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/emms.h"
22 #include "libavutil/intmath.h"
23 #include "libavutil/libm.h"
24 #include "libavutil/log.h"
25 #include "libavutil/opt.h"
26 #include "libavutil/pixdesc.h"
28 #include "codec_internal.h"
30 #include "internal.h" //For AVCodecInternal.recon_frame
32 #include "packet_internal.h"
37 #include "rangecoder.h"
40 #include "mpegvideo.h"
45 typedef struct SnowEncContext {
48 MpegvideoEncDSPContext mpvencdsp;
59 int iterative_dia_size;
60 int scenechange_threshold;
63 MpegEncContext m; // needed for motion estimation, should not be used for anything else, the idea is to eventually make the motion estimation independent of MpegEncContext, so this will be removed then (FIXME/XXX)
64 #define ME_CACHE_SIZE 1024
65 unsigned me_cache[ME_CACHE_SIZE];
66 unsigned me_cache_generation;
68 uint64_t encoding_error[SNOW_MAX_PLANES];
71 static void init_ref(MotionEstContext *c, const uint8_t *const src[3],
72 uint8_t *const ref[3], uint8_t *const ref2[3],
73 int x, int y, int ref_index)
75 SnowContext *s = c->avctx->priv_data;
76 const int offset[3] = {
78 ((y*c->uvstride + x) >> s->chroma_h_shift),
79 ((y*c->uvstride + x) >> s->chroma_h_shift),
81 for (int i = 0; i < 3; i++) {
82 c->src[0][i] = src [i];
83 c->ref[0][i] = ref [i] + offset[i];
85 av_assert2(!ref_index);
88 static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed)
91 const int a = FFABS(v);
92 const int e = av_log2(a);
93 const int el = FFMIN(e, 10);
96 put_rac(c, state + 0, 0);
98 for (i = 0; i < el; i++)
99 put_rac(c, state + 1 + i, 1); //1..10
101 put_rac(c, state + 1 + 9, 1); //1..10
102 put_rac(c, state + 1 + FFMIN(i, 9), 0);
104 for (i = e - 1; i >= el; i--)
105 put_rac(c, state + 22 + 9, (a >> i) & 1); //22..31
107 put_rac(c, state + 22 + i, (a >> i) & 1); //22..31
110 put_rac(c, state + 11 + el, v < 0); //11..21
112 put_rac(c, state + 0, 1);
116 static inline void put_symbol2(RangeCoder *c, uint8_t *state, int v, int log2)
118 int r = log2 >= 0 ? 1<<log2 : 1;
121 av_assert2(log2 >= -4);
124 put_rac(c, state + 4 + log2, 1);
127 if (log2 > 0) r += r;
129 put_rac(c, state + 4 + log2, 0);
131 for (int i = log2 - 1; i >= 0; i--)
132 put_rac(c, state + 31 - i, (v >> i) & 1);
135 static int get_encode_buffer(SnowContext *s, AVFrame *frame)
139 frame->width = s->avctx->width + 2 * EDGE_WIDTH;
140 frame->height = s->avctx->height + 2 * EDGE_WIDTH;
142 ret = ff_encode_alloc_frame(s->avctx, frame);
145 for (int i = 0; frame->data[i]; i++) {
146 int offset = (EDGE_WIDTH >> (i ? s->chroma_v_shift : 0)) *
148 (EDGE_WIDTH >> (i ? s->chroma_h_shift : 0));
149 frame->data[i] += offset;
151 frame->width = s->avctx->width;
152 frame->height = s->avctx->height;
157 static av_cold int encode_init(AVCodecContext *avctx)
159 SnowEncContext *const enc = avctx->priv_data;
160 SnowContext *const s = &enc->com;
161 MpegEncContext *const mpv = &enc->m;
162 int plane_index, ret;
165 if (enc->pred == DWT_97
166 && (avctx->flags & AV_CODEC_FLAG_QSCALE)
167 && avctx->global_quality == 0){
168 av_log(avctx, AV_LOG_ERROR, "The 9/7 wavelet is incompatible with lossless mode.\n");
169 return AVERROR(EINVAL);
172 s->spatial_decomposition_type = enc->pred; //FIXME add decorrelator type r transform_type
174 s->mv_scale = (avctx->flags & AV_CODEC_FLAG_QPEL) ? 2 : 4;
175 s->block_max_depth= (avctx->flags & AV_CODEC_FLAG_4MV ) ? 1 : 0;
177 for(plane_index=0; plane_index<3; plane_index++){
178 s->plane[plane_index].diag_mc= 1;
179 s->plane[plane_index].htaps= 6;
180 s->plane[plane_index].hcoeff[0]= 40;
181 s->plane[plane_index].hcoeff[1]= -10;
182 s->plane[plane_index].hcoeff[2]= 2;
183 s->plane[plane_index].fast_mc= 1;
186 // Must be before ff_snow_common_init()
187 ff_hpeldsp_init(&s->hdsp, avctx->flags);
188 if ((ret = ff_snow_common_init(avctx)) < 0) {
193 enc->qdsp.put_qpel_pixels_tab [0][dy+dx/4]=\
194 enc->qdsp.put_no_rnd_qpel_pixels_tab[0][dy+dx/4]=\
195 s->h264qpel.put_h264_qpel_pixels_tab[0][dy+dx/4];\
196 enc->qdsp.put_qpel_pixels_tab [1][dy+dx/4]=\
197 enc->qdsp.put_no_rnd_qpel_pixels_tab[1][dy+dx/4]=\
198 s->h264qpel.put_h264_qpel_pixels_tab[1][dy+dx/4];
217 ff_me_cmp_init(&enc->mecc, avctx);
218 ff_mpegvideoencdsp_init(&enc->mpvencdsp, avctx);
220 ff_snow_alloc_blocks(s);
225 mpv->bit_rate= avctx->bit_rate;
226 mpv->lmin = avctx->mb_lmin;
227 mpv->lmax = avctx->mb_lmax;
228 mpv->mb_num = (avctx->width * avctx->height + 255) / 256; // For ratecontrol
231 mpv->me.scratchpad = av_calloc(avctx->width + 64, 2*16*2*sizeof(uint8_t));
232 mpv->sc.obmc_scratchpad= av_mallocz(MB_SIZE*MB_SIZE*12*sizeof(uint32_t));
233 mpv->me.map = av_mallocz(2 * ME_MAP_SIZE * sizeof(*mpv->me.map));
234 if (!mpv->me.scratchpad || !mpv->me.map || !mpv->sc.obmc_scratchpad)
235 return AVERROR(ENOMEM);
236 mpv->me.score_map = mpv->me.map + ME_MAP_SIZE;
238 ff_h263_encode_init(mpv); //mv_penalty
240 s->max_ref_frames = av_clip(avctx->refs, 1, MAX_REF_FRAMES);
242 if(avctx->flags&AV_CODEC_FLAG_PASS1){
243 if(!avctx->stats_out)
244 avctx->stats_out = av_mallocz(256);
246 if (!avctx->stats_out)
247 return AVERROR(ENOMEM);
249 if((avctx->flags&AV_CODEC_FLAG_PASS2) || !(avctx->flags&AV_CODEC_FLAG_QSCALE)){
250 ret = ff_rate_control_init(mpv);
254 enc->pass1_rc = !(avctx->flags & (AV_CODEC_FLAG_QSCALE|AV_CODEC_FLAG_PASS2));
256 switch(avctx->pix_fmt){
257 case AV_PIX_FMT_YUV444P:
258 // case AV_PIX_FMT_YUV422P:
259 case AV_PIX_FMT_YUV420P:
260 // case AV_PIX_FMT_YUV411P:
261 case AV_PIX_FMT_YUV410P:
263 s->colorspace_type= 0;
265 case AV_PIX_FMT_GRAY8:
267 s->colorspace_type = 1;
269 /* case AV_PIX_FMT_RGB32:
274 ret = av_pix_fmt_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_h_shift,
279 ret = ff_set_cmp(&enc->mecc, enc->mecc.me_cmp, s->avctx->me_cmp);
280 ret |= ff_set_cmp(&enc->mecc, enc->mecc.me_sub_cmp, s->avctx->me_sub_cmp);
282 return AVERROR(EINVAL);
284 s->input_picture = av_frame_alloc();
285 if (!s->input_picture)
286 return AVERROR(ENOMEM);
288 if ((ret = get_encode_buffer(s, s->input_picture)) < 0)
291 if (enc->motion_est == FF_ME_ITER) {
292 int size= s->b_width * s->b_height << 2*s->block_max_depth;
293 for(i=0; i<s->max_ref_frames; i++){
294 s->ref_mvs[i] = av_calloc(size, sizeof(*s->ref_mvs[i]));
295 s->ref_scores[i] = av_calloc(size, sizeof(*s->ref_scores[i]));
296 if (!s->ref_mvs[i] || !s->ref_scores[i])
297 return AVERROR(ENOMEM);
304 //near copy & paste from dsputil, FIXME
305 static int pix_sum(const uint8_t * pix, int line_size, int w, int h)
310 for (i = 0; i < h; i++) {
311 for (j = 0; j < w; j++) {
315 pix += line_size - w;
320 //near copy & paste from dsputil, FIXME
321 static int pix_norm1(const uint8_t * pix, int line_size, int w)
324 const uint32_t *sq = ff_square_tab + 256;
327 for (i = 0; i < w; i++) {
328 for (j = 0; j < w; j ++) {
332 pix += line_size - w;
337 static inline int get_penalty_factor(int lambda, int lambda2, int type){
341 return lambda>>FF_LAMBDA_SHIFT;
343 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
345 return (4*lambda)>>(FF_LAMBDA_SHIFT);
347 return (2*lambda)>>(FF_LAMBDA_SHIFT);
350 return (2*lambda)>>FF_LAMBDA_SHIFT;
355 return lambda2>>FF_LAMBDA_SHIFT;
364 #define P_TOPRIGHT P[3]
365 #define P_MEDIAN P[4]
367 #define FLAG_QPEL 1 //must be 1
369 static int encode_q_branch(SnowEncContext *enc, int level, int x, int y)
371 SnowContext *const s = &enc->com;
372 MotionEstContext *const c = &enc->m.me;
373 uint8_t p_buffer[1024];
374 uint8_t i_buffer[1024];
375 uint8_t p_state[sizeof(s->block_state)];
376 uint8_t i_state[sizeof(s->block_state)];
378 uint8_t *pbbak= s->c.bytestream;
379 uint8_t *pbbak_start= s->c.bytestream_start;
380 int score, score2, iscore, i_len, p_len, block_s, sum, base_bits;
381 const int w= s->b_width << s->block_max_depth;
382 const int h= s->b_height << s->block_max_depth;
383 const int rem_depth= s->block_max_depth - level;
384 const int index= (x + y*w) << rem_depth;
385 const int block_w= 1<<(LOG2_MB_SIZE - level);
386 int trx= (x+1)<<rem_depth;
387 int try= (y+1)<<rem_depth;
388 const BlockNode *left = x ? &s->block[index-1] : &null_block;
389 const BlockNode *top = y ? &s->block[index-w] : &null_block;
390 const BlockNode *right = trx<w ? &s->block[index+1] : &null_block;
391 const BlockNode *bottom= try<h ? &s->block[index+w] : &null_block;
392 const BlockNode *tl = y && x ? &s->block[index-w-1] : left;
393 const BlockNode *tr = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
394 int pl = left->color[0];
395 int pcb= left->color[1];
396 int pcr= left->color[2];
400 const int stride= s->current_picture->linesize[0];
401 const int uvstride= s->current_picture->linesize[1];
402 const uint8_t *const current_data[3] = { s->input_picture->data[0] + (x + y* stride)*block_w,
403 s->input_picture->data[1] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift),
404 s->input_picture->data[2] + ((x*block_w)>>s->chroma_h_shift) + ((y*uvstride*block_w)>>s->chroma_v_shift)};
406 int16_t last_mv[3][2];
407 int qpel= !!(s->avctx->flags & AV_CODEC_FLAG_QPEL); //unused
408 const int shift= 1+qpel;
409 int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
410 int mx_context= av_log2(2*FFABS(left->mx - top->mx));
411 int my_context= av_log2(2*FFABS(left->my - top->my));
412 int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
413 int ref, best_ref, ref_score, ref_mx, ref_my;
415 av_assert0(sizeof(s->block_state) >= 256);
417 set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
421 // clip predictors / edge ?
