1 /***************************************************-*- coding: iso-8859-1 -*-
2 * ratecontrol.c: h264 encoder library (Rate Control)
3 *****************************************************************************
4 * Copyright (C) 2005 x264 project
5 * $Id: ratecontrol.c,v 1.1 2004/06/03 19:27:08 fenrir Exp $
7 * Authors: Loren Merritt <lorenm@u.washington.edu>
8 * Michael Niedermayer <michaelni@gmx.at>
9 * Måns Rullgård <mru@mru.ath.cx>
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2 of the License, or
14 * (at your option) any later version.
16 * This program is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with this program; if not, write to the Free Software
23 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111, USA.
24 *****************************************************************************/
26 #define _ISOC99_SOURCE
27 #undef NDEBUG // always check asserts, the speed effect is far too small to disable them
34 #include "common/common.h"
35 #include "common/cpu.h"
36 #include "ratecontrol.h"
38 #if defined(SYS_OPENBSD)
39 #define isfinite finite
42 #define isfinite _finite
44 #if defined(_MSC_VER) || defined(SYS_SunOS) || defined(SYS_MACOSX)
47 #ifdef WIN32 // POSIX says that rename() removes the destination, but win32 doesn't.
48 #define rename(src,dst) (unlink(dst), rename(src,dst))
60 uint64_t expected_bits;
66 float blurred_complexity;
68 } ratecontrol_entry_t;
77 struct x264_ratecontrol_t
86 double rate_tolerance;
87 int nmb; /* number of macroblocks in a frame */
91 ratecontrol_entry_t *rce;
92 int qp; /* qp for current frame */
93 int qpm; /* qp for current macroblock */
94 float qpa; /* average of macroblocks' qp */
101 double buffer_rate; /* # of bits added to buffer_fill after each frame */
102 predictor_t pred[5]; /* predict frame size from satd */
107 double cplxr_sum; /* sum of bits*qscale/rceq */
108 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow */
109 double wanted_bits_window; /* target bitrate * window */
111 double short_term_cplxsum;
112 double short_term_cplxcount;
113 double rate_factor_constant;
118 FILE *p_stat_file_out;
119 char *psz_stat_file_tmpname;
121 int num_entries; /* number of ratecontrol_entry_ts */
122 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
124 double last_qscale_for[5]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
125 int last_non_b_pict_type;
126 double accum_p_qp; /* for determining I-frame quant */
128 double last_accum_p_norm;
129 double lmin[5]; /* min qscale by frame type */
131 double lstep; /* max change (multiply) in qscale per frame */
132 double i_cplx_sum[5]; /* estimated total texture bits in intra MBs at qscale=1 */
133 double p_cplx_sum[5];
134 double mv_bits_sum[5];
135 int frame_count[5]; /* number of frames of each type */
138 double frame_size_planned;
139 int first_row, last_row; /* region of the frame to be encoded by this thread */
140 predictor_t *row_pred;
141 predictor_t row_preds[5];
142 predictor_t pred_b_from_p; /* predict B-frame size from P-frame satd */
143 int bframes; /* # consecutive B-frames before this P-frame */
144 int bframe_bits; /* total cost of those frames */
151 static int parse_zones( x264_t *h );
152 static int init_pass2(x264_t *);
153 static float rate_estimate_qscale( x264_t *h, int pict_type );
154 static void update_vbv( x264_t *h, int bits );
155 static double predict_size( predictor_t *p, double q, double var );
156 static void update_predictor( predictor_t *p, double q, double var, double bits );
157 int x264_rc_analyse_slice( x264_t *h );
160 * qp = h.264's quantizer
161 * qscale = linearized quantizer = Lagrange multiplier
163 static inline double qp2qscale(double qp)
165 return 0.85 * pow(2.0, ( qp - 12.0 ) / 6.0);
167 static inline double qscale2qp(double qscale)
169 return 12.0 + 6.0 * log(qscale/0.85) / log(2.0);
172 /* Texture bitrate is not quite inversely proportional to qscale,
173 * probably due the the changing number of SKIP blocks.
