2 * Copyright (C) 2004 Michael Niedermayer <michaelni@gmx.at>
4 * This file is part of FFmpeg.
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.
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.
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
21 #include "libavutil/intmath.h"
22 #include "libavutil/log.h"
23 #include "libavutil/opt.h"
30 #include "rangecoder.h"
33 #include "mpegvideo.h"
44 static void find_sse(SnowContext *s, Plane *p, int *score, int score_stride, IDWTELEM *r0, IDWTELEM *r1, int level, int orientation){
45 SubBand *b= &p->band[level][orientation];
49 int step= 1 << (s->spatial_decomposition_count - level);
56 //FIXME bias for nonzero ?
58 memset(score, 0, sizeof(*score)*score_stride*((p->height + Q2_STEP-1)/Q2_STEP));
59 for(y=0; y<p->height; y++){
60 for(x=0; x<p->width; x++){
61 int sx= (x-xo + step/2) / step / Q2_STEP;
62 int sy= (y-yo + step/2) / step / Q2_STEP;
63 int v= r0[x + y*p->width] - r1[x + y*p->width];
64 assert(sx>=0 && sy>=0 && sx < score_stride);
66 score[sx + sy*score_stride] += v*v;
67 assert(score[sx + sy*score_stride] >= 0);
72 static void dequantize_all(SnowContext *s, Plane *p, IDWTELEM *buffer, int width, int height){
73 int level, orientation;
75 for(level=0; level<s->spatial_decomposition_count; level++){
76 for(orientation=level ? 1 : 0; orientation<4; orientation++){
77 SubBand *b= &p->band[level][orientation];
78 IDWTELEM *dst= buffer + (b->ibuf - s->spatial_idwt_buffer);
80 dequantize(s, b, dst, b->stride);
85 static void dwt_quantize(SnowContext *s, Plane *p, DWTELEM *buffer, int width, int height, int stride, int type){
86 int level, orientation, ys, xs, x, y, pass;
87 IDWTELEM best_dequant[height * stride];
88 IDWTELEM idwt2_buffer[height * stride];
89 const int score_stride= (width + 10)/Q2_STEP;
90 int best_score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
91 int score[(width + 10)/Q2_STEP * (height + 10)/Q2_STEP]; //FIXME size
92 int threshold= (s->m.lambda * s->m.lambda) >> 6;
94 //FIXME pass the copy cleanly ?
96 // memcpy(dwt_buffer, buffer, height * stride * sizeof(DWTELEM));
97 ff_spatial_dwt(buffer, width, height, stride, type, s->spatial_decomposition_count);
99 for(level=0; level<s->spatial_decomposition_count; level++){
100 for(orientation=level ? 1 : 0; orientation<4; orientation++){
101 SubBand *b= &p->band[level][orientation];
102 IDWTELEM *dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
103 DWTELEM *src= buffer + (b-> buf - s->spatial_dwt_buffer);
104 assert(src == b->buf); // code does not depend on this but it is true currently
106 quantize(s, b, dst, src, b->stride, s->qbias);
109 for(pass=0; pass<1; pass++){
110 if(s->qbias == 0) //keyframe
112 for(level=0; level<s->spatial_decomposition_count; level++){
113 for(orientation=level ? 1 : 0; orientation<4; orientation++){
114 SubBand *b= &p->band[level][orientation];
115 IDWTELEM *dst= idwt2_buffer + (b->ibuf - s->spatial_idwt_buffer);
116 IDWTELEM *best_dst= best_dequant + (b->ibuf - s->spatial_idwt_buffer);
118 for(ys= 0; ys<Q2_STEP; ys++){
119 for(xs= 0; xs<Q2_STEP; xs++){
120 memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
121 dequantize_all(s, p, idwt2_buffer, width, height);
122 ff_spatial_idwt(idwt2_buffer, width, height, stride, type, s->spatial_decomposition_count);
123 find_sse(s, p, best_score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
124 memcpy(idwt2_buffer, best_dequant, height * stride * sizeof(IDWTELEM));
125 for(y=ys; y<b->height; y+= Q2_STEP){
126 for(x=xs; x<b->width; x+= Q2_STEP){
127 if(dst[x + y*b->stride]<0) dst[x + y*b->stride]++;
128 if(dst[x + y*b->stride]>0) dst[x + y*b->stride]--;
129 //FIXME try more than just --
132 dequantize_all(s, p, idwt2_buffer, width, height);
133 ff_spatial_idwt(idwt2_buffer, width, height, stride, type, s->spatial_decomposition_count);
134 find_sse(s, p, score, score_stride, idwt2_buffer, s->spatial_idwt_buffer, level, orientation);
135 for(y=ys; y<b->height; y+= Q2_STEP){
136 for(x=xs; x<b->width; x+= Q2_STEP){
137 int score_idx= x/Q2_STEP + (y/Q2_STEP)*score_stride;
138 if(score[score_idx] <= best_score[score_idx] + threshold){
139 best_score[score_idx]= score[score_idx];
140 if(best_dst[x + y*b->stride]<0) best_dst[x + y*b->stride]++;
141 if(best_dst[x + y*b->stride]>0) best_dst[x + y*b->stride]--;
151 memcpy(s->spatial_idwt_buffer, best_dequant, height * stride * sizeof(IDWTELEM)); //FIXME work with that directly instead of copy at the end
154 #endif /* QUANTIZE2==1 */
156 static av_cold int encode_init(AVCodecContext *avctx)
158 SnowContext *s = avctx->priv_data;
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");
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");
174 s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
176 s->mv_scale = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
177 s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
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;
188 ff_snow_common_init(avctx);
189 ff_snow_alloc_blocks(s);
194 s->m.flags = avctx->flags;
195 s->m.bit_rate= avctx->bit_rate;
198 s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
199 s->m.me.map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
200 s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
201 s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
202 ff_h263_encode_init(&s->m); //mv_penalty
204 s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
206 if(avctx->flags&CODEC_FLAG_PASS1){
207 if(!avctx->stats_out)
208 avctx->stats_out = av_mallocz(256);
210 if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
211 if(ff_rate_control_init(&s->m) < 0)
214 s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
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;
226 /* case PIX_FMT_RGB32:
230 av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
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;
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);
240 s->avctx->get_buffer(s->avctx, &s->input_picture);
242 if(s->avctx->me_method == ME_ITER){
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));
251 s->runs = av_malloc(avctx->width * avctx->height * sizeof(*s->runs));
256 //near copy & paste from dsputil, FIXME
257 static int pix_sum(uint8_t * pix, int line_size, int w)
262 for (i = 0; i < w; i++) {
263 for (j = 0; j < w; j++) {
267 pix += line_size - w;
272 //near copy & paste from dsputil, FIXME
273 static int pix_norm1(uint8_t * pix, int line_size, int w)
276 uint32_t *sq = ff_squareTbl + 256;
279 for (i = 0; i < w; i++) {
280 for (j = 0; j < w; j ++) {
284 pix += line_size - w;
292 #define P_TOPRIGHT P[3]
293 #define P_MEDIAN P[4]
295 #define FLAG_QPEL 1 //must be 1
297 static int encode_q_branch(SnowContext *s, int level, int x, int y){
298 uint8_t p_buffer[1024];
299 uint8_t i_buffer[1024];
300 uint8_t p_state[sizeof(s->block_state)];
301 uint8_t i_state[sizeof(s->block_state)];
303 uint8_t *pbbak= s->c.bytestream;
304 uint8_t *pbbak_start= s->c.bytestream_start;
305 int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
306 const int w= s->b_width << s->block_max_depth;
