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"
31 #include "rangecoder.h"
34 #include "mpegvideo.h"
37 static av_cold int encode_init(AVCodecContext *avctx)
39 SnowContext *s = avctx->priv_data;
42 if(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL){
43 av_log(avctx, AV_LOG_ERROR, "This codec is under development, files encoded with it may not be decodable with future versions!!!\n"
44 "Use vstrict=-2 / -strict -2 to use it anyway.\n");
48 if(avctx->prediction_method == DWT_97
49 && (avctx->flags & CODEC_FLAG_QSCALE)
50 && avctx->global_quality == 0){
51 av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
55 s->spatial_decomposition_type= avctx->prediction_method; //FIXME add decorrelator type r transform_type
57 s->mv_scale = (avctx->flags & CODEC_FLAG_QPEL) ? 2 : 4;
58 s->block_max_depth= (avctx->flags & CODEC_FLAG_4MV ) ? 1 : 0;
60 for(plane_index=0; plane_index<3; plane_index++){
61 s->plane[plane_index].diag_mc= 1;
62 s->plane[plane_index].htaps= 6;
63 s->plane[plane_index].hcoeff[0]= 40;
64 s->plane[plane_index].hcoeff[1]= -10;
65 s->plane[plane_index].hcoeff[2]= 2;
66 s->plane[plane_index].fast_mc= 1;
69 if ((ret = ff_snow_common_init(avctx)) < 0) {
70 ff_snow_common_end(avctx->priv_data);
73 ff_snow_alloc_blocks(s);
78 s->m.flags = avctx->flags;
79 s->m.bit_rate= avctx->bit_rate;
82 s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
83 s->m.me.map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
84 s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
85 s->m.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
86 if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.obmc_scratchpad)
87 return AVERROR(ENOMEM);
89 ff_h263_encode_init(&s->m); //mv_penalty
91 s->max_ref_frames = FFMAX(FFMIN(avctx->refs, MAX_REF_FRAMES), 1);
93 if(avctx->flags&CODEC_FLAG_PASS1){
95 avctx->stats_out = av_mallocz(256);
97 if (!avctx->stats_out)
98 return AVERROR(ENOMEM);
100 if((avctx->flags&CODEC_FLAG_PASS2) || !(avctx->flags&CODEC_FLAG_QSCALE)){
101 if(ff_rate_control_init(&s->m) < 0)
104 s->pass1_rc= !(avctx->flags & (CODEC_FLAG_QSCALE|CODEC_FLAG_PASS2));
106 switch(avctx->pix_fmt){
107 case AV_PIX_FMT_YUV444P:
108 // case AV_PIX_FMT_YUV422P:
109 case AV_PIX_FMT_YUV420P:
110 // case AV_PIX_FMT_YUV411P:
111 case AV_PIX_FMT_YUV410P:
113 s->colorspace_type= 0;
115 case AV_PIX_FMT_GRAY8:
117 s->colorspace_type = 1;
119 /* case AV_PIX_FMT_RGB32:
123 av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
126 avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift, &s->chroma_v_shift);
128 ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
129 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
131 s->input_picture = av_frame_alloc();
132 if (!s->input_picture)
133 return AVERROR(ENOMEM);
134 if ((ret = ff_get_buffer(s->avctx, s->input_picture, AV_GET_BUFFER_FLAG_REF)) < 0)
137 if(s->avctx->me_method == ME_ITER){
139 int size= s->b_width * s->b_height << 2*s->block_max_depth;
140 for(i=0; i<s->max_ref_frames; i++){
141 s->ref_mvs[i]= av_mallocz(size*sizeof(int16_t[2]));
142 s->ref_scores[i]= av_mallocz(size*sizeof(uint32_t));
143 if (!s->ref_mvs[i] || !s->ref_scores[i])
144 return AVERROR(ENOMEM);
151 //near copy & paste from dsputil, FIXME
152 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
157 for (i = 0; i < h; i++) {
158 for (j = 0; j < w; j++) {
162 pix += line_size - w;
167 //near copy & paste from dsputil, FIXME
168 static int pix_norm1(uint8_t * pix, int line_size, int w)
171 uint32_t *sq = ff_squareTbl + 256;
174 for (i = 0; i < w; i++) {
175 for (j = 0; j < w; j ++) {
179 pix += line_size - w;
184 static inline int get_penalty_factor(int lambda, int lambda2, int type){
188 return lambda>>FF_LAMBDA_SHIFT;
190 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
192 return (4*lambda)>>(FF_LAMBDA_SHIFT);
194 return (2*lambda)>>(FF_LAMBDA_SHIFT);
197 return (2*lambda)>>FF_LAMBDA_SHIFT;
202 return lambda2>>FF_LAMBDA_SHIFT;
211 #define P_TOPRIGHT P[3]
212 #define P_MEDIAN P[4]
214 #define FLAG_QPEL 1 //must be 1
216 static int encode_q_branch(SnowContext *s, int level, int x, int y){
217 uint8_t p_buffer[1024];
218 uint8_t i_buffer[1024];
219 uint8_t p_state[sizeof(s->block_state)];
220 uint8_t i_state[sizeof(s->block_state)];
222 uint8_t *pbbak= s->c.bytestream;
223 uint8_t *pbbak_start= s->c.bytestream_start;
224 int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
225 const int w= s->b_width << s->block_max_depth;
226 const int h= s->b_height << s->block_max_depth;
227 const int rem_depth= s->block_max_depth - level;
228 const int index= (x + y*w) << rem_depth;
229 const int block_w= 1<<(LOG2_MB_SIZE - level);
230 int trx= (x+1)<<rem_depth;
231 int try= (y+1)<<rem_depth;
232 const BlockNode *left = x ? &s->block[index-1] : &null_block;
233 const BlockNode *top = y ? &s->block[index-w] : &null_block;
234 const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
235 const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
236 const BlockNode *tl = y && x ? &s->block[index-w-1] : left;
237 const BlockNode *tr = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
238 int pl = left->color[0];
239 int pcb= left->color[1];
240 int pcr= left->color[2];
244 const int stride= s->current_picture->linesize[0];
245 const int uvstride= s->current_picture->linesize[1];
246 uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y* stride)*block_w,
247 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
248 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
250 int16_t last_mv[3][2];
251 int qpel= !!(s->avctx->flags & CODEC_FLAG_QPEL); //unused
252 const int shift= 1+qpel;
253 MotionEstContext *c= &s->m.me;
254 int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
255 int mx_context= av_log2(2*FFABS(left->mx - top->mx));
256 int my_context= av_log2(2*FFABS(left->my - top->my));
257 int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
258 int ref, best_ref, ref_score, ref_mx, ref_my;
260 av_assert0(sizeof(s->block_state) >= 256);
262 set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
266 // clip predictors / edge ?
