2 * Rate control for video encoders
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
6 * This file is part of FFmpeg.
8 * FFmpeg 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 * FFmpeg 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 FFmpeg; 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.
29 #include "ratecontrol.h"
30 #include "mpegvideo.h"
31 #include "libavutil/eval.h"
33 #undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
37 #define M_E 2.718281828
40 static int init_pass2(MpegEncContext *s);
41 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
42 double rate_factor, int frame_num);
44 void ff_write_pass1_stats(MpegEncContext *s)
46 snprintf(s->avctx->stats_out, 256,
47 "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d "
48 "fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d;\n",
49 s->current_picture_ptr->f.display_picture_number,
50 s->current_picture_ptr->f.coded_picture_number,
52 s->current_picture.f.quality,
59 s->current_picture.mc_mb_var_sum,
60 s->current_picture.mb_var_sum,
61 s->i_count, s->skip_count,
65 static double get_fps(AVCodecContext *avctx)
67 return 1.0 / av_q2d(avctx->time_base) / FFMAX(avctx->ticks_per_frame, 1);
70 static inline double qp2bits(RateControlEntry *rce, double qp)
73 av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
75 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp;
78 static inline double bits2qp(RateControlEntry *rce, double bits)
81 av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
83 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits;
86 int ff_rate_control_init(MpegEncContext *s)
88 RateControlContext *rcc = &s->rc_context;
90 static const char * const const_names[] = {
119 static double (* const func1[])(void *, double) = {
124 static const char * const func1_names[] = {
131 if (!s->avctx->rc_max_available_vbv_use && s->avctx->rc_buffer_size) {
132 if (s->avctx->rc_max_rate) {
133 s->avctx->rc_max_available_vbv_use = av_clipf(s->avctx->rc_max_rate/(s->avctx->rc_buffer_size*get_fps(s->avctx)), 1.0/3, 1.0);
135 s->avctx->rc_max_available_vbv_use = 1.0;
138 res = av_expr_parse(&rcc->rc_eq_eval,
139 s->avctx->rc_eq ? s->avctx->rc_eq : "tex^qComp",
140 const_names, func1_names, func1,
141 NULL, NULL, 0, s->avctx);
143 av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->avctx->rc_eq);
147 for (i = 0; i < 5; i++) {
148 rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0;
149 rcc->pred[i].count = 1.0;
150 rcc->pred[i].decay = 0.4;
152 rcc->i_cplx_sum [i] =
153 rcc->p_cplx_sum [i] =
154 rcc->mv_bits_sum[i] =
155 rcc->qscale_sum [i] =
156 rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such
158 rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5;
160 rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy;
161 if (!rcc->buffer_index)
162 rcc->buffer_index = s->avctx->rc_buffer_size * 3 / 4;
164 if (s->flags & CODEC_FLAG_PASS2) {
168 /* find number of pics */
169 p = s->avctx->stats_in;
171 p = strchr(p + 1, ';');
172 i += s->max_b_frames;
173 if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry))
175 rcc->entry = av_mallocz(i * sizeof(RateControlEntry));
176 rcc->num_entries = i;
178 /* init all to skipped p frames
179 * (with b frames we might have a not encoded frame at the end FIXME) */
180 for (i = 0; i < rcc->num_entries; i++) {
181 RateControlEntry *rce = &rcc->entry[i];
183 rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P;
184 rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2;
185 rce->misc_bits = s->mb_num + 10;
186 rce->mb_var_sum = s->mb_num * 100;
190 p = s->avctx->stats_in;
191 for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
192 RateControlEntry *rce;
197 next = strchr(p, ';');
199 (*next) = 0; // sscanf in unbelievably slow on looong strings // FIXME copy / do not write
202 e = sscanf(p, " in:%d ", &picture_number);
204 assert(picture_number >= 0);
205 assert(picture_number < rcc->num_entries);
206 rce = &rcc->entry[picture_number];
208 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",
209 &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
210 &rce->mv_bits, &rce->misc_bits,
211 &rce->f_code, &rce->b_code,
212 &rce->mc_mb_var_sum, &rce->mb_var_sum,
213 &rce->i_count, &rce->skip_count, &rce->header_bits);
215 av_log(s->avctx, AV_LOG_ERROR,
216 "statistics are damaged at line %d, parser out=%d\n",
224 if (init_pass2(s) < 0)
227 // FIXME maybe move to end
228 if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) {
230 return ff_xvid_rate_control_init(s);