307 const int h= s->b_height << s->block_max_depth;
308 const int rem_depth= s->block_max_depth - level;
309 const int index= (x + y*w) << rem_depth;
310 const int block_w= 1<<(LOG2_MB_SIZE - level);
311 int trx= (x+1)<<rem_depth;
312 int try= (y+1)<<rem_depth;
313 const BlockNode *left = x ? &s->block[index-1] : &null_block;
314 const BlockNode *top = y ? &s->block[index-w] : &null_block;
315 const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
316 const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
317 const BlockNode *tl = y && x ? &s->block[index-w-1] : left;
318 const BlockNode *tr = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
319 int pl = left->color[0];
320 int pcb= left->color[1];
321 int pcr= left->color[2];
325 const int stride= s->current_picture.linesize[0];
326 const int uvstride= s->current_picture.linesize[1];
327 uint8_t *current_data[3]= { s->input_picture.data[0] + (x + y* stride)*block_w,
328 s->input_picture.data[1] + (x + y*uvstride)*block_w/2,
329 s->input_picture.data[2] + (x + y*uvstride)*block_w/2};
331 int16_t last_mv[3][2];
332 int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
333 const int shift= 1+qpel;
334 MotionEstContext *c= &s->m.me;
335 int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
336 int mx_context= av_log2(2*FFABS(left->mx - top->mx));
337 int my_context= av_log2(2*FFABS(left->my - top->my));
338 int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
339 int ref, best_ref, ref_score, ref_mx, ref_my;
341 assert(sizeof(s->block_state) >= 256);
343 set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
347 // clip predictors / edge ?
353 P_TOPRIGHT[0]= tr->mx;
354 P_TOPRIGHT[1]= tr->my;
356 last_mv[0][0]= s->block[index].mx;
357 last_mv[0][1]= s->block[index].my;
358 last_mv[1][0]= right->mx;
359 last_mv[1][1]= right->my;
360 last_mv[2][0]= bottom->mx;
361 last_mv[2][1]= bottom->my;
368 assert(c-> stride == stride);
369 assert(c->uvstride == uvstride);
371 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
372 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
373 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
374 c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
376 c->xmin = - x*block_w - 16+3;
377 c->ymin = - y*block_w - 16+3;
378 c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
379 c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
381 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
382 if(P_LEFT[1] > (c->ymax<<shift)) P_LEFT[1] = (c->ymax<<shift);
383 if(P_TOP[0] > (c->xmax<<shift)) P_TOP[0] = (c->xmax<<shift);
384 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
385 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
386 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
387 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
389 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
390 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
393 c->pred_x= P_LEFT[0];
394 c->pred_y= P_LEFT[1];
396 c->pred_x = P_MEDIAN[0];
397 c->pred_y = P_MEDIAN[1];
402 for(ref=0; ref<s->ref_frames; ref++){
403 init_ref(c, current_data, s->last_picture[ref].data, NULL, block_w*x, block_w*y, 0);
405 ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
406 (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
408 assert(ref_mx >= c->xmin);
409 assert(ref_mx <= c->xmax);
410 assert(ref_my >= c->ymin);
411 assert(ref_my <= c->ymax);
413 ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
414 ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
415 ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
417 s->ref_mvs[ref][index][0]= ref_mx;
418 s->ref_mvs[ref][index][1]= ref_my;
419 s->ref_scores[ref][index]= ref_score;
421 if(score > ref_score){
428 //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
431 base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
434 pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
435 memcpy(p_state, s->block_state, sizeof(s->block_state));
437 if(level!=s->block_max_depth)
438 put_rac(&pc, &p_state[4 + s_context], 1);
439 put_rac(&pc, &p_state[1 + left->type + top->type], 0);
440 if(s->ref_frames > 1)
441 put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
442 pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
443 put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
444 put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
445 p_len= pc.bytestream - pc.bytestream_start;
446 score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
448 block_s= block_w*block_w;
449 sum = pix_sum(current_data[0], stride, block_w);
450 l= (sum + block_s/2)/block_s;
451 iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
453 block_s= block_w*block_w>>2;
454 sum = pix_sum(current_data[1], uvstride, block_w>>1);
455 cb= (sum + block_s/2)/block_s;
456 // iscore += pix_norm1(¤t_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
457 sum = pix_sum(current_data[2], uvstride, block_w>>1);
458 cr= (sum + block_s/2)/block_s;
459 // iscore += pix_norm1(¤t_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
463 ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
464 memcpy(i_state, s->block_state, sizeof(s->block_state));
465 if(level!=s->block_max_depth)
466 put_rac(&ic, &i_state[4 + s_context], 1);
467 put_rac(&ic, &i_state[1 + left->type + top->type], 1);
468 put_symbol(&ic, &i_state[32], l-pl , 1);
469 put_symbol(&ic, &i_state[64], cb-pcb, 1);
470 put_symbol(&ic, &i_state[96], cr-pcr, 1);
471 i_len= ic.bytestream - ic.bytestream_start;
472 iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
474 // assert(score==256*256*256*64-1);
475 assert(iscore < 255*255*256 + s->lambda2*10);
477 assert(l>=0 && l<=255);
478 assert(pl>=0 && pl<=255);
481 int varc= iscore >> 8;
482 int vard= score >> 8;
483 if (vard <= 64 || vard < varc)
484 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
486 c->scene_change_score+= s->m.qscale;
489 if(level!=s->block_max_depth){
490 put_rac(&s->c, &s->block_state[4 + s_context], 0);
491 score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
492 score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
493 score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
494 score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
495 score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
497 if(score2 < score && score2 < iscore)
502 pred_mv(s, &pmx, &pmy, 0, left, top, tr);
503 memcpy(pbbak, i_buffer, i_len);
505 s->c.bytestream_start= pbbak_start;
506 s->c.bytestream= pbbak + i_len;
507 set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
508 memcpy(s->block_state, i_state, sizeof(s->block_state));
511 memcpy(pbbak, p_buffer, p_len);
513 s->c.bytestream_start= pbbak_start;
514 s->c.bytestream= pbbak + p_len;
515 set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
516 memcpy(s->block_state, p_state, sizeof(s->block_state));
521 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
522 const int w= s->b_width << s->block_max_depth;
523 const int rem_depth= s->block_max_depth - level;
524 const int index= (x + y*w) << rem_depth;
525 int trx= (x+1)<<rem_depth;
526 BlockNode *b= &s->block[index];
