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
32 #include "common/common.h"
33 #include "common/cpu.h"
34 #include "ratecontrol.h"
45 uint64_t expected_bits;
51 float blurred_complexity;
53 } ratecontrol_entry_t;
62 struct x264_ratecontrol_t
71 double rate_tolerance;
72 int nmb; /* number of macroblocks in a frame */
76 ratecontrol_entry_t *rce;
77 int qp; /* qp for current frame */
78 int qpm; /* qp for current macroblock */
79 float qpa_rc; /* average of macroblocks' qp before aq */
80 float qpa_aq; /* average of macroblocks' qp after aq */
85 double buffer_fill_final; /* real buffer as of the last finished frame */
86 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
87 double buffer_rate; /* # of bits added to buffer_fill after each frame */
88 predictor_t *pred; /* predict frame size from satd */
93 double cplxr_sum; /* sum of bits*qscale/rceq */
94 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow */
95 double wanted_bits_window; /* target bitrate * window */
97 double short_term_cplxsum;
98 double short_term_cplxcount;
99 double rate_factor_constant;
104 FILE *p_stat_file_out;
105 char *psz_stat_file_tmpname;
107 int num_entries; /* number of ratecontrol_entry_ts */
108 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
110 double last_qscale_for[5]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
111 int last_non_b_pict_type;
112 double accum_p_qp; /* for determining I-frame quant */
114 double last_accum_p_norm;
115 double lmin[5]; /* min qscale by frame type */
117 double lstep; /* max change (multiply) in qscale per frame */
118 double i_cplx_sum[5]; /* estimated total texture bits in intra MBs at qscale=1 */
119 double p_cplx_sum[5];
120 double mv_bits_sum[5];
121 int frame_count[5]; /* number of frames of each type */
124 double frame_size_planned;
125 predictor_t *row_pred;
126 predictor_t row_preds[5];
127 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
128 int bframes; /* # consecutive B-frames before this P-frame */
129 int bframe_bits; /* total cost of those frames */
133 x264_zone_t *prev_zone;
137 static int parse_zones( x264_t *h );
138 static int init_pass2(x264_t *);
139 static float rate_estimate_qscale( x264_t *h );
140 static void update_vbv( x264_t *h, int bits );
141 static void update_vbv_plan( x264_t *h );
142 static double predict_size( predictor_t *p, double q, double var );
143 static void update_predictor( predictor_t *p, double q, double var, double bits );
144 int x264_rc_analyse_slice( x264_t *h );
147 * qp = h.264's quantizer
148 * qscale = linearized quantizer = Lagrange multiplier
150 static inline double qp2qscale(double qp)
152 return 0.85 * pow(2.0, ( qp - 12.0 ) / 6.0);
154 static inline double qscale2qp(double qscale)
156 return 12.0 + 6.0 * log(qscale/0.85) / log(2.0);
159 /* Texture bitrate is not quite inversely proportional to qscale,
160 * probably due the the changing number of SKIP blocks.
