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 */
99 double buffer_fill_final; /* real buffer as of the last finished frame */
100 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
101 double buffer_rate; /* # of bits added to buffer_fill after each frame */
102 predictor_t *pred; /* 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 predictor_t *row_pred;
140 predictor_t row_preds[5];
141 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
142 int bframes; /* # consecutive B-frames before this P-frame */
143 int bframe_bits; /* total cost of those frames */
147 x264_zone_t *prev_zone;
151 static int parse_zones( x264_t *h );
152 static int init_pass2(x264_t *);
153 static float rate_estimate_qscale( x264_t *h );
154 static void update_vbv( x264_t *h, int bits );
155 static void update_vbv_plan( x264_t *h );
156 static double predict_size( predictor_t *p, double q, double var );
157 static void update_predictor( predictor_t *p, double q, double var, double bits );
158 int x264_rc_analyse_slice( x264_t *h );
161 * qp = h.264's quantizer
162 * qscale = linearized quantizer = Lagrange multiplier
164 static inline double qp2qscale(double qp)
166 return 0.85 * pow(2.0, ( qp - 12.0 ) / 6.0);
168 static inline double qscale2qp(double qscale)
170 return 12.0 + 6.0 * log(qscale/0.85) / log(2.0);
173 /* Texture bitrate is not quite inversely proportional to qscale,
174 * probably due the the changing number of SKIP blocks.
175 * MV bits level off at about qp<=12, because the lambda used
176 * for motion estimation is constant there. */
177 static inline double qscale2bits(ratecontrol_entry_t *rce, double qscale)
181 return (rce->i_tex_bits + rce->p_tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
182 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
187 int x264_ratecontrol_new( x264_t *h )
189 x264_ratecontrol_t *rc;
192 x264_cpu_restore( h->param.cpu );
194 rc = h->rc = x264_malloc( h->param.i_threads * sizeof(x264_ratecontrol_t) );
195 memset( rc, 0, h->param.i_threads * sizeof(x264_ratecontrol_t) );
197 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
198 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
200 /* FIXME: use integers */
201 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
202 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
206 rc->bitrate = h->param.rc.i_bitrate * 1000;
207 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
208 rc->nmb = h->mb.i_mb_count;
209 rc->last_non_b_pict_type = -1;
212 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
214 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
217 if( h->param.rc.i_vbv_buffer_size )
219 if( h->param.rc.i_rc_method == X264_RC_CQP )
220 x264_log(h, X264_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
221 else if( h->param.rc.i_vbv_max_bitrate == 0 )
223 x264_log( h, X264_LOG_DEBUG, "VBV maxrate unspecified, assuming CBR\n" );
224 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
227 if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate &&
228 h->param.rc.i_vbv_max_bitrate > 0)
229 x264_log(h, X264_LOG_WARNING, "max bitrate less than average bitrate, ignored.\n");
230 else if( h->param.rc.i_vbv_max_bitrate > 0 &&
231 h->param.rc.i_vbv_buffer_size > 0 )
233 if( h->param.rc.i_vbv_buffer_size < 3 * h->param.rc.i_vbv_max_bitrate / rc->fps )
235 h->param.rc.i_vbv_buffer_size = 3 * h->param.rc.i_vbv_max_bitrate / rc->fps;
236 x264_log( h, X264_LOG_WARNING, "VBV buffer size too small, using %d kbit\n",
237 h->param.rc.i_vbv_buffer_size );
239 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000 / rc->fps;
240 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000;
241 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
242 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
243 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
245 rc->b_vbv_min_rate = !rc->b_2pass
246 && h->param.rc.i_rc_method == X264_RC_ABR
247 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
249 else if( h->param.rc.i_vbv_max_bitrate )
251 x264_log(h, X264_LOG_WARNING, "VBV maxrate specified, but no bufsize.\n");
252 h->param.rc.i_vbv_max_bitrate = 0;
254 if(rc->rate_tolerance < 0.01) {
255 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
256 rc->rate_tolerance = 0.01;
259 h->mb.b_variable_qp = rc->b_vbv && !rc->b_2pass;
263 /* FIXME ABR_INIT_QP is actually used only in CRF */
264 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
265 rc->accum_p_norm = .01;
266 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
267 /* estimated ratio that produces a reasonable QP for the first I-frame */
268 rc->cplxr_sum = .01 * pow( 7.0e5, h->param.rc.f_qcompress ) * pow( h->mb.i_mb_count, 0.5 );
269 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
270 rc->last_non_b_pict_type = SLICE_TYPE_I;
273 if( h->param.rc.i_rc_method == X264_RC_CRF )
275 /* arbitrary rescaling to make CRF somewhat similar to QP */
276 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
277 rc->rate_factor_constant = pow( base_cplx, 1 - h->param.rc.f_qcompress )
278 / qp2qscale( h->param.rc.f_rf_constant );
281 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
282 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