527 const BlockNode *left = x ? &s->block[index-1] : &null_block;
528 const BlockNode *top = y ? &s->block[index-w] : &null_block;
529 const BlockNode *tl = y && x ? &s->block[index-w-1] : left;
530 const BlockNode *tr = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
531 int pl = left->color[0];
532 int pcb= left->color[1];
533 int pcr= left->color[2];
535 int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
536 int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
537 int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
538 int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
541 set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
545 if(level!=s->block_max_depth){
546 if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
547 put_rac(&s->c, &s->block_state[4 + s_context], 1);
549 put_rac(&s->c, &s->block_state[4 + s_context], 0);
550 encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
551 encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
552 encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
553 encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
557 if(b->type & BLOCK_INTRA){
558 pred_mv(s, &pmx, &pmy, 0, left, top, tr);
559 put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
560 put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
561 put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
562 put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
563 set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
565 pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
566 put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
567 if(s->ref_frames > 1)
568 put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
569 put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
570 put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
571 set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
575 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
577 Plane *p= &s->plane[plane_index];
578 const int block_size = MB_SIZE >> s->block_max_depth;
579 const int block_w = plane_index ? block_size/2 : block_size;
580 const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+1] : ff_obmc_tab[s->block_max_depth];
581 const int obmc_stride= plane_index ? block_size : 2*block_size;
582 const int ref_stride= s->current_picture.linesize[plane_index];
583 uint8_t *src= s-> input_picture.data[plane_index];
584 IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
585 const int b_stride = s->b_width << s->block_max_depth;
586 const int w= p->width;
587 const int h= p->height;
588 int index= mb_x + mb_y*b_stride;
589 BlockNode *b= &s->block[index];
590 BlockNode backup= *b;
594 b->type|= BLOCK_INTRA;
595 b->color[plane_index]= 0;
596 memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
599 int mb_x2= mb_x + (i &1) - 1;
600 int mb_y2= mb_y + (i>>1) - 1;
601 int x= block_w*mb_x2 + block_w/2;
602 int y= block_w*mb_y2 + block_w/2;
604 add_yblock(s, 0, NULL, dst + ((i&1)+(i>>1)*obmc_stride)*block_w, NULL, obmc,
605 x, y, block_w, block_w, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
607 for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_w); y2++){
608 for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
609 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_w*mb_y - block_w/2))*obmc_stride;
610 int obmc_v= obmc[index];
612 if(y<0) obmc_v += obmc[index + block_w*obmc_stride];
613 if(x<0) obmc_v += obmc[index + block_w];
614 if(y+block_w>h) obmc_v += obmc[index - block_w*obmc_stride];
615 if(x+block_w>w) obmc_v += obmc[index - block_w];
616 //FIXME precalculate this or simplify it somehow else
618 d = -dst[index] + (1<<(FRAC_BITS-1));
620 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
621 aa += obmc_v * obmc_v; //FIXME precalculate this
627 return av_clip(((ab<<LOG2_OBMC_MAX) + aa/2)/aa, 0, 255); //FIXME we should not need clipping
630 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
631 const int b_stride = s->b_width << s->block_max_depth;
632 const int b_height = s->b_height<< s->block_max_depth;
633 int index= x + y*b_stride;
634 const BlockNode *b = &s->block[index];
635 const BlockNode *left = x ? &s->block[index-1] : &null_block;
636 const BlockNode *top = y ? &s->block[index-b_stride] : &null_block;
637 const BlockNode *tl = y && x ? &s->block[index-b_stride-1] : left;
638 const BlockNode *tr = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
640 // int mx_context= av_log2(2*FFABS(left->mx - top->mx));
641 // int my_context= av_log2(2*FFABS(left->my - top->my));
643 if(x<0 || x>=b_stride || y>=b_height)
652 //FIXME try accurate rate
653 //FIXME intra and inter predictors if surrounding blocks are not the same type
654 if(b->type & BLOCK_INTRA){
655 return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
656 + av_log2(2*FFABS(left->color[1] - b->color[1]))
657 + av_log2(2*FFABS(left->color[2] - b->color[2])));
659 pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
662 return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
663 + av_log2(2*FFABS(dmy))
664 + av_log2(2*b->ref));
668 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, const uint8_t *obmc_edged){
669 Plane *p= &s->plane[plane_index];
670 const int block_size = MB_SIZE >> s->block_max_depth;
671 const int block_w = plane_index ? block_size/2 : block_size;
672 const int obmc_stride= plane_index ? block_size : 2*block_size;
673 const int ref_stride= s->current_picture.linesize[plane_index];
674 uint8_t *dst= s->current_picture.data[plane_index];
675 uint8_t *src= s-> input_picture.data[plane_index];
676 IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
677 uint8_t *cur = s->scratchbuf;
678 uint8_t tmp[ref_stride*(2*MB_SIZE+HTAPS_MAX-1)];
679 const int b_stride = s->b_width << s->block_max_depth;
680 const int b_height = s->b_height<< s->block_max_depth;
681 const int w= p->width;
682 const int h= p->height;
685 const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
686 int sx= block_w*mb_x - block_w/2;
687 int sy= block_w*mb_y - block_w/2;
688 int x0= FFMAX(0,-sx);
689 int y0= FFMAX(0,-sy);
690 int x1= FFMIN(block_w*2, w-sx);
691 int y1= FFMIN(block_w*2, h-sy);
694 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);
696 for(y=y0; y<y1; y++){
697 const uint8_t *obmc1= obmc_edged + y*obmc_stride;
698 const IDWTELEM *pred1 = pred + y*obmc_stride;
699 uint8_t *cur1 = cur + y*ref_stride;
700 uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
701 for(x=x0; x<x1; x++){
702 #if FRAC_BITS >= LOG2_OBMC_MAX
703 int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
705 int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
707 v = (v + pred1[x]) >> FRAC_BITS;
708 if(v&(~255)) v= ~(v>>31);
713 /* copy the regions where obmc[] = (uint8_t)256 */
714 if(LOG2_OBMC_MAX == 8
715 && (mb_x == 0 || mb_x == b_stride-1)
716 && (mb_y == 0 || mb_y == b_height-1)){
726 memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
730 /* FIXME rearrange dsputil to fit 32x32 cmp functions */
731 /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
732 /* FIXME cmps overlap but do not cover the wavelet's whole support.
