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 inline double qp2bits(RateControlEntry *rce, double qp)
48 av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
50 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp;
53 static inline double bits2qp(RateControlEntry *rce, double bits)
56 av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
58 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits;
61 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
63 RateControlContext *rcc = &s->rc_context;
64 AVCodecContext *a = s->avctx;
65 const int pict_type = rce->new_pict_type;
66 const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
67 const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];
69 if (pict_type == AV_PICTURE_TYPE_I &&
70 (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P))
71 q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
72 else if (pict_type == AV_PICTURE_TYPE_B &&
73 a->b_quant_factor > 0.0)
74 q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
78 /* last qscale / qdiff stuff */
79 if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
80 double last_q = rcc->last_qscale_for[pict_type];
81 const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;
83 if (q > last_q + maxdiff)
85 else if (q < last_q - maxdiff)
89 rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring
91 if (pict_type != AV_PICTURE_TYPE_B)
92 rcc->last_non_b_pict_type = pict_type;
98 * Get the qmin & qmax for pict_type.
100 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
105 assert(qmin <= qmax);
108 case AV_PICTURE_TYPE_B:
109 qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
110 qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
112 case AV_PICTURE_TYPE_I:
113 qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
114 qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
118 qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
119 qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);
128 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce,
129 double q, int frame_num)
131 RateControlContext *rcc = &s->rc_context;
132 const double buffer_size = s->avctx->rc_buffer_size;
133 const double fps = 1 / av_q2d(s->avctx->time_base);
134 const double min_rate = s->avctx->rc_min_rate / fps;
135 const double max_rate = s->avctx->rc_max_rate / fps;
136 const int pict_type = rce->new_pict_type;
139 get_qminmax(&qmin, &qmax, s, pict_type);
142 if (s->rc_qmod_freq &&
143 frame_num % s->rc_qmod_freq == 0 &&
144 pict_type == AV_PICTURE_TYPE_P)
147 /* buffer overflow/underflow protection */
149 double expected_size = rcc->buffer_index;
153 double d = 2 * (buffer_size - expected_size) / buffer_size;
158 q *= pow(d, 1.0 / s->rc_buffer_aggressivity);
160 q_limit = bits2qp(rce,
161 FFMAX((min_rate - buffer_size + rcc->buffer_index) *
162 s->avctx->rc_min_vbv_overflow_use, 1));
165 if (s->avctx->debug & FF_DEBUG_RC)
166 av_log(s->avctx, AV_LOG_DEBUG,
167 "limiting QP %f -> %f\n", q, q_limit);
173 double d = 2 * expected_size / buffer_size;
178 q /= pow(d, 1.0 / s->rc_buffer_aggressivity);
180 q_limit = bits2qp(rce,
181 FFMAX(rcc->buffer_index *
182 s->avctx->rc_max_available_vbv_use,
185 if (s->avctx->debug & FF_DEBUG_RC)
186 av_log(s->avctx, AV_LOG_DEBUG,
187 "limiting QP %f -> %f\n", q, q_limit);
192 ff_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
193 q, max_rate, min_rate, buffer_size, rcc->buffer_index,
194 s->rc_buffer_aggressivity);
195 if (s->rc_qsquish == 0.0 || qmin == qmax) {
201 double min2 = log(qmin);
202 double max2 = log(qmax);
205 q = (q - min2) / (max2 - min2) - 0.5;
207 q = 1.0 / (1.0 + exp(q));
208 q = q * (max2 - min2) + min2;
217 * Modify the bitrate curve from pass1 for one frame.