232 av_log(s->avctx, AV_LOG_ERROR,
233 "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n");
239 if (!(s->flags & CODEC_FLAG_PASS2)) {
240 rcc->short_term_qsum = 0.001;
241 rcc->short_term_qcount = 0.001;
243 rcc->pass1_rc_eq_output_sum = 0.001;
244 rcc->pass1_wanted_bits = 0.001;
246 if (s->avctx->qblur > 1.0) {
247 av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
250 /* init stuff with the user specified complexity */
251 if (s->avctx->rc_initial_cplx) {
252 for (i = 0; i < 60 * 30; i++) {
253 double bits = s->avctx->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
254 RateControlEntry rce;
256 if (i % ((s->gop_size + 3) / 4) == 0)
257 rce.pict_type = AV_PICTURE_TYPE_I;
258 else if (i % (s->max_b_frames + 1))
259 rce.pict_type = AV_PICTURE_TYPE_B;
261 rce.pict_type = AV_PICTURE_TYPE_P;
263 rce.new_pict_type = rce.pict_type;
264 rce.mc_mb_var_sum = bits * s->mb_num / 100000;
265 rce.mb_var_sum = s->mb_num;
267 rce.qscale = FF_QP2LAMBDA * 2;
272 if (s->pict_type == AV_PICTURE_TYPE_I) {
273 rce.i_count = s->mb_num;
274 rce.i_tex_bits = bits;
278 rce.i_count = 0; // FIXME we do know this approx
280 rce.p_tex_bits = bits * 0.9;
281 rce.mv_bits = bits * 0.1;
283 rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale;
284 rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale;
285 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
286 rcc->frame_count[rce.pict_type]++;
288 get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
290 // FIXME misbehaves a little for variable fps
291 rcc->pass1_wanted_bits += s->bit_rate / get_fps(s->avctx);
299 void ff_rate_control_uninit(MpegEncContext *s)
301 RateControlContext *rcc = &s->rc_context;
304 av_expr_free(rcc->rc_eq_eval);
305 av_freep(&rcc->entry);
308 if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
309 ff_xvid_rate_control_uninit(s);
313 int ff_vbv_update(MpegEncContext *s, int frame_size)
315 RateControlContext *rcc = &s->rc_context;
316 const double fps = get_fps(s->avctx);
317 const int buffer_size = s->avctx->rc_buffer_size;
318 const double min_rate = s->avctx->rc_min_rate / fps;
319 const double max_rate = s->avctx->rc_max_rate / fps;
321 av_dlog(s, "%d %f %d %f %f\n",
322 buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
327 rcc->buffer_index -= frame_size;
328 if (rcc->buffer_index < 0) {
329 av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
330 rcc->buffer_index = 0;
333 left = buffer_size - rcc->buffer_index - 1;
334 rcc->buffer_index += av_clip(left, min_rate, max_rate);
336 if (rcc->buffer_index > buffer_size) {
337 int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
339 if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
341 rcc->buffer_index -= 8 * stuffing;
343 if (s->avctx->debug & FF_DEBUG_RC)
344 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
353 * Modify the bitrate curve from pass1 for one frame.
355 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
356 double rate_factor, int frame_num)
358 RateControlContext *rcc = &s->rc_context;
359 AVCodecContext *a = s->avctx;
360 const int pict_type = rce->new_pict_type;
361 const double mb_num = s->mb_num;
365 double const_values[] = {
368 rce->i_tex_bits * rce->qscale,
369 rce->p_tex_bits * rce->qscale,
370 (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
371 rce->mv_bits / mb_num,
372 rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
373 rce->i_count / mb_num,
374 rce->mc_mb_var_sum / mb_num,
375 rce->mb_var_sum / mb_num,
376 rce->pict_type == AV_PICTURE_TYPE_I,
377 rce->pict_type == AV_PICTURE_TYPE_P,
378 rce->pict_type == AV_PICTURE_TYPE_B,
379 rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
382 rcc->last_qscale_for[AV_PICTURE_TYPE_I],
383 rcc->last_qscale_for[AV_PICTURE_TYPE_P],
384 rcc->last_qscale_for[AV_PICTURE_TYPE_B],
385 rcc->next_non_b_qscale,
387 rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
388 rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
389 rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
390 rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
391 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
395 bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
397 av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq);
401 rcc->pass1_rc_eq_output_sum += bits;
405 bits += 1.0; // avoid 1/0 issues
408 for (i = 0; i < s->avctx->rc_override_count; i++) {
409 RcOverride *rco = s->avctx->rc_override;
410 if (rco[i].start_frame > frame_num)
412 if (rco[i].end_frame < frame_num)
416 bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it?
