, kstate, s->keyframe);
1456     if(s->keyframe || s->always_reset){
1457         ff_snow_reset_contexts(s);
1458         s->last_spatial_decomposition_type=
1459         s->last_qlog=
1460         s->last_qbias=
1461         s->last_mv_scale=
1462         s->last_block_max_depth= 0;
1463         for(plane_index=0; plane_index<2; plane_index++){
1464             Plane *p= &s->plane[plane_index];
1465             p->last_htaps=0;
1466             p->last_diag_mc=0;
1467             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1468         }
1469     }
1470     if(s->keyframe){
1471         put_symbol(&s->c, s->header_state, s->version, 0);
1472         put_rac(&s->c, s->header_state, s->always_reset);
1473         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1474         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1475         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1476         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1477         put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1478         put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1479         put_rac(&s->c, s->header_state, s->spatial_scalability);
1480 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1481         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1482
1483         encode_qlogs(s);
1484     }
1485
1486     if(!s->keyframe){
1487         int update_mc=0;
1488         for(plane_index=0; plane_index<2; plane_index++){
1489             Plane *p= &s->plane[plane_index];
1490             update_mc |= p->last_htaps   != p->htaps;
1491             update_mc |= p->last_diag_mc != p->diag_mc;
1492             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1493         }
1494         put_rac(&s->c, s->header_state, update_mc);
1495         if(update_mc){
1496             for(plane_index=0; plane_index<2; plane_index++){
1497                 Plane *p= &s->plane[plane_index];
1498                 put_rac(&s->c, s->header_state, p->diag_mc);
1499                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1500                 for(i= p->htaps/2; i; i--)
1501                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1502             }
1503         }
1504         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1505             put_rac(&s->c, s->header_state, 1);
1506             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1507             encode_qlogs(s);
1508         }else
1509             put_rac(&s->c, s->header_state, 0);
1510     }
1511
1512     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1513     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1514     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1515     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1516     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1517
1518 }
1519
1520 static void update_last_header_values(SnowContext *s){
1521     int plane_index;
1522
1523     if(!s->keyframe){
1524         for(plane_index=0; plane_index<2; plane_index++){
1525             Plane *p= &s->plane[plane_index];
1526             p->last_diag_mc= p->diag_mc;
1527             p->last_htaps  = p->htaps;
1528             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1529         }
1530     }
1531
1532     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1533     s->last_qlog                        = s->qlog;
1534     s->last_qbias                       = s->qbias;
1535     s->last_mv_scale                    = s->mv_scale;
1536     s->last_block_max_depth             = s->block_max_depth;
1537     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1538 }
1539
1540 static int qscale2qlog(int qscale){
1541     return rint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1542            + 61*QROOT/8; ///< 64 > 60
1543 }
1544
1545 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1546 {
1547     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1548      * FIXME we know exact mv bits at this point,
1549      * but ratecontrol isn't set up to include them. */
1550     uint32_t coef_sum= 0;
1551     int level, orientation, delta_qlog;
1552
1553     for(level=0; level<s->spatial_decomposition_count; level++){
1554         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1555             SubBand *b= &s->plane[0].band[level][orientation];
1556             IDWTELEM *buf= b->ibuf;
1557             const int w= b->width;
1558             const int h= b->height;
1559             const int stride= b->stride;
1560             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1561             const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1562             const int qdiv= (1<<16)/qmul;
1563             int x, y;
1564             //FIXME this is ugly
1565             for(y=0; y<h; y++)
1566                 for(x=0; x<w; x++)
1567                     buf[x+y*stride]= b->buf[x+y*stride];
1568             if(orientation==0)
1569                 decorrelate(s, b, buf, stride, 1, 0);
1570             for(y=0; y<h; y++)
