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"
29 #include "libavutil/internal.h"
33 #include "ratecontrol.h"
34 #include "mpegutils.h"
35 #include "mpegvideo.h"
36 #include "libavutil/eval.h"
38 #undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
42 #define M_E 2.718281828
45 static int init_pass2(MpegEncContext *s);
46 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
47 double rate_factor, int frame_num);
49 static inline double qp2bits(RateControlEntry *rce, double qp)
52 av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
54 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp;
57 static inline double bits2qp(RateControlEntry *rce, double bits)
60 av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
62 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits;
65 av_cold int ff_rate_control_init(MpegEncContext *s)
67 RateControlContext *rcc = &s->rc_context;
69 static const char * const const_names[] = {
92 static double (* const func1[])(void *, double) = {
97 static const char * const func1_names[] = {
104 res = av_expr_parse(&rcc->rc_eq_eval,
105 s->rc_eq ? s->rc_eq : "tex^qComp",
106 const_names, func1_names, func1,
107 NULL, NULL, 0, s->avctx);
109 av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->rc_eq);
113 for (i = 0; i < 5; i++) {
114 rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0;
115 rcc->pred[i].count = 1.0;
116 rcc->pred[i].decay = 0.4;
118 rcc->i_cplx_sum [i] =
119 rcc->p_cplx_sum [i] =
120 rcc->mv_bits_sum[i] =
121 rcc->qscale_sum [i] =
122 rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such
124 rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5;
126 rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy;
128 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
132 /* find number of pics */
133 p = s->avctx->stats_in;
135 p = strchr(p + 1, ';');
136 i += s->max_b_frames;
137 if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry))
139 rcc->entry = av_mallocz(i * sizeof(RateControlEntry));
140 rcc->num_entries = i;
142 return AVERROR(ENOMEM);
144 /* init all to skipped P-frames
145 * (with B-frames we might have a not encoded frame at the end FIXME) */
146 for (i = 0; i < rcc->num_entries; i++) {
147 RateControlEntry *rce = &rcc->entry[i];
149 rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P;
150 rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2;
151 rce->misc_bits = s->mb_num + 10;
152 rce->mb_var_sum = s->mb_num * 100;
156 p = s->avctx->stats_in;
157 for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
158 RateControlEntry *rce;
163 next = strchr(p, ';');
165 (*next) = 0; // sscanf is unbelievably slow on looong strings // FIXME copy / do not write
168 e = sscanf(p, " in:%d ", &picture_number);
170 assert(picture_number >= 0);
171 assert(picture_number < rcc->num_entries);
172 rce = &rcc->entry[picture_number];
174 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",
175 &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
176 &rce->mv_bits, &rce->misc_bits,
177 &rce->f_code, &rce->b_code,
178 &rce->mc_mb_var_sum, &rce->mb_var_sum,
179 &rce->i_count, &rce->skip_count, &rce->header_bits);
181 av_log(s->avctx, AV_LOG_ERROR,
182 "statistics are damaged at line %d, parser out=%d\n",
190 if (init_pass2(s) < 0) {
191 ff_rate_control_uninit(s);
196 if (!(s->avctx->flags & AV_CODEC_FLAG_PASS2)) {
197 rcc->short_term_qsum = 0.001;
198 rcc->short_term_qcount = 0.001;
200 rcc->pass1_rc_eq_output_sum = 0.001;
201 rcc->pass1_wanted_bits = 0.001;
203 if (s->avctx->qblur > 1.0) {
204 av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
207 /* init stuff with the user specified complexity */
208 if (s->rc_initial_cplx) {
209 for (i = 0; i < 60 * 30; i++) {
210 double bits = s->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
