2 * Rate control for video encoders
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
6 * This file is part of FFmpeg.
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 * Rate control for video encoders.
28 #include "libavutil/attributes.h"
29 #include "libavutil/internal.h"
33 #include "ratecontrol.h"
34 #include "mpegutils.h"
35 #include "mpegvideo.h"
36 #include "libavutil/eval.h"
38 void ff_write_pass1_stats(MpegEncContext *s)
40 snprintf(s->avctx->stats_out, 256,
41 "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d "
42 "fcode:%d bcode:%d mc-var:%"PRId64" var:%"PRId64" icount:%d skipcount:%d hbits:%d;\n",
43 s->current_picture_ptr->f->display_picture_number,
44 s->current_picture_ptr->f->coded_picture_number,
46 s->current_picture.f->quality,
53 s->current_picture.mc_mb_var_sum,
54 s->current_picture.mb_var_sum,
55 s->i_count, s->skip_count,
59 static double get_fps(AVCodecContext *avctx)
61 return 1.0 / av_q2d(avctx->time_base) / FFMAX(avctx->ticks_per_frame, 1);
64 static inline double qp2bits(RateControlEntry *rce, double qp)
67 av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
69 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp;
72 static inline double bits2qp(RateControlEntry *rce, double bits)
75 av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
77 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits;
80 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
82 RateControlContext *rcc = &s->rc_context;
83 AVCodecContext *a = s->avctx;
84 const int pict_type = rce->new_pict_type;
85 const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
86 const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];
88 if (pict_type == AV_PICTURE_TYPE_I &&
89 (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P))
90 q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
91 else if (pict_type == AV_PICTURE_TYPE_B &&
92 a->b_quant_factor > 0.0)
93 q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
97 /* last qscale / qdiff stuff */
98 if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
99 double last_q = rcc->last_qscale_for[pict_type];
100 const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;
102 if (q > last_q + maxdiff)
103 q = last_q + maxdiff;
104 else if (q < last_q - maxdiff)
105 q = last_q - maxdiff;
108 rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring
110 if (pict_type != AV_PICTURE_TYPE_B)
111 rcc->last_non_b_pict_type = pict_type;
117 * Get the qmin & qmax for pict_type.
119 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
124 av_assert0(qmin <= qmax);
127 case AV_PICTURE_TYPE_B:
128 qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
129 qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
131 case AV_PICTURE_TYPE_I:
132 qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
133 qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
137 qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
138 qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);
147 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce,
148 double q, int frame_num)
150 RateControlContext *rcc = &s->rc_context;
151 const double buffer_size = s->avctx->rc_buffer_size;
152 const double fps = get_fps(s->avctx);
153 const double min_rate = s->avctx->rc_min_rate / fps;
154 const double max_rate = s->avctx->rc_max_rate / fps;
155 const int pict_type = rce->new_pict_type;
158 get_qminmax(&qmin, &qmax, s, pict_type);
161 if (s->rc_qmod_freq &&
162 frame_num % s->rc_qmod_freq == 0 &&
163 pict_type == AV_PICTURE_TYPE_P)
166 /* buffer overflow/underflow protection */
168 double expected_size = rcc->buffer_index;
172 double d = 2 * (buffer_size - expected_size) / buffer_size;
177 q *= pow(d, 1.0 / s->rc_buffer_aggressivity);
179 q_limit = bits2qp(rce,
180 FFMAX((min_rate - buffer_size + rcc->buffer_index) *
181 s->avctx->rc_min_vbv_overflow_use, 1));
184 if (s->avctx->debug & FF_DEBUG_RC)
185 av_log(s->avctx, AV_LOG_DEBUG,
186 "limiting QP %f -> %f\n", q, q_limit);
192 double d = 2 * expected_size / buffer_size;
197 q /= pow(d, 1.0 / s->rc_buffer_aggressivity);
199 q_limit = bits2qp(rce,
200 FFMAX(rcc->buffer_index *
201 s->avctx->rc_max_available_vbv_use,
204 if (s->avctx->debug & FF_DEBUG_RC)
205 av_log(s->avctx, AV_LOG_DEBUG,
206 "limiting QP %f -> %f\n", q, q_limit);
211 ff_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
212 q, max_rate, min_rate, buffer_size, rcc->buffer_index,
213 s->rc_buffer_aggressivity);
214 if (s->rc_qsquish == 0.0 || qmin == qmax) {
220 double min2 = log(qmin);
221 double max2 = log(qmax);
224 q = (q - min2) / (max2 - min2) - 0.5;
226 q = 1.0 / (1.0 + exp(q));
227 q = q * (max2 - min2) + min2;
236 * Modify the bitrate curve from pass1 for one frame.
