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->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 /* init all to skipped p frames
168 * (with b frames we might have a not encoded frame at the end FIXME) */
169 for (i = 0; i < rcc->num_entries; i++) {
170 RateControlEntry *rce = &rcc->entry[i];
172 rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P;
173 rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2;
174 rce->misc_bits = s->mb_num + 10;
175 rce->mb_var_sum = s->mb_num * 100;
179 p = s->avctx->stats_in;
180 for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
181 RateControlEntry *rce;
186 next = strchr(p, ';');
188 (*next) = 0; // sscanf in unbelievably slow on looong strings // FIXME copy / do not write
191 e = sscanf(p, " in:%d ", &picture_number);
193 assert(picture_number >= 0);
194 assert(picture_number < rcc->num_entries);
195 rce = &rcc->entry[picture_number];
197 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",
198 &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
199 &rce->mv_bits, &rce->misc_bits,
200 &rce->f_code, &rce->b_code,
201 &rce->mc_mb_var_sum, &rce->mb_var_sum,
202 &rce->i_count, &rce->skip_count, &rce->header_bits);
204 av_log(s->avctx, AV_LOG_ERROR,
205 "statistics are damaged at line %d, parser out=%d\n",
213 if (init_pass2(s) < 0)
216 // FIXME maybe move to end
217 if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) {
219 return ff_xvid_rate_control_init(s);
221 av_log(s->avctx, AV_LOG_ERROR,
222 "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n");
228 if (!(s->flags & CODEC_FLAG_PASS2)) {
229 rcc->short_term_qsum = 0.001;
230 rcc->short_term_qcount = 0.001;
232 rcc->pass1_rc_eq_output_sum = 0.001;
233 rcc->pass1_wanted_bits = 0.001;
235 if (s->avctx->qblur > 1.0) {
236 av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
239 /* init stuff with the user specified complexity */
240 if (s->rc_initial_cplx) {
241 for (i = 0; i < 60 * 30; i++) {
242 double bits = s->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
243 RateControlEntry rce;
245 if (i % ((s->gop_size + 3) / 4) == 0)
246 rce.pict_type = AV_PICTURE_TYPE_I;
247 else if (i % (s->max_b_frames + 1))
248 rce.pict_type = AV_PICTURE_TYPE_B;
250 rce.pict_type = AV_PICTURE_TYPE_P;
252 rce.new_pict_type = rce.pict_type;
253 rce.mc_mb_var_sum = bits * s->mb_num / 100000;
254 rce.mb_var_sum = s->mb_num;
256 rce.qscale = FF_QP2LAMBDA * 2;
261 if (s->pict_type == AV_PICTURE_TYPE_I) {
262 rce.i_count = s->mb_num;
263 rce.i_tex_bits = bits;
267 rce.i_count = 0; // FIXME we do know this approx
269 rce.p_tex_bits = bits * 0.9;
270 rce.mv_bits = bits * 0.1;
272 rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale;
273 rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale;
274 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
275 rcc->frame_count[rce.pict_type]++;
277 get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
279 // FIXME misbehaves a little for variable fps
280 rcc->pass1_wanted_bits += s->bit_rate / (1 / av_q2d(s->avctx->time_base));
288 av_cold void ff_rate_control_uninit(MpegEncContext *s)
290 RateControlContext *rcc = &s->rc_context;
293 av_expr_free(rcc->rc_eq_eval);
294 av_freep(&rcc->entry);
297 if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
298 ff_xvid_rate_control_uninit(s);
302 int ff_vbv_update(MpegEncContext *s, int frame_size)
304 RateControlContext *rcc = &s->rc_context;
305 const double fps = 1 / av_q2d(s->avctx->time_base);
306 const int buffer_size = s->avctx->rc_buffer_size;
307 const double min_rate = s->avctx->rc_min_rate / fps;
308 const double max_rate = s->avctx->rc_max_rate / fps;
310 ff_dlog(s, "%d %f %d %f %f\n",
311 buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
316 rcc->buffer_index -= frame_size;
317 if (rcc->buffer_index < 0) {
318 av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
319 rcc->buffer_index = 0;
322 left = buffer_size - rcc->buffer_index - 1;
323 rcc->buffer_index += av_clip(left, min_rate, max_rate);
325 if (rcc->buffer_index > buffer_size) {
326 int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
328 if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
330 rcc->buffer_index -= 8 * stuffing;
332 if (s->avctx->debug & FF_DEBUG_RC)
333 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
342 * Modify the bitrate curve from pass1 for one frame.
