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"
39 #define M_E 2.718281828
42 static int init_pass2(MpegEncContext *s);
43 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
44 double rate_factor, int frame_num);
46 void ff_write_pass1_stats(MpegEncContext *s)
48 snprintf(s->avctx->stats_out, 256,
49 "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d "
50 "fcode:%d bcode:%d mc-var:%"PRId64" var:%"PRId64" icount:%d skipcount:%d hbits:%d;\n",
51 s->current_picture_ptr->f->display_picture_number,
52 s->current_picture_ptr->f->coded_picture_number,
54 s->current_picture.f->quality,
61 s->current_picture.mc_mb_var_sum,
62 s->current_picture.mb_var_sum,
63 s->i_count, s->skip_count,
67 static double get_fps(AVCodecContext *avctx)
69 return 1.0 / av_q2d(avctx->time_base) / FFMAX(avctx->ticks_per_frame, 1);
72 static inline double qp2bits(RateControlEntry *rce, double qp)
75 av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
77 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp;
80 static inline double bits2qp(RateControlEntry *rce, double bits)
83 av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
85 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits;
88 av_cold int ff_rate_control_init(MpegEncContext *s)
90 RateControlContext *rcc = &s->rc_context;
92 static const char * const const_names[] = {
121 static double (* const func1[])(void *, double) = {
126 static const char * const func1_names[] = {
133 if (!s->avctx->rc_max_available_vbv_use && s->avctx->rc_buffer_size) {
134 if (s->avctx->rc_max_rate) {
135 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);
137 s->avctx->rc_max_available_vbv_use = 1.0;
140 res = av_expr_parse(&rcc->rc_eq_eval,
141 s->rc_eq ? s->rc_eq : "tex^qComp",
142 const_names, func1_names, func1,
143 NULL, NULL, 0, s->avctx);
145 av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->rc_eq);
149 #if FF_API_RC_STRATEGY
150 FF_DISABLE_DEPRECATION_WARNINGS
152 s->rc_strategy = s->avctx->rc_strategy;
153 FF_ENABLE_DEPRECATION_WARNINGS
156 for (i = 0; i < 5; i++) {
157 rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0;
158 rcc->pred[i].count = 1.0;
159 rcc->pred[i].decay = 0.4;
161 rcc->i_cplx_sum [i] =
162 rcc->p_cplx_sum [i] =
163 rcc->mv_bits_sum[i] =
164 rcc->qscale_sum [i] =
165 rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such
167 rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5;
169 rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy;
170 if (!rcc->buffer_index)
171 rcc->buffer_index = s->avctx->rc_buffer_size * 3 / 4;
173 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
177 /* find number of pics */
178 p = s->avctx->stats_in;
180 p = strchr(p + 1, ';');
181 i += s->max_b_frames;
182 if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry))
184 rcc->entry = av_mallocz(i * sizeof(RateControlEntry));
186 return AVERROR(ENOMEM);
187 rcc->num_entries = i;
189 /* init all to skipped p frames
190 * (with b frames we might have a not encoded frame at the end FIXME) */
191 for (i = 0; i < rcc->num_entries; i++) {
192 RateControlEntry *rce = &rcc->entry[i];
194 rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P;
195 rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2;
196 rce->misc_bits = s->mb_num + 10;
197 rce->mb_var_sum = s->mb_num * 100;
201 p = s->avctx->stats_in;
202 for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
203 RateControlEntry *rce;
208 next = strchr(p, ';');
210 (*next) = 0; // sscanf in unbelievably slow on looong strings // FIXME copy / do not write
213 e = sscanf(p, " in:%d ", &picture_number);
215 av_assert0(picture_number >= 0);
216 av_assert0(picture_number < rcc->num_entries);
217 rce = &rcc->entry[picture_number];
219 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",
220 &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
221 &rce->mv_bits, &rce->misc_bits,
222 &rce->f_code, &rce->b_code,
223 &rce->mc_mb_var_sum, &rce->mb_var_sum,
224 &rce->i_count, &rce->skip_count, &rce->header_bits);
226 av_log(s->avctx, AV_LOG_ERROR,
227 "statistics are damaged at line %d, parser out=%d\n",
235 if (init_pass2(s) < 0) {
236 ff_rate_control_uninit(s);
240 // FIXME maybe move to end
241 if ((s->avctx->flags & AV_CODEC_FLAG_PASS2) && s->rc_strategy == 1) {
243 return ff_xvid_rate_control_init(s);
245 av_log(s->avctx, AV_LOG_ERROR,
246 "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n");
252 if (!(s->avctx->flags & AV_CODEC_FLAG_PASS2)) {
253 rcc->short_term_qsum = 0.001;
254 rcc->short_term_qcount = 0.001;