238 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
239 double rate_factor, int frame_num)
241 RateControlContext *rcc = &s->rc_context;
242 AVCodecContext *a = s->avctx;
243 const int pict_type = rce->new_pict_type;
244 const double mb_num = s->mb_num;
248 double const_values[] = {
251 rce->i_tex_bits * rce->qscale,
252 rce->p_tex_bits * rce->qscale,
253 (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
254 rce->mv_bits / mb_num,
255 rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
256 rce->i_count / mb_num,
257 rce->mc_mb_var_sum / mb_num,
258 rce->mb_var_sum / mb_num,
259 rce->pict_type == AV_PICTURE_TYPE_I,
260 rce->pict_type == AV_PICTURE_TYPE_P,
261 rce->pict_type == AV_PICTURE_TYPE_B,
262 rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
264 rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
265 rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
266 rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
267 rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
268 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
272 bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
274 av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->rc_eq);
278 rcc->pass1_rc_eq_output_sum += bits;
282 bits += 1.0; // avoid 1/0 issues
285 for (i = 0; i < s->avctx->rc_override_count; i++) {
286 RcOverride *rco = s->avctx->rc_override;
287 if (rco[i].start_frame > frame_num)
289 if (rco[i].end_frame < frame_num)
293 bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it?
295 bits *= rco[i].quality_factor;
298 q = bits2qp(rce, bits);
301 if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
302 q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
303 else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
304 q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
311 static int init_pass2(MpegEncContext *s)
313 RateControlContext *rcc = &s->rc_context;
314 AVCodecContext *a = s->avctx;
316 double fps = get_fps(s->avctx);
317 double complexity[5] = { 0 }; // approximate bits at quant=1
318 uint64_t const_bits[5] = { 0 }; // quantizer independent bits
319 uint64_t all_const_bits;
320 uint64_t all_available_bits = (uint64_t)(s->bit_rate *
321 (double)rcc->num_entries / fps);
322 double rate_factor = 0;
324 const int filter_size = (int)(a->qblur * 4) | 1;
325 double expected_bits = 0; // init to silence gcc warning
326 double *qscale, *blurred_qscale, qscale_sum;
328 /* find complexity & const_bits & decide the pict_types */
329 for (i = 0; i < rcc->num_entries; i++) {
330 RateControlEntry *rce = &rcc->entry[i];
332 rce->new_pict_type = rce->pict_type;
333 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
334 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
335 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
336 rcc->frame_count[rce->pict_type]++;
338 complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
340 const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
343 all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
344 const_bits[AV_PICTURE_TYPE_P] +
345 const_bits[AV_PICTURE_TYPE_B];
347 if (all_available_bits < all_const_bits) {
348 av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
352 qscale = av_malloc_array(rcc->num_entries, sizeof(double));
353 blurred_qscale = av_malloc_array(rcc->num_entries, sizeof(double));
354 if (!qscale || !blurred_qscale) {
356 av_free(blurred_qscale);
357 return AVERROR(ENOMEM);
361 for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
365 rcc->buffer_index = s->avctx->rc_buffer_size / 2;
368 for (i = 0; i < rcc->num_entries; i++) {
369 RateControlEntry *rce = &rcc->entry[i];
371 qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
372 rcc->last_qscale_for[rce->pict_type] = qscale[i];
374 av_assert0(filter_size % 2 == 1);
376 /* fixed I/B QP relative to P mode */
377 for (i = FFMAX(0, rcc->num_entries - 300); i < rcc->num_entries; i++) {
378 RateControlEntry *rce = &rcc->entry[i];
