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;
86 s->m.me.scratchpad= av_mallocz_array((avctx->width+64), 2*16*2*sizeof(uint8_t));
87 s->m.me.map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
88 s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
89 s->m.sc.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
90 if (!s->m.me.scratchpad || !s->m.me.map || !s->m.me.score_map || !s->m.sc.obmc_scratchpad)
91 return AVERROR(ENOMEM);
93 ff_h263_encode_init(&s->m); //mv_penalty
95 s->max_ref_frames = av_clip(avctx->refs, 1, MAX_REF_FRAMES);
97 if(avctx->flags&AV_CODEC_FLAG_PASS1){
99 avctx->stats_out = av_mallocz(256);
101 if (!avctx->stats_out)
102 return AVERROR(ENOMEM);
104 if((avctx->flags&AV_CODEC_FLAG_PASS2) || !(avctx->flags&AV_CODEC_FLAG_QSCALE)){
105 ret = ff_rate_control_init(&s->m);
109 s->pass1_rc= !(avctx->flags & (AV_CODEC_FLAG_QSCALE|AV_CODEC_FLAG_PASS2));
111 switch(avctx->pix_fmt){
112 case AV_PIX_FMT_YUV444P:
113 // case AV_PIX_FMT_YUV422P:
114 case AV_PIX_FMT_YUV420P:
115 // case AV_PIX_FMT_YUV411P:
116 case AV_PIX_FMT_YUV410P:
118 s->colorspace_type= 0;
120 case AV_PIX_FMT_GRAY8:
122 s->colorspace_type = 1;
124 /* case AV_PIX_FMT_RGB32:
128 av_log(avctx, AV_LOG_ERROR, "pixel format not supported\n");
129 return AVERROR_PATCHWELCOME;
132 ret = av_pix_fmt_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift,
135 av_log(avctx, AV_LOG_ERROR, "pixel format invalid or unknown\n");
139 ff_set_cmp(&s->mecc, s->mecc.me_cmp, s->avctx->me_cmp);
140 ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, s->avctx->me_sub_cmp);
142 s->input_picture = av_frame_alloc();
143 if (!s->input_picture)
144 return AVERROR(ENOMEM);
146 if ((ret = ff_snow_get_buffer(s, s->input_picture)) < 0)
149 if(s->motion_est == FF_ME_ITER){
150 int size= s->b_width * s->b_height << 2*s->block_max_depth;
151 for(i=0; i<s->max_ref_frames; i++){
152 s->ref_mvs[i]= av_mallocz_array(size, sizeof(int16_t[2]));
153 s->ref_scores[i]= av_mallocz_array(size, sizeof(uint32_t));
154 if (!s->ref_mvs[i] || !s->ref_scores[i])
155 return AVERROR(ENOMEM);
162 //near copy & paste from dsputil, FIXME
163 static int pix_sum(uint8_t * pix, int line_size, int w, int h)
168 for (i = 0; i < h; i++) {
169 for (j = 0; j < w; j++) {
173 pix += line_size - w;
178 //near copy & paste from dsputil, FIXME
179 static int pix_norm1(uint8_t * pix, int line_size, int w)
182 uint32_t *sq = ff_square_tab + 256;
185 for (i = 0; i < w; i++) {
186 for (j = 0; j < w; j ++) {
190 pix += line_size - w;
195 static inline int get_penalty_factor(int lambda, int lambda2, int type){
199 return lambda>>FF_LAMBDA_SHIFT;
201 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
203 return (4*lambda)>>(FF_LAMBDA_SHIFT);
205 return (2*lambda)>>(FF_LAMBDA_SHIFT);
208 return (2*lambda)>>FF_LAMBDA_SHIFT;
213 return lambda2>>FF_LAMBDA_SHIFT;
222 #define P_TOPRIGHT P[3]
223 #define P_MEDIAN P[4]
225 #define FLAG_QPEL 1 //must be 1
227 static int encode_q_branch(SnowContext *s, int level, int x, int y){
228 uint8_t p_buffer[1024];
229 uint8_t i_buffer[1024];
230 uint8_t p_state[sizeof(s->block_state)];
231 uint8_t i_state[sizeof(s->block_state)];
233 uint8_t *pbbak= s->c.bytestream;
234 uint8_t *pbbak_start= s->c.bytestream_start;
235 int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
236 const int w= s->b_width << s->block_max_depth;
237 const int h= s->b_height << s->block_max_depth;
238 const int rem_depth= s->block_max_depth - level;
239 const int index= (x + y*w) << rem_depth;
240 const int block_w= 1<<(LOG2_MB_SIZE - level);
241 int trx= (x+1)<<rem_depth;
242 int try= (y+1)<<rem_depth;
243 const BlockNode *left = x ? &s->block[index-1] : &null_block;
244 const BlockNode *top = y ? &s->block[index-w] : &null_block;
245 const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
246 const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
247 const BlockNode *tl = y && x ? &s->block[index-w-1] : left;
248 const BlockNode *tr = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
249 int pl = left->color[0];
250 int pcb= left->color[1];
251 int pcr= left->color[2];
255 const int stride= s->current_picture->linesize[0];
256 const int uvstride= s->current_picture->linesize[1];
257 uint8_t *current_data[3]= { s->input_picture->data[0] + (x + y* stride)*block_w,
258 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
259 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
261 int16_t last_mv[3][2];
262 int qpel= !!(s->avctx->flags & AV_CODEC_FLAG_QPEL); //unused
263 const int shift= 1+qpel;
264 MotionEstContext *c= &s->m.me;
265 int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
266 int mx_context= av_log2(2*FFABS(left->mx - top->mx));
267 int my_context= av_log2(2*FFABS(left->my - top->my));
268 int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
269 int ref, best_ref, ref_score, ref_mx, ref_my;
271 av_assert0(sizeof(s->block_state) >= 256);
273 set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
277 // clip predictors / edge ?