283 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
284 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
285 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
287 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
288 rc->last_qscale = qp2qscale(26);
289 rc->pred = x264_malloc( 5*sizeof(predictor_t) );
290 rc->pred_b_from_p = x264_malloc( sizeof(predictor_t) );
291 for( i = 0; i < 5; i++ )
293 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
294 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
295 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
296 rc->pred[i].coeff= 2.0;
297 rc->pred[i].count= 1.0;
298 rc->pred[i].decay= 0.5;
299 rc->row_preds[i].coeff= .25;
300 rc->row_preds[i].count= 1.0;
301 rc->row_preds[i].decay= 0.5;
303 *rc->pred_b_from_p = rc->pred[0];
305 if( parse_zones( h ) < 0 )
307 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
311 /* Load stat file and init 2pass algo */
312 if( h->param.rc.b_stat_read )
314 char *p, *stats_in, *stats_buf;
316 /* read 1st pass stats */
317 assert( h->param.rc.psz_stat_in );
318 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
321 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
325 /* check whether 1st pass options were compatible with current options */
326 if( !strncmp( stats_buf, "#options:", 9 ) )
329 char *opts = stats_buf;
330 stats_in = strchr( stats_buf, '\n' );
336 if( ( p = strstr( opts, "bframes=" ) ) && sscanf( p, "bframes=%d", &i )
337 && h->param.i_bframe != i )
339 x264_log( h, X264_LOG_ERROR, "different number of B-frames than 1st pass (%d vs %d)\n",
340 h->param.i_bframe, i );
344 /* since B-adapt doesn't (yet) take into account B-pyramid,
345 * the converse is not a problem */
346 if( strstr( opts, "b_pyramid=1" ) && !h->param.b_bframe_pyramid )
347 x264_log( h, X264_LOG_WARNING, "1st pass used B-pyramid, 2nd doesn't\n" );
349 if( ( p = strstr( opts, "keyint=" ) ) && sscanf( p, "keyint=%d", &i )
350 && h->param.i_keyint_max != i )
351 x264_log( h, X264_LOG_WARNING, "different keyint than 1st pass (%d vs %d)\n",
352 h->param.i_keyint_max, i );
354 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
355 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
358 /* find number of pics */
361 p = strchr(p+1, ';');
364 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
369 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
371 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
372 h->param.i_frame_total, rc->num_entries );
374 if( h->param.i_frame_total > rc->num_entries + h->param.i_bframe )
376 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
377 h->param.i_frame_total, rc->num_entries );
381 /* FIXME: ugly padding because VfW drops delayed B-frames */
382 rc->num_entries += h->param.i_bframe;
384 rc->entry = (ratecontrol_entry_t*) x264_malloc(rc->num_entries * sizeof(ratecontrol_entry_t));
385 memset(rc->entry, 0, rc->num_entries * sizeof(ratecontrol_entry_t));
387 /* init all to skipped p frames */
388 for(i=0; i<rc->num_entries; i++){
389 ratecontrol_entry_t *rce = &rc->entry[i];
390 rce->pict_type = SLICE_TYPE_P;
391 rce->qscale = rce->new_qscale = qp2qscale(20);
392 rce->misc_bits = rc->nmb + 10;
398 for(i=0; i < rc->num_entries - h->param.i_bframe; i++){
399 ratecontrol_entry_t *rce;
406 next= strchr(p, ';');
408 (*next)=0; //sscanf is unbelievably slow on looong strings
411 e = sscanf(p, " in:%d ", &frame_number);
413 if(frame_number < 0 || frame_number >= rc->num_entries)
415 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
418 rce = &rc->entry[frame_number];
419 rce->direct_mode = 0;
421 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",
422 &pict_type, &qp, &rce->i_tex_bits, &rce->p_tex_bits,
423 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
424 &rce->s_count, &rce->direct_mode);
427 case 'I': rce->kept_as_ref = 1;
428 case 'i': rce->pict_type = SLICE_TYPE_I; break;
429 case 'P': rce->pict_type = SLICE_TYPE_P; break;
430 case 'B': rce->kept_as_ref = 1;
431 case 'b': rce->pict_type = SLICE_TYPE_B; break;
432 default: e = -1; break;
435 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
438 rce->qscale = qp2qscale(qp);
442 x264_free(stats_buf);
444 if(h->param.rc.i_rc_method == X264_RC_ABR)
446 if(init_pass2(h) < 0) return -1;
447 } /* else we're using constant quant, so no need to run the bitrate allocation */
450 /* Open output file */
451 /* If input and output files are the same, output to a temp file
452 * and move it to the real name only when it's complete */
453 if( h->param.rc.b_stat_write )
457 rc->psz_stat_file_tmpname = x264_malloc( strlen(h->param.rc.psz_stat_out) + 6 );
458 strcpy( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
459 strcat( rc->psz_stat_file_tmpname, ".temp" );
461 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
462 if( rc->p_stat_file_out == NULL )
464 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
468 p = x264_param2string( &h->param, 1 );
469 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
473 for( i=1; i<h->param.i_threads; i++ )
475 h->thread[i]->rc = rc+i;
482 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
487 z->f_bitrate_factor = 1;
488 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