161 * MV bits level off at about qp<=12, because the lambda used
162 * for motion estimation is constant there. */
163 static inline double qscale2bits(ratecontrol_entry_t *rce, double qscale)
167 return (rce->i_tex_bits + rce->p_tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
168 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
173 int x264_ratecontrol_new( x264_t *h )
175 x264_ratecontrol_t *rc;
178 x264_cpu_restore( h->param.cpu );
180 rc = h->rc = x264_malloc( h->param.i_threads * sizeof(x264_ratecontrol_t) );
181 memset( rc, 0, h->param.i_threads * sizeof(x264_ratecontrol_t) );
183 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
184 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
186 /* FIXME: use integers */
187 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
188 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
192 rc->bitrate = h->param.rc.i_bitrate * 1000.;
193 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
194 rc->nmb = h->mb.i_mb_count;
195 rc->last_non_b_pict_type = -1;
198 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
200 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
203 if( h->param.rc.i_vbv_buffer_size )
205 if( h->param.rc.i_rc_method == X264_RC_CQP )
206 x264_log(h, X264_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
207 else if( h->param.rc.i_vbv_max_bitrate == 0 )
209 x264_log( h, X264_LOG_DEBUG, "VBV maxrate unspecified, assuming CBR\n" );
210 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
213 if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate &&
214 h->param.rc.i_vbv_max_bitrate > 0)
215 x264_log(h, X264_LOG_WARNING, "max bitrate less than average bitrate, ignored.\n");
216 else if( h->param.rc.i_vbv_max_bitrate > 0 &&
217 h->param.rc.i_vbv_buffer_size > 0 )
219 if( h->param.rc.i_vbv_buffer_size < 3 * h->param.rc.i_vbv_max_bitrate / rc->fps )
221 h->param.rc.i_vbv_buffer_size = 3 * h->param.rc.i_vbv_max_bitrate / rc->fps;
222 x264_log( h, X264_LOG_WARNING, "VBV buffer size too small, using %d kbit\n",
223 h->param.rc.i_vbv_buffer_size );
225 if( h->param.rc.f_vbv_buffer_init > 1. )
226 h->param.rc.f_vbv_buffer_init = x264_clip3f( h->param.rc.f_vbv_buffer_init / h->param.rc.i_vbv_buffer_size, 0, 1 );
227 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000. / rc->fps;
228 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000.;
229 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
230 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
231 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
233 rc->b_vbv_min_rate = !rc->b_2pass
234 && h->param.rc.i_rc_method == X264_RC_ABR
235 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
237 else if( h->param.rc.i_vbv_max_bitrate )
239 x264_log(h, X264_LOG_WARNING, "VBV maxrate specified, but no bufsize.\n");
240 h->param.rc.i_vbv_max_bitrate = 0;
242 if(rc->rate_tolerance < 0.01) {
243 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
244 rc->rate_tolerance = 0.01;
247 h->mb.b_variable_qp = rc->b_vbv && !rc->b_2pass;
251 /* FIXME ABR_INIT_QP is actually used only in CRF */
252 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
253 rc->accum_p_norm = .01;
254 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
255 /* estimated ratio that produces a reasonable QP for the first I-frame */
256 rc->cplxr_sum = .01 * pow( 7.0e5, h->param.rc.f_qcompress ) * pow( h->mb.i_mb_count, 0.5 );
257 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
258 rc->last_non_b_pict_type = SLICE_TYPE_I;
261 if( h->param.rc.i_rc_method == X264_RC_CRF )
263 /* arbitrary rescaling to make CRF somewhat similar to QP */
264 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
265 rc->rate_factor_constant = pow( base_cplx, 1 - h->param.rc.f_qcompress )
266 / qp2qscale( h->param.rc.f_rf_constant );
269 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
270 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
271 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
272 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
273 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
275 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
276 rc->last_qscale = qp2qscale(26);
277 rc->pred = x264_malloc( 5*sizeof(predictor_t) );
278 rc->pred_b_from_p = x264_malloc( sizeof(predictor_t) );
279 for( i = 0; i < 5; i++ )
281 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
282 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
