2 * Rate control for video encoders
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
6 * This file is part of Libav.
8 * Libav is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * Libav is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with Libav; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 * Rate control for video encoders.
28 #include "libavutil/attributes.h"
30 #include "ratecontrol.h"
31 #include "mpegutils.h"
32 #include "mpegvideo.h"
33 #include "libavutil/eval.h"
35 #undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
39 #define M_E 2.718281828
42 static int init_pass2(MpegEncContext *s);
43 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
44 double rate_factor, int frame_num);
46 void ff_write_pass1_stats(MpegEncContext *s)
48 snprintf(s->avctx->stats_out, 256,
49 "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d "
50 "fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d;\n",
51 s->current_picture_ptr->f->display_picture_number,
52 s->current_picture_ptr->f->coded_picture_number,
54 s->current_picture.f->quality,
61 s->current_picture.mc_mb_var_sum,
62 s->current_picture.mb_var_sum,
63 s->i_count, s->skip_count,
67 static inline double qp2bits(RateControlEntry *rce, double qp)
70 av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
72 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp;
75 static inline double bits2qp(RateControlEntry *rce, double bits)
78 av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
80 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits;
83 av_cold int ff_rate_control_init(MpegEncContext *s)
85 RateControlContext *rcc = &s->rc_context;
87 static const char * const const_names[] = {
116 static double (* const func1[])(void *, double) = {
121 static const char * const func1_names[] = {
128 res = av_expr_parse(&rcc->rc_eq_eval,
129 s->avctx->rc_eq ? s->avctx->rc_eq : "tex^qComp",
130 const_names, func1_names, func1,
131 NULL, NULL, 0, s->avctx);
133 av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->avctx->rc_eq);
137 for (i = 0; i < 5; i++) {
138 rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0;
139 rcc->pred[i].count = 1.0;
140 rcc->pred[i].decay = 0.4;
142 rcc->i_cplx_sum [i] =
143 rcc->p_cplx_sum [i] =
144 rcc->mv_bits_sum[i] =
145 rcc->qscale_sum [i] =
146 rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such
148 rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5;
150 rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy;
152 if (s->flags & CODEC_FLAG_PASS2) {
156 /* find number of pics */
157 p = s->avctx->stats_in;
159 p = strchr(p + 1, ';');
160 i += s->max_b_frames;
161 if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry))
163 rcc->entry = av_mallocz(i * sizeof(RateControlEntry));
164 rcc->num_entries = i;
166 /* init all to skipped p frames
167 * (with b frames we might have a not encoded frame at the end FIXME) */
168 for (i = 0; i < rcc->num_entries; i++) {
169 RateControlEntry *rce = &rcc->entry[i];
171 rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P;
172 rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2;
173 rce->misc_bits = s->mb_num + 10;
174 rce->mb_var_sum = s->mb_num * 100;
178 p = s->avctx->stats_in;
179 for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
180 RateControlEntry *rce;
185 next = strchr(p, ';');
187 (*next) = 0; // sscanf in unbelievably slow on looong strings // FIXME copy / do not write
190 e = sscanf(p, " in:%d ", &picture_number);
192 assert(picture_number >= 0);
193 assert(picture_number < rcc->num_entries);
194 rce = &rcc->entry[picture_number];
196 e += sscanf(p, " in:%*d out:%*d type:%d q:%f itex:%d ptex:%d mv:%d misc:%d fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d",
197 &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
198 &rce->mv_bits, &rce->misc_bits,
199 &rce->f_code, &rce->b_code,
200 &rce->mc_mb_var_sum, &rce->mb_var_sum,
201 &rce->i_count, &rce->skip_count, &rce->header_bits);
203 av_log(s->avctx, AV_LOG_ERROR,
204 "statistics are damaged at line %d, parser out=%d\n",
212 if (init_pass2(s) < 0)
215 // FIXME maybe move to end
216 if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) {
218 return ff_xvid_rate_control_init(s);
220 av_log(s->avctx, AV_LOG_ERROR,
221 "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n");
227 if (!(s->flags & CODEC_FLAG_PASS2)) {
228 rcc->short_term_qsum = 0.001;