256 rcc->pass1_rc_eq_output_sum = 0.001;
257 rcc->pass1_wanted_bits = 0.001;
259 if (s->avctx->qblur > 1.0) {
260 av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
263 /* init stuff with the user specified complexity */
264 if (s->rc_initial_cplx) {
265 for (i = 0; i < 60 * 30; i++) {
266 double bits = s->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
267 RateControlEntry rce;
269 if (i % ((s->gop_size + 3) / 4) == 0)
270 rce.pict_type = AV_PICTURE_TYPE_I;
271 else if (i % (s->max_b_frames + 1))
272 rce.pict_type = AV_PICTURE_TYPE_B;
274 rce.pict_type = AV_PICTURE_TYPE_P;
276 rce.new_pict_type = rce.pict_type;
277 rce.mc_mb_var_sum = bits * s->mb_num / 100000;
278 rce.mb_var_sum = s->mb_num;
280 rce.qscale = FF_QP2LAMBDA * 2;
285 if (s->pict_type == AV_PICTURE_TYPE_I) {
286 rce.i_count = s->mb_num;
287 rce.i_tex_bits = bits;
291 rce.i_count = 0; // FIXME we do know this approx
293 rce.p_tex_bits = bits * 0.9;
294 rce.mv_bits = bits * 0.1;
296 rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale;
297 rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale;
298 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
299 rcc->frame_count[rce.pict_type]++;
301 get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
303 // FIXME misbehaves a little for variable fps
304 rcc->pass1_wanted_bits += s->bit_rate / get_fps(s->avctx);
312 av_cold void ff_rate_control_uninit(MpegEncContext *s)
314 RateControlContext *rcc = &s->rc_context;
317 av_expr_free(rcc->rc_eq_eval);
318 av_freep(&rcc->entry);
321 if ((s->avctx->flags & AV_CODEC_FLAG_PASS2) && s->rc_strategy == 1)
322 ff_xvid_rate_control_uninit(s);
326 int ff_vbv_update(MpegEncContext *s, int frame_size)
328 RateControlContext *rcc = &s->rc_context;
329 const double fps = get_fps(s->avctx);
330 const int buffer_size = s->avctx->rc_buffer_size;
331 const double min_rate = s->avctx->rc_min_rate / fps;
332 const double max_rate = s->avctx->rc_max_rate / fps;
334 ff_dlog(s, "%d %f %d %f %f\n",
335 buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
340 rcc->buffer_index -= frame_size;
341 if (rcc->buffer_index < 0) {
342 av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
343 if (frame_size > max_rate && s->qscale == s->avctx->qmax) {
344 av_log(s->avctx, AV_LOG_ERROR, "max bitrate possibly too small or try trellis with large lmax or increase qmax\n");
346 rcc->buffer_index = 0;
349 left = buffer_size - rcc->buffer_index - 1;
350 rcc->buffer_index += av_clip(left, min_rate, max_rate);
352 if (rcc->buffer_index > buffer_size) {
353 int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
355 if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
357 rcc->buffer_index -= 8 * stuffing;
359 if (s->avctx->debug & FF_DEBUG_RC)
360 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
369 * Modify the bitrate curve from pass1 for one frame.
371 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
372 double rate_factor, int frame_num)
374 RateControlContext *rcc = &s->rc_context;
375 AVCodecContext *a = s->avctx;
376 const int pict_type = rce->new_pict_type;
377 const double mb_num = s->mb_num;
381 double const_values[] = {
384 rce->i_tex_bits * rce->qscale,
385 rce->p_tex_bits * rce->qscale,
386 (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
387 rce->mv_bits / mb_num,
388 rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
389 rce->i_count / mb_num,
390 rce->mc_mb_var_sum / mb_num,
391 rce->mb_var_sum / mb_num,
392 rce->pict_type == AV_PICTURE_TYPE_I,
393 rce->pict_type == AV_PICTURE_TYPE_P,
394 rce->pict_type == AV_PICTURE_TYPE_B,
395 rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
398 rcc->last_qscale_for[AV_PICTURE_TYPE_I],
399 rcc->last_qscale_for[AV_PICTURE_TYPE_P],
400 rcc->last_qscale_for[AV_PICTURE_TYPE_B],
401 rcc->next_non_b_qscale,
403 rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
404 rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
405 rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
406 rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
407 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
411 bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
413 av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->rc_eq);
417 rcc->pass1_rc_eq_output_sum += bits;
421 bits += 1.0; // avoid 1/0 issues
424 for (i = 0; i < s->avctx->rc_override_count; i++) {
425 RcOverride *rco = s->avctx->rc_override;
426 if (rco[i].start_frame > frame_num)
428 if (rco[i].end_frame < frame_num)
432 bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it?