344 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
345 double rate_factor, int frame_num)
347 RateControlContext *rcc = &s->rc_context;
348 AVCodecContext *a = s->avctx;
349 const int pict_type = rce->new_pict_type;
350 const double mb_num = s->mb_num;
354 double const_values[] = {
357 rce->i_tex_bits * rce->qscale,
358 rce->p_tex_bits * rce->qscale,
359 (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
360 rce->mv_bits / mb_num,
361 rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
362 rce->i_count / mb_num,
363 rce->mc_mb_var_sum / mb_num,
364 rce->mb_var_sum / mb_num,
365 rce->pict_type == AV_PICTURE_TYPE_I,
366 rce->pict_type == AV_PICTURE_TYPE_P,
367 rce->pict_type == AV_PICTURE_TYPE_B,
368 rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
371 rcc->last_qscale_for[AV_PICTURE_TYPE_I],
372 rcc->last_qscale_for[AV_PICTURE_TYPE_P],
373 rcc->last_qscale_for[AV_PICTURE_TYPE_B],
374 rcc->next_non_b_qscale,
376 rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
377 rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
378 rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
379 rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
380 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
384 bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
386 av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->rc_eq);
390 rcc->pass1_rc_eq_output_sum += bits;
394 bits += 1.0; // avoid 1/0 issues
397 for (i = 0; i < s->avctx->rc_override_count; i++) {
398 RcOverride *rco = s->avctx->rc_override;
399 if (rco[i].start_frame > frame_num)
401 if (rco[i].end_frame < frame_num)
405 bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it?
407 bits *= rco[i].quality_factor;
410 q = bits2qp(rce, bits);
413 if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
414 q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
415 else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
416 q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
423 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
425 RateControlContext *rcc = &s->rc_context;
426 AVCodecContext *a = s->avctx;
427 const int pict_type = rce->new_pict_type;
428 const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
429 const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];
431 if (pict_type == AV_PICTURE_TYPE_I &&
432 (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P))
433 q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
434 else if (pict_type == AV_PICTURE_TYPE_B &&
435 a->b_quant_factor > 0.0)
436 q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
440 /* last qscale / qdiff stuff */
441 if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
442 double last_q = rcc->last_qscale_for[pict_type];
443 const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;
445 if (q > last_q + maxdiff)
446 q = last_q + maxdiff;
447 else if (q < last_q - maxdiff)
448 q = last_q - maxdiff;
451 rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring
453 if (pict_type != AV_PICTURE_TYPE_B)
454 rcc->last_non_b_pict_type = pict_type;
460 * Get the qmin & qmax for pict_type.
462 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
467 assert(qmin <= qmax);
470 case AV_PICTURE_TYPE_B:
471 qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
472 qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
474 case AV_PICTURE_TYPE_I:
475 qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
476 qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
480 qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
481 qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);
490 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce,
491 double q, int frame_num)
493 RateControlContext *rcc = &s->rc_context;
494 const double buffer_size = s->avctx->rc_buffer_size;
495 const double fps = 1 / av_q2d(s->avctx->time_base);
496 const double min_rate = s->avctx->rc_min_rate / fps;
497 const double max_rate = s->avctx->rc_max_rate / fps;
498 const int pict_type = rce->new_pict_type;
501 get_qminmax(&qmin, &qmax, s, pict_type);
504 if (s->rc_qmod_freq &&
505 frame_num % s->rc_qmod_freq == 0 &&
506 pict_type == AV_PICTURE_TYPE_P)
509 /* buffer overflow/underflow protection */
511 double expected_size = rcc->buffer_index;
515 double d = 2 * (buffer_size - expected_size) / buffer_size;