380 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
383 for (i = rcc->num_entries - 1; i >= 0; i--) {
384 RateControlEntry *rce = &rcc->entry[i];
386 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
390 for (i = 0; i < rcc->num_entries; i++) {
391 RateControlEntry *rce = &rcc->entry[i];
392 const int pict_type = rce->new_pict_type;
394 double q = 0.0, sum = 0.0;
396 for (j = 0; j < filter_size; j++) {
397 int index = i + j - filter_size / 2;
398 double d = index - i;
399 double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));
401 if (index < 0 || index >= rcc->num_entries)
403 if (pict_type != rcc->entry[index].new_pict_type)
405 q += qscale[index] * coeff;
408 blurred_qscale[i] = q / sum;
411 /* find expected bits */
412 for (i = 0; i < rcc->num_entries; i++) {
413 RateControlEntry *rce = &rcc->entry[i];
416 rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);
418 bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
419 bits += 8 * ff_vbv_update(s, bits);
421 rce->expected_bits = expected_bits;
422 expected_bits += bits;
426 "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
427 expected_bits, (int)all_available_bits, rate_factor);
428 if (expected_bits > all_available_bits) {
434 av_free(blurred_qscale);
436 /* check bitrate calculations and print info */
438 for (i = 0; i < rcc->num_entries; i++) {
439 ff_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
441 rcc->entry[i].new_qscale,
442 rcc->entry[i].new_qscale / FF_QP2LAMBDA);
443 qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
444 s->avctx->qmin, s->avctx->qmax);
446 av_assert0(toobig <= 40);
447 av_log(s->avctx, AV_LOG_DEBUG,
448 "[lavc rc] requested bitrate: %"PRId64" bps expected bitrate: %"PRId64" bps\n",
450 (int64_t)(expected_bits / ((double)all_available_bits / s->bit_rate)));
451 av_log(s->avctx, AV_LOG_DEBUG,
452 "[lavc rc] estimated target average qp: %.3f\n",
453 (float)qscale_sum / rcc->num_entries);
455 av_log(s->avctx, AV_LOG_INFO,
456 "[lavc rc] Using all of requested bitrate is not "
457 "necessary for this video with these parameters.\n");
458 } else if (toobig == 40) {
459 av_log(s->avctx, AV_LOG_ERROR,
460 "[lavc rc] Error: bitrate too low for this video "
461 "with these parameters.\n");
463 } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
464 av_log(s->avctx, AV_LOG_ERROR,
465 "[lavc rc] Error: 2pass curve failed to converge\n");
472 av_cold int ff_rate_control_init(MpegEncContext *s)
474 RateControlContext *rcc = &s->rc_context;
476 static const char * const const_names[] = {
499 static double (* const func1[])(void *, double) = {
500 (double (*)(void *, double)) bits2qp,
501 (double (*)(void *, double)) qp2bits,
504 static const char * const func1_names[] = {
511 if (!s->avctx->rc_max_available_vbv_use && s->avctx->rc_buffer_size) {
512 if (s->avctx->rc_max_rate) {
513 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);
515 s->avctx->rc_max_available_vbv_use = 1.0;
518 res = av_expr_parse(&rcc->rc_eq_eval,
519 s->rc_eq ? s->rc_eq : "tex^qComp",
520 const_names, func1_names, func1,
521 NULL, NULL, 0, s->avctx);
523 av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->rc_eq);
527 for (i = 0; i < 5; i++) {
528 rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0;
529 rcc->pred[i].count = 1.0;
530 rcc->pred[i].decay = 0.4;
532 rcc->i_cplx_sum [i] =
533 rcc->p_cplx_sum [i] =
534 rcc->mv_bits_sum[i] =
535 rcc->qscale_sum [i] =
536 rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such
538 rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5;
540 rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy;
541 if (!rcc->buffer_index)
542 rcc->buffer_index = s->avctx->rc_buffer_size * 3 / 4;
544 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
548 /* find number of pics */
549 p = s->avctx->stats_in;
551 p = strchr(p + 1, ';');
552 i += s->max_b_frames;
553 if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry))