211 RateControlEntry rce;
213 if (i % ((s->gop_size + 3) / 4) == 0)
214 rce.pict_type = AV_PICTURE_TYPE_I;
215 else if (i % (s->max_b_frames + 1))
216 rce.pict_type = AV_PICTURE_TYPE_B;
218 rce.pict_type = AV_PICTURE_TYPE_P;
220 rce.new_pict_type = rce.pict_type;
221 rce.mc_mb_var_sum = bits * s->mb_num / 100000;
222 rce.mb_var_sum = s->mb_num;
224 rce.qscale = FF_QP2LAMBDA * 2;
229 if (s->pict_type == AV_PICTURE_TYPE_I) {
230 rce.i_count = s->mb_num;
231 rce.i_tex_bits = bits;
235 rce.i_count = 0; // FIXME we do know this approx
237 rce.p_tex_bits = bits * 0.9;
238 rce.mv_bits = bits * 0.1;
240 rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale;
241 rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale;
242 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
243 rcc->frame_count[rce.pict_type]++;
245 get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
247 // FIXME misbehaves a little for variable fps
248 rcc->pass1_wanted_bits += s->bit_rate / (1 / av_q2d(s->avctx->time_base));
256 av_cold void ff_rate_control_uninit(MpegEncContext *s)
258 RateControlContext *rcc = &s->rc_context;
261 av_expr_free(rcc->rc_eq_eval);
262 av_freep(&rcc->entry);
265 int ff_vbv_update(MpegEncContext *s, int frame_size)
267 RateControlContext *rcc = &s->rc_context;
268 const double fps = 1 / av_q2d(s->avctx->time_base);
269 const int buffer_size = s->avctx->rc_buffer_size;
270 const double min_rate = s->avctx->rc_min_rate / fps;
271 const double max_rate = s->avctx->rc_max_rate / fps;
273 ff_dlog(s, "%d %f %d %f %f\n",
274 buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
279 rcc->buffer_index -= frame_size;
280 if (rcc->buffer_index < 0) {
281 av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
282 rcc->buffer_index = 0;
285 left = buffer_size - rcc->buffer_index - 1;
286 rcc->buffer_index += av_clip(left, min_rate, max_rate);
288 if (rcc->buffer_index > buffer_size) {
289 int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
291 if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
293 rcc->buffer_index -= 8 * stuffing;
295 if (s->avctx->debug & FF_DEBUG_RC)
296 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
305 * Modify the bitrate curve from pass1 for one frame.
307 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
308 double rate_factor, int frame_num)
310 RateControlContext *rcc = &s->rc_context;
311 AVCodecContext *a = s->avctx;
312 const int pict_type = rce->new_pict_type;
313 const double mb_num = s->mb_num;
317 double const_values[] = {
320 rce->i_tex_bits * rce->qscale,
321 rce->p_tex_bits * rce->qscale,
322 (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
323 rce->mv_bits / mb_num,
324 rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
325 rce->i_count / mb_num,
326 rce->mc_mb_var_sum / mb_num,
327 rce->mb_var_sum / mb_num,
328 rce->pict_type == AV_PICTURE_TYPE_I,
329 rce->pict_type == AV_PICTURE_TYPE_P,
330 rce->pict_type == AV_PICTURE_TYPE_B,
331 rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
333 rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
334 rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
335 rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
336 rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
337 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
341 bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
343 av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->rc_eq);
347 rcc->pass1_rc_eq_output_sum += bits;
351 bits += 1.0; // avoid 1/0 issues
354 for (i = 0; i < s->avctx->rc_override_count; i++) {
355 RcOverride *rco = s->avctx->rc_override;
356 if (rco[i].start_frame > frame_num)
358 if (rco[i].end_frame < frame_num)
362 bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it?