272 P_TOPRIGHT[0]= tr->mx;
273 P_TOPRIGHT[1]= tr->my;
275 last_mv[0][0]= s->block[index].mx;
276 last_mv[0][1]= s->block[index].my;
277 last_mv[1][0]= right->mx;
278 last_mv[1][1]= right->my;
279 last_mv[2][0]= bottom->mx;
280 last_mv[2][1]= bottom->my;
287 av_assert1(c-> stride == stride);
288 av_assert1(c->uvstride == uvstride);
290 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
291 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
292 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
293 c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_MV;
295 c->xmin = - x*block_w - 16+3;
296 c->ymin = - y*block_w - 16+3;
297 c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
298 c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
300 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
301 if(P_LEFT[1] > (c->ymax<<shift)) P_LEFT[1] = (c->ymax<<shift);
302 if(P_TOP[0] > (c->xmax<<shift)) P_TOP[0] = (c->xmax<<shift);
303 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
304 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
305 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
306 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
308 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
309 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
312 c->pred_x= P_LEFT[0];
313 c->pred_y= P_LEFT[1];
315 c->pred_x = P_MEDIAN[0];
316 c->pred_y = P_MEDIAN[1];
321 for(ref=0; ref<s->ref_frames; ref++){
322 init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
324 ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
325 (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
327 av_assert2(ref_mx >= c->xmin);
328 av_assert2(ref_mx <= c->xmax);
329 av_assert2(ref_my >= c->ymin);
330 av_assert2(ref_my <= c->ymax);
332 ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
333 ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
334 ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
336 s->ref_mvs[ref][index][0]= ref_mx;
337 s->ref_mvs[ref][index][1]= ref_my;
338 s->ref_scores[ref][index]= ref_score;
340 if(score > ref_score){
347 //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
350 base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
353 pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
354 memcpy(p_state, s->block_state, sizeof(s->block_state));
356 if(level!=s->block_max_depth)
357 put_rac(&pc, &p_state[4 + s_context], 1);
358 put_rac(&pc, &p_state[1 + left->type + top->type], 0);
359 if(s->ref_frames > 1)
360 put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
361 pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
362 put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
363 put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
364 p_len= pc.bytestream - pc.bytestream_start;
365 score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
367 block_s= block_w*block_w;
368 sum = pix_sum(current_data[0], stride, block_w, block_w);
369 l= (sum + block_s/2)/block_s;
370 iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
372 if (s->nb_planes > 2) {
373 block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
374 sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
375 cb= (sum + block_s/2)/block_s;
376 // iscore += pix_norm1(¤t_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
377 sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
378 cr= (sum + block_s/2)/block_s;
379 // iscore += pix_norm1(¤t_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
385 ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
386 memcpy(i_state, s->block_state, sizeof(s->block_state));
387 if(level!=s->block_max_depth)
388 put_rac(&ic, &i_state[4 + s_context], 1);
389 put_rac(&ic, &i_state[1 + left->type + top->type], 1);
390 put_symbol(&ic, &i_state[32], l-pl , 1);
391 if (s->nb_planes > 2) {
392 put_symbol(&ic, &i_state[64], cb-pcb, 1);
393 put_symbol(&ic, &i_state[96], cr-pcr, 1);
395 i_len= ic.bytestream - ic.bytestream_start;
396 iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
398 // assert(score==256*256*256*64-1);
399 av_assert1(iscore < 255*255*256 + s->lambda2*10);
400 av_assert1(iscore >= 0);
401 av_assert1(l>=0 && l<=255);
402 av_assert1(pl>=0 && pl<=255);
405 int varc= iscore >> 8;
406 int vard= score >> 8;
407 if (vard <= 64 || vard < varc)
408 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
410 c->scene_change_score+= s->m.qscale;
413 if(level!=s->block_max_depth){
414 put_rac(&s->c, &s->block_state[4 + s_context], 0);
415 score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
416 score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
417 score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
418 score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
419 score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
421 if(score2 < score && score2 < iscore)
426 pred_mv(s, &pmx, &pmy, 0, left, top, tr);
427 memcpy(pbbak, i_buffer, i_len);
429 s->c.bytestream_start= pbbak_start;
430 s->c.bytestream= pbbak + i_len;
431 set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
432 memcpy(s->block_state, i_state, sizeof(s->block_state));
435 memcpy(pbbak, p_buffer, p_len);
437 s->c.bytestream_start= pbbak_start;
438 s->c.bytestream= pbbak + p_len;
439 set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
440 memcpy(s->block_state, p_state, sizeof(s->block_state));
445 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
446 const int w= s->b_width << s->block_max_depth;
447 const int rem_depth= s->block_max_depth - level;
448 const int index= (x + y*w) << rem_depth;
449 int trx= (x+1)<<rem_depth;
450 BlockNode *b= &s->block[index];
451 const BlockNode *left = x ? &s->block[index-1] : &null_block;
452 const BlockNode *top = y ? &s->block[index-w] : &null_block;
453 const BlockNode *tl = y && x ? &s->block[index-w-1] : left;
454 const BlockNode *tr = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
455 int pl = left->color[0];
456 int pcb= left->color[1];
457 int pcr= left->color[2];
459 int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
460 int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
461 int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
462 int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
465 set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
469 if(level!=s->block_max_depth){
470 if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
471 put_rac(&s->c, &s->block_state[4 + s_context], 1);