174 * MV bits level off at about qp<=12, because the lambda used
175 * for motion estimation is constant there. */
176 static inline double qscale2bits(ratecontrol_entry_t *rce, double qscale)
180 return (rce->i_tex_bits + rce->p_tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
181 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
186 int x264_ratecontrol_new( x264_t *h )
188 x264_ratecontrol_t *rc;
191 x264_cpu_restore( h->param.cpu );
193 h->rc = rc = x264_malloc( h->param.i_threads * sizeof(x264_ratecontrol_t) );
194 memset( rc, 0, h->param.i_threads * sizeof(x264_ratecontrol_t) );
196 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
197 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
199 /* FIXME: use integers */
200 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
201 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
205 rc->bitrate = h->param.rc.i_bitrate * 1000;
206 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
207 rc->nmb = h->mb.i_mb_count;
208 rc->last_non_b_pict_type = -1;
211 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
213 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
216 if( h->param.rc.i_vbv_buffer_size )
218 if( h->param.rc.i_rc_method == X264_RC_CQP )
219 x264_log(h, X264_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
220 else if( h->param.rc.i_vbv_max_bitrate == 0 )
222 x264_log( h, X264_LOG_DEBUG, "VBV maxrate unspecified, assuming CBR\n" );
223 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
226 if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate &&
227 h->param.rc.i_vbv_max_bitrate > 0)
228 x264_log(h, X264_LOG_WARNING, "max bitrate less than average bitrate, ignored.\n");
229 else if( h->param.rc.i_vbv_max_bitrate > 0 &&
230 h->param.rc.i_vbv_buffer_size > 0 )
232 if( h->param.rc.i_vbv_buffer_size < 3 * h->param.rc.i_vbv_max_bitrate / rc->fps )
234 h->param.rc.i_vbv_buffer_size = 3 * h->param.rc.i_vbv_max_bitrate / rc->fps;
235 x264_log( h, X264_LOG_WARNING, "VBV buffer size too small, using %d kbit\n",
236 h->param.rc.i_vbv_buffer_size );
238 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000 / rc->fps;
239 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000;
240 rc->buffer_fill = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
241 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
242 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
244 rc->b_vbv_min_rate = !rc->b_2pass
245 && h->param.rc.i_rc_method == X264_RC_ABR
246 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
248 else if( h->param.rc.i_vbv_max_bitrate )
250 x264_log(h, X264_LOG_WARNING, "VBV maxrate specified, but no bufsize.\n");
251 h->param.rc.i_vbv_max_bitrate = 0;
253 if(rc->rate_tolerance < 0.01) {
254 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
255 rc->rate_tolerance = 0.01;
258 h->mb.b_variable_qp = rc->b_vbv && !rc->b_2pass;
262 /* FIXME ABR_INIT_QP is actually used only in CRF */
263 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
264 rc->accum_p_norm = .01;
265 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
266 /* estimated ratio that produces a reasonable QP for the first I-frame */
267 rc->cplxr_sum = .01 * pow( 7.0e5, h->param.rc.f_qcompress ) * pow( h->mb.i_mb_count, 0.5 );
268 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
269 rc->last_non_b_pict_type = SLICE_TYPE_I;
272 if( h->param.rc.i_rc_method == X264_RC_CRF )
274 /* arbitrary rescaling to make CRF somewhat similar to QP */
275 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
276 rc->rate_factor_constant = pow( base_cplx, 1 - h->param.rc.f_qcompress )
277 / qp2qscale( h->param.rc.f_rf_constant );
280 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
281 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
282 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
283 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
284 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
286 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
287 rc->last_qscale = qp2qscale(26);
288 for( i = 0; i < 5; i++ )
290 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
291 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
292 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
293 rc->pred[i].coeff= 2.0;
294 rc->pred[i].count= 1.0;
295 rc->pred[i].decay= 0.5;
296 rc->row_preds[i].coeff= .25;
297 rc->row_preds[i].count= 1.0;
298 rc->row_preds[i].decay= 0.5;
300 rc->pred_b_from_p = rc->pred[0];
302 if( parse_zones( h ) < 0 )
305 /* Load stat file and init 2pass algo */
306 if( h->param.rc.b_stat_read )
308 char *p, *stats_in, *stats_buf;
310 /* read 1st pass stats */
311 assert( h->param.rc.psz_stat_in );
312 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
315 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
319 /* check whether 1st pass options were compatible with current options */
320 if( !strncmp( stats_buf, "#options:", 9 ) )
323 char *opts = stats_buf;
324 stats_in = strchr( stats_buf, '\n' );
330 if( ( p = strstr( opts, "bframes=" ) ) && sscanf( p, "bframes=%d", &i )
331 && h->param.i_bframe != i )
333 x264_log( h, X264_LOG_ERROR, "different number of B-frames than 1st pass (%d vs %d)\n",
334 h->param.i_bframe, i );
338 /* since B-adapt doesn't (yet) take into account B-pyramid,
339 * the converse is not a problem */
340 if( strstr( opts, "b_pyramid=1" ) && !h->param.b_bframe_pyramid )
341 x264_log( h, X264_LOG_WARNING, "1st pass used B-pyramid, 2nd doesn't\n" );
343 if( ( p = strstr( opts, "keyint=" ) ) && sscanf( p, "keyint=%d", &i )
344 && h->param.i_keyint_max != i )
345 x264_log( h, X264_LOG_WARNING, "different keyint than 1st pass (%d vs %d)\n",
346 h->param.i_keyint_max, i );
348 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
349 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
352 /* find number of pics */
355 p = strchr(p+1, ';');
