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