283 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
284 rc->pred[i].coeff= 2.0;
285 rc->pred[i].count= 1.0;
286 rc->pred[i].decay= 0.5;
287 rc->row_preds[i].coeff= .25;
288 rc->row_preds[i].count= 1.0;
289 rc->row_preds[i].decay= 0.5;
291 *rc->pred_b_from_p = rc->pred[0];
293 if( parse_zones( h ) < 0 )
295 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
299 /* Load stat file and init 2pass algo */
300 if( h->param.rc.b_stat_read )
302 char *p, *stats_in, *stats_buf;
304 /* read 1st pass stats */
305 assert( h->param.rc.psz_stat_in );
306 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
309 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
313 /* check whether 1st pass options were compatible with current options */
314 if( !strncmp( stats_buf, "#options:", 9 ) )
317 char *opts = stats_buf;
318 stats_in = strchr( stats_buf, '\n' );
324 if( ( p = strstr( opts, "bframes=" ) ) && sscanf( p, "bframes=%d", &i )
325 && h->param.i_bframe != i )
327 x264_log( h, X264_LOG_ERROR, "different number of B-frames than 1st pass (%d vs %d)\n",
328 h->param.i_bframe, i );
332 /* since B-adapt doesn't (yet) take into account B-pyramid,
333 * the converse is not a problem */
334 if( strstr( opts, "b_pyramid=1" ) && !h->param.b_bframe_pyramid )
335 x264_log( h, X264_LOG_WARNING, "1st pass used B-pyramid, 2nd doesn't\n" );
337 if( ( p = strstr( opts, "keyint=" ) ) && sscanf( p, "keyint=%d", &i )
338 && h->param.i_keyint_max != i )
339 x264_log( h, X264_LOG_WARNING, "different keyint than 1st pass (%d vs %d)\n",
340 h->param.i_keyint_max, i );
342 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
343 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
346 /* find number of pics */
349 p = strchr(p+1, ';');
352 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
357 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
359 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
360 h->param.i_frame_total, rc->num_entries );
362 if( h->param.i_frame_total > rc->num_entries + h->param.i_bframe )
364 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
365 h->param.i_frame_total, rc->num_entries );
369 /* FIXME: ugly padding because VfW drops delayed B-frames */
370 rc->num_entries += h->param.i_bframe;
372 rc->entry = (ratecontrol_entry_t*) x264_malloc(rc->num_entries * sizeof(ratecontrol_entry_t));
373 memset(rc->entry, 0, rc->num_entries * sizeof(ratecontrol_entry_t));
375 /* init all to skipped p frames */
376 for(i=0; i<rc->num_entries; i++){
377 ratecontrol_entry_t *rce = &rc->entry[i];
378 rce->pict_type = SLICE_TYPE_P;
379 rce->qscale = rce->new_qscale = qp2qscale(20);
380 rce->misc_bits = rc->nmb + 10;
386 for(i=0; i < rc->num_entries - h->param.i_bframe; i++){
387 ratecontrol_entry_t *rce;
394 next= strchr(p, ';');
396 (*next)=0; //sscanf is unbelievably slow on looong strings
399 e = sscanf(p, " in:%d ", &frame_number);
401 if(frame_number < 0 || frame_number >= rc->num_entries)
403 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
406 rce = &rc->entry[frame_number];
407 rce->direct_mode = 0;
409 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",
410 &pict_type, &qp, &rce->i_tex_bits, &rce->p_tex_bits,
411 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
412 &rce->s_count, &rce->direct_mode);
415 case 'I': rce->kept_as_ref = 1;
416 case 'i': rce->pict_type = SLICE_TYPE_I; break;
417 case 'P': rce->pict_type = SLICE_TYPE_P; break;
418 case 'B': rce->kept_as_ref = 1;
419 case 'b': rce->pict_type = SLICE_TYPE_B; break;
420 default: e = -1; break;
423 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
426 rce->qscale = qp2qscale(qp);
430 x264_free(stats_buf);
432 if(h->param.rc.i_rc_method == X264_RC_ABR)
434 if(init_pass2(h) < 0) return -1;
435 } /* else we're using constant quant, so no need to run the bitrate allocation */
438 /* Open output file */
439 /* If input and output files are the same, output to a temp file
440 * and move it to the real name only when it's complete */
441 if( h->param.rc.b_stat_write )
445 rc->psz_stat_file_tmpname = x264_malloc( strlen(h->param.rc.psz_stat_out) + 6 );
446 strcpy( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
447 strcat( rc->psz_stat_file_tmpname, ".temp" );
449 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
450 if( rc->p_stat_file_out == NULL )
452 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
456 p = x264_param2string( &h->param, 1 );