358 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
363 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
365 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
366 h->param.i_frame_total, rc->num_entries );
368 if( h->param.i_frame_total > rc->num_entries + h->param.i_bframe )
370 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
371 h->param.i_frame_total, rc->num_entries );
375 /* FIXME: ugly padding because VfW drops delayed B-frames */
376 rc->num_entries += h->param.i_bframe;
378 rc->entry = (ratecontrol_entry_t*) x264_malloc(rc->num_entries * sizeof(ratecontrol_entry_t));
379 memset(rc->entry, 0, rc->num_entries * sizeof(ratecontrol_entry_t));
381 /* init all to skipped p frames */
382 for(i=0; i<rc->num_entries; i++){
383 ratecontrol_entry_t *rce = &rc->entry[i];
384 rce->pict_type = SLICE_TYPE_P;
385 rce->qscale = rce->new_qscale = qp2qscale(20);
386 rce->misc_bits = rc->nmb + 10;
392 for(i=0; i < rc->num_entries - h->param.i_bframe; i++){
393 ratecontrol_entry_t *rce;
400 next= strchr(p, ';');
402 (*next)=0; //sscanf is unbelievably slow on looong strings
405 e = sscanf(p, " in:%d ", &frame_number);
407 if(frame_number < 0 || frame_number >= rc->num_entries)
409 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
412 rce = &rc->entry[frame_number];
413 rce->direct_mode = 0;
415 e += sscanf(p, " in:%*d out:%*d type:%c q:%f itex:%d ptex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c",
416 &pict_type, &qp, &rce->i_tex_bits, &rce->p_tex_bits,
417 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
418 &rce->s_count, &rce->direct_mode);
421 case 'I': rce->kept_as_ref = 1;
422 case 'i': rce->pict_type = SLICE_TYPE_I; break;
423 case 'P': rce->pict_type = SLICE_TYPE_P; break;
424 case 'B': rce->kept_as_ref = 1;
425 case 'b': rce->pict_type = SLICE_TYPE_B; break;
426 default: e = -1; break;
429 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
432 rce->qscale = qp2qscale(qp);
436 x264_free(stats_buf);
438 if(h->param.rc.i_rc_method == X264_RC_ABR)
440 if(init_pass2(h) < 0) return -1;
441 } /* else we're using constant quant, so no need to run the bitrate allocation */
444 /* Open output file */
445 /* If input and output files are the same, output to a temp file
446 * and move it to the real name only when it's complete */
447 if( h->param.rc.b_stat_write )
451 rc->psz_stat_file_tmpname = x264_malloc( strlen(h->param.rc.psz_stat_out) + 6 );
452 strcpy( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
453 strcat( rc->psz_stat_file_tmpname, ".temp" );
455 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
456 if( rc->p_stat_file_out == NULL )
458 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
462 p = x264_param2string( &h->param, 1 );
463 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
470 static int parse_zones( x264_t *h )
472 x264_ratecontrol_t *rc = h->rc;
474 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
477 h->param.rc.i_zones = 1;
478 for( p = h->param.rc.psz_zones; *p; p++ )
479 h->param.rc.i_zones += (*p == '/');
480 h->param.rc.zones = x264_malloc( h->param.rc.i_zones * sizeof(x264_zone_t) );
481 p = h->param.rc.psz_zones;
482 for( i = 0; i < h->param.rc.i_zones; i++)
484 x264_zone_t *z = &h->param.rc.zones[i];
485 if( 3 == sscanf(p, "%u,%u,q=%u", &z->i_start, &z->i_end, &z->i_qp) )
487 else if( 3 == sscanf(p, "%u,%u,b=%f", &z->i_start, &z->i_end, &z->f_bitrate_factor) )
491 char *slash = strchr(p, '/');
492 if(slash) *slash = '\0';
493 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
496 p = strchr(p, '/') + 1;
500 if( h->param.rc.i_zones > 0 )
502 for( i = 0; i < h->param.rc.i_zones; i++ )
504 x264_zone_t z = h->param.rc.zones[i];
505 if( z.i_start < 0 || z.i_start > z.i_end )
507 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
508 z.i_start, z.i_end );
511 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
513 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
514 z.f_bitrate_factor );
519 rc->i_zones = h->param.rc.i_zones;
520 rc->zones = x264_malloc( rc->i_zones * sizeof(x264_zone_t) );
521 memcpy( rc->zones, h->param.rc.zones, rc->i_zones * sizeof(x264_zone_t) );
527 x264_zone_t *get_zone( x264_t *h, int frame_num )
530 for( i = h->rc->i_zones-1; i >= 0; i-- )
532 x264_zone_t *z = &h->rc->zones[i];
533 if( frame_num >= z->i_start && frame_num <= z->i_end )
539 void x264_ratecontrol_summary( x264_t *h )
541 x264_ratecontrol_t *rc = h->rc;
542 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
544 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
545 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
546 qscale2qp( pow( base_cplx, 1 - h->param.rc.f_qcompress )
547 * rc->cplxr_sum / rc->wanted_bits_window ) );
551 void x264_ratecontrol_delete( x264_t *h )
553 x264_ratecontrol_t *rc = h->rc;
555 if( rc->p_stat_file_out )
557 fclose( rc->p_stat_file_out );
558 if( h->i_frame >= rc->num_entries - h->param.i_bframe )
559 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
561 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
562 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
564 x264_free( rc->psz_stat_file_tmpname );
566 x264_free( rc->entry );
567 x264_free( rc->zones );
571 /* Before encoding a frame, choose a QP for it */
572 void x264_ratecontrol_start( x264_t *h, int i_slice_type, int i_force_qp )
574 x264_ratecontrol_t *rc = h->rc;
575 ratecontrol_entry_t *rce = NULL;
577 x264_cpu_restore( h->param.cpu );
579 rc->qp_force = i_force_qp;
580 rc->slice_type = i_slice_type;
582 if( h->param.rc.b_stat_read )
584 int frame = h->fenc->i_frame;
585 assert( frame >= 0 && frame < rc->num_entries );
586 rce = h->rc->rce = &h->rc->entry[frame];
588 if( i_slice_type == SLICE_TYPE_B
589 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
591 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
592 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
596 if( h->fdec->i_row_bits )