733 * So improving the score of one block is not strictly guaranteed
734 * to improve the score of the whole frame, thus iterative motion
735 * estimation does not always converge. */
736 if(s->avctx->me_cmp == FF_CMP_W97)
737 distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
738 else if(s->avctx->me_cmp == FF_CMP_W53)
739 distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
743 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
744 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
749 distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
758 rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
760 if(mb_x == b_stride-2)
761 rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
763 return distortion + rate*penalty_factor;
766 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
768 Plane *p= &s->plane[plane_index];
769 const int block_size = MB_SIZE >> s->block_max_depth;
770 const int block_w = plane_index ? block_size/2 : block_size;
771 const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+1] : ff_obmc_tab[s->block_max_depth];
772 const int obmc_stride= plane_index ? block_size : 2*block_size;
773 const int ref_stride= s->current_picture.linesize[plane_index];
774 uint8_t *dst= s->current_picture.data[plane_index];
775 uint8_t *src= s-> input_picture.data[plane_index];
776 //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
777 // const has only been removed from zero_dst to suppress a warning
778 static IDWTELEM zero_dst[4096]; //FIXME
779 const int b_stride = s->b_width << s->block_max_depth;
780 const int w= p->width;
781 const int h= p->height;
784 const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
787 int mb_x2= mb_x + (i%3) - 1;
788 int mb_y2= mb_y + (i/3) - 1;
789 int x= block_w*mb_x2 + block_w/2;
790 int y= block_w*mb_y2 + block_w/2;
792 add_yblock(s, 0, NULL, zero_dst, dst, obmc,
793 x, y, block_w, block_w, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
795 //FIXME find a cleaner/simpler way to skip the outside stuff
796 for(y2= y; y2<0; y2++)
797 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
798 for(y2= h; y2<y+block_w; y2++)
799 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
801 for(y2= y; y2<y+block_w; y2++)
802 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
805 for(y2= y; y2<y+block_w; y2++)
806 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
809 assert(block_w== 8 || block_w==16);
810 distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_w);
814 BlockNode *b= &s->block[mb_x+mb_y*b_stride];
815 int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
823 rate = get_block_bits(s, mb_x, mb_y, 2);
824 for(i=merged?4:0; i<9; i++){
825 static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
826 rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
829 return distortion + rate*penalty_factor;
832 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
833 const int w= b->width;
834 const int h= b->height;
845 int /*ll=0, */l=0, lt=0, t=0, rt=0;
846 v= src[x + y*stride];
849 t= src[x + (y-1)*stride];
851 lt= src[x - 1 + (y-1)*stride];
854 rt= src[x + 1 + (y-1)*stride];
858 l= src[x - 1 + y*stride];
860 if(orientation==1) ll= src[y + (x-2)*stride];
861 else ll= src[x - 2 + y*stride];
867 if(px<b->parent->width && py<b->parent->height)
868 p= parent[px + py*2*stride];
870 if(!(/*ll|*/l|lt|t|rt|p)){
872 s->runs[run_index++]= run;
880 max_index= run_index;
881 s->runs[run_index++]= run;
883 run= s->runs[run_index++];
885 put_symbol2(&s->c, b->state[30], max_index, 0);
886 if(run_index <= max_index)
887 put_symbol2(&s->c, b->state[1], run, 3);
890 if(s->c.bytestream_end - s->c.bytestream < w*40){
891 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
896 int /*ll=0, */l=0, lt=0, t=0, rt=0;
897 v= src[x + y*stride];
900 t= src[x + (y-1)*stride];
902 lt= src[x - 1 + (y-1)*stride];
905 rt= src[x + 1 + (y-1)*stride];
909 l= src[x - 1 + y*stride];
911 if(orientation==1) ll= src[y + (x-2)*stride];
912 else ll= src[x - 2 + y*stride];
918 if(px<b->parent->width && py<b->parent->height)
919 p= parent[px + py*2*stride];
921 if(/*ll|*/l|lt|t|rt|p){
922 int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
924 put_rac(&s->c, &b->state[0][context], !!v);
927 run= s->runs[run_index++];
929 if(run_index <= max_index)
930 put_symbol2(&s->c, b->state[1], run, 3);
938 int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
939 int l2= 2*FFABS(l) + (l<0);
940 int t2= 2*FFABS(t) + (t<0);
942 put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
943 put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
951 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
952 // encode_subband_qtree(s, b, src, parent, stride, orientation);
953 // encode_subband_z0run(s, b, src, parent, stride, orientation);
954 return encode_subband_c0run(s, b, src, parent, stride, orientation);
955 // encode_subband_dzr(s, b, src, parent, stride, orientation);
958 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){
959 const int b_stride= s->b_width << s->block_max_depth;
960 BlockNode *block= &s->block[mb_x + mb_y * b_stride];
961 BlockNode backup= *block;
965 assert(mb_x>=0 && mb_y>=0);
966 assert(mb_x<b_stride);
969 block->color[0] = p[0];
970 block->color[1] = p[1];
971 block->color[2] = p[2];
972 block->type |= BLOCK_INTRA;
974 index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
975 value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
976 if(s->me_cache[index] == value)
978 s->me_cache[index]= value;
982 block->type &= ~BLOCK_INTRA;
985 rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
997 /* special case for int[2] args we discard afterwards,
998 * fixes compilation problem with gcc 2.95 */
999 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){
1000 int p[2] = {p0, p1};
1001 return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
1004 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){
1005 const int b_stride= s->b_width << s->block_max_depth;
1006 BlockNode *block= &s->block[mb_x + mb_y * b_stride];
1007 BlockNode backup[4]= {block[0], block[1], block[b_stride], block[b_stride+1]};
1011 assert(mb_x>=0 && mb_y>=0);
1012 assert(mb_x<b_stride);
1013 assert(((mb_x|mb_y)&1) == 0);
1015 index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
1016 value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