520 q *= pow(d, 1.0 / s->rc_buffer_aggressivity);
522 q_limit = bits2qp(rce,
523 FFMAX((min_rate - buffer_size + rcc->buffer_index) *
524 s->avctx->rc_min_vbv_overflow_use, 1));
527 if (s->avctx->debug & FF_DEBUG_RC)
528 av_log(s->avctx, AV_LOG_DEBUG,
529 "limiting QP %f -> %f\n", q, q_limit);
535 double d = 2 * expected_size / buffer_size;
540 q /= pow(d, 1.0 / s->rc_buffer_aggressivity);
542 q_limit = bits2qp(rce,
543 FFMAX(rcc->buffer_index *
544 s->avctx->rc_max_available_vbv_use,
547 if (s->avctx->debug & FF_DEBUG_RC)
548 av_log(s->avctx, AV_LOG_DEBUG,
549 "limiting QP %f -> %f\n", q, q_limit);
554 ff_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
555 q, max_rate, min_rate, buffer_size, rcc->buffer_index,
556 s->rc_buffer_aggressivity);
557 if (s->rc_qsquish == 0.0 || qmin == qmax) {
563 double min2 = log(qmin);
564 double max2 = log(qmax);
567 q = (q - min2) / (max2 - min2) - 0.5;
569 q = 1.0 / (1.0 + exp(q));
570 q = q * (max2 - min2) + min2;
578 // ----------------------------------
581 static double predict_size(Predictor *p, double q, double var)
583 return p->coeff * var / (q * p->count);
586 static void update_predictor(Predictor *p, double q, double var, double size)
588 double new_coeff = size * q / (var + 1);
592 p->count *= p->decay;
593 p->coeff *= p->decay;
595 p->coeff += new_coeff;
598 static void adaptive_quantization(MpegEncContext *s, double q)
601 const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0);
602 const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0);
603 const float temp_cplx_masking = s->avctx->temporal_cplx_masking;
604 const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
605 const float p_masking = s->avctx->p_masking;
606 const float border_masking = s->border_masking;
607 float bits_sum = 0.0;
608 float cplx_sum = 0.0;
609 float *cplx_tab = s->cplx_tab;
610 float *bits_tab = s->bits_tab;
611 const int qmin = s->avctx->mb_lmin;
612 const int qmax = s->avctx->mb_lmax;
613 Picture *const pic = &s->current_picture;
614 const int mb_width = s->mb_width;
615 const int mb_height = s->mb_height;
617 for (i = 0; i < s->mb_num; i++) {
618 const int mb_xy = s->mb_index2xy[i];
619 float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
620 float spat_cplx = sqrt(pic->mb_var[mb_xy]);
621 const int lumi = pic->mb_mean[mb_xy];
622 float bits, cplx, factor;
623 int mb_x = mb_xy % s->mb_stride;
624 int mb_y = mb_xy / s->mb_stride;
626 float mb_factor = 0.0;
628 spat_cplx = 4; // FIXME finetune
630 temp_cplx = 4; // FIXME finetune
632 if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
634 factor = 1.0 + p_masking;
637 factor = pow(temp_cplx, -temp_cplx_masking);
639 factor *= pow(spat_cplx, -spatial_cplx_masking);
642 factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
644 factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);
646 if (mb_x < mb_width / 5) {
647 mb_distance = mb_width / 5 - mb_x;
648 mb_factor = (float)mb_distance / (float)(mb_width / 5);
649 } else if (mb_x > 4 * mb_width / 5) {
650 mb_distance = mb_x - 4 * mb_width / 5;
651 mb_factor = (float)mb_distance / (float)(mb_width / 5);
653 if (mb_y < mb_height / 5) {
654 mb_distance = mb_height / 5 - mb_y;
655 mb_factor = FFMAX(mb_factor,
656 (float)mb_distance / (float)(mb_height / 5));
657 } else if (mb_y > 4 * mb_height / 5) {
658 mb_distance = mb_y - 4 * mb_height / 5;
659 mb_factor = FFMAX(mb_factor,
660 (float)mb_distance / (float)(mb_height / 5));
663 factor *= 1.0 - border_masking * mb_factor;
665 if (factor < 0.00001)
668 bits = cplx * factor;
675 /* handle qmin/qmax clipping */
676 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
677 float factor = bits_sum / cplx_sum;
678 for (i = 0; i < s->mb_num; i++) {
679 float newq = q * cplx_tab[i] / bits_tab[i];
683 bits_sum -= bits_tab[i];
684 cplx_sum -= cplx_tab[i] * q / qmax;
685 } else if (newq < qmin) {
686 bits_sum -= bits_tab[i];
687 cplx_sum -= cplx_tab[i] * q / qmin;
690 if (bits_sum < 0.001)
692 if (cplx_sum < 0.001)
696 for (i = 0; i < s->mb_num; i++) {
697 const int mb_xy = s->mb_index2xy[i];
698 float newq = q * cplx_tab[i] / bits_tab[i];
701 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
702 newq *= bits_sum / cplx_sum;
705 intq = (int)(newq + 0.5);
709 else if (intq < qmin)
711 s->lambda_table[mb_xy] = intq;