418 bits *= rco[i].quality_factor;
421 q = bits2qp(rce, bits);
424 if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
425 q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
426 else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
427 q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
434 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
436 RateControlContext *rcc = &s->rc_context;
437 AVCodecContext *a = s->avctx;
438 const int pict_type = rce->new_pict_type;
439 const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
440 const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];
442 if (pict_type == AV_PICTURE_TYPE_I &&
443 (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P))
444 q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
445 else if (pict_type == AV_PICTURE_TYPE_B &&
446 a->b_quant_factor > 0.0)
447 q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
451 /* last qscale / qdiff stuff */
452 if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
453 double last_q = rcc->last_qscale_for[pict_type];
454 const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;
456 if (q > last_q + maxdiff)
457 q = last_q + maxdiff;
458 else if (q < last_q - maxdiff)
459 q = last_q - maxdiff;
462 rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring
464 if (pict_type != AV_PICTURE_TYPE_B)
465 rcc->last_non_b_pict_type = pict_type;
471 * Get the qmin & qmax for pict_type.
473 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
475 int qmin = s->avctx->lmin;
476 int qmax = s->avctx->lmax;
478 assert(qmin <= qmax);
481 case AV_PICTURE_TYPE_B:
482 qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
483 qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
485 case AV_PICTURE_TYPE_I:
486 qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
487 qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
491 qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
492 qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);
501 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce,
502 double q, int frame_num)
504 RateControlContext *rcc = &s->rc_context;
505 const double buffer_size = s->avctx->rc_buffer_size;
506 const double fps = get_fps(s->avctx);
507 const double min_rate = s->avctx->rc_min_rate / fps;
508 const double max_rate = s->avctx->rc_max_rate / fps;
509 const int pict_type = rce->new_pict_type;
512 get_qminmax(&qmin, &qmax, s, pict_type);
515 if (s->avctx->rc_qmod_freq &&
516 frame_num % s->avctx->rc_qmod_freq == 0 &&
517 pict_type == AV_PICTURE_TYPE_P)
518 q *= s->avctx->rc_qmod_amp;
520 /* buffer overflow/underflow protection */
522 double expected_size = rcc->buffer_index;
526 double d = 2 * (buffer_size - expected_size) / buffer_size;
531 q *= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity);
533 q_limit = bits2qp(rce,
534 FFMAX((min_rate - buffer_size + rcc->buffer_index) *
535 s->avctx->rc_min_vbv_overflow_use, 1));
538 if (s->avctx->debug & FF_DEBUG_RC)
539 av_log(s->avctx, AV_LOG_DEBUG,
540 "limiting QP %f -> %f\n", q, q_limit);
546 double d = 2 * expected_size / buffer_size;
551 q /= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity);
553 q_limit = bits2qp(rce,
554 FFMAX(rcc->buffer_index *
555 s->avctx->rc_max_available_vbv_use,
558 if (s->avctx->debug & FF_DEBUG_RC)
559 av_log(s->avctx, AV_LOG_DEBUG,
560 "limiting QP %f -> %f\n", q, q_limit);
565 av_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
566 q, max_rate, min_rate, buffer_size, rcc->buffer_index,
567 s->avctx->rc_buffer_aggressivity);
568 if (s->avctx->rc_qsquish == 0.0 || qmin == qmax) {
574 double min2 = log(qmin);
575 double max2 = log(qmax);
578 q = (q - min2) / (max2 - min2) - 0.5;
580 q = 1.0 / (1.0 + exp(q));
581 q = q * (max2 - min2) + min2;
589 // ----------------------------------
592 static double predict_size(Predictor *p, double q, double var)
594 return p->coeff * var / (q * p->count);
597 static void update_predictor(Predictor *p, double q, double var, double size)