490 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
492 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
496 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
502 z->param = malloc( sizeof(x264_param_t) );
503 memcpy( z->param, &h->param, sizeof(x264_param_t) );
504 while( (tok = strtok_r( p, ",", &saveptr )) )
506 char *val = strchr( tok, '=' );
512 if( x264_param_parse( z->param, tok, val ) )
514 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
522 static int parse_zones( x264_t *h )
524 x264_ratecontrol_t *rc = h->rc;
526 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
528 char *p, *tok, *saveptr;
529 char *psz_zones = x264_malloc( strlen(h->param.rc.psz_zones)+1 );
530 strcpy( psz_zones, h->param.rc.psz_zones );
531 h->param.rc.i_zones = 1;
532 for( p = psz_zones; *p; p++ )
533 h->param.rc.i_zones += (*p == '/');
534 h->param.rc.zones = x264_malloc( h->param.rc.i_zones * sizeof(x264_zone_t) );
536 for( i = 0; i < h->param.rc.i_zones; i++ )
538 tok = strtok_r( p, "/", &saveptr );
539 if( !tok || parse_zone( h, &h->param.rc.zones[i], tok ) )
543 x264_free( psz_zones );
546 if( h->param.rc.i_zones > 0 )
548 for( i = 0; i < h->param.rc.i_zones; i++ )
550 x264_zone_t z = h->param.rc.zones[i];
551 if( z.i_start < 0 || z.i_start > z.i_end )
553 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
554 z.i_start, z.i_end );
557 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
559 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
560 z.f_bitrate_factor );
565 rc->i_zones = h->param.rc.i_zones + 1;
566 rc->zones = x264_malloc( rc->i_zones * sizeof(x264_zone_t) );
567 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
569 // default zone to fall back to if none of the others match
570 rc->zones[0].i_start = 0;
571 rc->zones[0].i_end = INT_MAX;
572 rc->zones[0].b_force_qp = 0;
573 rc->zones[0].f_bitrate_factor = 1;
574 rc->zones[0].param = x264_malloc( sizeof(x264_param_t) );
575 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
576 for( i = 1; i < rc->i_zones; i++ )
578 if( !rc->zones[i].param )
579 rc->zones[i].param = rc->zones[0].param;
586 x264_zone_t *get_zone( x264_t *h, int frame_num )
589 for( i = h->rc->i_zones-1; i >= 0; i-- )
591 x264_zone_t *z = &h->rc->zones[i];
592 if( frame_num >= z->i_start && frame_num <= z->i_end )
598 void x264_ratecontrol_summary( x264_t *h )
600 x264_ratecontrol_t *rc = h->rc;
601 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
603 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
604 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
605 qscale2qp( pow( base_cplx, 1 - h->param.rc.f_qcompress )
606 * rc->cplxr_sum / rc->wanted_bits_window ) );
610 void x264_ratecontrol_delete( x264_t *h )
612 x264_ratecontrol_t *rc = h->rc;
615 if( rc->p_stat_file_out )
617 fclose( rc->p_stat_file_out );
618 if( h->i_frame >= rc->num_entries - h->param.i_bframe )
619 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
621 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
622 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
624 x264_free( rc->psz_stat_file_tmpname );
626 x264_free( rc->pred );
627 x264_free( rc->pred_b_from_p );
628 x264_free( rc->entry );
631 x264_free( rc->zones[0].param );
632 if( h->param.rc.psz_zones )
633 for( i=1; i<rc->i_zones; i++ )
634 if( rc->zones[i].param != rc->zones[0].param )
635 x264_free( rc->zones[i].param );
636 x264_free( rc->zones );
641 static void accum_p_qp_update( x264_t *h, float qp )
643 x264_ratecontrol_t *rc = h->rc;
644 rc->accum_p_qp *= .95;
645 rc->accum_p_norm *= .95;
646 rc->accum_p_norm += 1;
647 if( h->sh.i_type == SLICE_TYPE_I )
648 rc->accum_p_qp += qp + rc->ip_offset;
650 rc->accum_p_qp += qp;
653 /* Before encoding a frame, choose a QP for it */
654 void x264_ratecontrol_start( x264_t *h, int i_force_qp )
656 x264_ratecontrol_t *rc = h->rc;
657 ratecontrol_entry_t *rce = NULL;
658 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
661 x264_cpu_restore( h->param.cpu );
663 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
664 x264_encoder_reconfig( h, zone->param );
665 rc->prev_zone = zone;
667 rc->qp_force = i_force_qp;
669 if( h->param.rc.b_stat_read )
671 int frame = h->fenc->i_frame;
672 assert( frame >= 0 && frame < rc->num_entries );
673 rce = h->rc->rce = &h->rc->entry[frame];
675 if( h->sh.i_type == SLICE_TYPE_B
676 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
678 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
679 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
685 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
686 rc->row_pred = &rc->row_preds[h->sh.i_type];
687 update_vbv_plan( h );
690 if( h->sh.i_type != SLICE_TYPE_B )
693 while( h->frames.current[rc->bframes] && IS_X264_TYPE_B(h->frames.current[rc->bframes]->i_type) )
705 q = qscale2qp( rate_estimate_qscale( h ) );
707 else if( rc->b_2pass )
709 rce->new_qscale = rate_estimate_qscale( h );
710 q = qscale2qp( rce->new_qscale );
714 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
715 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
717 q = rc->qp_constant[ h->sh.i_type ];
721 if( zone->b_force_qp )
722 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