434 bits *= rco[i].quality_factor;
437 q = bits2qp(rce, bits);
440 if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
441 q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
442 else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
443 q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
450 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
452 RateControlContext *rcc = &s->rc_context;
453 AVCodecContext *a = s->avctx;
454 const int pict_type = rce->new_pict_type;
455 const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
456 const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];
458 if (pict_type == AV_PICTURE_TYPE_I &&
459 (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P))
460 q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
461 else if (pict_type == AV_PICTURE_TYPE_B &&
462 a->b_quant_factor > 0.0)
463 q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
467 /* last qscale / qdiff stuff */
468 if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
469 double last_q = rcc->last_qscale_for[pict_type];
470 const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;
472 if (q > last_q + maxdiff)
473 q = last_q + maxdiff;
474 else if (q < last_q - maxdiff)
475 q = last_q - maxdiff;
478 rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring
480 if (pict_type != AV_PICTURE_TYPE_B)
481 rcc->last_non_b_pict_type = pict_type;
487 * Get the qmin & qmax for pict_type.
489 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
494 av_assert0(qmin <= qmax);
497 case AV_PICTURE_TYPE_B:
498 qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
499 qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
501 case AV_PICTURE_TYPE_I:
502 qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
503 qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
507 qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
508 qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);
517 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce,
518 double q, int frame_num)
520 RateControlContext *rcc = &s->rc_context;
521 const double buffer_size = s->avctx->rc_buffer_size;
522 const double fps = get_fps(s->avctx);
523 const double min_rate = s->avctx->rc_min_rate / fps;
524 const double max_rate = s->avctx->rc_max_rate / fps;
525 const int pict_type = rce->new_pict_type;
528 get_qminmax(&qmin, &qmax, s, pict_type);
531 if (s->rc_qmod_freq &&
532 frame_num % s->rc_qmod_freq == 0 &&
533 pict_type == AV_PICTURE_TYPE_P)
536 /* buffer overflow/underflow protection */
538 double expected_size = rcc->buffer_index;
542 double d = 2 * (buffer_size - expected_size) / buffer_size;
547 q *= pow(d, 1.0 / s->rc_buffer_aggressivity);
549 q_limit = bits2qp(rce,
550 FFMAX((min_rate - buffer_size + rcc->buffer_index) *
551 s->avctx->rc_min_vbv_overflow_use, 1));
554 if (s->avctx->debug & FF_DEBUG_RC)
555 av_log(s->avctx, AV_LOG_DEBUG,
556 "limiting QP %f -> %f\n", q, q_limit);
562 double d = 2 * expected_size / buffer_size;
567 q /= pow(d, 1.0 / s->rc_buffer_aggressivity);
569 q_limit = bits2qp(rce,
570 FFMAX(rcc->buffer_index *
571 s->avctx->rc_max_available_vbv_use,
574 if (s->avctx->debug & FF_DEBUG_RC)
575 av_log(s->avctx, AV_LOG_DEBUG,
576 "limiting QP %f -> %f\n", q, q_limit);
581 ff_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
582 q, max_rate, min_rate, buffer_size, rcc->buffer_index,
583 s->rc_buffer_aggressivity);
584 if (s->rc_qsquish == 0.0 || qmin == qmax) {
590 double min2 = log(qmin);
591 double max2 = log(qmax);
594 q = (q - min2) / (max2 - min2) - 0.5;
596 q = 1.0 / (1.0 + exp(q));
597 q = q * (max2 - min2) + min2;
605 // ----------------------------------
608 static double predict_size(Predictor *p, double q, double var)
610 return p->coeff * var / (q * p->count);
613 static void update_predictor(Predictor *p, double q, double var, double size)
615 double new_coeff = size * q / (var + 1);
619 p->count *= p->decay;
620 p->coeff *= p->decay;
622 p->coeff += new_coeff;
625 static void adaptive_quantization(MpegEncContext *s, double q)
