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"
30 #include "ratecontrol.h"
31 #include "mpegvideo.h"
32 #include "libavutil/eval.h"
34 #undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
38 #define M_E 2.718281828
41 static int init_pass2(MpegEncContext *s);
42 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
43 double rate_factor, int frame_num);
45 void ff_write_pass1_stats(MpegEncContext *s)
47 snprintf(s->avctx->stats_out, 256,
48 "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d "
49 "fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d;\n",
50 s->current_picture_ptr->f.display_picture_number,
51 s->current_picture_ptr->f.coded_picture_number,
53 s->current_picture.f.quality,
60 s->current_picture.mc_mb_var_sum,
61 s->current_picture.mb_var_sum,
62 s->i_count, s->skip_count,
66 static inline double qp2bits(RateControlEntry *rce, double qp)
69 av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
71 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp;
74 static inline double bits2qp(RateControlEntry *rce, double bits)
77 av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
79 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits;
82 av_cold int ff_rate_control_init(MpegEncContext *s)
84 RateControlContext *rcc = &s->rc_context;
86 static const char * const const_names[] = {
115 static double (* const func1[])(void *, double) = {
120 static const char * const func1_names[] = {
127 res = av_expr_parse(&rcc->rc_eq_eval,
128 s->avctx->rc_eq ? s->avctx->rc_eq : "tex^qComp",
129 const_names, func1_names, func1,
130 NULL, NULL, 0, s->avctx);
132 av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->avctx->rc_eq);
136 for (i = 0; i < 5; i++) {
137 rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0;
138 rcc->pred[i].count = 1.0;
139 rcc->pred[i].decay = 0.4;
141 rcc->i_cplx_sum [i] =
142 rcc->p_cplx_sum [i] =
143 rcc->mv_bits_sum[i] =
144 rcc->qscale_sum [i] =
145 rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such
147 rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5;
149 rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy;
151 if (s->flags & CODEC_FLAG_PASS2) {
155 /* find number of pics */
156 p = s->avctx->stats_in;
158 p = strchr(p + 1, ';');
159 i += s->max_b_frames;
160 if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry))
162 rcc->entry = av_mallocz(i * sizeof(RateControlEntry));
163 rcc->num_entries = i;
165 /* init all to skipped p frames
166 * (with b frames we might have a not encoded frame at the end FIXME) */
167 for (i = 0; i < rcc->num_entries; i++) {
168 RateControlEntry *rce = &rcc->entry[i];
170 rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P;
171 rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2;
172 rce->misc_bits = s->mb_num + 10;
173 rce->mb_var_sum = s->mb_num * 100;
177 p = s->avctx->stats_in;
178 for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
179 RateControlEntry *rce;
184 next = strchr(p, ';');
186 (*next) = 0; // sscanf in unbelievably slow on looong strings // FIXME copy / do not write
189 e = sscanf(p, " in:%d ", &picture_number);
191 assert(picture_number >= 0);
192 assert(picture_number < rcc->num_entries);
193 rce = &rcc->entry[picture_number];
195 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",
196 &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
197 &rce->mv_bits, &rce->misc_bits,
198 &rce->f_code, &rce->b_code,
199 &rce->mc_mb_var_sum, &rce->mb_var_sum,
200 &rce->i_count, &rce->skip_count, &rce->header_bits);
202 av_log(s->avctx, AV_LOG_ERROR,
203 "statistics are damaged at line %d, parser out=%d\n",
211 if (init_pass2(s) < 0)
214 // FIXME maybe move to end
215 if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) {
217 return ff_xvid_rate_control_init(s);
219 av_log(s->avctx, AV_LOG_ERROR,
220 "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n");
226 if (!(s->flags & CODEC_FLAG_PASS2)) {
227 rcc->short_term_qsum = 0.001;
228 rcc->short_term_qcount = 0.001;
230 rcc->pass1_rc_eq_output_sum = 0.001;
231 rcc->pass1_wanted_bits = 0.001;
233 if (s->avctx->qblur > 1.0) {
234 av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
237 /* init stuff with the user specified complexity */