427 P_TOPRIGHT[0]= tr->mx;
428 P_TOPRIGHT[1]= tr->my;
430 last_mv[0][0]= s->block[index].mx;
431 last_mv[0][1]= s->block[index].my;
432 last_mv[1][0]= right->mx;
433 last_mv[1][1]= right->my;
434 last_mv[2][0]= bottom->mx;
435 last_mv[2][1]= bottom->my;
437 enc->m.mb_stride = 2;
442 av_assert1(c-> stride == stride);
443 av_assert1(c->uvstride == uvstride);
445 c->penalty_factor = get_penalty_factor(enc->lambda, enc->lambda2, c->avctx->me_cmp);
446 c->sub_penalty_factor= get_penalty_factor(enc->lambda, enc->lambda2, c->avctx->me_sub_cmp);
447 c->mb_penalty_factor = get_penalty_factor(enc->lambda, enc->lambda2, c->avctx->mb_cmp);
448 c->current_mv_penalty = c->mv_penalty[enc->m.f_code=1] + MAX_DMV;
450 c->xmin = - x*block_w - 16+3;
451 c->ymin = - y*block_w - 16+3;
452 c->xmax = - (x+1)*block_w + (w<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
453 c->ymax = - (y+1)*block_w + (h<<(LOG2_MB_SIZE - s->block_max_depth)) + 16-3;
455 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
456 if(P_LEFT[1] > (c->ymax<<shift)) P_LEFT[1] = (c->ymax<<shift);
457 if(P_TOP[0] > (c->xmax<<shift)) P_TOP[0] = (c->xmax<<shift);
458 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
459 if(P_TOPRIGHT[0] < (c->xmin * (1<<shift))) P_TOPRIGHT[0]= (c->xmin * (1<<shift));
460 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); //due to pmx no clip
461 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
463 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
464 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
467 c->pred_x= P_LEFT[0];
468 c->pred_y= P_LEFT[1];
470 c->pred_x = P_MEDIAN[0];
471 c->pred_y = P_MEDIAN[1];
476 for(ref=0; ref<s->ref_frames; ref++){
477 init_ref(c, current_data, s->last_picture[ref]->data, NULL, block_w*x, block_w*y, 0);
479 ref_score= ff_epzs_motion_search(&enc->m, &ref_mx, &ref_my, P, 0, /*ref_index*/ 0, last_mv,
480 (1<<16)>>shift, level-LOG2_MB_SIZE+4, block_w);
482 av_assert2(ref_mx >= c->xmin);
483 av_assert2(ref_mx <= c->xmax);
484 av_assert2(ref_my >= c->ymin);
485 av_assert2(ref_my <= c->ymax);
487 ref_score= c->sub_motion_search(&enc->m, &ref_mx, &ref_my, ref_score, 0, 0, level-LOG2_MB_SIZE+4, block_w);
488 ref_score= ff_get_mb_score(&enc->m, ref_mx, ref_my, 0, 0, level-LOG2_MB_SIZE+4, block_w, 0);
489 ref_score+= 2*av_log2(2*ref)*c->penalty_factor;
491 s->ref_mvs[ref][index][0]= ref_mx;
492 s->ref_mvs[ref][index][1]= ref_my;
493 s->ref_scores[ref][index]= ref_score;
495 if(score > ref_score){
502 //FIXME if mb_cmp != SSE then intra cannot be compared currently and mb_penalty vs. lambda2
505 base_bits= get_rac_count(&s->c) - 8*(s->c.bytestream - s->c.bytestream_start);
508 pc.bytestream= p_buffer; //FIXME end/start? and at the other stoo
509 memcpy(p_state, s->block_state, sizeof(s->block_state));
511 if(level!=s->block_max_depth)
512 put_rac(&pc, &p_state[4 + s_context], 1);
513 put_rac(&pc, &p_state[1 + left->type + top->type], 0);
514 if(s->ref_frames > 1)
515 put_symbol(&pc, &p_state[128 + 1024 + 32*ref_context], best_ref, 0);
516 pred_mv(s, &pmx, &pmy, best_ref, left, top, tr);
517 put_symbol(&pc, &p_state[128 + 32*(mx_context + 16*!!best_ref)], mx - pmx, 1);
518 put_symbol(&pc, &p_state[128 + 32*(my_context + 16*!!best_ref)], my - pmy, 1);
519 p_len= pc.bytestream - pc.bytestream_start;
520 score += (enc->lambda2*(get_rac_count(&pc)-base_bits))>>FF_LAMBDA_SHIFT;
522 block_s= block_w*block_w;
523 sum = pix_sum(current_data[0], stride, block_w, block_w);
524 l= (sum + block_s/2)/block_s;
525 iscore = pix_norm1(current_data[0], stride, block_w) - 2*l*sum + l*l*block_s;
527 if (s->nb_planes > 2) {
528 block_s= block_w*block_w>>(s->chroma_h_shift + s->chroma_v_shift);
529 sum = pix_sum(current_data[1], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
530 cb= (sum + block_s/2)/block_s;
531 // iscore += pix_norm1(¤t_mb[1][0], uvstride, block_w>>1) - 2*cb*sum + cb*cb*block_s;
532 sum = pix_sum(current_data[2], uvstride, block_w>>s->chroma_h_shift, block_w>>s->chroma_v_shift);
533 cr= (sum + block_s/2)/block_s;
534 // iscore += pix_norm1(¤t_mb[2][0], uvstride, block_w>>1) - 2*cr*sum + cr*cr*block_s;
540 ic.bytestream= i_buffer; //FIXME end/start? and at the other stoo
541 memcpy(i_state, s->block_state, sizeof(s->block_state));
542 if(level!=s->block_max_depth)
543 put_rac(&ic, &i_state[4 + s_context], 1);
544 put_rac(&ic, &i_state[1 + left->type + top->type], 1);
545 put_symbol(&ic, &i_state[32], l-pl , 1);
546 if (s->nb_planes > 2) {
547 put_symbol(&ic, &i_state[64], cb-pcb, 1);
548 put_symbol(&ic, &i_state[96], cr-pcr, 1);
550 i_len= ic.bytestream - ic.bytestream_start;
551 iscore += (enc->lambda2*(get_rac_count(&ic)-base_bits))>>FF_LAMBDA_SHIFT;
553 av_assert1(iscore < 255*255*256 + enc->lambda2*10);
554 av_assert1(iscore >= 0);
555 av_assert1(l>=0 && l<=255);
556 av_assert1(pl>=0 && pl<=255);
559 int varc= iscore >> 8;
560 int vard= score >> 8;
561 if (vard <= 64 || vard < varc)
562 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
564 c->scene_change_score += enc->m.qscale;
567 if(level!=s->block_max_depth){
568 put_rac(&s->c, &s->block_state[4 + s_context], 0);
569 score2 = encode_q_branch(enc, level+1, 2*x+0, 2*y+0);
570 score2+= encode_q_branch(enc, level+1, 2*x+1, 2*y+0);
571 score2+= encode_q_branch(enc, level+1, 2*x+0, 2*y+1);
572 score2+= encode_q_branch(enc, level+1, 2*x+1, 2*y+1);
573 score2+= enc->lambda2>>FF_LAMBDA_SHIFT; //FIXME exact split overhead
575 if(score2 < score && score2 < iscore)
580 pred_mv(s, &pmx, &pmy, 0, left, top, tr);
581 memcpy(pbbak, i_buffer, i_len);
583 s->c.bytestream_start= pbbak_start;
584 s->c.bytestream= pbbak + i_len;
585 set_blocks(s, level, x, y, l, cb, cr, pmx, pmy, 0, BLOCK_INTRA);
586 memcpy(s->block_state, i_state, sizeof(s->block_state));
589 memcpy(pbbak, p_buffer, p_len);
591 s->c.bytestream_start= pbbak_start;
592 s->c.bytestream= pbbak + p_len;
593 set_blocks(s, level, x, y, pl, pcb, pcr, mx, my, best_ref, 0);
594 memcpy(s->block_state, p_state, sizeof(s->block_state));
599 static void encode_q_branch2(SnowContext *s, int level, int x, int y){
600 const int w= s->b_width << s->block_max_depth;
601 const int rem_depth= s->block_max_depth - level;
602 const int index= (x + y*w) << rem_depth;
603 int trx= (x+1)<<rem_depth;
604 BlockNode *b= &s->block[index];
605 const BlockNode *left = x ? &s->block[index-1] : &null_block;
606 const BlockNode *top = y ? &s->block[index-w] : &null_block;
607 const BlockNode *tl = y && x ? &s->block[index-w-1] : left;
608 const BlockNode *tr = y && trx<w && ((x&1)==0 || level==0) ? &s->block[index-w+(1<<rem_depth)] : tl; //FIXME use lt
609 int pl = left->color[0];
610 int pcb= left->color[1];
611 int pcr= left->color[2];
613 int ref_context= av_log2(2*left->ref) + av_log2(2*top->ref);
614 int mx_context= av_log2(2*FFABS(left->mx - top->mx)) + 16*!!b->ref;
615 int my_context= av_log2(2*FFABS(left->my - top->my)) + 16*!!b->ref;
616 int s_context= 2*left->level + 2*top->level + tl->level + tr->level;
619 set_blocks(s, level, x, y, pl, pcb, pcr, 0, 0, 0, BLOCK_INTRA);
623 if(level!=s->block_max_depth){