457 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
461 for( i=1; i<h->param.i_threads; i++ )
463 h->thread[i]->rc = rc+i;
470 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
475 z->f_bitrate_factor = 1;
476 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
478 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
480 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
484 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
490 z->param = malloc( sizeof(x264_param_t) );
491 memcpy( z->param, &h->param, sizeof(x264_param_t) );
492 while( (tok = strtok_r( p, ",", &saveptr )) )
494 char *val = strchr( tok, '=' );
500 if( x264_param_parse( z->param, tok, val ) )
502 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
510 static int parse_zones( x264_t *h )
512 x264_ratecontrol_t *rc = h->rc;
514 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
516 char *p, *tok, *saveptr;
517 char *psz_zones = x264_malloc( strlen(h->param.rc.psz_zones)+1 );
518 strcpy( psz_zones, h->param.rc.psz_zones );
519 h->param.rc.i_zones = 1;
520 for( p = psz_zones; *p; p++ )
521 h->param.rc.i_zones += (*p == '/');
522 h->param.rc.zones = x264_malloc( h->param.rc.i_zones * sizeof(x264_zone_t) );
524 for( i = 0; i < h->param.rc.i_zones; i++ )
526 tok = strtok_r( p, "/", &saveptr );
527 if( !tok || parse_zone( h, &h->param.rc.zones[i], tok ) )
531 x264_free( psz_zones );
534 if( h->param.rc.i_zones > 0 )
536 for( i = 0; i < h->param.rc.i_zones; i++ )
538 x264_zone_t z = h->param.rc.zones[i];
539 if( z.i_start < 0 || z.i_start > z.i_end )
541 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
542 z.i_start, z.i_end );
545 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
547 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
548 z.f_bitrate_factor );
553 rc->i_zones = h->param.rc.i_zones + 1;
554 rc->zones = x264_malloc( rc->i_zones * sizeof(x264_zone_t) );
555 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
557 // default zone to fall back to if none of the others match
558 rc->zones[0].i_start = 0;
559 rc->zones[0].i_end = INT_MAX;
560 rc->zones[0].b_force_qp = 0;
561 rc->zones[0].f_bitrate_factor = 1;
562 rc->zones[0].param = x264_malloc( sizeof(x264_param_t) );
563 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
564 for( i = 1; i < rc->i_zones; i++ )
566 if( !rc->zones[i].param )
567 rc->zones[i].param = rc->zones[0].param;
574 x264_zone_t *get_zone( x264_t *h, int frame_num )
577 for( i = h->rc->i_zones-1; i >= 0; i-- )
579 x264_zone_t *z = &h->rc->zones[i];
580 if( frame_num >= z->i_start && frame_num <= z->i_end )
586 void x264_ratecontrol_summary( x264_t *h )
588 x264_ratecontrol_t *rc = h->rc;
589 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
591 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
592 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
593 qscale2qp( pow( base_cplx, 1 - h->param.rc.f_qcompress )
594 * rc->cplxr_sum / rc->wanted_bits_window ) );
598 void x264_ratecontrol_delete( x264_t *h )
600 x264_ratecontrol_t *rc = h->rc;
603 if( rc->p_stat_file_out )
605 fclose( rc->p_stat_file_out );
606 if( h->i_frame >= rc->num_entries - h->param.i_bframe )
607 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
609 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
610 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
612 x264_free( rc->psz_stat_file_tmpname );
614 x264_free( rc->pred );
615 x264_free( rc->pred_b_from_p );
616 x264_free( rc->entry );
619 x264_free( rc->zones[0].param );
620 if( h->param.rc.psz_zones )
621 for( i=1; i<rc->i_zones; i++ )
622 if( rc->zones[i].param != rc->zones[0].param )
623 x264_free( rc->zones[i].param );
624 x264_free( rc->zones );
629 static void accum_p_qp_update( x264_t *h, float qp )
631 x264_ratecontrol_t *rc = h->rc;
632 rc->accum_p_qp *= .95;
633 rc->accum_p_norm *= .95;
634 rc->accum_p_norm += 1;
635 if( h->sh.i_type == SLICE_TYPE_I )
636 rc->accum_p_qp += qp + rc->ip_offset;
638 rc->accum_p_qp += qp;
641 /* Before encoding a frame, choose a QP for it */
642 void x264_ratecontrol_start( x264_t *h, int i_force_qp )
644 x264_ratecontrol_t *rc = h->rc;
645 ratecontrol_entry_t *rce = NULL;
646 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
649 x264_cpu_restore( h->param.cpu );
651 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
652 x264_encoder_reconfig( h, zone->param );
653 rc->prev_zone = zone;
655 rc->qp_force = i_force_qp;
657 if( h->param.rc.b_stat_read )