555 rcc->entry = av_mallocz(i * sizeof(RateControlEntry));
557 return AVERROR(ENOMEM);
558 rcc->num_entries = i;
560 /* init all to skipped P-frames
561 * (with B-frames we might have a not encoded frame at the end FIXME) */
562 for (i = 0; i < rcc->num_entries; i++) {
563 RateControlEntry *rce = &rcc->entry[i];
565 rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P;
566 rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2;
567 rce->misc_bits = s->mb_num + 10;
568 rce->mb_var_sum = s->mb_num * 100;
572 p = s->avctx->stats_in;
573 for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
574 RateControlEntry *rce;
579 next = strchr(p, ';');
581 (*next) = 0; // sscanf is unbelievably slow on looong strings // FIXME copy / do not write
584 e = sscanf(p, " in:%d ", &picture_number);
586 av_assert0(picture_number >= 0);
587 av_assert0(picture_number < rcc->num_entries);
588 rce = &rcc->entry[picture_number];
590 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",
591 &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
592 &rce->mv_bits, &rce->misc_bits,
593 &rce->f_code, &rce->b_code,
594 &rce->mc_mb_var_sum, &rce->mb_var_sum,
595 &rce->i_count, &rce->skip_count, &rce->header_bits);
597 av_log(s->avctx, AV_LOG_ERROR,
598 "statistics are damaged at line %d, parser out=%d\n",
606 if (init_pass2(s) < 0) {
607 ff_rate_control_uninit(s);
612 if (!(s->avctx->flags & AV_CODEC_FLAG_PASS2)) {
613 rcc->short_term_qsum = 0.001;
614 rcc->short_term_qcount = 0.001;
616 rcc->pass1_rc_eq_output_sum = 0.001;
617 rcc->pass1_wanted_bits = 0.001;
619 if (s->avctx->qblur > 1.0) {
620 av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
623 /* init stuff with the user specified complexity */
624 if (s->rc_initial_cplx) {
625 for (i = 0; i < 60 * 30; i++) {
626 double bits = s->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
627 RateControlEntry rce;
629 if (i % ((s->gop_size + 3) / 4) == 0)
630 rce.pict_type = AV_PICTURE_TYPE_I;
631 else if (i % (s->max_b_frames + 1))
632 rce.pict_type = AV_PICTURE_TYPE_B;
634 rce.pict_type = AV_PICTURE_TYPE_P;
636 rce.new_pict_type = rce.pict_type;
637 rce.mc_mb_var_sum = bits * s->mb_num / 100000;
638 rce.mb_var_sum = s->mb_num;
640 rce.qscale = FF_QP2LAMBDA * 2;
645 if (s->pict_type == AV_PICTURE_TYPE_I) {
646 rce.i_count = s->mb_num;
647 rce.i_tex_bits = bits;
651 rce.i_count = 0; // FIXME we do know this approx
653 rce.p_tex_bits = bits * 0.9;
654 rce.mv_bits = bits * 0.1;
656 rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale;
657 rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale;
658 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
659 rcc->frame_count[rce.pict_type]++;
661 get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
663 // FIXME misbehaves a little for variable fps
664 rcc->pass1_wanted_bits += s->bit_rate / get_fps(s->avctx);
672 av_cold void ff_rate_control_uninit(MpegEncContext *s)
674 RateControlContext *rcc = &s->rc_context;
677 av_expr_free(rcc->rc_eq_eval);
678 av_freep(&rcc->entry);
681 int ff_vbv_update(MpegEncContext *s, int frame_size)
683 RateControlContext *rcc = &s->rc_context;
684 const double fps = get_fps(s->avctx);
685 const int buffer_size = s->avctx->rc_buffer_size;
686 const double min_rate = s->avctx->rc_min_rate / fps;
687 const double max_rate = s->avctx->rc_max_rate / fps;
689 ff_dlog(s, "%d %f %d %f %f\n",
690 buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
695 rcc->buffer_index -= frame_size;
696 if (rcc->buffer_index < 0) {
697 av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
698 if (frame_size > max_rate && s->qscale == s->avctx->qmax) {
699 av_log(s->avctx, AV_LOG_ERROR, "max bitrate possibly too small or try trellis with large lmax or increase qmax\n");
701 rcc->buffer_index = 0;
704 left = buffer_size - rcc->buffer_index - 1;
705 rcc->buffer_index += av_clip(left, min_rate, max_rate);
707 if (rcc->buffer_index > buffer_size) {