628 const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0);
629 const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0);
630 const float temp_cplx_masking = s->avctx->temporal_cplx_masking;
631 const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
632 const float p_masking = s->avctx->p_masking;
633 const float border_masking = s->border_masking;
634 float bits_sum = 0.0;
635 float cplx_sum = 0.0;
636 float *cplx_tab = s->cplx_tab;
637 float *bits_tab = s->bits_tab;
638 const int qmin = s->avctx->mb_lmin;
639 const int qmax = s->avctx->mb_lmax;
640 Picture *const pic = &s->current_picture;
641 const int mb_width = s->mb_width;
642 const int mb_height = s->mb_height;
644 for (i = 0; i < s->mb_num; i++) {
645 const int mb_xy = s->mb_index2xy[i];
646 float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
647 float spat_cplx = sqrt(pic->mb_var[mb_xy]);
648 const int lumi = pic->mb_mean[mb_xy];
649 float bits, cplx, factor;
650 int mb_x = mb_xy % s->mb_stride;
651 int mb_y = mb_xy / s->mb_stride;
653 float mb_factor = 0.0;
655 spat_cplx = 4; // FIXME finetune
657 temp_cplx = 4; // FIXME finetune
659 if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
661 factor = 1.0 + p_masking;
664 factor = pow(temp_cplx, -temp_cplx_masking);
666 factor *= pow(spat_cplx, -spatial_cplx_masking);
669 factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
671 factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);
673 if (mb_x < mb_width / 5) {
674 mb_distance = mb_width / 5 - mb_x;
675 mb_factor = (float)mb_distance / (float)(mb_width / 5);
676 } else if (mb_x > 4 * mb_width / 5) {
677 mb_distance = mb_x - 4 * mb_width / 5;
678 mb_factor = (float)mb_distance / (float)(mb_width / 5);
680 if (mb_y < mb_height / 5) {
681 mb_distance = mb_height / 5 - mb_y;
682 mb_factor = FFMAX(mb_factor,
683 (float)mb_distance / (float)(mb_height / 5));
684 } else if (mb_y > 4 * mb_height / 5) {
685 mb_distance = mb_y - 4 * mb_height / 5;
686 mb_factor = FFMAX(mb_factor,
687 (float)mb_distance / (float)(mb_height / 5));
690 factor *= 1.0 - border_masking * mb_factor;
692 if (factor < 0.00001)
695 bits = cplx * factor;
702 /* handle qmin/qmax clipping */
703 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
704 float factor = bits_sum / cplx_sum;
705 for (i = 0; i < s->mb_num; i++) {
706 float newq = q * cplx_tab[i] / bits_tab[i];
710 bits_sum -= bits_tab[i];
711 cplx_sum -= cplx_tab[i] * q / qmax;
712 } else if (newq < qmin) {
713 bits_sum -= bits_tab[i];
714 cplx_sum -= cplx_tab[i] * q / qmin;
717 if (bits_sum < 0.001)
719 if (cplx_sum < 0.001)
723 for (i = 0; i < s->mb_num; i++) {
724 const int mb_xy = s->mb_index2xy[i];
725 float newq = q * cplx_tab[i] / bits_tab[i];
728 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
729 newq *= bits_sum / cplx_sum;
732 intq = (int)(newq + 0.5);
736 else if (intq < qmin)
738 s->lambda_table[mb_xy] = intq;
742 void ff_get_2pass_fcode(MpegEncContext *s)
744 RateControlContext *rcc = &s->rc_context;
745 RateControlEntry *rce = &rcc->entry[s->picture_number];
747 s->f_code = rce->f_code;
748 s->b_code = rce->b_code;
751 // FIXME rd or at least approx for dquant
753 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
757 float br_compensation;
761 int picture_number = s->picture_number;
763 RateControlContext *rcc = &s->rc_context;
764 AVCodecContext *a = s->avctx;
765 RateControlEntry local_rce, *rce;
769 const int pict_type = s->pict_type;
770 Picture * const pic = &s->current_picture;
774 if ((s->avctx->flags & AV_CODEC_FLAG_PASS2) && s->rc_strategy == 1)
775 return ff_xvid_rate_estimate_qscale(s, dry_run);
778 get_qminmax(&qmin, &qmax, s, pict_type);
780 fps = get_fps(s->avctx);
781 /* update predictors */
782 if (picture_number > 2 && !dry_run) {
783 const int64_t last_var =
784 s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
785 : rcc->last_mc_mb_var_sum;
786 av_assert1(s->frame_bits >= s->stuffing_bits);
787 update_predictor(&rcc->pred[s->last_pict_type],