598 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
601 if( i_slice_type != SLICE_TYPE_B )
605 while( h->frames.current[rc->bframes] && IS_X264_TYPE_B(h->frames.current[rc->bframes]->i_type) )
613 rc->qpm = rc->qp = i_force_qp - 1;
618 x264_clip3( (int)(qscale2qp( rate_estimate_qscale( h, i_slice_type ) ) + .5), 0, 51 );
620 else if( rc->b_2pass )
622 rce->new_qscale = rate_estimate_qscale( h, i_slice_type );
623 rc->qpm = rc->qp = rce->new_qp =
624 x264_clip3( (int)(qscale2qp(rce->new_qscale) + 0.5), 0, 51 );
628 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
630 if( i_slice_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
631 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
633 q = rc->qp_constant[ i_slice_type ];
637 if( zone->b_force_qp )
638 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
640 q -= 6*log(zone->f_bitrate_factor)/log(2);
643 rc->qpm = rc->qp = (int)(q + 0.5);
647 double predict_row_size( x264_t *h, int y, int qp )
649 /* average between two predictors:
650 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
651 x264_ratecontrol_t *rc = h->rc;
652 double pred_s = predict_size( rc->row_pred, qp2qscale(qp), h->fdec->i_row_satd[y] );
654 if( rc->slice_type != SLICE_TYPE_I
655 && h->fref0[0]->i_type == h->fdec->i_type
656 && h->fref0[0]->i_row_satd[y] > 0 )
658 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
659 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
664 return (pred_s + pred_t) / 2;
667 double predict_row_size_sum( x264_t *h, int y, int qp )
671 for( i = h->rc->first_row; i <= y; i++ )
672 bits += h->fdec->i_row_bits[i];
673 for( i = y+1; i <= h->rc->last_row; i++ )
674 bits += predict_row_size( h, i, qp );
678 void x264_ratecontrol_mb( x264_t *h, int bits )
680 x264_ratecontrol_t *rc = h->rc;
681 const int y = h->mb.i_mb_y;
683 x264_cpu_restore( h->param.cpu );
685 h->fdec->i_row_bits[y] += bits;
688 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !h->mb.b_variable_qp )
691 h->fdec->i_row_qp[y] = rc->qpm;
693 if( rc->slice_type == SLICE_TYPE_B )
695 /* B-frames shouldn't use lower QP than their reference frames */
696 if( y < rc->last_row )
698 rc->qpm = X264_MAX( rc->qp,
699 X264_MIN( h->fref0[0]->i_row_qp[y+1],
700 h->fref1[0]->i_row_qp[y+1] ));
705 update_predictor( rc->row_pred, qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
707 /* tweak quality based on difference from predicted size */
708 if( y < rc->last_row && h->stat.i_slice_count[rc->slice_type] > 0 )
710 int prev_row_qp = h->fdec->i_row_qp[y];
711 int b0 = predict_row_size_sum( h, y, rc->qpm );
713 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
714 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
715 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
717 if( !rc->b_vbv_min_rate )
718 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
720 while( rc->qpm < i_qp_max
721 && (b1 > rc->frame_size_planned * 1.15
722 || (rc->buffer_fill - b1 < buffer_left_planned * 0.5)))
725 b1 = predict_row_size_sum( h, y, rc->qpm );
728 while( rc->qpm > i_qp_min
729 && buffer_left_planned > rc->buffer_size * 0.4
730 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
731 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
734 b1 = predict_row_size_sum( h, y, rc->qpm );
740 int x264_ratecontrol_qp( x264_t *h )
745 /* In 2pass, force the same frame types as in the 1st pass */
746 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
748 x264_ratecontrol_t *rc = h->rc;
749 if( h->param.rc.b_stat_read )
751 if( frame_num >= rc->num_entries )
753 /* We could try to initialize everything required for ABR and
754 * adaptive B-frames, but that would be complicated.
755 * So just calculate the average QP used so far. */
757 h->param.rc.i_qp_constant = (h->stat.i_slice_count[SLICE_TYPE_P] == 0) ? 24
758 : 1 + h->stat.i_slice_qp[SLICE_TYPE_P] / h->stat.i_slice_count[SLICE_TYPE_P];
759 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
760 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( (int)( qscale2qp( qp2qscale( h->param.rc.i_qp_constant ) / fabs( h->param.rc.f_ip_factor )) + 0.5 ), 0, 51 );
761 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( (int)( qscale2qp( qp2qscale( h->param.rc.i_qp_constant ) * fabs( h->param.rc.f_pb_factor )) + 0.5 ), 0, 51 );
763 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
764 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
765 if( h->param.b_bframe_adaptive )
766 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
770 h->param.rc.i_rc_method = X264_RC_CQP;
771 h->param.rc.b_stat_read = 0;
772 h->param.b_bframe_adaptive = 0;
773 if( h->param.i_bframe > 1 )
774 h->param.i_bframe = 1;
777 switch( rc->entry[frame_num].pict_type )
780 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I;
783 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B;
792 return X264_TYPE_AUTO;
796 /* After encoding one frame, save stats and update ratecontrol state */
797 void x264_ratecontrol_end( x264_t *h, int bits )
799 x264_ratecontrol_t *rc = h->rc;
800 const int *mbs = h->stat.frame.i_mb_count;
803 x264_cpu_restore( h->param.cpu );
805 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
806 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
807 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
808 for( i = B_DIRECT; i < B_8x8; i++ )
809 h->stat.frame.i_mb_count_p += mbs[i];
811 if( h->mb.b_variable_qp )
813 for( i = 1; i < h->param.i_threads; i++ )
814 rc->qpa += rc[i].qpa;
815 rc->qpa /= h->mb.i_mb_count;
819 h->fdec->f_qp_avg = rc->qpa;
821 if( h->param.rc.b_stat_write )
823 char c_type = rc->slice_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
824 : rc->slice_type==SLICE_TYPE_P ? 'P'
825 : h->fenc->b_kept_as_ref ? 'B' : 'b';
826 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
827 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
828 char c_direct = h->mb.b_direct_auto_write ?
