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"
31 #include "ratecontrol.h"
32 #include "mpegutils.h"
33 #include "mpegvideo.h"
34 #include "libavutil/eval.h"
36 #undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
40 #define M_E 2.718281828
43 static int init_pass2(MpegEncContext *s);
44 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
45 double rate_factor, int frame_num);
47 void ff_write_pass1_stats(MpegEncContext *s)
49 snprintf(s->avctx->stats_out, 256,
50 "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d "
51 "fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d;\n",
52 s->current_picture_ptr->f->display_picture_number,
53 s->current_picture_ptr->f->coded_picture_number,
55 s->current_picture.f->quality,
62 s->current_picture.mc_mb_var_sum,
63 s->current_picture.mb_var_sum,
64 s->i_count, s->skip_count,
68 static inline double qp2bits(RateControlEntry *rce, double qp)
71 av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
73 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp;
76 static inline double bits2qp(RateControlEntry *rce, double bits)
79 av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
81 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits;
84 av_cold int ff_rate_control_init(MpegEncContext *s)
86 RateControlContext *rcc = &s->rc_context;
88 static const char * const const_names[] = {
117 static double (* const func1[])(void *, double) = {
122 static const char * const func1_names[] = {
129 res = av_expr_parse(&rcc->rc_eq_eval,
130 s->rc_eq ? s->rc_eq : "tex^qComp",
131 const_names, func1_names, func1,
132 NULL, NULL, 0, s->avctx);
134 av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->rc_eq);
138 for (i = 0; i < 5; i++) {
139 rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0;
140 rcc->pred[i].count = 1.0;
141 rcc->pred[i].decay = 0.4;
143 rcc->i_cplx_sum [i] =
144 rcc->p_cplx_sum [i] =
145 rcc->mv_bits_sum[i] =
146 rcc->qscale_sum [i] =
147 rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such
149 rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5;
151 rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy;
153 if (s->avctx->flags & CODEC_FLAG_PASS2) {
157 /* find number of pics */
158 p = s->avctx->stats_in;
160 p = strchr(p + 1, ';');
161 i += s->max_b_frames;
162 if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry))
164 rcc->entry = av_mallocz(i * sizeof(RateControlEntry));
165 rcc->num_entries = i;
167 return AVERROR(ENOMEM);
169 /* init all to skipped p frames
170 * (with b frames we might have a not encoded frame at the end FIXME) */
171 for (i = 0; i < rcc->num_entries; i++) {
172 RateControlEntry *rce = &rcc->entry[i];
174 rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P;
175 rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2;
176 rce->misc_bits = s->mb_num + 10;
177 rce->mb_var_sum = s->mb_num * 100;
181 p = s->avctx->stats_in;
182 for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
183 RateControlEntry *rce;
188 next = strchr(p, ';');
190 (*next) = 0; // sscanf in unbelievably slow on looong strings // FIXME copy / do not write
193 e = sscanf(p, " in:%d ", &picture_number);
195 assert(picture_number >= 0);
196 assert(picture_number < rcc->num_entries);
197 rce = &rcc->entry[picture_number];
199 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",
200 &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
201 &rce->mv_bits, &rce->misc_bits,
202 &rce->f_code, &rce->b_code,
203 &rce->mc_mb_var_sum, &rce->mb_var_sum,
204 &rce->i_count, &rce->skip_count, &rce->header_bits);
206 av_log(s->avctx, AV_LOG_ERROR,
207 "statistics are damaged at line %d, parser out=%d\n",
215 if (init_pass2(s) < 0) {
216 ff_rate_control_uninit(s);
220 // FIXME maybe move to end
221 if ((s->avctx->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) {
223 return ff_xvid_rate_control_init(s);
225 av_log(s->avctx, AV_LOG_ERROR,
226 "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n");
232 if (!(s->avctx->flags & CODEC_FLAG_PASS2)) {
233 rcc->short_term_qsum = 0.001;
234 rcc->short_term_qcount = 0.001;
236 rcc->pass1_rc_eq_output_sum = 0.001;
237 rcc->pass1_wanted_bits = 0.001;
239 if (s->avctx->qblur > 1.0) {
240 av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