473 put_rac(&s->c, &s->block_state[4 + s_context], 0);
474 encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
475 encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
476 encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
477 encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
481 if(b->type & BLOCK_INTRA){
482 pred_mv(s, &pmx, &pmy, 0, left, top, tr);
483 put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
484 put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
485 if (s->nb_planes > 2) {
486 put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
487 put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
489 set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
491 pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
492 put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
493 if(s->ref_frames > 1)
494 put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
495 put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
496 put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
497 set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
501 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
503 Plane *p= &s->plane[plane_index];
504 const int block_size = MB_SIZE >> s->block_max_depth;
505 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
506 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
507 const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
508 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
509 const int ref_stride= s->current_picture->linesize[plane_index];
510 uint8_t *src= s-> input_picture->data[plane_index];
511 IDWTELEM *dst= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
512 const int b_stride = s->b_width << s->block_max_depth;
513 const int w= p->width;
514 const int h= p->height;
515 int index= mb_x + mb_y*b_stride;
516 BlockNode *b= &s->block[index];
517 BlockNode backup= *b;
521 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
523 b->type|= BLOCK_INTRA;
524 b->color[plane_index]= 0;
525 memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
528 int mb_x2= mb_x + (i &1) - 1;
529 int mb_y2= mb_y + (i>>1) - 1;
530 int x= block_w*mb_x2 + block_w/2;
531 int y= block_h*mb_y2 + block_h/2;
533 add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
534 x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
536 for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
537 for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
538 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
539 int obmc_v= obmc[index];
541 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
542 if(x<0) obmc_v += obmc[index + block_w];
543 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
544 if(x+block_w>w) obmc_v += obmc[index - block_w];
545 //FIXME precalculate this or simplify it somehow else
547 d = -dst[index] + (1<<(FRAC_BITS-1));
549 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
550 aa += obmc_v * obmc_v; //FIXME precalculate this
556 return av_clip( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa), 0, 255); //FIXME we should not need clipping
559 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
560 const int b_stride = s->b_width << s->block_max_depth;
561 const int b_height = s->b_height<< s->block_max_depth;
562 int index= x + y*b_stride;
563 const BlockNode *b = &s->block[index];
564 const BlockNode *left = x ? &s->block[index-1] : &null_block;
565 const BlockNode *top = y ? &s->block[index-b_stride] : &null_block;
566 const BlockNode *tl = y && x ? &s->block[index-b_stride-1] : left;
567 const BlockNode *tr = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
569 // int mx_context= av_log2(2*FFABS(left->mx - top->mx));
570 // int my_context= av_log2(2*FFABS(left->my - top->my));
572 if(x<0 || x>=b_stride || y>=b_height)
581 //FIXME try accurate rate
582 //FIXME intra and inter predictors if surrounding blocks are not the same type
583 if(b->type & BLOCK_INTRA){
584 return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
585 + av_log2(2*FFABS(left->color[1] - b->color[1]))
586 + av_log2(2*FFABS(left->color[2] - b->color[2])));
588 pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
591 return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
592 + av_log2(2*FFABS(dmy))
593 + av_log2(2*b->ref));
597 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
598 Plane *p= &s->plane[plane_index];
599 const int block_size = MB_SIZE >> s->block_max_depth;
600 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
601 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
602 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
603 const int ref_stride= s->current_picture->linesize[plane_index];
604 uint8_t *dst= s->current_picture->data[plane_index];
605 uint8_t *src= s-> input_picture->data[plane_index];
606 IDWTELEM *pred= (IDWTELEM*)s->m.obmc_scratchpad + plane_index*block_size*block_size*4;
607 uint8_t *cur = s->scratchbuf;
608 uint8_t *tmp = s->emu_edge_buffer;
609 const int b_stride = s->b_width << s->block_max_depth;
610 const int b_height = s->b_height<< s->block_max_depth;
611 const int w= p->width;
612 const int h= p->height;
615 const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
616 int sx= block_w*mb_x - block_w/2;
617 int sy= block_h*mb_y - block_h/2;
618 int x0= FFMAX(0,-sx);
619 int y0= FFMAX(0,-sy);
620 int x1= FFMIN(block_w*2, w-sx);
621 int y1= FFMIN(block_h*2, h-sy);
624 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
626 ff_snow_pred_block(s, cur, tmp, ref_stride, sx, sy, block_w*2, block_h*2, &s->block[mb_x + mb_y*b_stride], plane_index, w, h);
628 for(y=y0; y<y1; y++){
629 const uint8_t *obmc1= obmc_edged[y];
630 const IDWTELEM *pred1 = pred + y*obmc_stride;
631 uint8_t *cur1 = cur + y*ref_stride;
632 uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
633 for(x=x0; x<x1; x++){
634 #if FRAC_BITS >= LOG2_OBMC_MAX
635 int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
637 int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
639 v = (v + pred1[x]) >> FRAC_BITS;
640 if(v&(~255)) v= ~(v>>31);
645 /* copy the regions where obmc[] = (uint8_t)256 */
646 if(LOG2_OBMC_MAX == 8
647 && (mb_x == 0 || mb_x == b_stride-1)
648 && (mb_y == 0 || mb_y == b_height-1)){
658 memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
662 /* FIXME rearrange dsputil to fit 32x32 cmp functions */
663 /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
664 /* FIXME cmps overlap but do not cover the wavelet's whole support.
