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/libm.h"
23 #include "libavutil/log.h"
24 #include "libavutil/opt.h"
25 #include "libavutil/pixdesc.h"
31 #include "rangecoder.h"
34 #include "mpegvideo.h"
37 static av_cold int encode_init(AVCodecContext *avctx)
39 SnowContext *s = avctx->priv_data;
43 #if FF_API_PRIVATE_OPT
44 FF_DISABLE_DEPRECATION_WARNINGS
45 if (avctx->prediction_method)
46 s->pred = avctx->prediction_method;
47 FF_ENABLE_DEPRECATION_WARNINGS
51 && (avctx->flags & AV_CODEC_FLAG_QSCALE)
52 && avctx->global_quality == 0){
53 av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
54 return AVERROR(EINVAL);
57 s->spatial_decomposition_type= s->pred; //FIXME add decorrelator type r transform_type
59 s->mv_scale = (avctx->flags & AV_CODEC_FLAG_QPEL) ? 2 : 4;
60 s->block_max_depth= (avctx->flags & AV_CODEC_FLAG_4MV ) ? 1 : 0;
62 for(plane_index=0; plane_index<3; plane_index++){
63 s->plane[plane_index].diag_mc= 1;
64 s->plane[plane_index].htaps= 6;
65 s->plane[plane_index].hcoeff[0]= 40;
66 s->plane[plane_index].hcoeff[1]= -10;
67 s->plane[plane_index].hcoeff[2]= 2;
68 s->plane[plane_index].fast_mc= 1;
71 if ((ret = ff_snow_common_init(avctx)) < 0) {
74 ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
76 ff_snow_alloc_blocks(s);
81 s->m.bit_rate= avctx->bit_rate;
82 s->m.lmin = avctx->mb_lmin;
83 s->m.lmax = avctx->mb_lmax;
84 s->m.mb_num = (avctx->width * avctx->height + 255) / 256; // For ratecontrol
87 s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
88 s->m.me.map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
89 s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
90 s->m.sc.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
91 if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.sc.obmc_scratchpad)
92 return AVERROR(ENOMEM);
94 ff_h263_encode_init(&s->m); //mv_penalty
96 s->max_ref_frames = av_clip(avctx->refs, 1, MAX_REF_FRAMES);
98 if(avctx->flags&AV_CODEC_FLAG_PASS1){
100 avctx->stats_out = av_mallocz(256);
102 if (!avctx->stats_out)
103 return AVERROR(ENOMEM);
105 if((avctx->flags&AV_CODEC_FLAG_PASS2) || !(avctx->flags&AV_CODEC_FLAG_QSCALE)){
106 ret = ff_rate_control_init(&s->m);
110 s->pass1_rc= !(avctx->flags & (AV_CODEC_FLAG_QSCALE|AV_CODEC_FLAG_PASS2));
112 switch(avctx->pix_fmt){
113 case AV_PIX_FMT_YUV444P:
114 // case AV_PIX_FMT_YUV422P:
115 case AV_PIX_FMT_YUV420P:
116 // case AV_PIX_FMT_YUV411P:
117 case AV_PIX_FMT_YUV410P:
119 s->colorspace_type= 0;
121 case AV_PIX_FMT_GRAY8:
123 s->colorspace_type = 1;
125 /* case AV_PIX_FMT_RGB32:
129 av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
130 return AVERROR_PATCHWELCOME;
133 ret = av_pix_fmt_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift,
136 av_log(avctx, AV_LOG_ERROR, "pixel format invalid or unknown\n");
140 ff_set_cmp(&s->mecc, s->mecc.me_cmp, s->avctx->me_cmp);
141 ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, s->avctx->me_sub_cmp);
143 s->input_picture = av_frame_alloc();
144 if (!s->input_picture)
145 return AVERROR(ENOMEM);
147 if ((ret = ff_snow_get_buffer(s, s->input_picture)) < 0)
150 if(s->motion_est == FF_ME_ITER){
151 int size= s->b_width * s->b_height << 2*s->block_max_depth;
152 for(i=0; i<s->max_ref_frames; i++){
153 s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
154 s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
155 if (!s->ref_mvs[i] || !s->ref_scores[i])
156 return AVERROR(ENOMEM);
163 //near copy & paste from dsputil, FIXME
164 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
169 for (i = 0; i < h; i++) {
170 for (j = 0; j < w; j++) {
174 pix += line_size - w;
179 //near copy & paste from dsputil, FIXME
180 static int pix_norm1(uint8_t * pix, int line_size, int w)
183 const uint32_t *sq = ff_square_tab + 256;
186 for (i = 0; i < w; i++) {
187 for (j = 0; j < w; j ++) {
191 pix += line_size - w;
196 static inline int get_penalty_factor(int lambda, int lambda2, int type){
200 return lambda>>FF_LAMBDA_SHIFT;
202 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
204 return (4*lambda)>>(FF_LAMBDA_SHIFT);
206 return (2*lambda)>>(FF_LAMBDA_SHIFT);
209 return (2*lambda)>>FF_LAMBDA_SHIFT;
214 return lambda2>>FF_LAMBDA_SHIFT;
223 #define P_TOPRIGHT P[3]
224 #define P_MEDIAN P[4]
226 #define FLAG_QPEL 1 //must be 1
228 static int encode_q_branch(SnowContext *s, int level, int x, int y){
229 uint8_t p_buffer[1024];
230 uint8_t i_buffer[1024];
231 uint8_t p_state[sizeof(s->block_state)];
232 uint8_t i_state[sizeof(s->block_state)];
234 uint8_t *pbbak= s->c.bytestream;
235 uint8_t *pbbak_start= s->c.bytestream_start;
236 int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
237 const int w= s->b_width << s->block_max_depth;
238 const int h= s->b_height << s->block_max_depth;
239 const int rem_depth= s->block_max_depth - level;
240 const int index= (x + y*w) << rem_depth;
241 const int block_w= 1<<(LOG2_MB_SIZE - level);
242 int trx= (x+1)<<rem_depth;
243 int try= (y+1)<<rem_depth;
244 const BlockNode *left = x ? &s->block[index-1] : &null_block;
245 const BlockNode *top = y ? &s->block[index-w] : &null_block;
246 const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
247 const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
248 const BlockNode *tl = y && x ? &s->block[index-w-1] : left;
249 const BlockNode *tr = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
250 int pl = left->color[0];
251 int pcb= left->color[1];
252 int pcr= left->color[2];
256 const int stride= s->current_picture->linesize[0];
257 const int uvstride= s->current_picture->linesize[1];
258 uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y* stride)*block_w,
259 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
260 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
262 int16_t last_mv[3][2];
263 int qpel= !!(s->avctx->flags & AV_CODEC_FLAG_QPEL); //unused
264 const int shift= 1+qpel;
265 MotionEstContext *c= &s->m.me;
266 int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
267 int mx_context= av_log2(2*FFABS(left->mx - top->mx));
268 int my_context= av_log2(2*FFABS(left->my - top->my));
269 int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
270 int ref, best_ref, ref_score, ref_mx, ref_my;
272 av_assert0(sizeof(s->block_state) >= 256);
274 set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
278 // clip predictors / edge ?