624 if(same_block(b,b+1) && same_block(b,b+w) && same_block(b,b+w+1)){
625 put_rac(&s->c, &s->block_state[4 + s_context], 1);
627 put_rac(&s->c, &s->block_state[4 + s_context], 0);
628 encode_q_branch2(s, level+1, 2*x+0, 2*y+0);
629 encode_q_branch2(s, level+1, 2*x+1, 2*y+0);
630 encode_q_branch2(s, level+1, 2*x+0, 2*y+1);
631 encode_q_branch2(s, level+1, 2*x+1, 2*y+1);
635 if(b->type & BLOCK_INTRA){
636 pred_mv(s, &pmx, &pmy, 0, left, top, tr);
637 put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 1);
638 put_symbol(&s->c, &s->block_state[32], b->color[0]-pl , 1);
639 if (s->nb_planes > 2) {
640 put_symbol(&s->c, &s->block_state[64], b->color[1]-pcb, 1);
641 put_symbol(&s->c, &s->block_state[96], b->color[2]-pcr, 1);
643 set_blocks(s, level, x, y, b->color[0], b->color[1], b->color[2], pmx, pmy, 0, BLOCK_INTRA);
645 pred_mv(s, &pmx, &pmy, b->ref, left, top, tr);
646 put_rac(&s->c, &s->block_state[1 + (left->type&1) + (top->type&1)], 0);
647 if(s->ref_frames > 1)
648 put_symbol(&s->c, &s->block_state[128 + 1024 + 32*ref_context], b->ref, 0);
649 put_symbol(&s->c, &s->block_state[128 + 32*mx_context], b->mx - pmx, 1);
650 put_symbol(&s->c, &s->block_state[128 + 32*my_context], b->my - pmy, 1);
651 set_blocks(s, level, x, y, pl, pcb, pcr, b->mx, b->my, b->ref, 0);
655 static int get_dc(SnowEncContext *enc, int mb_x, int mb_y, int plane_index)
657 SnowContext *const s = &enc->com;
659 Plane *p= &s->plane[plane_index];
660 const int block_size = MB_SIZE >> s->block_max_depth;
661 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
662 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
663 const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
664 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
665 const int ref_stride= s->current_picture->linesize[plane_index];
666 const uint8_t *src = s->input_picture->data[plane_index];
667 IDWTELEM *dst= (IDWTELEM*)enc->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4; //FIXME change to unsigned
668 const int b_stride = s->b_width << s->block_max_depth;
669 const int w= p->width;
670 const int h= p->height;
671 int index= mb_x + mb_y*b_stride;
672 BlockNode *b= &s->block[index];
673 BlockNode backup= *b;
677 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc stuff above
679 b->type|= BLOCK_INTRA;
680 b->color[plane_index]= 0;
681 memset(dst, 0, obmc_stride*obmc_stride*sizeof(IDWTELEM));
684 int mb_x2= mb_x + (i &1) - 1;
685 int mb_y2= mb_y + (i>>1) - 1;
686 int x= block_w*mb_x2 + block_w/2;
687 int y= block_h*mb_y2 + block_h/2;
689 add_yblock(s, 0, NULL, dst + (i&1)*block_w + (i>>1)*obmc_stride*block_h, NULL, obmc,
690 x, y, block_w, block_h, w, h, obmc_stride, ref_stride, obmc_stride, mb_x2, mb_y2, 0, 0, plane_index);
692 for(y2= FFMAX(y, 0); y2<FFMIN(h, y+block_h); y2++){
693 for(x2= FFMAX(x, 0); x2<FFMIN(w, x+block_w); x2++){
694 int index= x2-(block_w*mb_x - block_w/2) + (y2-(block_h*mb_y - block_h/2))*obmc_stride;
695 int obmc_v= obmc[index];
697 if(y<0) obmc_v += obmc[index + block_h*obmc_stride];
698 if(x<0) obmc_v += obmc[index + block_w];
699 if(y+block_h>h) obmc_v += obmc[index - block_h*obmc_stride];
700 if(x+block_w>w) obmc_v += obmc[index - block_w];
701 //FIXME precalculate this or simplify it somehow else
703 d = -dst[index] + (1<<(FRAC_BITS-1));
705 ab += (src[x2 + y2*ref_stride] - (d>>FRAC_BITS)) * obmc_v;
706 aa += obmc_v * obmc_v; //FIXME precalculate this
712 return av_clip_uint8( ROUNDED_DIV(ab<<LOG2_OBMC_MAX, aa) ); //FIXME we should not need clipping
715 static inline int get_block_bits(SnowContext *s, int x, int y, int w){
716 const int b_stride = s->b_width << s->block_max_depth;
717 const int b_height = s->b_height<< s->block_max_depth;
718 int index= x + y*b_stride;
719 const BlockNode *b = &s->block[index];
720 const BlockNode *left = x ? &s->block[index-1] : &null_block;
721 const BlockNode *top = y ? &s->block[index-b_stride] : &null_block;
722 const BlockNode *tl = y && x ? &s->block[index-b_stride-1] : left;
723 const BlockNode *tr = y && x+w<b_stride ? &s->block[index-b_stride+w] : tl;
725 // int mx_context= av_log2(2*FFABS(left->mx - top->mx));
726 // int my_context= av_log2(2*FFABS(left->my - top->my));
728 if(x<0 || x>=b_stride || y>=b_height)
737 //FIXME try accurate rate
738 //FIXME intra and inter predictors if surrounding blocks are not the same type
739 if(b->type & BLOCK_INTRA){
740 return 3+2*( av_log2(2*FFABS(left->color[0] - b->color[0]))
741 + av_log2(2*FFABS(left->color[1] - b->color[1]))
742 + av_log2(2*FFABS(left->color[2] - b->color[2])));
744 pred_mv(s, &dmx, &dmy, b->ref, left, top, tr);
747 return 2*(1 + av_log2(2*FFABS(dmx)) //FIXME kill the 2* can be merged in lambda
748 + av_log2(2*FFABS(dmy))
749 + av_log2(2*b->ref));
753 static int get_block_rd(SnowEncContext *enc, int mb_x, int mb_y,
754 int plane_index, uint8_t (*obmc_edged)[MB_SIZE * 2])
756 SnowContext *const s = &enc->com;
757 Plane *p= &s->plane[plane_index];
758 const int block_size = MB_SIZE >> s->block_max_depth;
759 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
760 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
761 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
762 const int ref_stride= s->current_picture->linesize[plane_index];
763 uint8_t *dst= s->current_picture->data[plane_index];
764 const uint8_t *src = s->input_picture->data[plane_index];
765 IDWTELEM *pred= (IDWTELEM*)enc->m.sc.obmc_scratchpad + plane_index*block_size*block_size*4;
766 uint8_t *cur = s->scratchbuf;
767 uint8_t *tmp = s->emu_edge_buffer;
768 const int b_stride = s->b_width << s->block_max_depth;
769 const int b_height = s->b_height<< s->block_max_depth;
770 const int w= p->width;
771 const int h= p->height;
774 const int penalty_factor = get_penalty_factor(enc->lambda, enc->lambda2, s->avctx->me_cmp);
775 int sx= block_w*mb_x - block_w/2;
776 int sy= block_h*mb_y - block_h/2;
777 int x0= FFMAX(0,-sx);
778 int y0= FFMAX(0,-sy);
779 int x1= FFMIN(block_w*2, w-sx);
780 int y1= FFMIN(block_h*2, h-sy);
783 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below chckinhg only block_w
785 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);
787 for(y=y0; y<y1; y++){
788 const uint8_t *obmc1= obmc_edged[y];
789 const IDWTELEM *pred1 = pred + y*obmc_stride;
790 uint8_t *cur1 = cur + y*ref_stride;
791 uint8_t *dst1 = dst + sx + (sy+y)*ref_stride;
792 for(x=x0; x<x1; x++){
793 #if FRAC_BITS >= LOG2_OBMC_MAX
794 int v = (cur1[x] * obmc1[x]) << (FRAC_BITS - LOG2_OBMC_MAX);
796 int v = (cur1[x] * obmc1[x] + (1<<(LOG2_OBMC_MAX - FRAC_BITS-1))) >> (LOG2_OBMC_MAX - FRAC_BITS);
798 v = (v + pred1[x]) >> FRAC_BITS;
799 if(v&(~255)) v= ~(v>>31);
804 /* copy the regions where obmc[] = (uint8_t)256 */
805 if(LOG2_OBMC_MAX == 8
806 && (mb_x == 0 || mb_x == b_stride-1)
807 && (mb_y == 0 || mb_y == b_height-1)){
817 memcpy(dst + sx+x0 + (sy+y)*ref_stride, cur + x0 + y*ref_stride, x1-x0);
821 /* FIXME rearrange dsputil to fit 32x32 cmp functions */
822 /* FIXME check alignment of the cmp wavelet vs the encoding wavelet */
823 /* FIXME cmps overlap but do not cover the wavelet's whole support.