364 bits *= rco[i].quality_factor;
367 q = bits2qp(rce, bits);
370 if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
371 q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
372 else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
373 q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
380 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
382 RateControlContext *rcc = &s->rc_context;
383 AVCodecContext *a = s->avctx;
384 const int pict_type = rce->new_pict_type;
385 const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
386 const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];
388 if (pict_type == AV_PICTURE_TYPE_I &&
389 (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P))
390 q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
391 else if (pict_type == AV_PICTURE_TYPE_B &&
392 a->b_quant_factor > 0.0)
393 q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
397 /* last qscale / qdiff stuff */
398 if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
399 double last_q = rcc->last_qscale_for[pict_type];
400 const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;
402 if (q > last_q + maxdiff)
403 q = last_q + maxdiff;
404 else if (q < last_q - maxdiff)
405 q = last_q - maxdiff;
408 rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring
410 if (pict_type != AV_PICTURE_TYPE_B)
411 rcc->last_non_b_pict_type = pict_type;
417 * Get the qmin & qmax for pict_type.
419 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
424 assert(qmin <= qmax);
427 case AV_PICTURE_TYPE_B:
428 qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
429 qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
431 case AV_PICTURE_TYPE_I:
432 qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
433 qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
437 qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
438 qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);
447 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce,
448 double q, int frame_num)
450 RateControlContext *rcc = &s->rc_context;
451 const double buffer_size = s->avctx->rc_buffer_size;
452 const double fps = 1 / av_q2d(s->avctx->time_base);
453 const double min_rate = s->avctx->rc_min_rate / fps;
454 const double max_rate = s->avctx->rc_max_rate / fps;
455 const int pict_type = rce->new_pict_type;
458 get_qminmax(&qmin, &qmax, s, pict_type);
461 if (s->rc_qmod_freq &&
462 frame_num % s->rc_qmod_freq == 0 &&
463 pict_type == AV_PICTURE_TYPE_P)
466 /* buffer overflow/underflow protection */
468 double expected_size = rcc->buffer_index;
472 double d = 2 * (buffer_size - expected_size) / buffer_size;
477 q *= pow(d, 1.0 / s->rc_buffer_aggressivity);
479 q_limit = bits2qp(rce,
480 FFMAX((min_rate - buffer_size + rcc->buffer_index) *
481 s->avctx->rc_min_vbv_overflow_use, 1));
484 if (s->avctx->debug & FF_DEBUG_RC)
485 av_log(s->avctx, AV_LOG_DEBUG,
486 "limiting QP %f -> %f\n", q, q_limit);
492 double d = 2 * expected_size / buffer_size;
497 q /= pow(d, 1.0 / s->rc_buffer_aggressivity);
499 q_limit = bits2qp(rce,
500 FFMAX(rcc->buffer_index *
501 s->avctx->rc_max_available_vbv_use,
504 if (s->avctx->debug & FF_DEBUG_RC)
505 av_log(s->avctx, AV_LOG_DEBUG,
506 "limiting QP %f -> %f\n", q, q_limit);
511 ff_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
512 q, max_rate, min_rate, buffer_size, rcc->buffer_index,
513 s->rc_buffer_aggressivity);
514 if (s->rc_qsquish == 0.0 || qmin == qmax) {
520 double min2 = log(qmin);
521 double max2 = log(qmax);
524 q = (q - min2) / (max2 - min2) - 0.5;
526 q = 1.0 / (1.0 + exp(q));
527 q = q * (max2 - min2) + min2;
535 // ----------------------------------
538 static double predict_size(Predictor *p, double q, double var)
540 return p->coeff * var / (q * p->count);
543 static void update_predictor(Predictor *p, double q, double var, double size)
545 double new_coeff = size * q / (var + 1);