243 /* init stuff with the user specified complexity */
244 if (s->rc_initial_cplx) {
245 for (i = 0; i < 60 * 30; i++) {
246 double bits = s->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
247 RateControlEntry rce;
249 if (i % ((s->gop_size + 3) / 4) == 0)
250 rce.pict_type = AV_PICTURE_TYPE_I;
251 else if (i % (s->max_b_frames + 1))
252 rce.pict_type = AV_PICTURE_TYPE_B;
254 rce.pict_type = AV_PICTURE_TYPE_P;
256 rce.new_pict_type = rce.pict_type;
257 rce.mc_mb_var_sum = bits * s->mb_num / 100000;
258 rce.mb_var_sum = s->mb_num;
260 rce.qscale = FF_QP2LAMBDA * 2;
265 if (s->pict_type == AV_PICTURE_TYPE_I) {
266 rce.i_count = s->mb_num;
267 rce.i_tex_bits = bits;
271 rce.i_count = 0; // FIXME we do know this approx
273 rce.p_tex_bits = bits * 0.9;
274 rce.mv_bits = bits * 0.1;
276 rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale;
277 rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale;
278 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
279 rcc->frame_count[rce.pict_type]++;
281 get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
283 // FIXME misbehaves a little for variable fps
284 rcc->pass1_wanted_bits += s->bit_rate / (1 / av_q2d(s->avctx->time_base));
292 av_cold void ff_rate_control_uninit(MpegEncContext *s)
294 RateControlContext *rcc = &s->rc_context;
297 av_expr_free(rcc->rc_eq_eval);
298 av_freep(&rcc->entry);
301 if ((s->avctx->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
302 ff_xvid_rate_control_uninit(s);
306 int ff_vbv_update(MpegEncContext *s, int frame_size)
308 RateControlContext *rcc = &s->rc_context;
309 const double fps = 1 / av_q2d(s->avctx->time_base);
310 const int buffer_size = s->avctx->rc_buffer_size;
311 const double min_rate = s->avctx->rc_min_rate / fps;
312 const double max_rate = s->avctx->rc_max_rate / fps;
314 ff_dlog(s, "%d %f %d %f %f\n",
315 buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
320 rcc->buffer_index -= frame_size;
321 if (rcc->buffer_index < 0) {
322 av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
323 rcc->buffer_index = 0;
326 left = buffer_size - rcc->buffer_index - 1;
327 rcc->buffer_index += av_clip(left, min_rate, max_rate);
329 if (rcc->buffer_index > buffer_size) {
330 int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
332 if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
334 rcc->buffer_index -= 8 * stuffing;
336 if (s->avctx->debug & FF_DEBUG_RC)
337 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
346 * Modify the bitrate curve from pass1 for one frame.
348 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
349 double rate_factor, int frame_num)
351 RateControlContext *rcc = &s->rc_context;
352 AVCodecContext *a = s->avctx;
353 const int pict_type = rce->new_pict_type;
354 const double mb_num = s->mb_num;
358 double const_values[] = {
361 rce->i_tex_bits * rce->qscale,
362 rce->p_tex_bits * rce->qscale,
363 (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
364 rce->mv_bits / mb_num,
365 rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
366 rce->i_count / mb_num,
367 rce->mc_mb_var_sum / mb_num,
368 rce->mb_var_sum / mb_num,
369 rce->pict_type == AV_PICTURE_TYPE_I,
370 rce->pict_type == AV_PICTURE_TYPE_P,
371 rce->pict_type == AV_PICTURE_TYPE_B,
372 rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
375 rcc->last_qscale_for[AV_PICTURE_TYPE_I],
376 rcc->last_qscale_for[AV_PICTURE_TYPE_P],
377 rcc->last_qscale_for[AV_PICTURE_TYPE_B],
378 rcc->next_non_b_qscale,
380 rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
381 rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
382 rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
383 rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
384 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
388 bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
390 av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->rc_eq);
394 rcc->pass1_rc_eq_output_sum += bits;
398 bits += 1.0; // avoid 1/0 issues
401 for (i = 0; i < s->avctx->rc_override_count; i++) {
402 RcOverride *rco = s->avctx->rc_override;
403 if (rco[i].start_frame > frame_num)
405 if (rco[i].end_frame < frame_num)
409 bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it?