284 P_TOPRIGHT[0]= tr->mx;
285 P_TOPRIGHT[1]= tr->my;
287 last_mv[0][0]= s->block[index].mx;
288 last_mv[0][1]= s->block[index].my;
289 last_mv[1][0]= right->mx;
290 last_mv[1][1]= right->my;
291 last_mv[2][0]= bottom->mx;
292 last_mv[2][1]= bottom->my;
299 av_assert1(c-> stride == stride);
300 av_assert1(c->uvstride == uvstride);
302 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
303 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
304 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
305 c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_DMV;
307 c->xmin = - x*block_w - 16+3;
308 c->ymin = - y*block_w - 16+3;
309 c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
310 c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
312 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
313 if(P_LEFT[1] > (c->ymax<<shift)) P_LEFT[1] = (c->ymax<<shift);
314 if(P_TOP[0] > (c->xmax<<shift)) P_TOP[0] = (c->xmax<<shift);
315 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
316 if(P_TOPRIGHT[0] < (c->xmin * (1<<shift))) P_TOPRIGHT[0]= (c->xmin * (1<<shift));
317 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
318 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
320 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
321 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
324 c->pred_x= P_LEFT[0];
325 c->pred_y= P_LEFT[1];
327 c->pred_x = P_MEDIAN[0];
328 c->pred_y = P_MEDIAN[1];
333 for(ref=0; ref<s->ref_frames; ref++){
334 init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
336 ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
337 (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
339 av_assert2(ref_mx >= c->xmin);
340 av_assert2(ref_mx <= c->xmax);
341 av_assert2(ref_my >= c->ymin);
342 av_assert2(ref_my <= c->ymax);
344 ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
345 ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
346 ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
348 s->ref_mvs[ref][index][0]= ref_mx;
349 s->ref_mvs[ref][index][1]= ref_my;
350 s->ref_scores[ref][index]= ref_score;
352 if(score > ref_score){
359 //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
362 base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
365 pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
366 memcpy(p_state, s->block_state, sizeof(s->block_state));
368 if(level!=s->block_max_depth)
369 put_rac(&pc, &p_state[4 + s_context], 1);
370 put_rac(&pc, &p_state[1 + left->type + top->type], 0);
371 if(s->ref_frames > 1)
372 put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
373 pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
374 put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
375 put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
376 p_len= pc.bytestream - pc.bytestream_start;
377 score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
379 block_s= block_w*block_w;
380 sum = pix_sum(current_data[0], stride, block_w, block_w);
381 l= (sum + block_s/2)/block_s;
382 iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
384 if (s->nb_planes > 2) {
385 block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
386 sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
387 cb= (sum + block_s/2)/block_s;
388 // iscore += pix_norm1(¤t_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
389 sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
390 cr= (sum + block_s/2)/block_s;
391 // iscore += pix_norm1(¤t_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
397 ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
398 memcpy(i_state, s->block_state, sizeof(s->block_state));
399 if(level!=s->block_max_depth)
400 put_rac(&ic, &i_state[4 + s_context], 1);
401 put_rac(&ic, &i_state[1 + left->type + top->type], 1);
402 put_symbol(&ic, &i_state[32], l-pl , 1);
403 if (s->nb_planes > 2) {
404 put_symbol(&ic, &i_state[64], cb-pcb, 1);
405 put_symbol(&ic, &i_state[96], cr-pcr, 1);
407 i_len= ic.bytestream - ic.bytestream_start;
408 iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
410 av_assert1(iscore < 255*255*256 + s->lambda2*10);
411 av_assert1(iscore >= 0);
412 av_assert1(l>=0 && l<=255);
413 av_assert1(pl>=0 && pl<=255);
416 int varc= iscore >> 8;
417 int vard= score >> 8;
418 if (vard <= 64 || vard < varc)
419 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
421 c->scene_change_score+= s->m.qscale;
424 if(level!=s->block_max_depth){
425 put_rac(&s->c, &s->block_state[4 + s_context], 0);
426 score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
427 score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
428 score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
429 score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
430 score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
432 if(score2 < score && score2 < iscore)
437 pred_mv(s, &pmx, &pmy, 0, left, top, tr);
438 memcpy(pbbak, i_buffer, i_len);
440 s->c.bytestream_start= pbbak_start;
441 s->c.bytestream= pbbak + i_len;
442 set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
443 memcpy(s->block_state, i_state, sizeof(s->block_state));
446 memcpy(pbbak, p_buffer, p_len);
448 s->c.bytestream_start= pbbak_start;
449 s->c.bytestream= pbbak + p_len;
450 set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
451 memcpy(s->block_state, p_state, sizeof(s->block_state));
456 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
457 const int w= s->b_width << s->block_max_depth;
458 const int rem_depth= s->block_max_depth - level;
459 const int index= (x + y*w) << rem_depth;
460 int trx= (x+1)<<rem_depth;
461 BlockNode *b= &s->block[index];
462 const BlockNode *left = x ? &s->block[index-1] : &null_block;
463 const BlockNode *top = y ? &s->block[index-w] : &null_block;
464 const BlockNode *tl = y && x ? &s->block[index-w-1] : left;
465 const BlockNode *tr = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
466 int pl = left->color[0];
467 int pcb= left->color[1];
468 int pcr= left->color[2];
470 int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
471 int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
472 int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
473 int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
476 set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
480 if(level!=s->block_max_depth){
481 if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
482 put_rac(&s->c, &s->block_state[4 + s_context], 1);