1571                 for(x=0; x<w; x++)
1572                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1573         }
1574     }
1575
1576     /* ugly, ratecontrol just takes a sqrt again */
1577     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1578     assert(coef_sum < INT_MAX);
1579
1580     if(pict->pict_type == AV_PICTURE_TYPE_I){
1581         s->m.current_picture.mb_var_sum= coef_sum;
1582         s->m.current_picture.mc_mb_var_sum= 0;
1583     }else{
1584         s->m.current_picture.mc_mb_var_sum= coef_sum;
1585         s->m.current_picture.mb_var_sum= 0;
1586     }
1587
1588     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1589     if (pict->quality < 0)
1590         return INT_MIN;
1591     s->lambda= pict->quality * 3/2;
1592     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1593     s->qlog+= delta_qlog;
1594     return delta_qlog;
1595 }
1596
1597 static void calculate_visual_weight(SnowContext *s, Plane *p){
1598     int width = p->width;
1599     int height= p->height;
1600     int level, orientation, x, y;
1601
1602     for(level=0; level<s->spatial_decomposition_count; level++){
1603         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1604             SubBand *b= &p->band[level][orientation];
1605             IDWTELEM *ibuf= b->ibuf;
1606             int64_t error=0;
1607
1608             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1609             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1610             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1611             for(y=0; y<height; y++){
1612                 for(x=0; x<width; x++){
1613                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1614                     error += d*d;
1615                 }
1616             }
1617
1618             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1619         }
1620     }
1621 }
1622
1623 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1624                         const AVFrame *pict, int *got_packet)
1625 {
1626     SnowContext *s = avctx->priv_data;
1627     RangeCoder * const c= &s->c;
1628     AVFrame *pic = &s->new_picture;
1629     const int width= s->avctx->width;
1630     const int height= s->avctx->height;
1631     int level, orientation, plane_index, i, y, ret;
1632     uint8_t rc_header_bak[sizeof(s->header_state)];
1633     uint8_t rc_block_bak[sizeof(s->block_state)];
1634
1635     if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0)
1636         return ret;
1637
1638     ff_init_range_encoder(c, pkt->data, pkt->size);
1639     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1640
1641     for(i=0; i<3; i++){
1642         int hshift= i ? s->chroma_h_shift : 0;
1643         int vshift= i ? s->chroma_v_shift : 0;
1644         for(y=0; y<(height>>vshift); y++)
1645             memcpy(&s->input_picture.data[i][y * s->input_picture.linesize[i]],
1646                    &pict->data[i][y * pict->linesize[i]],
1647                    width>>hshift);
1648     }
1649     s->new_picture = *pict;
1650
1651     s->m.picture_number= avctx->frame_number;
1652     if(avctx->flags&CODEC_FLAG_PASS2){
1653         s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1654         s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1655         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1656             pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1657             if (pic->quality < 0)
1658                 return -1;
1659         }
1660     }else{
1661         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1662         s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1663     }
1664
1665     if(s->pass1_rc && avctx->frame_number == 0)
1666         pic->quality = 2*FF_QP2LAMBDA;
1667     if (pic->quality) {
1668         s->qlog   = qscale2qlog(pic->quality);
1669         s->lambda = pic->quality * 3/2;
1670     }
1671     if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1672         s->qlog= LOSSLESS_QLOG;
1673         s->lambda = 0;
1674     }//else keep previous frame's qlog until after motion estimation
1675
1676     ff_snow_frame_start(s);
1677
1678     s->m.current_picture_ptr= &s->m.current_picture;
1679     s->m.last_picture.f.pts = s->m.current_picture.f.pts;
1680     s->m.current_picture.f.pts = pict->pts;
1681     if(pic->pict_type == AV_PICTURE_TYPE_P){
1682         int block_width = (width +15)>>4;
1683         int block_height= (height+15)>>4;
1684         int stride= s->current_picture.linesize[0];
1685
1686         assert(s->current_picture.data[0]);
1687         assert(s->last_picture[0].data[0]);
1688
1689         s->m.avctx= s->avctx;
1690         s->m.current_picture.f.data[0] = s->current_picture.data[0];
1691         s->m.   last_picture.f.data[0] = s->last_picture[0].data[0];
1692         s->m.    new_picture.f.data[0] = s->  input_picture.data[0];
1693         s->m.   last_picture_ptr= &s->m.   last_picture;
1694         s->m.linesize=
1695         s->m.   last_picture.f.linesize[0] =