724 q -= 6*log(zone->f_bitrate_factor)/log(2);
730 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
732 rce->new_qp = rc->qp;
734 /* accum_p_qp needs to be here so that future frames can benefit from the
735 * data before this frame is done. but this only works because threading
736 * guarantees to not re-encode any frames. so the non-threaded case does
737 * accum_p_qp later. */
738 if( h->param.i_threads > 1 )
739 accum_p_qp_update( h, rc->qp );
741 if( h->sh.i_type != SLICE_TYPE_B )
742 rc->last_non_b_pict_type = h->sh.i_type;
745 double predict_row_size( x264_t *h, int y, int qp )
747 /* average between two predictors:
748 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
749 x264_ratecontrol_t *rc = h->rc;
750 double pred_s = predict_size( rc->row_pred, qp2qscale(qp), h->fdec->i_row_satd[y] );
752 if( h->sh.i_type != SLICE_TYPE_I
753 && h->fref0[0]->i_type == h->fdec->i_type
754 && h->fref0[0]->i_row_satd[y] > 0 )
756 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
757 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
762 return (pred_s + pred_t) / 2;
765 double predict_row_size_sum( x264_t *h, int y, int qp )
769 for( i = 0; i <= y; i++ )
770 bits += h->fdec->i_row_bits[i];
771 for( i = y+1; i < h->sps->i_mb_height; i++ )
772 bits += predict_row_size( h, i, qp );
776 void x264_ratecontrol_mb( x264_t *h, int bits )
778 x264_ratecontrol_t *rc = h->rc;
779 const int y = h->mb.i_mb_y;
781 x264_cpu_restore( h->param.cpu );
783 h->fdec->i_row_bits[y] += bits;
786 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !h->mb.b_variable_qp )
789 h->fdec->i_row_qp[y] = rc->qpm;
791 if( h->sh.i_type == SLICE_TYPE_B )
793 /* B-frames shouldn't use lower QP than their reference frames */
794 if( y < h->sps->i_mb_height-1 )
796 rc->qpm = X264_MAX( rc->qp,
797 X264_MIN( h->fref0[0]->i_row_qp[y+1],
798 h->fref1[0]->i_row_qp[y+1] ));
803 update_predictor( rc->row_pred, qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
805 /* tweak quality based on difference from predicted size */
806 if( y < h->sps->i_mb_height-1 && h->stat.i_slice_count[h->sh.i_type] > 0 )
808 int prev_row_qp = h->fdec->i_row_qp[y];
809 int b0 = predict_row_size_sum( h, y, rc->qpm );
811 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
812 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
813 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
815 if( !rc->b_vbv_min_rate )
816 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
818 while( rc->qpm < i_qp_max
819 && (b1 > rc->frame_size_planned * 1.15
820 || (rc->buffer_fill - b1 < buffer_left_planned * 0.5)))
823 b1 = predict_row_size_sum( h, y, rc->qpm );
826 while( rc->qpm > i_qp_min
827 && buffer_left_planned > rc->buffer_size * 0.4
828 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
829 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
832 b1 = predict_row_size_sum( h, y, rc->qpm );
838 int x264_ratecontrol_qp( x264_t *h )
843 /* In 2pass, force the same frame types as in the 1st pass */
844 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
846 x264_ratecontrol_t *rc = h->rc;
847 if( h->param.rc.b_stat_read )
849 if( frame_num >= rc->num_entries )
851 /* We could try to initialize everything required for ABR and
852 * adaptive B-frames, but that would be complicated.
853 * So just calculate the average QP used so far. */
855 h->param.rc.i_qp_constant = (h->stat.i_slice_count[SLICE_TYPE_P] == 0) ? 24
856 : 1 + h->stat.i_slice_qp[SLICE_TYPE_P] / h->stat.i_slice_count[SLICE_TYPE_P];
857 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
858 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 );
859 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 );
861 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
862 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
863 if( h->param.b_bframe_adaptive )
864 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
868 h->param.rc.i_rc_method = X264_RC_CQP;
869 h->param.rc.b_stat_read = 0;
870 h->param.b_bframe_adaptive = 0;
871 if( h->param.i_bframe > 1 )
872 h->param.i_bframe = 1;
875 switch( rc->entry[frame_num].pict_type )
878 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I;
881 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B;
890 return X264_TYPE_AUTO;
894 /* After encoding one frame, save stats and update ratecontrol state */
895 void x264_ratecontrol_end( x264_t *h, int bits )
897 x264_ratecontrol_t *rc = h->rc;
898 const int *mbs = h->stat.frame.i_mb_count;
901 x264_cpu_restore( h->param.cpu );
903 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
904 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
905 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
906 for( i = B_DIRECT; i < B_8x8; i++ )
907 h->stat.frame.i_mb_count_p += mbs[i];
909 if( h->mb.b_variable_qp )
910 rc->qpa /= h->mb.i_mb_count;
913 h->fdec->f_qp_avg = rc->qpa;
915 if( h->param.rc.b_stat_write )
917 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
918 : h->sh.i_type==SLICE_TYPE_P ? 'P'
919 : h->fenc->b_kept_as_ref ? 'B' : 'b';
920 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
921 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
922 char c_direct = h->mb.b_direct_auto_write ?