229 rcc->short_term_qcount = 0.001;
231 rcc->pass1_rc_eq_output_sum = 0.001;
232 rcc->pass1_wanted_bits = 0.001;
234 if (s->avctx->qblur > 1.0) {
235 av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
238 /* init stuff with the user specified complexity */
239 if (s->avctx->rc_initial_cplx) {
240 for (i = 0; i < 60 * 30; i++) {
241 double bits = s->avctx->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
242 RateControlEntry rce;
244 if (i % ((s->gop_size + 3) / 4) == 0)
245 rce.pict_type = AV_PICTURE_TYPE_I;
246 else if (i % (s->max_b_frames + 1))
247 rce.pict_type = AV_PICTURE_TYPE_B;
249 rce.pict_type = AV_PICTURE_TYPE_P;
251 rce.new_pict_type = rce.pict_type;
252 rce.mc_mb_var_sum = bits * s->mb_num / 100000;
253 rce.mb_var_sum = s->mb_num;
255 rce.qscale = FF_QP2LAMBDA * 2;
260 if (s->pict_type == AV_PICTURE_TYPE_I) {
261 rce.i_count = s->mb_num;
262 rce.i_tex_bits = bits;
266 rce.i_count = 0; // FIXME we do know this approx
268 rce.p_tex_bits = bits * 0.9;
269 rce.mv_bits = bits * 0.1;
271 rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale;
272 rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale;
273 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
274 rcc->frame_count[rce.pict_type]++;
276 get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
278 // FIXME misbehaves a little for variable fps
279 rcc->pass1_wanted_bits += s->bit_rate / (1 / av_q2d(s->avctx->time_base));
287 av_cold void ff_rate_control_uninit(MpegEncContext *s)
289 RateControlContext *rcc = &s->rc_context;
292 av_expr_free(rcc->rc_eq_eval);
293 av_freep(&rcc->entry);
296 if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
297 ff_xvid_rate_control_uninit(s);
301 int ff_vbv_update(MpegEncContext *s, int frame_size)
303 RateControlContext *rcc = &s->rc_context;
304 const double fps = 1 / av_q2d(s->avctx->time_base);
305 const int buffer_size = s->avctx->rc_buffer_size;
306 const double min_rate = s->avctx->rc_min_rate / fps;
307 const double max_rate = s->avctx->rc_max_rate / fps;
309 av_dlog(s, "%d %f %d %f %f\n",
310 buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
315 rcc->buffer_index -= frame_size;
316 if (rcc->buffer_index < 0) {
317 av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
318 rcc->buffer_index = 0;
321 left = buffer_size - rcc->buffer_index - 1;
322 rcc->buffer_index += av_clip(left, min_rate, max_rate);
324 if (rcc->buffer_index > buffer_size) {
325 int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
327 if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
329 rcc->buffer_index -= 8 * stuffing;
331 if (s->avctx->debug & FF_DEBUG_RC)
332 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
341 * Modify the bitrate curve from pass1 for one frame.
343 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
344 double rate_factor, int frame_num)
346 RateControlContext *rcc = &s->rc_context;
347 AVCodecContext *a = s->avctx;
348 const int pict_type = rce->new_pict_type;
349 const double mb_num = s->mb_num;
353 double const_values[] = {
356 rce->i_tex_bits * rce->qscale,
357 rce->p_tex_bits * rce->qscale,
358 (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
359 rce->mv_bits / mb_num,
360 rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
361 rce->i_count / mb_num,
362 rce->mc_mb_var_sum / mb_num,
363 rce->mb_var_sum / mb_num,
364 rce->pict_type == AV_PICTURE_TYPE_I,
365 rce->pict_type == AV_PICTURE_TYPE_P,
366 rce->pict_type == AV_PICTURE_TYPE_B,
367 rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
370 rcc->last_qscale_for[AV_PICTURE_TYPE_I],
371 rcc->last_qscale_for[AV_PICTURE_TYPE_P],
372 rcc->last_qscale_for[AV_PICTURE_TYPE_B],
373 rcc->next_non_b_qscale,
375 rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
376 rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
377 rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
378 rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
379 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
383 bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
385 av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq);
389 rcc->pass1_rc_eq_output_sum += bits;
393 bits += 1.0; // avoid 1/0 issues
396 for (i = 0; i < s->avctx->rc_override_count; i++) {
397 RcOverride *rco = s->avctx->rc_override;
398 if (rco[i].start_frame > frame_num)
400 if (rco[i].end_frame < frame_num)
404 bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it?