715 void ff_get_2pass_fcode(MpegEncContext *s)
717 RateControlContext *rcc = &s->rc_context;
718 RateControlEntry *rce = &rcc->entry[s->picture_number];
720 s->f_code = rce->f_code;
721 s->b_code = rce->b_code;
724 // FIXME rd or at least approx for dquant
726 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
730 float br_compensation;
734 int picture_number = s->picture_number;
736 RateControlContext *rcc = &s->rc_context;
737 AVCodecContext *a = s->avctx;
738 RateControlEntry local_rce, *rce;
742 const int pict_type = s->pict_type;
743 Picture * const pic = &s->current_picture;
747 if ((s->flags & CODEC_FLAG_PASS2) &&
748 s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
749 return ff_xvid_rate_estimate_qscale(s, dry_run);
752 get_qminmax(&qmin, &qmax, s, pict_type);
754 fps = 1 / av_q2d(s->avctx->time_base);
755 /* update predictors */
756 if (picture_number > 2 && !dry_run) {
757 const int last_var = s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
758 : rcc->last_mc_mb_var_sum;
759 update_predictor(&rcc->pred[s->last_pict_type],
761 sqrt(last_var), s->frame_bits);
764 if (s->flags & CODEC_FLAG_PASS2) {
765 assert(picture_number >= 0);
766 assert(picture_number < rcc->num_entries);
767 rce = &rcc->entry[picture_number];
768 wanted_bits = rce->expected_bits;
773 /* FIXME add a dts field to AVFrame and ensure it is set and use it
774 * here instead of reordering but the reordering is simpler for now
775 * until H.264 B-pyramid must be handled. */
776 if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
777 dts_pic = s->current_picture_ptr;
779 dts_pic = s->last_picture_ptr;
781 if (!dts_pic || dts_pic->f->pts == AV_NOPTS_VALUE)
782 wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
784 wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f->pts / fps);
787 diff = s->total_bits - wanted_bits;
788 br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
789 if (br_compensation <= 0.0)
790 br_compensation = 0.001;
792 var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
794 short_term_q = 0; /* avoid warning */
795 if (s->flags & CODEC_FLAG_PASS2) {
796 if (pict_type != AV_PICTURE_TYPE_I)
797 assert(pict_type == rce->new_pict_type);
799 q = rce->new_qscale / br_compensation;
800 ff_dlog(s, "%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale,
801 br_compensation, s->frame_bits, var, pict_type);
804 rce->new_pict_type = pict_type;
805 rce->mc_mb_var_sum = pic->mc_mb_var_sum;
806 rce->mb_var_sum = pic->mb_var_sum;
807 rce->qscale = FF_QP2LAMBDA * 2;
808 rce->f_code = s->f_code;
809 rce->b_code = s->b_code;
812 bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
813 if (pict_type == AV_PICTURE_TYPE_I) {
814 rce->i_count = s->mb_num;
815 rce->i_tex_bits = bits;
819 rce->i_count = 0; // FIXME we do know this approx
821 rce->p_tex_bits = bits * 0.9;
822 rce->mv_bits = bits * 0.1;
824 rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale;
825 rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale;
826 rcc->mv_bits_sum[pict_type] += rce->mv_bits;
827 rcc->frame_count[pict_type]++;
829 bits = rce->i_tex_bits + rce->p_tex_bits;
830 rate_factor = rcc->pass1_wanted_bits /
831 rcc->pass1_rc_eq_output_sum * br_compensation;
833 q = get_qscale(s, rce, rate_factor, picture_number);
838 q = get_diff_limited_q(s, rce, q);
841 // FIXME type dependent blur like in 2-pass
842 if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
843 rcc->short_term_qsum *= a->qblur;
844 rcc->short_term_qcount *= a->qblur;
846 rcc->short_term_qsum += q;
847 rcc->short_term_qcount++;
848 q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
852 q = modify_qscale(s, rce, q, picture_number);
854 rcc->pass1_wanted_bits += s->bit_rate / fps;
859 if (s->avctx->debug & FF_DEBUG_RC) {
860 av_log(s->avctx, AV_LOG_DEBUG,
861 "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
862 "size:%d var:%d/%d br:%d fps:%d\n",
863 av_get_picture_type_char(pict_type),
864 qmin, q, qmax, picture_number,
865 (int)wanted_bits / 1000, (int)s->total_bits / 1000,
866 br_compensation, short_term_q, s->frame_bits,
867 pic->mb_var_sum, pic->mc_mb_var_sum,
868 s->bit_rate / 1000, (int)fps);
876 if (s->adaptive_quant)
877 adaptive_quantization(s, q);
882 rcc->last_qscale = q;
883 rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