923 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
924 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
926 fprintf( rc->p_stat_file_out,
927 "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",
928 h->fenc->i_frame, h->i_frame,
930 h->stat.frame.i_itex_bits, h->stat.frame.i_ptex_bits,
931 h->stat.frame.i_hdr_bits, h->stat.frame.i_misc_bits,
932 h->stat.frame.i_mb_count_i,
933 h->stat.frame.i_mb_count_p,
934 h->stat.frame.i_mb_count_skip,
940 if( h->sh.i_type != SLICE_TYPE_B )
941 rc->cplxr_sum += bits * qp2qscale(rc->qpa) / rc->last_rceq;
944 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
945 * Not perfectly accurate with B-refs, but good enough. */
946 rc->cplxr_sum += bits * qp2qscale(rc->qpa) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
948 rc->cplxr_sum *= rc->cbr_decay;
949 rc->wanted_bits_window += rc->bitrate / rc->fps;
950 rc->wanted_bits_window *= rc->cbr_decay;
952 if( h->param.i_threads == 1 )
953 accum_p_qp_update( h, rc->qpa );
958 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
961 if( h->mb.b_variable_qp )
963 if( h->sh.i_type == SLICE_TYPE_B )
965 rc->bframe_bits += bits;
966 if( !h->frames.current[0] || !IS_X264_TYPE_B(h->frames.current[0]->i_type) )
968 update_predictor( rc->pred_b_from_p, qp2qscale(rc->qpa),
969 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
975 update_vbv( h, bits );
978 /****************************************************************************
980 ***************************************************************************/
982 double x264_eval( char *s, double *const_value, const char **const_name,
983 double (**func1)(void *, double), const char **func1_name,
984 double (**func2)(void *, double, double), char **func2_name,
988 * modify the bitrate curve from pass1 for one frame
990 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
992 x264_ratecontrol_t *rcc= h->rc;
993 const int pict_type = rce->pict_type;
995 x264_zone_t *zone = get_zone( h, frame_num );
997 double const_values[]={
998 rce->i_tex_bits * rce->qscale,
999 rce->p_tex_bits * rce->qscale,
1000 (rce->i_tex_bits + rce->p_tex_bits) * rce->qscale,
1001 rce->mv_bits * rce->qscale,
1002 (double)rce->i_count / rcc->nmb,
1003 (double)rce->p_count / rcc->nmb,
1004 (double)rce->s_count / rcc->nmb,
1005 rce->pict_type == SLICE_TYPE_I,
1006 rce->pict_type == SLICE_TYPE_P,
1007 rce->pict_type == SLICE_TYPE_B,
1008 h->param.rc.f_qcompress,
1009 rcc->i_cplx_sum[SLICE_TYPE_I] / rcc->frame_count[SLICE_TYPE_I],
1010 rcc->i_cplx_sum[SLICE_TYPE_P] / rcc->frame_count[SLICE_TYPE_P],
1011 rcc->p_cplx_sum[SLICE_TYPE_P] / rcc->frame_count[SLICE_TYPE_P],
1012 rcc->p_cplx_sum[SLICE_TYPE_B] / rcc->frame_count[SLICE_TYPE_B],
1013 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / rcc->frame_count[pict_type],
1014 rce->blurred_complexity,
1017 static const char *const_names[]={
1037 static double (*func1[])(void *, double)={
1038 // (void *)bits2qscale,
1039 (void *)qscale2bits,
1042 static const char *func1_names[]={
1048 q = x264_eval((char*)h->param.rc.psz_rc_eq, const_values, const_names, func1, func1_names, NULL, NULL, rce);
1050 // avoid NaN's in the rc_eq
1051 if(!isfinite(q) || rce->i_tex_bits + rce->p_tex_bits + rce->mv_bits == 0)
1052 q = rcc->last_qscale;
1056 rcc->last_qscale = q;
1061 if( zone->b_force_qp )
1062 q = qp2qscale(zone->i_qp);
1064 q /= zone->f_bitrate_factor;
1070 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1072 x264_ratecontrol_t *rcc = h->rc;
1073 const int pict_type = rce->pict_type;
1075 // force I/B quants as a function of P quants
1076 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1077 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1078 if( pict_type == SLICE_TYPE_I )