411 bits *= rco[i].quality_factor;
414 q = bits2qp(rce, bits);
417 if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
418 q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
419 else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
420 q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
427 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
429 RateControlContext *rcc = &s->rc_context;
430 AVCodecContext *a = s->avctx;
431 const int pict_type = rce->new_pict_type;
432 const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
433 const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];
435 if (pict_type == AV_PICTURE_TYPE_I &&
436 (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P))
437 q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
438 else if (pict_type == AV_PICTURE_TYPE_B &&
439 a->b_quant_factor > 0.0)
440 q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
444 /* last qscale / qdiff stuff */
445 if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
446 double last_q = rcc->last_qscale_for[pict_type];
447 const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;
449 if (q > last_q + maxdiff)
450 q = last_q + maxdiff;
451 else if (q < last_q - maxdiff)
452 q = last_q - maxdiff;
455 rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring
457 if (pict_type != AV_PICTURE_TYPE_B)
458 rcc->last_non_b_pict_type = pict_type;
464 * Get the qmin & qmax for pict_type.
466 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
471 assert(qmin <= qmax);
474 case AV_PICTURE_TYPE_B:
475 qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
476 qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
478 case AV_PICTURE_TYPE_I:
479 qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
480 qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
484 qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
485 qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);
494 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce,
495 double q, int frame_num)
497 RateControlContext *rcc = &s->rc_context;
498 const double buffer_size = s->avctx->rc_buffer_size;
499 const double fps = 1 / av_q2d(s->avctx->time_base);
500 const double min_rate = s->avctx->rc_min_rate / fps;
501 const double max_rate = s->avctx->rc_max_rate / fps;
502 const int pict_type = rce->new_pict_type;
505 get_qminmax(&qmin, &qmax, s, pict_type);
508 if (s->rc_qmod_freq &&
509 frame_num % s->rc_qmod_freq == 0 &&
510 pict_type == AV_PICTURE_TYPE_P)
513 /* buffer overflow/underflow protection */
515 double expected_size = rcc->buffer_index;
519 double d = 2 * (buffer_size - expected_size) / buffer_size;
524 q *= pow(d, 1.0 / s->rc_buffer_aggressivity);
526 q_limit = bits2qp(rce,
527 FFMAX((min_rate - buffer_size + rcc->buffer_index) *
528 s->avctx->rc_min_vbv_overflow_use, 1));
531 if (s->avctx->debug & FF_DEBUG_RC)
532 av_log(s->avctx, AV_LOG_DEBUG,
533 "limiting QP %f -> %f\n", q, q_limit);
539 double d = 2 * expected_size / buffer_size;
544 q /= pow(d, 1.0 / s->rc_buffer_aggressivity);
546 q_limit = bits2qp(rce,
547 FFMAX(rcc->buffer_index *
548 s->avctx->rc_max_available_vbv_use,
551 if (s->avctx->debug & FF_DEBUG_RC)
552 av_log(s->avctx, AV_LOG_DEBUG,
553 "limiting QP %f -> %f\n", q, q_limit);
558 ff_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
559 q, max_rate, min_rate, buffer_size, rcc->buffer_index,
560 s->rc_buffer_aggressivity);
561 if (s->rc_qsquish == 0.0 || qmin == qmax) {
567 double min2 = log(qmin);
568 double max2 = log(qmax);
571 q = (q - min2) / (max2 - min2) - 0.5;
573 q = 1.0 / (1.0 + exp(q));
574 q = q * (max2 - min2) + min2;
582 // ----------------------------------
585 static double predict_size(Predictor *p, double q, double var)
587 return p->coeff * var / (q * p->count);
590 static void update_predictor(Predictor *p, double q, double var, double size)
592 double new_coeff = size * q / (var + 1);
596 p->count *= p->decay;