708 int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
710 if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
712 rcc->buffer_index -= 8 * stuffing;
714 if (s->avctx->debug & FF_DEBUG_RC)
715 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
723 static double predict_size(Predictor *p, double q, double var)
725 return p->coeff * var / (q * p->count);
728 static void update_predictor(Predictor *p, double q, double var, double size)
730 double new_coeff = size * q / (var + 1);
734 p->count *= p->decay;
735 p->coeff *= p->decay;
737 p->coeff += new_coeff;
740 static void adaptive_quantization(MpegEncContext *s, double q)
743 const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0);
744 const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0);
745 const float temp_cplx_masking = s->avctx->temporal_cplx_masking;
746 const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
747 const float p_masking = s->avctx->p_masking;
748 const float border_masking = s->border_masking;
749 float bits_sum = 0.0;
750 float cplx_sum = 0.0;
751 float *cplx_tab = s->cplx_tab;
752 float *bits_tab = s->bits_tab;
753 const int qmin = s->avctx->mb_lmin;
754 const int qmax = s->avctx->mb_lmax;
755 Picture *const pic = &s->current_picture;
756 const int mb_width = s->mb_width;
757 const int mb_height = s->mb_height;
759 for (i = 0; i < s->mb_num; i++) {
760 const int mb_xy = s->mb_index2xy[i];
761 float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
762 float spat_cplx = sqrt(pic->mb_var[mb_xy]);
763 const int lumi = pic->mb_mean[mb_xy];
764 float bits, cplx, factor;
765 int mb_x = mb_xy % s->mb_stride;
766 int mb_y = mb_xy / s->mb_stride;
768 float mb_factor = 0.0;
770 spat_cplx = 4; // FIXME fine-tune
772 temp_cplx = 4; // FIXME fine-tune
774 if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
776 factor = 1.0 + p_masking;
779 factor = pow(temp_cplx, -temp_cplx_masking);
781 factor *= pow(spat_cplx, -spatial_cplx_masking);
784 factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
786 factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);
788 if (mb_x < mb_width / 5) {
789 mb_distance = mb_width / 5 - mb_x;
790 mb_factor = (float)mb_distance / (float)(mb_width / 5);
791 } else if (mb_x > 4 * mb_width / 5) {
792 mb_distance = mb_x - 4 * mb_width / 5;
793 mb_factor = (float)mb_distance / (float)(mb_width / 5);
795 if (mb_y < mb_height / 5) {
796 mb_distance = mb_height / 5 - mb_y;
797 mb_factor = FFMAX(mb_factor,
798 (float)mb_distance / (float)(mb_height / 5));
799 } else if (mb_y > 4 * mb_height / 5) {
800 mb_distance = mb_y - 4 * mb_height / 5;
801 mb_factor = FFMAX(mb_factor,
802 (float)mb_distance / (float)(mb_height / 5));
805 factor *= 1.0 - border_masking * mb_factor;
807 if (factor < 0.00001)
810 bits = cplx * factor;
817 /* handle qmin/qmax clipping */
818 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
819 float factor = bits_sum / cplx_sum;
820 for (i = 0; i < s->mb_num; i++) {
821 float newq = q * cplx_tab[i] / bits_tab[i];
825 bits_sum -= bits_tab[i];
826 cplx_sum -= cplx_tab[i] * q / qmax;
827 } else if (newq < qmin) {
828 bits_sum -= bits_tab[i];
829 cplx_sum -= cplx_tab[i] * q / qmin;
832 if (bits_sum < 0.001)
834 if (cplx_sum < 0.001)
838 for (i = 0; i < s->mb_num; i++) {
839 const int mb_xy = s->mb_index2xy[i];
840 float newq = q * cplx_tab[i] / bits_tab[i];
843 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
844 newq *= bits_sum / cplx_sum;
847 intq = (int)(newq + 0.5);
851 else if (intq < qmin)
853 s->lambda_table[mb_xy] = intq;
857 void ff_get_2pass_fcode(MpegEncContext *s)
859 RateControlContext *rcc = &s->rc_context;
860 RateControlEntry *rce = &rcc->entry[s->picture_number];
862 s->f_code = rce->f_code;
863 s->b_code = rce->b_code;
866 // FIXME rd or at least approx for dquant
868 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
872 float br_compensation;
876 int picture_number = s->picture_number;