219 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
220 double rate_factor, int frame_num)
222 RateControlContext *rcc = &s->rc_context;
223 AVCodecContext *a = s->avctx;
224 const int pict_type = rce->new_pict_type;
225 const double mb_num = s->mb_num;
229 double const_values[] = {
232 rce->i_tex_bits * rce->qscale,
233 rce->p_tex_bits * rce->qscale,
234 (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
235 rce->mv_bits / mb_num,
236 rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
237 rce->i_count / mb_num,
238 rce->mc_mb_var_sum / mb_num,
239 rce->mb_var_sum / mb_num,
240 rce->pict_type == AV_PICTURE_TYPE_I,
241 rce->pict_type == AV_PICTURE_TYPE_P,
242 rce->pict_type == AV_PICTURE_TYPE_B,
243 rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
245 rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
246 rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
247 rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
248 rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
249 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
253 bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
255 av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->rc_eq);
259 rcc->pass1_rc_eq_output_sum += bits;
263 bits += 1.0; // avoid 1/0 issues
266 for (i = 0; i < s->avctx->rc_override_count; i++) {
267 RcOverride *rco = s->avctx->rc_override;
268 if (rco[i].start_frame > frame_num)
270 if (rco[i].end_frame < frame_num)
274 bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it?
276 bits *= rco[i].quality_factor;
279 q = bits2qp(rce, bits);
282 if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
283 q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
284 else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
285 q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
292 static int init_pass2(MpegEncContext *s)
294 RateControlContext *rcc = &s->rc_context;
295 AVCodecContext *a = s->avctx;
297 double fps = 1 / av_q2d(s->avctx->time_base);
298 double complexity[5] = { 0 }; // approximate bits at quant=1
299 uint64_t const_bits[5] = { 0 }; // quantizer independent bits
300 uint64_t all_const_bits;
301 uint64_t all_available_bits = (uint64_t)(s->bit_rate *
302 (double)rcc->num_entries / fps);
303 double rate_factor = 0;
305 const int filter_size = (int)(a->qblur * 4) | 1;
306 double expected_bits;
307 double *qscale, *blurred_qscale, qscale_sum;
309 /* find complexity & const_bits & decide the pict_types */
310 for (i = 0; i < rcc->num_entries; i++) {
311 RateControlEntry *rce = &rcc->entry[i];
313 rce->new_pict_type = rce->pict_type;
314 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
315 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
316 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
317 rcc->frame_count[rce->pict_type]++;
319 complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
321 const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
324 all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
325 const_bits[AV_PICTURE_TYPE_P] +
326 const_bits[AV_PICTURE_TYPE_B];
328 if (all_available_bits < all_const_bits) {
329 av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
333 qscale = av_malloc(sizeof(double) * rcc->num_entries);
334 blurred_qscale = av_malloc(sizeof(double) * rcc->num_entries);
335 if (!qscale || !blurred_qscale) {
337 av_free(blurred_qscale);
338 return AVERROR(ENOMEM);
342 for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
346 rcc->buffer_index = s->avctx->rc_buffer_size / 2;
349 for (i = 0; i < rcc->num_entries; i++) {
350 RateControlEntry *rce = &rcc->entry[i];
352 qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
353 rcc->last_qscale_for[rce->pict_type] = qscale[i];
355 assert(filter_size % 2 == 1);