597 p->coeff *= p->decay;
599 p->coeff += new_coeff;
602 static void adaptive_quantization(MpegEncContext *s, double q)
605 const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0);
606 const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0);
607 const float temp_cplx_masking = s->avctx->temporal_cplx_masking;
608 const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
609 const float p_masking = s->avctx->p_masking;
610 const float border_masking = s->border_masking;
611 float bits_sum = 0.0;
612 float cplx_sum = 0.0;
613 float *cplx_tab = s->cplx_tab;
614 float *bits_tab = s->bits_tab;
615 const int qmin = s->avctx->mb_lmin;
616 const int qmax = s->avctx->mb_lmax;
617 Picture *const pic = &s->current_picture;
618 const int mb_width = s->mb_width;
619 const int mb_height = s->mb_height;
621 for (i = 0; i < s->mb_num; i++) {
622 const int mb_xy = s->mb_index2xy[i];
623 float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
624 float spat_cplx = sqrt(pic->mb_var[mb_xy]);
625 const int lumi = pic->mb_mean[mb_xy];
626 float bits, cplx, factor;
627 int mb_x = mb_xy % s->mb_stride;
628 int mb_y = mb_xy / s->mb_stride;
630 float mb_factor = 0.0;
632 spat_cplx = 4; // FIXME finetune
634 temp_cplx = 4; // FIXME finetune
636 if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
638 factor = 1.0 + p_masking;
641 factor = pow(temp_cplx, -temp_cplx_masking);
643 factor *= pow(spat_cplx, -spatial_cplx_masking);
646 factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
648 factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);
650 if (mb_x < mb_width / 5) {
651 mb_distance = mb_width / 5 - mb_x;
652 mb_factor = (float)mb_distance / (float)(mb_width / 5);
653 } else if (mb_x > 4 * mb_width / 5) {
654 mb_distance = mb_x - 4 * mb_width / 5;
655 mb_factor = (float)mb_distance / (float)(mb_width / 5);
657 if (mb_y < mb_height / 5) {
658 mb_distance = mb_height / 5 - mb_y;
659 mb_factor = FFMAX(mb_factor,
660 (float)mb_distance / (float)(mb_height / 5));
661 } else if (mb_y > 4 * mb_height / 5) {
662 mb_distance = mb_y - 4 * mb_height / 5;
663 mb_factor = FFMAX(mb_factor,
664 (float)mb_distance / (float)(mb_height / 5));
667 factor *= 1.0 - border_masking * mb_factor;
669 if (factor < 0.00001)
672 bits = cplx * factor;
679 /* handle qmin/qmax clipping */
680 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
681 float factor = bits_sum / cplx_sum;
682 for (i = 0; i < s->mb_num; i++) {
683 float newq = q * cplx_tab[i] / bits_tab[i];
687 bits_sum -= bits_tab[i];
688 cplx_sum -= cplx_tab[i] * q / qmax;
689 } else if (newq < qmin) {
690 bits_sum -= bits_tab[i];
691 cplx_sum -= cplx_tab[i] * q / qmin;
694 if (bits_sum < 0.001)
696 if (cplx_sum < 0.001)
700 for (i = 0; i < s->mb_num; i++) {
701 const int mb_xy = s->mb_index2xy[i];
702 float newq = q * cplx_tab[i] / bits_tab[i];
705 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
706 newq *= bits_sum / cplx_sum;
709 intq = (int)(newq + 0.5);
713 else if (intq < qmin)
715 s->lambda_table[mb_xy] = intq;
719 void ff_get_2pass_fcode(MpegEncContext *s)
721 RateControlContext *rcc = &s->rc_context;
722 RateControlEntry *rce = &rcc->entry[s->picture_number];
724 s->f_code = rce->f_code;
725 s->b_code = rce->b_code;
728 // FIXME rd or at least approx for dquant
730 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
734 float br_compensation;
738 int picture_number = s->picture_number;
740 RateControlContext *rcc = &s->rc_context;
741 AVCodecContext *a = s->avctx;
742 RateControlEntry local_rce, *rce;
746 const int pict_type = s->pict_type;
747 Picture * const pic = &s->current_picture;
751 if ((s->avctx->flags & CODEC_FLAG_PASS2) &&
752 s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
753 return ff_xvid_rate_estimate_qscale(s, dry_run);
756 get_qminmax(&qmin, &qmax, s, pict_type);
758 fps = 1 / av_q2d(s->avctx->time_base);
759 /* update predictors */
760 if (picture_number > 2 && !dry_run) {