665 * So improving the score of one block is not strictly guaranteed
666 * to improve the score of the whole frame, thus iterative motion
667 * estimation does not always converge. */
668 if(s->avctx->me_cmp == FF_CMP_W97)
669 distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
670 else if(s->avctx->me_cmp == FF_CMP_W53)
671 distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
675 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
676 distortion += s->dsp.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
680 av_assert2(block_w==8);
681 distortion = s->dsp.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
690 rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
692 if(mb_x == b_stride-2)
693 rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
695 return distortion + rate*penalty_factor;
698 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
700 Plane *p= &s->plane[plane_index];
701 const int block_size = MB_SIZE >> s->block_max_depth;
702 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
703 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
704 const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
705 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
706 const int ref_stride= s->current_picture->linesize[plane_index];
707 uint8_t *dst= s->current_picture->data[plane_index];
708 uint8_t *src= s-> input_picture->data[plane_index];
709 //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
710 // const has only been removed from zero_dst to suppress a warning
711 static IDWTELEM zero_dst[4096]; //FIXME
712 const int b_stride = s->b_width << s->block_max_depth;
713 const int w= p->width;
714 const int h= p->height;
717 const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
719 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
722 int mb_x2= mb_x + (i%3) - 1;
723 int mb_y2= mb_y + (i/3) - 1;
724 int x= block_w*mb_x2 + block_w/2;
725 int y= block_h*mb_y2 + block_h/2;
727 add_yblock(s, 0, NULL, zero_dst, dst, obmc,
728 x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
730 //FIXME find a cleaner/simpler way to skip the outside stuff
731 for(y2= y; y2<0; y2++)
732 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
733 for(y2= h; y2<y+block_h; y2++)
734 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
736 for(y2= y; y2<y+block_h; y2++)
737 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
740 for(y2= y; y2<y+block_h; y2++)
741 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
744 av_assert1(block_w== 8 || block_w==16);
745 distortion += s->dsp.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
749 BlockNode *b= &s->block[mb_x+mb_y*b_stride];
750 int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
758 rate = get_block_bits(s, mb_x, mb_y, 2);
759 for(i=merged?4:0; i<9; i++){
760 static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
761 rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
764 return distortion + rate*penalty_factor;
767 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
768 const int w= b->width;
769 const int h= b->height;
774 int *runs = s->run_buffer;
781 int /*ll=0, */l=0, lt=0, t=0, rt=0;
782 v= src[x + y*stride];
785 t= src[x + (y-1)*stride];
787 lt= src[x - 1 + (y-1)*stride];
790 rt= src[x + 1 + (y-1)*stride];
794 l= src[x - 1 + y*stride];
796 if(orientation==1) ll= src[y + (x-2)*stride];
797 else ll= src[x - 2 + y*stride];
803 if(px<b->parent->width && py<b->parent->height)
804 p= parent[px + py*2*stride];
806 if(!(/*ll|*/l|lt|t|rt|p)){
808 runs[run_index++]= run;
816 max_index= run_index;
817 runs[run_index++]= run;
819 run= runs[run_index++];
821 put_symbol2(&s->c, b->state[30], max_index, 0);
822 if(run_index <= max_index)
823 put_symbol2(&s->c, b->state[1], run, 3);
826 if(s->c.bytestream_end - s->c.bytestream < w*40){
827 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
832 int /*ll=0, */l=0, lt=0, t=0, rt=0;
833 v= src[x + y*stride];
836 t= src[x + (y-1)*stride];
838 lt= src[x - 1 + (y-1)*stride];
841 rt= src[x + 1 + (y-1)*stride];
845 l= src[x - 1 + y*stride];
847 if(orientation==1) ll= src[y + (x-2)*stride];
848 else ll= src[x - 2 + y*stride];
854 if(px<b->parent->width && py<b->parent->height)
855 p= parent[px + py*2*stride];
857 if(/*ll|*/l|lt|t|rt|p){
858 int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
860 put_rac(&s->c, &b->state[0][context], !!v);
863 run= runs[run_index++];
865 if(run_index <= max_index)
866 put_symbol2(&s->c, b->state[1], run, 3);
874 int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
875 int l2= 2*FFABS(l) + (l<0);
876 int t2= 2*FFABS(t) + (t<0);
878 put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
879 put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
887 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
888 // encode_subband_qtree(s, b, src, parent, stride, orientation);
889 // encode_subband_z0run(s, b, src, parent, stride, orientation);
890 return encode_subband_c0run(s, b, src, parent, stride, orientation);
891 // encode_subband_dzr(s, b, src, parent, stride, orientation);
894 static av_always_inline int check_block(SnowContext *s, int mb_x, int mb_y, int p[3], int intra, uint8_t (*obmc_edged)[MB_SIZE * 2], int *best_rd){
895 const int b_stride= s->b_width << s->block_max_depth;
896 BlockNode *block= &s->block[mb_x + mb_y * b_stride];
897 BlockNode backup= *block;
901 av_assert2(mb_x>=0 && mb_y>=0);
902 av_assert2(mb_x<b_stride);
905 block->color[0] = p[0];
906 block->color[1] = p[1];
907 block->color[2] = p[2];
908 block->type |= BLOCK_INTRA;
910 index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
911 value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
912 if(s->me_cache[index] == value)
914 s->me_cache[index]= value;
918 block->type &= ~BLOCK_INTRA;
921 rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged);
933 /* special case for int[2] args we discard afterwards,
934 * fixes compilation problem with gcc 2.95 */
935 static av_always_inline int check_block_inter(SnowContext *s, int mb_x, int mb_y, int p0, int p1, uint8_t (*obmc_edged)[MB_SIZE * 2], int *best_rd){
937 return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
940 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){
941 const int b_stride= s->b_width << s->block_max_depth;
942 BlockNode *block= &s->block[mb_x + mb_y * b_stride];
947 /* We don't initialize backup[] during variable declaration, because
948 * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
950 backup[0] = block[0];
951 backup[1] = block[1];
952 backup[2] = block[b_stride];