599 double new_coeff = size * q / (var + 1);
603 p->count *= p->decay;
604 p->coeff *= p->decay;
606 p->coeff += new_coeff;
609 static void adaptive_quantization(MpegEncContext *s, double q)
612 const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0);
613 const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0);
614 const float temp_cplx_masking = s->avctx->temporal_cplx_masking;
615 const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
616 const float p_masking = s->avctx->p_masking;
617 const float border_masking = s->avctx->border_masking;
618 float bits_sum = 0.0;
619 float cplx_sum = 0.0;
620 float *cplx_tab = s->cplx_tab;
621 float *bits_tab = s->bits_tab;
622 const int qmin = s->avctx->mb_lmin;
623 const int qmax = s->avctx->mb_lmax;
624 Picture *const pic = &s->current_picture;
625 const int mb_width = s->mb_width;
626 const int mb_height = s->mb_height;
628 for (i = 0; i < s->mb_num; i++) {
629 const int mb_xy = s->mb_index2xy[i];
630 float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
631 float spat_cplx = sqrt(pic->mb_var[mb_xy]);
632 const int lumi = pic->mb_mean[mb_xy];
633 float bits, cplx, factor;
634 int mb_x = mb_xy % s->mb_stride;
635 int mb_y = mb_xy / s->mb_stride;
637 float mb_factor = 0.0;
639 spat_cplx = 4; // FIXME finetune
641 temp_cplx = 4; // FIXME finetune
643 if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
645 factor = 1.0 + p_masking;
648 factor = pow(temp_cplx, -temp_cplx_masking);
650 factor *= pow(spat_cplx, -spatial_cplx_masking);
653 factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
655 factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);
657 if (mb_x < mb_width / 5) {
658 mb_distance = mb_width / 5 - mb_x;
659 mb_factor = (float)mb_distance / (float)(mb_width / 5);
660 } else if (mb_x > 4 * mb_width / 5) {
661 mb_distance = mb_x - 4 * mb_width / 5;
662 mb_factor = (float)mb_distance / (float)(mb_width / 5);
664 if (mb_y < mb_height / 5) {
665 mb_distance = mb_height / 5 - mb_y;
666 mb_factor = FFMAX(mb_factor,
667 (float)mb_distance / (float)(mb_height / 5));
668 } else if (mb_y > 4 * mb_height / 5) {
669 mb_distance = mb_y - 4 * mb_height / 5;
670 mb_factor = FFMAX(mb_factor,
671 (float)mb_distance / (float)(mb_height / 5));
674 factor *= 1.0 - border_masking * mb_factor;
676 if (factor < 0.00001)
679 bits = cplx * factor;
686 /* handle qmin/qmax clipping */
687 if (s->flags & CODEC_FLAG_NORMALIZE_AQP) {
688 float factor = bits_sum / cplx_sum;
689 for (i = 0; i < s->mb_num; i++) {
690 float newq = q * cplx_tab[i] / bits_tab[i];
694 bits_sum -= bits_tab[i];
695 cplx_sum -= cplx_tab[i] * q / qmax;
696 } else if (newq < qmin) {
697 bits_sum -= bits_tab[i];
698 cplx_sum -= cplx_tab[i] * q / qmin;
701 if (bits_sum < 0.001)
703 if (cplx_sum < 0.001)
707 for (i = 0; i < s->mb_num; i++) {
708 const int mb_xy = s->mb_index2xy[i];
709 float newq = q * cplx_tab[i] / bits_tab[i];
712 if (s->flags & CODEC_FLAG_NORMALIZE_AQP) {
713 newq *= bits_sum / cplx_sum;
716 intq = (int)(newq + 0.5);
720 else if (intq < qmin)
722 s->lambda_table[mb_xy] = intq;
726 void ff_get_2pass_fcode(MpegEncContext *s)
728 RateControlContext *rcc = &s->rc_context;
729 RateControlEntry *rce = &rcc->entry[s->picture_number];
731 s->f_code = rce->f_code;
732 s->b_code = rce->b_code;
735 // FIXME rd or at least approx for dquant
737 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
741 float br_compensation;
745 int picture_number = s->picture_number;
747 RateControlContext *rcc = &s->rc_context;
748 AVCodecContext *a = s->avctx;
749 RateControlEntry local_rce, *rce;
753 const int pict_type = s->pict_type;
754 Picture * const pic = &s->current_picture;
758 if ((s->flags & CODEC_FLAG_PASS2) &&
759 s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
760 return ff_xvid_rate_estimate_qscale(s, dry_run);
763 get_qminmax(&qmin, &qmax, s, pict_type);
765 fps = get_fps(s->avctx);
766 /* update predictors */
767 if (picture_number > 2 && !dry_run) {