238 if (s->avctx->rc_initial_cplx) {
239 for (i = 0; i < 60 * 30; i++) {
240 double bits = s->avctx->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
241 RateControlEntry rce;
243 if (i % ((s->gop_size + 3) / 4) == 0)
244 rce.pict_type = AV_PICTURE_TYPE_I;
245 else if (i % (s->max_b_frames + 1))
246 rce.pict_type = AV_PICTURE_TYPE_B;
248 rce.pict_type = AV_PICTURE_TYPE_P;
250 rce.new_pict_type = rce.pict_type;
251 rce.mc_mb_var_sum = bits * s->mb_num / 100000;
252 rce.mb_var_sum = s->mb_num;
254 rce.qscale = FF_QP2LAMBDA * 2;
259 if (s->pict_type == AV_PICTURE_TYPE_I) {
260 rce.i_count = s->mb_num;
261 rce.i_tex_bits = bits;
265 rce.i_count = 0; // FIXME we do know this approx
267 rce.p_tex_bits = bits * 0.9;
268 rce.mv_bits = bits * 0.1;
270 rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale;
271 rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale;
272 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
273 rcc->frame_count[rce.pict_type]++;
275 get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
277 // FIXME misbehaves a little for variable fps
278 rcc->pass1_wanted_bits += s->bit_rate / (1 / av_q2d(s->avctx->time_base));
286 av_cold void ff_rate_control_uninit(MpegEncContext *s)
288 RateControlContext *rcc = &s->rc_context;
291 av_expr_free(rcc->rc_eq_eval);
292 av_freep(&rcc->entry);
295 if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
296 ff_xvid_rate_control_uninit(s);
300 int ff_vbv_update(MpegEncContext *s, int frame_size)
302 RateControlContext *rcc = &s->rc_context;
303 const double fps = 1 / av_q2d(s->avctx->time_base);
304 const int buffer_size = s->avctx->rc_buffer_size;
305 const double min_rate = s->avctx->rc_min_rate / fps;
306 const double max_rate = s->avctx->rc_max_rate / fps;
308 av_dlog(s, "%d %f %d %f %f\n",
309 buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
314 rcc->buffer_index -= frame_size;
315 if (rcc->buffer_index < 0) {
316 av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
317 rcc->buffer_index = 0;
320 left = buffer_size - rcc->buffer_index - 1;
321 rcc->buffer_index += av_clip(left, min_rate, max_rate);
323 if (rcc->buffer_index > buffer_size) {
324 int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
326 if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
328 rcc->buffer_index -= 8 * stuffing;
330 if (s->avctx->debug & FF_DEBUG_RC)
331 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
340 * Modify the bitrate curve from pass1 for one frame.
342 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
343 double rate_factor, int frame_num)
345 RateControlContext *rcc = &s->rc_context;
346 AVCodecContext *a = s->avctx;
347 const int pict_type = rce->new_pict_type;
348 const double mb_num = s->mb_num;
352 double const_values[] = {
355 rce->i_tex_bits * rce->qscale,
356 rce->p_tex_bits * rce->qscale,
357 (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
358 rce->mv_bits / mb_num,
359 rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
360 rce->i_count / mb_num,
361 rce->mc_mb_var_sum / mb_num,
362 rce->mb_var_sum / mb_num,
363 rce->pict_type == AV_PICTURE_TYPE_I,
364 rce->pict_type == AV_PICTURE_TYPE_P,
365 rce->pict_type == AV_PICTURE_TYPE_B,
366 rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
369 rcc->last_qscale_for[AV_PICTURE_TYPE_I],
370 rcc->last_qscale_for[AV_PICTURE_TYPE_P],
371 rcc->last_qscale_for[AV_PICTURE_TYPE_B],
372 rcc->next_non_b_qscale,
374 rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
375 rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
376 rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
377 rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
378 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
382 bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
384 av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq);
388 rcc->pass1_rc_eq_output_sum += bits;
392 bits += 1.0; // avoid 1/0 issues
395 for (i = 0; i < s->avctx->rc_override_count; i++) {
396 RcOverride *rco = s->avctx->rc_override;
397 if (rco[i].start_frame > frame_num)
399 if (rco[i].end_frame < frame_num)
403 bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it?
