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 "dsputil.h"
26 #include "dwt.h"
27 #include "snow.h"
28
29 #include "rangecoder.h"
30 #include "mathops.h"
31
32 #include "mpegvideo.h"
33 #include "h263.h"
34
35 #undef NDEBUG
36 #include <assert.h>
37
38 #define QUANTIZE2 0
39
40 #if QUANTIZE2==1
41 #define Q2_STEP 8
42
43 static void find_sse(SnowContext *s, Plane *p, int *score, int score_stride, IDWTELEM *r0, IDWTELEM *r1, int level, int orientation){
44     SubBand *b= &p->band[level][orientation];
45     int x, y;
46     int xo=0;
47     int yo=0;
48     int step= 1 << (s->spatial_decomposition_count - level);
49
50     if(orientation&1)
51         xo= step>>1;
52     if(orientation&2)
53         yo= step>>1;
54
55     //FIXME bias for nonzero ?
56     //FIXME optimize
57     memset(score, 0, sizeof(*score)*score_stride*((p->height + Q2_STEP-1)/Q2_STEP));
58     for(y=0; y<p->height; y++){
59         for(x=0; x<p->width; x++){
60             int sx= (x-xo + step/2) / step / Q2_STEP;
61             int sy= (y-yo + step/2) / step / Q2_STEP;
62             int v= r0[x + y*p->width] - r1[x + y*p->width];
63             assert(sx>=0 && sy>=0 && sx < score_stride);
64             v= ((v+8)>>4)<<4;
65             score[sx + sy*score_stride] += v*v;
66             assert(score[sx + sy*score_stride] >= 0);
67         }
68     }
69 }
70
71 static void dequantize_all(SnowContext *s, Plane *p, IDWTELEM *buffer, int width, int height){
72     int level, orientation;
73
74     for(level=0; level<s->spatial_decomposition_count; level++){
75         for(orientation=level ? 1 : 0; orientation<4; orientation++){
76             SubBand *b= &p->band[level][orientation];
77             IDWTELEM *dst= buffer + (b->ibuf - s->spatial_idwt_buffer);
78
79             dequantize(s, b, dst, b->stride);
80         }
81     }
82 }
83
84 static void dwt_quantize(SnowContext *s, Plane *p, DWTELEM *buffer, int width, int height, int stride, int type){
85     int level, orientation, ys, xs, x, y, pass;
86     IDWTELEM best_dequant[height * stride];
87     IDWTELEM idwt2_buffer[height * stride];
88     const int score_stride= (width + 10)/Q2_STEP;
89     int best_score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
90     int score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
91     int threshold= (s->m.lambda * s->m.lambda) >> 6;
92
93     //FIXME pass the copy cleanly ?
94
95 //    memcpy(dwt_buffer, buffer, height * stride * sizeof(DWTELEM));
96     ff_spatial_dwt(buffer, width, height, stride, type, s->spatial_decomposition_count);
97
98     for(level=0; level<s->spatial_decomposition_count; level++){
99         for(orientation=level ? 1 : 0; orientation<4; orientation++){
100             SubBand *b= &p->band[level][orientation];
101             IDWTELEM *dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
102              DWTELEM *src=       buffer + (b-> buf - s->spatial_dwt_buffer);
103             assert(src == b->buf); // code does not depend on this but it is true currently
104
105             quantize(s, b, dst, src, b->stride, s->qbias);
106         }
107     }
108     for(pass=0; pass<1; pass++){
109         if(s->qbias == 0) //keyframe
110             continue;
111         for(level=0; level<s->spatial_decomposition_count; level++){
112             for(orientation=level ? 1 : 0; orientation<4; orientation++){
113                 SubBand *b= &p->band[level][orientation];
114                 IDWTELEM *dst= idwt2_buffer + (b->ibuf - s->spatial_idwt_buffer);
115                 IDWTELEM *best_dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
116
117                 for(ys= 0; ys<Q2_STEP; ys++){
118                     for(xs= 0; xs<Q2_STEP; xs++){
119                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
120                         dequantize_all(s, p, idwt2_buffer, width, height);
121                         ff_spatial_idwt(idwt2_buffer, width, height, stride, type, s->spatial_decomposition_count);
122                         find_sse(s, p, best_score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
123                         memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
124                         for(y=ys; y<b->height; y+= Q2_STEP){
125                             for(x=xs; x<b->width; x+= Q2_STEP){
126                                 if(dst[x + y*b->stride]<0) dst[x + y*b->stride]++;
127                                 if(dst[x + y*b->stride]>0) dst[x + y*b->stride]--;
128                                 //FIXME try more than just --
129                             }
130                         }
131                         dequantize_all(s, p, idwt2_buffer, width, height);
132                         ff_spatial_idwt(idwt2_buffer, width, height, stride, type, s->spatial_decomposition_count);
133                         find_sse(s, p, score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
134                         for(y=ys; y<b->height; y+= Q2_STEP){
135                             for(x=xs; x<b->width; x+= Q2_STEP){
136                                 int score_idx= x/Q2_STEP + (y/Q2_STEP)*score_stride;
137                                 if(score[score_idx] <= best_score[score_idx] + threshold){
138                                     best_score[score_idx]= score[score_idx];
139                                     if(best_dst[x + y*b->stride]<0) best_dst[x + y*b->stride]++;
140                                     if(best_dst[x + y*b->stride]>0) best_dst[x + y*b->stride]--;
141                                     //FIXME copy instead
142                                 }
143                             }
144                         }
145                     }
146                 }
147             }
148         }
149     }
150     memcpy(s->spatial_idwt_buffer, best_dequant, height * stride * sizeof(IDWTELEM)); //FIXME work with that directly instead of copy at the end
151 }
152
153 #endif /* QUANTIZE2==1 */
154
155 #if CONFIG_SNOW_ENCODER
156 static av_cold int encode_init(AVCodecContext *avctx)
157 {
158     SnowContext *s = avctx->priv_data;
159     int plane_index;
160
161     if(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL){
162         av_log(avctx, AV_LOG_ERROR, "This codec is under development, files encoded with it may not be decodable with future versions!!!\n"
163                "Use vstrict=-2 / -strict -2 to use it anyway.\n");
164         return -1;
165     }
166
167     if(avctx->prediction_method == DWT_97
168        && (avctx->flags & CODEC_FLAG_QSCALE)
169        && avctx->global_quality == 0){
170         av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
171         return -1;
172     }
173
174     s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
175
176     s->mv_scale       = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
177     s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
178
179     for(plane_index=0; plane_index<3; plane_index++){
180         s->plane[plane_index].diag_mc= 1;
181         s->plane[plane_index].htaps= 6;
182         s->plane[plane_index].hcoeff[0]=  40;
183         s->plane[plane_index].hcoeff[1]= -10;
184         s->plane[plane_index].hcoeff[2]=   2;
185         s->plane[plane_index].fast_mc= 1;
186     }
187
188     ff_snow_common_init(avctx);
189     ff_snow_alloc_blocks(s);
190
191     s->version=0;
192
193     s->m.avctx   = avctx;
194     s->m.flags   = avctx->flags;
195     s->m.bit_rate= avctx->bit_rate;
196
197     s->m.me.temp      =
198     s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
199     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
200     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
201     s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
202     h263_encode_init(&s->m); //mv_penalty
203
204     s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
205
206     if(avctx->flags&CODEC_FLAG_PASS1){
207         if(!avctx->stats_out)
208             avctx->stats_out = av_mallocz(256);
209     }
210     if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
211         if(ff_rate_control_init(&s->m) < 0)
212             return -1;
213     }
214     s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
215
216     avctx->coded_frame= &s->current_picture;
217     switch(avctx->pix_fmt){
218 //    case PIX_FMT_YUV444P:
219 //    case PIX_FMT_YUV422P:
220     case PIX_FMT_YUV420P:
221 //     case PIX_FMT_GRAY8:
222 //    case PIX_FMT_YUV411P:
223 //    case PIX_FMT_YUV410P:
224         s->colorspace_type= 0;
225         break;
226 /*    case PIX_FMT_RGB32:
227         s->colorspace= 1;
228         break;*/
229     default:
230         av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
231         return -1;
232     }
233 //    avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
234     s->chroma_h_shift= 1;
235     s->chroma_v_shift= 1;
236
237     ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
238     ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
239
240     s->avctx->get_buffer(s->avctx, &s->input_picture);
241
242     if(s->avctx->me_method == ME_ITER){
243         int i;
244         int size= s->b_width * s->b_height << 2*s->block_max_depth;
245         for(i=0; i<s->max_ref_frames; i++){
246             s->ref_mvs[i]= av_mallocz(size*sizeof(int16_t[2]));
247             s->ref_scores[i]= av_mallocz(size*sizeof(uint32_t));
248         }
249     }
250
251     return 0;
252 }
253
254 //near copy & paste from dsputil, FIXME
255 static int pix_sum(uint8_t * pix, int line_size, int w)
256 {
257     int s, i, j;
258
259     s = 0;
260     for (i = 0; i < w; i++) {
261         for (j = 0; j < w; j++) {
262             s += pix[0];
263             pix ++;
264         }
265         pix += line_size - w;
266     }
267     return s;
268 }
269
270 //near copy & paste from dsputil, FIXME
271 static int pix_norm1(uint8_t * pix, int line_size, int w)
272 {
273     int s, i, j;
274     uint32_t *sq = ff_squareTbl + 256;
275
276     s = 0;
277     for (i = 0; i < w; i++) {
278         for (j = 0; j < w; j ++) {