761 const int last_var = s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
762 : rcc->last_mc_mb_var_sum;
763 update_predictor(&rcc->pred[s->last_pict_type],
765 sqrt(last_var), s->frame_bits);
768 if (s->avctx->flags & CODEC_FLAG_PASS2) {
769 assert(picture_number >= 0);
770 assert(picture_number < rcc->num_entries);
771 rce = &rcc->entry[picture_number];
772 wanted_bits = rce->expected_bits;
777 /* FIXME add a dts field to AVFrame and ensure it is set and use it
778 * here instead of reordering but the reordering is simpler for now
779 * until H.264 B-pyramid must be handled. */
780 if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
781 dts_pic = s->current_picture_ptr;
783 dts_pic = s->last_picture_ptr;
785 if (!dts_pic || dts_pic->f->pts == AV_NOPTS_VALUE)
786 wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
788 wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f->pts / fps);
791 diff = s->total_bits - wanted_bits;
792 br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
793 if (br_compensation <= 0.0)
794 br_compensation = 0.001;
796 var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
798 short_term_q = 0; /* avoid warning */
799 if (s->avctx->flags & CODEC_FLAG_PASS2) {
800 if (pict_type != AV_PICTURE_TYPE_I)
801 assert(pict_type == rce->new_pict_type);
803 q = rce->new_qscale / br_compensation;
804 ff_dlog(s, "%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale,
805 br_compensation, s->frame_bits, var, pict_type);
808 rce->new_pict_type = pict_type;
809 rce->mc_mb_var_sum = pic->mc_mb_var_sum;
810 rce->mb_var_sum = pic->mb_var_sum;
811 rce->qscale = FF_QP2LAMBDA * 2;
812 rce->f_code = s->f_code;
813 rce->b_code = s->b_code;
816 bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
817 if (pict_type == AV_PICTURE_TYPE_I) {
818 rce->i_count = s->mb_num;
819 rce->i_tex_bits = bits;
823 rce->i_count = 0; // FIXME we do know this approx
825 rce->p_tex_bits = bits * 0.9;
826 rce->mv_bits = bits * 0.1;
828 rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale;
829 rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale;
830 rcc->mv_bits_sum[pict_type] += rce->mv_bits;
831 rcc->frame_count[pict_type]++;
833 bits = rce->i_tex_bits + rce->p_tex_bits;
834 rate_factor = rcc->pass1_wanted_bits /
835 rcc->pass1_rc_eq_output_sum * br_compensation;
837 q = get_qscale(s, rce, rate_factor, picture_number);
842 q = get_diff_limited_q(s, rce, q);
845 // FIXME type dependent blur like in 2-pass
846 if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
847 rcc->short_term_qsum *= a->qblur;
848 rcc->short_term_qcount *= a->qblur;
850 rcc->short_term_qsum += q;
851 rcc->short_term_qcount++;
852 q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
856 q = modify_qscale(s, rce, q, picture_number);
858 rcc->pass1_wanted_bits += s->bit_rate / fps;
863 if (s->avctx->debug & FF_DEBUG_RC) {
864 av_log(s->avctx, AV_LOG_DEBUG,
865 "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
866 "size:%d var:%d/%d br:%d fps:%d\n",
867 av_get_picture_type_char(pict_type),
868 qmin, q, qmax, picture_number,
869 (int)wanted_bits / 1000, (int)s->total_bits / 1000,
870 br_compensation, short_term_q, s->frame_bits,
871 pic->mb_var_sum, pic->mc_mb_var_sum,
872 s->bit_rate / 1000, (int)fps);
880 if (s->adaptive_quant)
881 adaptive_quantization(s, q);
886 rcc->last_qscale = q;
887 rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
888 rcc->last_mb_var_sum = pic->mb_var_sum;
893 // ----------------------------------------------
896 static int init_pass2(MpegEncContext *s)
898 RateControlContext *rcc = &s->rc_context;
899 AVCodecContext *a = s->avctx;
901 double fps = 1 / av_q2d(s->avctx->time_base);
902 double complexity[5] = { 0 }; // approximate bits at quant=1
903 uint64_t const_bits[5] = { 0 }; // quantizer independent bits
904 uint64_t all_const_bits;
905 uint64_t all_available_bits = (uint64_t)(s->bit_rate *
906 (double)rcc->num_entries / fps);