405 bits *= rco[i].quality_factor;
408 q = bits2qp(rce, bits);
411 if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
412 q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
413 else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
414 q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
421 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
423 RateControlContext *rcc = &s->rc_context;
424 AVCodecContext *a = s->avctx;
425 const int pict_type = rce->new_pict_type;
426 const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
427 const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];
429 if (pict_type == AV_PICTURE_TYPE_I &&
430 (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P))
431 q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
432 else if (pict_type == AV_PICTURE_TYPE_B &&
433 a->b_quant_factor > 0.0)
434 q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
438 /* last qscale / qdiff stuff */
439 if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
440 double last_q = rcc->last_qscale_for[pict_type];
441 const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;
443 if (q > last_q + maxdiff)
444 q = last_q + maxdiff;
445 else if (q < last_q - maxdiff)
446 q = last_q - maxdiff;
449 rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring
451 if (pict_type != AV_PICTURE_TYPE_B)
452 rcc->last_non_b_pict_type = pict_type;
458 * Get the qmin & qmax for pict_type.
460 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
462 int qmin = s->avctx->lmin;
463 int qmax = s->avctx->lmax;
465 assert(qmin <= qmax);
468 case AV_PICTURE_TYPE_B:
469 qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
470 qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
472 case AV_PICTURE_TYPE_I:
473 qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
474 qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
478 qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
479 qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);
488 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce,
489 double q, int frame_num)
491 RateControlContext *rcc = &s->rc_context;
492 const double buffer_size = s->avctx->rc_buffer_size;
493 const double fps = 1 / av_q2d(s->avctx->time_base);
494 const double min_rate = s->avctx->rc_min_rate / fps;
495 const double max_rate = s->avctx->rc_max_rate / fps;
496 const int pict_type = rce->new_pict_type;
499 get_qminmax(&qmin, &qmax, s, pict_type);
502 if (s->avctx->rc_qmod_freq &&
503 frame_num % s->avctx->rc_qmod_freq == 0 &&
504 pict_type == AV_PICTURE_TYPE_P)
505 q *= s->avctx->rc_qmod_amp;
507 /* buffer overflow/underflow protection */
509 double expected_size = rcc->buffer_index;
513 double d = 2 * (buffer_size - expected_size) / buffer_size;
518 q *= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity);
520 q_limit = bits2qp(rce,
521 FFMAX((min_rate - buffer_size + rcc->buffer_index) *
522 s->avctx->rc_min_vbv_overflow_use, 1));
525 if (s->avctx->debug & FF_DEBUG_RC)
526 av_log(s->avctx, AV_LOG_DEBUG,
527 "limiting QP %f -> %f\n", q, q_limit);
533 double d = 2 * expected_size / buffer_size;
538 q /= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity);
540 q_limit = bits2qp(rce,
541 FFMAX(rcc->buffer_index *
542 s->avctx->rc_max_available_vbv_use,
545 if (s->avctx->debug & FF_DEBUG_RC)
546 av_log(s->avctx, AV_LOG_DEBUG,
547 "limiting QP %f -> %f\n", q, q_limit);
552 av_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
553 q, max_rate, min_rate, buffer_size, rcc->buffer_index,
554 s->avctx->rc_buffer_aggressivity);
555 if (s->avctx->rc_qsquish == 0.0 || qmin == qmax) {
561 double min2 = log(qmin);
562 double max2 = log(qmax);
565 q = (q - min2) / (max2 - min2) - 0.5;
567 q = 1.0 / (1.0 + exp(q));
568 q = q * (max2 - min2) + min2;
576 // ----------------------------------
579 static double predict_size(Predictor *p, double q, double var)
581 return p->coeff * var / (q * p->count);