484 put_rac(&s->c, &s->block_state[4 + s_context], 0);
485 encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
486 encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
487 encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
488 encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
492 if(b->type & BLOCK_INTRA){
493 pred_mv(s, &pmx, &pmy, 0, left, top, tr);
494 put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
495 put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
496 if (s->nb_planes > 2) {
497 put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
498 put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
500 set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
502 pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
503 put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
504 if(s->ref_frames > 1)
505 put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
506 put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
507 put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
508 set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
512 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
514 Plane *p= &s->plane[plane_index];
515 const int block_size = MB_SIZE >> s->block_max_depth;
516 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
517 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
518 const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
519 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
520 const int ref_stride= s->current_picture->linesize[plane_index];
521 uint8_t *src= s-> input_picture->data[plane_index];
522 IDWTELEM *dst= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
523 const int b_stride = s->b_width << s->block_max_depth;
524 const int w= p->width;
525 const int h= p->height;
526 int index= mb_x + mb_y*b_stride;
527 BlockNode *b= &s->block[index];
528 BlockNode backup= *b;
532 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
534 b->type|= BLOCK_INTRA;
535 b->color[plane_index]= 0;
536 memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
539 int mb_x2= mb_x + (i &1) - 1;
540 int mb_y2= mb_y + (i>>1) - 1;
541 int x= block_w*mb_x2 + block_w/2;
542 int y= block_h*mb_y2 + block_h/2;
544 add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
545 x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
547 for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
548 for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
549 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
550 int obmc_v= obmc[index];
552 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
553 if(x<0) obmc_v += obmc[index + block_w];
554 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
555 if(x+block_w>w) obmc_v += obmc[index - block_w];
556 //FIXME precalculate this or simplify it somehow else
558 d = -dst[index] + (1<<(FRAC_BITS-1));
560 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
561 aa += obmc_v * obmc_v; //FIXME precalculate this
567 return av_clip_uint8( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa) ); //FIXME we should not need clipping
570 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
571 const int b_stride = s->b_width << s->block_max_depth;
572 const int b_height = s->b_height<< s->block_max_depth;
573 int index= x + y*b_stride;
574 const BlockNode *b = &s->block[index];
575 const BlockNode *left = x ? &s->block[index-1] : &null_block;
576 const BlockNode *top = y ? &s->block[index-b_stride] : &null_block;
577 const BlockNode *tl = y && x ? &s->block[index-b_stride-1] : left;
578 const BlockNode *tr = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
580 // int mx_context= av_log2(2*FFABS(left->mx - top->mx));
581 // int my_context= av_log2(2*FFABS(left->my - top->my));
583 if(x<0 || x>=b_stride || y>=b_height)
592 //FIXME try accurate rate
593 //FIXME intra and inter predictors if surrounding blocks are not the same type
594 if(b->type & BLOCK_INTRA){
595 return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
596 + av_log2(2*FFABS(left->color[1] - b->color[1]))
597 + av_log2(2*FFABS(left->color[2] - b->color[2])));
599 pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
602 return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
603 + av_log2(2*FFABS(dmy))
604 + av_log2(2*b->ref));
608 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
609 Plane *p= &s->plane[plane_index];
610 const int block_size = MB_SIZE >> s->block_max_depth;
611 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
612 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
613 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
614 const int ref_stride= s->current_picture->linesize[plane_index];
615 uint8_t *dst= s->current_picture->data[plane_index];
616 uint8_t *src= s-> input_picture->data[plane_index];
617 IDWTELEM *pred= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4;
618 uint8_t *cur = s->scratchbuf;
619 uint8_t *tmp = s->emu_edge_buffer;
620 const int b_stride = s->b_width << s->block_max_depth;
621 const int b_height = s->b_height<< s->block_max_depth;
622 const int w= p->width;
623 const int h= p->height;
626 const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
627 int sx= block_w*mb_x - block_w/2;
628 int sy= block_h*mb_y - block_h/2;
629 int x0= FFMAX(0,-sx);
630 int y0= FFMAX(0,-sy);
631 int x1= FFMIN(block_w*2, w-sx);
632 int y1= FFMIN(block_h*2, h-sy);
635 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
637 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);
639 for(y=y0; y<y1; y++){
640 const uint8_t *obmc1= obmc_edged[y];
641 const IDWTELEM *pred1 = pred + y*obmc_stride;
642 uint8_t *cur1 = cur + y*ref_stride;
643 uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
644 for(x=x0; x<x1; x++){
645 #if FRAC_BITS >= LOG2_OBMC_MAX
646 int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
648 int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
650 v = (v + pred1[x]) >> FRAC_BITS;
651 if(v&(~255)) v= ~(v>>31);
656 /* copy the regions where obmc[] = (uint8_t)256 */
657 if(LOG2_OBMC_MAX == 8
658 && (mb_x == 0 || mb_x == b_stride-1)
659 && (mb_y == 0 || mb_y == b_height-1)){
669 memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
673 /* FIXME rearrange dsputil to fit 32x32 cmp functions */
674 /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
675 /* FIXME cmps overlap but do not cover the wavelet's whole support.