283 P_TOPRIGHT[0]= tr->mx;
284 P_TOPRIGHT[1]= tr->my;
286 last_mv[0][0]= s->block[index].mx;
287 last_mv[0][1]= s->block[index].my;
288 last_mv[1][0]= right->mx;
289 last_mv[1][1]= right->my;
290 last_mv[2][0]= bottom->mx;
291 last_mv[2][1]= bottom->my;
298 av_assert1(c-> stride == stride);
299 av_assert1(c->uvstride == uvstride);
301 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
302 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
303 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
304 c->current_mv_penalty= c->mv_penalty[s->m.f_code=1] + MAX_DMV;
306 c->xmin = - x*block_w - 16+3;
307 c->ymin = - y*block_w - 16+3;
308 c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
309 c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
311 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
312 if(P_LEFT[1] > (c->ymax<<shift)) P_LEFT[1] = (c->ymax<<shift);
313 if(P_TOP[0] > (c->xmax<<shift)) P_TOP[0] = (c->xmax<<shift);
314 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
315 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
316 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
317 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
319 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
320 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
323 c->pred_x= P_LEFT[0];
324 c->pred_y= P_LEFT[1];
326 c->pred_x = P_MEDIAN[0];
327 c->pred_y = P_MEDIAN[1];
332 for(ref=0; ref<s->ref_frames; ref++){
333 init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
335 ref_score= ff_epzs_motion_search(&s->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
336 (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
338 av_assert2(ref_mx >= c->xmin);
339 av_assert2(ref_mx <= c->xmax);
340 av_assert2(ref_my >= c->ymin);
341 av_assert2(ref_my <= c->ymax);
343 ref_score= c->sub_motion_search(&s->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
344 ref_score= ff_get_mb_score(&s->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
345 ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
347 s->ref_mvs[ref][index][0]= ref_mx;
348 s->ref_mvs[ref][index][1]= ref_my;
349 s->ref_scores[ref][index]= ref_score;
351 if(score > ref_score){
358 //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
361 base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
364 pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
365 memcpy(p_state, s->block_state, sizeof(s->block_state));
367 if(level!=s->block_max_depth)
368 put_rac(&pc, &p_state[4 + s_context], 1);
369 put_rac(&pc, &p_state[1 + left->type + top->type], 0);
370 if(s->ref_frames > 1)
371 put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
372 pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
373 put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
374 put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
375 p_len= pc.bytestream - pc.bytestream_start;
376 score += (s->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
378 block_s= block_w*block_w;
379 sum = pix_sum(current_data[0], stride, block_w, block_w);
380 l= (sum + block_s/2)/block_s;
381 iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
383 if (s->nb_planes > 2) {
384 block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
385 sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
386 cb= (sum + block_s/2)/block_s;
387 // iscore += pix_norm1(¤t_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
388 sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
389 cr= (sum + block_s/2)/block_s;
390 // iscore += pix_norm1(¤t_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
396 ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
397 memcpy(i_state, s->block_state, sizeof(s->block_state));
398 if(level!=s->block_max_depth)
399 put_rac(&ic, &i_state[4 + s_context], 1);
400 put_rac(&ic, &i_state[1 + left->type + top->type], 1);
401 put_symbol(&ic, &i_state[32], l-pl , 1);
402 if (s->nb_planes > 2) {
403 put_symbol(&ic, &i_state[64], cb-pcb, 1);
404 put_symbol(&ic, &i_state[96], cr-pcr, 1);
406 i_len= ic.bytestream - ic.bytestream_start;
407 iscore += (s->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
409 av_assert1(iscore < 255*255*256 + s->lambda2*10);
410 av_assert1(iscore >= 0);
411 av_assert1(l>=0 && l<=255);
412 av_assert1(pl>=0 && pl<=255);
415 int varc= iscore >> 8;
416 int vard= score >> 8;
417 if (vard <= 64 || vard < varc)
418 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
420 c->scene_change_score+= s->m.qscale;
423 if(level!=s->block_max_depth){
424 put_rac(&s->c, &s->block_state[4 + s_context], 0);
425 score2 = encode_q_branch(s, level+1, 2*x+0, 2*y+0);
426 score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+0);
427 score2+= encode_q_branch(s, level+1, 2*x+0, 2*y+1);
428 score2+= encode_q_branch(s, level+1, 2*x+1, 2*y+1);
429 score2+= s->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
431 if(score2 < score && score2 < iscore)
436 pred_mv(s, &pmx, &pmy, 0, left, top, tr);
437 memcpy(pbbak, i_buffer, i_len);
439 s->c.bytestream_start= pbbak_start;
440 s->c.bytestream= pbbak + i_len;
441 set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
442 memcpy(s->block_state, i_state, sizeof(s->block_state));
445 memcpy(pbbak, p_buffer, p_len);
447 s->c.bytestream_start= pbbak_start;
448 s->c.bytestream= pbbak + p_len;
449 set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
450 memcpy(s->block_state, p_state, sizeof(s->block_state));
455 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
456 const int w= s->b_width << s->block_max_depth;
457 const int rem_depth= s->block_max_depth - level;
458 const int index= (x + y*w) << rem_depth;
459 int trx= (x+1)<<rem_depth;
460 BlockNode *b= &s->block[index];
461 const BlockNode *left = x ? &s->block[index-1] : &null_block;
462 const BlockNode *top = y ? &s->block[index-w] : &null_block;
463 const BlockNode *tl = y && x ? &s->block[index-w-1] : left;
464 const BlockNode *tr = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
465 int pl = left->color[0];
466 int pcb= left->color[1];
467 int pcr= left->color[2];
469 int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
470 int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
471 int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
472 int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
475 set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
479 if(level!=s->block_max_depth){
480 if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
481 put_rac(&s->c, &s->block_state[4 + s_context], 1);
483 put_rac(&s->c, &s->block_state[4 + s_context], 0);