584 static void update_predictor(Predictor *p, double q, double var, double size)
586 double new_coeff = size * q / (var + 1);
590 p->count *= p->decay;
591 p->coeff *= p->decay;
593 p->coeff += new_coeff;
596 static void adaptive_quantization(MpegEncContext *s, double q)
599 const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0);
600 const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0);
601 const float temp_cplx_masking = s->avctx->temporal_cplx_masking;
602 const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
603 const float p_masking = s->avctx->p_masking;
604 const float border_masking = s->avctx->border_masking;
605 float bits_sum = 0.0;
606 float cplx_sum = 0.0;
607 float *cplx_tab = s->cplx_tab;
608 float *bits_tab = s->bits_tab;
609 const int qmin = s->avctx->mb_lmin;
610 const int qmax = s->avctx->mb_lmax;
611 Picture *const pic = &s->current_picture;
612 const int mb_width = s->mb_width;
613 const int mb_height = s->mb_height;
615 for (i = 0; i < s->mb_num; i++) {
616 const int mb_xy = s->mb_index2xy[i];
617 float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
618 float spat_cplx = sqrt(pic->mb_var[mb_xy]);
619 const int lumi = pic->mb_mean[mb_xy];
620 float bits, cplx, factor;
621 int mb_x = mb_xy % s->mb_stride;
622 int mb_y = mb_xy / s->mb_stride;
624 float mb_factor = 0.0;
626 spat_cplx = 4; // FIXME finetune
628 temp_cplx = 4; // FIXME finetune
630 if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
632 factor = 1.0 + p_masking;
635 factor = pow(temp_cplx, -temp_cplx_masking);
637 factor *= pow(spat_cplx, -spatial_cplx_masking);
640 factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
642 factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);
644 if (mb_x < mb_width / 5) {
645 mb_distance = mb_width / 5 - mb_x;
646 mb_factor = (float)mb_distance / (float)(mb_width / 5);
647 } else if (mb_x > 4 * mb_width / 5) {
648 mb_distance = mb_x - 4 * mb_width / 5;
649 mb_factor = (float)mb_distance / (float)(mb_width / 5);
651 if (mb_y < mb_height / 5) {
652 mb_distance = mb_height / 5 - mb_y;
653 mb_factor = FFMAX(mb_factor,
654 (float)mb_distance / (float)(mb_height / 5));
655 } else if (mb_y > 4 * mb_height / 5) {
656 mb_distance = mb_y - 4 * mb_height / 5;
657 mb_factor = FFMAX(mb_factor,
658 (float)mb_distance / (float)(mb_height / 5));
661 factor *= 1.0 - border_masking * mb_factor;
663 if (factor < 0.00001)
666 bits = cplx * factor;
673 /* handle qmin/qmax clipping */
674 if (s->flags & CODEC_FLAG_NORMALIZE_AQP) {
675 float factor = bits_sum / cplx_sum;
676 for (i = 0; i < s->mb_num; i++) {
677 float newq = q * cplx_tab[i] / bits_tab[i];
681 bits_sum -= bits_tab[i];
682 cplx_sum -= cplx_tab[i] * q / qmax;
683 } else if (newq < qmin) {
684 bits_sum -= bits_tab[i];
685 cplx_sum -= cplx_tab[i] * q / qmin;
688 if (bits_sum < 0.001)
690 if (cplx_sum < 0.001)
694 for (i = 0; i < s->mb_num; i++) {
695 const int mb_xy = s->mb_index2xy[i];
696 float newq = q * cplx_tab[i] / bits_tab[i];
699 if (s->flags & CODEC_FLAG_NORMALIZE_AQP) {
700 newq *= bits_sum / cplx_sum;
703 intq = (int)(newq + 0.5);
707 else if (intq < qmin)
709 s->lambda_table[mb_xy] = intq;
713 void ff_get_2pass_fcode(MpegEncContext *s)
715 RateControlContext *rcc = &s->rc_context;
716 RateControlEntry *rce = &rcc->entry[s->picture_number];
718 s->f_code = rce->f_code;
719 s->b_code = rce->b_code;
722 // FIXME rd or at least approx for dquant
724 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
728 float br_compensation;
732 int picture_number = s->picture_number;
734 RateControlContext *rcc = &s->rc_context;
735 AVCodecContext *a = s->avctx;
736 RateControlEntry local_rce, *rce;
740 const int pict_type = s->pict_type;
741 Picture * const pic = &s->current_picture;
745 if ((s->flags & CODEC_FLAG_PASS2) &&
746 s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
747 return ff_xvid_rate_estimate_qscale(s, dry_run);