659 int frame = h->fenc->i_frame;
660 assert( frame >= 0 && frame < rc->num_entries );
661 rce = h->rc->rce = &h->rc->entry[frame];
663 if( h->sh.i_type == SLICE_TYPE_B
664 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
666 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
667 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
673 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
674 rc->row_pred = &rc->row_preds[h->sh.i_type];
675 update_vbv_plan( h );
678 if( h->sh.i_type != SLICE_TYPE_B )
681 while( h->frames.current[rc->bframes] && IS_X264_TYPE_B(h->frames.current[rc->bframes]->i_type) )
691 q = qscale2qp( rate_estimate_qscale( h ) );
693 else if( rc->b_2pass )
695 rce->new_qscale = rate_estimate_qscale( h );
696 q = qscale2qp( rce->new_qscale );
700 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
701 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
703 q = rc->qp_constant[ h->sh.i_type ];
707 if( zone->b_force_qp )
708 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
710 q -= 6*log(zone->f_bitrate_factor)/log(2);
716 h->fdec->f_qp_avg_rc =
717 h->fdec->f_qp_avg_aq =
719 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
721 rce->new_qp = rc->qp;
723 /* accum_p_qp needs to be here so that future frames can benefit from the
724 * data before this frame is done. but this only works because threading
725 * guarantees to not re-encode any frames. so the non-threaded case does
726 * accum_p_qp later. */
727 if( h->param.i_threads > 1 )
728 accum_p_qp_update( h, rc->qp );
730 if( h->sh.i_type != SLICE_TYPE_B )
731 rc->last_non_b_pict_type = h->sh.i_type;
734 double predict_row_size( x264_t *h, int y, int qp )
736 /* average between two predictors:
737 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
738 x264_ratecontrol_t *rc = h->rc;
739 double pred_s = predict_size( rc->row_pred, qp2qscale(qp), h->fdec->i_row_satd[y] );
741 if( h->sh.i_type != SLICE_TYPE_I
742 && h->fref0[0]->i_type == h->fdec->i_type
743 && h->fref0[0]->i_row_satd[y] > 0 )
745 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
746 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
751 return (pred_s + pred_t) / 2;
754 double predict_row_size_sum( x264_t *h, int y, int qp )
758 for( i = 0; i <= y; i++ )
759 bits += h->fdec->i_row_bits[i];
760 for( i = y+1; i < h->sps->i_mb_height; i++ )
761 bits += predict_row_size( h, i, qp );
765 void x264_ratecontrol_mb( x264_t *h, int bits )
767 x264_ratecontrol_t *rc = h->rc;
768 const int y = h->mb.i_mb_y;
770 x264_cpu_restore( h->param.cpu );
772 h->fdec->i_row_bits[y] += bits;
773 rc->qpa_rc += rc->qpm;
774 rc->qpa_aq += h->mb.i_qp;
776 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !rc->b_vbv || rc->b_2pass )
779 h->fdec->i_row_qp[y] = rc->qpm;
781 if( h->sh.i_type == SLICE_TYPE_B )
783 /* B-frames shouldn't use lower QP than their reference frames */
784 if( y < h->sps->i_mb_height-1 )
786 rc->qpm = X264_MAX( rc->qp,
787 X264_MIN( h->fref0[0]->i_row_qp[y+1],
788 h->fref1[0]->i_row_qp[y+1] ));
793 update_predictor( rc->row_pred, qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
795 /* tweak quality based on difference from predicted size */
796 if( y < h->sps->i_mb_height-1 && h->stat.i_slice_count[h->sh.i_type] > 0 )
798 int prev_row_qp = h->fdec->i_row_qp[y];
799 int b0 = predict_row_size_sum( h, y, rc->qpm );
801 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
802 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
803 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
805 if( !rc->b_vbv_min_rate )
806 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
808 while( rc->qpm < i_qp_max
809 && (b1 > rc->frame_size_planned * 1.15
810 || (rc->buffer_fill - b1 < buffer_left_planned * 0.5)))
813 b1 = predict_row_size_sum( h, y, rc->qpm );
816 while( rc->qpm > i_qp_min
817 && buffer_left_planned > rc->buffer_size * 0.4
818 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
819 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
822 b1 = predict_row_size_sum( h, y, rc->qpm );
828 int x264_ratecontrol_qp( x264_t *h )
833 /* In 2pass, force the same frame types as in the 1st pass */
834 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
836 x264_ratecontrol_t *rc = h->rc;
837 if( h->param.rc.b_stat_read )
839 if( frame_num >= rc->num_entries )
841 /* We could try to initialize everything required for ABR and
842 * adaptive B-frames, but that would be complicated.