824 * So improving the score of one block is not strictly guaranteed
825 * to improve the score of the whole frame, thus iterative motion
826 * estimation does not always converge. */
827 if(s->avctx->me_cmp == FF_CMP_W97)
828 distortion = ff_w97_32_c(&enc->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
829 else if(s->avctx->me_cmp == FF_CMP_W53)
830 distortion = ff_w53_32_c(&enc->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, 32);
834 int off = sx+16*(i&1) + (sy+16*(i>>1))*ref_stride;
835 distortion += enc->mecc.me_cmp[0](&enc->m, src + off, dst + off, ref_stride, 16);
839 av_assert2(block_w==8);
840 distortion = enc->mecc.me_cmp[0](&enc->m, src + sx + sy*ref_stride, dst + sx + sy*ref_stride, ref_stride, block_w*2);
849 rate += get_block_bits(s, mb_x + (i&1) - (i>>1), mb_y + (i>>1), 1);
851 if(mb_x == b_stride-2)
852 rate += get_block_bits(s, mb_x + 1, mb_y + 1, 1);
854 return distortion + rate*penalty_factor;
857 static int get_4block_rd(SnowEncContext *enc, int mb_x, int mb_y, int plane_index)
859 SnowContext *const s = &enc->com;
861 Plane *p= &s->plane[plane_index];
862 const int block_size = MB_SIZE >> s->block_max_depth;
863 const int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
864 const int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
865 const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
866 const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
867 const int ref_stride= s->current_picture->linesize[plane_index];
868 uint8_t *dst= s->current_picture->data[plane_index];
869 const uint8_t *src = s->input_picture->data[plane_index];
870 //FIXME zero_dst is const but add_yblock changes dst if add is 0 (this is never the case for dst=zero_dst
871 // const has only been removed from zero_dst to suppress a warning
872 static IDWTELEM zero_dst[4096]; //FIXME
873 const int b_stride = s->b_width << s->block_max_depth;
874 const int w= p->width;
875 const int h= p->height;
878 const int penalty_factor= get_penalty_factor(enc->lambda, enc->lambda2, s->avctx->me_cmp);
880 av_assert2(s->chroma_h_shift == s->chroma_v_shift); //obmc and square assumtions below
883 int mb_x2= mb_x + (i%3) - 1;
884 int mb_y2= mb_y + (i/3) - 1;
885 int x= block_w*mb_x2 + block_w/2;
886 int y= block_h*mb_y2 + block_h/2;
888 add_yblock(s, 0, NULL, zero_dst, dst, obmc,
889 x, y, block_w, block_h, w, h, /*dst_stride*/0, ref_stride, obmc_stride, mb_x2, mb_y2, 1, 1, plane_index);
891 //FIXME find a cleaner/simpler way to skip the outside stuff
892 for(y2= y; y2<0; y2++)
893 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
894 for(y2= h; y2<y+block_h; y2++)
895 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, block_w);
897 for(y2= y; y2<y+block_h; y2++)
898 memcpy(dst + x + y2*ref_stride, src + x + y2*ref_stride, -x);
901 for(y2= y; y2<y+block_h; y2++)
902 memcpy(dst + w + y2*ref_stride, src + w + y2*ref_stride, x+block_w - w);
905 av_assert1(block_w== 8 || block_w==16);
906 distortion += enc->mecc.me_cmp[block_w==8](&enc->m, src + x + y*ref_stride, dst + x + y*ref_stride, ref_stride, block_h);
910 BlockNode *b= &s->block[mb_x+mb_y*b_stride];
911 int merged= same_block(b,b+1) && same_block(b,b+b_stride) && same_block(b,b+b_stride+1);
919 rate = get_block_bits(s, mb_x, mb_y, 2);
920 for(i=merged?4:0; i<9; i++){
921 static const int dxy[9][2] = {{0,0},{1,0},{0,1},{1,1},{2,0},{2,1},{-1,2},{0,2},{1,2}};
922 rate += get_block_bits(s, mb_x + dxy[i][0], mb_y + dxy[i][1], 1);
925 return distortion + rate*penalty_factor;
928 static int encode_subband_c0run(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
929 const int w= b->width;
930 const int h= b->height;
935 int *runs = s->run_buffer;
942 int /*ll=0, */l=0, lt=0, t=0, rt=0;
943 v= src[x + y*stride];
946 t= src[x + (y-1)*stride];
948 lt= src[x - 1 + (y-1)*stride];
951 rt= src[x + 1 + (y-1)*stride];
955 l= src[x - 1 + y*stride];
957 if(orientation==1) ll= src[y + (x-2)*stride];
958 else ll= src[x - 2 + y*stride];
964 if(px<b->parent->width && py<b->parent->height)
965 p= parent[px + py*2*stride];
967 if(!(/*ll|*/l|lt|t|rt|p)){
969 runs[run_index++]= run;
977 max_index= run_index;
978 runs[run_index++]= run;
980 run= runs[run_index++];
982 put_symbol2(&s->c, b->state[30], max_index, 0);
983 if(run_index <= max_index)
984 put_symbol2(&s->c, b->state[1], run, 3);
987 if(s->c.bytestream_end - s->c.bytestream < w*40){
988 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
989 return AVERROR(ENOMEM);
993 int /*ll=0, */l=0, lt=0, t=0, rt=0;
994 v= src[x + y*stride];
997 t= src[x + (y-1)*stride];
999 lt= src[x - 1 + (y-1)*stride];
1002 rt= src[x + 1 + (y-1)*stride];
1006 l= src[x - 1 + y*stride];
1008 if(orientation==1) ll= src[y + (x-2)*stride];
1009 else ll= src[x - 2 + y*stride];
1015 if(px<b->parent->width && py<b->parent->height)
1016 p= parent[px + py*2*stride];
1018 if(/*ll|*/l|lt|t|rt|p){
1019 int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
1021 put_rac(&s->c, &b->state[0][context], !!v);
1024 run= runs[run_index++];
1026 if(run_index <= max_index)
1027 put_symbol2(&s->c, b->state[1], run, 3);
1035 int context= av_log2(/*FFABS(ll) + */3*FFABS(l) + FFABS(lt) + 2*FFABS(t) + FFABS(rt) + FFABS(p));
1036 int l2= 2*FFABS(l) + (l<0);
1037 int t2= 2*FFABS(t) + (t<0);
1039 put_symbol2(&s->c, b->state[context + 2], FFABS(v)-1, context-4);
1040 put_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l2&0xFF] + 3*ff_quant3bA[t2&0xFF]], v<0);
1048 static int encode_subband(SnowContext *s, SubBand *b, const IDWTELEM *src, const IDWTELEM *parent, int stride, int orientation){
1049 // encode_subband_qtree(s, b, src, parent, stride, orientation);
1050 // encode_subband_z0run(s, b, src, parent, stride, orientation);
1051 return encode_subband_c0run(s, b, src, parent, stride, orientation);
1052 // encode_subband_dzr(s, b, src, parent, stride, orientation);
1055 static av_always_inline int check_block_intra(SnowEncContext *enc, int mb_x, int mb_y, int p[3],
1056 uint8_t (*obmc_edged)[MB_SIZE * 2], int *best_rd)
1058 SnowContext *const s = &enc->com;
1059 const int b_stride= s->b_width << s->block_max_depth;
1060 BlockNode *block= &s->block[mb_x + mb_y * b_stride];
1061 BlockNode backup= *block;
1064 av_assert2(mb_x>=0 && mb_y>=0);
1065 av_assert2(mb_x<b_stride);
1067 block->color[0] = p[0];
1068 block->color[1] = p[1];
1069 block->color[2] = p[2];
1070 block->type |= BLOCK_INTRA;
1072 rd = get_block_rd(enc, mb_x, mb_y, 0, obmc_edged) + enc->intra_penalty;
1084 /* special case for int[2] args we discard afterwards,
1085 * fixes compilation problem with gcc 2.95 */
1086 static av_always_inline int check_block_inter(SnowEncContext *enc,
1087 int mb_x, int mb_y, int p0, int p1,
1088 uint8_t (*obmc_edged)[MB_SIZE * 2], int *best_rd)
1090 SnowContext *const s = &enc->com;
1091 const int b_stride = s->b_width << s->block_max_depth;
1092 BlockNode *block = &s->block[mb_x + mb_y * b_stride];
1093 BlockNode backup = *block;
1097 av_assert2(mb_x >= 0 && mb_y >= 0);
1098 av_assert2(mb_x < b_stride);
1100 index = (p0 + 31 * p1) & (ME_CACHE_SIZE-1);
1101 value = enc->me_cache_generation + (p0 >> 10) + p1 * (1 << 6) + (block->ref << 12);
1102 if (enc->me_cache[index] == value)
1104 enc->me_cache[index] = value;
1108 block->type &= ~BLOCK_INTRA;
1110 rd = get_block_rd(enc, mb_x, mb_y, 0, obmc_edged);
1113 if (rd < *best_rd) {
1122 static av_always_inline int check_4block_inter(SnowEncContext *enc, int mb_x, int mb_y,
1123 int p0, int p1, int ref, int *best_rd)
1125 SnowContext *const s = &enc->com;
1126 const int b_stride= s->b_width << s->block_max_depth;
1127 BlockNode *block= &s->block[mb_x + mb_y * b_stride];
1128 BlockNode backup[4];