829 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
830 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
832 fprintf( rc->p_stat_file_out,
833 "in:%d out:%d type:%c q:%.2f itex:%d ptex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c;\n",
834 h->fenc->i_frame, h->i_frame,
836 h->stat.frame.i_itex_bits, h->stat.frame.i_ptex_bits,
837 h->stat.frame.i_hdr_bits, h->stat.frame.i_misc_bits,
838 h->stat.frame.i_mb_count_i,
839 h->stat.frame.i_mb_count_p,
840 h->stat.frame.i_mb_count_skip,
846 if( rc->slice_type != SLICE_TYPE_B )
847 rc->cplxr_sum += bits * qp2qscale(rc->qpa) / rc->last_rceq;
850 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
851 * Not perfectly accurate with B-refs, but good enough. */
852 rc->cplxr_sum += bits * qp2qscale(rc->qpa) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
854 rc->cplxr_sum *= rc->cbr_decay;
855 rc->wanted_bits_window += rc->bitrate / rc->fps;
856 rc->wanted_bits_window *= rc->cbr_decay;
858 rc->accum_p_qp *= .95;
859 rc->accum_p_norm *= .95;
860 rc->accum_p_norm += 1;
861 if( rc->slice_type == SLICE_TYPE_I )
862 rc->accum_p_qp += rc->qpa * fabs(h->param.rc.f_ip_factor);
864 rc->accum_p_qp += rc->qpa;
869 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
872 if( h->mb.b_variable_qp )
874 if( rc->slice_type == SLICE_TYPE_B )
876 rc->bframe_bits += bits;
877 if( !h->frames.current[0] || !IS_X264_TYPE_B(h->frames.current[0]->i_type) )
878 update_predictor( &rc->pred_b_from_p, qp2qscale(rc->qpa), h->fref1[0]->i_satd, rc->bframe_bits / rc->bframes );
882 /* Update row predictor based on data collected by other threads. */
884 for( y = rc->last_row+1; y < h->sps->i_mb_height; y++ )
885 update_predictor( rc->row_pred, qp2qscale(h->fdec->i_row_qp[y]), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
886 rc->row_preds[rc->slice_type] = *rc->row_pred;
890 update_vbv( h, bits );
892 if( rc->slice_type != SLICE_TYPE_B )
893 rc->last_non_b_pict_type = rc->slice_type;
896 /****************************************************************************
898 ***************************************************************************/
900 double x264_eval( char *s, double *const_value, const char **const_name,
901 double (**func1)(void *, double), const char **func1_name,
902 double (**func2)(void *, double, double), char **func2_name,
906 * modify the bitrate curve from pass1 for one frame
908 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
910 x264_ratecontrol_t *rcc= h->rc;
911 const int pict_type = rce->pict_type;
913 x264_zone_t *zone = get_zone( h, frame_num );
915 double const_values[]={
916 rce->i_tex_bits * rce->qscale,
917 rce->p_tex_bits * rce->qscale,
918 (rce->i_tex_bits + rce->p_tex_bits) * rce->qscale,
919 rce->mv_bits * rce->qscale,
920 (double)rce->i_count / rcc->nmb,
921 (double)rce->p_count / rcc->nmb,
922 (double)rce->s_count / rcc->nmb,
923 rce->pict_type == SLICE_TYPE_I,
924 rce->pict_type == SLICE_TYPE_P,
925 rce->pict_type == SLICE_TYPE_B,
926 h->param.rc.f_qcompress,
927 rcc->i_cplx_sum[SLICE_TYPE_I] / rcc->frame_count[SLICE_TYPE_I],
928 rcc->i_cplx_sum[SLICE_TYPE_P] / rcc->frame_count[SLICE_TYPE_P],
929 rcc->p_cplx_sum[SLICE_TYPE_P] / rcc->frame_count[SLICE_TYPE_P],
930 rcc->p_cplx_sum[SLICE_TYPE_B] / rcc->frame_count[SLICE_TYPE_B],
931 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / rcc->frame_count[pict_type],
932 rce->blurred_complexity,
935 static const char *const_names[]={
955 static double (*func1[])(void *, double)={
956 // (void *)bits2qscale,
960 static const char *func1_names[]={
966 q = x264_eval((char*)h->param.rc.psz_rc_eq, const_values, const_names, func1, func1_names, NULL, NULL, rce);