1696         s->m.    new_picture.f.linesize[0] =
1697         s->m.current_picture.f.linesize[0] = stride;
1698         s->m.uvlinesize= s->current_picture.linesize[1];
1699         s->m.width = width;
1700         s->m.height= height;
1701         s->m.mb_width = block_width;
1702         s->m.mb_height= block_height;
1703         s->m.mb_stride=   s->m.mb_width+1;
1704         s->m.b8_stride= 2*s->m.mb_width+1;
1705         s->m.f_code=1;
1706         s->m.pict_type = pic->pict_type;
1707         s->m.me_method= s->avctx->me_method;
1708         s->m.me.scene_change_score=0;
1709         s->m.flags= s->avctx->flags;
1710         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1711         s->m.out_format= FMT_H263;
1712         s->m.unrestricted_mv= 1;
1713
1714         s->m.lambda = s->lambda;
1715         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1716         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1717
1718         s->m.dsp= s->dsp; //move
1719         ff_init_me(&s->m);
1720         s->dsp= s->m.dsp;
1721     }
1722
1723     if(s->pass1_rc){
1724         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1725         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1726     }
1727
1728 redo_frame:
1729
1730     if (pic->pict_type == AV_PICTURE_TYPE_I)
1731         s->spatial_decomposition_count= 5;
1732     else
1733         s->spatial_decomposition_count= 5;
1734
1735     while(   !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1736           || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1737         s->spatial_decomposition_count--;
1738
1739     s->m.pict_type = pic->pict_type;
1740     s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1741
1742     ff_snow_common_init_after_header(avctx);
1743
1744     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1745         for(plane_index=0; plane_index<3; plane_index++){
1746             calculate_visual_weight(s, &s->plane[plane_index]);
1747         }
1748     }
1749
1750     encode_header(s);
1751     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1752     encode_blocks(s, 1);
1753     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1754
1755     for(plane_index=0; plane_index<3; plane_index++){
1756         Plane *p= &s->plane[plane_index];
1757         int w= p->width;
1758         int h= p->height;
1759         int x, y;
1760 //        int bits= put_bits_count(&s->c.pb);
1761
1762         if (!s->memc_only) {
1763             //FIXME optimize
1764             if(pict->data[plane_index]) //FIXME gray hack
1765                 for(y=0; y<h; y++){
1766                     for(x=0; x<w; x++){
1767                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1768                     }
1769                 }
1770             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1771
1772             if(   plane_index==0
1773                && pic->pict_type == AV_PICTURE_TYPE_P
1774                && !(avctx->flags&CODEC_FLAG_PASS2)
1775                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1776                 ff_init_range_encoder(c, pkt->data, pkt->size);
1777                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1778                 pic->pict_type= AV_PICTURE_TYPE_I;
1779                 s->keyframe=1;
1780                 s->current_picture.key_frame=1;
1781                 goto redo_frame;
1782             }
1783
1784             if(s->qlog == LOSSLESS_QLOG){
1785                 for(y=0; y<h; y++){
1786                     for(x=0; x<w; x++){
1787                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1788                     }
1789                 }
1790             }else{
1791                 for(y=0; y<h; y++){
1792                     for(x=0; x<w; x++){
1793                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1794                     }
1795                 }
1796             }
1797
1798             /*  if(QUANTIZE2)
1799                 dwt_quantize(s, p, s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type);
1800             else*/
1801                 ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1802
1803             if(s->pass1_rc && plane_index==0){
1804                 int delta_qlog = ratecontrol_1pass(s, pic);
1805                 if (delta_qlog <= INT_MIN)
1806                     return -1;
1807                 if(delta_qlog){
1808                     //reordering qlog in the bitstream would eliminate this reset
1809                     ff_init_range_encoder(c, pkt->data, pkt->size);
1810                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1811                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1812                     encode_header(s);
1813                     encode_blocks(s, 0);
1814                 }
1815             }
1816
1817             for(level=0; level<s->spatial_decomposition_count; level++){
1818                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1819                     SubBand *b= &p->band[level][orientation];