676 * So improving the score of one block is not strictly guaranteed
677 * to improve the score of the whole frame, thus iterative motion
678 * estimation does not always converge. */
679 if(s->avctx->me_cmp == FF_CMP_W97)
680 distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
681 else if(s->avctx->me_cmp == FF_CMP_W53)
682 distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
686 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
687 distortion += s->mecc.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
691 av_assert2(block_w==8);
692 distortion = s->mecc.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
701 rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
703 if(mb_x == b_stride-2)
704 rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
706 return distortion + rate*penalty_factor;
709 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
711 Plane *p= &s->plane[plane_index];
712 const int block_size = MB_SIZE >> s->block_max_depth;
713 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
714 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
715 const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
716 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
717 const int ref_stride= s->current_picture->linesize[plane_index];
718 uint8_t *dst= s->current_picture->data[plane_index];
719 uint8_t *src= s-> input_picture->data[plane_index];
720 //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
721 // const has only been removed from zero_dst to suppress a warning
722 static IDWTELEM zero_dst[4096]; //FIXME
723 const int b_stride = s->b_width << s->block_max_depth;
724 const int w= p->width;
725 const int h= p->height;
728 const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
730 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
733 int mb_x2= mb_x + (i%3) - 1;
734 int mb_y2= mb_y + (i/3) - 1;
735 int x= block_w*mb_x2 + block_w/2;
736 int y= block_h*mb_y2 + block_h/2;
738 add_yblock(s, 0, NULL, zero_dst, dst, obmc,
739 x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
741 //FIXME find a cleaner/simpler way to skip the outside stuff
742 for(y2= y; y2<0; y2++)
743 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
744 for(y2= h; y2<y+block_h; y2++)
745 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
747 for(y2= y; y2<y+block_h; y2++)
748 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
751 for(y2= y; y2<y+block_h; y2++)
752 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
755 av_assert1(block_w== 8 || block_w==16);
756 distortion += s->mecc.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
760 BlockNode *b= &s->block[mb_x+mb_y*b_stride];
761 int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
769 rate = get_block_bits(s, mb_x, mb_y, 2);
770 for(i=merged?4:0; i<9; i++){
771 static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
772 rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
775 return distortion + rate*penalty_factor;
778 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
779 const int w= b->width;
780 const int h= b->height;
785 int *runs = s->run_buffer;
792 int /*ll=0, */l=0, lt=0, t=0, rt=0;
793 v= src[x + y*stride];
796 t= src[x + (y-1)*stride];
798 lt= src[x - 1 + (y-1)*stride];
801 rt= src[x + 1 + (y-1)*stride];
805 l= src[x - 1 + y*stride];
807 if(orientation==1) ll= src[y + (x-2)*stride];
808 else ll= src[x - 2 + y*stride];
814 if(px<b->parent->width && py<b->parent->height)
815 p= parent[px + py*2*stride];
817 if(!(/*ll|*/l|lt|t|rt|p)){
819 runs[run_index++]= run;
827 max_index= run_index;
828 runs[run_index++]= run;
830 run= runs[run_index++];
832 put_symbol2(&s->c, b->state[30], max_index, 0);
833 if(run_index <= max_index)
834 put_symbol2(&s->c, b->state[1], run, 3);
837 if(s->c.bytestream_end - s->c.bytestream < w*40){
838 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
839 return AVERROR(ENOMEM);
843 int /*ll=0, */l=0, lt=0, t=0, rt=0;
844 v= src[x + y*stride];
847 t= src[x + (y-1)*stride];
849 lt= src[x - 1 + (y-1)*stride];
852 rt= src[x + 1 + (y-1)*stride];
856 l= src[x - 1 + y*stride];
858 if(orientation==1) ll= src[y + (x-2)*stride];
859 else ll= src[x - 2 + y*stride];
865 if(px<b->parent->width && py<b->parent->height)
866 p= parent[px + py*2*stride];
868 if(/*ll|*/l|lt|t|rt|p){
869 int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
871 put_rac(&s->c, &b->state[0][context], !!v);
874 run= runs[run_index++];
876 if(run_index <= max_index)
877 put_symbol2(&s->c, b->state[1], run, 3);
885 int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
886 int l2= 2*FFABS(l) + (l<0);
887 int t2= 2*FFABS(t) + (t<0);
889 put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
890 put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
898 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
899 // encode_subband_qtree(s, b, src, parent, stride, orientation);
900 // encode_subband_z0run(s, b, src, parent, stride, orientation);
901 return encode_subband_c0run(s, b, src, parent, stride, orientation);
902 // encode_subband_dzr(s, b, src, parent, stride, orientation);
905 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){
906 const int b_stride= s->b_width << s->block_max_depth;
907 BlockNode *block= &s->block[mb_x + mb_y * b_stride];
908 BlockNode backup= *block;
912 av_assert2(mb_x>=0 && mb_y>=0);
913 av_assert2(mb_x<b_stride);
916 block->color[0] = p[0];
917 block->color[1] = p[1];
918 block->color[2] = p[2];
919 block->type |= BLOCK_INTRA;
921 index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
922 value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
923 if(s->me_cache[index] == value)
925 s->me_cache[index]= value;
929 block->type &= ~BLOCK_INTRA;
932 rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged) + s->intra_penalty * !!intra;
944 /* special case for int[2] args we discard afterwards,
945 * fixes compilation problem with gcc 2.95 */
946 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){
948 return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
951 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){
952 const int b_stride= s->b_width << s->block_max_depth;
953 BlockNode *block= &s->block[mb_x + mb_y * b_stride];
958 /* We don't initialize backup[] during variable declaration, because
959 * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
961 backup[0] = block[0];
962 backup[1] = block[1];
963 backup[2] = block[b_stride];
964 backup[3] = block[b_stride + 1];
966 av_assert2(mb_x>=0 && mb_y>=0);
967 av_assert2(mb_x<b_stride);
968 av_assert2(((mb_x|mb_y)&1) == 0);
970 index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