357 /* fixed I/B QP relative to P mode */
358 for (i = rcc->num_entries - 1; i >= 0; i--) {
359 RateControlEntry *rce = &rcc->entry[i];
361 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
365 for (i = 0; i < rcc->num_entries; i++) {
366 RateControlEntry *rce = &rcc->entry[i];
367 const int pict_type = rce->new_pict_type;
369 double q = 0.0, sum = 0.0;
371 for (j = 0; j < filter_size; j++) {
372 int index = i + j - filter_size / 2;
373 double d = index - i;
374 double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));
376 if (index < 0 || index >= rcc->num_entries)
378 if (pict_type != rcc->entry[index].new_pict_type)
380 q += qscale[index] * coeff;
383 blurred_qscale[i] = q / sum;
386 /* find expected bits */
387 for (i = 0; i < rcc->num_entries; i++) {
388 RateControlEntry *rce = &rcc->entry[i];
391 rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);
393 bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
394 bits += 8 * ff_vbv_update(s, bits);
396 rce->expected_bits = expected_bits;
397 expected_bits += bits;
401 "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
402 expected_bits, (int)all_available_bits, rate_factor);
403 if (expected_bits > all_available_bits) {
409 av_free(blurred_qscale);
411 /* check bitrate calculations and print info */
413 for (i = 0; i < rcc->num_entries; i++) {
414 ff_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
416 rcc->entry[i].new_qscale,
417 rcc->entry[i].new_qscale / FF_QP2LAMBDA);
418 qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
419 s->avctx->qmin, s->avctx->qmax);
421 assert(toobig <= 40);
422 av_log(s->avctx, AV_LOG_DEBUG,
423 "[lavc rc] requested bitrate: %d bps expected bitrate: %d bps\n",
425 (int)(expected_bits / ((double)all_available_bits / s->bit_rate)));
426 av_log(s->avctx, AV_LOG_DEBUG,
427 "[lavc rc] estimated target average qp: %.3f\n",
428 (float)qscale_sum / rcc->num_entries);
430 av_log(s->avctx, AV_LOG_INFO,
431 "[lavc rc] Using all of requested bitrate is not "
432 "necessary for this video with these parameters.\n");
433 } else if (toobig == 40) {
434 av_log(s->avctx, AV_LOG_ERROR,
435 "[lavc rc] Error: bitrate too low for this video "
436 "with these parameters.\n");
438 } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
439 av_log(s->avctx, AV_LOG_ERROR,
440 "[lavc rc] Error: 2pass curve failed to converge\n");
447 av_cold int ff_rate_control_init(MpegEncContext *s)
449 RateControlContext *rcc = &s->rc_context;
451 static const char * const const_names[] = {
474 static double (* const func1[])(void *, double) = {
475 (double (*)(void *, double)) bits2qp,
476 (double (*)(void *, double)) qp2bits,
479 static const char * const func1_names[] = {
486 res = av_expr_parse(&rcc->rc_eq_eval,
487 s->rc_eq ? s->rc_eq : "tex^qComp",
488 const_names, func1_names, func1,
489 NULL, NULL, 0, s->avctx);
491 av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->rc_eq);
495 for (i = 0; i < 5; i++) {
496 rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0;
497 rcc->pred[i].count = 1.0;
498 rcc->pred[i].decay = 0.4;
500 rcc->i_cplx_sum [i] =
501 rcc->p_cplx_sum [i] =
502 rcc->mv_bits_sum[i] =
503 rcc->qscale_sum [i] =
504 rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such
506 rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5;
508 rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy;
510 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
514 /* find number of pics */
515 p = s->avctx->stats_in;
517 p = strchr(p + 1, ';');
518 i += s->max_b_frames;
519 if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry))
521 rcc->entry = av_mallocz(i * sizeof(RateControlEntry));
522 rcc->num_entries = i;
524 return AVERROR(ENOMEM);
526 /* init all to skipped P-frames
527 * (with B-frames we might have a not encoded frame at the end FIXME) */