1132 /* We don't initialize backup[] during variable declaration, because
1133 * that fails to compile on MSVC: "cannot convert from 'BlockNode' to
1135 backup[0] = block[0];
1136 backup[1] = block[1];
1137 backup[2] = block[b_stride];
1138 backup[3] = block[b_stride + 1];
1140 av_assert2(mb_x>=0 && mb_y>=0);
1141 av_assert2(mb_x<b_stride);
1142 av_assert2(((mb_x|mb_y)&1) == 0);
1144 index= (p0 + 31*p1) & (ME_CACHE_SIZE-1);
1145 value = enc->me_cache_generation + (p0>>10) + (p1<<6) + (block->ref<<12);
1146 if (enc->me_cache[index] == value)
1148 enc->me_cache[index] = value;
1153 block->type &= ~BLOCK_INTRA;
1154 block[1]= block[b_stride]= block[b_stride+1]= *block;
1156 rd = get_4block_rd(enc, mb_x, mb_y, 0);
1163 block[0]= backup[0];
1164 block[1]= backup[1];
1165 block[b_stride]= backup[2];
1166 block[b_stride+1]= backup[3];
1171 static void iterative_me(SnowEncContext *enc)
1173 SnowContext *const s = &enc->com;
1174 int pass, mb_x, mb_y;
1175 const int b_width = s->b_width << s->block_max_depth;
1176 const int b_height= s->b_height << s->block_max_depth;
1177 const int b_stride= b_width;
1181 RangeCoder r = s->c;
1182 uint8_t state[sizeof(s->block_state)];
1183 memcpy(state, s->block_state, sizeof(s->block_state));
1184 for(mb_y= 0; mb_y<s->b_height; mb_y++)
1185 for(mb_x= 0; mb_x<s->b_width; mb_x++)
1186 encode_q_branch(enc, 0, mb_x, mb_y);
1188 memcpy(s->block_state, state, sizeof(s->block_state));
1191 for(pass=0; pass<25; pass++){
1194 for(mb_y= 0; mb_y<b_height; mb_y++){
1195 for(mb_x= 0; mb_x<b_width; mb_x++){
1196 int dia_change, i, j, ref;
1197 int best_rd= INT_MAX, ref_rd;
1198 BlockNode backup, ref_b;
1199 const int index= mb_x + mb_y * b_stride;
1200 BlockNode *block= &s->block[index];
1201 BlockNode *tb = mb_y ? &s->block[index-b_stride ] : NULL;
1202 BlockNode *lb = mb_x ? &s->block[index -1] : NULL;
1203 BlockNode *rb = mb_x+1<b_width ? &s->block[index +1] : NULL;
1204 BlockNode *bb = mb_y+1<b_height ? &s->block[index+b_stride ] : NULL;
1205 BlockNode *tlb= mb_x && mb_y ? &s->block[index-b_stride-1] : NULL;
1206 BlockNode *trb= mb_x+1<b_width && mb_y ? &s->block[index-b_stride+1] : NULL;
1207 BlockNode *blb= mb_x && mb_y+1<b_height ? &s->block[index+b_stride-1] : NULL;
1208 BlockNode *brb= mb_x+1<b_width && mb_y+1<b_height ? &s->block[index+b_stride+1] : NULL;
1209 const int b_w= (MB_SIZE >> s->block_max_depth);
1210 uint8_t obmc_edged[MB_SIZE * 2][MB_SIZE * 2];
1212 if(pass && (block->type & BLOCK_OPT))
1214 block->type |= BLOCK_OPT;
1218 if (!enc->me_cache_generation)
1219 memset(enc->me_cache, 0, sizeof(enc->me_cache));
1220 enc->me_cache_generation += 1<<22;
1222 //FIXME precalculate
1225 for (y = 0; y < b_w * 2; y++)
1226 memcpy(obmc_edged[y], ff_obmc_tab[s->block_max_depth] + y * b_w * 2, b_w * 2);
1228 for(y=0; y<b_w*2; y++)
1229 memset(obmc_edged[y], obmc_edged[y][0] + obmc_edged[y][b_w-1], b_w);
1230 if(mb_x==b_stride-1)
1231 for(y=0; y<b_w*2; y++)
1232 memset(obmc_edged[y]+b_w, obmc_edged[y][b_w] + obmc_edged[y][b_w*2-1], b_w);
1234 for(x=0; x<b_w*2; x++)
1235 obmc_edged[0][x] += obmc_edged[b_w-1][x];
1236 for(y=1; y<b_w; y++)
1237 memcpy(obmc_edged[y], obmc_edged[0], b_w*2);
1239 if(mb_y==b_height-1){
1240 for(x=0; x<b_w*2; x++)
1241 obmc_edged[b_w*2-1][x] += obmc_edged[b_w][x];
1242 for(y=b_w; y<b_w*2-1; y++)
1243 memcpy(obmc_edged[y], obmc_edged[b_w*2-1], b_w*2);
1247 //skip stuff outside the picture
1248 if(mb_x==0 || mb_y==0 || mb_x==b_width-1 || mb_y==b_height-1){
1249 const uint8_t *src = s->input_picture->data[0];
1250 uint8_t *dst= s->current_picture->data[0];
1251 const int stride= s->current_picture->linesize[0];
1252 const int block_w= MB_SIZE >> s->block_max_depth;
1253 const int block_h= MB_SIZE >> s->block_max_depth;
1254 const int sx= block_w*mb_x - block_w/2;
1255 const int sy= block_h*mb_y - block_h/2;
1256 const int w= s->plane[0].width;
1257 const int h= s->plane[0].height;
1261 memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1262 for(y=h; y<sy+block_h*2; y++)
1263 memcpy(dst + sx + y*stride, src + sx + y*stride, block_w*2);
1265 for(y=sy; y<sy+block_h*2; y++)
1266 memcpy(dst + sx + y*stride, src + sx + y*stride, -sx);
1268 if(sx+block_w*2 > w){
1269 for(y=sy; y<sy+block_h*2; y++)
1270 memcpy(dst + w + y*stride, src + w + y*stride, sx+block_w*2 - w);
1274 // intra(black) = neighbors' contribution to the current block
1275 for(i=0; i < s->nb_planes; i++)
1276 color[i]= get_dc(enc, mb_x, mb_y, i);
1278 // get previous score (cannot be cached due to OBMC)
1279 if(pass > 0 && (block->type&BLOCK_INTRA)){
1280 int color0[3]= {block->color[0], block->color[1], block->color[2]};
1281 check_block_intra(enc, mb_x, mb_y, color0, obmc_edged, &best_rd);
1283 check_block_inter(enc, mb_x, mb_y, block->mx, block->my, obmc_edged, &best_rd);
1287 for(ref=0; ref < s->ref_frames; ref++){
1288 int16_t (*mvr)[2]= &s->ref_mvs[ref][index];
1289 if(s->ref_scores[ref][index] > s->ref_scores[ref_b.ref][index]*3/2) //FIXME tune threshold
1294 check_block_inter(enc, mb_x, mb_y, mvr[0][0], mvr[0][1], obmc_edged, &best_rd);
1295 check_block_inter(enc, mb_x, mb_y, 0, 0, obmc_edged, &best_rd);
1297 check_block_inter(enc, mb_x, mb_y, mvr[-b_stride][0], mvr[-b_stride][1], obmc_edged, &best_rd);
1299 check_block_inter(enc, mb_x, mb_y, mvr[-1][0], mvr[-1][1], obmc_edged, &best_rd);
1301 check_block_inter(enc, mb_x, mb_y, mvr[1][0], mvr[1][1], obmc_edged, &best_rd);
1303 check_block_inter(enc, mb_x, mb_y, mvr[b_stride][0], mvr[b_stride][1], obmc_edged, &best_rd);
1306 //FIXME avoid subpel interpolation / round to nearest integer
1308 int newx = block->mx;
1309 int newy = block->my;
1310 int dia_size = enc->iterative_dia_size ? enc->iterative_dia_size : FFMAX(s->avctx->dia_size, 1);
1312 for(i=0; i < dia_size; i++){
1314 dia_change |= check_block_inter(enc, mb_x, mb_y, newx+4*(i-j), newy+(4*j), obmc_edged, &best_rd);
1315 dia_change |= check_block_inter(enc, mb_x, mb_y, newx-4*(i-j), newy-(4*j), obmc_edged, &best_rd);
1316 dia_change |= check_block_inter(enc, mb_x, mb_y, newx-(4*j), newy+4*(i-j), obmc_edged, &best_rd);
1317 dia_change |= check_block_inter(enc, mb_x, mb_y, newx+(4*j), newy-4*(i-j), obmc_edged, &best_rd);
1323 static const int square[8][2]= {{+1, 0},{-1, 0},{ 0,+1},{ 0,-1},{+1,+1},{-1,-1},{+1,-1},{-1,+1},};
1326 dia_change |= check_block_inter(enc, mb_x, mb_y, block->mx+square[i][0], block->my+square[i][1], obmc_edged, &best_rd);
1328 //FIXME or try the standard 2 pass qpel or similar
1330 mvr[0][0]= block->mx;
1331 mvr[0][1]= block->my;
1332 if(ref_rd > best_rd){
1339 check_block_intra(enc, mb_x, mb_y, color, obmc_edged, &best_rd);
1340 //FIXME RD style color selection
1341 if(!same_block(block, &backup)){
1342 if(tb ) tb ->type &= ~BLOCK_OPT;
1343 if(lb ) lb ->type &= ~BLOCK_OPT;
1344 if(rb ) rb ->type &= ~BLOCK_OPT;
1345 if(bb ) bb ->type &= ~BLOCK_OPT;
1346 if(tlb) tlb->type &= ~BLOCK_OPT;
1347 if(trb) trb->type &= ~BLOCK_OPT;
1348 if(blb) blb->type &= ~BLOCK_OPT;
1349 if(brb) brb->type &= ~BLOCK_OPT;
1354 av_log(s->avctx, AV_LOG_DEBUG, "pass:%d changed:%d\n", pass, change);
1359 if(s->block_max_depth == 1){
1361 for(mb_y= 0; mb_y<b_height; mb_y+=2){
1362 for(mb_x= 0; mb_x<b_width; mb_x+=2){
1364 int best_rd, init_rd;
1365 const int index= mb_x + mb_y * b_stride;
1368 b[0]= &s->block[index];
1370 b[2]= b[0]+b_stride;
1372 if(same_block(b[0], b[1]) &&
1373 same_block(b[0], b[2]) &&
1374 same_block(b[0], b[3]))
1377 if (!enc->me_cache_generation)
1378 memset(enc->me_cache, 0, sizeof(enc->me_cache));
1379 enc->me_cache_generation += 1<<22;
1381 init_rd = best_rd = get_4block_rd(enc, mb_x, mb_y, 0);
1383 //FIXME more multiref search?