406 bits *= rco[i].quality_factor;
409 q = bits2qp(rce, bits);
412 if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
413 q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
414 else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
415 q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
422 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
424 RateControlContext *rcc = &s->rc_context;
425 AVCodecContext *a = s->avctx;
426 const int pict_type = rce->new_pict_type;
427 const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
428 const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];
430 if (pict_type == AV_PICTURE_TYPE_I &&
431 (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P))
432 q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
433 else if (pict_type == AV_PICTURE_TYPE_B &&
434 a->b_quant_factor > 0.0)
435 q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
439 /* last qscale / qdiff stuff */
440 if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
441 double last_q = rcc->last_qscale_for[pict_type];
442 const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;
444 if (q > last_q + maxdiff)
445 q = last_q + maxdiff;
446 else if (q < last_q - maxdiff)
447 q = last_q - maxdiff;
450 rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring
452 if (pict_type != AV_PICTURE_TYPE_B)
453 rcc->last_non_b_pict_type = pict_type;
459 * Get the qmin & qmax for pict_type.
461 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
463 int qmin = s->avctx->lmin;
464 int qmax = s->avctx->lmax;
466 assert(qmin <= qmax);
469 case AV_PICTURE_TYPE_B:
470 qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
471 qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
473 case AV_PICTURE_TYPE_I:
474 qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
475 qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
479 qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
480 qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);
489 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce,
490 double q, int frame_num)
492 RateControlContext *rcc = &s->rc_context;
493 const double buffer_size = s->avctx->rc_buffer_size;
494 const double fps = 1 / av_q2d(s->avctx->time_base);
495 const double min_rate = s->avctx->rc_min_rate / fps;
496 const double max_rate = s->avctx->rc_max_rate / fps;
497 const int pict_type = rce->new_pict_type;
500 get_qminmax(&qmin, &qmax, s, pict_type);
503 if (s->avctx->rc_qmod_freq &&
504 frame_num % s->avctx->rc_qmod_freq == 0 &&
505 pict_type == AV_PICTURE_TYPE_P)
506 q *= s->avctx->rc_qmod_amp;
508 /* buffer overflow/underflow protection */
510 double expected_size = rcc->buffer_index;
514 double d = 2 * (buffer_size - expected_size) / buffer_size;
519 q *= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity);
521 q_limit = bits2qp(rce,
522 FFMAX((min_rate - buffer_size + rcc->buffer_index) *
523 s->avctx->rc_min_vbv_overflow_use, 1));
526 if (s->avctx->debug & FF_DEBUG_RC)
527 av_log(s->avctx, AV_LOG_DEBUG,
528 "limiting QP %f -> %f\n", q, q_limit);
534 double d = 2 * expected_size / buffer_size;
539 q /= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity);
541 q_limit = bits2qp(rce,
542 FFMAX(rcc->buffer_index *
543 s->avctx->rc_max_available_vbv_use,
546 if (s->avctx->debug & FF_DEBUG_RC)
547 av_log(s->avctx, AV_LOG_DEBUG,
548 "limiting QP %f -> %f\n", q, q_limit);
553 av_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
554 q, max_rate, min_rate, buffer_size, rcc->buffer_index,
555 s->avctx->rc_buffer_aggressivity);
556 if (s->avctx->rc_qsquish == 0.0 || qmin == qmax) {
562 double min2 = log(qmin);
563 double max2 = log(qmax);
566 q = (q - min2) / (max2 - min2) - 0.5;
568 q = 1.0 / (1.0 + exp(q));
569 q = q * (max2 - min2) + min2;
577 // ----------------------------------
580 static double predict_size(Predictor *p, double q, double var)
582 return p->coeff * var / (q * p->count);