790 s->frame_bits - s->stuffing_bits);
793 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
794 av_assert0(picture_number >= 0);
795 if (picture_number >= rcc->num_entries) {
796 av_log(s, AV_LOG_ERROR, "Input is longer than 2-pass log file\n");
799 rce = &rcc->entry[picture_number];
800 wanted_bits = rce->expected_bits;
805 /* FIXME add a dts field to AVFrame and ensure it is set and use it
806 * here instead of reordering but the reordering is simpler for now
807 * until H.264 B-pyramid must be handled. */
808 if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
809 dts_pic = s->current_picture_ptr;
811 dts_pic = s->last_picture_ptr;
813 if (!dts_pic || dts_pic->f->pts == AV_NOPTS_VALUE)
814 wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
816 wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f->pts / fps);
819 diff = s->total_bits - wanted_bits;
820 br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
821 if (br_compensation <= 0.0)
822 br_compensation = 0.001;
824 var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
826 short_term_q = 0; /* avoid warning */
827 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
828 if (pict_type != AV_PICTURE_TYPE_I)
829 av_assert0(pict_type == rce->new_pict_type);
831 q = rce->new_qscale / br_compensation;
832 ff_dlog(s, "%f %f %f last:%d var:%"PRId64" type:%d//\n", q, rce->new_qscale,
833 br_compensation, s->frame_bits, var, pict_type);
836 rce->new_pict_type = pict_type;
837 rce->mc_mb_var_sum = pic->mc_mb_var_sum;
838 rce->mb_var_sum = pic->mb_var_sum;
839 rce->qscale = FF_QP2LAMBDA * 2;
840 rce->f_code = s->f_code;
841 rce->b_code = s->b_code;
844 bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
845 if (pict_type == AV_PICTURE_TYPE_I) {
846 rce->i_count = s->mb_num;
847 rce->i_tex_bits = bits;
851 rce->i_count = 0; // FIXME we do know this approx
853 rce->p_tex_bits = bits * 0.9;
854 rce->mv_bits = bits * 0.1;
856 rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale;
857 rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale;
858 rcc->mv_bits_sum[pict_type] += rce->mv_bits;
859 rcc->frame_count[pict_type]++;
861 rate_factor = rcc->pass1_wanted_bits /
862 rcc->pass1_rc_eq_output_sum * br_compensation;
864 q = get_qscale(s, rce, rate_factor, picture_number);
869 q = get_diff_limited_q(s, rce, q);
872 // FIXME type dependent blur like in 2-pass
873 if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
874 rcc->short_term_qsum *= a->qblur;
875 rcc->short_term_qcount *= a->qblur;
877 rcc->short_term_qsum += q;
878 rcc->short_term_qcount++;
879 q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
883 q = modify_qscale(s, rce, q, picture_number);
885 rcc->pass1_wanted_bits += s->bit_rate / fps;
890 if (s->avctx->debug & FF_DEBUG_RC) {
891 av_log(s->avctx, AV_LOG_DEBUG,
892 "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
893 "size:%d var:%"PRId64"/%"PRId64" br:%"PRId64" fps:%d\n",
894 av_get_picture_type_char(pict_type),
895 qmin, q, qmax, picture_number,
896 (int)wanted_bits / 1000, (int)s->total_bits / 1000,
897 br_compensation, short_term_q, s->frame_bits,
898 pic->mb_var_sum, pic->mc_mb_var_sum,
899 s->bit_rate / 1000, (int)fps);
907 if (s->adaptive_quant)
908 adaptive_quantization(s, q);
913 rcc->last_qscale = q;
914 rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
915 rcc->last_mb_var_sum = pic->mb_var_sum;
920 // ----------------------------------------------
923 static int init_pass2(MpegEncContext *s)
925 RateControlContext *rcc = &s->rc_context;
926 AVCodecContext *a = s->avctx;
928 double fps = get_fps(s->avctx);
929 double complexity[5] = { 0 }; // approximate bits at quant=1
930 uint64_t const_bits[5] = { 0 }; // quantizer independent bits
931 uint64_t all_const_bits;
932 uint64_t all_available_bits = (uint64_t)(s->bit_rate *
933 (double)rcc->num_entries / fps);
934 double rate_factor = 0;
936 const int filter_size = (int)(a->qblur * 4) | 1;
937 double expected_bits = 0; // init to silence gcc warning
938 double *qscale, *blurred_qscale, qscale_sum;