968 // avoid NaN's in the rc_eq
969 if(!isfinite(q) || rce->i_tex_bits + rce->p_tex_bits + rce->mv_bits == 0)
970 q = rcc->last_qscale;
974 rcc->last_qscale = q;
979 if( zone->b_force_qp )
980 q = qp2qscale(zone->i_qp);
982 q /= zone->f_bitrate_factor;
988 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
990 x264_ratecontrol_t *rcc = h->rc;
991 const int pict_type = rce->pict_type;
993 // force I/B quants as a function of P quants
994 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
995 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
996 if( pict_type == SLICE_TYPE_I )
999 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1000 double ip_factor = fabs( h->param.rc.f_ip_factor );
1001 /* don't apply ip_factor if the following frame is also I */
1002 if( rcc->accum_p_norm <= 0 )
1004 else if( h->param.rc.f_ip_factor < 0 )
1006 else if( rcc->accum_p_norm >= 1 )
1009 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1011 else if( pict_type == SLICE_TYPE_B )
1013 if( h->param.rc.f_pb_factor > 0 )
1015 if( !rce->kept_as_ref )
1016 q *= fabs( h->param.rc.f_pb_factor );
1018 else if( pict_type == SLICE_TYPE_P
1019 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1020 && rce->i_tex_bits + rce->p_tex_bits == 0 )
1025 /* last qscale / qdiff stuff */
1026 if(rcc->last_non_b_pict_type==pict_type
1027 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1029 double last_q = rcc->last_qscale_for[pict_type];
1030 double max_qscale = last_q * rcc->lstep;
1031 double min_qscale = last_q / rcc->lstep;
1033 if (q > max_qscale) q = max_qscale;
1034 else if(q < min_qscale) q = min_qscale;
1037 rcc->last_qscale_for[pict_type] = q;
1038 if(pict_type!=SLICE_TYPE_B)
1039 rcc->last_non_b_pict_type = pict_type;
1040 if(pict_type==SLICE_TYPE_I)
1042 rcc->last_accum_p_norm = rcc->accum_p_norm;
1043 rcc->accum_p_norm = 0;
1044 rcc->accum_p_qp = 0;
1046 if(pict_type==SLICE_TYPE_P)
1048 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1049 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1050 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1055 static double predict_size( predictor_t *p, double q, double var )
1057 return p->coeff*var / (q*p->count);
1060 static void update_predictor( predictor_t *p, double q, double var, double bits )
1064 p->count *= p->decay;
1065 p->coeff *= p->decay;
1067 p->coeff += bits*q / var;
1070 static void update_vbv( x264_t *h, int bits )
1072 x264_ratecontrol_t *rcc = h->rc;
1074 if( rcc->last_satd >= h->mb.i_mb_count )
1075 update_predictor( &rcc->pred[rcc->slice_type], qp2qscale(rcc->qpa), rcc->last_satd, bits );
1080 rcc->buffer_fill += rcc->buffer_rate - bits;
1081 if( rcc->buffer_fill < 0 && !rcc->b_2pass )
1082 x264_log( h, X264_LOG_WARNING, "VBV underflow (%.0f bits)\n", rcc->buffer_fill );
1083 rcc->buffer_fill = x264_clip3( rcc->buffer_fill, 0, rcc->buffer_size );
1086 // apply VBV constraints and clip qscale to between lmin and lmax
1087 static double clip_qscale( x264_t *h, int pict_type, double q )
1089 x264_ratecontrol_t *rcc = h->rc;
1090 double lmin = rcc->lmin[pict_type];
1091 double lmax = rcc->lmax[pict_type];
1094 /* B-frames are not directly subject to VBV,
1095 * since they are controlled by the P-frames' QPs.
1096 * FIXME: in 2pass we could modify previous frames' QP too,
1097 * instead of waiting for the buffer to fill */
1099 ( pict_type == SLICE_TYPE_P ||
1100 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) )
1102 if( rcc->buffer_fill/rcc->buffer_size < 0.5 )
1103 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1106 if( rcc->b_vbv && rcc->last_satd > 0 )
1108 /* Now a hard threshold to make sure the frame fits in VBV.