1384 check_4block_inter(enc, mb_x, mb_y,
1385 (b[0]->mx + b[1]->mx + b[2]->mx + b[3]->mx + 2) >> 2,
1386 (b[0]->my + b[1]->my + b[2]->my + b[3]->my + 2) >> 2, 0, &best_rd);
1389 if(!(b[i]->type&BLOCK_INTRA))
1390 check_4block_inter(enc, mb_x, mb_y, b[i]->mx, b[i]->my, b[i]->ref, &best_rd);
1392 if(init_rd != best_rd)
1396 av_log(s->avctx, AV_LOG_ERROR, "pass:4mv changed:%d\n", change*4);
1400 static void encode_blocks(SnowEncContext *enc, int search)
1402 SnowContext *const s = &enc->com;
1407 if (enc->motion_est == FF_ME_ITER && !s->keyframe && search)
1411 if(s->c.bytestream_end - s->c.bytestream < w*MB_SIZE*MB_SIZE*3){ //FIXME nicer limit
1412 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1416 if (enc->motion_est == FF_ME_ITER || !search)
1417 encode_q_branch2(s, 0, x, y);
1419 encode_q_branch (enc, 0, x, y);
1424 static void quantize(SnowContext *s, SubBand *b, IDWTELEM *dst, DWTELEM *src, int stride, int bias){
1425 const int w= b->width;
1426 const int h= b->height;
1427 const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1428 const int qmul= ff_qexp[qlog&(QROOT-1)]<<((qlog>>QSHIFT) + ENCODER_EXTRA_BITS);
1429 int x,y, thres1, thres2;
1431 if(s->qlog == LOSSLESS_QLOG){
1434 dst[x + y*stride]= src[x + y*stride];
1438 bias= bias ? 0 : (3*qmul)>>3;
1439 thres1= ((qmul - bias)>>QEXPSHIFT) - 1;
1445 int i= src[x + y*stride];
1447 if((unsigned)(i+thres1) > thres2){
1450 i/= qmul; //FIXME optimize
1451 dst[x + y*stride]= i;
1455 i/= qmul; //FIXME optimize
1456 dst[x + y*stride]= -i;
1459 dst[x + y*stride]= 0;
1465 int i= src[x + y*stride];
1467 if((unsigned)(i+thres1) > thres2){
1470 i= (i + bias) / qmul; //FIXME optimize
1471 dst[x + y*stride]= i;
1475 i= (i + bias) / qmul; //FIXME optimize
1476 dst[x + y*stride]= -i;
1479 dst[x + y*stride]= 0;
1485 static void dequantize(SnowContext *s, SubBand *b, IDWTELEM *src, int stride){
1486 const int w= b->width;
1487 const int h= b->height;
1488 const int qlog= av_clip(s->qlog + b->qlog, 0, QROOT*16);
1489 const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1490 const int qadd= (s->qbias*qmul)>>QBIAS_SHIFT;
1493 if(s->qlog == LOSSLESS_QLOG) return;
1497 int i= src[x + y*stride];
1499 src[x + y*stride]= -((-i*qmul + qadd)>>(QEXPSHIFT)); //FIXME try different bias
1501 src[x + y*stride]= (( i*qmul + qadd)>>(QEXPSHIFT));
1507 static void decorrelate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1508 const int w= b->width;
1509 const int h= b->height;
1512 for(y=h-1; y>=0; y--){
1513 for(x=w-1; x>=0; x--){
1514 int i= x + y*stride;
1518 if(y && x+1<w) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1519 else src[i] -= src[i - 1];
1521 if(y) src[i] -= mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1522 else src[i] -= src[i - 1];
1525 if(y) src[i] -= src[i - stride];
1531 static void correlate(SnowContext *s, SubBand *b, IDWTELEM *src, int stride, int inverse, int use_median){
1532 const int w= b->width;
1533 const int h= b->height;
1538 int i= x + y*stride;
1542 if(y && x+1<w) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - stride + 1]);
1543 else src[i] += src[i - 1];
1545 if(y) src[i] += mid_pred(src[i - 1], src[i - stride], src[i - 1] + src[i - stride] - src[i - 1 - stride]);
1546 else src[i] += src[i - 1];
1549 if(y) src[i] += src[i - stride];
1555 static void encode_qlogs(SnowContext *s){
1556 int plane_index, level, orientation;
1558 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1559 for(level=0; level<s->spatial_decomposition_count; level++){
1560 for(orientation=level ? 1:0; orientation<4; orientation++){
1561 if(orientation==2) continue;
1562 put_symbol(&s->c, s->header_state, s->plane[plane_index].band[level][orientation].qlog, 1);
1568 static void encode_header(SnowContext *s){
1572 memset(kstate, MID_STATE, sizeof(kstate));
1574 put_rac(&s->c, kstate, s->keyframe);
1575 if(s->keyframe || s->always_reset){
1576 ff_snow_reset_contexts(s);
1577 s->last_spatial_decomposition_type=
1581 s->last_block_max_depth= 0;
1582 for(plane_index=0; plane_index<2; plane_index++){
1583 Plane *p= &s->plane[plane_index];
1586 memset(p->last_hcoeff, 0, sizeof(p->last_hcoeff));
1590 put_symbol(&s->c, s->header_state, s->version, 0);
1591 put_rac(&s->c, s->header_state, s->always_reset);
1592 put_symbol(&s->c, s->header_state, s->temporal_decomposition_type, 0);
1593 put_symbol(&s->c, s->header_state, s->temporal_decomposition_count, 0);
1594 put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1595 put_symbol(&s->c, s->header_state, s->colorspace_type, 0);
1596 if (s->nb_planes > 2) {
1597 put_symbol(&s->c, s->header_state, s->chroma_h_shift, 0);
1598 put_symbol(&s->c, s->header_state, s->chroma_v_shift, 0);
1600 put_rac(&s->c, s->header_state, s->spatial_scalability);
1601 // put_rac(&s->c, s->header_state, s->rate_scalability);
1602 put_symbol(&s->c, s->header_state, s->max_ref_frames-1, 0);
1609 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1610 Plane *p= &s->plane[plane_index];
1611 update_mc |= p->last_htaps != p->htaps;
1612 update_mc |= p->last_diag_mc != p->diag_mc;
1613 update_mc |= !!memcmp(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1615 put_rac(&s->c, s->header_state, update_mc);
1617 for(plane_index=0; plane_index<FFMIN(s->nb_planes, 2); plane_index++){
1618 Plane *p= &s->plane[plane_index];
1619 put_rac(&s->c, s->header_state, p->diag_mc);
1620 put_symbol(&s->c, s->header_state, p->htaps/2-1, 0);
1621 for(i= p->htaps/2; i; i--)
1622 put_symbol(&s->c, s->header_state, FFABS(p->hcoeff[i]), 0);
1625 if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1626 put_rac(&s->c, s->header_state, 1);
1627 put_symbol(&s->c, s->header_state, s->spatial_decomposition_count, 0);
1630 put_rac(&s->c, s->header_state, 0);
1633 put_symbol(&s->c, s->header_state, s->spatial_decomposition_type - s->last_spatial_decomposition_type, 1);
1634 put_symbol(&s->c, s->header_state, s->qlog - s->last_qlog , 1);
1635 put_symbol(&s->c, s->header_state, s->mv_scale - s->last_mv_scale, 1);
1636 put_symbol(&s->c, s->header_state, s->qbias - s->last_qbias , 1);
1637 put_symbol(&s->c, s->header_state, s->block_max_depth - s->last_block_max_depth, 1);
1641 static void update_last_header_values(SnowContext *s){
1645 for(plane_index=0; plane_index<2; plane_index++){
1646 Plane *p= &s->plane[plane_index];
1647 p->last_diag_mc= p->diag_mc;
1648 p->last_htaps = p->htaps;
1649 memcpy(p->last_hcoeff, p->hcoeff, sizeof(p->hcoeff));
1653 s->last_spatial_decomposition_type = s->spatial_decomposition_type;
1654 s->last_qlog = s->qlog;
1655 s->last_qbias = s->qbias;
1656 s->last_mv_scale = s->mv_scale;
1657 s->last_block_max_depth = s->block_max_depth;
1658 s->last_spatial_decomposition_count = s->spatial_decomposition_count;
1661 static int qscale2qlog(int qscale){
1662 return lrint(QROOT*log2(qscale / (float)FF_QP2LAMBDA))
1663 + 61*QROOT/8; ///< 64 > 60
1666 static int ratecontrol_1pass(SnowEncContext *enc, AVFrame *pict)
1668 SnowContext *const s = &enc->com;
1669 /* Estimate the frame's complexity as a sum of weighted dwt coefficients.
