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.
28 #include "libavutil/attributes.h"
29 #include "libavutil/internal.h"
33 #include "ratecontrol.h"
34 #include "mpegutils.h"
35 #include "mpegvideo.h"
36 #include "libavutil/eval.h"
38 static int init_pass2(MpegEncContext *s);
39 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
40 double rate_factor, int frame_num);
42 void ff_write_pass1_stats(MpegEncContext *s)
44 snprintf(s->avctx->stats_out, 256,
45 "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d "
46 "fcode:%d bcode:%d mc-var:%"PRId64" var:%"PRId64" icount:%d skipcount:%d hbits:%d;\n",
47 s->current_picture_ptr->f->display_picture_number,
48 s->current_picture_ptr->f->coded_picture_number,
50 s->current_picture.f->quality,
57 s->current_picture.mc_mb_var_sum,
58 s->current_picture.mb_var_sum,
59 s->i_count, s->skip_count,
63 static double get_fps(AVCodecContext *avctx)
65 return 1.0 / av_q2d(avctx->time_base) / FFMAX(avctx->ticks_per_frame, 1);
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 if (!s->avctx->rc_max_available_vbv_use && s->avctx->rc_buffer_size) {
130 if (s->avctx->rc_max_rate) {
131 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);
133 s->avctx->rc_max_available_vbv_use = 1.0;
136 res = av_expr_parse(&rcc->rc_eq_eval,
137 s->rc_eq ? s->rc_eq : "tex^qComp",
138 const_names, func1_names, func1,
139 NULL, NULL, 0, s->avctx);
141 av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->rc_eq);
145 #if FF_API_RC_STRATEGY
146 FF_DISABLE_DEPRECATION_WARNINGS
148 s->rc_strategy = s->avctx->rc_strategy;
149 FF_ENABLE_DEPRECATION_WARNINGS
152 for (i = 0; i < 5; i++) {
153 rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0;
154 rcc->pred[i].count = 1.0;
155 rcc->pred[i].decay = 0.4;
157 rcc->i_cplx_sum [i] =
158 rcc->p_cplx_sum [i] =
159 rcc->mv_bits_sum[i] =
160 rcc->qscale_sum [i] =
161 rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such
163 rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5;
165 rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy;
166 if (!rcc->buffer_index)
167 rcc->buffer_index = s->avctx->rc_buffer_size * 3 / 4;
169 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
173 /* find number of pics */
174 p = s->avctx->stats_in;
176 p = strchr(p + 1, ';');
177 i += s->max_b_frames;
178 if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry))
180 rcc->entry = av_mallocz(i * sizeof(RateControlEntry));
182 return AVERROR(ENOMEM);
183 rcc->num_entries = i;
185 /* init all to skipped p frames
186 * (with b frames we might have a not encoded frame at the end FIXME) */
187 for (i = 0; i < rcc->num_entries; i++) {
188 RateControlEntry *rce = &rcc->entry[i];
190 rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P;
191 rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2;
192 rce->misc_bits = s->mb_num + 10;
193 rce->mb_var_sum = s->mb_num * 100;
197 p = s->avctx->stats_in;
198 for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
199 RateControlEntry *rce;
204 next = strchr(p, ';');
206 (*next) = 0; // sscanf in unbelievably slow on looong strings // FIXME copy / do not write
209 e = sscanf(p, " in:%d ", &picture_number);
211 av_assert0(picture_number >= 0);
212 av_assert0(picture_number < rcc->num_entries);
213 rce = &rcc->entry[picture_number];
215 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:%"SCNd64" var:%"SCNd64" icount:%d skipcount:%d hbits:%d",
216 &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
217 &rce->mv_bits, &rce->misc_bits,
218 &rce->f_code, &rce->b_code,
219 &rce->mc_mb_var_sum, &rce->mb_var_sum,
220 &rce->i_count, &rce->skip_count, &rce->header_bits);
222 av_log(s->avctx, AV_LOG_ERROR,
223 "statistics are damaged at line %d, parser out=%d\n",
231 if (init_pass2(s) < 0) {
232 ff_rate_control_uninit(s);
236 // FIXME maybe move to end
237 if ((s->avctx->flags & AV_CODEC_FLAG_PASS2) && s->rc_strategy == 1) {
239 return ff_xvid_rate_control_init(s);
241 av_log(s->avctx, AV_LOG_ERROR,
242 "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n");
248 if (!(s->avctx->flags & AV_CODEC_FLAG_PASS2)) {
249 rcc->short_term_qsum = 0.001;
250 rcc->short_term_qcount = 0.001;
252 rcc->pass1_rc_eq_output_sum = 0.001;
253 rcc->pass1_wanted_bits = 0.001;
255 if (s->avctx->qblur > 1.0) {
256 av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