279             s += sq[pix[0]];
280             pix ++;
281         }
282         pix += line_size - w;
283     }
284     return s;
285 }
286
287 //FIXME copy&paste
288 #define P_LEFT P[1]
289 #define P_TOP P[2]
290 #define P_TOPRIGHT P[3]
291 #define P_MEDIAN P[4]
292 #define P_MV1 P[9]
293 #define FLAG_QPEL   1 //must be 1
294
295 static int encode_q_branch(SnowContext *s, int level, int x, int y){
296     uint8_t p_buffer[1024];
297     uint8_t i_buffer[1024];
298     uint8_t p_state[sizeof(s->block_state)];
299     uint8_t i_state[sizeof(s->block_state)];
300     RangeCoder pc, ic;
301     uint8_t *pbbak= s->c.bytestream;
302     uint8_t *pbbak_start= s->c.bytestream_start;
303     int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
304     const int w= s->b_width  << s->block_max_depth;
305     const int h= s->b_height << s->block_max_depth;
306     const int rem_depth= s->block_max_depth - level;
307     const int index= (x + y*w) << rem_depth;
308     const int block_w= 1<<(LOG2_MB_SIZE - level);
309     int trx= (x+1)<<rem_depth;
310     int try= (y+1)<<rem_depth;
311     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
312     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
313     const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
314     const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
315     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
316     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
317     int pl = left->color[0];
318     int pcb= left->color[1];
319     int pcr= left->color[2];
320     int pmx, pmy;
321     int mx=0, my=0;
322     int l,cr,cb;
323     const int stride= s->current_picture.linesize[0];
324     const int uvstride= s->current_picture.linesize[1];
325     uint8_t *current_data[3]= { s->input_picture.data[0] + (x + y*  stride)*block_w,
326                                 s->input_picture.data[1] + (x + y*uvstride)*block_w/2,
327                                 s->input_picture.data[2] + (x + y*uvstride)*block_w/2};
328     int P[10][2];
329     int16_t last_mv[3][2];
330     int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
331     const int shift= 1+qpel;
332     MotionEstContext *c= &s->m.me;
333     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
334     int mx_context= av_log2(2*FFABS(left->mx - top->mx));
335     int my_context= av_log2(2*FFABS(left->my - top->my));
336     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
337     int ref, best_ref, ref_score, ref_mx, ref_my;
338
339     assert(sizeof(s->block_state) >= 256);
340     if(s->keyframe){
341         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
342         return 0;
343     }
344
345 //    clip predictors / edge ?
346
347     P_LEFT[0]= left->mx;
348     P_LEFT[1]= left->my;
349     P_TOP [0]= top->mx;
350     P_TOP [1]= top->my;
351     P_TOPRIGHT[0]= tr->mx;
352     P_TOPRIGHT[1]= tr->my;
353
354     last_mv[0][0]= s->block[index].mx;
355     last_mv[0][1]= s->block[index].my;
356     last_mv[1][0]= right->mx;
357     last_mv[1][1]= right->my;
358     last_mv[2][0]= bottom->mx;
359     last_mv[2][1]= bottom->my;
360
361     s->m.mb_stride=2;
362     s->m.mb_x=
363     s->m.mb_y= 0;
364     c->skip= 0;
365
366     assert(c->  stride ==   stride);
367     assert(c->uvstride == uvstride);
368
369     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
370     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
371     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
372     c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
373
374     c->xmin = - x*block_w - 16+3;
375     c->ymin = - y*block_w - 16+3;
376     c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
377     c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
378
379     if(P_LEFT[0]     > (c->xmax<<shift)) P_LEFT[0]    = (c->xmax<<shift);
380     if(P_LEFT[1]     > (c->ymax<<shift)) P_LEFT[1]    = (c->ymax<<shift);
381     if(P_TOP[0]      > (c->xmax<<shift)) P_TOP[0]     = (c->xmax<<shift);
382     if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
383     if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
384     if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
385     if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
386
387     P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
388     P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
389
390     if (!y) {
391         c->pred_x= P_LEFT[0];
392         c->pred_y= P_LEFT[1];
393     } else {
394         c->pred_x = P_MEDIAN[0];
395         c->pred_y = P_MEDIAN[1];
396     }
397
398     score= INT_MAX;
399     best_ref= 0;
400     for(ref=0; ref<s->ref_frames; ref++){
401         init_ref(c, current_data, s->last_picture[ref].data, NULL, block_w*x, block_w*y, 0);
402
403         ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
404                                          (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
405
406         assert(ref_mx >= c->xmin);
407         assert(ref_mx <= c->xmax);
408         assert(ref_my >= c->ymin);
409         assert(ref_my <= c->ymax);
410
411         ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
412         ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
413         ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
414         if(s->ref_mvs[ref]){
415             s->ref_mvs[ref][index][0]= ref_mx;
416             s->ref_mvs[ref][index][1]= ref_my;
417             s->ref_scores[ref][index]= ref_score;
418         }
419         if(score > ref_score){
420             score= ref_score;
421             best_ref= ref;
422             mx= ref_mx;
423             my= ref_my;
424         }
425     }
426     //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
427
428   //  subpel search
429     base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
430     pc= s->c;
431     pc.bytestream_start=
432     pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
433     memcpy(p_state, s->block_state, sizeof(s->block_state));
434
435     if(level!=s->block_max_depth)
436         put_rac(&pc, &p_state[4 + s_context], 1);
437     put_rac(&pc, &p_state[1 + left->type + top->type], 0);
438     if(s->ref_frames > 1)
439         put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
440     pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
441     put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
442     put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
443     p_len= pc.bytestream - pc.bytestream_start;
444     score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
445
446     block_s= block_w*block_w;
447     sum = pix_sum(current_data[0], stride, block_w);
448     l= (sum + block_s/2)/block_s;
449     iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
450
451     block_s= block_w*block_w>>2;
452     sum = pix_sum(current_data[1], uvstride, block_w>>1);
453     cb= (sum + block_s/2)/block_s;
454 //    iscore += pix_norm1(&current_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
455     sum = pix_sum(current_data[2], uvstride, block_w>>1);
456     cr= (sum + block_s/2)/block_s;
457 //    iscore += pix_norm1(&current_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
458
459     ic= s->c;
460     ic.bytestream_start=
461     ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
462     memcpy(i_state, s->block_state, sizeof(s->block_state));
463     if(level!=s->block_max_depth)
464         put_rac(&ic, &i_state[4 + s_context], 1);
465     put_rac(&ic, &i_state[1 + left->type + top->type], 1);
466     put_symbol(&ic, &i_state[32],  l-pl , 1);
467     put_symbol(&ic, &i_state[64], cb-pcb, 1);
468     put_symbol(&ic, &i_state[96], cr-pcr, 1);
469     i_len= ic.bytestream - ic.bytestream_start;
470     iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
471
472 //    assert(score==256*256*256*64-1);
473     assert(iscore < 255*255*256 + s->lambda2*10);
474     assert(iscore >= 0);
475     assert(l>=0 && l<=255);
476     assert(pl>=0 && pl<=255);
477
478     if(level==0){
479         int varc= iscore >> 8;
480         int vard= score >> 8;
481         if (vard <= 64 || vard < varc)
482             c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
483         else
484             c->scene_change_score+= s->m.qscale;
485     }
486
487     if(level!=s->block_max_depth){
488         put_rac(&s->c, &s->block_state[4 + s_context], 0);
489         score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
490         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
491         score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
492         score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
493         score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
494
495         if(score2 < score && score2 < iscore)
496             return score2;
497     }
498
499     if(iscore < score){
500         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
501         memcpy(pbbak, i_buffer, i_len);
502         s->c= ic;
503         s->c.bytestream_start= pbbak_start;
504         s->c.bytestream= pbbak + i_len;
505         set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
506         memcpy(s->block_state, i_state, sizeof(s->block_state));
507         return iscore;
508     }else{
509         memcpy(pbbak, p_buffer, p_len);
510         s->c= pc;
511         s->c.bytestream_start= pbbak_start;
512         s->c.bytestream= pbbak + p_len;
513         set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
514         memcpy(s->block_state, p_state, sizeof(s->block_state));
515         return score;
516     }
517 }
518
519 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
520     const int w= s->b_width  << s->block_max_depth;
521     const int rem_depth= s->block_max_depth - level;
522     const int index= (x + y*w) << rem_depth;
523     int trx= (x+1)<<rem_depth;
524     BlockNode *b= &s->block[index];
525     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
526     const BlockNode *top   = y ? &s->block[index-w] : &null_block;