971 value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
972 if(s->me_cache[index] == value)
974 s->me_cache[index]= value;
979 block->type &= ~BLOCK_INTRA;
980 block[1]= block[b_stride]= block[b_stride+1]= *block;
982 rd= get_4block_rd(s, mb_x, mb_y, 0);
991 block[b_stride]= backup[2];
992 block[b_stride+1]= backup[3];
997 static void iterative_me(SnowContext *s){
998 int pass, mb_x, mb_y;
999 const int b_width = s->b_width << s->block_max_depth;
1000 const int b_height= s->b_height << s->block_max_depth;
1001 const int b_stride= b_width;
1005 RangeCoder r = s->c;
1006 uint8_t state[sizeof(s->block_state)];
1007 memcpy(state, s->block_state, sizeof(s->block_state));
1008 for(mb_y= 0; mb_y<s->b_height; mb_y++)
1009 for(mb_x= 0; mb_x<s->b_width; mb_x++)
1010 encode_q_branch(s, 0, mb_x, mb_y);
1012 memcpy(s->block_state, state, sizeof(s->block_state));
1015 for(pass=0; pass<25; pass++){
1018 for(mb_y= 0; mb_y<b_height; mb_y++){
1019 for(mb_x= 0; mb_x<b_width; mb_x++){
1020 int dia_change, i, j, ref;
1021 int best_rd= INT_MAX, ref_rd;
1022 BlockNode backup, ref_b;
1023 const int index= mb_x + mb_y * b_stride;
1024 BlockNode *block= &s->block[index];
1025 BlockNode *tb = mb_y ? &s->block[index-b_stride ] : NULL;
1026 BlockNode *lb = mb_x ? &s->block[index -1] : NULL;
1027 BlockNode *rb = mb_x+1<b_width ? &s->block[index +1] : NULL;
1028 BlockNode *bb = mb_y+1<b_height ? &s->block[index+b_stride ] : NULL;
1029 BlockNode *tlb= mb_x && mb_y ? &s->block[index-b_stride-1] : NULL;
1030 BlockNode *trb= mb_x+1<b_width && mb_y ? &s->block[index-b_stride+1] : NULL;
1031 BlockNode *blb= mb_x && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1032 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1033 const int b_w= (MB_SIZE >> s->block_max_depth);
1034 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1036 if(pass && (block->type & BLOCK_OPT))
1038 block->type |= BLOCK_OPT;
1042 if(!s->me_cache_generation)
1043 memset(s->me_cache, 0, sizeof(s->me_cache));
1044 s->me_cache_generation += 1<<22;
1046 //FIXME precalculate
1049 for (y = 0; y < b_w * 2; y++)
1050 memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1052 for(y=0; y<b_w*2; y++)
1053 memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1054 if(mb_x==b_stride-1)
1055 for(y=0; y<b_w*2; y++)
1056 memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1058 for(x=0; x<b_w*2; x++)
1059 obmc_edged[0][x] += obmc_edged[b_w-1][x];
1060 for(y=1; y<b_w; y++)
1061 memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1063 if(mb_y==b_height-1){
1064 for(x=0; x<b_w*2; x++)
1065 obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1066 for(y=b_w; y<b_w*2-1; y++)
1067 memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1071 //skip stuff outside the picture
1072 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1073 uint8_t *src= s-> input_picture->data[0];
1074 uint8_t *dst= s->current_picture->data[0];
1075 const int stride= s->current_picture->linesize[0];
1076 const int block_w= MB_SIZE >> s->block_max_depth;
1077 const int block_h= MB_SIZE >> s->block_max_depth;
1078 const int sx= block_w*mb_x - block_w/2;
1079 const int sy= block_h*mb_y - block_h/2;
1080 const int w= s->plane[0].width;
1081 const int h= s->plane[0].height;
1085 memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1086 for(y=h; y<sy+block_h*2; y++)
1087 memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1089 for(y=sy; y<sy+block_h*2; y++)
1090 memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1092 if(sx+block_w*2 > w){
1093 for(y=sy; y<sy+block_h*2; y++)
1094 memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1098 // intra(black) = neighbors' contribution to the current block
1099 for(i=0; i < s->nb_planes; i++)
1100 color[i]= get_dc(s, mb_x, mb_y, i);
1102 // get previous score (cannot be cached due to OBMC)
1103 if(pass > 0 && (block->type&BLOCK_INTRA)){
1104 int color0[3]= {block->color[0], block->color[1], block->color[2]};
1105 check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1107 check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1111 for(ref=0; ref < s->ref_frames; ref++){
1112 int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1113 if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1118 check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1119 check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1121 check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1123 check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1125 check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1127 check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1130 //FIXME avoid subpel interpolation / round to nearest integer
1132 int newx = block->mx;
1133 int newy = block->my;
1134 int dia_size = s->iterative_dia_size ? s->iterative_dia_size : FFMAX(s->avctx->dia_size, 1);
1136 for(i=0; i < dia_size; i++){
1138 dia_change |= check_block_inter(s, mb_x, mb_y, newx+4*(i-j), newy+(4*j), obmc_edged, &best_rd);
1139 dia_change |= check_block_inter(s, mb_x, mb_y, newx-4*(i-j), newy-(4*j), obmc_edged, &best_rd);
1140 dia_change |= check_block_inter(s, mb_x, mb_y, newx-(4*j), newy+4*(i-j), obmc_edged, &best_rd);
1141 dia_change |= check_block_inter(s, mb_x, mb_y, newx+(4*j), newy-4*(i-j), obmc_edged, &best_rd);
1147 static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1150 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1152 //FIXME or try the standard 2 pass qpel or similar
1154 mvr[0][0]= block->mx;
1155 mvr[0][1]= block->my;
1156 if(ref_rd > best_rd){
1163 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1164 //FIXME RD style color selection
1165 if(!same_block(block, &backup)){
1166 if(tb ) tb ->type &= ~BLOCK_OPT;
1167 if(lb ) lb ->type &= ~BLOCK_OPT;
1168 if(rb ) rb ->type &= ~BLOCK_OPT;
1169 if(bb ) bb ->type &= ~BLOCK_OPT;
1170 if(tlb) tlb->type &= ~BLOCK_OPT;
1171 if(trb) trb->type &= ~BLOCK_OPT;
1172 if(blb) blb->type &= ~BLOCK_OPT;
1173 if(brb) brb->type &= ~BLOCK_OPT;
1178 av_log(s->avctx, AV_LOG_DEBUG, "pass:%d changed:%d\n", pass, change);
1183 if(s->block_max_depth == 1){
1185 for(mb_y= 0; mb_y<b_height; mb_y+=2){
1186 for(mb_x= 0; mb_x<b_width; mb_x+=2){
1188 int best_rd, init_rd;
1189 const int index= mb_x + mb_y * b_stride;
1192 b[0]= &s->block[index];
1194 b[2]= b[0]+b_stride;
1196 if(same_block(b[0], b[1]) &&
1197 same_block(b[0], b[2]) &&
1198 same_block(b[0], b[3]))
1201 if(!s->me_cache_generation)
1202 memset(s->me_cache, 0, sizeof(s->me_cache));
1203 s->me_cache_generation += 1<<22;
1205 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1207 //FIXME more multiref search?