1081 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1082 double ip_factor = fabs( h->param.rc.f_ip_factor );
1083 /* don't apply ip_factor if the following frame is also I */
1084 if( rcc->accum_p_norm <= 0 )
1086 else if( h->param.rc.f_ip_factor < 0 )
1088 else if( rcc->accum_p_norm >= 1 )
1091 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1093 else if( pict_type == SLICE_TYPE_B )
1095 if( h->param.rc.f_pb_factor > 0 )
1097 if( !rce->kept_as_ref )
1098 q *= fabs( h->param.rc.f_pb_factor );
1100 else if( pict_type == SLICE_TYPE_P
1101 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1102 && rce->i_tex_bits + rce->p_tex_bits == 0 )
1107 /* last qscale / qdiff stuff */
1108 if(rcc->last_non_b_pict_type==pict_type
1109 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1111 double last_q = rcc->last_qscale_for[pict_type];
1112 double max_qscale = last_q * rcc->lstep;
1113 double min_qscale = last_q / rcc->lstep;
1115 if (q > max_qscale) q = max_qscale;
1116 else if(q < min_qscale) q = min_qscale;
1119 rcc->last_qscale_for[pict_type] = q;
1120 if(pict_type!=SLICE_TYPE_B)
1121 rcc->last_non_b_pict_type = pict_type;
1122 if(pict_type==SLICE_TYPE_I)
1124 rcc->last_accum_p_norm = rcc->accum_p_norm;
1125 rcc->accum_p_norm = 0;
1126 rcc->accum_p_qp = 0;
1128 if(pict_type==SLICE_TYPE_P)
1130 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1131 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1132 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1137 static double predict_size( predictor_t *p, double q, double var )
1139 return p->coeff*var / (q*p->count);
1142 static void update_predictor( predictor_t *p, double q, double var, double bits )
1146 p->count *= p->decay;
1147 p->coeff *= p->decay;
1149 p->coeff += bits*q / var;
1152 // update VBV after encoding a frame
1153 static void update_vbv( x264_t *h, int bits )
1155 x264_ratecontrol_t *rcc = h->rc;
1156 x264_ratecontrol_t *rct = h->thread[0]->rc;
1158 if( rcc->last_satd >= h->mb.i_mb_count )
1159 update_predictor( &rct->pred[h->sh.i_type], qp2qscale(rcc->qpa), rcc->last_satd, bits );
1164 rct->buffer_fill_final += rct->buffer_rate - bits;
1165 if( rct->buffer_fill_final < 0 && !rct->b_2pass )
1166 x264_log( h, X264_LOG_WARNING, "VBV underflow (%.0f bits)\n", rct->buffer_fill_final );
1167 rct->buffer_fill_final = x264_clip3( rct->buffer_fill_final, 0, rct->buffer_size );
1170 // provisionally update VBV according to the planned size of all frames currently in progress
1171 static void update_vbv_plan( x264_t *h )
1173 x264_ratecontrol_t *rcc = h->rc;
1174 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final;
1175 if( h->param.i_threads > 1 )
1177 int j = h->rc - h->thread[0]->rc;
1179 for( i=1; i<h->param.i_threads; i++ )
1181 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1182 double bits = t->rc->frame_size_planned;
1183 if( !t->b_thread_active )
1185 rcc->buffer_fill += rcc->buffer_rate - bits;
1186 rcc->buffer_fill = x264_clip3( rcc->buffer_fill, 0, rcc->buffer_size );
1191 // apply VBV constraints and clip qscale to between lmin and lmax
1192 static double clip_qscale( x264_t *h, int pict_type, double q )
1194 x264_ratecontrol_t *rcc = h->rc;
1195 double lmin = rcc->lmin[pict_type];
1196 double lmax = rcc->lmax[pict_type];
1199 /* B-frames are not directly subject to VBV,
1200 * since they are controlled by the P-frames' QPs.
1201 * FIXME: in 2pass we could modify previous frames' QP too,
1202 * instead of waiting for the buffer to fill */
1204 ( pict_type == SLICE_TYPE_P ||
1205 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) )
1207 if( rcc->buffer_fill/rcc->buffer_size < 0.5 )
1208 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1211 if( rcc->b_vbv && rcc->last_satd > 0 )
1213 /* Now a hard threshold to make sure the frame fits in VBV.