484 encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
485 encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
486 encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
487 encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
491 if(b->type & BLOCK_INTRA){
492 pred_mv(s, &pmx, &pmy, 0, left, top, tr);
493 put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
494 put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
495 if (s->nb_planes > 2) {
496 put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
497 put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
499 set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
501 pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
502 put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
503 if(s->ref_frames > 1)
504 put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
505 put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
506 put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
507 set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
511 static int get_dc(SnowContext *s, int mb_x, int mb_y, int plane_index){
513 Plane *p= &s->plane[plane_index];
514 const int block_size = MB_SIZE >> s->block_max_depth;
515 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
516 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
517 const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
518 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
519 const int ref_stride= s->current_picture->linesize[plane_index];
520 uint8_t *src= s-> input_picture->data[plane_index];
521 IDWTELEM *dst= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
522 const int b_stride = s->b_width << s->block_max_depth;
523 const int w= p->width;
524 const int h= p->height;
525 int index= mb_x + mb_y*b_stride;
526 BlockNode *b= &s->block[index];
527 BlockNode backup= *b;
531 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
533 b->type|= BLOCK_INTRA;
534 b->color[plane_index]= 0;
535 memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
538 int mb_x2= mb_x + (i &1) - 1;
539 int mb_y2= mb_y + (i>>1) - 1;
540 int x= block_w*mb_x2 + block_w/2;
541 int y= block_h*mb_y2 + block_h/2;
543 add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
544 x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
546 for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
547 for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
548 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
549 int obmc_v= obmc[index];
551 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
552 if(x<0) obmc_v += obmc[index + block_w];
553 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
554 if(x+block_w>w) obmc_v += obmc[index - block_w];
555 //FIXME precalculate this or simplify it somehow else
557 d = -dst[index] + (1<<(FRAC_BITS-1));
559 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
560 aa += obmc_v * obmc_v; //FIXME precalculate this
566 return av_clip_uint8( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa) ); //FIXME we should not need clipping
569 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
570 const int b_stride = s->b_width << s->block_max_depth;
571 const int b_height = s->b_height<< s->block_max_depth;
572 int index= x + y*b_stride;
573 const BlockNode *b = &s->block[index];
574 const BlockNode *left = x ? &s->block[index-1] : &null_block;
575 const BlockNode *top = y ? &s->block[index-b_stride] : &null_block;
576 const BlockNode *tl = y && x ? &s->block[index-b_stride-1] : left;
577 const BlockNode *tr = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
579 // int mx_context= av_log2(2*FFABS(left->mx - top->mx));
580 // int my_context= av_log2(2*FFABS(left->my - top->my));
582 if(x<0 || x>=b_stride || y>=b_height)
591 //FIXME try accurate rate
592 //FIXME intra and inter predictors if surrounding blocks are not the same type
593 if(b->type & BLOCK_INTRA){
594 return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
595 + av_log2(2*FFABS(left->color[1] - b->color[1]))
596 + av_log2(2*FFABS(left->color[2] - b->color[2])));
598 pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
601 return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
602 + av_log2(2*FFABS(dmy))
603 + av_log2(2*b->ref));
607 static int get_block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2]){
608 Plane *p= &s->plane[plane_index];
609 const int block_size = MB_SIZE >> s->block_max_depth;
610 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
611 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
612 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
613 const int ref_stride= s->current_picture->linesize[plane_index];
614 uint8_t *dst= s->current_picture->data[plane_index];
615 uint8_t *src= s-> input_picture->data[plane_index];
616 IDWTELEM *pred= (IDWTELEM*)s->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4;
617 uint8_t *cur = s->scratchbuf;
618 uint8_t *tmp = s->emu_edge_buffer;
619 const int b_stride = s->b_width << s->block_max_depth;
620 const int b_height = s->b_height<< s->block_max_depth;
621 const int w= p->width;
622 const int h= p->height;
625 const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
626 int sx= block_w*mb_x - block_w/2;
627 int sy= block_h*mb_y - block_h/2;
628 int x0= FFMAX(0,-sx);
629 int y0= FFMAX(0,-sy);
630 int x1= FFMIN(block_w*2, w-sx);
631 int y1= FFMIN(block_h*2, h-sy);
634 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
636 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);
638 for(y=y0; y<y1; y++){
639 const uint8_t *obmc1= obmc_edged[y];
640 const IDWTELEM *pred1 = pred + y*obmc_stride;
641 uint8_t *cur1 = cur + y*ref_stride;
642 uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
643 for(x=x0; x<x1; x++){
644 #if FRAC_BITS >= LOG2_OBMC_MAX
645 int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
647 int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
649 v = (v + pred1[x]) >> FRAC_BITS;
650 if(v&(~255)) v= ~(v>>31);
655 /* copy the regions where obmc[] = (uint8_t)256 */
656 if(LOG2_OBMC_MAX == 8
657 && (mb_x == 0 || mb_x == b_stride-1)
658 && (mb_y == 0 || mb_y == b_height-1)){
668 memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
672 /* FIXME rearrange dsputil to fit 32x32 cmp functions */
673 /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
674 /* FIXME cmps overlap but do not cover the wavelet's whole support.
675 * So improving the score of one block is not strictly guaranteed
676 * to improve the score of the whole frame, thus iterative motion
677 * estimation does not always converge. */