878 RateControlContext *rcc = &s->rc_context;
879 AVCodecContext *a = s->avctx;
880 RateControlEntry local_rce, *rce;
884 const int pict_type = s->pict_type;
885 Picture * const pic = &s->current_picture;
888 get_qminmax(&qmin, &qmax, s, pict_type);
890 fps = get_fps(s->avctx);
891 /* update predictors */
892 if (picture_number > 2 && !dry_run) {
893 const int64_t last_var =
894 s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
895 : rcc->last_mc_mb_var_sum;
896 av_assert1(s->frame_bits >= s->stuffing_bits);
897 update_predictor(&rcc->pred[s->last_pict_type],
900 s->frame_bits - s->stuffing_bits);
903 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
904 av_assert0(picture_number >= 0);
905 if (picture_number >= rcc->num_entries) {
906 av_log(s, AV_LOG_ERROR, "Input is longer than 2-pass log file\n");
909 rce = &rcc->entry[picture_number];
910 wanted_bits = rce->expected_bits;
915 /* FIXME add a dts field to AVFrame and ensure it is set and use it
916 * here instead of reordering but the reordering is simpler for now
917 * until H.264 B-pyramid must be handled. */
918 if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
919 dts_pic = s->current_picture_ptr;
921 dts_pic = s->last_picture_ptr;
923 if (!dts_pic || dts_pic->f->pts == AV_NOPTS_VALUE)
924 wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
926 wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f->pts / fps);
929 diff = s->total_bits - wanted_bits;
930 br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
931 if (br_compensation <= 0.0)
932 br_compensation = 0.001;
934 var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
936 short_term_q = 0; /* avoid warning */
937 if (s->avctx->flags & AV_CODEC_FLAG_PASS2) {
938 if (pict_type != AV_PICTURE_TYPE_I)
939 av_assert0(pict_type == rce->new_pict_type);
941 q = rce->new_qscale / br_compensation;
942 ff_dlog(s, "%f %f %f last:%d var:%"PRId64" type:%d//\n", q, rce->new_qscale,
943 br_compensation, s->frame_bits, var, pict_type);
946 rce->new_pict_type = pict_type;
947 rce->mc_mb_var_sum = pic->mc_mb_var_sum;
948 rce->mb_var_sum = pic->mb_var_sum;
949 rce->qscale = FF_QP2LAMBDA * 2;
950 rce->f_code = s->f_code;
951 rce->b_code = s->b_code;
954 bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
955 if (pict_type == AV_PICTURE_TYPE_I) {
956 rce->i_count = s->mb_num;
957 rce->i_tex_bits = bits;
961 rce->i_count = 0; // FIXME we do know this approx
963 rce->p_tex_bits = bits * 0.9;
964 rce->mv_bits = bits * 0.1;
966 rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale;
967 rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale;
968 rcc->mv_bits_sum[pict_type] += rce->mv_bits;
969 rcc->frame_count[pict_type]++;
971 rate_factor = rcc->pass1_wanted_bits /
972 rcc->pass1_rc_eq_output_sum * br_compensation;
974 q = get_qscale(s, rce, rate_factor, picture_number);
979 q = get_diff_limited_q(s, rce, q);
982 // FIXME type dependent blur like in 2-pass
983 if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
984 rcc->short_term_qsum *= a->qblur;
985 rcc->short_term_qcount *= a->qblur;
987 rcc->short_term_qsum += q;
988 rcc->short_term_qcount++;
989 q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
993 q = modify_qscale(s, rce, q, picture_number);
995 rcc->pass1_wanted_bits += s->bit_rate / fps;
1000 if (s->avctx->debug & FF_DEBUG_RC) {
1001 av_log(s->avctx, AV_LOG_DEBUG,
1002 "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
1003 "size:%d var:%"PRId64"/%"PRId64" br:%"PRId64" fps:%d\n",
1004 av_get_picture_type_char(pict_type),
1005 qmin, q, qmax, picture_number,
1006 (int)wanted_bits / 1000, (int)s->total_bits / 1000,
1007 br_compensation, short_term_q, s->frame_bits,
1008 pic->mb_var_sum, pic->mc_mb_var_sum,
1009 s->bit_rate / 1000, (int)fps);
1017 if (s->adaptive_quant)
1018 adaptive_quantization(s, q);
1023 rcc->last_qscale = q;
1024 rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
1025 rcc->last_mb_var_sum = pic->mb_var_sum;