953 backup[3] = block[b_stride + 1];
955 av_assert2(mb_x>=0 && mb_y>=0);
956 av_assert2(mb_x<b_stride);
957 av_assert2(((mb_x|mb_y)&1) == 0);
959 index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
960 value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
961 if(s->me_cache[index] == value)
963 s->me_cache[index]= value;
968 block->type &= ~BLOCK_INTRA;
969 block[1]= block[b_stride]= block[b_stride+1]= *block;
971 rd= get_4block_rd(s, mb_x, mb_y, 0);
980 block[b_stride]= backup[2];
981 block[b_stride+1]= backup[3];
986 static void iterative_me(SnowContext *s){
987 int pass, mb_x, mb_y;
988 const int b_width = s->b_width << s->block_max_depth;
989 const int b_height= s->b_height << s->block_max_depth;
990 const int b_stride= b_width;
995 uint8_t state[sizeof(s->block_state)];
996 memcpy(state, s->block_state, sizeof(s->block_state));
997 for(mb_y= 0; mb_y<s->b_height; mb_y++)
998 for(mb_x= 0; mb_x<s->b_width; mb_x++)
999 encode_q_branch(s, 0, mb_x, mb_y);
1001 memcpy(s->block_state, state, sizeof(s->block_state));
1004 for(pass=0; pass<25; pass++){
1007 for(mb_y= 0; mb_y<b_height; mb_y++){
1008 for(mb_x= 0; mb_x<b_width; mb_x++){
1009 int dia_change, i, j, ref;
1010 int best_rd= INT_MAX, ref_rd;
1011 BlockNode backup, ref_b;
1012 const int index= mb_x + mb_y * b_stride;
1013 BlockNode *block= &s->block[index];
1014 BlockNode *tb = mb_y ? &s->block[index-b_stride ] : NULL;
1015 BlockNode *lb = mb_x ? &s->block[index -1] : NULL;
1016 BlockNode *rb = mb_x+1<b_width ? &s->block[index +1] : NULL;
1017 BlockNode *bb = mb_y+1<b_height ? &s->block[index+b_stride ] : NULL;
1018 BlockNode *tlb= mb_x && mb_y ? &s->block[index-b_stride-1] : NULL;
1019 BlockNode *trb= mb_x+1<b_width && mb_y ? &s->block[index-b_stride+1] : NULL;
1020 BlockNode *blb= mb_x && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1021 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1022 const int b_w= (MB_SIZE >> s->block_max_depth);
1023 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1025 if(pass && (block->type & BLOCK_OPT))
1027 block->type |= BLOCK_OPT;
1031 if(!s->me_cache_generation)
1032 memset(s->me_cache, 0, sizeof(s->me_cache));
1033 s->me_cache_generation += 1<<22;
1035 //FIXME precalculate
1038 for (y = 0; y < b_w * 2; y++)
1039 memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1041 for(y=0; y<b_w*2; y++)
1042 memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1043 if(mb_x==b_stride-1)
1044 for(y=0; y<b_w*2; y++)
1045 memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1047 for(x=0; x<b_w*2; x++)
1048 obmc_edged[0][x] += obmc_edged[b_w-1][x];
1049 for(y=1; y<b_w; y++)
1050 memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1052 if(mb_y==b_height-1){
1053 for(x=0; x<b_w*2; x++)
1054 obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1055 for(y=b_w; y<b_w*2-1; y++)
1056 memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1060 //skip stuff outside the picture
1061 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1062 uint8_t *src= s-> input_picture->data[0];
1063 uint8_t *dst= s->current_picture->data[0];
1064 const int stride= s->current_picture->linesize[0];
1065 const int block_w= MB_SIZE >> s->block_max_depth;
1066 const int block_h= MB_SIZE >> s->block_max_depth;
1067 const int sx= block_w*mb_x - block_w/2;
1068 const int sy= block_h*mb_y - block_h/2;
1069 const int w= s->plane[0].width;
1070 const int h= s->plane[0].height;
1074 memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1075 for(y=h; y<sy+block_h*2; y++)
1076 memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1078 for(y=sy; y<sy+block_h*2; y++)
1079 memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1081 if(sx+block_w*2 > w){
1082 for(y=sy; y<sy+block_h*2; y++)
1083 memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1087 // intra(black) = neighbors' contribution to the current block
1088 for(i=0; i < s->nb_planes; i++)
1089 color[i]= get_dc(s, mb_x, mb_y, i);
1091 // get previous score (cannot be cached due to OBMC)
1092 if(pass > 0 && (block->type&BLOCK_INTRA)){
1093 int color0[3]= {block->color[0], block->color[1], block->color[2]};
1094 check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1096 check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1100 for(ref=0; ref < s->ref_frames; ref++){
1101 int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1102 if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1107 check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1108 check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1110 check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1112 check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1114 check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1116 check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1119 //FIXME avoid subpel interpolation / round to nearest integer
1122 for(i=0; i<FFMAX(s->avctx->dia_size, 1); i++){
1124 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1125 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1126 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+4*(i-j), block->my-(4*j), obmc_edged, &best_rd);
1127 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx-4*(i-j), block->my+(4*j), obmc_edged, &best_rd);
1133 static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1136 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1138 //FIXME or try the standard 2 pass qpel or similar
1140 mvr[0][0]= block->mx;
1141 mvr[0][1]= block->my;
1142 if(ref_rd > best_rd){
1149 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1150 //FIXME RD style color selection
1151 if(!same_block(block, &backup)){
1152 if(tb ) tb ->type &= ~BLOCK_OPT;
1153 if(lb ) lb ->type &= ~BLOCK_OPT;
1154 if(rb ) rb ->type &= ~BLOCK_OPT;
1155 if(bb ) bb ->type &= ~BLOCK_OPT;
1156 if(tlb) tlb->type &= ~BLOCK_OPT;
1157 if(trb) trb->type &= ~BLOCK_OPT;
1158 if(blb) blb->type &= ~BLOCK_OPT;
1159 if(brb) brb->type &= ~BLOCK_OPT;
1164 av_log(s->avctx, AV_LOG_ERROR, "pass:%d changed:%d\n", pass, change);
1169 if(s->block_max_depth == 1){
1171 for(mb_y= 0; mb_y<b_height; mb_y+=2){
1172 for(mb_x= 0; mb_x<b_width; mb_x+=2){
1174 int best_rd, init_rd;
1175 const int index= mb_x + mb_y * b_stride;
1178 b[0]= &s->block[index];
1180 b[2]= b[0]+b_stride;
1182 if(same_block(b[0], b[1]) &&
1183 same_block(b[0], b[2]) &&
1184 same_block(b[0], b[3]))
1187 if(!s->me_cache_generation)
1188 memset(s->me_cache, 0, sizeof(s->me_cache));
1189 s->me_cache_generation += 1<<22;
1191 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1193 //FIXME more multiref search?