678 if(s->avctx->me_cmp == FF_CMP_W97)
679 distortion = ff_w97_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
680 else if(s->avctx->me_cmp == FF_CMP_W53)
681 distortion = ff_w53_32_c(&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
685 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
686 distortion += s->mecc.me_cmp[0](&s->m, src + off, dst + off, ref_stride, 16);
690 av_assert2(block_w==8);
691 distortion = s->mecc.me_cmp[0](&s->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
700 rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
702 if(mb_x == b_stride-2)
703 rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
705 return distortion + rate*penalty_factor;
708 static int get_4block_rd(SnowContext *s, int mb_x, int mb_y, int plane_index){
710 Plane *p= &s->plane[plane_index];
711 const int block_size = MB_SIZE >> s->block_max_depth;
712 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
713 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
714 const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
715 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
716 const int ref_stride= s->current_picture->linesize[plane_index];
717 uint8_t *dst= s->current_picture->data[plane_index];
718 uint8_t *src= s-> input_picture->data[plane_index];
719 //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
720 // const has only been removed from zero_dst to suppress a warning
721 static IDWTELEM zero_dst[4096]; //FIXME
722 const int b_stride = s->b_width << s->block_max_depth;
723 const int w= p->width;
724 const int h= p->height;
727 const int penalty_factor= get_penalty_factor(s->lambda, s->lambda2, s->avctx->me_cmp);
729 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
732 int mb_x2= mb_x + (i%3) - 1;
733 int mb_y2= mb_y + (i/3) - 1;
734 int x= block_w*mb_x2 + block_w/2;
735 int y= block_h*mb_y2 + block_h/2;
737 add_yblock(s, 0, NULL, zero_dst, dst, obmc,
738 x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
740 //FIXME find a cleaner/simpler way to skip the outside stuff
741 for(y2= y; y2<0; y2++)
742 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
743 for(y2= h; y2<y+block_h; y2++)
744 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
746 for(y2= y; y2<y+block_h; y2++)
747 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
750 for(y2= y; y2<y+block_h; y2++)
751 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
754 av_assert1(block_w== 8 || block_w==16);
755 distortion += s->mecc.me_cmp[block_w==8](&s->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
759 BlockNode *b= &s->block[mb_x+mb_y*b_stride];
760 int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
768 rate = get_block_bits(s, mb_x, mb_y, 2);
769 for(i=merged?4:0; i<9; i++){
770 static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
771 rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
774 return distortion + rate*penalty_factor;
777 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
778 const int w= b->width;
779 const int h= b->height;
784 int *runs = s->run_buffer;
791 int /*ll=0, */l=0, lt=0, t=0, rt=0;
792 v= src[x + y*stride];
795 t= src[x + (y-1)*stride];
797 lt= src[x - 1 + (y-1)*stride];
800 rt= src[x + 1 + (y-1)*stride];
804 l= src[x - 1 + y*stride];
806 if(orientation==1) ll= src[y + (x-2)*stride];
807 else ll= src[x - 2 + y*stride];
813 if(px<b->parent->width && py<b->parent->height)
814 p= parent[px + py*2*stride];
816 if(!(/*ll|*/l|lt|t|rt|p)){
818 runs[run_index++]= run;
826 max_index= run_index;
827 runs[run_index++]= run;
829 run= runs[run_index++];
831 put_symbol2(&s->c, b->state[30], max_index, 0);
832 if(run_index <= max_index)
833 put_symbol2(&s->c, b->state[1], run, 3);
836 if(s->c.bytestream_end - s->c.bytestream < w*40){
837 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
838 return AVERROR(ENOMEM);
842 int /*ll=0, */l=0, lt=0, t=0, rt=0;
843 v= src[x + y*stride];
846 t= src[x + (y-1)*stride];
848 lt= src[x - 1 + (y-1)*stride];
851 rt= src[x + 1 + (y-1)*stride];
855 l= src[x - 1 + y*stride];
857 if(orientation==1) ll= src[y + (x-2)*stride];
858 else ll= src[x - 2 + y*stride];
864 if(px<b->parent->width && py<b->parent->height)
865 p= parent[px + py*2*stride];
867 if(/*ll|*/l|lt|t|rt|p){
868 int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
870 put_rac(&s->c, &b->state[0][context], !!v);
873 run= runs[run_index++];
875 if(run_index <= max_index)
876 put_symbol2(&s->c, b->state[1], run, 3);
884 int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
885 int l2= 2*FFABS(l) + (l<0);
886 int t2= 2*FFABS(t) + (t<0);
888 put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
889 put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
897 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
898 // encode_subband_qtree(s, b, src, parent, stride, orientation);
899 // encode_subband_z0run(s, b, src, parent, stride, orientation);
900 return encode_subband_c0run(s, b, src, parent, stride, orientation);
901 // encode_subband_dzr(s, b, src, parent, stride, orientation);
904 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){
905 const int b_stride= s->b_width << s->block_max_depth;
906 BlockNode *block= &s->block[mb_x + mb_y * b_stride];
907 BlockNode backup= *block;
911 av_assert2(mb_x>=0 && mb_y>=0);
912 av_assert2(mb_x<b_stride);
915 block->color[0] = p[0];
916 block->color[1] = p[1];
917 block->color[2] = p[2];
918 block->type |= BLOCK_INTRA;
920 index= (p[0] + 31*p[1]) & (ME_CACHE_SIZE-1);
921 value= s->me_cache_generation + (p[0]>>10) + (p[1]<<6) + (block->ref<<12);
922 if(s->me_cache[index] == value)
924 s->me_cache[index]= value;
928 block->type &= ~BLOCK_INTRA;
931 rd= get_block_rd(s, mb_x, mb_y, 0, obmc_edged) + s->intra_penalty * !!intra;
943 /* special case for int[2] args we discard afterwards,
944 * fixes compilation problem with gcc 2.95 */
945 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){
947 return check_block(s, mb_x, mb_y, p, 0, obmc_edged, best_rd);
950 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){
951 const int b_stride= s->b_width << s->block_max_depth;
952 BlockNode *block= &s->block[mb_x + mb_y * b_stride];
957 /* We don't initialize backup[] during variable declaration, because
958 * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
960 backup[0] = block[0];
961 backup[1] = block[1];
962 backup[2] = block[b_stride];
963 backup[3] = block[b_stride + 1];
965 av_assert2(mb_x>=0 && mb_y>=0);
966 av_assert2(mb_x<b_stride);
967 av_assert2(((mb_x|mb_y)&1) == 0);
969 index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