259 /* init stuff with the user specified complexity */
260 if (s->rc_initial_cplx) {
261 for (i = 0; i < 60 * 30; i++) {
262 double bits = s->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
263 RateControlEntry rce;
265 if (i % ((s->gop_size + 3) / 4) == 0)
266 rce.pict_type = AV_PICTURE_TYPE_I;
267 else if (i % (s->max_b_frames + 1))
268 rce.pict_type = AV_PICTURE_TYPE_B;
270 rce.pict_type = AV_PICTURE_TYPE_P;
272 rce.new_pict_type = rce.pict_type;
273 rce.mc_mb_var_sum = bits * s->mb_num / 100000;
274 rce.mb_var_sum = s->mb_num;
276 rce.qscale = FF_QP2LAMBDA * 2;
281 if (s->pict_type == AV_PICTURE_TYPE_I) {
282 rce.i_count = s->mb_num;
283 rce.i_tex_bits = bits;
287 rce.i_count = 0; // FIXME we do know this approx
289 rce.p_tex_bits = bits * 0.9;
290 rce.mv_bits = bits * 0.1;
292 rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale;
293 rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale;
294 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
295 rcc->frame_count[rce.pict_type]++;
297 get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
299 // FIXME misbehaves a little for variable fps
300 rcc->pass1_wanted_bits += s->bit_rate / get_fps(s->avctx);
308 av_cold void ff_rate_control_uninit(MpegEncContext *s)
310 RateControlContext *rcc = &s->rc_context;
313 av_expr_free(rcc->rc_eq_eval);
314 av_freep(&rcc->entry);
317 if ((s->avctx->flags & AV_CODEC_FLAG_PASS2) && s->rc_strategy == 1)
318 ff_xvid_rate_control_uninit(s);
322 int ff_vbv_update(MpegEncContext *s, int frame_size)
324 RateControlContext *rcc = &s->rc_context;
325 const double fps = get_fps(s->avctx);
326 const int buffer_size = s->avctx->rc_buffer_size;
327 const double min_rate = s->avctx->rc_min_rate / fps;
328 const double max_rate = s->avctx->rc_max_rate / fps;
330 ff_dlog(s, "%d %f %d %f %f\n",
331 buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
336 rcc->buffer_index -= frame_size;
337 if (rcc->buffer_index < 0) {
338 av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
339 if (frame_size > max_rate && s->qscale == s->avctx->qmax) {
340 av_log(s->avctx, AV_LOG_ERROR, "max bitrate possibly too small or try trellis with large lmax or increase qmax\n");
342 rcc->buffer_index = 0;
345 left = buffer_size - rcc->buffer_index - 1;
346 rcc->buffer_index += av_clip(left, min_rate, max_rate);
348 if (rcc->buffer_index > buffer_size) {
349 int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
351 if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
353 rcc->buffer_index -= 8 * stuffing;
355 if (s->avctx->debug & FF_DEBUG_RC)
356 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
365 * Modify the bitrate curve from pass1 for one frame.
367 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
368 double rate_factor, int frame_num)
370 RateControlContext *rcc = &s->rc_context;
371 AVCodecContext *a = s->avctx;
372 const int pict_type = rce->new_pict_type;
373 const double mb_num = s->mb_num;
377 double const_values[] = {
380 rce->i_tex_bits * rce->qscale,
381 rce->p_tex_bits * rce->qscale,
382 (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
383 rce->mv_bits / mb_num,
384 rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
385 rce->i_count / mb_num,
386 rce->mc_mb_var_sum / mb_num,
387 rce->mb_var_sum / mb_num,
388 rce->pict_type == AV_PICTURE_TYPE_I,
389 rce->pict_type == AV_PICTURE_TYPE_P,
390 rce->pict_type == AV_PICTURE_TYPE_B,
391 rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
394 rcc->last_qscale_for[AV_PICTURE_TYPE_I],
395 rcc->last_qscale_for[AV_PICTURE_TYPE_P],
396 rcc->last_qscale_for[AV_PICTURE_TYPE_B],
397 rcc->next_non_b_qscale,
399 rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
400 rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
401 rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
402 rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
403 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
407 bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
409 av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->rc_eq);
413 rcc->pass1_rc_eq_output_sum += bits;
417 bits += 1.0; // avoid 1/0 issues
420 for (i = 0; i < s->avctx->rc_override_count; i++) {
421 RcOverride *rco = s->avctx->rc_override;
422 if (rco[i].start_frame > frame_num)
424 if (rco[i].end_frame < frame_num)
428 bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it?