1109 * This one is mostly for I-frames. */
1110 double bits = predict_size( &rcc->pred[rcc->slice_type], q, rcc->last_satd );
1112 if( bits > rcc->buffer_fill/2 )
1113 qf = x264_clip3f( rcc->buffer_fill/(2*bits), 0.2, 1.0 );
1116 if( bits < rcc->buffer_rate/2 )
1117 q *= bits*2/rcc->buffer_rate;
1118 q = X264_MAX( q0, q );
1120 /* Check B-frame complexity, and use up any bits that would
1121 * overflow before the next P-frame. */
1122 if( rcc->slice_type == SLICE_TYPE_P )
1124 int nb = rcc->bframes;
1125 double pbbits = bits;
1126 double bbits = predict_size( &rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1129 if( bbits > rcc->buffer_rate )
1131 pbbits += nb * bbits;
1133 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1134 if( pbbits < space )
1136 q *= X264_MAX( pbbits / space,
1137 bits / (0.5 * rcc->buffer_size) );
1139 q = X264_MAX( q0-5, q );
1142 if( !rcc->b_vbv_min_rate )
1143 q = X264_MAX( q0, q );
1148 else if(rcc->b_2pass)
1150 double min2 = log(lmin);
1151 double max2 = log(lmax);
1152 q = (log(q) - min2)/(max2-min2) - 0.5;
1153 q = 1.0/(1.0 + exp(-4*q));
1154 q = q*(max2-min2) + min2;
1158 return x264_clip3f(q, lmin, lmax);
1161 // update qscale for 1 frame based on actual bits used so far
1162 static float rate_estimate_qscale(x264_t *h, int pict_type)
1165 x264_ratecontrol_t *rcc = h->rc;
1166 ratecontrol_entry_t rce;
1167 double lmin = rcc->lmin[pict_type];
1168 double lmax = rcc->lmax[pict_type];
1169 int64_t total_bits = 8*(h->stat.i_slice_size[SLICE_TYPE_I]
1170 + h->stat.i_slice_size[SLICE_TYPE_P]
1171 + h->stat.i_slice_size[SLICE_TYPE_B]);
1176 if(pict_type != rce.pict_type)
1178 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1179 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1183 if( pict_type == SLICE_TYPE_B )
1185 /* B-frames don't have independent ratecontrol, but rather get the
1186 * average QP of the two adjacent P-frames + an offset */
1188 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1189 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1190 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1191 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1192 float q0 = h->fref0[0]->f_qp_avg;
1193 float q1 = h->fref1[0]->f_qp_avg;
1195 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1196 q0 -= rcc->pb_offset/2;
1197 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1198 q1 -= rcc->pb_offset/2;
1201 q = (q0 + q1) / 2 + rcc->ip_offset;
1207 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1209 if(h->fenc->b_kept_as_ref)
1210 q += rcc->pb_offset/2;
1212 q += rcc->pb_offset;
1215 return qp2qscale(q);
1219 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1222 //FIXME adjust abr_buffer based on distance to the end of the video
1223 int64_t diff = total_bits - (int64_t)rce.expected_bits;
1225 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1226 if( h->fenc->i_frame > 30 )
1228 /* Adjust quant based on the difference between
1229 * achieved and expected bitrate so far */
1230 double time = (double)h->fenc->i_frame / rcc->num_entries;
1231 double w = x264_clip3f( time*100, 0.0, 1.0 );
1232 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1234 q = x264_clip3f( q, lmin, lmax );
1236 else /* 1pass ABR */
1238 /* Calculate the quantizer which would have produced the desired
1239 * average bitrate if it had been applied to all frames so far.
1240 * Then modulate that quant based on the current frame's complexity
1241 * relative to the average complexity so far (using the 2pass RCEQ).
1242 * Then bias the quant up or down if total size so far was far from
1244 * Result: Depending on the value of rate_tolerance, there is a
1245 * tradeoff between quality and bitrate precision. But at large
1246 * tolerances, the bit distribution approaches that of 2pass. */
1248 double wanted_bits, overflow, lmin, lmax;
1250 rcc->last_satd = x264_rc_analyse_slice( h );
1251 rcc->short_term_cplxsum *= 0.5;
1252 rcc->short_term_cplxcount *= 0.5;
1253 rcc->short_term_cplxsum += rcc->last_satd;
1254 rcc->short_term_cplxcount ++;
1256 rce.p_tex_bits = rcc->last_satd;
1257 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1260 rce.p_count = rcc->nmb;
1264 rce.pict_type = pict_type;
1266 if( h->param.rc.i_rc_method == X264_RC_CRF )
1268 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1273 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1275 wanted_bits = h->fenc->i_frame * rcc->bitrate / rcc->fps;
1276 abr_buffer *= X264_MAX( 1, sqrt(h->fenc->i_frame/25) );
1277 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1281 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1282 /* should test _next_ pict type, but that isn't decided yet */
1283 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1285 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1286 q /= fabs( h->param.rc.f_ip_factor );
1288 else if( h->i_frame > 0 )
1290 /* Asymmetric clipping, because symmetric would prevent
1291 * overflow control in areas of rapidly oscillating complexity */
1292 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1293 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1294 if( overflow > 1.1 && h->i_frame > 3 )
1296 else if( overflow < 0.9 )
1299 q = x264_clip3f(q, lmin, lmax);
1301 else if( h->param.rc.i_rc_method == X264_RC_CRF )
1303 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1306 //FIXME use get_diff_limited_q() ?