1820
1821                     if(!QUANTIZE2)
1822                         quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1823                     if(orientation==0)
1824                         decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1825                     if (!s->no_bitstream)
1826                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1827                     assert(b->parent==NULL || b->parent->stride == b->stride*2);
1828                     if(orientation==0)
1829                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1830                 }
1831             }
1832
1833             for(level=0; level<s->spatial_decomposition_count; level++){
1834                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1835                     SubBand *b= &p->band[level][orientation];
1836
1837                     dequantize(s, b, b->ibuf, b->stride);
1838                 }
1839             }
1840
1841             ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1842             if(s->qlog == LOSSLESS_QLOG){
1843                 for(y=0; y<h; y++){
1844                     for(x=0; x<w; x++){
1845                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1846                     }
1847                 }
1848             }
1849             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1850         }else{
1851             //ME/MC only
1852             if(pic->pict_type == AV_PICTURE_TYPE_I){
1853                 for(y=0; y<h; y++){
1854                     for(x=0; x<w; x++){
1855                         s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]=
1856                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1857                     }
1858                 }
1859             }else{
1860                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1861                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1862             }
1863         }
1864         if(s->avctx->flags&CODEC_FLAG_PSNR){
1865             int64_t error= 0;
1866
1867             if(pict->data[plane_index]) //FIXME gray hack
1868                 for(y=0; y<h; y++){
1869                     for(x=0; x<w; x++){
1870                         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];
1871                         error += d*d;
1872                     }
1873                 }
1874             s->avctx->error[plane_index] += error;
1875             s->current_picture.error[plane_index] = error;
1876         }
1877
1878     }
1879
1880     update_last_header_values(s);
1881
1882     ff_snow_release_buffer(avctx);
1883
1884     s->current_picture.coded_picture_number = avctx->frame_number;
1885     s->current_picture.pict_type = pict->pict_type;
1886     s->current_picture.quality = pict->quality;
1887     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1888     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1889     s->m.current_picture.f.display_picture_number =
1890     s->m.current_picture.f.coded_picture_number   = avctx->frame_number;
1891     s->m.current_picture.f.quality                = pic->quality;
1892     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1893     if(s->pass1_rc)
1894         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1895             return -1;
1896     if(avctx->flags&CODEC_FLAG_PASS1)
1897         ff_write_pass1_stats(&s->m);
1898     s->m.last_pict_type = s->m.pict_type;
1899     avctx->frame_bits = s->m.frame_bits;
1900     avctx->mv_bits = s->m.mv_bits;
1901     avctx->misc_bits = s->m.misc_bits;
1902     avctx->p_tex_bits = s->m.p_tex_bits;
1903
1904     emms_c();
1905
1906     pkt->size = ff_rac_terminate(c);
1907     if (avctx->coded_frame->key_frame)
1908         pkt->flags |= AV_PKT_FLAG_KEY;
1909     *got_packet = 1;
1910
1911     return 0;
1912 }
1913
1914 static av_cold int encode_end(AVCodecContext *avctx)
1915 {
1916     SnowContext *s = avctx->priv_data;
1917
1918     ff_snow_common_end(s);
1919     if (s->input_picture.data[0])
1920         avctx->release_buffer(avctx, &s->input_picture);
1921     av_free(avctx->stats_out);
1922
1923     return 0;
1924 }
1925
1926 #define OFFSET(x) offsetof(SnowContext, x)
1927 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1928 static const AVOption options[] = {
1929     { "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 },
1930     { "no_bitstream",   "Skip final bitstream writeout.",                    OFFSET(no_bitstream), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1931     { NULL },
1932 };
1933
1934 static const AVClass snowenc_class = {
1935     .class_name = "snow encoder",
1936     .item_name  = av_default_item_name,
1937     .option     = options,
1938     .version    = LIBAVUTIL_VERSION_INT,
1939 };
1940
1941 AVCodec ff_snow_encoder = {
1942     .name           = "snow",
1943     .type           = AVMEDIA_TYPE_VIDEO,
1944     .id             = AV_CODEC_ID_SNOW,
1945     .priv_data_size = sizeof(SnowContext),
1946     .init           = encode_init,