970 value= s->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
971 if(s->me_cache[index] == value)
973 s->me_cache[index]= value;
978 block->type &= ~BLOCK_INTRA;
979 block[1]= block[b_stride]= block[b_stride+1]= *block;
981 rd= get_4block_rd(s, mb_x, mb_y, 0);
990 block[b_stride]= backup[2];
991 block[b_stride+1]= backup[3];
996 static void iterative_me(SnowContext *s){
997 int pass, mb_x, mb_y;
998 const int b_width = s->b_width << s->block_max_depth;
999 const int b_height= s->b_height << s->block_max_depth;
1000 const int b_stride= b_width;
1004 RangeCoder r = s->c;
1005 uint8_t state[sizeof(s->block_state)];
1006 memcpy(state, s->block_state, sizeof(s->block_state));
1007 for(mb_y= 0; mb_y<s->b_height; mb_y++)
1008 for(mb_x= 0; mb_x<s->b_width; mb_x++)
1009 encode_q_branch(s, 0, mb_x, mb_y);
1011 memcpy(s->block_state, state, sizeof(s->block_state));
1014 for(pass=0; pass<25; pass++){
1017 for(mb_y= 0; mb_y<b_height; mb_y++){
1018 for(mb_x= 0; mb_x<b_width; mb_x++){
1019 int dia_change, i, j, ref;
1020 int best_rd= INT_MAX, ref_rd;
1021 BlockNode backup, ref_b;
1022 const int index= mb_x + mb_y * b_stride;
1023 BlockNode *block= &s->block[index];
1024 BlockNode *tb = mb_y ? &s->block[index-b_stride ] : NULL;
1025 BlockNode *lb = mb_x ? &s->block[index -1] : NULL;
1026 BlockNode *rb = mb_x+1<b_width ? &s->block[index +1] : NULL;
1027 BlockNode *bb = mb_y+1<b_height ? &s->block[index+b_stride ] : NULL;
1028 BlockNode *tlb= mb_x && mb_y ? &s->block[index-b_stride-1] : NULL;
1029 BlockNode *trb= mb_x+1<b_width && mb_y ? &s->block[index-b_stride+1] : NULL;
1030 BlockNode *blb= mb_x && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1031 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1032 const int b_w= (MB_SIZE >> s->block_max_depth);
1033 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1035 if(pass && (block->type & BLOCK_OPT))
1037 block->type |= BLOCK_OPT;
1041 if(!s->me_cache_generation)
1042 memset(s->me_cache, 0, sizeof(s->me_cache));
1043 s->me_cache_generation += 1<<22;
1045 //FIXME precalculate
1048 for (y = 0; y < b_w * 2; y++)
1049 memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1051 for(y=0; y<b_w*2; y++)
1052 memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1053 if(mb_x==b_stride-1)
1054 for(y=0; y<b_w*2; y++)
1055 memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1057 for(x=0; x<b_w*2; x++)
1058 obmc_edged[0][x] += obmc_edged[b_w-1][x];
1059 for(y=1; y<b_w; y++)
1060 memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1062 if(mb_y==b_height-1){
1063 for(x=0; x<b_w*2; x++)
1064 obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1065 for(y=b_w; y<b_w*2-1; y++)
1066 memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1070 //skip stuff outside the picture
1071 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1072 uint8_t *src= s-> input_picture->data[0];
1073 uint8_t *dst= s->current_picture->data[0];
1074 const int stride= s->current_picture->linesize[0];
1075 const int block_w= MB_SIZE >> s->block_max_depth;
1076 const int block_h= MB_SIZE >> s->block_max_depth;
1077 const int sx= block_w*mb_x - block_w/2;
1078 const int sy= block_h*mb_y - block_h/2;
1079 const int w= s->plane[0].width;
1080 const int h= s->plane[0].height;
1084 memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1085 for(y=h; y<sy+block_h*2; y++)
1086 memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1088 for(y=sy; y<sy+block_h*2; y++)
1089 memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1091 if(sx+block_w*2 > w){
1092 for(y=sy; y<sy+block_h*2; y++)
1093 memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1097 // intra(black) = neighbors' contribution to the current block
1098 for(i=0; i < s->nb_planes; i++)
1099 color[i]= get_dc(s, mb_x, mb_y, i);
1101 // get previous score (cannot be cached due to OBMC)
1102 if(pass > 0 && (block->type&BLOCK_INTRA)){
1103 int color0[3]= {block->color[0], block->color[1], block->color[2]};
1104 check_block(s, mb_x, mb_y, color0, 1, obmc_edged, &best_rd);
1106 check_block_inter(s, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1110 for(ref=0; ref < s->ref_frames; ref++){
1111 int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1112 if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1117 check_block_inter(s, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1118 check_block_inter(s, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1120 check_block_inter(s, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1122 check_block_inter(s, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1124 check_block_inter(s, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1126 check_block_inter(s, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1129 //FIXME avoid subpel interpolation / round to nearest integer
1131 int newx = block->mx;
1132 int newy = block->my;
1133 int dia_size = s->iterative_dia_size ? s->iterative_dia_size : FFMAX(s->avctx->dia_size, 1);
1135 for(i=0; i < dia_size; i++){
1137 dia_change |= check_block_inter(s, mb_x, mb_y, newx+4*(i-j), newy+(4*j), obmc_edged, &best_rd);
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*j), newy+4*(i-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);
1146 static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1149 dia_change |= check_block_inter(s, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1151 //FIXME or try the standard 2 pass qpel or similar
1153 mvr[0][0]= block->mx;
1154 mvr[0][1]= block->my;
1155 if(ref_rd > best_rd){
1162 check_block(s, mb_x, mb_y, color, 1, obmc_edged, &best_rd);
1163 //FIXME RD style color selection
1164 if(!same_block(block, &backup)){
1165 if(tb ) tb ->type &= ~BLOCK_OPT;
1166 if(lb ) lb ->type &= ~BLOCK_OPT;
1167 if(rb ) rb ->type &= ~BLOCK_OPT;
1168 if(bb ) bb ->type &= ~BLOCK_OPT;
1169 if(tlb) tlb->type &= ~BLOCK_OPT;
1170 if(trb) trb->type &= ~BLOCK_OPT;
1171 if(blb) blb->type &= ~BLOCK_OPT;
1172 if(brb) brb->type &= ~BLOCK_OPT;
1177 av_log(s->avctx, AV_LOG_DEBUG, "pass:%d changed:%d\n", pass, change);
1182 if(s->block_max_depth == 1){
1184 for(mb_y= 0; mb_y<b_height; mb_y+=2){
1185 for(mb_x= 0; mb_x<b_width; mb_x+=2){
1187 int best_rd, init_rd;
1188 const int index= mb_x + mb_y * b_stride;
1191 b[0]= &s->block[index];
1193 b[2]= b[0]+b_stride;
1195 if(same_block(b[0], b[1]) &&
1196 same_block(b[0], b[2]) &&
1197 same_block(b[0], b[3]))
1200 if(!s->me_cache_generation)
1201 memset(s->me_cache, 0, sizeof(s->me_cache));
1202 s->me_cache_generation += 1<<22;
1204 init_rd= best_rd= get_4block_rd(s, mb_x, mb_y, 0);
1206 //FIXME more multiref search?