549 p->count *= p->decay;
550 p->coeff *= p->decay;
552 p->coeff += new_coeff;
555 static void adaptive_quantization(MpegEncContext *s, double q)
558 const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0);
559 const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0);
560 const float temp_cplx_masking = s->avctx->temporal_cplx_masking;
561 const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
562 const float p_masking = s->avctx->p_masking;
563 const float border_masking = s->border_masking;
564 float bits_sum = 0.0;
565 float cplx_sum = 0.0;
566 float *cplx_tab = s->cplx_tab;
567 float *bits_tab = s->bits_tab;
568 const int qmin = s->avctx->mb_lmin;
569 const int qmax = s->avctx->mb_lmax;
570 Picture *const pic = &s->current_picture;
571 const int mb_width = s->mb_width;
572 const int mb_height = s->mb_height;
574 for (i = 0; i < s->mb_num; i++) {
575 const int mb_xy = s->mb_index2xy[i];
576 float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
577 float spat_cplx = sqrt(pic->mb_var[mb_xy]);
578 const int lumi = pic->mb_mean[mb_xy];
579 float bits, cplx, factor;
580 int mb_x = mb_xy % s->mb_stride;
581 int mb_y = mb_xy / s->mb_stride;
583 float mb_factor = 0.0;
585 spat_cplx = 4; // FIXME fine-tune
587 temp_cplx = 4; // FIXME fine-tune
589 if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
591 factor = 1.0 + p_masking;
594 factor = pow(temp_cplx, -temp_cplx_masking);
596 factor *= pow(spat_cplx, -spatial_cplx_masking);
599 factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
601 factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);
603 if (mb_x < mb_width / 5) {
604 mb_distance = mb_width / 5 - mb_x;
605 mb_factor = (float)mb_distance / (float)(mb_width / 5);
606 } else if (mb_x > 4 * mb_width / 5) {
607 mb_distance = mb_x - 4 * mb_width / 5;
608 mb_factor = (float)mb_distance / (float)(mb_width / 5);
610 if (mb_y < mb_height / 5) {
611 mb_distance = mb_height / 5 - mb_y;
612 mb_factor = FFMAX(mb_factor,
613 (float)mb_distance / (float)(mb_height / 5));
614 } else if (mb_y > 4 * mb_height / 5) {
615 mb_distance = mb_y - 4 * mb_height / 5;
616 mb_factor = FFMAX(mb_factor,
617 (float)mb_distance / (float)(mb_height / 5));
620 factor *= 1.0 - border_masking * mb_factor;
622 if (factor < 0.00001)
625 bits = cplx * factor;
632 /* handle qmin/qmax clipping */
633 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
634 float factor = bits_sum / cplx_sum;
635 for (i = 0; i < s->mb_num; i++) {
636 float newq = q * cplx_tab[i] / bits_tab[i];
640 bits_sum -= bits_tab[i];
641 cplx_sum -= cplx_tab[i] * q / qmax;
642 } else if (newq < qmin) {
643 bits_sum -= bits_tab[i];
644 cplx_sum -= cplx_tab[i] * q / qmin;
647 if (bits_sum < 0.001)
649 if (cplx_sum < 0.001)
653 for (i = 0; i < s->mb_num; i++) {
654 const int mb_xy = s->mb_index2xy[i];
655 float newq = q * cplx_tab[i] / bits_tab[i];
658 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
659 newq *= bits_sum / cplx_sum;
662 intq = (int)(newq + 0.5);
666 else if (intq < qmin)
668 s->lambda_table[mb_xy] = intq;
672 void ff_get_2pass_fcode(MpegEncContext *s)
674 RateControlContext *rcc = &s->rc_context;
675 RateControlEntry *rce = &rcc->entry[s->picture_number];
677 s->f_code = rce->f_code;
678 s->b_code = rce->b_code;
681 // FIXME rd or at least approx for dquant
683 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
687 float br_compensation;
691 int picture_number = s->picture_number;
693 RateControlContext *rcc = &s->rc_context;
694 AVCodecContext *a = s->avctx;
695 RateControlEntry local_rce, *rce;
699 const int pict_type = s->pict_type;
700 Picture * const pic = &s->current_picture;
703 get_qminmax(&qmin, &qmax, s, pict_type);
705 fps = 1 / av_q2d(s->avctx->time_base);
706 /* update predictors */
707 if (picture_number > 2 && !dry_run) {
708 const int last_var = s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