430 bits *= rco[i].quality_factor;
433 q = bits2qp(rce, bits);
436 if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
437 q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
438 else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
439 q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
446 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
448 RateControlContext *rcc = &s->rc_context;
449 AVCodecContext *a = s->avctx;
450 const int pict_type = rce->new_pict_type;
451 const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
452 const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];
454 if (pict_type == AV_PICTURE_TYPE_I &&
455 (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P))
456 q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
457 else if (pict_type == AV_PICTURE_TYPE_B &&
458 a->b_quant_factor > 0.0)
459 q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
463 /* last qscale / qdiff stuff */
464 if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
465 double last_q = rcc->last_qscale_for[pict_type];
466 const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;
468 if (q > last_q + maxdiff)
469 q = last_q + maxdiff;
470 else if (q < last_q - maxdiff)
471 q = last_q - maxdiff;
474 rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring
476 if (pict_type != AV_PICTURE_TYPE_B)
477 rcc->last_non_b_pict_type = pict_type;
483 * Get the qmin & qmax for pict_type.
485 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
490 av_assert0(qmin <= qmax);
493 case AV_PICTURE_TYPE_B:
494 qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
495 qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
497 case AV_PICTURE_TYPE_I:
498 qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
499 qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
503 qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
504 qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);
513 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce,
514 double q, int frame_num)
516 RateControlContext *rcc = &s->rc_context;
517 const double buffer_size = s->avctx->rc_buffer_size;
518 const double fps = get_fps(s->avctx);
519 const double min_rate = s->avctx->rc_min_rate / fps;
520 const double max_rate = s->avctx->rc_max_rate / fps;
521 const int pict_type = rce->new_pict_type;
524 get_qminmax(&qmin, &qmax, s, pict_type);
527 if (s->rc_qmod_freq &&
528 frame_num % s->rc_qmod_freq == 0 &&
529 pict_type == AV_PICTURE_TYPE_P)
532 /* buffer overflow/underflow protection */
534 double expected_size = rcc->buffer_index;
538 double d = 2 * (buffer_size - expected_size) / buffer_size;
543 q *= pow(d, 1.0 / s->rc_buffer_aggressivity);
545 q_limit = bits2qp(rce,
546 FFMAX((min_rate - buffer_size + rcc->buffer_index) *
547 s->avctx->rc_min_vbv_overflow_use, 1));
550 if (s->avctx->debug & FF_DEBUG_RC)
551 av_log(s->avctx, AV_LOG_DEBUG,
552 "limiting QP %f -> %f\n", q, q_limit);
558 double d = 2 * expected_size / buffer_size;
563 q /= pow(d, 1.0 / s->rc_buffer_aggressivity);
565 q_limit = bits2qp(rce,
566 FFMAX(rcc->buffer_index *
567 s->avctx->rc_max_available_vbv_use,
570 if (s->avctx->debug & FF_DEBUG_RC)
571 av_log(s->avctx, AV_LOG_DEBUG,
572 "limiting QP %f -> %f\n", q, q_limit);
577 ff_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
578 q, max_rate, min_rate, buffer_size, rcc->buffer_index,
579 s->rc_buffer_aggressivity);
580 if (s->rc_qsquish == 0.0 || qmin == qmax) {
586 double min2 = log(qmin);
587 double max2 = log(qmax);
590 q = (q - min2) / (max2 - min2) - 0.5;
592 q = 1.0 / (1.0 + exp(q));
593 q = q * (max2 - min2) + min2;
601 // ----------------------------------
604 static double predict_size(Predictor *p, double q, double var)
606 return p->coeff * var / (q * p->count);
609 static void update_predictor(Predictor *p, double q, double var, double size)
611 double new_coeff = size * q / (var + 1);
615 p->count *= p->decay;
616 p->coeff *= p->decay;
618 p->coeff += new_coeff;
621 static void adaptive_quantization(MpegEncContext *s, double q)