1207 check_4block_inter(s, mb_x, mb_y,
1208 (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1209 (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1212 if(!(b[i]->type&BLOCK_INTRA))
1213 check_4block_inter(s, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1215 if(init_rd != best_rd)
1219 av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1223 static void encode_blocks(SnowContext *s, int search){
1228 if(s->motion_est == FF_ME_ITER && !s->keyframe && search)
1232 if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1233 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1237 if(s->motion_est == FF_ME_ITER || !search)
1238 encode_q_branch2(s, 0, x, y);
1240 encode_q_branch (s, 0, x, y);
1245 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1246 const int w= b->width;
1247 const int h= b->height;
1248 const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1249 const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1250 int x,y, thres1, thres2;
1252 if(s->qlog == LOSSLESS_QLOG){
1255 dst[x + y*stride]= src[x + y*stride];
1259 bias= bias ? 0 : (3*qmul)>>3;
1260 thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1266 int i= src[x + y*stride];
1268 if((unsigned)(i+thres1) > thres2){
1271 i/= qmul; //FIXME optimize
1272 dst[x + y*stride]= i;
1276 i/= qmul; //FIXME optimize
1277 dst[x + y*stride]= -i;
1280 dst[x + y*stride]= 0;
1286 int i= src[x + y*stride];
1288 if((unsigned)(i+thres1) > thres2){
1291 i= (i + bias) / qmul; //FIXME optimize
1292 dst[x + y*stride]= i;
1296 i= (i + bias) / qmul; //FIXME optimize
1297 dst[x + y*stride]= -i;
1300 dst[x + y*stride]= 0;
1306 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1307 const int w= b->width;
1308 const int h= b->height;
1309 const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1310 const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1311 const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1314 if(s->qlog == LOSSLESS_QLOG) return;
1318 int i= src[x + y*stride];
1320 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1322 src[x + y*stride]= (( i*qmul + qadd)>>(QEXPSHIFT));
1328 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1329 const int w= b->width;
1330 const int h= b->height;
1333 for(y=h-1; y>=0; y--){
1334 for(x=w-1; x>=0; x--){
1335 int i= x + y*stride;
1339 if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1340 else src[i] -= src[i - 1];
1342 if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1343 else src[i] -= src[i - 1];
1346 if(y) src[i] -= src[i - stride];
1352 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1353 const int w= b->width;
1354 const int h= b->height;
1359 int i= x + y*stride;
1363 if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1364 else src[i] += src[i - 1];
1366 if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1367 else src[i] += src[i - 1];
1370 if(y) src[i] += src[i - stride];
1376 static void encode_qlogs(SnowContext *s){
1377 int plane_index, level, orientation;
1379 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1380 for(level=0; level<s->spatial_decomposition_count; level++){
1381 for(orientation=level ? 1:0; orientation<4; orientation++){
1382 if(orientation==2) continue;
1383 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1389 static void encode_header(SnowContext *s){
1393 memset(kstate, MID_STATE, sizeof(kstate));
1395 put_rac(&s->c, kstate, s->keyframe);
1396 if(s->keyframe || s->always_reset){
1397 ff_snow_reset_contexts(s);
1398 s->last_spatial_decomposition_type=
1402 s->last_block_max_depth= 0;
1403 for(plane_index=0; plane_index<2; plane_index++){
1404 Plane *p= &s->plane[plane_index];
1407 memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1411 put_symbol(&s->c, s->header_state, s->version, 0);
1412 put_rac(&s->c, s->header_state, s->always_reset);
1413 put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1414 put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1415 put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1416 put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1417 if (s->nb_planes > 2) {
1418 put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1419 put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1421 put_rac(&s->c, s->header_state, s->spatial_scalability);
1422 // put_rac(&s->c, s->header_state, s->rate_scalability);
1423 put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1430 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1431 Plane *p= &s->plane[plane_index];
1432 update_mc |= p->last_htaps != p->htaps;
1433 update_mc |= p->last_diag_mc != p->diag_mc;
1434 update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1436 put_rac(&s->c, s->header_state, update_mc);
1438 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1439 Plane *p= &s->plane[plane_index];
1440 put_rac(&s->c, s->header_state, p->diag_mc);
1441 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1442 for(i= p->htaps/2; i; i--)
1443 put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1446 if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1447 put_rac(&s->c, s->header_state, 1);
1448 put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1451 put_rac(&s->c, s->header_state, 0);
1454 put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1455 put_symbol(&s->c, s->header_state, s->qlog - s->last_qlog , 1);
1456 put_symbol(&s->c, s->header_state, s->mv_scale - s->last_mv_scale, 1);
1457 put_symbol(&s->c, s->header_state, s->qbias - s->last_qbias , 1);
1458 put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1462 static void update_last_header_values(SnowContext *s){
1466 for(plane_index=0; plane_index<2; plane_index++){
1467 Plane *p= &s->plane[plane_index];
1468 p->last_diag_mc= p->diag_mc;
1469 p->last_htaps = p->htaps;
1470 memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1474 s->last_spatial_decomposition_type = s->spatial_decomposition_type;
1475 s->last_qlog = s->qlog;
1476 s->last_qbias = s->qbias;
1477 s->last_mv_scale = s->mv_scale;
1478 s->last_block_max_depth = s->block_max_depth;
1479 s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1482 static int qscale2qlog(int qscale){
1483 return lrint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1484 + 61*QROOT/8; ///< 64 > 60
1487 static int ratecontrol_1pass(SnowContext *s, AVFrame *pict)
1489 /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1490 * FIXME we know exact mv bits at this point,
1491 * but ratecontrol isn't set up to include them. */
1492 uint32_t coef_sum= 0;
1493 int level, orientation, delta_qlog;
1495 for(level=0; level<s->spatial_decomposition_count; level++){
1496 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1497 SubBand *b= &s->plane[0].band[level][orientation];
1498 IDWTELEM *buf= b->ibuf;
1499 const int w= b->width;
1500 const int h= b->height;
1501 const int stride= b->stride;
1502 const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1503 const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1504 const int qdiv= (1<<16)/qmul;