709 : rcc->last_mc_mb_var_sum;
710 update_predictor(&rcc->pred[s->last_pict_type],
712 sqrt(last_var), s->frame_bits);
715 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
716 assert(picture_number >= 0);
717 assert(picture_number < rcc->num_entries);
718 rce = &rcc->entry[picture_number];
719 wanted_bits = rce->expected_bits;
724 /* FIXME add a dts field to AVFrame and ensure it is set and use it
725 * here instead of reordering but the reordering is simpler for now
726 * until H.264 B-pyramid must be handled. */
727 if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
728 dts_pic = s->current_picture_ptr;
730 dts_pic = s->last_picture_ptr;
732 if (!dts_pic || dts_pic->f->pts == AV_NOPTS_VALUE)
733 wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
735 wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f->pts / fps);
738 diff = s->total_bits - wanted_bits;
739 br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
740 if (br_compensation <= 0.0)
741 br_compensation = 0.001;
743 var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
745 short_term_q = 0; /* avoid warning */
746 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
747 if (pict_type != AV_PICTURE_TYPE_I)
748 assert(pict_type == rce->new_pict_type);
750 q = rce->new_qscale / br_compensation;
751 ff_dlog(s, "%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale,
752 br_compensation, s->frame_bits, var, pict_type);
755 rce->new_pict_type = pict_type;
756 rce->mc_mb_var_sum = pic->mc_mb_var_sum;
757 rce->mb_var_sum = pic->mb_var_sum;
758 rce->qscale = FF_QP2LAMBDA * 2;
759 rce->f_code = s->f_code;
760 rce->b_code = s->b_code;
763 bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
764 if (pict_type == AV_PICTURE_TYPE_I) {
765 rce->i_count = s->mb_num;
766 rce->i_tex_bits = bits;
770 rce->i_count = 0; // FIXME we do know this approx
772 rce->p_tex_bits = bits * 0.9;
773 rce->mv_bits = bits * 0.1;
775 rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale;
776 rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale;
777 rcc->mv_bits_sum[pict_type] += rce->mv_bits;
778 rcc->frame_count[pict_type]++;
780 bits = rce->i_tex_bits + rce->p_tex_bits;
781 rate_factor = rcc->pass1_wanted_bits /
782 rcc->pass1_rc_eq_output_sum * br_compensation;
784 q = get_qscale(s, rce, rate_factor, picture_number);
789 q = get_diff_limited_q(s, rce, q);
792 // FIXME type dependent blur like in 2-pass
793 if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
794 rcc->short_term_qsum *= a->qblur;
795 rcc->short_term_qcount *= a->qblur;
797 rcc->short_term_qsum += q;
798 rcc->short_term_qcount++;
799 q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
803 q = modify_qscale(s, rce, q, picture_number);
805 rcc->pass1_wanted_bits += s->bit_rate / fps;
810 if (s->avctx->debug & FF_DEBUG_RC) {
811 av_log(s->avctx, AV_LOG_DEBUG,
812 "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
813 "size:%d var:%d/%d br:%d fps:%d\n",
814 av_get_picture_type_char(pict_type),
815 qmin, q, qmax, picture_number,
816 (int)wanted_bits / 1000, (int)s->total_bits / 1000,
817 br_compensation, short_term_q, s->frame_bits,
818 pic->mb_var_sum, pic->mc_mb_var_sum,
819 s->bit_rate / 1000, (int)fps);
827 if (s->adaptive_quant)
828 adaptive_quantization(s, q);
833 rcc->last_qscale = q;
834 rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
835 rcc->last_mb_var_sum = pic->mb_var_sum;
840 // ----------------------------------------------
843 static int init_pass2(MpegEncContext *s)
845 RateControlContext *rcc = &s->rc_context;
846 AVCodecContext *a = s->avctx;
848 double fps = 1 / av_q2d(s->avctx->time_base);
849 double complexity[5] = { 0 }; // approximate bits at quant=1
850 uint64_t const_bits[5] = { 0 }; // quantizer independent bits
851 uint64_t all_const_bits;
852 uint64_t all_available_bits = (uint64_t)(s->bit_rate *