1208 check_4block_inter(s, mb_x, mb_y,
1209 (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1210 (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1213 if(!(b[i]->type&BLOCK_INTRA))
1214 check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1216 if(init_rd != best_rd)
1220 av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1224 static void encode_blocks(SnowContext *s, int search){
1229 if(s->motion_est == FF_ME_ITER && !s->keyframe && search)
1233 if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1234 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1238 if(s->motion_est == FF_ME_ITER || !search)
1239 encode_q_branch2(s, 0, x, y);
1241 encode_q_branch (s, 0, x, y);
1246 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1247 const int w= b->width;
1248 const int h= b->height;
1249 const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1250 const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1251 int x,y, thres1, thres2;
1253 if(s->qlog == LOSSLESS_QLOG){
1256 dst[x + y*stride]= src[x + y*stride];
1260 bias= bias ? 0 : (3*qmul)>>3;
1261 thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1267 int i= src[x + y*stride];
1269 if((unsigned)(i+thres1) > thres2){
1272 i/= qmul; //FIXME optimize
1273 dst[x + y*stride]= i;
1277 i/= qmul; //FIXME optimize
1278 dst[x + y*stride]= -i;
1281 dst[x + y*stride]= 0;
1287 int i= src[x + y*stride];
1289 if((unsigned)(i+thres1) > thres2){
1292 i= (i + bias) / qmul; //FIXME optimize
1293 dst[x + y*stride]= i;
1297 i= (i + bias) / qmul; //FIXME optimize
1298 dst[x + y*stride]= -i;
1301 dst[x + y*stride]= 0;
1307 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1308 const int w= b->width;
1309 const int h= b->height;
1310 const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1311 const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1312 const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1315 if(s->qlog == LOSSLESS_QLOG) return;
1319 int i= src[x + y*stride];
1321 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1323 src[x + y*stride]= (( i*qmul + qadd)>>(QEXPSHIFT));
1329 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1330 const int w= b->width;
1331 const int h= b->height;
1334 for(y=h-1; y>=0; y--){
1335 for(x=w-1; x>=0; x--){
1336 int i= x + y*stride;
1340 if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1341 else src[i] -= src[i - 1];
1343 if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1344 else src[i] -= src[i - 1];
1347 if(y) src[i] -= src[i - stride];
1353 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1354 const int w= b->width;
1355 const int h= b->height;
1360 int i= x + y*stride;
1364 if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1365 else src[i] += src[i - 1];
1367 if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1368 else src[i] += src[i - 1];
1371 if(y) src[i] += src[i - stride];
1377 static void encode_qlogs(SnowContext *s){
1378 int plane_index, level, orientation;
1380 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1381 for(level=0; level<s->spatial_decomposition_count; level++){
1382 for(orientation=level ? 1:0; orientation<4; orientation++){
1383 if(orientation==2) continue;
1384 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1390 static void encode_header(SnowContext *s){
1394 memset(kstate, MID_STATE, sizeof(kstate));
1396 put_rac(&s->c, kstate, s->keyframe);
1397 if(s->keyframe || s->always_reset){
1398 ff_snow_reset_contexts(s);
1399 s->last_spatial_decomposition_type=
1403 s->last_block_max_depth= 0;
1404 for(plane_index=0; plane_index<2; plane_index++){
1405 Plane *p= &s->plane[plane_index];
1408 memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1412 put_symbol(&s->c, s->header_state, s->version, 0);
1413 put_rac(&s->c, s->header_state, s->always_reset);
1414 put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1415 put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1416 put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1417 put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1418 if (s->nb_planes > 2) {
1419 put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1420 put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1422 put_rac(&s->c, s->header_state, s->spatial_scalability);
1423 // put_rac(&s->c, s->header_state, s->rate_scalability);
1424 put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1431 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1432 Plane *p= &s->plane[plane_index];
1433 update_mc |= p->last_htaps != p->htaps;
1434 update_mc |= p->last_diag_mc != p->diag_mc;
1435 update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1437 put_rac(&s->c, s->header_state, update_mc);
1439 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1440 Plane *p= &s->plane[plane_index];
1441 put_rac(&s->c, s->header_state, p->diag_mc);
1442 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1443 for(i= p->htaps/2; i; i--)
1444 put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1447 if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1448 put_rac(&s->c, s->header_state, 1);
1449 put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1452 put_rac(&s->c, s->header_state, 0);
1455 put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1456 put_symbol(&s->c, s->header_state, s->qlog - s->last_qlog , 1);
1457 put_symbol(&s->c, s->header_state, s->mv_scale - s->last_mv_scale, 1);
1458 put_symbol(&s->c, s->header_state, s->qbias - s->last_qbias , 1);
1459 put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1463 static void update_last_header_values(SnowContext *s){
1467 for(plane_index=0; plane_index<2; plane_index++){
1468 Plane *p= &s->plane[plane_index];
1469 p->last_diag_mc= p->diag_mc;
1470 p->last_htaps = p->htaps;
1471 memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1475 s->last_spatial_decomposition_type = s->spatial_decomposition_type;
1476 s->last_qlog = s->qlog;
1477 s->last_qbias = s->qbias;
1478 s->last_mv_scale = s->mv_scale;
1479 s->last_block_max_depth = s->block_max_depth;
1480 s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1483 static int qscale2qlog(int qscale){
1484 return lrint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1485 + 61*QROOT/8; ///< 64 > 60
1488 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1490 /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1491 * FIXME we know exact mv bits at this point,
1492 * but ratecontrol isn't set up to include them. */
1493 uint32_t coef_sum= 0;
1494 int level, orientation, delta_qlog;
1496 for(level=0; level<s->spatial_decomposition_count; level++){
1497 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1498 SubBand *b= &s->plane[0].band[level][orientation];
1499 IDWTELEM *buf= b->ibuf;
1500 const int w= b->width;
1501 const int h= b->height;
1502 const int stride= b->stride;
1503 const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1504 const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1505 const int qdiv= (1<<16)/qmul;