1670 * FIXME we know exact mv bits at this point,
1671 * but ratecontrol isn't set up to include them. */
1672 uint32_t coef_sum= 0;
1673 int level, orientation, delta_qlog;
1675 for(level=0; level<s->spatial_decomposition_count; level++){
1676 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1677 SubBand *b= &s->plane[0].band[level][orientation];
1678 IDWTELEM *buf= b->ibuf;
1679 const int w= b->width;
1680 const int h= b->height;
1681 const int stride= b->stride;
1682 const int qlog= av_clip(2*QROOT + b->qlog, 0, QROOT*16);
1683 const int qmul= ff_qexp[qlog&(QROOT-1)]<<(qlog>>QSHIFT);
1684 const int qdiv= (1<<16)/qmul;
1686 //FIXME this is ugly
1689 buf[x+y*stride]= b->buf[x+y*stride];
1691 decorrelate(s, b, buf, stride, 1, 0);
1694 coef_sum+= abs(buf[x+y*stride]) * qdiv >> 16;
1698 /* ugly, ratecontrol just takes a sqrt again */
1699 av_assert0(coef_sum < INT_MAX);
1700 coef_sum = (uint64_t)coef_sum * coef_sum >> 16;
1702 if(pict->pict_type == AV_PICTURE_TYPE_I){
1703 enc->m.mb_var_sum = coef_sum;
1704 enc->m.mc_mb_var_sum = 0;
1706 enc->m.mc_mb_var_sum = coef_sum;
1707 enc->m.mb_var_sum = 0;
1710 pict->quality= ff_rate_estimate_qscale(&enc->m, 1);
1711 if (pict->quality < 0)
1713 enc->lambda= pict->quality * 3/2;
1714 delta_qlog= qscale2qlog(pict->quality) - s->qlog;
1715 s->qlog+= delta_qlog;
1719 static void calculate_visual_weight(SnowContext *s, Plane *p){
1720 int width = p->width;
1721 int height= p->height;
1722 int level, orientation, x, y;
1724 for(level=0; level<s->spatial_decomposition_count; level++){
1726 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1727 SubBand *b= &p->band[level][orientation];
1728 IDWTELEM *ibuf= b->ibuf;
1730 memset(s->spatial_idwt_buffer, 0, sizeof(*s->spatial_idwt_buffer)*width*height);
1731 ibuf[b->width/2 + b->height/2*b->stride]= 256*16;
1732 ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, width, height, width, s->spatial_decomposition_type, s->spatial_decomposition_count);
1733 for(y=0; y<height; y++){
1734 for(x=0; x<width; x++){
1735 int64_t d= s->spatial_idwt_buffer[x + y*width]*16;
1739 if (orientation == 2)
1741 b->qlog= (int)(QROOT * log2(352256.0/sqrt(error)) + 0.5);
1742 if (orientation != 1)
1745 p->band[level][1].qlog = p->band[level][2].qlog;
1749 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1750 const AVFrame *pict, int *got_packet)
1752 SnowEncContext *const enc = avctx->priv_data;
1753 SnowContext *const s = &enc->com;
1754 MpegEncContext *const mpv = &enc->m;
1755 RangeCoder * const c= &s->c;
1756 AVCodecInternal *avci = avctx->internal;
1758 const int width= s->avctx->width;
1759 const int height= s->avctx->height;
1760 int level, orientation, plane_index, i, y, ret;
1761 uint8_t rc_header_bak[sizeof(s->header_state)];
1762 uint8_t rc_block_bak[sizeof(s->block_state)];
1764 if ((ret = ff_alloc_packet(avctx, pkt, s->b_width*s->b_height*MB_SIZE*MB_SIZE*3 + AV_INPUT_BUFFER_MIN_SIZE)) < 0)
1767 ff_init_range_encoder(c, pkt->data, pkt->size);
1768 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1770 for(i=0; i < s->nb_planes; i++){
1771 int hshift= i ? s->chroma_h_shift : 0;
1772 int vshift= i ? s->chroma_v_shift : 0;
1773 for(y=0; y<AV_CEIL_RSHIFT(height, vshift); y++)
1774 memcpy(&s->input_picture->data[i][y * s->input_picture->linesize[i]],
1775 &pict->data[i][y * pict->linesize[i]],
1776 AV_CEIL_RSHIFT(width, hshift));
1777 enc->mpvencdsp.draw_edges(s->input_picture->data[i], s->input_picture->linesize[i],
1778 AV_CEIL_RSHIFT(width, hshift), AV_CEIL_RSHIFT(height, vshift),
1779 EDGE_WIDTH >> hshift, EDGE_WIDTH >> vshift,
1780 EDGE_TOP | EDGE_BOTTOM);
1784 pic = s->input_picture;
1785 pic->pict_type = pict->pict_type;
1786 pic->quality = pict->quality;
1788 mpv->picture_number = avctx->frame_num;
1789 if(avctx->flags&AV_CODEC_FLAG_PASS2){
1790 mpv->pict_type = pic->pict_type = mpv->rc_context.entry[avctx->frame_num].new_pict_type;
1791 s->keyframe = pic->pict_type == AV_PICTURE_TYPE_I;
1792 if(!(avctx->flags&AV_CODEC_FLAG_QSCALE)) {
1793 pic->quality = ff_rate_estimate_qscale(mpv, 0);
1794 if (pic->quality < 0)
1798 s->keyframe= avctx->gop_size==0 || avctx->frame_num % avctx->gop_size == 0;
1799 mpv->pict_type = pic->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
1802 if (enc->pass1_rc && avctx->frame_num == 0)
1803 pic->quality = 2*FF_QP2LAMBDA;
1805 s->qlog = qscale2qlog(pic->quality);
1806 enc->lambda = pic->quality * 3/2;
1808 if (s->qlog < 0 || (!pic->quality && (avctx->flags & AV_CODEC_FLAG_QSCALE))) {
1809 s->qlog= LOSSLESS_QLOG;
1811 }//else keep previous frame's qlog until after motion estimation
1813 if (s->current_picture->data[0]) {
1814 int w = s->avctx->width;
1815 int h = s->avctx->height;
1817 enc->mpvencdsp.draw_edges(s->current_picture->data[0],
1818 s->current_picture->linesize[0], w , h ,
1819 EDGE_WIDTH , EDGE_WIDTH , EDGE_TOP | EDGE_BOTTOM);
1820 if (s->current_picture->data[2]) {
1821 enc->mpvencdsp.draw_edges(s->current_picture->data[1],
1822 s->current_picture->linesize[1], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1823 EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1824 enc->mpvencdsp.draw_edges(s->current_picture->data[2],
1825 s->current_picture->linesize[2], w>>s->chroma_h_shift, h>>s->chroma_v_shift,
1826 EDGE_WIDTH>>s->chroma_h_shift, EDGE_WIDTH>>s->chroma_v_shift, EDGE_TOP | EDGE_BOTTOM);
1831 ff_snow_frames_prepare(s);
1832 ret = get_encode_buffer(s, s->current_picture);
1836 mpv->current_picture_ptr = &mpv->current_picture;
1837 mpv->current_picture.f = s->current_picture;
1838 mpv->current_picture.f->pts = pict->pts;
1839 if(pic->pict_type == AV_PICTURE_TYPE_P){
1840 int block_width = (width +15)>>4;
1841 int block_height= (height+15)>>4;
1842 int stride= s->current_picture->linesize[0];
1844 av_assert0(s->current_picture->data[0]);
1845 av_assert0(s->last_picture[0]->data[0]);
1847 mpv->avctx = s->avctx;
1848 mpv->last_picture.f = s->last_picture[0];
1849 mpv-> new_picture = s->input_picture;
1850 mpv->last_picture_ptr = &mpv->last_picture;
1851 mpv->linesize = stride;
1852 mpv->uvlinesize = s->current_picture->linesize[1];
1854 mpv->height = height;
1855 mpv->mb_width = block_width;
1856 mpv->mb_height = block_height;
1857 mpv->mb_stride = mpv->mb_width + 1;
1858 mpv->b8_stride = 2 * mpv->mb_width + 1;
1860 mpv->pict_type = pic->pict_type;
1861 mpv->motion_est = enc->motion_est;
1862 mpv->me.scene_change_score = 0;
1863 mpv->me.dia_size = avctx->dia_size;
1864 mpv->quarter_sample = (s->avctx->flags & AV_CODEC_FLAG_QPEL)!=0;
1865 mpv->out_format = FMT_H263;
1866 mpv->unrestricted_mv = 1;
1868 mpv->lambda = enc->lambda;
1869 mpv->qscale = (mpv->lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1870 enc->lambda2 = mpv->lambda2 = (mpv->lambda*mpv->lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1872 mpv->mecc = enc->mecc; //move
1873 mpv->qdsp = enc->qdsp; //move
1874 mpv->hdsp = s->hdsp;
1875 ff_init_me(&enc->m);
1876 s->hdsp = mpv->hdsp;
1877 enc->mecc = mpv->mecc;
1880 if (enc->pass1_rc) {
1881 memcpy(rc_header_bak, s->header_state, sizeof(s->header_state));
1882 memcpy(rc_block_bak, s->block_state, sizeof(s->block_state));
1887 s->spatial_decomposition_count= 5;
1889 while( !(width >>(s->chroma_h_shift + s->spatial_decomposition_count))
1890 || !(height>>(s->chroma_v_shift + s->spatial_decomposition_count)))
1891 s->spatial_decomposition_count--;
1893 if (s->spatial_decomposition_count <= 0) {
1894 av_log(avctx, AV_LOG_ERROR, "Resolution too low\n");
1895 return AVERROR(EINVAL);
1898 mpv->pict_type = pic->pict_type;
1899 s->qbias = pic->pict_type == AV_PICTURE_TYPE_P ? 2 : 0;
1901 ff_snow_common_init_after_header(avctx);
1903 if(s->last_spatial_decomposition_count != s->spatial_decomposition_count){
1904 for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1905 calculate_visual_weight(s, &s->plane[plane_index]);