1214 * This one is mostly for I-frames. */
1215 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1217 if( bits > rcc->buffer_fill/2 )
1218 qf = x264_clip3f( rcc->buffer_fill/(2*bits), 0.2, 1.0 );
1221 if( bits < rcc->buffer_rate/2 )
1222 q *= bits*2/rcc->buffer_rate;
1223 q = X264_MAX( q0, q );
1225 /* Check B-frame complexity, and use up any bits that would
1226 * overflow before the next P-frame. */
1227 if( h->sh.i_type == SLICE_TYPE_P )
1229 int nb = rcc->bframes;
1230 double pbbits = bits;
1231 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1234 if( bbits > rcc->buffer_rate )
1236 pbbits += nb * bbits;
1238 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1239 if( pbbits < space )
1241 q *= X264_MAX( pbbits / space,
1242 bits / (0.5 * rcc->buffer_size) );
1244 q = X264_MAX( q0-5, q );
1247 if( !rcc->b_vbv_min_rate )
1248 q = X264_MAX( q0, q );
1253 else if(rcc->b_2pass)
1255 double min2 = log(lmin);
1256 double max2 = log(lmax);
1257 q = (log(q) - min2)/(max2-min2) - 0.5;
1258 q = 1.0/(1.0 + exp(-4*q));
1259 q = q*(max2-min2) + min2;
1263 return x264_clip3f(q, lmin, lmax);
1266 // update qscale for 1 frame based on actual bits used so far
1267 static float rate_estimate_qscale( x264_t *h )
1270 x264_ratecontrol_t *rcc = h->rc;
1271 ratecontrol_entry_t rce;
1272 int pict_type = h->sh.i_type;
1273 double lmin = rcc->lmin[pict_type];
1274 double lmax = rcc->lmax[pict_type];
1275 int64_t total_bits = 8*(h->stat.i_slice_size[SLICE_TYPE_I]
1276 + h->stat.i_slice_size[SLICE_TYPE_P]
1277 + h->stat.i_slice_size[SLICE_TYPE_B]);
1282 if(pict_type != rce.pict_type)
1284 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1285 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1289 if( pict_type == SLICE_TYPE_B )
1291 /* B-frames don't have independent ratecontrol, but rather get the
1292 * average QP of the two adjacent P-frames + an offset */
1294 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1295 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1296 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1297 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1298 float q0 = h->fref0[0]->f_qp_avg;
1299 float q1 = h->fref1[0]->f_qp_avg;
1301 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1302 q0 -= rcc->pb_offset/2;
1303 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1304 q1 -= rcc->pb_offset/2;
1307 q = (q0 + q1) / 2 + rcc->ip_offset;
1313 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1315 if(h->fenc->b_kept_as_ref)
1316 q += rcc->pb_offset/2;
1318 q += rcc->pb_offset;
1320 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1322 return qp2qscale(q);
1326 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1329 //FIXME adjust abr_buffer based on distance to the end of the video
1330 int64_t diff = total_bits - (int64_t)rce.expected_bits;
1332 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1333 if( h->fenc->i_frame > 30 )
1335 /* Adjust quant based on the difference between
1336 * achieved and expected bitrate so far */
1337 double time = (double)h->fenc->i_frame / rcc->num_entries;
1338 double w = x264_clip3f( time*100, 0.0, 1.0 );
1339 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1341 q = x264_clip3f( q, lmin, lmax );
1343 else /* 1pass ABR */
1345 /* Calculate the quantizer which would have produced the desired
1346 * average bitrate if it had been applied to all frames so far.
1347 * Then modulate that quant based on the current frame's complexity
1348 * relative to the average complexity so far (using the 2pass RCEQ).
1349 * Then bias the quant up or down if total size so far was far from
1351 * Result: Depending on the value of rate_tolerance, there is a
1352 * tradeoff between quality and bitrate precision. But at large
1353 * tolerances, the bit distribution approaches that of 2pass. */
1355 double wanted_bits, overflow=1, lmin, lmax;
1357 rcc->last_satd = x264_rc_analyse_slice( h );
1358 rcc->short_term_cplxsum *= 0.5;
1359 rcc->short_term_cplxcount *= 0.5;
1360 rcc->short_term_cplxsum += rcc->last_satd;
1361 rcc->short_term_cplxcount ++;
1363 rce.p_tex_bits = rcc->last_satd;
1364 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1367 rce.p_count = rcc->nmb;
1371 rce.pict_type = pict_type;
1373 if( h->param.rc.i_rc_method == X264_RC_CRF )
1375 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1379 int i_frame_done = h->fenc->i_frame + 1 - h->param.i_threads;
1381 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1383 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
1384 wanted_bits = i_frame_done * rcc->bitrate / rcc->fps;
1385 if( wanted_bits > 0 )
1387 abr_buffer *= X264_MAX( 1, sqrt(i_frame_done/25) );
1388 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1393 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1394 /* should test _next_ pict type, but that isn't decided yet */
1395 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1397 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1398 q /= fabs( h->param.rc.f_ip_factor );
1400 else if( h->i_frame > 0 )
1402 /* Asymmetric clipping, because symmetric would prevent
1403 * overflow control in areas of rapidly oscillating complexity */
1404 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1405 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1406 if( overflow > 1.1 && h->i_frame > 3 )
1408 else if( overflow < 0.9 )
1411 q = x264_clip3f(q, lmin, lmax);
1413 else if( h->param.rc.i_rc_method == X264_RC_CRF )
1415 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1418 //FIXME use get_diff_limited_q() ?
1419 q = clip_qscale( h, pict_type, q );
1422 rcc->last_qscale_for[pict_type] =
1423 rcc->last_qscale = q;
1425 if( !rcc->b_2pass && h->fenc->i_frame == 0 )
1426 rcc->last_qscale_for[SLICE_TYPE_P] = q;
1428 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1433 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
1437 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
1438 /* these vars are updated in x264_ratecontrol_start()
1439 * so copy them from the context that most recently started (prev)
1440 * to the context that's about to start (cur).