585 static void update_predictor(Predictor *p, double q, double var, double size)
587 double new_coeff = size * q / (var + 1);
591 p->count *= p->decay;
592 p->coeff *= p->decay;
594 p->coeff += new_coeff;
597 static void adaptive_quantization(MpegEncContext *s, double q)
600 const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0);
601 const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0);
602 const float temp_cplx_masking = s->avctx->temporal_cplx_masking;
603 const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
604 const float p_masking = s->avctx->p_masking;
605 const float border_masking = s->avctx->border_masking;
606 float bits_sum = 0.0;
607 float cplx_sum = 0.0;
608 float *cplx_tab = s->cplx_tab;
609 float *bits_tab = s->bits_tab;
610 const int qmin = s->avctx->mb_lmin;
611 const int qmax = s->avctx->mb_lmax;
612 Picture *const pic = &s->current_picture;
613 const int mb_width = s->mb_width;
614 const int mb_height = s->mb_height;
616 for (i = 0; i < s->mb_num; i++) {
617 const int mb_xy = s->mb_index2xy[i];
618 float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
619 float spat_cplx = sqrt(pic->mb_var[mb_xy]);
620 const int lumi = pic->mb_mean[mb_xy];
621 float bits, cplx, factor;
622 int mb_x = mb_xy % s->mb_stride;
623 int mb_y = mb_xy / s->mb_stride;
625 float mb_factor = 0.0;
627 spat_cplx = 4; // FIXME finetune
629 temp_cplx = 4; // FIXME finetune
631 if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
633 factor = 1.0 + p_masking;
636 factor = pow(temp_cplx, -temp_cplx_masking);
638 factor *= pow(spat_cplx, -spatial_cplx_masking);
641 factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
643 factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);
645 if (mb_x < mb_width / 5) {
646 mb_distance = mb_width / 5 - mb_x;
647 mb_factor = (float)mb_distance / (float)(mb_width / 5);
648 } else if (mb_x > 4 * mb_width / 5) {
649 mb_distance = mb_x - 4 * mb_width / 5;
650 mb_factor = (float)mb_distance / (float)(mb_width / 5);
652 if (mb_y < mb_height / 5) {
653 mb_distance = mb_height / 5 - mb_y;
654 mb_factor = FFMAX(mb_factor,
655 (float)mb_distance / (float)(mb_height / 5));
656 } else if (mb_y > 4 * mb_height / 5) {
657 mb_distance = mb_y - 4 * mb_height / 5;
658 mb_factor = FFMAX(mb_factor,
659 (float)mb_distance / (float)(mb_height / 5));
662 factor *= 1.0 - border_masking * mb_factor;
664 if (factor < 0.00001)
667 bits = cplx * factor;
674 /* handle qmin/qmax clipping */
675 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
676 float factor = bits_sum / cplx_sum;
677 for (i = 0; i < s->mb_num; i++) {
678 float newq = q * cplx_tab[i] / bits_tab[i];
682 bits_sum -= bits_tab[i];
683 cplx_sum -= cplx_tab[i] * q / qmax;
684 } else if (newq < qmin) {
685 bits_sum -= bits_tab[i];
686 cplx_sum -= cplx_tab[i] * q / qmin;
689 if (bits_sum < 0.001)
691 if (cplx_sum < 0.001)
695 for (i = 0; i < s->mb_num; i++) {
696 const int mb_xy = s->mb_index2xy[i];
697 float newq = q * cplx_tab[i] / bits_tab[i];
700 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
701 newq *= bits_sum / cplx_sum;
704 intq = (int)(newq + 0.5);
708 else if (intq < qmin)
710 s->lambda_table[mb_xy] = intq;
714 void ff_get_2pass_fcode(MpegEncContext *s)
716 RateControlContext *rcc = &s->rc_context;
717 RateControlEntry *rce = &rcc->entry[s->picture_number];
719 s->f_code = rce->f_code;
720 s->b_code = rce->b_code;
723 // FIXME rd or at least approx for dquant
725 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
729 float br_compensation;
733 int picture_number = s->picture_number;
735 RateControlContext *rcc = &s->rc_context;
736 AVCodecContext *a = s->avctx;
737 RateControlEntry local_rce, *rce;
741 const int pict_type = s->pict_type;
742 Picture * const pic = &s->current_picture;
746 if ((s->flags & CODEC_FLAG_PASS2) &&
747 s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
748 return ff_xvid_rate_estimate_qscale(s, dry_run);