624 const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0);
625 const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0);
626 const float temp_cplx_masking = s->avctx->temporal_cplx_masking;
627 const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
628 const float p_masking = s->avctx->p_masking;
629 const float border_masking = s->border_masking;
630 float bits_sum = 0.0;
631 float cplx_sum = 0.0;
632 float *cplx_tab = s->cplx_tab;
633 float *bits_tab = s->bits_tab;
634 const int qmin = s->avctx->mb_lmin;
635 const int qmax = s->avctx->mb_lmax;
636 Picture *const pic = &s->current_picture;
637 const int mb_width = s->mb_width;
638 const int mb_height = s->mb_height;
640 for (i = 0; i < s->mb_num; i++) {
641 const int mb_xy = s->mb_index2xy[i];
642 float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
643 float spat_cplx = sqrt(pic->mb_var[mb_xy]);
644 const int lumi = pic->mb_mean[mb_xy];
645 float bits, cplx, factor;
646 int mb_x = mb_xy % s->mb_stride;
647 int mb_y = mb_xy / s->mb_stride;
649 float mb_factor = 0.0;
651 spat_cplx = 4; // FIXME finetune
653 temp_cplx = 4; // FIXME finetune
655 if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
657 factor = 1.0 + p_masking;
660 factor = pow(temp_cplx, -temp_cplx_masking);
662 factor *= pow(spat_cplx, -spatial_cplx_masking);
665 factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
667 factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);
669 if (mb_x < mb_width / 5) {
670 mb_distance = mb_width / 5 - mb_x;
671 mb_factor = (float)mb_distance / (float)(mb_width / 5);
672 } else if (mb_x > 4 * mb_width / 5) {
673 mb_distance = mb_x - 4 * mb_width / 5;
674 mb_factor = (float)mb_distance / (float)(mb_width / 5);
676 if (mb_y < mb_height / 5) {
677 mb_distance = mb_height / 5 - mb_y;
678 mb_factor = FFMAX(mb_factor,
679 (float)mb_distance / (float)(mb_height / 5));
680 } else if (mb_y > 4 * mb_height / 5) {
681 mb_distance = mb_y - 4 * mb_height / 5;
682 mb_factor = FFMAX(mb_factor,
683 (float)mb_distance / (float)(mb_height / 5));
686 factor *= 1.0 - border_masking * mb_factor;
688 if (factor < 0.00001)
691 bits = cplx * factor;
698 /* handle qmin/qmax clipping */
699 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
700 float factor = bits_sum / cplx_sum;
701 for (i = 0; i < s->mb_num; i++) {
702 float newq = q * cplx_tab[i] / bits_tab[i];
706 bits_sum -= bits_tab[i];
707 cplx_sum -= cplx_tab[i] * q / qmax;
708 } else if (newq < qmin) {
709 bits_sum -= bits_tab[i];
710 cplx_sum -= cplx_tab[i] * q / qmin;
713 if (bits_sum < 0.001)
715 if (cplx_sum < 0.001)
719 for (i = 0; i < s->mb_num; i++) {
720 const int mb_xy = s->mb_index2xy[i];
721 float newq = q * cplx_tab[i] / bits_tab[i];
724 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
725 newq *= bits_sum / cplx_sum;
728 intq = (int)(newq + 0.5);
732 else if (intq < qmin)
734 s->lambda_table[mb_xy] = intq;
738 void ff_get_2pass_fcode(MpegEncContext *s)
740 RateControlContext *rcc = &s->rc_context;
741 RateControlEntry *rce = &rcc->entry[s->picture_number];
743 s->f_code = rce->f_code;
744 s->b_code = rce->b_code;
747 // FIXME rd or at least approx for dquant
749 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
753 float br_compensation;
757 int picture_number = s->picture_number;
759 RateControlContext *rcc = &s->rc_context;
760 AVCodecContext *a = s->avctx;
761 RateControlEntry local_rce, *rce;
765 const int pict_type = s->pict_type;
766 Picture * const pic = &s->current_picture;
770 if ((s->avctx->flags & AV_CODEC_FLAG_PASS2) && s->rc_strategy == 1)
771 return ff_xvid_rate_estimate_qscale(s, dry_run);
774 get_qminmax(&qmin, &qmax, s, pict_type);
776 fps = get_fps(s->avctx);
777 /* update predictors */
778 if (picture_number > 2 && !dry_run) {
779 const int64_t last_var =
780 s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
781 : rcc->last_mc_mb_var_sum;
782 av_assert1(s->frame_bits >= s->stuffing_bits);
783 update_predictor(&rcc->pred[s->last_pict_type],