1017 if(s->me_cache[index] == value)
1019 s->me_cache[index]= value;
1024 block->type &= ~BLOCK_INTRA;
1025 block[1]= block[b_stride]= block[b_stride+1]= *block;
1027 rd= get_4block_rd(s, mb_x, mb_y, 0);
1034 block[0]= backup[0];
1035 block[1]= backup[1];
1036 block[b_stride]= backup[2];
1037 block[b_stride+1]= backup[3];
1042 static void iterative_me(SnowContext *s){
1043 int pass, mb_x, mb_y;
1044 const int b_width = s->b_width << s->block_max_depth;
1045 const int b_height= s->b_height << s->block_max_depth;
1046 const int b_stride= b_width;
1050 RangeCoder r = s->c;
1051 uint8_t state[sizeof(s->block_state)];
1052 memcpy(state, s->block_state, sizeof(s->block_state));
1053 for(mb_y= 0; mb_y<s->b_height; mb_y++)
1054 for(mb_x= 0; mb_x<s->b_width; mb_x++)
1055 encode_q_branch(s, 0, mb_x, mb_y);
1057 memcpy(s->block_state, state, sizeof(s->block_state));
1060 for(pass=0; pass<25; pass++){
1063 for(mb_y= 0; mb_y<b_height; mb_y++){
1064 for(mb_x= 0; mb_x<b_width; mb_x++){
1065 int dia_change, i, j, ref;
1066 int best_rd= INT_MAX, ref_rd;
1067 BlockNode backup, ref_b;
1068 const int index= mb_x + mb_y * b_stride;
1069 BlockNode *block= &s->block[index];
1070 BlockNode *tb = mb_y ? &s->block[index-b_stride ] : NULL;
1071 BlockNode *lb = mb_x ? &s->block[index -1] : NULL;
1072 BlockNode *rb = mb_x+1<b_width ? &s->block[index +1] : NULL;
1073 BlockNode *bb = mb_y+1<b_height ? &s->block[index+b_stride ] : NULL;
1074 BlockNode *tlb= mb_x && mb_y ? &s->block[index-b_stride-1] : NULL;
1075 BlockNode *trb= mb_x+1<b_width && mb_y ? &s->block[index-b_stride+1] : NULL;
1076 BlockNode *blb= mb_x && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1077 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1078 const int b_w= (MB_SIZE >> s->block_max_depth);
1079 uint8_t obmc_edged[b_w*2][b_w*2];
1081 if(pass && (block->type & BLOCK_OPT))
1083 block->type |= BLOCK_OPT;
1087 if(!s->me_cache_generation)
1088 memset(s->me_cache, 0, sizeof(s->me_cache));
1089 s->me_cache_generation += 1<<22;
1091 //FIXME precalculate
1094 memcpy(obmc_edged, ff_obmc_tab[s->block_max_depth], b_w*b_w*4);
1096 for(y=0; y<b_w*2; y++)
1097 memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1098 if(mb_x==b_stride-1)
1099 for(y=0; y<b_w*2; y++)
1100 memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1102 for(x=0; x<b_w*2; x++)
1103 obmc_edged[0][x] += obmc_edged[b_w-1][x];
1104 for(y=1; y<b_w; y++)
1105 memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1107 if(mb_y==b_height-1){
1108 for(x=0; x<b_w*2; x++)
1109 obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1110 for(y=b_w; y<b_w*2-1; y++)
1111 memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1115 //skip stuff outside the picture
1116 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1117 uint8_t *src= s-> input_picture.data[0];
1118 uint8_t *dst= s->current_picture.data[0];
1119 const int stride= s->current_picture.linesize[0];
1120 const int block_w= MB_SIZE >> s->block_max_depth;
1121 const int sx= block_w*mb_x - block_w/2;
1122 const int sy= block_w*mb_y - block_w/2;
1123 const int w= s->plane[0].width;
1124 const int h= s->plane[0].height;
1128 memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1129 for(y=h; y<sy+block_w*2; y++)
1130 memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1132 for(y=sy; y<sy+block_w*2; y++)
1133 memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1135 if(sx+block_w*2 > w){
1136 for(y=sy; y<sy+block_w*2; y++)
1137 memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1141 // intra(black) = neighbors' contribution to the current block
1143 color[i]= get_dc(s, mb_x, mb_y, i);
1145 // get previous score (cannot be cached due to OBMC)
1146 if(pass > 0 && (block->type&BLOCK_INTRA)){
1147 int color0[3]= {block->color[0], block->color[1], block->color[2]};
1148 check_block(s, mb_x, mb_y, color0, 1, *obmc_edged, &best_rd);
1150 check_block_inter(s, mb_x, mb_y, block->mx, block->my, *obmc_edged, &best_rd);
1154 for(ref=0; ref < s->ref_frames; ref++){
1155 int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1156 if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1161 check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], *obmc_edged, &best_rd);
1162 check_block_inter(s, mb_x, mb_y, 0, 0, *obmc_edged, &best_rd);
1164 check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], *obmc_edged, &best_rd);
1166 check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], *obmc_edged, &best_rd);
1168 check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], *obmc_edged, &best_rd);
1170 check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], *obmc_edged, &best_rd);
1173 //FIXME avoid subpel interpolation / round to nearest integer
1176 for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
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 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), *obmc_edged, &best_rd);
1187 static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1190 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], *obmc_edged, &best_rd);
1192 //FIXME or try the standard 2 pass qpel or similar
1194 mvr[0][0]= block->mx;
1195 mvr[0][1]= block->my;
1196 if(ref_rd > best_rd){
1203 check_block(s, mb_x, mb_y, color, 1, *obmc_edged, &best_rd);
1204 //FIXME RD style color selection
1205 if(!same_block(block, &backup)){
1206 if(tb ) tb ->type &= ~BLOCK_OPT;
1207 if(lb ) lb ->type &= ~BLOCK_OPT;
1208 if(rb ) rb ->type &= ~BLOCK_OPT;
1209 if(bb ) bb ->type &= ~BLOCK_OPT;
1210 if(tlb) tlb->type &= ~BLOCK_OPT;
1211 if(trb) trb->type &= ~BLOCK_OPT;
1212 if(blb) blb->type &= ~BLOCK_OPT;
1213 if(brb) brb->type &= ~BLOCK_OPT;
1218 av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1223 if(s->block_max_depth == 1){
1225 for(mb_y= 0; mb_y<b_height; mb_y+=2){
1226 for(mb_x= 0; mb_x<b_width; mb_x+=2){
1228 int best_rd, init_rd;
1229 const int index= mb_x + mb_y * b_stride;
1232 b[0]= &s->block[index];
1234 b[2]= b[0]+b_stride;
1236 if(same_block(b[0], b[1]) &&
1237 same_block(b[0], b[2]) &&
1238 same_block(b[0], b[3]))
1241 if(!s->me_cache_generation)
1242 memset(s->me_cache, 0, sizeof(s->me_cache));
1243 s->me_cache_generation += 1<<22;
1245 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1247 //FIXME more multiref search?