528 for (i = 0; i < rcc->num_entries; i++) {
529 RateControlEntry *rce = &rcc->entry[i];
531 rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P;
532 rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2;
533 rce->misc_bits = s->mb_num + 10;
534 rce->mb_var_sum = s->mb_num * 100;
538 p = s->avctx->stats_in;
539 for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
540 RateControlEntry *rce;
545 next = strchr(p, ';');
547 (*next) = 0; // sscanf is unbelievably slow on looong strings // FIXME copy / do not write
550 e = sscanf(p, " in:%d ", &picture_number);
552 assert(picture_number >= 0);
553 assert(picture_number < rcc->num_entries);
554 rce = &rcc->entry[picture_number];
556 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",
557 &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
558 &rce->mv_bits, &rce->misc_bits,
559 &rce->f_code, &rce->b_code,
560 &rce->mc_mb_var_sum, &rce->mb_var_sum,
561 &rce->i_count, &rce->skip_count, &rce->header_bits);
563 av_log(s->avctx, AV_LOG_ERROR,
564 "statistics are damaged at line %d, parser out=%d\n",
572 if (init_pass2(s) < 0) {
573 ff_rate_control_uninit(s);
578 if (!(s->avctx->flags & AV_CODEC_FLAG_PASS2)) {
579 rcc->short_term_qsum = 0.001;
580 rcc->short_term_qcount = 0.001;
582 rcc->pass1_rc_eq_output_sum = 0.001;
583 rcc->pass1_wanted_bits = 0.001;
585 if (s->avctx->qblur > 1.0) {
586 av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
589 /* init stuff with the user specified complexity */
590 if (s->rc_initial_cplx) {
591 for (i = 0; i < 60 * 30; i++) {
592 double bits = s->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
593 RateControlEntry rce;
595 if (i % ((s->gop_size + 3) / 4) == 0)
596 rce.pict_type = AV_PICTURE_TYPE_I;
597 else if (i % (s->max_b_frames + 1))
598 rce.pict_type = AV_PICTURE_TYPE_B;
600 rce.pict_type = AV_PICTURE_TYPE_P;
602 rce.new_pict_type = rce.pict_type;
603 rce.mc_mb_var_sum = bits * s->mb_num / 100000;
604 rce.mb_var_sum = s->mb_num;
606 rce.qscale = FF_QP2LAMBDA * 2;
611 if (s->pict_type == AV_PICTURE_TYPE_I) {
612 rce.i_count = s->mb_num;
613 rce.i_tex_bits = bits;
617 rce.i_count = 0; // FIXME we do know this approx
619 rce.p_tex_bits = bits * 0.9;
620 rce.mv_bits = bits * 0.1;
622 rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale;
623 rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale;
624 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
625 rcc->frame_count[rce.pict_type]++;
627 get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
629 // FIXME misbehaves a little for variable fps
630 rcc->pass1_wanted_bits += s->bit_rate / (1 / av_q2d(s->avctx->time_base));
638 av_cold void ff_rate_control_uninit(MpegEncContext *s)
640 RateControlContext *rcc = &s->rc_context;
643 av_expr_free(rcc->rc_eq_eval);
644 av_freep(&rcc->entry);
647 int ff_vbv_update(MpegEncContext *s, int frame_size)
649 RateControlContext *rcc = &s->rc_context;
650 const double fps = 1 / av_q2d(s->avctx->time_base);
651 const int buffer_size = s->avctx->rc_buffer_size;
652 const double min_rate = s->avctx->rc_min_rate / fps;
653 const double max_rate = s->avctx->rc_max_rate / fps;
655 ff_dlog(s, "%d %f %d %f %f\n",
656 buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
661 rcc->buffer_index -= frame_size;
662 if (rcc->buffer_index < 0) {
663 av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
664 rcc->buffer_index = 0;
667 left = buffer_size - rcc->buffer_index - 1;
668 rcc->buffer_index += av_clip(left, min_rate, max_rate);
670 if (rcc->buffer_index > buffer_size) {
671 int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
673 if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
675 rcc->buffer_index -= 8 * stuffing;
677 if (s->avctx->debug & FF_DEBUG_RC)