786 s->frame_bits - s->stuffing_bits);
789 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
790 av_assert0(picture_number >= 0);
791 if (picture_number >= rcc->num_entries) {
792 av_log(s, AV_LOG_ERROR, "Input is longer than 2-pass log file\n");
795 rce = &rcc->entry[picture_number];
796 wanted_bits = rce->expected_bits;
801 /* FIXME add a dts field to AVFrame and ensure it is set and use it
802 * here instead of reordering but the reordering is simpler for now
803 * until H.264 B-pyramid must be handled. */
804 if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
805 dts_pic = s->current_picture_ptr;
807 dts_pic = s->last_picture_ptr;
809 if (!dts_pic || dts_pic->f->pts == AV_NOPTS_VALUE)
810 wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
812 wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f->pts / fps);
815 diff = s->total_bits - wanted_bits;
816 br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
817 if (br_compensation <= 0.0)
818 br_compensation = 0.001;
820 var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
822 short_term_q = 0; /* avoid warning */
823 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
824 if (pict_type != AV_PICTURE_TYPE_I)
825 av_assert0(pict_type == rce->new_pict_type);
827 q = rce->new_qscale / br_compensation;
828 ff_dlog(s, "%f %f %f last:%d var:%"PRId64" type:%d//\n", q, rce->new_qscale,
829 br_compensation, s->frame_bits, var, pict_type);
832 rce->new_pict_type = pict_type;
833 rce->mc_mb_var_sum = pic->mc_mb_var_sum;
834 rce->mb_var_sum = pic->mb_var_sum;
835 rce->qscale = FF_QP2LAMBDA * 2;
836 rce->f_code = s->f_code;
837 rce->b_code = s->b_code;
840 bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
841 if (pict_type == AV_PICTURE_TYPE_I) {
842 rce->i_count = s->mb_num;
843 rce->i_tex_bits = bits;
847 rce->i_count = 0; // FIXME we do know this approx
849 rce->p_tex_bits = bits * 0.9;
850 rce->mv_bits = bits * 0.1;
852 rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale;
853 rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale;
854 rcc->mv_bits_sum[pict_type] += rce->mv_bits;
855 rcc->frame_count[pict_type]++;
857 rate_factor = rcc->pass1_wanted_bits /
858 rcc->pass1_rc_eq_output_sum * br_compensation;
860 q = get_qscale(s, rce, rate_factor, picture_number);
865 q = get_diff_limited_q(s, rce, q);
868 // FIXME type dependent blur like in 2-pass
869 if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
870 rcc->short_term_qsum *= a->qblur;
871 rcc->short_term_qcount *= a->qblur;
873 rcc->short_term_qsum += q;
874 rcc->short_term_qcount++;
875 q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
879 q = modify_qscale(s, rce, q, picture_number);
881 rcc->pass1_wanted_bits += s->bit_rate / fps;
886 if (s->avctx->debug & FF_DEBUG_RC) {
887 av_log(s->avctx, AV_LOG_DEBUG,
888 "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
889 "size:%d var:%"PRId64"/%"PRId64" br:%"PRId64" fps:%d\n",
890 av_get_picture_type_char(pict_type),
891 qmin, q, qmax, picture_number,
892 (int)wanted_bits / 1000, (int)s->total_bits / 1000,
893 br_compensation, short_term_q, s->frame_bits,
894 pic->mb_var_sum, pic->mc_mb_var_sum,
895 s->bit_rate / 1000, (int)fps);
903 if (s->adaptive_quant)
904 adaptive_quantization(s, q);
909 rcc->last_qscale = q;
910 rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
911 rcc->last_mb_var_sum = pic->mb_var_sum;
916 // ----------------------------------------------
919 static int init_pass2(MpegEncContext *s)
921 RateControlContext *rcc = &s->rc_context;
922 AVCodecContext *a = s->avctx;
924 double fps = get_fps(s->avctx);
925 double complexity[5] = { 0 }; // approximate bits at quant=1
926 uint64_t const_bits[5] = { 0 }; // quantizer independent bits
927 uint64_t all_const_bits;
928 uint64_t all_available_bits = (uint64_t)(s->bit_rate *
929 (double)rcc->num_entries / fps);
930 double rate_factor = 0;
932 const int filter_size = (int)(a->qblur * 4) | 1;
933 double expected_bits = 0; // init to silence gcc warning
934 double *qscale, *blurred_qscale, qscale_sum;