1506 //FIXME this is ugly
1509 buf[x+y*stride]= b->buf[x+y*stride];
1511 decorrelate(s, b, buf, stride, 1, 0);
1514 coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1518 /* ugly, ratecontrol just takes a sqrt again */
1519 av_assert0(coef_sum < INT_MAX);
1520 coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1522 if(pict->pict_type == AV_PICTURE_TYPE_I){
1523 s->m.current_picture.mb_var_sum= coef_sum;
1524 s->m.current_picture.mc_mb_var_sum= 0;
1526 s->m.current_picture.mc_mb_var_sum= coef_sum;
1527 s->m.current_picture.mb_var_sum= 0;
1530 pict->quality= ff_rate_estimate_qscale(&s->m, 1);
1531 if (pict->quality < 0)
1533 s->lambda= pict->quality * 3/2;
1534 delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1535 s->qlog+= delta_qlog;
1539 static void calculate_visual_weight(SnowContext *s, Plane *p){
1540 int width = p->width;
1541 int height= p->height;
1542 int level, orientation, x, y;
1544 for(level=0; level<s->spatial_decomposition_count; level++){
1545 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1546 SubBand *b= &p->band[level][orientation];
1547 IDWTELEM *ibuf= b->ibuf;
1550 memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1551 ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1552 ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1553 for(y=0; y<height; y++){
1554 for(x=0; x<width; x++){
1555 int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1560 b->qlog= (int)(QROOT * log2(352256.0/sqrt(error)) + 0.5);
1565 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1566 const AVFrame *pict, int *got_packet)
1568 SnowContext *s = avctx->priv_data;
1569 RangeCoder * const c= &s->c;
1571 const int width= s->avctx->width;
1572 const int height= s->avctx->height;
1573 int level, orientation, plane_index, i, y, ret;
1574 uint8_t rc_header_bak[sizeof(s->header_state)];
1575 uint8_t rc_block_bak[sizeof(s->block_state)];
1577 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)
1580 ff_init_range_encoder(c, pkt->data, pkt->size);
1581 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1583 for(i=0; i < s->nb_planes; i++){
1584 int hshift= i ? s->chroma_h_shift : 0;
1585 int vshift= i ? s->chroma_v_shift : 0;
1586 for(y=0; y<AV_CEIL_RSHIFT(height, vshift); y++)
1587 memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1588 &pict->data[i][y * pict->linesize[i]],
1589 AV_CEIL_RSHIFT(width, hshift));
1590 s->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1591 AV_CEIL_RSHIFT(width, hshift), AV_CEIL_RSHIFT(height, vshift),
1592 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1593 EDGE_TOP | EDGE_BOTTOM);
1597 pic = s->input_picture;
1598 pic->pict_type = pict->pict_type;
1599 pic->quality = pict->quality;
1601 s->m.picture_number= avctx->frame_number;
1602 if(avctx->flags&AV_CODEC_FLAG_PASS2){
1603 s->m.pict_type = pic->pict_type = s->m.rc_context.entry[avctx->frame_number].new_pict_type;
1604 s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1605 if(!(avctx->flags&AV_CODEC_FLAG_QSCALE)) {
1606 pic->quality = ff_rate_estimate_qscale(&s->m, 0);
1607 if (pic->quality < 0)
1611 s->keyframe= avctx->gop_size==0 || avctx->frame_number % avctx->gop_size == 0;
1612 s->m.pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1615 if(s->pass1_rc && avctx->frame_number == 0)
1616 pic->quality = 2*FF_QP2LAMBDA;
1618 s->qlog = qscale2qlog(pic->quality);
1619 s->lambda = pic->quality * 3/2;
1621 if (s->qlog < 0 || (!pic->quality && (avctx->flags & AV_CODEC_FLAG_QSCALE))) {
1622 s->qlog= LOSSLESS_QLOG;
1624 }//else keep previous frame's qlog until after motion estimation
1626 if (s->current_picture->data[0]) {
1627 int w = s->avctx->width;
1628 int h = s->avctx->height;
1630 s->mpvencdsp.draw_edges(s->current_picture->data[0],
1631 s->current_picture->linesize[0], w , h ,
1632 EDGE_WIDTH , EDGE_WIDTH , EDGE_TOP | EDGE_BOTTOM);
1633 if (s->current_picture->data[2]) {
1634 s->mpvencdsp.draw_edges(s->current_picture->data[1],
1635 s->current_picture->linesize[1], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1636 EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1637 s->mpvencdsp.draw_edges(s->current_picture->data[2],
1638 s->current_picture->linesize[2], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1639 EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1644 ff_snow_frame_start(s);
1645 #if FF_API_CODED_FRAME
1646 FF_DISABLE_DEPRECATION_WARNINGS
1647 av_frame_unref(avctx->coded_frame);
1648 ret = av_frame_ref(avctx->coded_frame, s->current_picture);
1649 FF_ENABLE_DEPRECATION_WARNINGS
1654 s->m.current_picture_ptr= &s->m.current_picture;
1655 s->m.current_picture.f = s->current_picture;
1656 s->m.current_picture.f->pts = pict->pts;
1657 if(pic->pict_type == AV_PICTURE_TYPE_P){
1658 int block_width = (width +15)>>4;
1659 int block_height= (height+15)>>4;
1660 int stride= s->current_picture->linesize[0];
1662 av_assert0(s->current_picture->data[0]);
1663 av_assert0(s->last_picture[0]->data[0]);
1665 s->m.avctx= s->avctx;
1666 s->m. last_picture.f = s->last_picture[0];
1667 s->m. new_picture.f = s->input_picture;
1668 s->m. last_picture_ptr= &s->m. last_picture;
1669 s->m.linesize = stride;
1670 s->m.uvlinesize= s->current_picture->linesize[1];
1672 s->m.height= height;
1673 s->m.mb_width = block_width;
1674 s->m.mb_height= block_height;
1675 s->m.mb_stride= s->m.mb_width+1;
1676 s->m.b8_stride= 2*s->m.mb_width+1;
1678 s->m.pict_type = pic->pict_type;
1679 s->m.motion_est= s->motion_est;
1680 s->m.me.scene_change_score=0;
1681 s->m.me.dia_size = avctx->dia_size;
1682 s->m.quarter_sample= (s->avctx->flags & AV_CODEC_FLAG_QPEL)!=0;
1683 s->m.out_format= FMT_H263;
1684 s->m.unrestricted_mv= 1;
1686 s->m.lambda = s->lambda;
1687 s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1688 s->lambda2= s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1690 s->m.mecc= s->mecc; //move
1691 s->m.qdsp= s->qdsp; //move
1692 s->m.hdsp = s->hdsp;
1694 s->hdsp = s->m.hdsp;
1699 memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1700 memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1705 s->spatial_decomposition_count= 5;
1707 while( !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1708 || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1709 s->spatial_decomposition_count--;
1711 if (s->spatial_decomposition_count <= 0) {
1712 av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1713 return AVERROR(EINVAL);
1716 s->m.pict_type = pic->pict_type;
1717 s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1719 ff_snow_common_init_after_header(avctx);
1721 if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1722 for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1723 calculate_visual_weight(s, &s->plane[plane_index]);
1728 s->m.misc_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1729 encode_blocks(s, 1);
1730 s->m.mv_bits = 8*(s->c.bytestream - s->c.bytestream_start) - s->m.misc_bits;