1248 check_4block_inter(s, mb_x, mb_y,
1249 (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1250 (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1253 if(!(b[i]->type&BLOCK_INTRA))
1254 check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1256 if(init_rd != best_rd)
1260 av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1264 static void encode_blocks(SnowContext *s, int search){
1269 if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1273 if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1274 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1278 if(s->avctx->me_method == ME_ITER || !search)
1279 encode_q_branch2(s, 0, x, y);
1281 encode_q_branch (s, 0, x, y);
1286 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1287 const int w= b->width;
1288 const int h= b->height;
1289 const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1290 const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1291 int x,y, thres1, thres2;
1293 if(s->qlog == LOSSLESS_QLOG){
1296 dst[x + y*stride]= src[x + y*stride];
1300 bias= bias ? 0 : (3*qmul)>>3;
1301 thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1307 int i= src[x + y*stride];
1309 if((unsigned)(i+thres1) > thres2){
1312 i/= qmul; //FIXME optimize
1313 dst[x + y*stride]= i;
1317 i/= qmul; //FIXME optimize
1318 dst[x + y*stride]= -i;
1321 dst[x + y*stride]= 0;
1327 int i= src[x + y*stride];
1329 if((unsigned)(i+thres1) > thres2){
1332 i= (i + bias) / qmul; //FIXME optimize
1333 dst[x + y*stride]= i;
1337 i= (i + bias) / qmul; //FIXME optimize
1338 dst[x + y*stride]= -i;
1341 dst[x + y*stride]= 0;
1347 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1348 const int w= b->width;
1349 const int h= b->height;
1350 const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1351 const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1352 const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1355 if(s->qlog == LOSSLESS_QLOG) return;
1359 int i= src[x + y*stride];
1361 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1363 src[x + y*stride]= (( i*qmul + qadd)>>(QEXPSHIFT));
1369 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1370 const int w= b->width;
1371 const int h= b->height;
1374 for(y=h-1; y>=0; y--){
1375 for(x=w-1; x>=0; x--){
1376 int i= x + y*stride;
1380 if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1381 else src[i] -= src[i - 1];
1383 if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1384 else src[i] -= src[i - 1];
1387 if(y) src[i] -= src[i - stride];
1393 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1394 const int w= b->width;
1395 const int h= b->height;
1400 int i= x + y*stride;
1404 if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1405 else src[i] += src[i - 1];
1407 if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1408 else src[i] += src[i - 1];
1411 if(y) src[i] += src[i - stride];
1417 static void encode_qlogs(SnowContext *s){
1418 int plane_index, level, orientation;
1420 for(plane_index=0; plane_index<2; plane_index++){
1421 for(level=0; level<s->spatial_decomposition_count; level++){
1422 for(orientation=level ? 1:0; orientation<4; orientation++){
1423 if(orientation==2) continue;
1424 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1430 static void encode_header(SnowContext *s){
1434 memset(kstate, MID_STATE, sizeof(kstate));
1436 put_rac(&s->c, kstate, s->keyframe);
1437 if(s->keyframe || s->always_reset){
1438 ff_snow_reset_contexts(s);
1439 s->last_spatial_decomposition_type=
1443 s->last_block_max_depth= 0;
1444 for(plane_index=0; plane_index<2; plane_index++){
1445 Plane *p= &s->plane[plane_index];
1448 memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1452 put_symbol(&s->c, s->header_state, s->version, 0);
1453 put_rac(&s->c, s->header_state, s->always_reset);
1454 put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1455 put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1456 put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1457 put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1458 put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1459 put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1460 put_rac(&s->c, s->header_state, s->spatial_scalability);
1461 // put_rac(&s->c, s->header_state, s->rate_scalability);
1462 put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1469 for(plane_index=0; plane_index<2; plane_index++){
1470 Plane *p= &s->plane[plane_index];
1471 update_mc |= p->last_htaps != p->htaps;
1472 update_mc |= p->last_diag_mc != p->diag_mc;
1473 update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1475 put_rac(&s->c, s->header_state, update_mc);
1477 for(plane_index=0; plane_index<2; plane_index++){
1478 Plane *p= &s->plane[plane_index];
1479 put_rac(&s->c, s->header_state, p->diag_mc);
1480 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1481 for(i= p->htaps/2; i; i--)
1482 put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1485 if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1486 put_rac(&s->c, s->header_state, 1);
1487 put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1490 put_rac(&s->c, s->header_state, 0);
1493 put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1494 put_symbol(&s->c, s->header_state, s->qlog - s->last_qlog , 1);
1495 put_symbol(&s->c, s->header_state, s->mv_scale - s->last_mv_scale, 1);
1496 put_symbol(&s->c, s->header_state, s->qbias - s->last_qbias , 1);
1497 put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1501 static void update_last_header_values(SnowContext *s){
1505 for(plane_index=0; plane_index<2; plane_index++){
1506 Plane *p= &s->plane[plane_index];
1507 p->last_diag_mc= p->diag_mc;
1508 p->last_htaps = p->htaps;
1509 memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1513 s->last_spatial_decomposition_type = s->spatial_decomposition_type;
1514 s->last_qlog = s->qlog;
1515 s->last_qbias = s->qbias;
1516 s->last_mv_scale = s->mv_scale;
1517 s->last_block_max_depth = s->block_max_depth;
1518 s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1521 static int qscale2qlog(int qscale){
1522 return rint(QROOT*log(qscale / (float)FF_QP2LAMBDA)/log(2))
1523 + 61*QROOT/8; ///< 64 > 60
1526 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1528 /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1529 * FIXME we know exact mv bits at this point,
1530 * but ratecontrol isn't set up to include them. */
1531 uint32_t coef_sum= 0;
1532 int level, orientation, delta_qlog;
1534 for(level=0; level<s->spatial_decomposition_count; level++){
1535 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1536 SubBand *b= &s->plane[0].band[level][orientation];
1537 IDWTELEM *buf= b->ibuf;
1538 const int w= b->width;
1539 const int h= b->height;
1540 const int stride= b->stride;