884 rcc->last_mb_var_sum = pic->mb_var_sum;
889 // ----------------------------------------------
892 static int init_pass2(MpegEncContext *s)
894 RateControlContext *rcc = &s->rc_context;
895 AVCodecContext *a = s->avctx;
897 double fps = 1 / av_q2d(s->avctx->time_base);
898 double complexity[5] = { 0 }; // approximate bits at quant=1
899 uint64_t const_bits[5] = { 0 }; // quantizer independent bits
900 uint64_t all_const_bits;
901 uint64_t all_available_bits = (uint64_t)(s->bit_rate *
902 (double)rcc->num_entries / fps);
903 double rate_factor = 0;
905 const int filter_size = (int)(a->qblur * 4) | 1;
906 double expected_bits;
907 double *qscale, *blurred_qscale, qscale_sum;
909 /* find complexity & const_bits & decide the pict_types */
910 for (i = 0; i < rcc->num_entries; i++) {
911 RateControlEntry *rce = &rcc->entry[i];
913 rce->new_pict_type = rce->pict_type;
914 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
915 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
916 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
917 rcc->frame_count[rce->pict_type]++;
919 complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
921 const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
924 all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
925 const_bits[AV_PICTURE_TYPE_P] +
926 const_bits[AV_PICTURE_TYPE_B];
928 if (all_available_bits < all_const_bits) {
929 av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
933 qscale = av_malloc(sizeof(double) * rcc->num_entries);
934 blurred_qscale = av_malloc(sizeof(double) * rcc->num_entries);
937 for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
941 rcc->buffer_index = s->avctx->rc_buffer_size / 2;
944 for (i = 0; i < rcc->num_entries; i++) {
945 RateControlEntry *rce = &rcc->entry[i];
947 qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
948 rcc->last_qscale_for[rce->pict_type] = qscale[i];
950 assert(filter_size % 2 == 1);
952 /* fixed I/B QP relative to P mode */
953 for (i = rcc->num_entries - 1; i >= 0; i--) {
954 RateControlEntry *rce = &rcc->entry[i];
956 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
960 for (i = 0; i < rcc->num_entries; i++) {
961 RateControlEntry *rce = &rcc->entry[i];
962 const int pict_type = rce->new_pict_type;
964 double q = 0.0, sum = 0.0;
966 for (j = 0; j < filter_size; j++) {
967 int index = i + j - filter_size / 2;
968 double d = index - i;
969 double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));
971 if (index < 0 || index >= rcc->num_entries)
973 if (pict_type != rcc->entry[index].new_pict_type)
975 q += qscale[index] * coeff;
978 blurred_qscale[i] = q / sum;
981 /* find expected bits */
982 for (i = 0; i < rcc->num_entries; i++) {
983 RateControlEntry *rce = &rcc->entry[i];
986 rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);
988 bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
989 bits += 8 * ff_vbv_update(s, bits);
991 rce->expected_bits = expected_bits;
992 expected_bits += bits;
996 "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
997 expected_bits, (int)all_available_bits, rate_factor);
998 if (expected_bits > all_available_bits) {
1004 av_free(blurred_qscale);
1006 /* check bitrate calculations and print info */
1008 for (i = 0; i < rcc->num_entries; i++) {
1009 ff_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
1011 rcc->entry[i].new_qscale,
1012 rcc->entry[i].new_qscale / FF_QP2LAMBDA);
1013 qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
1014 s->avctx->qmin, s->avctx->qmax);
1016 assert(toobig <= 40);
1017 av_log(s->avctx, AV_LOG_DEBUG,
1018 "[lavc rc] requested bitrate: %d bps expected bitrate: %d bps\n",
1020 (int)(expected_bits / ((double)all_available_bits / s->bit_rate)));
1021 av_log(s->avctx, AV_LOG_DEBUG,
1022 "[lavc rc] estimated target average qp: %.3f\n",
1023 (float)qscale_sum / rcc->num_entries);
1025 av_log(s->avctx, AV_LOG_INFO,
1026 "[lavc rc] Using all of requested bitrate is not "
1027 "necessary for this video with these parameters.\n");
1028 } else if (toobig == 40) {
1029 av_log(s->avctx, AV_LOG_ERROR,
1030 "[lavc rc] Error: bitrate too low for this video "
1031 "with these parameters.\n");
1033 } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
1034 av_log(s->avctx, AV_LOG_ERROR,
1035 "[lavc rc] Error: 2pass curve failed to converge\n");