1910 mpv->misc_bits = 8 * (s->c.bytestream - s->c.bytestream_start);
1911 encode_blocks(enc, 1);
1912 mpv->mv_bits = 8 * (s->c.bytestream - s->c.bytestream_start) - mpv->misc_bits;
1914 for(plane_index=0; plane_index < s->nb_planes; plane_index++){
1915 Plane *p= &s->plane[plane_index];
1919 // int bits= put_bits_count(&s->c.pb);
1921 if (!enc->memc_only) {
1923 if(pict->data[plane_index]) //FIXME gray hack
1926 s->spatial_idwt_buffer[y*w + x]= pict->data[plane_index][y*pict->linesize[plane_index] + x]<<FRAC_BITS;
1929 predict_plane(s, s->spatial_idwt_buffer, plane_index, 0);
1932 && pic->pict_type == AV_PICTURE_TYPE_P
1933 && !(avctx->flags&AV_CODEC_FLAG_PASS2)
1934 && mpv->me.scene_change_score > enc->scenechange_threshold) {
1935 ff_init_range_encoder(c, pkt->data, pkt->size);
1936 ff_build_rac_states(c, (1LL<<32)/20, 256-8);
1937 pic->pict_type= AV_PICTURE_TYPE_I;
1939 s->current_picture->flags |= AV_FRAME_FLAG_KEY;
1943 if(s->qlog == LOSSLESS_QLOG){
1946 s->spatial_dwt_buffer[y*w + x]= (s->spatial_idwt_buffer[y*w + x] + (1<<(FRAC_BITS-1))-1)>>FRAC_BITS;
1952 s->spatial_dwt_buffer[y*w + x]= s->spatial_idwt_buffer[y*w + x] * (1 << ENCODER_EXTRA_BITS);
1957 ff_spatial_dwt(s->spatial_dwt_buffer, s->temp_dwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1959 if (enc->pass1_rc && plane_index==0) {
1960 int delta_qlog = ratecontrol_1pass(enc, pic);
1961 if (delta_qlog <= INT_MIN)
1964 //reordering qlog in the bitstream would eliminate this reset
1965 ff_init_range_encoder(c, pkt->data, pkt->size);
1966 memcpy(s->header_state, rc_header_bak, sizeof(s->header_state));
1967 memcpy(s->block_state, rc_block_bak, sizeof(s->block_state));
1969 encode_blocks(enc, 0);
1973 for(level=0; level<s->spatial_decomposition_count; level++){
1974 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1975 SubBand *b= &p->band[level][orientation];
1977 quantize(s, b, b->ibuf, b->buf, b->stride, s->qbias);
1979 decorrelate(s, b, b->ibuf, b->stride, pic->pict_type == AV_PICTURE_TYPE_P, 0);
1980 if (!enc->no_bitstream)
1981 encode_subband(s, b, b->ibuf, b->parent ? b->parent->ibuf : NULL, b->stride, orientation);
1982 av_assert0(b->parent==NULL || b->parent->stride == b->stride*2);
1984 correlate(s, b, b->ibuf, b->stride, 1, 0);
1988 for(level=0; level<s->spatial_decomposition_count; level++){
1989 for(orientation=level ? 1 : 0; orientation<4; orientation++){
1990 SubBand *b= &p->band[level][orientation];
1992 dequantize(s, b, b->ibuf, b->stride);
1996 ff_spatial_idwt(s->spatial_idwt_buffer, s->temp_idwt_buffer, w, h, w, s->spatial_decomposition_type, s->spatial_decomposition_count);
1997 if(s->qlog == LOSSLESS_QLOG){
2000 s->spatial_idwt_buffer[y*w + x] *= 1 << FRAC_BITS;
2004 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
2007 if(pic->pict_type == AV_PICTURE_TYPE_I){
2010 s->current_picture->data[plane_index][y*s->current_picture->linesize[plane_index] + x]=
2011 pict->data[plane_index][y*pict->linesize[plane_index] + x];
2015 memset(s->spatial_idwt_buffer, 0, sizeof(IDWTELEM)*w*h);
2016 predict_plane(s, s->spatial_idwt_buffer, plane_index, 1);
2019 if(s->avctx->flags&AV_CODEC_FLAG_PSNR){
2022 if(pict->data[plane_index]) //FIXME gray hack
2025 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];
2029 s->avctx->error[plane_index] += error;
2030 enc->encoding_error[plane_index] = error;
2036 update_last_header_values(s);
2038 ff_snow_release_buffer(avctx);
2040 s->current_picture->pict_type = pic->pict_type;
2041 s->current_picture->quality = pic->quality;
2042 mpv->frame_bits = 8 * (s->c.bytestream - s->c.bytestream_start);
2043 mpv->p_tex_bits = mpv->frame_bits - mpv->misc_bits - mpv->mv_bits;
2044 mpv->total_bits += 8*(s->c.bytestream - s->c.bytestream_start);
2045 mpv->current_picture.display_picture_number =
2046 mpv->current_picture.coded_picture_number = avctx->frame_num;
2047 mpv->current_picture.f->quality = pic->quality;
2049 if (ff_rate_estimate_qscale(mpv, 0) < 0)
2051 if(avctx->flags&AV_CODEC_FLAG_PASS1)
2052 ff_write_pass1_stats(mpv);
2053 mpv->last_pict_type = mpv->pict_type;
2057 ff_side_data_set_encoder_stats(pkt, s->current_picture->quality,
2058 enc->encoding_error,
2059 (s->avctx->flags&AV_CODEC_FLAG_PSNR) ? SNOW_MAX_PLANES : 0,
2060 s->current_picture->pict_type);
2061 if (s->avctx->flags & AV_CODEC_FLAG_RECON_FRAME) {
2062 av_frame_replace(avci->recon_frame, s->current_picture);
2065 pkt->size = ff_rac_terminate(c, 0);
2066 if (s->current_picture->flags & AV_FRAME_FLAG_KEY)
2067 pkt->flags |= AV_PKT_FLAG_KEY;
2073 static av_cold int encode_end(AVCodecContext *avctx)
2075 SnowEncContext *const enc = avctx->priv_data;
2076 SnowContext *const s = &enc->com;
2078 ff_snow_common_end(s);
2079 ff_rate_control_uninit(&enc->m);
2080 av_frame_free(&s->input_picture);
2082 for (int i = 0; i < MAX_REF_FRAMES; i++) {
2083 av_freep(&s->ref_mvs[i]);
2084 av_freep(&s->ref_scores[i]);
2087 enc->m.me.temp = NULL;
2088 av_freep(&enc->m.me.scratchpad);
2089 av_freep(&enc->m.me.map);
2090 av_freep(&enc->m.sc.obmc_scratchpad);
2092 av_freep(&avctx->stats_out);
2097 #define OFFSET(x) offsetof(SnowEncContext, x)
2098 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
2099 static const AVOption options[] = {
2100 {"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" },
2101 { "zero", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ZERO }, 0, 0, VE, "motion_est" },
2102 { "epzs", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_EPZS }, 0, 0, VE, "motion_est" },
2103 { "xone", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_XONE }, 0, 0, VE, "motion_est" },
2104 { "iter", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_ME_ITER }, 0, 0, VE, "motion_est" },
2105 { "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 },
2106 { "no_bitstream", "Skip final bitstream writeout.", OFFSET(no_bitstream), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
2107 { "intra_penalty", "Penalty for intra blocks in block decission", OFFSET(intra_penalty), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
2108 { "iterative_dia_size", "Dia size for the iterative ME", OFFSET(iterative_dia_size), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
2109 { "sc_threshold", "Scene change threshold", OFFSET(scenechange_threshold), AV_OPT_TYPE_INT, { .i64 = 0 }, INT_MIN, INT_MAX, VE },
2110 { "pred", "Spatial decomposition type", OFFSET(pred), AV_OPT_TYPE_INT, { .i64 = 0 }, DWT_97, DWT_53, VE, "pred" },
2111 { "dwt97", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, INT_MIN, INT_MAX, VE, "pred" },
2112 { "dwt53", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN, INT_MAX, VE, "pred" },
2113 { "rc_eq", "Set rate control equation. When computing the expression, besides the standard functions "
2114 "defined in the section 'Expression Evaluation', the following functions are available: "
2115 "bits2qp(bits), qp2bits(qp). Also the following constants are available: iTex pTex tex mv "
2116 "fCode iCount mcVar var isI isP isB avgQP qComp avgIITex avgPITex avgPPTex avgBPTex avgTex.",
2117 OFFSET(m.rc_eq), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VE },
2121 static const AVClass snowenc_class = {
2122 .class_name = "snow encoder",
2123 .item_name = av_default_item_name,
2125 .version = LIBAVUTIL_VERSION_INT,
2128 const FFCodec ff_snow_encoder = {
2130 CODEC_LONG_NAME("Snow"),
2131 .p.type = AVMEDIA_TYPE_VIDEO,
2132 .p.id = AV_CODEC_ID_SNOW,
2133 .p.capabilities = AV_CODEC_CAP_DR1 |
2134 AV_CODEC_CAP_ENCODER_REORDERED_OPAQUE |
2135 AV_CODEC_CAP_ENCODER_RECON_FRAME,
2136 .priv_data_size = sizeof(SnowEncContext),
2137 .init = encode_init,
2138 FF_CODEC_ENCODE_CB(encode_frame),
2139 .close = encode_end,
2140 .p.pix_fmts = (const enum AVPixelFormat[]){
2141 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV444P,
2145 .p.priv_class = &snowenc_class,
2146 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,