1507 //FIXME this is ugly
1510 buf[x+y*stride]= b->buf[x+y*stride];
1512 decorrelate(s, b, buf, stride, 1, 0);
1515 coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1519 /* ugly, ratecontrol just takes a sqrt again */
1520 av_assert0(coef_sum < INT_MAX);
1521 coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1523 if(pict->pict_type == AV_PICTURE_TYPE_I){
1524 s->m.current_picture.mb_var_sum= coef_sum;
1525 s->m.current_picture.mc_mb_var_sum= 0;
1527 s->m.current_picture.mc_mb_var_sum= coef_sum;
1528 s->m.current_picture.mb_var_sum= 0;
1531 pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1532 if (pict->quality < 0)
1534 s->lambda= pict->quality * 3/2;
1535 delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1536 s->qlog+= delta_qlog;
1540 static void calculate_visual_weight(SnowContext *s, Plane *p){
1541 int width = p->width;
1542 int height= p->height;
1543 int level, orientation, x, y;
1545 for(level=0; level<s->spatial_decomposition_count; level++){
1546 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1547 SubBand *b= &p->band[level][orientation];
1548 IDWTELEM *ibuf= b->ibuf;
1551 memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1552 ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1553 ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1554 for(y=0; y<height; y++){
1555 for(x=0; x<width; x++){
1556 int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1561 b->qlog= (int)(QROOT * log2(352256.0/sqrt(error)) + 0.5);
1566 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1567 const AVFrame *pict, int *got_packet)
1569 SnowContext *s = avctx->priv_data;
1570 RangeCoder * const c= &s->c;
1572 const int width= s->avctx->width;
1573 const int height= s->avctx->height;
1574 int level, orientation, plane_index, i, y, ret;
1575 uint8_t rc_header_bak[sizeof(s->header_state)];
1576 uint8_t rc_block_bak[sizeof(s->block_state)];
1578 if ((ret = ff_alloc_packet2(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + AV_INPUT_BUFFER_MIN_SIZE, 0)) < 0)
1581 ff_init_range_encoder(c, pkt->data, pkt->size);
1582 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1584 for(i=0; i < s->nb_planes; i++){
1585 int hshift= i ? s->chroma_h_shift : 0;
1586 int vshift= i ? s->chroma_v_shift : 0;
1587 for(y=0; y<AV_CEIL_RSHIFT(height, vshift); y++)
1588 memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1589 &pict->data[i][y * pict->linesize[i]],
1590 AV_CEIL_RSHIFT(width, hshift));
1591 s->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1592 AV_CEIL_RSHIFT(width, hshift), AV_CEIL_RSHIFT(height, vshift),
1593 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1594 EDGE_TOP | EDGE_BOTTOM);
1598 pic = s->input_picture;
1599 pic->pict_type = pict->pict_type;
1600 pic->quality = pict->quality;
1602 s->m.picture_number= avctx->frame_number;
1603 if(avctx->flags&AV_CODEC_FLAG_PASS2){
1604 s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1605 s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1606 if(!(avctx->flags&AV_CODEC_FLAG_QSCALE)) {
1607 pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1608 if (pic->quality < 0)
1612 s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1613 s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1616 if(s->pass1_rc && avctx->frame_number == 0)
1617 pic->quality = 2*FF_QP2LAMBDA;
1619 s->qlog = qscale2qlog(pic->quality);
1620 s->lambda = pic->quality * 3/2;
1622 if (s->qlog < 0 || (!pic->quality && (avctx->flags & AV_CODEC_FLAG_QSCALE))) {
1623 s->qlog= LOSSLESS_QLOG;
1625 }//else keep previous frame's qlog until after motion estimation
1627 if (s->current_picture->data[0]) {
1628 int w = s->avctx->width;
1629 int h = s->avctx->height;
1631 s->mpvencdsp.draw_edges(s->current_picture->data[0],
1632 s->current_picture->linesize[0], w , h ,
1633 EDGE_WIDTH , EDGE_WIDTH , EDGE_TOP | EDGE_BOTTOM);
1634 if (s->current_picture->data[2]) {
1635 s->mpvencdsp.draw_edges(s->current_picture->data[1],
1636 s->current_picture->linesize[1], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1637 EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1638 s->mpvencdsp.draw_edges(s->current_picture->data[2],
1639 s->current_picture->linesize[2], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1640 EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1645 ff_snow_frame_start(s);
1646 #if FF_API_CODED_FRAME
1647 FF_DISABLE_DEPRECATION_WARNINGS
1648 av_frame_unref(avctx->coded_frame);
1649 ret = av_frame_ref(avctx->coded_frame, s->current_picture);
1650 FF_ENABLE_DEPRECATION_WARNINGS
1655 s->m.current_picture_ptr= &s->m.current_picture;
1656 s->m.current_picture.f = s->current_picture;
1657 s->m.current_picture.f->pts = pict->pts;
1658 if(pic->pict_type == AV_PICTURE_TYPE_P){
1659 int block_width = (width +15)>>4;
1660 int block_height= (height+15)>>4;
1661 int stride= s->current_picture->linesize[0];
1663 av_assert0(s->current_picture->data[0]);
1664 av_assert0(s->last_picture[0]->data[0]);
1666 s->m.avctx= s->avctx;
1667 s->m. last_picture.f = s->last_picture[0];
1668 s->m. new_picture.f = s->input_picture;
1669 s->m. last_picture_ptr= &s->m. last_picture;
1670 s->m.linesize = stride;
1671 s->m.uvlinesize= s->current_picture->linesize[1];
1673 s->m.height= height;
1674 s->m.mb_width = block_width;
1675 s->m.mb_height= block_height;
1676 s->m.mb_stride= s->m.mb_width+1;
1677 s->m.b8_stride= 2*s->m.mb_width+1;
1679 s->m.pict_type = pic->pict_type;
1680 s->m.motion_est= s->motion_est;
1681 s->m.me.scene_change_score=0;
1682 s->m.me.dia_size = avctx->dia_size;
1683 s->m.quarter_sample= (s->avctx->flags & AV_CODEC_FLAG_QPEL)!=0;
1684 s->m.out_format= FMT_H263;
1685 s->m.unrestricted_mv= 1;
1687 s->m.lambda = s->lambda;
1688 s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1689 s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1691 s->m.mecc= s->mecc; //move
1692 s->m.qdsp= s->qdsp; //move
1693 s->m.hdsp = s->hdsp;
1695 s->hdsp = s->m.hdsp;
1700 memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1701 memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1706 s->spatial_decomposition_count= 5;
1708 while( !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1709 || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1710 s->spatial_decomposition_count--;
1712 if (s->spatial_decomposition_count <= 0) {
1713 av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1714 return AVERROR(EINVAL);
1717 s->m.pict_type = pic->pict_type;
1718 s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1720 ff_snow_common_init_after_header(avctx);
1722 if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1723 for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1724 calculate_visual_weight(s, &s->plane[plane_index]);
1729 s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1730 encode_blocks(s, 1);
1731 s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1733 for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1734 Plane *p= &s->plane[plane_index];