768 const int last_var = s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
769 : rcc->last_mc_mb_var_sum;
770 av_assert1(s->frame_bits >= s->stuffing_bits);
771 update_predictor(&rcc->pred[s->last_pict_type],
774 s->frame_bits - s->stuffing_bits);
777 if (s->flags & CODEC_FLAG_PASS2) {
778 assert(picture_number >= 0);
779 if (picture_number >= rcc->num_entries) {
780 av_log(s, AV_LOG_ERROR, "Input is longer than 2-pass log file\n");
783 rce = &rcc->entry[picture_number];
784 wanted_bits = rce->expected_bits;
789 /* FIXME add a dts field to AVFrame and ensure it is set and use it
790 * here instead of reordering but the reordering is simpler for now
791 * until H.264 B-pyramid must be handled. */
792 if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
793 dts_pic = s->current_picture_ptr;
795 dts_pic = s->last_picture_ptr;
797 if (!dts_pic || dts_pic->f.pts == AV_NOPTS_VALUE)
798 wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
800 wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f.pts / fps);
803 diff = s->total_bits - wanted_bits;
804 br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
805 if (br_compensation <= 0.0)
806 br_compensation = 0.001;
808 var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
810 short_term_q = 0; /* avoid warning */
811 if (s->flags & CODEC_FLAG_PASS2) {
812 if (pict_type != AV_PICTURE_TYPE_I)
813 assert(pict_type == rce->new_pict_type);
815 q = rce->new_qscale / br_compensation;
816 av_dlog(s, "%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale,
817 br_compensation, s->frame_bits, var, pict_type);
820 rce->new_pict_type = pict_type;
821 rce->mc_mb_var_sum = pic->mc_mb_var_sum;
822 rce->mb_var_sum = pic->mb_var_sum;
823 rce->qscale = FF_QP2LAMBDA * 2;
824 rce->f_code = s->f_code;
825 rce->b_code = s->b_code;
828 bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
829 if (pict_type == AV_PICTURE_TYPE_I) {
830 rce->i_count = s->mb_num;
831 rce->i_tex_bits = bits;
835 rce->i_count = 0; // FIXME we do know this approx
837 rce->p_tex_bits = bits * 0.9;
838 rce->mv_bits = bits * 0.1;
840 rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale;
841 rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale;
842 rcc->mv_bits_sum[pict_type] += rce->mv_bits;
843 rcc->frame_count[pict_type]++;
845 bits = rce->i_tex_bits + rce->p_tex_bits;
846 rate_factor = rcc->pass1_wanted_bits /
847 rcc->pass1_rc_eq_output_sum * br_compensation;
849 q = get_qscale(s, rce, rate_factor, picture_number);
854 q = get_diff_limited_q(s, rce, q);
857 // FIXME type dependent blur like in 2-pass
858 if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
859 rcc->short_term_qsum *= a->qblur;
860 rcc->short_term_qcount *= a->qblur;
862 rcc->short_term_qsum += q;
863 rcc->short_term_qcount++;
864 q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
868 q = modify_qscale(s, rce, q, picture_number);
870 rcc->pass1_wanted_bits += s->bit_rate / fps;
875 if (s->avctx->debug & FF_DEBUG_RC) {
876 av_log(s->avctx, AV_LOG_DEBUG,
877 "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
878 "size:%d var:%d/%d br:%d fps:%d\n",
879 av_get_picture_type_char(pict_type),
880 qmin, q, qmax, picture_number,
881 (int)wanted_bits / 1000, (int)s->total_bits / 1000,
882 br_compensation, short_term_q, s->frame_bits,
883 pic->mb_var_sum, pic->mc_mb_var_sum,
884 s->bit_rate / 1000, (int)fps);
892 if (s->adaptive_quant)
893 adaptive_quantization(s, q);
898 rcc->last_qscale = q;
899 rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
900 rcc->last_mb_var_sum = pic->mb_var_sum;
905 // ----------------------------------------------
908 static int init_pass2(MpegEncContext *s)
910 RateControlContext *rcc = &s->rc_context;
911 AVCodecContext *a = s->avctx;
913 double fps = get_fps(s->avctx);
914 double complexity[5] = { 0 }; // approximate bits at quant=1
915 uint64_t const_bits[5] = { 0 }; // quantizer independent bits
916 uint64_t all_const_bits;
917 uint64_t all_available_bits = (uint64_t)(s->bit_rate *
918 (double)rcc->num_entries / fps);