1732 for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1733 Plane *p= &s->plane[plane_index];
1737 // int bits= put_bits_count(&s->c.pb);
1739 if (!s->memc_only) {
1741 if(pict->data[plane_index]) //FIXME gray hack
1744 s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1747 predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1749 #if FF_API_PRIVATE_OPT
1750 FF_DISABLE_DEPRECATION_WARNINGS
1751 if(s->avctx->scenechange_threshold)
1752 s->scenechange_threshold = s->avctx->scenechange_threshold;
1753 FF_ENABLE_DEPRECATION_WARNINGS
1757 && pic->pict_type == AV_PICTURE_TYPE_P
1758 && !(avctx->flags&AV_CODEC_FLAG_PASS2)
1759 && s->m.me.scene_change_score > s->scenechange_threshold){
1760 ff_init_range_encoder(c, pkt->data, pkt->size);
1761 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1762 pic->pict_type= AV_PICTURE_TYPE_I;
1764 s->current_picture->key_frame=1;
1768 if(s->qlog == LOSSLESS_QLOG){
1771 s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1777 s->spatial_dwt_buffer[y*w + x]=s->spatial_idwt_buffer[y*w + x]<<ENCODER_EXTRA_BITS;
1782 ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1784 if(s->pass1_rc && plane_index==0){
1785 int delta_qlog = ratecontrol_1pass(s, pic);
1786 if (delta_qlog <= INT_MIN)
1789 //reordering qlog in the bitstream would eliminate this reset
1790 ff_init_range_encoder(c, pkt->data, pkt->size);
1791 memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1792 memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1794 encode_blocks(s, 0);
1798 for(level=0; level<s->spatial_decomposition_count; level++){
1799 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1800 SubBand *b= &p->band[level][orientation];
1802 quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1804 decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1805 if (!s->no_bitstream)
1806 encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1807 av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1809 correlate(s, b, b->ibuf, b->stride, 1, 0);
1813 for(level=0; level<s->spatial_decomposition_count; level++){
1814 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1815 SubBand *b= &p->band[level][orientation];
1817 dequantize(s, b, b->ibuf, b->stride);
1821 ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1822 if(s->qlog == LOSSLESS_QLOG){
1825 s->spatial_idwt_buffer[y*w + x]<<=FRAC_BITS;
1829 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1832 if(pic->pict_type == AV_PICTURE_TYPE_I){
1835 s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
1836 pict->data[plane_index][y*pict->linesize[plane_index] + x];
1840 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
1841 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
1844 if(s->avctx->flags&AV_CODEC_FLAG_PSNR){
1847 if(pict->data[plane_index]) //FIXME gray hack
1850 int d= s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x] - pict->data[plane_index][y*pict->linesize[plane_index] + x];
1854 s->avctx->error[plane_index] += error;
1855 s->encoding_error[plane_index] = error;
1861 update_last_header_values(s);
1863 ff_snow_release_buffer(avctx);
1865 s->current_picture->coded_picture_number = avctx->frame_number;
1866 s->current_picture->pict_type = pic->pict_type;
1867 s->current_picture->quality = pic->quality;
1868 s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);
1869 s->m.p_tex_bits = s->m.frame_bits - s->m.misc_bits - s->m.mv_bits;
1870 s->m.current_picture.f->display_picture_number =
1871 s->m.current_picture.f->coded_picture_number = avctx->frame_number;
1872 s->m.current_picture.f->quality = pic->quality;
1873 s->m.total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
1875 if (ff_rate_estimate_qscale(&s->m, 0) < 0)
1877 if(avctx->flags&AV_CODEC_FLAG_PASS1)
1878 ff_write_pass1_stats(&s->m);
1879 s->m.last_pict_type = s->m.pict_type;
1880 #if FF_API_STAT_BITS
1881 FF_DISABLE_DEPRECATION_WARNINGS
1882 avctx->frame_bits = s->m.frame_bits;
1883 avctx->mv_bits = s->m.mv_bits;
1884 avctx->misc_bits = s->m.misc_bits;
1885 avctx->p_tex_bits = s->m.p_tex_bits;
1886 FF_ENABLE_DEPRECATION_WARNINGS
1891 ff_side_data_set_encoder_stats(pkt, s->current_picture->quality,
1893 (s->avctx->flags&AV_CODEC_FLAG_PSNR) ? 4 : 0,
1894 s->current_picture->pict_type);
1896 #if FF_API_ERROR_FRAME
1897 FF_DISABLE_DEPRECATION_WARNINGS
1898 memcpy(s->current_picture->error, s->encoding_error, sizeof(s->encoding_error));
1899 FF_ENABLE_DEPRECATION_WARNINGS
1902 pkt->size = ff_rac_terminate(c);
1903 if (s->current_picture->key_frame)
1904 pkt->flags |= AV_PKT_FLAG_KEY;
1910 static av_cold int encode_end(AVCodecContext *avctx)
1912 SnowContext *s = avctx->priv_data;
1914 ff_snow_common_end(s);
1915 ff_rate_control_uninit(&s->m);
1916 av_frame_free(&s->input_picture);
1917 av_freep(&avctx->stats_out);
1922 #define OFFSET(x) offsetof(SnowContext, x)
1923 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1924 static const AVOption options[] = {
1925 {"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" },
1926 { "zero", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ZERO }, 0, 0, VE, "motion_est" },
1927 { "epzs", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_EPZS }, 0, 0, VE, "motion_est" },
1928 { "xone", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_XONE }, 0, 0, VE, "motion_est" },
1929 { "iter", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ITER }, 0, 0, VE, "motion_est" },
1930 { "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 },
1931 { "no_bitstream", "Skip final bitstream writeout.", OFFSET(no_bitstream), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
1932 { "intra_penalty", "Penalty for intra blocks in block decission", OFFSET(intra_penalty), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1933 { "iterative_dia_size", "Dia size for the iterative ME", OFFSET(iterative_dia_size), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1934 { "sc_threshold", "Scene change threshold", OFFSET(scenechange_threshold), AV_OPT_TYPE_INT, { .i64 = 0 }, INT_MIN, INT_MAX, VE },
1935 { "pred", "Spatial decomposition type", OFFSET(pred), AV_OPT_TYPE_INT, { .i64 = 0 }, DWT_97, DWT_53, VE, "pred" },
1936 { "dwt97", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, INT_MIN, INT_MAX, VE, "pred" },
1937 { "dwt53", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN, INT_MAX, VE, "pred" },
1941 static const AVClass snowenc_class = {
1942 .class_name = "snow encoder",
1943 .item_name = av_default_item_name,
1945 .version = LIBAVUTIL_VERSION_INT,
1948 AVCodec ff_snow_encoder = {
1950 .long_name = NULL_IF_CONFIG_SMALL("Snow"),
1951 .type = AVMEDIA_TYPE_VIDEO,
1952 .id = AV_CODEC_ID_SNOW,
1953 .priv_data_size = sizeof(SnowContext),
1954 .init = encode_init,
1955 .encode2 = encode_frame,
1956 .close = encode_end,
1957 .pix_fmts = (const enum AVPixelFormat[]){
1958 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
1962 .priv_class = &snowenc_class,
1963 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE |
1964 FF_CODEC_CAP_INIT_CLEANUP,