1541 const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1542 const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1543 const int qdiv= (1<<16)/qmul;
1545 //FIXME this is ugly
1548 buf[x+y*stride]= b->buf[x+y*stride];
1550 decorrelate(s, b, buf, stride, 1, 0);
1553 coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1557 /* ugly, ratecontrol just takes a sqrt again */
1558 coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1559 assert(coef_sum < INT_MAX);
1561 if(pict->pict_type == AV_PICTURE_TYPE_I){
1562 s->m.current_picture.mb_var_sum= coef_sum;
1563 s->m.current_picture.mc_mb_var_sum= 0;
1565 s->m.current_picture.mc_mb_var_sum= coef_sum;
1566 s->m.current_picture.mb_var_sum= 0;
1569 pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1570 if (pict->quality < 0)
1572 s->lambda= pict->quality * 3/2;
1573 delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1574 s->qlog+= delta_qlog;
1578 static void calculate_visual_weight(SnowContext *s, Plane *p){
1579 int width = p->width;
1580 int height= p->height;
1581 int level, orientation, x, y;
1583 for(level=0; level<s->spatial_decomposition_count; level++){
1584 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1585 SubBand *b= &p->band[level][orientation];
1586 IDWTELEM *ibuf= b->ibuf;
1589 memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1590 ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1591 ff_spatial_idwt(s->spatial_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1592 for(y=0; y<height; y++){
1593 for(x=0; x<width; x++){
1594 int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1599 b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1604 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1605 const AVFrame *pict, int *got_packet)
1607 SnowContext *s = avctx->priv_data;
1608 RangeCoder * const c= &s->c;
1609 AVFrame *pic = &s->new_picture;
1610 const int width= s->avctx->width;
1611 const int height= s->avctx->height;
1612 int level, orientation, plane_index, i, y, ret;
1613 uint8_t rc_header_bak[sizeof(s->header_state)];
1614 uint8_t rc_block_bak[sizeof(s->block_state)];
1616 if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0)
1619 ff_init_range_encoder(c, pkt->data, pkt->size);
1620 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1624 for(y=0; y<(height>>shift); y++)
1625 memcpy(&s->input_picture.data[i][y * s->input_picture.linesize[i]],
1626 &pict->data[i][y * pict->linesize[i]],
1629 s->new_picture = *pict;
1631 s->m.picture_number= avctx->frame_number;
1632 if(avctx->flags&CODEC_FLAG_PASS2){
1633 s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1634 s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1635 if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1636 pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1637 if (pic->quality < 0)
1641 s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1642 s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1645 if(s->pass1_rc && avctx->frame_number == 0)
1646 pic->quality = 2*FF_QP2LAMBDA;
1648 s->qlog = qscale2qlog(pic->quality);
1649 s->lambda = pic->quality * 3/2;
1651 if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1652 s->qlog= LOSSLESS_QLOG;
1654 }//else keep previous frame's qlog until after motion estimation
1656 ff_snow_frame_start(s);
1658 s->m.current_picture_ptr= &s->m.current_picture;
1659 s->m.last_picture.f.pts = s->m.current_picture.f.pts;
1660 s->m.current_picture.f.pts = pict->pts;
1661 if(pic->pict_type == AV_PICTURE_TYPE_P){
1662 int block_width = (width +15)>>4;
1663 int block_height= (height+15)>>4;
1664 int stride= s->current_picture.linesize[0];
1666 assert(s->current_picture.data[0]);
1667 assert(s->last_picture[0].data[0]);
1669 s->m.avctx= s->avctx;
1670 s->m.current_picture.f.data[0] = s->current_picture.data[0];
1671 s->m. last_picture.f.data[0] = s->last_picture[0].data[0];
1672 s->m. new_picture.f.data[0] = s-> input_picture.data[0];
1673 s->m. last_picture_ptr= &s->m. last_picture;
1675 s->m. last_picture.f.linesize[0] =
1676 s->m. new_picture.f.linesize[0] =
1677 s->m.current_picture.f.linesize[0] = stride;
1678 s->m.uvlinesize= s->current_picture.linesize[1];
1680 s->m.height= height;
1681 s->m.mb_width = block_width;
1682 s->m.mb_height= block_height;
1683 s->m.mb_stride= s->m.mb_width+1;
1684 s->m.b8_stride= 2*s->m.mb_width+1;
1686 s->m.pict_type = pic->pict_type;
1687 s->m.me_method= s->avctx->me_method;
1688 s->m.me.scene_change_score=0;
1689 s->m.flags= s->avctx->flags;
1690 s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1691 s->m.out_format= FMT_H263;
1692 s->m.unrestricted_mv= 1;
1694 s->m.lambda = s->lambda;
1695 s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1696 s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1698 s->m.dsp= s->dsp; //move
1704 memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1705 memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1710 if (pic->pict_type == AV_PICTURE_TYPE_I)
1711 s->spatial_decomposition_count= 5;
1713 s->spatial_decomposition_count= 5;
1715 while( !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1716 || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1717 s->spatial_decomposition_count--;
1719 s->m.pict_type = pic->pict_type;
1720 s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1722 ff_snow_common_init_after_header(avctx);
1724 if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1725 for(plane_index=0; plane_index<3; plane_index++){
1726 calculate_visual_weight(s, &s->plane[plane_index]);
1731 s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1732 encode_blocks(s, 1);
1733 s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1735 for(plane_index=0; plane_index<3; plane_index++){
1736 Plane *p= &s->plane[plane_index];
1740 // int bits= put_bits_count(&s->c.pb);
1742 if (!s->memc_only) {
1744 if(pict->data[plane_index]) //FIXME gray hack
1747 s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1750 predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1753 && pic->pict_type == AV_PICTURE_TYPE_P
1754 && !(avctx->flags&CODEC_FLAG_PASS2)
1755 && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1756 ff_init_range_encoder(c, pkt->data, pkt->size);
1757 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1758 pic->pict_type= AV_PICTURE_TYPE_I;
1760 s->current_picture.key_frame=1;
1764 if(s->qlog == LOSSLESS_QLOG){
1767 s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1773 s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1779 dwt_quantize(s, p, s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type);
1781 ff_spatial_dwt(s->spatial_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1783 if(s->pass1_rc && plane_index==0){
1784 int delta_qlog = ratecontrol_1pass(s, pic);
1785 if (delta_qlog <= INT_MIN)
1788 //reordering qlog in the bitstream would eliminate this reset
1789 ff_init_range_encoder(c, pkt->data, pkt->size);