843 * So just calculate the average QP used so far. */
845 h->param.rc.i_qp_constant = (h->stat.i_slice_count[SLICE_TYPE_P] == 0) ? 24
846 : 1 + h->stat.f_slice_qp[SLICE_TYPE_P] / h->stat.i_slice_count[SLICE_TYPE_P];
847 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
848 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 );
849 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 );
851 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
852 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
853 if( h->param.b_bframe_adaptive )
854 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
858 h->param.rc.i_rc_method = X264_RC_CQP;
859 h->param.rc.b_stat_read = 0;
860 h->param.b_bframe_adaptive = 0;
861 if( h->param.i_bframe > 1 )
862 h->param.i_bframe = 1;
865 switch( rc->entry[frame_num].pict_type )
868 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I;
871 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B;
880 return X264_TYPE_AUTO;
884 /* After encoding one frame, save stats and update ratecontrol state */
885 void x264_ratecontrol_end( x264_t *h, int bits )
887 x264_ratecontrol_t *rc = h->rc;
888 const int *mbs = h->stat.frame.i_mb_count;
891 x264_cpu_restore( h->param.cpu );
893 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
894 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
895 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
896 for( i = B_DIRECT; i < B_8x8; i++ )
897 h->stat.frame.i_mb_count_p += mbs[i];
899 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
900 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
902 if( h->param.rc.b_stat_write )
904 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
905 : h->sh.i_type==SLICE_TYPE_P ? 'P'
906 : h->fenc->b_kept_as_ref ? 'B' : 'b';
907 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
908 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
909 char c_direct = h->mb.b_direct_auto_write ?
910 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
911 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
913 fprintf( rc->p_stat_file_out,
914 "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",
915 h->fenc->i_frame, h->i_frame,
917 h->stat.frame.i_itex_bits, h->stat.frame.i_ptex_bits,
918 h->stat.frame.i_hdr_bits, h->stat.frame.i_misc_bits,
919 h->stat.frame.i_mb_count_i,
920 h->stat.frame.i_mb_count_p,
921 h->stat.frame.i_mb_count_skip,
927 if( h->sh.i_type != SLICE_TYPE_B )
928 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / rc->last_rceq;
931 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
932 * Not perfectly accurate with B-refs, but good enough. */
933 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
935 rc->cplxr_sum *= rc->cbr_decay;
936 rc->wanted_bits_window += rc->bitrate / rc->fps;
937 rc->wanted_bits_window *= rc->cbr_decay;
939 if( h->param.i_threads == 1 )
940 accum_p_qp_update( h, rc->qpa_rc );
945 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
948 if( h->mb.b_variable_qp )
950 if( h->sh.i_type == SLICE_TYPE_B )
952 rc->bframe_bits += bits;
953 if( !h->frames.current[0] || !IS_X264_TYPE_B(h->frames.current[0]->i_type) )
955 update_predictor( rc->pred_b_from_p, qp2qscale(rc->qpa_rc),
956 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
962 update_vbv( h, bits );
965 /****************************************************************************