527     const BlockNode *tl    = y && x ? &s->block[index-w-1] : left;
528     const BlockNode *tr    = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
529     int pl = left->color[0];
530     int pcb= left->color[1];
531     int pcr= left->color[2];
532     int pmx, pmy;
533     int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
534     int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
535     int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
536     int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
537
538     if(s->keyframe){
539         set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
540         return;
541     }
542
543     if(level!=s->block_max_depth){
544         if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
545             put_rac(&s->c, &s->block_state[4 + s_context], 1);
546         }else{
547             put_rac(&s->c, &s->block_state[4 + s_context], 0);
548             encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
549             encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
550             encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
551             encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
552             return;
553         }
554     }
555     if(b->type & BLOCK_INTRA){
556         pred_mv(s, &pmx, &pmy, 0, left, top, tr);
557         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
558         put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
559         put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
560         put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
561         set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
562     }else{
563         pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
564         put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
565         if(s->ref_frames > 1)
566             put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
567         put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
568         put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
569         set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
570     }
571 }
572
573 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
574     int i, x2, y2;
575     Plane *p= &s->plane[plane_index];
576     const int block_size = MB_SIZE >> s->block_max_depth;
577     const int block_w    = plane_index ? block_size/2 : block_size;
578     const uint8_t *obmc  = plane_index ? obmc_tab[s->block_max_depth+1] : obmc_tab[s->block_max_depth];
579     const int obmc_stride= plane_index ? block_size : 2*block_size;
580     const int ref_stride= s->current_picture.linesize[plane_index];
581     uint8_t *src= s-> input_picture.data[plane_index];
582     IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
583     const int b_stride = s->b_width << s->block_max_depth;
584     const int w= p->width;
585     const int h= p->height;
586     int index= mb_x + mb_y*b_stride;
587     BlockNode *b= &s->block[index];
588     BlockNode backup= *b;
589     int ab=0;
590     int aa=0;
591
592     b->type|= BLOCK_INTRA;
593     b->color[plane_index]= 0;
594     memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
595
596     for(i=0; i<4; i++){
597         int mb_x2= mb_x + (i &1) - 1;
598         int mb_y2= mb_y + (i>>1) - 1;
599         int x= block_w*mb_x2 + block_w/2;
600         int y= block_w*mb_y2 + block_w/2;
601
602         add_yblock(s, 0, NULL, dst + ((i&1)+(i>>1)*obmc_stride)*block_w, NULL, obmc,
603                     x, y, block_w, block_w, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
604
605         for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_w); y2++){
606             for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
607                 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_w*mb_y - block_w/2))*obmc_stride;
608                 int obmc_v= obmc[index];
609                 int d;
610                 if(y<0) obmc_v += obmc[index + block_w*obmc_stride];
611                 if(x<0) obmc_v += obmc[index + block_w];
612                 if(y+block_w>h) obmc_v += obmc[index - block_w*obmc_stride];
613                 if(x+block_w>w) obmc_v += obmc[index - block_w];
614                 //FIXME precalculate this or simplify it somehow else
615
616                 d = -dst[index] + (1<<(FRAC_BITS-1));
617                 dst[index] = d;
618                 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
619                 aa += obmc_v * obmc_v; //FIXME precalculate this
620             }
621         }
622     }
623     *b= backup;
624
625     return av_clip(((ab<<LOG2_OBMC_MAX) + aa/2)/aa, 0, 255); //FIXME we should not need clipping
626 }
627
628 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
629     const int b_stride = s->b_width << s->block_max_depth;
630     const int b_height = s->b_height<< s->block_max_depth;
631     int index= x + y*b_stride;
632     const BlockNode *b     = &s->block[index];
633     const BlockNode *left  = x ? &s->block[index-1] : &null_block;
634     const BlockNode *top   = y ? &s->block[index-b_stride] : &null_block;
635     const BlockNode *tl    = y && x ? &s->block[index-b_stride-1] : left;
636     const BlockNode *tr    = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
637     int dmx, dmy;
638 //  int mx_context= av_log2(2*FFABS(left->mx - top->mx));
639 //  int my_context= av_log2(2*FFABS(left->my - top->my));
640
641     if(x<0 || x>=b_stride || y>=b_height)
642         return 0;
643 /*
644 1            0      0
645 01X          1-2    1
646 001XX        3-6    2-3
647 0001XXX      7-14   4-7
648 00001XXXX   15-30   8-15
649 */
650 //FIXME try accurate rate
651 //FIXME intra and inter predictors if surrounding blocks are not the same type
652     if(b->type & BLOCK_INTRA){
653         return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
654                    + av_log2(2*FFABS(left->color[1] - b->color[1]))
655                    + av_log2(2*FFABS(left->color[2] - b->color[2])));
656     }else{
657         pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
658         dmx-= b->mx;
659         dmy-= b->my;
660         return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
661                     + av_log2(2*FFABS(dmy))
662                     + av_log2(2*b->ref));
663     }
664 }
665
666 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, const uint8_t *obmc_edged){
667     Plane *p= &s->plane[plane_index];
668     const int block_size = MB_SIZE >> s->block_max_depth;
669     const int block_w    = plane_index ? block_size/2 : block_size;
670     const int obmc_stride= plane_index ? block_size : 2*block_size;
671     const int ref_stride= s->current_picture.linesize[plane_index];
672     uint8_t *dst= s->current_picture.data[plane_index];
673     uint8_t *src= s->  input_picture.data[plane_index];
674     IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
675     uint8_t *cur = s->scratchbuf;
676     uint8_t tmp[ref_stride*(2*MB_SIZE+HTAPS_MAX-1)];
677     const int b_stride = s->b_width << s->block_max_depth;
678     const int b_height = s->b_height<< s->block_max_depth;
679     const int w= p->width;
680     const int h= p->height;
681     int distortion;
682     int rate= 0;
683     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
684     int sx= block_w*mb_x - block_w/2;
685     int sy= block_w*mb_y - block_w/2;
686     int x0= FFMAX(0,-sx);
687     int y0= FFMAX(0,-sy);
688     int x1= FFMIN(block_w*2, w-sx);
689     int y1= FFMIN(block_w*2, h-sy);
690     int i,x,y;
691
692     ff_snow_pred_block(s, cur, tmp, ref_stride, sx, sy, block_w*2, block_w*2, &s->block[mb_x + mb_y*b_stride], plane_index, w, h);
693
694     for(y=y0; y<y1; y++){
695         const uint8_t *obmc1= obmc_edged + y*obmc_stride;
696         const IDWTELEM *pred1 = pred + y*obmc_stride;
697         uint8_t *cur1 = cur + y*ref_stride;
698         uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
699         for(x=x0; x<x1; x++){
700 #if FRAC_BITS >= LOG2_OBMC_MAX
701             int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
702 #else
703             int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
704 #endif
705             v = (v + pred1[x]) >> FRAC_BITS;
706             if(v&(~255)) v= ~(v>>31);
707             dst1[x] = v;
708         }
709     }
710
711     /* copy the regions where obmc[] = (uint8_t)256 */
712     if(LOG2_OBMC_MAX == 8
713         && (mb_x == 0 || mb_x == b_stride-1)
714         && (mb_y == 0 || mb_y == b_height-1)){
715         if(mb_x == 0)
716             x1 = block_w;
717         else
718             x0 = block_w;
719         if(mb_y == 0)
720             y1 = block_w;
721         else
722             y0 = block_w;
723         for(y=y0; y<y1; y++)
724             memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
725     }
726
727     if(block_w==16){
728         /* FIXME rearrange dsputil to fit 32x32 cmp functions */
729         /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
730         /* FIXME cmps overlap but do not cover the wavelet's whole support.
731          * So improving the score of one block is not strictly guaranteed
732          * to improve the score of the whole frame, thus iterative motion
733          * estimation does not always converge. */
734         if(s->avctx->me_cmp == FF_CMP_W97)
735             distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
736         else if(s->avctx->me_cmp == FF_CMP_W53)
737             distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
738         else{
739             distortion = 0;
740             for(i=0; i<4; i++){
741                 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
742                 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
743             }
744         }
745     }else{
746         assert(block_w==8);
747         distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
748     }
749
750     if(plane_index==0){
751         for(i=0; i<4; i++){
752 /* ..RRr
753  * .RXx.
754  * rxx..