1790 memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1791 memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1793 encode_blocks(s, 0);
1797 for(level=0; level<s->spatial_decomposition_count; level++){
1798 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1799 SubBand *b= &p->band[level][orientation];
1802 quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1804 decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1805 if (!s->no_bitstream)
1806 encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1807 assert(b->parent==NULL || b->parent->stride == b->stride*2);
1809 correlate(s, b, b->ibuf, b->stride, 1, 0);
1813 for(level=0; level<s->spatial_decomposition_count; level++){
1814 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1815 SubBand *b= &p->band[level][orientation];
1817 dequantize(s, b, b->ibuf, b->stride);
1821 ff_spatial_idwt(s->spatial_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1822 if(s->qlog == LOSSLESS_QLOG){
1825 s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1829 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1832 if(pic->pict_type == AV_PICTURE_TYPE_I){
1835 s->current_picture.data[plane_index][y*s->current_picture.linesize[plane_index] + x]=
1836 pict->data[plane_index][y*pict->linesize[plane_index] + x];
1840 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1841 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1844 if(s->avctx->flags&CODEC_FLAG_PSNR){
1847 if(pict->data[plane_index]) //FIXME gray hack
1850 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];
1854 s->avctx->error[plane_index] += error;
1855 s->current_picture.error[plane_index] = error;
1860 update_last_header_values(s);
1862 ff_snow_release_buffer(avctx);
1864 s->current_picture.coded_picture_number = avctx->frame_number;
1865 s->current_picture.pict_type = pict->pict_type;
1866 s->current_picture.quality = pict->quality;
1867 s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1868 s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1869 s->m.current_picture.f.display_picture_number =
1870 s->m.current_picture.f.coded_picture_number = avctx->frame_number;
1871 s->m.current_picture.f.quality = pic->quality;
1872 s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1874 if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1876 if(avctx->flags&CODEC_FLAG_PASS1)
1877 ff_write_pass1_stats(&s->m);
1878 s->m.last_pict_type = s->m.pict_type;
1879 avctx->frame_bits = s->m.frame_bits;
1880 avctx->mv_bits = s->m.mv_bits;
1881 avctx->misc_bits = s->m.misc_bits;
1882 avctx->p_tex_bits = s->m.p_tex_bits;
1886 pkt->size = ff_rac_terminate(c);
1887 if (avctx->coded_frame->key_frame)
1888 pkt->flags |= AV_PKT_FLAG_KEY;
1894 static av_cold int encode_end(AVCodecContext *avctx)
1896 SnowContext *s = avctx->priv_data;
1898 ff_snow_common_end(s);
1899 if (s->input_picture.data[0])
1900 avctx->release_buffer(avctx, &s->input_picture);
1901 av_free(avctx->stats_out);
1907 #define OFFSET(x) offsetof(SnowContext, x)
1908 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1909 static const AVOption options[] = {
1910 { "memc_only", "Only do ME/MC (I frames -> ref, P frame -> ME+MC).", OFFSET(memc_only), AV_OPT_TYPE_INT, { 0 }, 0, 1, VE },
1911 { "no_bitstream", "Skip final bitstream writeout.", OFFSET(no_bitstream), AV_OPT_TYPE_INT, { 0 }, 0, 1, VE },
1915 static const AVClass snowenc_class = {
1916 .class_name = "snow encoder",
1917 .item_name = av_default_item_name,
1919 .version = LIBAVUTIL_VERSION_INT,
1922 AVCodec ff_snow_encoder = {
1924 .type = AVMEDIA_TYPE_VIDEO,
1925 .id = CODEC_ID_SNOW,
1926 .priv_data_size = sizeof(SnowContext),
1927 .init = encode_init,
1928 .encode2 = encode_frame,
1929 .close = encode_end,
1930 .long_name = NULL_IF_CONFIG_SMALL("Snow"),
1931 .priv_class = &snowenc_class,
1940 #include "libavutil/lfg.h"
1941 #include "libavutil/mathematics.h"
1946 int buffer[2][width*height];
1950 s.spatial_decomposition_count=6;
1951 s.spatial_decomposition_type=1;
1953 av_lfg_init(&prng, 1);
1955 printf("testing 5/3 DWT\n");
1956 for(i=0; i<width*height; i++)
1957 buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1959 ff_spatial_dwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1960 ff_spatial_idwt((IDWTELEM*)buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1962 for(i=0; i<width*height; i++)
1963 if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1965 printf("testing 9/7 DWT\n");
1966 s.spatial_decomposition_type=0;
1967 for(i=0; i<width*height; i++)
1968 buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1970 ff_spatial_dwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1971 ff_spatial_idwt((IDWTELEM*)buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1973 for(i=0; i<width*height; i++)
1974 if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1977 int level, orientation, x, y;
1978 int64_t errors[8][4];
1981 memset(errors, 0, sizeof(errors));
1982 s.spatial_decomposition_count=3;
1983 s.spatial_decomposition_type=0;
1984 for(level=0; level<s.spatial_decomposition_count; level++){
1985 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1986 int w= width >> (s.spatial_decomposition_count-level);
1987 int h= height >> (s.spatial_decomposition_count-level);
1988 int stride= width << (s.spatial_decomposition_count-level);
1989 DWTELEM *buf= buffer[0];
1992 if(orientation&1) buf+=w;
1993 if(orientation>1) buf+=stride>>1;
1995 memset(buffer[0], 0, sizeof(int)*width*height);
1996 buf[w/2 + h/2*stride]= 256*256;
1997 ff_spatial_idwt((IDWTELEM*)buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1998 for(y=0; y<height; y++){
1999 for(x=0; x<width; x++){
2000 int64_t d= buffer[0][x + y*width];
2002 if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
2004 if(FFABS(height/2-y)<9 && level==2) printf("\n");
2006 error= (int)(sqrt(error)+0.5);
2007 errors[level][orientation]= error;
2008 if(g) g=av_gcd(g, error);
2012 printf("static int const visual_weight[][4]={\n");
2013 for(level=0; level<s.spatial_decomposition_count; level++){
2015 for(orientation=0; orientation<4; orientation++){
2016 printf("%8"PRId64",", errors[level][orientation]/g);
2023 int w= width >> (s.spatial_decomposition_count-level);
2024 //int h= height >> (s.spatial_decomposition_count-level);
2025 int stride= width << (s.spatial_decomposition_count-level);
2026 DWTELEM *buf= buffer[0];
2032 memset(buffer[0], 0, sizeof(int)*width*height);
2033 for(y=0; y<height; y++){
2034 for(x=0; x<width; x++){
2035 int tab[4]={0,2,3,1};
2036 buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2039 ff_spatial_dwt(buffer[0], width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2040 for(y=0; y<height; y++){
2041 for(x=0; x<width; x++){
2042 int64_t d= buffer[0][x + y*width];
2044 if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2046 if(FFABS(height/2-y)<9) printf("\n");