1194 check_4block_inter(s, mb_x, mb_y,
1195 (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1196 (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1199 if(!(b[i]->type&BLOCK_INTRA))
1200 check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1202 if(init_rd != best_rd)
1206 av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1210 static void encode_blocks(SnowContext *s, int search){
1215 if(s->avctx->me_method == ME_ITER && !s->keyframe && search)
1219 if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1220 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1224 if(s->avctx->me_method == ME_ITER || !search)
1225 encode_q_branch2(s, 0, x, y);
1227 encode_q_branch (s, 0, x, y);
1232 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1233 const int w= b->width;
1234 const int h= b->height;
1235 const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1236 const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1237 int x,y, thres1, thres2;
1239 if(s->qlog == LOSSLESS_QLOG){
1242 dst[x + y*stride]= src[x + y*stride];
1246 bias= bias ? 0 : (3*qmul)>>3;
1247 thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1253 int i= src[x + y*stride];
1255 if((unsigned)(i+thres1) > thres2){
1258 i/= qmul; //FIXME optimize
1259 dst[x + y*stride]= i;
1263 i/= qmul; //FIXME optimize
1264 dst[x + y*stride]= -i;
1267 dst[x + y*stride]= 0;
1273 int i= src[x + y*stride];
1275 if((unsigned)(i+thres1) > thres2){
1278 i= (i + bias) / qmul; //FIXME optimize
1279 dst[x + y*stride]= i;
1283 i= (i + bias) / qmul; //FIXME optimize
1284 dst[x + y*stride]= -i;
1287 dst[x + y*stride]= 0;
1293 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1294 const int w= b->width;
1295 const int h= b->height;
1296 const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1297 const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1298 const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1301 if(s->qlog == LOSSLESS_QLOG) return;
1305 int i= src[x + y*stride];
1307 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1309 src[x + y*stride]= (( i*qmul + qadd)>>(QEXPSHIFT));
1315 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1316 const int w= b->width;
1317 const int h= b->height;
1320 for(y=h-1; y>=0; y--){
1321 for(x=w-1; x>=0; x--){
1322 int i= x + y*stride;
1326 if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1327 else src[i] -= src[i - 1];
1329 if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1330 else src[i] -= src[i - 1];
1333 if(y) src[i] -= src[i - stride];
1339 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1340 const int w= b->width;
1341 const int h= b->height;
1346 int i= x + y*stride;
1350 if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1351 else src[i] += src[i - 1];
1353 if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1354 else src[i] += src[i - 1];
1357 if(y) src[i] += src[i - stride];
1363 static void encode_qlogs(SnowContext *s){
1364 int plane_index, level, orientation;
1366 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1367 for(level=0; level<s->spatial_decomposition_count; level++){
1368 for(orientation=level ? 1:0; orientation<4; orientation++){
1369 if(orientation==2) continue;
1370 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1376 static void encode_header(SnowContext *s){
1380 memset(kstate, MID_STATE, sizeof(kstate));
1382 put_rac(&s->c, kstate, s->keyframe);
1383 if(s->keyframe || s->always_reset){
1384 ff_snow_reset_contexts(s);
1385 s->last_spatial_decomposition_type=
1389 s->last_block_max_depth= 0;
1390 for(plane_index=0; plane_index<2; plane_index++){
1391 Plane *p= &s->plane[plane_index];
1394 memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1398 put_symbol(&s->c, s->header_state, s->version, 0);
1399 put_rac(&s->c, s->header_state, s->always_reset);
1400 put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1401 put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1402 put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1403 put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1404 if (s->nb_planes > 2) {
1405 put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1406 put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1408 put_rac(&s->c, s->header_state, s->spatial_scalability);
1409 // put_rac(&s->c, s->header_state, s->rate_scalability);
1410 put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1417 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1418 Plane *p= &s->plane[plane_index];
1419 update_mc |= p->last_htaps != p->htaps;
1420 update_mc |= p->last_diag_mc != p->diag_mc;
1421 update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1423 put_rac(&s->c, s->header_state, update_mc);
1425 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1426 Plane *p= &s->plane[plane_index];
1427 put_rac(&s->c, s->header_state, p->diag_mc);
1428 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1429 for(i= p->htaps/2; i; i--)
1430 put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1433 if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1434 put_rac(&s->c, s->header_state, 1);
1435 put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1438 put_rac(&s->c, s->header_state, 0);
1441 put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1442 put_symbol(&s->c, s->header_state, s->qlog - s->last_qlog , 1);
1443 put_symbol(&s->c, s->header_state, s->mv_scale - s->last_mv_scale, 1);
1444 put_symbol(&s->c, s->header_state, s->qbias - s->last_qbias , 1);
1445 put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1449 static void update_last_header_values(SnowContext *s){
1453 for(plane_index=0; plane_index<2; plane_index++){
1454 Plane *p= &s->plane[plane_index];
1455 p->last_diag_mc= p->diag_mc;
1456 p->last_htaps = p->htaps;
1457 memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1461 s->last_spatial_decomposition_type = s->spatial_decomposition_type;
1462 s->last_qlog = s->qlog;
1463 s->last_qbias = s->qbias;
1464 s->last_mv_scale = s->mv_scale;
1465 s->last_block_max_depth = s->block_max_depth;
1466 s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1469 static int qscale2qlog(int qscale){
1470 return rint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1471 + 61*QROOT/8; ///< 64 > 60
1474 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1476 /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1477 * FIXME we know exact mv bits at this point,
1478 * but ratecontrol isn't set up to include them. */
1479 uint32_t coef_sum= 0;
1480 int level, orientation, delta_qlog;
1482 for(level=0; level<s->spatial_decomposition_count; level++){
1483 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1484 SubBand *b= &s->plane[0].band[level][orientation];
1485 IDWTELEM *buf= b->ibuf;
1486 const int w= b->width;
1487 const int h= b->height;
1488 const int stride= b->stride;
1489 const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1490 const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1491 const int qdiv= (1<<16)/qmul;
1493 //FIXME this is ugly
1496 buf[x+y*stride]= b->buf[x+y*stride];
1498 decorrelate(s, b, buf, stride, 1, 0);