1947     .encode2        = encode_frame,
1948     .close          = encode_end,
1949     .pix_fmts       = (const enum AVPixelFormat[]){
1950         AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1951         AV_PIX_FMT_NONE
1952     },
1953     .long_name      = NULL_IF_CONFIG_SMALL("Snow"),
1954     .priv_class     = &snowenc_class,
1955 };
1956
1957
1958 #ifdef TEST
1959 #undef malloc
1960 #undef free
1961 #undef printf
1962
1963 #include "libavutil/lfg.h"
1964 #include "libavutil/mathematics.h"
1965
1966 int main(void){
1967 #define width  256
1968 #define height 256
1969     int buffer[2][width*height];
1970     SnowContext s;
1971     int i;
1972     AVLFG prng;
1973     s.spatial_decomposition_count=6;
1974     s.spatial_decomposition_type=1;
1975
1976     s.temp_dwt_buffer  = av_mallocz(width * sizeof(DWTELEM));
1977     s.temp_idwt_buffer = av_mallocz(width * sizeof(IDWTELEM));
1978
1979     av_lfg_init(&prng, 1);
1980
1981     printf("testing 5/3 DWT\n");
1982     for(i=0; i<width*height; i++)
1983         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1984
1985     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1986     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1987
1988     for(i=0; i<width*height; i++)
1989         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1990
1991     printf("testing 9/7 DWT\n");
1992     s.spatial_decomposition_type=0;
1993     for(i=0; i<width*height; i++)
1994         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1995
1996     ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1997     ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1998
1999     for(i=0; i<width*height; i++)
2000         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
2001
2002     {
2003     int level, orientation, x, y;
2004     int64_t errors[8][4];
2005     int64_t g=0;
2006
2007         memset(errors, 0, sizeof(errors));
2008         s.spatial_decomposition_count=3;
2009         s.spatial_decomposition_type=0;
2010         for(level=0; level<s.spatial_decomposition_count; level++){
2011             for(orientation=level ? 1 : 0; orientation<4; orientation++){
2012                 int w= width  >> (s.spatial_decomposition_count-level);
2013                 int h= height >> (s.spatial_decomposition_count-level);
2014                 int stride= width  << (s.spatial_decomposition_count-level);
2015                 DWTELEM *buf= buffer[0];
2016                 int64_t error=0;
2017
2018                 if(orientation&1) buf+=w;
2019                 if(orientation>1) buf+=stride>>1;
2020
2021                 memset(buffer[0], 0, sizeof(int)*width*height);
2022                 buf[w/2 + h/2*stride]= 256*256;
2023                 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2024                 for(y=0; y<height; y++){
2025                     for(x=0; x<width; x++){
2026                         int64_t d= buffer[0][x + y*width];
2027                         error += d*d;
2028                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
2029                     }
2030                     if(FFABS(height/2-y)<9 && level==2) printf("\n");
2031                 }
2032                 error= (int)(sqrt(error)+0.5);
2033                 errors[level][orientation]= error;
2034                 if(g) g=av_gcd(g, error);
2035                 else g= error;
2036             }
2037         }
2038         printf("static int const visual_weight[][4]={\n");
2039         for(level=0; level<s.spatial_decomposition_count; level++){
2040             printf("  {");
2041             for(orientation=0; orientation<4; orientation++){
2042                 printf("%8"PRId64",", errors[level][orientation]/g);
2043             }
2044             printf("},\n");
2045         }
2046         printf("};\n");
2047         {
2048             int level=2;
2049             int w= width  >> (s.spatial_decomposition_count-level);
2050             //int h= height >> (s.spatial_decomposition_count-level);
2051             int stride= width  << (s.spatial_decomposition_count-level);
2052             DWTELEM *buf= buffer[0];
2053             int64_t error=0;
2054
2055             buf+=w;
2056             buf+=stride>>1;
2057
2058             memset(buffer[0], 0, sizeof(int)*width*height);
2059             for(y=0; y<height; y++){
2060                 for(x=0; x<width; x++){
2061                     int tab[4]={0,2,3,1};
2062                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2063                 }
2064             }
2065             ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2066             for(y=0; y<height; y++){
2067                 for(x=0; x<width; x++){
2068                     int64_t d= buffer[0][x + y*width];
2069                     error += d*d;
2070                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2071                 }
2072                 if(FFABS(height/2-y)<9) printf("\n");
2073             }
2074         }
2075
2076     }
2077     return 0;
2078 }
2079 #endif /* TEST */