853 (double)rcc->num_entries / fps);
854 double rate_factor = 0;
856 const int filter_size = (int)(a->qblur * 4) | 1;
857 double expected_bits;
858 double *qscale, *blurred_qscale, qscale_sum;
860 /* find complexity & const_bits & decide the pict_types */
861 for (i = 0; i < rcc->num_entries; i++) {
862 RateControlEntry *rce = &rcc->entry[i];
864 rce->new_pict_type = rce->pict_type;
865 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
866 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
867 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
868 rcc->frame_count[rce->pict_type]++;
870 complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
872 const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
875 all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
876 const_bits[AV_PICTURE_TYPE_P] +
877 const_bits[AV_PICTURE_TYPE_B];
879 if (all_available_bits < all_const_bits) {
880 av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
884 qscale = av_malloc(sizeof(double) * rcc->num_entries);
885 blurred_qscale = av_malloc(sizeof(double) * rcc->num_entries);
886 if (!qscale || !blurred_qscale) {
888 av_free(blurred_qscale);
889 return AVERROR(ENOMEM);
893 for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
897 rcc->buffer_index = s->avctx->rc_buffer_size / 2;
900 for (i = 0; i < rcc->num_entries; i++) {
901 RateControlEntry *rce = &rcc->entry[i];
903 qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
904 rcc->last_qscale_for[rce->pict_type] = qscale[i];
906 assert(filter_size % 2 == 1);
908 /* fixed I/B QP relative to P mode */
909 for (i = rcc->num_entries - 1; i >= 0; i--) {
910 RateControlEntry *rce = &rcc->entry[i];
912 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
916 for (i = 0; i < rcc->num_entries; i++) {
917 RateControlEntry *rce = &rcc->entry[i];
918 const int pict_type = rce->new_pict_type;
920 double q = 0.0, sum = 0.0;
922 for (j = 0; j < filter_size; j++) {
923 int index = i + j - filter_size / 2;
924 double d = index - i;
925 double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));
927 if (index < 0 || index >= rcc->num_entries)
929 if (pict_type != rcc->entry[index].new_pict_type)
931 q += qscale[index] * coeff;
934 blurred_qscale[i] = q / sum;
937 /* find expected bits */
938 for (i = 0; i < rcc->num_entries; i++) {
939 RateControlEntry *rce = &rcc->entry[i];
942 rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);
944 bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
945 bits += 8 * ff_vbv_update(s, bits);
947 rce->expected_bits = expected_bits;
948 expected_bits += bits;
952 "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
953 expected_bits, (int)all_available_bits, rate_factor);
954 if (expected_bits > all_available_bits) {
960 av_free(blurred_qscale);
962 /* check bitrate calculations and print info */
964 for (i = 0; i < rcc->num_entries; i++) {
965 ff_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
967 rcc->entry[i].new_qscale,
968 rcc->entry[i].new_qscale / FF_QP2LAMBDA);
969 qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
970 s->avctx->qmin, s->avctx->qmax);
972 assert(toobig <= 40);
973 av_log(s->avctx, AV_LOG_DEBUG,
974 "[lavc rc] requested bitrate: %d bps expected bitrate: %d bps\n",
976 (int)(expected_bits / ((double)all_available_bits / s->bit_rate)));
977 av_log(s->avctx, AV_LOG_DEBUG,
978 "[lavc rc] estimated target average qp: %.3f\n",
979 (float)qscale_sum / rcc->num_entries);
981 av_log(s->avctx, AV_LOG_INFO,
982 "[lavc rc] Using all of requested bitrate is not "
983 "necessary for this video with these parameters.\n");
984 } else if (toobig == 40) {
985 av_log(s->avctx, AV_LOG_ERROR,
986 "[lavc rc] Error: bitrate too low for this video "
987 "with these parameters.\n");
989 } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
990 av_log(s->avctx, AV_LOG_ERROR,
991 "[lavc rc] Error: 2pass curve failed to converge\n");