940 /* find complexity & const_bits & decide the pict_types */
941 for (i = 0; i < rcc->num_entries; i++) {
942 RateControlEntry *rce = &rcc->entry[i];
944 rce->new_pict_type = rce->pict_type;
945 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
946 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
947 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
948 rcc->frame_count[rce->pict_type]++;
950 complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
952 const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
955 all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
956 const_bits[AV_PICTURE_TYPE_P] +
957 const_bits[AV_PICTURE_TYPE_B];
959 if (all_available_bits < all_const_bits) {
960 av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
964 qscale = av_malloc_array(rcc->num_entries, sizeof(double));
965 blurred_qscale = av_malloc_array(rcc->num_entries, sizeof(double));
966 if (!qscale || !blurred_qscale) {
968 av_free(blurred_qscale);
969 return AVERROR(ENOMEM);
973 for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
977 rcc->buffer_index = s->avctx->rc_buffer_size / 2;
980 for (i = 0; i < rcc->num_entries; i++) {
981 RateControlEntry *rce = &rcc->entry[i];
983 qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
984 rcc->last_qscale_for[rce->pict_type] = qscale[i];
986 av_assert0(filter_size % 2 == 1);
988 /* fixed I/B QP relative to P mode */
989 for (i = FFMAX(0, rcc->num_entries - 300); i < rcc->num_entries; i++) {
990 RateControlEntry *rce = &rcc->entry[i];
992 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
995 for (i = rcc->num_entries - 1; i >= 0; i--) {
996 RateControlEntry *rce = &rcc->entry[i];
998 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
1002 for (i = 0; i < rcc->num_entries; i++) {
1003 RateControlEntry *rce = &rcc->entry[i];
1004 const int pict_type = rce->new_pict_type;
1006 double q = 0.0, sum = 0.0;
1008 for (j = 0; j < filter_size; j++) {
1009 int index = i + j - filter_size / 2;
1010 double d = index - i;
1011 double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));
1013 if (index < 0 || index >= rcc->num_entries)
1015 if (pict_type != rcc->entry[index].new_pict_type)
1017 q += qscale[index] * coeff;
1020 blurred_qscale[i] = q / sum;
1023 /* find expected bits */
1024 for (i = 0; i < rcc->num_entries; i++) {
1025 RateControlEntry *rce = &rcc->entry[i];
1028 rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);
1030 bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
1031 bits += 8 * ff_vbv_update(s, bits);
1033 rce->expected_bits = expected_bits;
1034 expected_bits += bits;
1038 "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
1039 expected_bits, (int)all_available_bits, rate_factor);
1040 if (expected_bits > all_available_bits) {
1041 rate_factor -= step;
1046 av_free(blurred_qscale);
1048 /* check bitrate calculations and print info */
1050 for (i = 0; i < rcc->num_entries; i++) {
1051 ff_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
1053 rcc->entry[i].new_qscale,
1054 rcc->entry[i].new_qscale / FF_QP2LAMBDA);
1055 qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
1056 s->avctx->qmin, s->avctx->qmax);
1058 av_assert0(toobig <= 40);
1059 av_log(s->avctx, AV_LOG_DEBUG,
1060 "[lavc rc] requested bitrate: %"PRId64" bps expected bitrate: %"PRId64" bps\n",
1062 (int64_t)(expected_bits / ((double)all_available_bits / s->bit_rate)));
1063 av_log(s->avctx, AV_LOG_DEBUG,
1064 "[lavc rc] estimated target average qp: %.3f\n",
1065 (float)qscale_sum / rcc->num_entries);
1067 av_log(s->avctx, AV_LOG_INFO,
1068 "[lavc rc] Using all of requested bitrate is not "
1069 "necessary for this video with these parameters.\n");
1070 } else if (toobig == 40) {
1071 av_log(s->avctx, AV_LOG_ERROR,
1072 "[lavc rc] Error: bitrate too low for this video "
1073 "with these parameters.\n");
1075 } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
1076 av_log(s->avctx, AV_LOG_ERROR,
1077 "[lavc rc] Error: 2pass curve failed to converge\n");