1501 coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1505 /* ugly, ratecontrol just takes a sqrt again */
1506 coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1507 av_assert0(coef_sum < INT_MAX);
1509 if(pict->pict_type == AV_PICTURE_TYPE_I){
1510 s->m.current_picture.mb_var_sum= coef_sum;
1511 s->m.current_picture.mc_mb_var_sum= 0;
1513 s->m.current_picture.mc_mb_var_sum= coef_sum;
1514 s->m.current_picture.mb_var_sum= 0;
1517 pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1518 if (pict->quality < 0)
1520 s->lambda= pict->quality * 3/2;
1521 delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1522 s->qlog+= delta_qlog;
1526 static void calculate_visual_weight(SnowContext *s, Plane *p){
1527 int width = p->width;
1528 int height= p->height;
1529 int level, orientation, x, y;
1531 for(level=0; level<s->spatial_decomposition_count; level++){
1532 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1533 SubBand *b= &p->band[level][orientation];
1534 IDWTELEM *ibuf= b->ibuf;
1537 memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1538 ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1539 ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1540 for(y=0; y<height; y++){
1541 for(x=0; x<width; x++){
1542 int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1547 b->qlog= (int)(log(352256.0/sqrt(error)) / log(pow(2.0, 1.0/QROOT))+0.5);
1552 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1553 AVFrame *pict, int *got_packet)
1555 SnowContext *s = avctx->priv_data;
1556 RangeCoder * const c= &s->c;
1557 AVFrame *pic = pict;
1558 const int width= s->avctx->width;
1559 const int height= s->avctx->height;
1560 int level, orientation, plane_index, i, y, ret;
1561 uint8_t rc_header_bak[sizeof(s->header_state)];
1562 uint8_t rc_block_bak[sizeof(s->block_state)];
1564 if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + FF_MIN_BUFFER_SIZE)) < 0)
1567 ff_init_range_encoder(c, pkt->data, pkt->size);
1568 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1570 for(i=0; i < s->nb_planes; i++){
1571 int hshift= i ? s->chroma_h_shift : 0;
1572 int vshift= i ? s->chroma_v_shift : 0;
1573 for(y=0; y<(height>>vshift); y++)
1574 memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1575 &pict->data[i][y * pict->linesize[i]],
1577 s->dsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1578 width >> hshift, height >> vshift,
1579 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1580 EDGE_TOP | EDGE_BOTTOM);
1584 s->new_picture = pict;
1586 s->m.picture_number= avctx->frame_number;
1587 if(avctx->flags&CODEC_FLAG_PASS2){
1588 s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1589 s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1590 if(!(avctx->flags&CODEC_FLAG_QSCALE)) {
1591 pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1592 if (pic->quality < 0)
1596 s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1597 s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1600 if(s->pass1_rc && avctx->frame_number == 0)
1601 pic->quality = 2*FF_QP2LAMBDA;
1603 s->qlog = qscale2qlog(pic->quality);
1604 s->lambda = pic->quality * 3/2;
1606 if (s->qlog < 0 || (!pic->quality && (avctx->flags & CODEC_FLAG_QSCALE))) {
1607 s->qlog= LOSSLESS_QLOG;
1609 }//else keep previous frame's qlog until after motion estimation
1611 ff_snow_frame_start(s);
1612 avctx->coded_frame= s->current_picture;
1614 s->m.current_picture_ptr= &s->m.current_picture;
1615 s->m.last_picture.f.pts = s->m.current_picture.f.pts;
1616 s->m.current_picture.f.pts = pict->pts;
1617 if(pic->pict_type == AV_PICTURE_TYPE_P){
1618 int block_width = (width +15)>>4;
1619 int block_height= (height+15)>>4;
1620 int stride= s->current_picture->linesize[0];
1622 av_assert0(s->current_picture->data[0]);
1623 av_assert0(s->last_picture[0]->data[0]);
1625 s->m.avctx= s->avctx;
1626 s->m.current_picture.f.data[0] = s->current_picture->data[0];
1627 s->m. last_picture.f.data[0] = s->last_picture[0]->data[0];
1628 s->m. new_picture.f.data[0] = s-> input_picture->data[0];
1629 s->m. last_picture_ptr= &s->m. last_picture;
1631 s->m. last_picture.f.linesize[0] =
1632 s->m. new_picture.f.linesize[0] =
1633 s->m.current_picture.f.linesize[0] = stride;
1634 s->m.uvlinesize= s->current_picture->linesize[1];
1636 s->m.height= height;
1637 s->m.mb_width = block_width;
1638 s->m.mb_height= block_height;
1639 s->m.mb_stride= s->m.mb_width+1;
1640 s->m.b8_stride= 2*s->m.mb_width+1;
1642 s->m.pict_type = pic->pict_type;
1643 s->m.me_method= s->avctx->me_method;
1644 s->m.me.scene_change_score=0;
1645 s->m.flags= s->avctx->flags;
1646 s->m.quarter_sample= (s->avctx->flags & CODEC_FLAG_QPEL)!=0;
1647 s->m.out_format= FMT_H263;
1648 s->m.unrestricted_mv= 1;
1650 s->m.lambda = s->lambda;
1651 s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1652 s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1654 s->m.dsp= s->dsp; //move
1655 s->m.hdsp = s->hdsp;
1657 s->hdsp = s->m.hdsp;
1662 memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1663 memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1668 if (pic->pict_type == AV_PICTURE_TYPE_I)
1669 s->spatial_decomposition_count= 5;
1671 s->spatial_decomposition_count= 5;
1673 while( !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1674 || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1675 s->spatial_decomposition_count--;
1677 if (s->spatial_decomposition_count <= 0) {
1678 av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1679 return AVERROR(EINVAL);
1682 s->m.pict_type = pic->pict_type;
1683 s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1685 ff_snow_common_init_after_header(avctx);
1687 if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1688 for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1689 calculate_visual_weight(s, &s->plane[plane_index]);
1694 s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1695 encode_blocks(s, 1);
1696 s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1698 for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1699 Plane *p= &s->plane[plane_index];
1703 // int bits= put_bits_count(&s->c.pb);
1705 if (!s->memc_only) {
1707 if(pict->data[plane_index]) //FIXME gray hack
1710 s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1713 predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1716 && pic->pict_type == AV_PICTURE_TYPE_P
1717 && !(avctx->flags&CODEC_FLAG_PASS2)
1718 && s->m.me.scene_change_score > s->avctx->scenechange_threshold){
1719 ff_init_range_encoder(c, pkt->data, pkt->size);
1720 ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);
1721 pic->pict_type= AV_PICTURE_TYPE_I;
1723 s->current_picture->key_frame=1;
1727 if(s->qlog == LOSSLESS_QLOG){
1730 s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1736 s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1741 ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1743 if(s->pass1_rc && plane_index==0){
1744 int delta_qlog = ratecontrol_1pass(s, pic);
1745 if (delta_qlog <= INT_MIN)
1748 //reordering qlog in the bitstream would eliminate this reset
1749 ff_init_range_encoder(c, pkt->data, pkt->size);