1307 q = clip_qscale( h, pict_type, q );
1310 rcc->last_qscale_for[pict_type] =
1311 rcc->last_qscale = q;
1313 if( !rcc->b_2pass && h->fenc->i_frame == 0 )
1314 rcc->last_qscale_for[SLICE_TYPE_P] = q;
1316 rcc->frame_size_planned = predict_size( &rcc->pred[rcc->slice_type], q, rcc->last_satd );
1322 /* Distribute bits among the slices, proportional to their estimated complexity */
1323 void x264_ratecontrol_threads_start( x264_t *h )
1325 x264_ratecontrol_t *rc = h->rc;
1328 double frame_size_planned = rc->frame_size_planned;
1330 for( t = 0; t < h->param.i_threads; t++ )
1332 h->thread[t]->rc = &rc[t];
1337 if( !h->mb.b_variable_qp || rc->slice_type == SLICE_TYPE_B )
1340 for( t = 0; t < h->param.i_threads; t++ )
1342 rc[t].first_row = h->thread[t]->sh.i_first_mb / h->sps->i_mb_width;
1343 rc[t].last_row = (h->thread[t]->sh.i_last_mb-1) / h->sps->i_mb_width;
1344 rc[t].frame_size_planned = 1;
1345 rc[t].row_pred = &rc[t].row_preds[rc->slice_type];
1346 if( h->param.i_threads > 1 )
1348 for( y = rc[t].first_row; y<= rc[t].last_row; y++ )
1349 rc[t].frame_size_planned += predict_row_size( h, y, qscale2qp(rc[t].qp) );
1351 den += rc[t].frame_size_planned;
1353 for( t = 0; t < h->param.i_threads; t++ )
1354 rc[t].frame_size_planned *= frame_size_planned / den;
1357 static int init_pass2( x264_t *h )
1359 x264_ratecontrol_t *rcc = h->rc;
1360 uint64_t all_const_bits = 0;
1361 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000 * (double)rcc->num_entries / rcc->fps);
1362 double rate_factor, step, step_mult;
1363 double qblur = h->param.rc.f_qblur;
1364 double cplxblur = h->param.rc.f_complexity_blur;
1365 const int filter_size = (int)(qblur*4) | 1;
1366 double expected_bits;
1367 double *qscale, *blurred_qscale;
1370 /* find total/average complexity & const_bits */
1371 for(i=0; i<rcc->num_entries; i++){
1372 ratecontrol_entry_t *rce = &rcc->entry[i];
1373 all_const_bits += rce->misc_bits;
1374 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
1375 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
1376 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits * rce->qscale;
1377 rcc->frame_count[rce->pict_type] ++;
1380 if( all_available_bits < all_const_bits)
1382 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
1383 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000)));
1387 /* Blur complexities, to reduce local fluctuation of QP.
1388 * We don't blur the QPs directly, because then one very simple frame
1389 * could drag down the QP of a nearby complex frame and give it more
1390 * bits than intended. */
1391 for(i=0; i<rcc->num_entries; i++){
1392 ratecontrol_entry_t *rce = &rcc->entry[i];
1393 double weight_sum = 0;
1394 double cplx_sum = 0;
1395 double weight = 1.0;
1397 /* weighted average of cplx of future frames */
1398 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++){
1399 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
1400 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1403 weight_sum += weight;
1404 cplx_sum += weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1406 /* weighted average of cplx of past frames */
1408 for(j=0; j<=cplxblur*2 && j<=i; j++){
1409 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
1410 weight_sum += weight;
1411 cplx_sum += weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1412 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1416 rce->blurred_complexity = cplx_sum / weight_sum;
1419 qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1421 blurred_qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1423 blurred_qscale = qscale;
1425 /* Search for a factor which, when multiplied by the RCEQ values from
1426 * each frame, adds up to the desired total size.
1427 * There is no exact closed-form solution because of VBV constraints and
1428 * because qscale2bits is not invertible, but we can start with the simple
1429 * approximation of scaling the 1st pass by the ratio of bitrates.
1430 * The search range is probably overkill, but speed doesn't matter here. */
1433 for(i=0; i<rcc->num_entries; i++)
1434 expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i));
1435 step_mult = all_available_bits / expected_bits;
1438 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5){
1440 rate_factor += step;
1442 rcc->last_non_b_pict_type = -1;
1443 rcc->last_accum_p_norm = 1;
1444 rcc->accum_p_norm = 0;
1445 rcc->buffer_fill = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
1448 for(i=0; i<rcc->num_entries; i++){
1449 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
1452 /* fixed I/B qscale relative to P */
1453 for(i=rcc->num_entries-1; i>=0; i--){
1454 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
1455 assert(qscale[i] >= 0);
1459 if(filter_size > 1){
1460 assert(filter_size%2==1);
1461 for(i=0; i<rcc->num_entries; i++){
1462 ratecontrol_entry_t *rce = &rcc->entry[i];
1464 double q=0.0, sum=0.0;
1466 for(j=0; j<filter_size; j++){
1467 int index = i+j-filter_size/2;
1469 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
1470 if(index < 0 || index >= rcc->num_entries) continue;
1471 if(rce->pict_type != rcc->entry[index].pict_type) continue;
1472 q += qscale[index] * coeff;
1475 blurred_qscale[i] = q/sum;
1479 /* find expected bits */
1480 for(i=0; i<rcc->num_entries; i++){
1481 ratecontrol_entry_t *rce = &rcc->entry[i];
1483 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
1484 assert(rce->new_qscale >= 0);
1485 bits = qscale2bits(rce, rce->new_qscale);
1487 rce->expected_bits = expected_bits;
1488 expected_bits += bits;
1489 update_vbv(h, bits);
1492 //printf("expected:%llu available:%llu factor:%lf avgQ:%lf\n", (uint64_t)expected_bits, all_available_bits, rate_factor);
1493 if(expected_bits > all_available_bits) rate_factor -= step;
1498 x264_free(blurred_qscale);
1500 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
1503 for(i=0; i<rcc->num_entries; i++)
1504 avgq += rcc->entry[i].new_qscale;
1505 avgq = qscale2qp(avgq / rcc->num_entries);
1507 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
1508 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
1509 (float)h->param.rc.i_bitrate,
1510 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
1512 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
1514 if(h->param.rc.i_qp_min > 0)
1515 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
1517 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
1519 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
1521 if(h->param.rc.i_qp_max < 51)
1522 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
1524 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
1527 x264_log(h, X264_LOG_WARNING, "internal error\n");