755  */
756             rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
757         }
758         if(mb_x == b_stride-2)
759             rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
760     }
761     return distortion + rate*penalty_factor;
762 }
763
764 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
765     int i, y2;
766     Plane *p= &s->plane[plane_index];
767     const int block_size = MB_SIZE >> s->block_max_depth;
768     const int block_w    = plane_index ? block_size/2 : block_size;
769     const uint8_t *obmc  = plane_index ? obmc_tab[s->block_max_depth+1] : obmc_tab[s->block_max_depth];
770     const int obmc_stride= plane_index ? block_size : 2*block_size;
771     const int ref_stride= s->current_picture.linesize[plane_index];
772     uint8_t *dst= s->current_picture.data[plane_index];
773     uint8_t *src= s-> input_picture.data[plane_index];
774     //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
775     // const has only been removed from zero_dst to suppress a warning
776     static IDWTELEM zero_dst[4096]; //FIXME
777     const int b_stride = s->b_width << s->block_max_depth;
778     const int w= p->width;
779     const int h= p->height;
780     int distortion= 0;
781     int rate= 0;
782     const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
783
784     for(i=0; i<9; i++){
785         int mb_x2= mb_x + (i%3) - 1;
786         int mb_y2= mb_y + (i/3) - 1;
787         int x= block_w*mb_x2 + block_w/2;
788         int y= block_w*mb_y2 + block_w/2;
789
790         add_yblock(s, 0, NULL, zero_dst, dst, obmc,
791                    x, y, block_w, block_w, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
792
793         //FIXME find a cleaner/simpler way to skip the outside stuff
794         for(y2= y; y2<0; y2++)
795             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
796         for(y2= h; y2<y+block_w; y2++)
797             memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
798         if(x<0){
799             for(y2= y; y2<y+block_w; y2++)
800                 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
801         }
802         if(x+block_w > w){
803             for(y2= y; y2<y+block_w; y2++)
804                 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
805         }
806
807         assert(block_w== 8 || block_w==16);
808         distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_w);
809     }
810
811     if(plane_index==0){
812         BlockNode *b= &s->block[mb_x+mb_y*b_stride];
813         int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
814
815 /* ..RRRr
816  * .RXXx.
817  * .RXXx.
818  * rxxx.
819  */
820         if(merged)
821             rate = get_block_bits(s, mb_x, mb_y, 2);
822         for(i=merged?4:0; i<9; i++){
823             static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
824             rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
825         }
826     }
827     return distortion + rate*penalty_factor;
828 }
829
830 static int encode_subband_c0run(SnowContext *s, SubBand *b, IDWTELEM *src, IDWTELEM *parent, int stride, int orientation){
831     const int w= b->width;
832     const int h= b->height;
833     int x, y;
834
835     if(1){
836         int run=0;
837         int runs[w*h];
838         int run_index=0;
839         int max_index;
840
841         for(y=0; y<h; y++){
842             for(x=0; x<w; x++){
843                 int v, p=0;
844                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
845                 v= src[x + y*stride];
846
847                 if(y){
848                     t= src[x + (y-1)*stride];
849                     if(x){
850                         lt= src[x - 1 + (y-1)*stride];
851                     }
852                     if(x + 1 < w){
853                         rt= src[x + 1 + (y-1)*stride];
854                     }
855                 }
856                 if(x){
857                     l= src[x - 1 + y*stride];
858                     /*if(x > 1){
859                         if(orientation==1) ll= src[y + (x-2)*stride];
860                         else               ll= src[x - 2 + y*stride];
861                     }*/
862                 }
863                 if(parent){
864                     int px= x>>1;
865                     int py= y>>1;
866                     if(px<b->parent->width && py<b->parent->height)
867                         p= parent[px + py*2*stride];
868                 }
869                 if(!(/*ll|*/l|lt|t|rt|p)){
870                     if(v){
871                         runs[run_index++]= run;
872                         run=0;
873                     }else{
874                         run++;
875                     }
876                 }
877             }
878         }
879         max_index= run_index;
880         runs[run_index++]= run;
881         run_index=0;
882         run= runs[run_index++];
883
884         put_symbol2(&s->c, b->state[30], max_index, 0);
885         if(run_index <= max_index)
886             put_symbol2(&s->c, b->state[1], run, 3);
887
888         for(y=0; y<h; y++){
889             if(s->c.bytestream_end - s->c.bytestream < w*40){
890                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
891                 return -1;
892             }
893             for(x=0; x<w; x++){
894                 int v, p=0;
895                 int /*ll=0, */l=0, lt=0, t=0, rt=0;
896                 v= src[x + y*stride];
897
898                 if(y){
899                     t= src[x + (y-1)*stride];
900                     if(x){
901                         lt= src[x - 1 + (y-1)*stride];
902                     }
903                     if(x + 1 < w){
904                         rt= src[x + 1 + (y-1)*stride];
905                     }
906                 }
907                 if(x){
908                     l= src[x - 1 + y*stride];
909                     /*if(x > 1){
910                         if(orientation==1) ll= src[y + (x-2)*stride];
911                         else               ll= src[x - 2 + y*stride];
912                     }*/
913                 }
914                 if(parent){
915                     int px= x>>1;
916                     int py= y>>1;
917                     if(px<b->parent->width && py<b->parent->height)
918                         p= parent[px + py*2*stride];
919                 }
920                 if(/*ll|*/l|lt|t|rt|p){
921                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
922
923                     put_rac(&s->c, &b->state[0][context], !!v);
924                 }else{
925                     if(!run){
926                         run= runs[run_index++];
927
928                         if(run_index <= max_index)
929                             put_symbol2(&s->c, b->state[1], run, 3);
930                         assert(v);
931                     }else{
932                         run--;
933                         assert(!v);
934                     }
935                 }
936                 if(v){
937                     int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
938                     int l2= 2*FFABS(l) + (l<0);
939                     int t2= 2*FFABS(t) + (t<0);
940
941                     put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
942                     put_rac(&s->c, &b->state[0][16 + 1 + 3 + quant3bA[l2&0xFF] + 3*quant3bA[t2&0xFF]], v<0);
943                 }
944             }
945         }
946     }
947     return 0;
948 }
949
950 static int encode_subband(SnowContext *s, SubBand *b, IDWTELEM *src, IDWTELEM *parent, int stride, int orientation){
951 //    encode_subband_qtree(s, b, src, parent, stride, orientation);
952 //    encode_subband_z0run(s, b, src, parent, stride, orientation);
953     return encode_subband_c0run(s, b, src, parent, stride, orientation);
954 //    encode_subband_dzr(s, b, src, parent, stride, orientation);
955 }
956
957 static av_always_inline int check_block(SnowContext *s, int mb_x, int mb_y, int p[3], int intra, const uint8_t *obmc_edged, int *best_rd){
958     const int b_stride= s->b_width << s->block_max_depth;
959     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
960     BlockNode backup= *block;
961     unsigned value;
962     int rd, index;
963
964     assert(mb_x>=0 && mb_y>=0);
965     assert(mb_x<b_stride);
966
967     if(intra){
968         block->color[0] = p[0];
969         block->color[1] = p[1];
970         block->color[2] = p[2];
971         block->type |= BLOCK_INTRA;
972     }else{
973         index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
974         value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
975         if(s->me_cache[index] == value)
976             return 0;
977         s->me_cache[index]= value;
978
979         block->mx= p[0];
980         block->my= p[1];
981         block->type &= ~BLOCK_INTRA;
982     }
983
984     rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
985
986 //FIXME chroma
987     if(rd < *best_rd){
988         *best_rd= rd;
989         return 1;
990     }else{
991         *block= backup;
992         return 0;
993     }
994 }
995
996 /* special case for int[2] args we discard afterwards,
997  * fixes compilation problem with gcc 2.95 */
998 static av_always_inline int check_block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, const uint8_t *obmc_edged, int *best_rd){
999     int p[2] = {p0, p1};
1000     return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
1001 }
1002
1003 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){
1004     const int b_stride= s->b_width << s->block_max_depth;
1005     BlockNode *block= &s->block[mb_x + mb_y * b_stride];
1006     BlockNode backup[4]= {block[0], block[1], block[b_stride], block[b_stride+1]};
1007     unsigned value;
1008     int rd, index;
1009
1010     assert(mb_x>=0 && mb_y>=0);
1011     assert(mb_x<b_stride);
1012     assert(((mb_x|mb_y)&1) == 0);
1013
1014     index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
1015     value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
1016     if(s->me_cache[index] == value)
1017         return 0;
1018     s->me_cache[index]= value;
1019
1020     block->mx= p0;
1021     block->my= p1;
1022     block->ref= ref;
1023     block->type &= ~BLOCK_INTRA;
1024     block[1]= block[b_stride]= block[b_stride+1]= *block;
1025
1026     rd= get_4block_rd(s, mb_x, mb_y, 0);
1027
1028 //FIXME chroma
1029     if(rd < *best_rd){
1030         *best_rd= rd;
1031         return 1;
1032     }else{
1033         block[0]= backup[0];
1034         block[1]= backup[1];
1035         block[b_stride]= backup[2];