751 get_qminmax(&qmin, &qmax, s, pict_type);
753 fps = 1 / av_q2d(s->avctx->time_base);
754 /* update predictors */
755 if (picture_number > 2 && !dry_run) {
756 const int last_var = s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
757 : rcc->last_mc_mb_var_sum;
758 update_predictor(&rcc->pred[s->last_pict_type],
760 sqrt(last_var), s->frame_bits);
763 if (s->flags & CODEC_FLAG_PASS2) {
764 assert(picture_number >= 0);
765 assert(picture_number < rcc->num_entries);
766 rce = &rcc->entry[picture_number];
767 wanted_bits = rce->expected_bits;
772 /* FIXME add a dts field to AVFrame and ensure it is set and use it
773 * here instead of reordering but the reordering is simpler for now
774 * until H.264 B-pyramid must be handled. */
775 if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
776 dts_pic = s->current_picture_ptr;
778 dts_pic = s->last_picture_ptr;
780 if (!dts_pic || dts_pic->f->pts == AV_NOPTS_VALUE)
781 wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
783 wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f->pts / fps);
786 diff = s->total_bits - wanted_bits;
787 br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
788 if (br_compensation <= 0.0)
789 br_compensation = 0.001;
791 var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
793 short_term_q = 0; /* avoid warning */
794 if (s->flags & CODEC_FLAG_PASS2) {
795 if (pict_type != AV_PICTURE_TYPE_I)
796 assert(pict_type == rce->new_pict_type);
798 q = rce->new_qscale / br_compensation;
799 av_dlog(s, "%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale,
800 br_compensation, s->frame_bits, var, pict_type);
803 rce->new_pict_type = pict_type;
804 rce->mc_mb_var_sum = pic->mc_mb_var_sum;
805 rce->mb_var_sum = pic->mb_var_sum;
806 rce->qscale = FF_QP2LAMBDA * 2;
807 rce->f_code = s->f_code;
808 rce->b_code = s->b_code;
811 bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
812 if (pict_type == AV_PICTURE_TYPE_I) {
813 rce->i_count = s->mb_num;
814 rce->i_tex_bits = bits;
818 rce->i_count = 0; // FIXME we do know this approx
820 rce->p_tex_bits = bits * 0.9;
821 rce->mv_bits = bits * 0.1;
823 rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale;
824 rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale;
825 rcc->mv_bits_sum[pict_type] += rce->mv_bits;
826 rcc->frame_count[pict_type]++;
828 bits = rce->i_tex_bits + rce->p_tex_bits;
829 rate_factor = rcc->pass1_wanted_bits /
830 rcc->pass1_rc_eq_output_sum * br_compensation;
832 q = get_qscale(s, rce, rate_factor, picture_number);
837 q = get_diff_limited_q(s, rce, q);
840 // FIXME type dependent blur like in 2-pass
841 if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
842 rcc->short_term_qsum *= a->qblur;
843 rcc->short_term_qcount *= a->qblur;
845 rcc->short_term_qsum += q;
846 rcc->short_term_qcount++;
847 q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
851 q = modify_qscale(s, rce, q, picture_number);
853 rcc->pass1_wanted_bits += s->bit_rate / fps;
858 if (s->avctx->debug & FF_DEBUG_RC) {
859 av_log(s->avctx, AV_LOG_DEBUG,
860 "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
861 "size:%d var:%d/%d br:%d fps:%d\n",
862 av_get_picture_type_char(pict_type),
863 qmin, q, qmax, picture_number,
864 (int)wanted_bits / 1000, (int)s->total_bits / 1000,
865 br_compensation, short_term_q, s->frame_bits,
866 pic->mb_var_sum, pic->mc_mb_var_sum,
867 s->bit_rate / 1000, (int)fps);
875 if (s->adaptive_quant)
876 adaptive_quantization(s, q);
881 rcc->last_qscale = q;
882 rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
883 rcc->last_mb_var_sum = pic->mb_var_sum;
888 // ----------------------------------------------
891 static int init_pass2(MpegEncContext *s)
893 RateControlContext *rcc = &s->rc_context;
894 AVCodecContext *a = s->avctx;
896 double fps = 1 / av_q2d(s->avctx->time_base);
897 double complexity[5] = { 0 }; // approximate bits at quant=1
898 uint64_t const_bits[5] = { 0 }; // quantizer independent bits