1750 memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1751 memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1753 encode_blocks(s, 0);
1757 for(level=0; level<s->spatial_decomposition_count; level++){
1758 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1759 SubBand *b= &p->band[level][orientation];
1761 quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1763 decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1764 if (!s->no_bitstream)
1765 encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1766 av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1768 correlate(s, b, b->ibuf, b->stride, 1, 0);
1772 for(level=0; level<s->spatial_decomposition_count; level++){
1773 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1774 SubBand *b= &p->band[level][orientation];
1776 dequantize(s, b, b->ibuf, b->stride);
1780 ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1781 if(s->qlog == LOSSLESS_QLOG){
1784 s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1788 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1791 if(pic->pict_type == AV_PICTURE_TYPE_I){
1794 s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1795 pict->data[plane_index][y*pict->linesize[plane_index] + x];
1799 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1800 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1803 if(s->avctx->flags&CODEC_FLAG_PSNR){
1806 if(pict->data[plane_index]) //FIXME gray hack
1809 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];
1813 s->avctx->error[plane_index] += error;
1814 s->current_picture->error[plane_index] = error;
1819 update_last_header_values(s);
1821 ff_snow_release_buffer(avctx);
1823 s->current_picture->coded_picture_number = avctx->frame_number;
1824 s->current_picture->pict_type = pict->pict_type;
1825 s->current_picture->quality = pict->quality;
1826 s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1827 s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1828 s->m.current_picture.f.display_picture_number =
1829 s->m.current_picture.f.coded_picture_number = avctx->frame_number;
1830 s->m.current_picture.f.quality = pic->quality;
1831 s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1833 if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1835 if(avctx->flags&CODEC_FLAG_PASS1)
1836 ff_write_pass1_stats(&s->m);
1837 s->m.last_pict_type = s->m.pict_type;
1838 avctx->frame_bits = s->m.frame_bits;
1839 avctx->mv_bits = s->m.mv_bits;
1840 avctx->misc_bits = s->m.misc_bits;
1841 avctx->p_tex_bits = s->m.p_tex_bits;
1845 pkt->size = ff_rac_terminate(c);
1846 if (avctx->coded_frame->key_frame)
1847 pkt->flags |= AV_PKT_FLAG_KEY;
1853 static av_cold int encode_end(AVCodecContext *avctx)
1855 SnowContext *s = avctx->priv_data;
1857 ff_snow_common_end(s);
1858 ff_rate_control_uninit(&s->m);
1859 av_frame_free(&s->input_picture);
1860 av_free(avctx->stats_out);
1865 #define OFFSET(x) offsetof(SnowContext, x)
1866 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1867 static const AVOption options[] = {
1868 { "memc_only", "Only do ME/MC (I frames -> ref, P frame -> ME+MC).", OFFSET(memc_only), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1869 { "no_bitstream", "Skip final bitstream writeout.", OFFSET(no_bitstream), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1873 static const AVClass snowenc_class = {
1874 .class_name = "snow encoder",
1875 .item_name = av_default_item_name,
1877 .version = LIBAVUTIL_VERSION_INT,
1880 AVCodec ff_snow_encoder = {
1882 .type = AVMEDIA_TYPE_VIDEO,
1883 .id = AV_CODEC_ID_SNOW,
1884 .priv_data_size = sizeof(SnowContext),
1885 .init = encode_init,
1886 .encode2 = encode_frame,
1887 .close = encode_end,
1888 .pix_fmts = (const enum AVPixelFormat[]){
1889 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1893 .long_name = NULL_IF_CONFIG_SMALL("Snow"),
1894 .priv_class = &snowenc_class,
1903 #include "libavutil/lfg.h"
1904 #include "libavutil/mathematics.h"
1909 int buffer[2][width*height];
1913 s.spatial_decomposition_count=6;
1914 s.spatial_decomposition_type=1;
1916 s.temp_dwt_buffer = av_mallocz(width * sizeof(DWTELEM));
1917 s.temp_idwt_buffer = av_mallocz(width * sizeof(IDWTELEM));
1919 av_lfg_init(&prng, 1);
1921 printf("testing 5/3 DWT\n");
1922 for(i=0; i<width*height; i++)
1923 buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1925 ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1926 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1928 for(i=0; i<width*height; i++)
1929 if(buffer[0][i]!= buffer[1][i]) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1931 printf("testing 9/7 DWT\n");
1932 s.spatial_decomposition_type=0;
1933 for(i=0; i<width*height; i++)
1934 buffer[0][i] = buffer[1][i] = av_lfg_get(&prng) % 54321 - 12345;
1936 ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1937 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1939 for(i=0; i<width*height; i++)
1940 if(FFABS(buffer[0][i] - buffer[1][i])>20) printf("fsck: %6d %12d %7d\n",i, buffer[0][i], buffer[1][i]);
1943 int level, orientation, x, y;
1944 int64_t errors[8][4];
1947 memset(errors, 0, sizeof(errors));
1948 s.spatial_decomposition_count=3;
1949 s.spatial_decomposition_type=0;
1950 for(level=0; level<s.spatial_decomposition_count; level++){
1951 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1952 int w= width >> (s.spatial_decomposition_count-level);
1953 int h= height >> (s.spatial_decomposition_count-level);
1954 int stride= width << (s.spatial_decomposition_count-level);
1955 DWTELEM *buf= buffer[0];
1958 if(orientation&1) buf+=w;
1959 if(orientation>1) buf+=stride>>1;
1961 memset(buffer[0], 0, sizeof(int)*width*height);
1962 buf[w/2 + h/2*stride]= 256*256;
1963 ff_spatial_idwt((IDWTELEM*)buffer[0], s.temp_idwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
1964 for(y=0; y<height; y++){
1965 for(x=0; x<width; x++){
1966 int64_t d= buffer[0][x + y*width];
1968 if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9 && level==2) printf("%8"PRId64" ", d);
1970 if(FFABS(height/2-y)<9 && level==2) printf("\n");
1972 error= (int)(sqrt(error)+0.5);
1973 errors[level][orientation]= error;
1974 if(g) g=av_gcd(g, error);
1978 printf("static int const visual_weight[][4]={\n");
1979 for(level=0; level<s.spatial_decomposition_count; level++){
1981 for(orientation=0; orientation<4; orientation++){
1982 printf("%8"PRId64",", errors[level][orientation]/g);
1989 int w= width >> (s.spatial_decomposition_count-level);
1990 //int h= height >> (s.spatial_decomposition_count-level);
1991 int stride= width << (s.spatial_decomposition_count-level);
1992 DWTELEM *buf= buffer[0];
1998 memset(buffer[0], 0, sizeof(int)*width*height);
1999 for(y=0; y<height; y++){
2000 for(x=0; x<width; x++){
2001 int tab[4]={0,2,3,1};
2002 buffer[0][x+width*y]= 256*256*tab[(x&1) + 2*(y&1)];
2005 ff_spatial_dwt(buffer[0], s.temp_dwt_buffer, width, height, width, s.spatial_decomposition_type, s.spatial_decomposition_count);
2006 for(y=0; y<height; y++){
2007 for(x=0; x<width; x++){
2008 int64_t d= buffer[0][x + y*width];
2010 if(FFABS(width/2-x)<9 && FFABS(height/2-y)<9) printf("%8"PRId64" ", d);
2012 if(FFABS(height/2-y)<9) printf("\n");