1036         block[b_stride+1]= backup[3];
1037         return 0;
1038     }
1039 }
1040
1041 static void iterative_me(SnowContext *s){
1042     int pass, mb_x, mb_y;
1043     const int b_width = s->b_width  << s->block_max_depth;
1044     const int b_height= s->b_height << s->block_max_depth;
1045     const int b_stride= b_width;
1046     int color[3];
1047
1048     {
1049         RangeCoder r = s->c;
1050         uint8_t state[sizeof(s->block_state)];
1051         memcpy(state, s->block_state, sizeof(s->block_state));
1052         for(mb_y= 0; mb_y<s->b_height; mb_y++)
1053             for(mb_x= 0; mb_x<s->b_width; mb_x++)
1054                 encode_q_branch(s, 0, mb_x, mb_y);
1055         s->c = r;
1056         memcpy(s->block_state, state, sizeof(s->block_state));
1057     }
1058
1059     for(pass=0; pass<25; pass++){
1060         int change= 0;
1061
1062         for(mb_y= 0; mb_y<b_height; mb_y++){
1063             for(mb_x= 0; mb_x<b_width; mb_x++){
1064                 int dia_change, i, j, ref;
1065                 int best_rd= INT_MAX, ref_rd;
1066                 BlockNode backup, ref_b;
1067                 const int index= mb_x + mb_y * b_stride;
1068                 BlockNode *block= &s->block[index];
1069                 BlockNode *tb =                   mb_y            ? &s->block[index-b_stride  ] : NULL;
1070                 BlockNode *lb = mb_x                              ? &s->block[index         -1] : NULL;
1071                 BlockNode *rb = mb_x+1<b_width                    ? &s->block[index         +1] : NULL;
1072                 BlockNode *bb =                   mb_y+1<b_height ? &s->block[index+b_stride  ] : NULL;
1073                 BlockNode *tlb= mb_x           && mb_y            ? &s->block[index-b_stride-1] : NULL;
1074                 BlockNode *trb= mb_x+1<b_width && mb_y            ? &s->block[index-b_stride+1] : NULL;
1075                 BlockNode *blb= mb_x           && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1076                 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1077                 const int b_w= (MB_SIZE >> s->block_max_depth);
1078                 uint8_t obmc_edged[b_w*2][b_w*2];
1079
1080                 if(pass && (block->type & BLOCK_OPT))
1081                     continue;
1082                 block->type |= BLOCK_OPT;
1083
1084                 backup= *block;
1085
1086                 if(!s->me_cache_generation)
1087                     memset(s->me_cache, 0, sizeof(s->me_cache));
1088                 s->me_cache_generation += 1<<22;
1089
1090                 //FIXME precalculate
1091                 {
1092                     int x, y;
1093                     memcpy(obmc_edged, obmc_tab[s->block_max_depth], b_w*b_w*4);
1094                     if(mb_x==0)
1095                         for(y=0; y<b_w*2; y++)
1096                             memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1097                     if(mb_x==b_stride-1)
1098                         for(y=0; y<b_w*2; y++)
1099                             memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1100                     if(mb_y==0){
1101                         for(x=0; x<b_w*2; x++)
1102                             obmc_edged[0][x] += obmc_edged[b_w-1][x];
1103                         for(y=1; y<b_w; y++)
1104                             memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1105                     }
1106                     if(mb_y==b_height-1){
1107                         for(x=0; x<b_w*2; x++)
1108                             obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1109                         for(y=b_w; y<b_w*2-1; y++)
1110                             memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1111                     }
1112                 }
1113
1114                 //skip stuff outside the picture
1115                 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1116                     uint8_t *src= s->  input_picture.data[0];
1117                     uint8_t *dst= s->current_picture.data[0];
1118                     const int stride= s->current_picture.linesize[0];
1119                     const int block_w= MB_SIZE >> s->block_max_depth;
1120                     const int sx= block_w*mb_x - block_w/2;
1121                     const int sy= block_w*mb_y - block_w/2;
1122                     const int w= s->plane[0].width;
1123                     const int h= s->plane[0].height;
1124                     int y;
1125
1126                     for(y=sy; y<0; y++)
1127                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1128                     for(y=h; y<sy+block_w*2; y++)
1129                         memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1130                     if(sx<0){
1131                         for(y=sy; y<sy+block_w*2; y++)
1132                             memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1133                     }
1134                     if(sx+block_w*2 > w){
1135                         for(y=sy; y<sy+block_w*2; y++)
1136                             memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1137                     }
1138                 }
1139
1140                 // intra(black) = neighbors' contribution to the current block
1141                 for(i=0; i<3; i++)
1142                     color[i]= get_dc(s, mb_x, mb_y, i);
1143
1144                 // get previous score (cannot be cached due to OBMC)
1145                 if(pass > 0 && (block->type&BLOCK_INTRA)){
1146                     int color0[3]= {block->color[0], block->color[1], block->color[2]};
1147                     check_block(s, mb_x, mb_y, color0, 1, *obmc_edged, &best_rd);
1148                 }else
1149                     check_block_inter(s, mb_x, mb_y, block->mx, block->my, *obmc_edged, &best_rd);
1150
1151                 ref_b= *block;
1152                 ref_rd= best_rd;
1153                 for(ref=0; ref < s->ref_frames; ref++){
1154                     int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1155                     if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1156                         continue;
1157                     block->ref= ref;
1158                     best_rd= INT_MAX;
1159
1160                     check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], *obmc_edged, &best_rd);
1161                     check_block_inter(s, mb_x, mb_y, 0, 0, *obmc_edged, &best_rd);
1162                     if(tb)
1163                         check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], *obmc_edged, &best_rd);
1164                     if(lb)
1165                         check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], *obmc_edged, &best_rd);
1166                     if(rb)
1167                         check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], *obmc_edged, &best_rd);
1168                     if(bb)
1169                         check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], *obmc_edged, &best_rd);
1170
1171                     /* fullpel ME */
1172                     //FIXME avoid subpel interpolation / round to nearest integer
1173                     do{
1174                         dia_change=0;
1175                         for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1176                             for(j=0; j<i; j++){
1177                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), *obmc_edged, &best_rd);
1178                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), *obmc_edged, &best_rd);
1179                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), *obmc_edged, &best_rd);
1180                                 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), *obmc_edged, &best_rd);
1181                             }
1182                         }
1183                     }while(dia_change);
1184                     /* subpel ME */
1185                     do{
1186                         static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1187                         dia_change=0;
1188                         for(i=0; i<8; i++)
1189                             dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], *obmc_edged, &best_rd);
1190                     }while(dia_change);
1191                     //FIXME or try the standard 2 pass qpel or similar
1192
1193                     mvr[0][0]= block->mx;
1194                     mvr[0][1]= block->my;
1195                     if(ref_rd > best_rd){
1196                         ref_rd= best_rd;
1197                         ref_b= *block;
1198                     }
1199                 }
1200                 best_rd= ref_rd;
1201                 *block= ref_b;
1202                 check_block(s, mb_x, mb_y, color, 1, *obmc_edged, &best_rd);
1203                 //FIXME RD style color selection
1204                 if(!same_block(block, &backup)){
1205                     if(tb ) tb ->type &= ~BLOCK_OPT;
1206                     if(lb ) lb ->type &= ~BLOCK_OPT;
1207                     if(rb ) rb ->type &= ~BLOCK_OPT;
1208                     if(bb ) bb ->type &= ~BLOCK_OPT;
1209                     if(tlb) tlb->type &= ~BLOCK_OPT;
1210                     if(trb) trb->type &= ~BLOCK_OPT;
1211                     if(blb) blb->type &= ~BLOCK_OPT;
1212                     if(brb) brb->type &= ~BLOCK_OPT;
1213                     change ++;
1214                 }
1215             }
1216         }
1217         av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1218         if(!change)
1219             break;
1220     }
1221
1222     if(s->block_max_depth == 1){
1223         int change= 0;
1224         for(mb_y= 0; mb_y<b_height; mb_y+=2){
1225             for(mb_x= 0; mb_x<b_width; mb_x+=2){
1226                 int i;
1227                 int best_rd, init_rd;
1228                 const int index= mb_x + mb_y * b_stride;
1229                 BlockNode *b[4];
1230
1231                 b[0]= &s->block[index];
1232                 b[1]= b[0]+1;
1233                 b[2]= b[0]+b_stride;
1234                 b[3]= b[2]+1;
1235                 if(same_block(b[0], b[1]) &&
1236                    same_block(b[0], b[2]) &&
1237                    same_block(b[0], b[3]))
1238                     continue;
1239
1240                 if(!s->me_cache_generation)
1241                     memset(s->me_cache, 0, sizeof(s->me_cache));
1242                 s->me_cache_generation += 1<<22;
1243
1244                 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1245
1246                 //FIXME more multiref search?