907 double rate_factor = 0;
909 const int filter_size = (int)(a->qblur * 4) | 1;
910 double expected_bits;
911 double *qscale, *blurred_qscale, qscale_sum;
913 /* find complexity & const_bits & decide the pict_types */
914 for (i = 0; i < rcc->num_entries; i++) {
915 RateControlEntry *rce = &rcc->entry[i];
917 rce->new_pict_type = rce->pict_type;
918 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
919 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
920 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
921 rcc->frame_count[rce->pict_type]++;
923 complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
925 const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
928 all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
929 const_bits[AV_PICTURE_TYPE_P] +
930 const_bits[AV_PICTURE_TYPE_B];
932 if (all_available_bits < all_const_bits) {
933 av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
937 qscale = av_malloc(sizeof(double) * rcc->num_entries);
938 blurred_qscale = av_malloc(sizeof(double) * rcc->num_entries);
939 if (!qscale || !blurred_qscale) {
941 av_free(blurred_qscale);
942 return AVERROR(ENOMEM);
946 for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
950 rcc->buffer_index = s->avctx->rc_buffer_size / 2;
953 for (i = 0; i < rcc->num_entries; i++) {
954 RateControlEntry *rce = &rcc->entry[i];
956 qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
957 rcc->last_qscale_for[rce->pict_type] = qscale[i];
959 assert(filter_size % 2 == 1);
961 /* fixed I/B QP relative to P mode */
962 for (i = rcc->num_entries - 1; i >= 0; i--) {
963 RateControlEntry *rce = &rcc->entry[i];
965 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
969 for (i = 0; i < rcc->num_entries; i++) {
970 RateControlEntry *rce = &rcc->entry[i];
971 const int pict_type = rce->new_pict_type;
973 double q = 0.0, sum = 0.0;
975 for (j = 0; j < filter_size; j++) {
976 int index = i + j - filter_size / 2;
977 double d = index - i;
978 double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));
980 if (index < 0 || index >= rcc->num_entries)
982 if (pict_type != rcc->entry[index].new_pict_type)
984 q += qscale[index] * coeff;
987 blurred_qscale[i] = q / sum;
990 /* find expected bits */
991 for (i = 0; i < rcc->num_entries; i++) {
992 RateControlEntry *rce = &rcc->entry[i];
995 rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);
997 bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
998 bits += 8 * ff_vbv_update(s, bits);
1000 rce->expected_bits = expected_bits;
1001 expected_bits += bits;
1005 "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
1006 expected_bits, (int)all_available_bits, rate_factor);
1007 if (expected_bits > all_available_bits) {
1008 rate_factor -= step;
1013 av_free(blurred_qscale);
1015 /* check bitrate calculations and print info */
1017 for (i = 0; i < rcc->num_entries; i++) {
1018 ff_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
1020 rcc->entry[i].new_qscale,
1021 rcc->entry[i].new_qscale / FF_QP2LAMBDA);
1022 qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
1023 s->avctx->qmin, s->avctx->qmax);
1025 assert(toobig <= 40);
1026 av_log(s->avctx, AV_LOG_DEBUG,
1027 "[lavc rc] requested bitrate: %d bps expected bitrate: %d bps\n",
1029 (int)(expected_bits / ((double)all_available_bits / s->bit_rate)));
1030 av_log(s->avctx, AV_LOG_DEBUG,
1031 "[lavc rc] estimated target average qp: %.3f\n",
1032 (float)qscale_sum / rcc->num_entries);
1034 av_log(s->avctx, AV_LOG_INFO,
1035 "[lavc rc] Using all of requested bitrate is not "
1036 "necessary for this video with these parameters.\n");
1037 } else if (toobig == 40) {
1038 av_log(s->avctx, AV_LOG_ERROR,
1039 "[lavc rc] Error: bitrate too low for this video "
1040 "with these parameters.\n");
1042 } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
1043 av_log(s->avctx, AV_LOG_ERROR,
1044 "[lavc rc] Error: 2pass curve failed to converge\n");