1446 COPY(last_qscale_for);
1447 COPY(last_non_b_pict_type);
1448 COPY(short_term_cplxsum);
1449 COPY(short_term_cplxcount);
1456 #define COPY(var) next->rc->var = cur->rc->var
1457 /* these vars are updated in x264_ratecontrol_end()
1458 * so copy them from the context that most recently ended (cur)
1459 * to the context that's about to end (next)
1462 COPY(expected_bits_sum);
1463 COPY(wanted_bits_window);
1467 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
1468 /* the rest of the variables are either constant or thread-local */
1471 static int init_pass2( x264_t *h )
1473 x264_ratecontrol_t *rcc = h->rc;
1474 uint64_t all_const_bits = 0;
1475 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000 * (double)rcc->num_entries / rcc->fps);
1476 double rate_factor, step, step_mult;
1477 double qblur = h->param.rc.f_qblur;
1478 double cplxblur = h->param.rc.f_complexity_blur;
1479 const int filter_size = (int)(qblur*4) | 1;
1480 double expected_bits;
1481 double *qscale, *blurred_qscale;
1484 /* find total/average complexity & const_bits */
1485 for(i=0; i<rcc->num_entries; i++){
1486 ratecontrol_entry_t *rce = &rcc->entry[i];
1487 all_const_bits += rce->misc_bits;
1488 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
1489 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
1490 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits * rce->qscale;
1491 rcc->frame_count[rce->pict_type] ++;
1494 if( all_available_bits < all_const_bits)
1496 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
1497 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000)));
1501 /* Blur complexities, to reduce local fluctuation of QP.
1502 * We don't blur the QPs directly, because then one very simple frame
1503 * could drag down the QP of a nearby complex frame and give it more
1504 * bits than intended. */
1505 for(i=0; i<rcc->num_entries; i++){
1506 ratecontrol_entry_t *rce = &rcc->entry[i];
1507 double weight_sum = 0;
1508 double cplx_sum = 0;
1509 double weight = 1.0;
1511 /* weighted average of cplx of future frames */
1512 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++){
1513 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
1514 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1517 weight_sum += weight;
1518 cplx_sum += weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1520 /* weighted average of cplx of past frames */
1522 for(j=0; j<=cplxblur*2 && j<=i; j++){
1523 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
1524 weight_sum += weight;
1525 cplx_sum += weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1526 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1530 rce->blurred_complexity = cplx_sum / weight_sum;
1533 qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1535 blurred_qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1537 blurred_qscale = qscale;
1539 /* Search for a factor which, when multiplied by the RCEQ values from
1540 * each frame, adds up to the desired total size.
1541 * There is no exact closed-form solution because of VBV constraints and
1542 * because qscale2bits is not invertible, but we can start with the simple
1543 * approximation of scaling the 1st pass by the ratio of bitrates.
1544 * The search range is probably overkill, but speed doesn't matter here. */
1547 for(i=0; i<rcc->num_entries; i++)
1548 expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i));
1549 step_mult = all_available_bits / expected_bits;
1552 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5){
1554 rate_factor += step;
1556 rcc->last_non_b_pict_type = -1;
1557 rcc->last_accum_p_norm = 1;
1558 rcc->accum_p_norm = 0;
1559 rcc->buffer_fill = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
1562 for(i=0; i<rcc->num_entries; i++){
1563 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
1566 /* fixed I/B qscale relative to P */
1567 for(i=rcc->num_entries-1; i>=0; i--){
1568 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
1569 assert(qscale[i] >= 0);
1573 if(filter_size > 1){
1574 assert(filter_size%2==1);
1575 for(i=0; i<rcc->num_entries; i++){
1576 ratecontrol_entry_t *rce = &rcc->entry[i];
1578 double q=0.0, sum=0.0;
1580 for(j=0; j<filter_size; j++){
1581 int index = i+j-filter_size/2;
1583 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
1584 if(index < 0 || index >= rcc->num_entries) continue;
1585 if(rce->pict_type != rcc->entry[index].pict_type) continue;
1586 q += qscale[index] * coeff;
1589 blurred_qscale[i] = q/sum;
1593 /* find expected bits */
1594 for(i=0; i<rcc->num_entries; i++){
1595 ratecontrol_entry_t *rce = &rcc->entry[i];
1597 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
1598 assert(rce->new_qscale >= 0);
1599 bits = qscale2bits(rce, rce->new_qscale);
1601 rce->expected_bits = expected_bits;
1602 expected_bits += bits;
1603 update_vbv(h, bits);
1604 rcc->buffer_fill = rcc->buffer_fill_final;
1607 //printf("expected:%llu available:%llu factor:%lf avgQ:%lf\n", (uint64_t)expected_bits, all_available_bits, rate_factor);
1608 if(expected_bits > all_available_bits) rate_factor -= step;
1613 x264_free(blurred_qscale);
1615 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
1618 for(i=0; i<rcc->num_entries; i++)
1619 avgq += rcc->entry[i].new_qscale;
1620 avgq = qscale2qp(avgq / rcc->num_entries);
1622 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
1623 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
1624 (float)h->param.rc.i_bitrate,
1625 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
1627 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
1629 if(h->param.rc.i_qp_min > 0)
1630 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
1632 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
1634 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
1636 if(h->param.rc.i_qp_max < 51)
1637 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
1639 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
1642 x264_log(h, X264_LOG_WARNING, "internal error\n");