1247                 check_4block_inter(s, mb_x, mb_y,
1248                                    (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1249                                    (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1250
1251                 for(i=0; i<4; i++)
1252                     if(!(b[i]->type&BLOCK_INTRA))
1253                         check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1254
1255                 if(init_rd != best_rd)
1256                     change++;
1257             }
1258         }
1259         av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1260     }
1261 }
1262
1263 static void encode_blocks(SnowContext *s, int search){
1264     int x, y;
1265     int w= s->b_width;
1266     int h= s->b_height;
1267
1268     if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1269         iterative_me(s);
1270
1271     for(y=0; y<h; y++){
1272         if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1273             av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1274             return;
1275         }
1276         for(x=0; x<w; x++){
1277             if(s->avctx->me_method == ME_ITER || !search)
1278                 encode_q_branch2(s, 0, x, y);
1279             else
1280                 encode_q_branch (s, 0, x, y);
1281         }
1282     }
1283 }
1284
1285 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1286     const int w= b->width;
1287     const int h= b->height;
1288     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1289     const int qmul= qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1290     int x,y, thres1, thres2;
1291
1292     if(s->qlog == LOSSLESS_QLOG){
1293         for(y=0; y<h; y++)
1294             for(x=0; x<w; x++)
1295                 dst[x + y*stride]= src[x + y*stride];
1296         return;
1297     }
1298
1299     bias= bias ? 0 : (3*qmul)>>3;
1300     thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1301     thres2= 2*thres1;
1302
1303     if(!bias){
1304         for(y=0; y<h; y++){
1305             for(x=0; x<w; x++){
1306                 int i= src[x + y*stride];
1307
1308                 if((unsigned)(i+thres1) > thres2){
1309                     if(i>=0){
1310                         i<<= QEXPSHIFT;
1311                         i/= qmul; //FIXME optimize
1312                         dst[x + y*stride]=  i;
1313                     }else{
1314                         i= -i;
1315                         i<<= QEXPSHIFT;
1316                         i/= qmul; //FIXME optimize
1317                         dst[x + y*stride]= -i;
1318                     }
1319                 }else
1320                     dst[x + y*stride]= 0;
1321             }
1322         }
1323     }else{
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= (i + bias) / qmul; //FIXME optimize
1332                         dst[x + y*stride]=  i;
1333                     }else{
1334                         i= -i;
1335                         i<<= QEXPSHIFT;
1336                         i= (i + bias) / qmul; //FIXME optimize
1337                         dst[x + y*stride]= -i;
1338                     }
1339                 }else
1340                     dst[x + y*stride]= 0;
1341             }
1342         }
1343     }
1344 }
1345
1346 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1347     const int w= b->width;
1348     const int h= b->height;
1349     const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1350     const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1351     const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1352     int x,y;
1353
1354     if(s->qlog == LOSSLESS_QLOG) return;
1355
1356     for(y=0; y<h; y++){
1357         for(x=0; x<w; x++){
1358             int i= src[x + y*stride];
1359             if(i<0){
1360                 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1361             }else if(i>0){
1362                 src[x + y*stride]=  (( i*qmul + qadd)>>(QEXPSHIFT));
1363             }
1364         }
1365     }
1366 }
1367
1368 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1369     const int w= b->width;
1370     const int h= b->height;
1371     int x,y;
1372
1373     for(y=h-1; y>=0; y--){
1374         for(x=w-1; x>=0; x--){
1375             int i= x + y*stride;
1376
1377             if(x){
1378                 if(use_median){
1379                     if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1380                     else  src[i] -= src[i - 1];
1381                 }else{
1382                     if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1383                     else  src[i] -= src[i - 1];
1384                 }
1385             }else{
1386                 if(y) src[i] -= src[i - stride];
1387             }
1388         }
1389     }
1390 }
1391
1392 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1393     const int w= b->width;
1394     const int h= b->height;
1395     int x,y;
1396
1397     for(y=0; y<h; y++){
1398         for(x=0; x<w; x++){
1399             int i= x + y*stride;
1400
1401             if(x){
1402                 if(use_median){
1403                     if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1404                     else  src[i] += src[i - 1];
1405                 }else{
1406                     if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1407                     else  src[i] += src[i - 1];
1408                 }
1409             }else{
1410                 if(y) src[i] += src[i - stride];
1411             }
1412         }
1413     }
1414 }
1415
1416 static void encode_qlogs(SnowContext *s){
1417     int plane_index, level, orientation;
1418
1419     for(plane_index=0; plane_index<2; plane_index++){
1420         for(level=0; level<s->spatial_decomposition_count; level++){
1421             for(orientation=level ? 1:0; orientation<4; orientation++){
1422                 if(orientation==2) continue;
1423                 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1424             }
1425         }
1426     }
1427 }
1428
1429 static void encode_header(SnowContext *s){
1430     int plane_index, i;
1431     uint8_t kstate[32];
1432
1433     memset(kstate, MID_STATE, sizeof(kstate));
1434
1435     put_rac(&s->c, kstate, s->keyframe);
1436     if(s->keyframe || s->always_reset){
1437         ff_snow_reset_contexts(s);
1438         s->last_spatial_decomposition_type=
1439         s->last_qlog=
1440         s->last_qbias=
1441         s->last_mv_scale=
1442         s->last_block_max_depth= 0;
1443         for(plane_index=0; plane_index<2; plane_index++){
1444             Plane *p= &s->plane[plane_index];
1445             p->last_htaps=0;
1446             p->last_diag_mc=0;
1447             memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1448         }
1449     }
1450     if(s->keyframe){
1451         put_symbol(&s->c, s->header_state, s->version, 0);
1452         put_rac(&s->c, s->header_state, s->always_reset);
1453         put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1454         put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1455         put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1456         put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1457         put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1458         put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1459         put_rac(&s->c, s->header_state, s->spatial_scalability);
1460 //        put_rac(&s->c, s->header_state, s->rate_scalability);
1461         put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1462
1463         encode_qlogs(s);
1464     }
1465
1466     if(!s->keyframe){
1467         int update_mc=0;
1468         for(plane_index=0; plane_index<2; plane_index++){
1469             Plane *p= &s->plane[plane_index];
1470             update_mc |= p->last_htaps   != p->htaps;
1471             update_mc |= p->last_diag_mc != p->diag_mc;
1472             update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1473         }
1474         put_rac(&s->c, s->header_state, update_mc);
1475         if(update_mc){
1476             for(plane_index=0; plane_index<2; plane_index++){
1477                 Plane *p= &s->plane[plane_index];
1478                 put_rac(&s->c, s->header_state, p->diag_mc);
1479                 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1480                 for(i= p->htaps/2; i; i--)
1481                     put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1482             }
1483         }
1484         if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1485             put_rac(&s->c, s->header_state, 1);
1486             put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1487             encode_qlogs(s);
1488         }else
1489             put_rac(&s->c, s->header_state, 0);
1490     }
1491
1492     put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1493     put_symbol(&s->c, s->header_state, s->qlog            - s->last_qlog    , 1);
1494     put_symbol(&s->c, s->header_state, s->mv_scale        - s->last_mv_scale, 1);
1495     put_symbol(&s->c, s->header_state, s->qbias           - s->last_qbias   , 1);
1496     put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1497
1498 }
1499
1500 static void update_last_header_values(SnowContext *s){
1501     int plane_index;
1502
1503     if(!s->keyframe){
1504         for(plane_index=0; plane_index<2; plane_index++){
1505             Plane *p= &s->plane[plane_index];
1506             p->last_diag_mc= p->diag_mc;
1507             p->last_htaps  = p->htaps;
1508             memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1509         }
1510     }
1511
1512     s->last_spatial_decomposition_type  = s->spatial_decomposition_type;
1513     s->last_qlog                        = s->qlog;
1514     s->last_qbias                       = s->qbias;
1515     s->last_mv_scale                    = s->mv_scale;
1516     s->last_block_max_depth             = s->block_max_depth;
1517     s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1518 }
1519
1520 static int qscale2qlog(int qscale){
1521     return rint(QROOT*log(qscale / (float)FF_QP2LAMBDA)/log(2))
1522            + 61*QROOT/8; ///< 64 > 60
1523 }
1524
1525 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1526 {
1527     /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1528      * FIXME we know exact mv bits at this point,