750 get_qminmax(&qmin, &qmax, s, pict_type);
752 fps = 1 / av_q2d(s->avctx->time_base);
753 /* update predictors */
754 if (picture_number > 2 && !dry_run) {
755 const int last_var = s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
756 : rcc->last_mc_mb_var_sum;
757 update_predictor(&rcc->pred[s->last_pict_type],
759 sqrt(last_var), s->frame_bits);
762 if (s->flags & CODEC_FLAG_PASS2) {
763 assert(picture_number >= 0);
764 assert(picture_number < rcc->num_entries);
765 rce = &rcc->entry[picture_number];
766 wanted_bits = rce->expected_bits;
771 /* FIXME add a dts field to AVFrame and ensure it is set and use it
772 * here instead of reordering but the reordering is simpler for now
773 * until H.264 B-pyramid must be handled. */
774 if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
775 dts_pic = s->current_picture_ptr;
777 dts_pic = s->last_picture_ptr;
779 if (!dts_pic || dts_pic->f.pts == AV_NOPTS_VALUE)
780 wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
782 wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f.pts / fps);
785 diff = s->total_bits - wanted_bits;
786 br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
787 if (br_compensation <= 0.0)
788 br_compensation = 0.001;
790 var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
792 short_term_q = 0; /* avoid warning */
793 if (s->flags & CODEC_FLAG_PASS2) {
794 if (pict_type != AV_PICTURE_TYPE_I)
795 assert(pict_type == rce->new_pict_type);
797 q = rce->new_qscale / br_compensation;
798 av_dlog(s, "%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale,
799 br_compensation, s->frame_bits, var, pict_type);
802 rce->new_pict_type = pict_type;
803 rce->mc_mb_var_sum = pic->mc_mb_var_sum;
804 rce->mb_var_sum = pic->mb_var_sum;
805 rce->qscale = FF_QP2LAMBDA * 2;
806 rce->f_code = s->f_code;
807 rce->b_code = s->b_code;
810 bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
811 if (pict_type == AV_PICTURE_TYPE_I) {
812 rce->i_count = s->mb_num;
813 rce->i_tex_bits = bits;
817 rce->i_count = 0; // FIXME we do know this approx
819 rce->p_tex_bits = bits * 0.9;
820 rce->mv_bits = bits * 0.1;
822 rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale;
823 rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale;
824 rcc->mv_bits_sum[pict_type] += rce->mv_bits;
825 rcc->frame_count[pict_type]++;
827 bits = rce->i_tex_bits + rce->p_tex_bits;
828 rate_factor = rcc->pass1_wanted_bits /
829 rcc->pass1_rc_eq_output_sum * br_compensation;
831 q = get_qscale(s, rce, rate_factor, picture_number);
836 q = get_diff_limited_q(s, rce, q);
839 // FIXME type dependent blur like in 2-pass
840 if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
841 rcc->short_term_qsum *= a->qblur;
842 rcc->short_term_qcount *= a->qblur;
844 rcc->short_term_qsum += q;
845 rcc->short_term_qcount++;
846 q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
850 q = modify_qscale(s, rce, q, picture_number);
852 rcc->pass1_wanted_bits += s->bit_rate / fps;
857 if (s->avctx->debug & FF_DEBUG_RC) {
858 av_log(s->avctx, AV_LOG_DEBUG,
859 "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
860 "size:%d var:%d/%d br:%d fps:%d\n",
861 av_get_picture_type_char(pict_type),
862 qmin, q, qmax, picture_number,
863 (int)wanted_bits / 1000, (int)s->total_bits / 1000,
864 br_compensation, short_term_q, s->frame_bits,
865 pic->mb_var_sum, pic->mc_mb_var_sum,
866 s->bit_rate / 1000, (int)fps);
874 if (s->adaptive_quant)
875 adaptive_quantization(s, q);
880 rcc->last_qscale = q;
881 rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
882 rcc->last_mb_var_sum = pic->mb_var_sum;
887 // ----------------------------------------------
890 static int init_pass2(MpegEncContext *s)
892 RateControlContext *rcc = &s->rc_context;
893 AVCodecContext *a = s->avctx;
895 double fps = 1 / av_q2d(s->avctx->time_base);
896 double complexity[5] = { 0 }; // approximate bits at quant=1
897 uint64_t const_bits[5] = { 0 }; // quantizer independent bits