1738 // int bits= put_bits_count(&s->c.pb);
1740 if (!s->memc_only) {
1742 if(pict->data[plane_index]) //FIXME gray hack
1745 s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1748 predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1750 #if FF_API_PRIVATE_OPT
1751 FF_DISABLE_DEPRECATION_WARNINGS
1752 if(s->avctx->scenechange_threshold)
1753 s->scenechange_threshold = s->avctx->scenechange_threshold;
1754 FF_ENABLE_DEPRECATION_WARNINGS
1758 && pic->pict_type == AV_PICTURE_TYPE_P
1759 && !(avctx->flags&AV_CODEC_FLAG_PASS2)
1760 && s->m.me.scene_change_score > s->scenechange_threshold){
1761 ff_init_range_encoder(c, pkt->data, pkt->size);
1762 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1763 pic->pict_type= AV_PICTURE_TYPE_I;
1765 s->current_picture->key_frame=1;
1769 if(s->qlog == LOSSLESS_QLOG){
1772 s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1778 s->spatial_dwt_buffer[y*w + x]= s->spatial_idwt_buffer[y*w + x] * (1 << ENCODER_EXTRA_BITS);
1783 ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1785 if(s->pass1_rc && plane_index==0){
1786 int delta_qlog = ratecontrol_1pass(s, pic);
1787 if (delta_qlog <= INT_MIN)
1790 //reordering qlog in the bitstream would eliminate this reset
1791 ff_init_range_encoder(c, pkt->data, pkt->size);
1792 memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1793 memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1795 encode_blocks(s, 0);
1799 for(level=0; level<s->spatial_decomposition_count; level++){
1800 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1801 SubBand *b= &p->band[level][orientation];
1803 quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1805 decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1806 if (!s->no_bitstream)
1807 encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1808 av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1810 correlate(s, b, b->ibuf, b->stride, 1, 0);
1814 for(level=0; level<s->spatial_decomposition_count; level++){
1815 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1816 SubBand *b= &p->band[level][orientation];
1818 dequantize(s, b, b->ibuf, b->stride);
1822 ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1823 if(s->qlog == LOSSLESS_QLOG){
1826 s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1830 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1833 if(pic->pict_type == AV_PICTURE_TYPE_I){
1836 s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1837 pict->data[plane_index][y*pict->linesize[plane_index] + x];
1841 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1842 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1845 if(s->avctx->flags&AV_CODEC_FLAG_PSNR){
1848 if(pict->data[plane_index]) //FIXME gray hack
1851 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];
1855 s->avctx->error[plane_index] += error;
1856 s->encoding_error[plane_index] = error;
1862 update_last_header_values(s);
1864 ff_snow_release_buffer(avctx);
1866 s->current_picture->coded_picture_number = avctx->frame_number;
1867 s->current_picture->pict_type = pic->pict_type;
1868 s->current_picture->quality = pic->quality;
1869 s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1870 s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1871 s->m.current_picture.f->display_picture_number =
1872 s->m.current_picture.f->coded_picture_number = avctx->frame_number;
1873 s->m.current_picture.f->quality = pic->quality;
1874 s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1876 if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1878 if(avctx->flags&AV_CODEC_FLAG_PASS1)
1879 ff_write_pass1_stats(&s->m);
1880 s->m.last_pict_type = s->m.pict_type;
1881 #if FF_API_STAT_BITS
1882 FF_DISABLE_DEPRECATION_WARNINGS
1883 avctx->frame_bits = s->m.frame_bits;
1884 avctx->mv_bits = s->m.mv_bits;
1885 avctx->misc_bits = s->m.misc_bits;
1886 avctx->p_tex_bits = s->m.p_tex_bits;
1887 FF_ENABLE_DEPRECATION_WARNINGS
1892 ff_side_data_set_encoder_stats(pkt, s->current_picture->quality,
1894 (s->avctx->flags&AV_CODEC_FLAG_PSNR) ? 4 : 0,
1895 s->current_picture->pict_type);
1897 #if FF_API_ERROR_FRAME
1898 FF_DISABLE_DEPRECATION_WARNINGS
1899 memcpy(s->current_picture->error, s->encoding_error, sizeof(s->encoding_error));
1900 FF_ENABLE_DEPRECATION_WARNINGS
1903 pkt->size = ff_rac_terminate(c, 0);
1904 if (s->current_picture->key_frame)
1905 pkt->flags |= AV_PKT_FLAG_KEY;
1911 static av_cold int encode_end(AVCodecContext *avctx)
1913 SnowContext *s = avctx->priv_data;
1915 ff_snow_common_end(s);
1916 ff_rate_control_uninit(&s->m);
1917 av_frame_free(&s->input_picture);
1918 av_freep(&avctx->stats_out);
1923 #define OFFSET(x) offsetof(SnowContext, x)
1924 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1925 static const AVOption options[] = {
1926 {"motion_est", "motion estimation algorithm", OFFSET(motion_est), AV_OPT_TYPE_INT, {.i64 = FF_ME_EPZS }, FF_ME_ZERO, FF_ME_ITER, VE, "motion_est" },
1927 { "zero", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ZERO }, 0, 0, VE, "motion_est" },
1928 { "epzs", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_EPZS }, 0, 0, VE, "motion_est" },
1929 { "xone", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_XONE }, 0, 0, VE, "motion_est" },
1930 { "iter", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ITER }, 0, 0, VE, "motion_est" },
1931 { "memc_only", "Only do ME/MC (I frames -> ref, P frame -> ME+MC).", OFFSET(memc_only), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
1932 { "no_bitstream", "Skip final bitstream writeout.", OFFSET(no_bitstream), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
1933 { "intra_penalty", "Penalty for intra blocks in block decission", OFFSET(intra_penalty), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1934 { "iterative_dia_size", "Dia size for the iterative ME", OFFSET(iterative_dia_size), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1935 { "sc_threshold", "Scene change threshold", OFFSET(scenechange_threshold), AV_OPT_TYPE_INT, { .i64 = 0 }, INT_MIN, INT_MAX, VE },
1936 { "pred", "Spatial decomposition type", OFFSET(pred), AV_OPT_TYPE_INT, { .i64 = 0 }, DWT_97, DWT_53, VE, "pred" },
1937 { "dwt97", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, INT_MIN, INT_MAX, VE, "pred" },
1938 { "dwt53", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN, INT_MAX, VE, "pred" },
1942 static const AVClass snowenc_class = {
1943 .class_name = "snow encoder",
1944 .item_name = av_default_item_name,
1946 .version = LIBAVUTIL_VERSION_INT,
1949 AVCodec ff_snow_encoder = {
1951 .long_name = NULL_IF_CONFIG_SMALL("Snow"),
1952 .type = AVMEDIA_TYPE_VIDEO,
1953 .id = AV_CODEC_ID_SNOW,
1954 .priv_data_size = sizeof(SnowContext),
1955 .init = encode_init,
1956 .encode2 = encode_frame,
1957 .close = encode_end,
1958 .pix_fmts = (const enum AVPixelFormat[]){
1959 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1963 .priv_class = &snowenc_class,
1964 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE |
1965 FF_CODEC_CAP_INIT_CLEANUP,