936 /* find complexity & const_bits & decide the pict_types */
937 for (i = 0; i < rcc->num_entries; i++) {
938 RateControlEntry *rce = &rcc->entry[i];
940 rce->new_pict_type = rce->pict_type;
941 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
942 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
943 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
944 rcc->frame_count[rce->pict_type]++;
946 complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
948 const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
951 all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
952 const_bits[AV_PICTURE_TYPE_P] +
953 const_bits[AV_PICTURE_TYPE_B];
955 if (all_available_bits < all_const_bits) {
956 av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
960 qscale = av_malloc_array(rcc->num_entries, sizeof(double));
961 blurred_qscale = av_malloc_array(rcc->num_entries, sizeof(double));
962 if (!qscale || !blurred_qscale) {
964 av_free(blurred_qscale);
965 return AVERROR(ENOMEM);
969 for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
973 rcc->buffer_index = s->avctx->rc_buffer_size / 2;
976 for (i = 0; i < rcc->num_entries; i++) {
977 RateControlEntry *rce = &rcc->entry[i];
979 qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
980 rcc->last_qscale_for[rce->pict_type] = qscale[i];
982 av_assert0(filter_size % 2 == 1);
984 /* fixed I/B QP relative to P mode */
985 for (i = FFMAX(0, rcc->num_entries - 300); i < rcc->num_entries; i++) {
986 RateControlEntry *rce = &rcc->entry[i];
988 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
991 for (i = rcc->num_entries - 1; i >= 0; i--) {
992 RateControlEntry *rce = &rcc->entry[i];
994 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
998 for (i = 0; i < rcc->num_entries; i++) {
999 RateControlEntry *rce = &rcc->entry[i];
1000 const int pict_type = rce->new_pict_type;
1002 double q = 0.0, sum = 0.0;
1004 for (j = 0; j < filter_size; j++) {
1005 int index = i + j - filter_size / 2;
1006 double d = index - i;
1007 double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));
1009 if (index < 0 || index >= rcc->num_entries)
1011 if (pict_type != rcc->entry[index].new_pict_type)
1013 q += qscale[index] * coeff;
1016 blurred_qscale[i] = q / sum;
1019 /* find expected bits */
1020 for (i = 0; i < rcc->num_entries; i++) {
1021 RateControlEntry *rce = &rcc->entry[i];
1024 rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);
1026 bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
1027 bits += 8 * ff_vbv_update(s, bits);
1029 rce->expected_bits = expected_bits;
1030 expected_bits += bits;
1034 "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
1035 expected_bits, (int)all_available_bits, rate_factor);
1036 if (expected_bits > all_available_bits) {
1037 rate_factor -= step;
1042 av_free(blurred_qscale);
1044 /* check bitrate calculations and print info */
1046 for (i = 0; i < rcc->num_entries; i++) {
1047 ff_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
1049 rcc->entry[i].new_qscale,
1050 rcc->entry[i].new_qscale / FF_QP2LAMBDA);
1051 qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
1052 s->avctx->qmin, s->avctx->qmax);
1054 av_assert0(toobig <= 40);
1055 av_log(s->avctx, AV_LOG_DEBUG,
1056 "[lavc rc] requested bitrate: %"PRId64" bps expected bitrate: %"PRId64" bps\n",
1058 (int64_t)(expected_bits / ((double)all_available_bits / s->bit_rate)));
1059 av_log(s->avctx, AV_LOG_DEBUG,
1060 "[lavc rc] estimated target average qp: %.3f\n",
1061 (float)qscale_sum / rcc->num_entries);
1063 av_log(s->avctx, AV_LOG_INFO,
1064 "[lavc rc] Using all of requested bitrate is not "
1065 "necessary for this video with these parameters.\n");
1066 } else if (toobig == 40) {
1067 av_log(s->avctx, AV_LOG_ERROR,
1068 "[lavc rc] Error: bitrate too low for this video "
1069 "with these parameters.\n");
1071 } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
1072 av_log(s->avctx, AV_LOG_ERROR,
1073 "[lavc rc] Error: 2pass curve failed to converge\n");