1529      * but ratecontrol isn't set up to include them. */
1530     uint32_t coef_sum= 0;
1531     int level, orientation, delta_qlog;
1532
1533     for(level=0; level<s->spatial_decomposition_count; level++){
1534         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1535             SubBand *b= &s->plane[0].band[level][orientation];
1536             IDWTELEM *buf= b->ibuf;
1537             const int w= b->width;
1538             const int h= b->height;
1539             const int stride= b->stride;
1540             const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1541             const int qmul= qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1542             const int qdiv= (1<<16)/qmul;
1543             int x, y;
1544             //FIXME this is ugly
1545             for(y=0; y<h; y++)
1546                 for(x=0; x<w; x++)
1547                     buf[x+y*stride]= b->buf[x+y*stride];
1548             if(orientation==0)
1549                 decorrelate(s, b, buf, stride, 1, 0);
1550             for(y=0; y<h; y++)
1551                 for(x=0; x<w; x++)
1552                     coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1553         }
1554     }
1555
1556     /* ugly, ratecontrol just takes a sqrt again */
1557     coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1558     assert(coef_sum < INT_MAX);
1559
1560     if(pict->pict_type == AV_PICTURE_TYPE_I){
1561         s->m.current_picture.mb_var_sum= coef_sum;
1562         s->m.current_picture.mc_mb_var_sum= 0;
1563     }else{
1564         s->m.current_picture.mc_mb_var_sum= coef_sum;
1565         s->m.current_picture.mb_var_sum= 0;
1566     }
1567
1568     pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1569     if (pict->quality < 0)
1570         return INT_MIN;
1571     s->lambda= pict->quality * 3/2;
1572     delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1573     s->qlog+= delta_qlog;
1574     return delta_qlog;
1575 }
1576
1577 static void calculate_visual_weight(SnowContext *s, Plane *p){
1578     int width = p->width;
1579     int height= p->height;
1580     int level, orientation, x, y;
1581
1582     for(level=0; level<s->spatial_decomposition_count; level++){
1583         for(orientation=level ? 1 : 0; orientation<4; orientation++){
1584             SubBand *b= &p->band[level][orientation];
1585             IDWTELEM *ibuf= b->ibuf;
1586             int64_t error=0;
1587
1588             memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1589             ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1590             ff_spatial_idwt(s->spatial_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1591             for(y=0; y<height; y++){
1592                 for(x=0; x<width; x++){
1593                     int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1594                     error += d*d;
1595                 }
1596             }
1597
1598             b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1599         }
1600     }
1601 }
1602
1603 static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data){
1604     SnowContext *s = avctx->priv_data;
1605     RangeCoder * const c= &s->c;
1606     AVFrame *pict = data;
1607     const int width= s->avctx->width;
1608     const int height= s->avctx->height;
1609     int level, orientation, plane_index, i, y;
1610     uint8_t rc_header_bak[sizeof(s->header_state)];
1611     uint8_t rc_block_bak[sizeof(s->block_state)];
1612
1613     ff_init_range_encoder(c, buf, buf_size);
1614     ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1615
1616     for(i=0; i<3; i++){
1617         int shift= !!i;
1618         for(y=0; y<(height>>shift); y++)
1619             memcpy(&s->input_picture.data[i][y * s->input_picture.linesize[i]],
1620                    &pict->data[i][y * pict->linesize[i]],
1621                    width>>shift);
1622     }
1623     s->new_picture = *pict;
1624
1625     s->m.picture_number= avctx->frame_number;
1626     if(avctx->flags&CODEC_FLAG_PASS2){
1627         s->m.pict_type =
1628         pict->pict_type= s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1629         s->keyframe= pict->pict_type==AV_PICTURE_TYPE_I;
1630         if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1631             pict->quality= ff_rate_estimate_qscale(&s->m, 0);
1632             if (pict->quality < 0)
1633                 return -1;
1634         }
1635     }else{
1636         s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1637         s->m.pict_type=
1638         pict->pict_type= s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1639     }
1640
1641     if(s->pass1_rc && avctx->frame_number == 0)
1642         pict->quality= 2*FF_QP2LAMBDA;
1643     if(pict->quality){
1644         s->qlog= qscale2qlog(pict->quality);
1645         s->lambda = pict->quality * 3/2;
1646     }
1647     if(s->qlog < 0 || (!pict->quality && (avctx->flags & CODEC_FLAG_QSCALE))){
1648         s->qlog= LOSSLESS_QLOG;
1649         s->lambda = 0;
1650     }//else keep previous frame's qlog until after motion estimation
1651
1652     ff_snow_frame_start(s);
1653
1654     s->m.current_picture_ptr= &s->m.current_picture;
1655     s->m.last_picture.f.pts = s->m.current_picture.f.pts;
1656     s->m.current_picture.f.pts = pict->pts;
1657     if(pict->pict_type == AV_PICTURE_TYPE_P){
1658         int block_width = (width +15)>>4;
1659         int block_height= (height+15)>>4;
1660         int stride= s->current_picture.linesize[0];
1661
1662         assert(s->current_picture.data[0]);
1663         assert(s->last_picture[0].data[0]);
1664
1665         s->m.avctx= s->avctx;
1666         s->m.current_picture.f.data[0] = s->current_picture.data[0];
1667         s->m.   last_picture.f.data[0] = s->last_picture[0].data[0];
1668         s->m.    new_picture.f.data[0] = s->  input_picture.data[0];
1669         s->m.   last_picture_ptr= &s->m.   last_picture;
1670         s->m.linesize=
1671         s->m.   last_picture.f.linesize[0] =
1672         s->m.    new_picture.f.linesize[0] =
1673         s->m.current_picture.f.linesize[0] = stride;
1674         s->m.uvlinesize= s->current_picture.linesize[1];
1675         s->m.width = width;
1676         s->m.height= height;
1677         s->m.mb_width = block_width;
1678         s->m.mb_height= block_height;
1679         s->m.mb_stride=   s->m.mb_width+1;
1680         s->m.b8_stride= 2*s->m.mb_width+1;
1681         s->m.f_code=1;
1682         s->m.pict_type= pict->pict_type;
1683         s->m.me_method= s->avctx->me_method;
1684         s->m.me.scene_change_score=0;
1685         s->m.flags= s->avctx->flags;
1686         s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1687         s->m.out_format= FMT_H263;
1688         s->m.unrestricted_mv= 1;
1689
1690         s->m.lambda = s->lambda;
1691         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1692         s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1693
1694         s->m.dsp= s->dsp; //move
1695         ff_init_me(&s->m);
1696         s->dsp= s->m.dsp;
1697     }
1698
1699     if(s->pass1_rc){
1700         memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1701         memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1702     }
1703
1704 redo_frame:
1705
1706     if(pict->pict_type == AV_PICTURE_TYPE_I)
1707         s->spatial_decomposition_count= 5;
1708     else
1709         s->spatial_decomposition_count= 5;
1710
1711     s->m.pict_type = pict->pict_type;
1712     s->qbias= pict->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1713
1714     ff_snow_common_init_after_header(avctx);
1715
1716     if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1717         for(plane_index=0; plane_index<3; plane_index++){
1718             calculate_visual_weight(s, &s->plane[plane_index]);
1719         }
1720     }
1721
1722     encode_header(s);
1723     s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1724     encode_blocks(s, 1);
1725     s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1726
1727     for(plane_index=0; plane_index<3; plane_index++){
1728         Plane *p= &s->plane[plane_index];
1729         int w= p->width;
1730         int h= p->height;
1731         int x, y;
1732 //        int bits= put_bits_count(&s->c.pb);
1733
1734         if (!s->memc_only) {
1735             //FIXME optimize
1736             if(pict->data[plane_index]) //FIXME gray hack
1737                 for(y=0; y<h; y++){
1738                     for(x=0; x<w; x++){
1739                         s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1740                     }
1741                 }
1742             predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1743
1744             if(   plane_index==0
1745                && pict->pict_type == AV_PICTURE_TYPE_P
1746                && !(avctx->flags&CODEC_FLAG_PASS2)
1747                && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1748                 ff_init_range_encoder(c, buf, buf_size);
1749                 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1750                 pict->pict_type= AV_PICTURE_TYPE_I;
1751                 s->keyframe=1;
1752                 s->current_picture.key_frame=1;
1753                 goto redo_frame;
1754             }
1755
1756             if(s->qlog == LOSSLESS_QLOG){
1757                 for(y=0; y<h; y++){
1758                     for(x=0; x<w; x++){
1759                         s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1760                     }
1761                 }
1762             }else{
1763                 for(y=0; y<h; y++){
1764                     for(x=0; x<w; x++){
1765                         s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1766                     }
1767                 }
1768             }
1769
1770             /*  if(QUANTIZE2)
1771                 dwt_quantize(s, p, s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type);
1772             else*/
1773                 ff_spatial_dwt(s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1774
1775             if(s->pass1_rc && plane_index==0){
1776                 int delta_qlog = ratecontrol_1pass(s, pict);
1777                 if (delta_qlog <= INT_MIN)
1778                     return -1;
1779                 if(delta_qlog){