678 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
686 static double predict_size(Predictor *p, double q, double var)
688 return p->coeff * var / (q * p->count);
691 static void update_predictor(Predictor *p, double q, double var, double size)
693 double new_coeff = size * q / (var + 1);
697 p->count *= p->decay;
698 p->coeff *= p->decay;
700 p->coeff += new_coeff;
703 static void adaptive_quantization(MpegEncContext *s, double q)
706 const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0);
707 const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0);
708 const float temp_cplx_masking = s->avctx->temporal_cplx_masking;
709 const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
710 const float p_masking = s->avctx->p_masking;
711 const float border_masking = s->border_masking;
712 float bits_sum = 0.0;
713 float cplx_sum = 0.0;
714 float *cplx_tab = s->cplx_tab;
715 float *bits_tab = s->bits_tab;
716 const int qmin = s->avctx->mb_lmin;
717 const int qmax = s->avctx->mb_lmax;
718 Picture *const pic = &s->current_picture;
719 const int mb_width = s->mb_width;
720 const int mb_height = s->mb_height;
722 for (i = 0; i < s->mb_num; i++) {
723 const int mb_xy = s->mb_index2xy[i];
724 float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
725 float spat_cplx = sqrt(pic->mb_var[mb_xy]);
726 const int lumi = pic->mb_mean[mb_xy];
727 float bits, cplx, factor;
728 int mb_x = mb_xy % s->mb_stride;
729 int mb_y = mb_xy / s->mb_stride;
731 float mb_factor = 0.0;
733 spat_cplx = 4; // FIXME fine-tune
735 temp_cplx = 4; // FIXME fine-tune
737 if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
739 factor = 1.0 + p_masking;
742 factor = pow(temp_cplx, -temp_cplx_masking);
744 factor *= pow(spat_cplx, -spatial_cplx_masking);
747 factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
749 factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);
751 if (mb_x < mb_width / 5) {
752 mb_distance = mb_width / 5 - mb_x;
753 mb_factor = (float)mb_distance / (float)(mb_width / 5);
754 } else if (mb_x > 4 * mb_width / 5) {
755 mb_distance = mb_x - 4 * mb_width / 5;
756 mb_factor = (float)mb_distance / (float)(mb_width / 5);
758 if (mb_y < mb_height / 5) {
759 mb_distance = mb_height / 5 - mb_y;
760 mb_factor = FFMAX(mb_factor,
761 (float)mb_distance / (float)(mb_height / 5));
762 } else if (mb_y > 4 * mb_height / 5) {
763 mb_distance = mb_y - 4 * mb_height / 5;
764 mb_factor = FFMAX(mb_factor,
765 (float)mb_distance / (float)(mb_height / 5));
768 factor *= 1.0 - border_masking * mb_factor;
770 if (factor < 0.00001)
773 bits = cplx * factor;
780 /* handle qmin/qmax clipping */
781 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
782 float factor = bits_sum / cplx_sum;
783 for (i = 0; i < s->mb_num; i++) {
784 float newq = q * cplx_tab[i] / bits_tab[i];
788 bits_sum -= bits_tab[i];
789 cplx_sum -= cplx_tab[i] * q / qmax;
790 } else if (newq < qmin) {
791 bits_sum -= bits_tab[i];
792 cplx_sum -= cplx_tab[i] * q / qmin;
795 if (bits_sum < 0.001)
797 if (cplx_sum < 0.001)
801 for (i = 0; i < s->mb_num; i++) {
802 const int mb_xy = s->mb_index2xy[i];
803 float newq = q * cplx_tab[i] / bits_tab[i];
806 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
807 newq *= bits_sum / cplx_sum;
810 intq = (int)(newq + 0.5);
814 else if (intq < qmin)
816 s->lambda_table[mb_xy] = intq;
820 void ff_get_2pass_fcode(MpegEncContext *s)
822 RateControlContext *rcc = &s->rc_context;
823 RateControlEntry *rce = &rcc->entry[s->picture_number];
825 s->f_code = rce->f_code;
826 s->b_code = rce->b_code;
829 // FIXME rd or at least approx for dquant
831 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
835 float br_compensation;
839 int picture_number = s->picture_number;
841 RateControlContext *rcc = &s->rc_context;