898 uint64_t all_const_bits;
899 uint64_t all_available_bits = (uint64_t)(s->bit_rate *
900 (double)rcc->num_entries / fps);
901 double rate_factor = 0;
903 const int filter_size = (int)(a->qblur * 4) | 1;
904 double expected_bits;
905 double *qscale, *blurred_qscale, qscale_sum;
907 /* find complexity & const_bits & decide the pict_types */
908 for (i = 0; i < rcc->num_entries; i++) {
909 RateControlEntry *rce = &rcc->entry[i];
911 rce->new_pict_type = rce->pict_type;
912 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
913 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
914 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
915 rcc->frame_count[rce->pict_type]++;
917 complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
919 const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
922 all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
923 const_bits[AV_PICTURE_TYPE_P] +
924 const_bits[AV_PICTURE_TYPE_B];
926 if (all_available_bits < all_const_bits) {
927 av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
931 qscale = av_malloc(sizeof(double) * rcc->num_entries);
932 blurred_qscale = av_malloc(sizeof(double) * rcc->num_entries);
935 for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
939 rcc->buffer_index = s->avctx->rc_buffer_size / 2;
942 for (i = 0; i < rcc->num_entries; i++) {
943 RateControlEntry *rce = &rcc->entry[i];
945 qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
946 rcc->last_qscale_for[rce->pict_type] = qscale[i];
948 assert(filter_size % 2 == 1);
950 /* fixed I/B QP relative to P mode */
951 for (i = rcc->num_entries - 1; i >= 0; i--) {
952 RateControlEntry *rce = &rcc->entry[i];
954 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
958 for (i = 0; i < rcc->num_entries; i++) {
959 RateControlEntry *rce = &rcc->entry[i];
960 const int pict_type = rce->new_pict_type;
962 double q = 0.0, sum = 0.0;
964 for (j = 0; j < filter_size; j++) {
965 int index = i + j - filter_size / 2;
966 double d = index - i;
967 double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));
969 if (index < 0 || index >= rcc->num_entries)
971 if (pict_type != rcc->entry[index].new_pict_type)
973 q += qscale[index] * coeff;
976 blurred_qscale[i] = q / sum;
979 /* find expected bits */
980 for (i = 0; i < rcc->num_entries; i++) {
981 RateControlEntry *rce = &rcc->entry[i];
984 rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);
986 bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
987 bits += 8 * ff_vbv_update(s, bits);
989 rce->expected_bits = expected_bits;
990 expected_bits += bits;
994 "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
995 expected_bits, (int)all_available_bits, rate_factor);
996 if (expected_bits > all_available_bits) {
1002 av_free(blurred_qscale);
1004 /* check bitrate calculations and print info */
1006 for (i = 0; i < rcc->num_entries; i++) {
1007 av_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
1009 rcc->entry[i].new_qscale,
1010 rcc->entry[i].new_qscale / FF_QP2LAMBDA);
1011 qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
1012 s->avctx->qmin, s->avctx->qmax);
1014 assert(toobig <= 40);
1015 av_log(s->avctx, AV_LOG_DEBUG,
1016 "[lavc rc] requested bitrate: %d bps expected bitrate: %d bps\n",
1018 (int)(expected_bits / ((double)all_available_bits / s->bit_rate)));
1019 av_log(s->avctx, AV_LOG_DEBUG,
1020 "[lavc rc] estimated target average qp: %.3f\n",
1021 (float)qscale_sum / rcc->num_entries);
1023 av_log(s->avctx, AV_LOG_INFO,
1024 "[lavc rc] Using all of requested bitrate is not "
1025 "necessary for this video with these parameters.\n");
1026 } else if (toobig == 40) {
1027 av_log(s->avctx, AV_LOG_ERROR,
1028 "[lavc rc] Error: bitrate too low for this video "
1029 "with these parameters.\n");
1031 } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
1032 av_log(s->avctx, AV_LOG_ERROR,
1033 "[lavc rc] Error: 2pass curve failed to converge\n");