919 double rate_factor = 0;
921 const int filter_size = (int)(a->qblur * 4) | 1;
922 double expected_bits;
923 double *qscale, *blurred_qscale, qscale_sum;
925 /* find complexity & const_bits & decide the pict_types */
926 for (i = 0; i < rcc->num_entries; i++) {
927 RateControlEntry *rce = &rcc->entry[i];
929 rce->new_pict_type = rce->pict_type;
930 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
931 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
932 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
933 rcc->frame_count[rce->pict_type]++;
935 complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
937 const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
940 all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
941 const_bits[AV_PICTURE_TYPE_P] +
942 const_bits[AV_PICTURE_TYPE_B];
944 if (all_available_bits < all_const_bits) {
945 av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
949 qscale = av_malloc(sizeof(double) * rcc->num_entries);
950 blurred_qscale = av_malloc(sizeof(double) * rcc->num_entries);
953 for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
957 rcc->buffer_index = s->avctx->rc_buffer_size / 2;
960 for (i = 0; i < rcc->num_entries; i++) {
961 RateControlEntry *rce = &rcc->entry[i];
963 qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
964 rcc->last_qscale_for[rce->pict_type] = qscale[i];
966 assert(filter_size % 2 == 1);
968 /* fixed I/B QP relative to P mode */
969 for (i = FFMAX(0, rcc->num_entries - 300); i < rcc->num_entries; i++) {
970 RateControlEntry *rce = &rcc->entry[i];
972 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
975 for (i = rcc->num_entries - 1; i >= 0; i--) {
976 RateControlEntry *rce = &rcc->entry[i];
978 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
982 for (i = 0; i < rcc->num_entries; i++) {
983 RateControlEntry *rce = &rcc->entry[i];
984 const int pict_type = rce->new_pict_type;
986 double q = 0.0, sum = 0.0;
988 for (j = 0; j < filter_size; j++) {
989 int index = i + j - filter_size / 2;
990 double d = index - i;
991 double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));
993 if (index < 0 || index >= rcc->num_entries)
995 if (pict_type != rcc->entry[index].new_pict_type)
997 q += qscale[index] * coeff;
1000 blurred_qscale[i] = q / sum;
1003 /* find expected bits */
1004 for (i = 0; i < rcc->num_entries; i++) {
1005 RateControlEntry *rce = &rcc->entry[i];
1008 rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);
1010 bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
1011 bits += 8 * ff_vbv_update(s, bits);
1013 rce->expected_bits = expected_bits;
1014 expected_bits += bits;
1018 "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
1019 expected_bits, (int)all_available_bits, rate_factor);
1020 if (expected_bits > all_available_bits) {
1021 rate_factor -= step;
1026 av_free(blurred_qscale);
1028 /* check bitrate calculations and print info */
1030 for (i = 0; i < rcc->num_entries; i++) {
1031 av_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
1033 rcc->entry[i].new_qscale,
1034 rcc->entry[i].new_qscale / FF_QP2LAMBDA);
1035 qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
1036 s->avctx->qmin, s->avctx->qmax);
1038 assert(toobig <= 40);
1039 av_log(s->avctx, AV_LOG_DEBUG,
1040 "[lavc rc] requested bitrate: %d bps expected bitrate: %d bps\n",
1042 (int)(expected_bits / ((double)all_available_bits / s->bit_rate)));
1043 av_log(s->avctx, AV_LOG_DEBUG,
1044 "[lavc rc] estimated target average qp: %.3f\n",
1045 (float)qscale_sum / rcc->num_entries);
1047 av_log(s->avctx, AV_LOG_INFO,
1048 "[lavc rc] Using all of requested bitrate is not "
1049 "necessary for this video with these parameters.\n");
1050 } else if (toobig == 40) {
1051 av_log(s->avctx, AV_LOG_ERROR,
1052 "[lavc rc] Error: bitrate too low for this video "
1053 "with these parameters.\n");
1055 } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
1056 av_log(s->avctx, AV_LOG_ERROR,
1057 "[lavc rc] Error: 2pass curve failed to converge\n");