842 AVCodecContext *a = s->avctx;
843 RateControlEntry local_rce, *rce;
847 const int pict_type = s->pict_type;
848 Picture * const pic = &s->current_picture;
851 get_qminmax(&qmin, &qmax, s, pict_type);
853 fps = 1 / av_q2d(s->avctx->time_base);
854 /* update predictors */
855 if (picture_number > 2 && !dry_run) {
856 const int last_var = s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
857 : rcc->last_mc_mb_var_sum;
858 update_predictor(&rcc->pred[s->last_pict_type],
860 sqrt(last_var), s->frame_bits);
863 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
864 assert(picture_number >= 0);
865 assert(picture_number < rcc->num_entries);
866 rce = &rcc->entry[picture_number];
867 wanted_bits = rce->expected_bits;
872 /* FIXME add a dts field to AVFrame and ensure it is set and use it
873 * here instead of reordering but the reordering is simpler for now
874 * until H.264 B-pyramid must be handled. */
875 if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
876 dts_pic = s->current_picture_ptr;
878 dts_pic = s->last_picture_ptr;
880 if (!dts_pic || dts_pic->f->pts == AV_NOPTS_VALUE)
881 wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
883 wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f->pts / fps);
886 diff = s->total_bits - wanted_bits;
887 br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
888 if (br_compensation <= 0.0)
889 br_compensation = 0.001;
891 var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
893 short_term_q = 0; /* avoid warning */
894 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
895 if (pict_type != AV_PICTURE_TYPE_I)
896 assert(pict_type == rce->new_pict_type);
898 q = rce->new_qscale / br_compensation;
899 ff_dlog(s, "%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale,
900 br_compensation, s->frame_bits, var, pict_type);
903 rce->new_pict_type = pict_type;
904 rce->mc_mb_var_sum = pic->mc_mb_var_sum;
905 rce->mb_var_sum = pic->mb_var_sum;
906 rce->qscale = FF_QP2LAMBDA * 2;
907 rce->f_code = s->f_code;
908 rce->b_code = s->b_code;
911 bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
912 if (pict_type == AV_PICTURE_TYPE_I) {
913 rce->i_count = s->mb_num;
914 rce->i_tex_bits = bits;
918 rce->i_count = 0; // FIXME we do know this approx
920 rce->p_tex_bits = bits * 0.9;
921 rce->mv_bits = bits * 0.1;
923 rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale;
924 rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale;
925 rcc->mv_bits_sum[pict_type] += rce->mv_bits;
926 rcc->frame_count[pict_type]++;
928 bits = rce->i_tex_bits + rce->p_tex_bits;
929 rate_factor = rcc->pass1_wanted_bits /
930 rcc->pass1_rc_eq_output_sum * br_compensation;
932 q = get_qscale(s, rce, rate_factor, picture_number);
937 q = get_diff_limited_q(s, rce, q);
940 // FIXME type dependent blur like in 2-pass
941 if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
942 rcc->short_term_qsum *= a->qblur;
943 rcc->short_term_qcount *= a->qblur;
945 rcc->short_term_qsum += q;
946 rcc->short_term_qcount++;
947 q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
951 q = modify_qscale(s, rce, q, picture_number);
953 rcc->pass1_wanted_bits += s->bit_rate / fps;
958 if (s->avctx->debug & FF_DEBUG_RC) {
959 av_log(s->avctx, AV_LOG_DEBUG,
960 "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
961 "size:%d var:%d/%d br:%d fps:%d\n",
962 av_get_picture_type_char(pict_type),
963 qmin, q, qmax, picture_number,
964 (int)wanted_bits / 1000, (int)s->total_bits / 1000,
965 br_compensation, short_term_q, s->frame_bits,
966 pic->mb_var_sum, pic->mc_mb_var_sum,
967 s->bit_rate / 1000, (int)fps);
975 if (s->adaptive_quant)
976 adaptive_quantization(s, q);
981 rcc->last_qscale = q;
982 rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
983 rcc->last_mb_var_sum = pic->mb_var_sum;