899 uint64_t all_const_bits;
900 uint64_t all_available_bits = (uint64_t)(s->bit_rate *
901 (double)rcc->num_entries / fps);
902 double rate_factor = 0;
904 const int filter_size = (int)(a->qblur * 4) | 1;
905 double expected_bits;
906 double *qscale, *blurred_qscale, qscale_sum;
908 /* find complexity & const_bits & decide the pict_types */
909 for (i = 0; i < rcc->num_entries; i++) {
910 RateControlEntry *rce = &rcc->entry[i];
912 rce->new_pict_type = rce->pict_type;
913 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
914 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
915 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
916 rcc->frame_count[rce->pict_type]++;
918 complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
920 const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
923 all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
924 const_bits[AV_PICTURE_TYPE_P] +
925 const_bits[AV_PICTURE_TYPE_B];
927 if (all_available_bits < all_const_bits) {
928 av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
932 qscale = av_malloc(sizeof(double) * rcc->num_entries);
933 blurred_qscale = av_malloc(sizeof(double) * rcc->num_entries);
936 for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
940 rcc->buffer_index = s->avctx->rc_buffer_size / 2;
943 for (i = 0; i < rcc->num_entries; i++) {
944 RateControlEntry *rce = &rcc->entry[i];
946 qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
947 rcc->last_qscale_for[rce->pict_type] = qscale[i];
949 assert(filter_size % 2 == 1);
951 /* fixed I/B QP relative to P mode */
952 for (i = rcc->num_entries - 1; i >= 0; i--) {
953 RateControlEntry *rce = &rcc->entry[i];
955 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
959 for (i = 0; i < rcc->num_entries; i++) {
960 RateControlEntry *rce = &rcc->entry[i];
961 const int pict_type = rce->new_pict_type;
963 double q = 0.0, sum = 0.0;
965 for (j = 0; j < filter_size; j++) {
966 int index = i + j - filter_size / 2;
967 double d = index - i;
968 double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));
970 if (index < 0 || index >= rcc->num_entries)
972 if (pict_type != rcc->entry[index].new_pict_type)
974 q += qscale[index] * coeff;
977 blurred_qscale[i] = q / sum;
980 /* find expected bits */
981 for (i = 0; i < rcc->num_entries; i++) {
982 RateControlEntry *rce = &rcc->entry[i];
985 rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);
987 bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
988 bits += 8 * ff_vbv_update(s, bits);
990 rce->expected_bits = expected_bits;
991 expected_bits += bits;
995 "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
996 expected_bits, (int)all_available_bits, rate_factor);
997 if (expected_bits > all_available_bits) {
1003 av_free(blurred_qscale);
1005 /* check bitrate calculations and print info */
1007 for (i = 0; i < rcc->num_entries; i++) {
1008 av_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
1010 rcc->entry[i].new_qscale,
1011 rcc->entry[i].new_qscale / FF_QP2LAMBDA);
1012 qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
1013 s->avctx->qmin, s->avctx->qmax);
1015 assert(toobig <= 40);
1016 av_log(s->avctx, AV_LOG_DEBUG,
1017 "[lavc rc] requested bitrate: %d bps expected bitrate: %d bps\n",
1019 (int)(expected_bits / ((double)all_available_bits / s->bit_rate)));
1020 av_log(s->avctx, AV_LOG_DEBUG,
1021 "[lavc rc] estimated target average qp: %.3f\n",
1022 (float)qscale_sum / rcc->num_entries);
1024 av_log(s->avctx, AV_LOG_INFO,
1025 "[lavc rc] Using all of requested bitrate is not "
1026 "necessary for this video with these parameters.\n");
1027 } else if (toobig == 40) {
1028 av_log(s->avctx, AV_LOG_ERROR,
1029 "[lavc rc] Error: bitrate too low for this video "
1030 "with these parameters.\n");
1032 } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
1033 av_log(s->avctx, AV_LOG_ERROR,
1034 "[lavc rc] Error: 2pass curve failed to converge\n");