1780                     //reordering qlog in the bitstream would eliminate this reset
1781                     ff_init_range_encoder(c, buf, buf_size);
1782                     memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1783                     memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1784                     encode_header(s);
1785                     encode_blocks(s, 0);
1786                 }
1787             }
1788
1789             for(level=0; level<s->spatial_decomposition_count; level++){
1790                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1791                     SubBand *b= &p->band[level][orientation];
1792
1793                     if(!QUANTIZE2)
1794                         quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1795                     if(orientation==0)
1796                         decorrelate(s, b, b->ibuf, b->stride, pict->pict_type == AV_PICTURE_TYPE_P, 0);
1797                     encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1798                     assert(b->parent==NULL || b->parent->stride == b->stride*2);
1799                     if(orientation==0)
1800                         correlate(s, b, b->ibuf, b->stride, 1, 0);
1801                 }
1802             }
1803
1804             for(level=0; level<s->spatial_decomposition_count; level++){
1805                 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1806                     SubBand *b= &p->band[level][orientation];
1807
1808                     dequantize(s, b, b->ibuf, b->stride);
1809                 }
1810             }
1811
1812             ff_spatial_idwt(s->spatial_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1813             if(s->qlog == LOSSLESS_QLOG){
1814                 for(y=0; y<h; y++){
1815                     for(x=0; x<w; x++){
1816                         s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1817                     }
1818                 }
1819             }
1820             predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1821         }else{
1822             //ME/MC only
1823             if(pict->pict_type == AV_PICTURE_TYPE_I){
1824                 for(y=0; y<h; y++){
1825                     for(x=0; x<w; x++){
1826                         s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]=
1827                             pict->data[plane_index][y*pict->linesize[plane_index] + x];
1828                     }
1829                 }
1830             }else{
1831                 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1832                 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1833             }
1834         }
1835         if(s->avctx->flags&CODEC_FLAG_PSNR){
1836             int64_t error= 0;
1837
1838             if(pict->data[plane_index]) //FIXME gray hack
1839                 for(y=0; y<h; y++){
1840                     for(x=0; x<w; x++){
1841                         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];
1842                         error += d*d;
1843                     }
1844                 }
1845             s->avctx->error[plane_index] += error;
1846             s->current_picture.error[plane_index] = error;
1847         }
1848
1849     }
1850
1851     update_last_header_values(s);
1852
1853     ff_snow_release_buffer(avctx);
1854
1855     s->current_picture.coded_picture_number = avctx->frame_number;
1856     s->current_picture.pict_type = pict->pict_type;
1857     s->current_picture.quality = pict->quality;
1858     s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1859     s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1860     s->m.current_picture.f.display_picture_number =
1861     s->m.current_picture.f.coded_picture_number   = avctx->frame_number;
1862     s->m.current_picture.f.quality                = pict->quality;
1863     s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1864     if(s->pass1_rc)
1865         if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1866             return -1;
1867     if(avctx->flags&CODEC_FLAG_PASS1)
1868         ff_write_pass1_stats(&s->m);
1869     s->m.last_pict_type = s->m.pict_type;
1870     avctx->frame_bits = s->m.frame_bits;
1871     avctx->mv_bits = s->m.mv_bits;
1872     avctx->misc_bits = s->m.misc_bits;
1873     avctx->p_tex_bits = s->m.p_tex_bits;
1874
1875     emms_c();
1876
1877     return ff_rac_terminate(c);
1878 }
1879
1880 static av_cold int encode_end(AVCodecContext *avctx)
1881 {
1882     SnowContext *s = avctx->priv_data;
1883
1884     ff_snow_common_end(s);
1885     if (s->input_picture.data[0])
1886         avctx->release_buffer(avctx, &s->input_picture);
1887     av_free(avctx->stats_out);
1888
1889     return 0;
1890 }
1891
1892 #define OFFSET(x) offsetof(SnowContext, x)
1893 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1894 static const AVOption options[] = {
1895     { "memc_only",      "Only do ME/MC (I frames -> ref, P frame -> ME+MC).",   OFFSET(memc_only), AV_OPT_TYPE_INT, { 0 }, 0, 1, VE },
1896     { NULL },
1897 };
1898
1899 static const AVClass snowenc_class = {
1900     .class_name = "snow encoder",
1901     .item_name  = av_default_item_name,
1902     .option     = options,
1903     .version    = LIBAVUTIL_VERSION_INT,
1904 };
1905
1906 AVCodec ff_snow_encoder = {
1907     .name           = "snow",
1908     .type           = AVMEDIA_TYPE_VIDEO,
1909     .id             = CODEC_ID_SNOW,
1910     .priv_data_size = sizeof(SnowContext),
1911     .init           = encode_init,
1912     .encode         = encode_frame,
1913     .close          = encode_end,
1914     .long_name = NULL_IF_CONFIG_SMALL("Snow"),
1915     .priv_class     = &snowenc_class,
1916 };
1917 #endif
1918
1919
1920 #ifdef TEST
1921 #undef malloc
1922 #undef free
1923 #undef printf
1924
1925 #include "libavutil/lfg.h"
1926 #include "libavutil/mathematics.h"
1927
1928 int main(void){
1929     int width=256;
1930     int height=256;
1931     int buffer[2][width*height];
1932     SnowContext s;
1933     int i;
1934     AVLFG prng;
1935     s.spatial_decomposition_count=6;
1936     s.spatial_decomposition_type=1;
1937
1938     av_lfg_init(&prng, 1);
1939
1940     printf("testing 5/3 DWT\n");
1941     for(i=0; i<width*height; i++)
1942         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1943
1944     ff_spatial_dwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1945     ff_spatial_idwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1946
1947     for(i=0; i<width*height; i++)
1948         if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1949
1950     printf("testing 9/7 DWT\n");
1951     s.spatial_decomposition_type=0;
1952     for(i=0; i<width*height; i++)
1953         buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1954
1955     ff_spatial_dwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1956     ff_spatial_idwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1957
1958     for(i=0; i<width*height; i++)
1959         if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1960
1961     {
1962     int level, orientation, x, y;
1963     int64_t errors[8][4];
1964     int64_t g=0;
1965
1966         memset(errors, 0, sizeof(errors));
1967         s.spatial_decomposition_count=3;
1968         s.spatial_decomposition_type=0;
1969         for(level=0; level<s.spatial_decomposition_count; level++){
1970             for(orientation=level ? 1 : 0; orientation<4; orientation++){
1971                 int w= width  >> (s.spatial_decomposition_count-level);
1972                 int h= height >> (s.spatial_decomposition_count-level);
1973                 int stride= width  << (s.spatial_decomposition_count-level);
1974                 DWTELEM *buf= buffer[0];
1975                 int64_t error=0;
1976
1977                 if(orientation&1) buf+=w;
1978                 if(orientation>1) buf+=stride>>1;
1979
1980                 memset(buffer[0], 0, sizeof(int)*width*height);
1981                 buf[w/2 + h/2*stride]= 256*256;
1982                 ff_spatial_idwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1983                 for(y=0; y<height; y++){
1984                     for(x=0; x<width; x++){
1985                         int64_t d= buffer[0][x + y*width];
1986                         error += d*d;
1987                         if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
1988                     }
1989                     if(FFABS(height/2-y)<9 && level==2) printf("\n");
1990                 }
1991                 error= (int)(sqrt(error)+0.5);
1992                 errors[level][orientation]= error;
1993                 if(g) g=av_gcd(g, error);
1994                 else g= error;
1995             }
1996         }
1997         printf("static int const visual_weight[][4]={\n");
1998         for(level=0; level<s.spatial_decomposition_count; level++){
1999             printf("  {");
2000             for(orientation=0; orientation<4; orientation++){
2001                 printf("%8"PRId64",", errors[level][orientation]/g);
2002             }
2003             printf("},\n");
2004         }
2005         printf("};\n");
2006         {
2007             int level=2;
2008             int w= width  >> (s.spatial_decomposition_count-level);
2009             //int h= height >> (s.spatial_decomposition_count-level);
2010             int stride= width  << (s.spatial_decomposition_count-level);
2011             DWTELEM *buf= buffer[0];
2012             int64_t error=0;
2013
2014             buf+=w;
2015             buf+=stride>>1;
2016
2017             memset(buffer[0], 0, sizeof(int)*width*height);
2018             for(y=0; y<height; y++){
2019                 for(x=0; x<width; x++){
2020                     int tab[4]={0,2,3,1};
2021                     buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2022                 }
2023             }
2024             ff_spatial_dwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2025             for(y=0; y<height; y++){
2026                 for(x=0; x<width; x++){
2027                     int64_t d= buffer[0][x + y*width];
2028                     error += d*d;
2029                     if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2030                 }
2031                 if(FFABS(height/2-y)<9) printf("\n");
2032             }
2033         }
2034
2035     }
2036     return 0;
2037 }
2038 #endif /* TEST */