967 ***************************************************************************/
969 double x264_eval( char *s, double *const_value, const char **const_name,
970 double (**func1)(void *, double), const char **func1_name,
971 double (**func2)(void *, double, double), char **func2_name,
975 * modify the bitrate curve from pass1 for one frame
977 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
979 x264_ratecontrol_t *rcc= h->rc;
980 const int pict_type = rce->pict_type;
982 x264_zone_t *zone = get_zone( h, frame_num );
984 double const_values[]={
985 rce->i_tex_bits * rce->qscale,
986 rce->p_tex_bits * rce->qscale,
987 (rce->i_tex_bits + rce->p_tex_bits) * rce->qscale,
988 rce->mv_bits * rce->qscale,
989 (double)rce->i_count / rcc->nmb,
990 (double)rce->p_count / rcc->nmb,
991 (double)rce->s_count / rcc->nmb,
992 rce->pict_type == SLICE_TYPE_I,
993 rce->pict_type == SLICE_TYPE_P,
994 rce->pict_type == SLICE_TYPE_B,
995 h->param.rc.f_qcompress,
996 rcc->i_cplx_sum[SLICE_TYPE_I] / rcc->frame_count[SLICE_TYPE_I],
997 rcc->i_cplx_sum[SLICE_TYPE_P] / rcc->frame_count[SLICE_TYPE_P],
998 rcc->p_cplx_sum[SLICE_TYPE_P] / rcc->frame_count[SLICE_TYPE_P],
999 rcc->p_cplx_sum[SLICE_TYPE_B] / rcc->frame_count[SLICE_TYPE_B],
1000 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / rcc->frame_count[pict_type],
1001 rce->blurred_complexity,
1004 static const char *const_names[]={
1024 static double (*func1[])(void *, double)={
1025 // (void *)bits2qscale,
1026 (void *)qscale2bits,
1029 static const char *func1_names[]={
1035 q = x264_eval((char*)h->param.rc.psz_rc_eq, const_values, const_names, func1, func1_names, NULL, NULL, rce);
1037 // avoid NaN's in the rc_eq
1038 if(!isfinite(q) || rce->i_tex_bits + rce->p_tex_bits + rce->mv_bits == 0)
1039 q = rcc->last_qscale;
1043 rcc->last_qscale = q;
1048 if( zone->b_force_qp )
1049 q = qp2qscale(zone->i_qp);
1051 q /= zone->f_bitrate_factor;
1057 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1059 x264_ratecontrol_t *rcc = h->rc;
1060 const int pict_type = rce->pict_type;
1062 // force I/B quants as a function of P quants
1063 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1064 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1065 if( pict_type == SLICE_TYPE_I )
1068 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1069 double ip_factor = fabs( h->param.rc.f_ip_factor );
1070 /* don't apply ip_factor if the following frame is also I */
1071 if( rcc->accum_p_norm <= 0 )
1073 else if( h->param.rc.f_ip_factor < 0 )
1075 else if( rcc->accum_p_norm >= 1 )
1078 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1080 else if( pict_type == SLICE_TYPE_B )
1082 if( h->param.rc.f_pb_factor > 0 )
1084 if( !rce->kept_as_ref )
1085 q *= fabs( h->param.rc.f_pb_factor );
1087 else if( pict_type == SLICE_TYPE_P
1088 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1089 && rce->i_tex_bits + rce->p_tex_bits == 0 )
1094 /* last qscale / qdiff stuff */
1095 if(rcc->last_non_b_pict_type==pict_type
1096 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1098 double last_q = rcc->last_qscale_for[pict_type];
1099 double max_qscale = last_q * rcc->lstep;
1100 double min_qscale = last_q / rcc->lstep;
1102 if (q > max_qscale) q = max_qscale;
1103 else if(q < min_qscale) q = min_qscale;
1106 rcc->last_qscale_for[pict_type] = q;
1107 if(pict_type!=SLICE_TYPE_B)
1108 rcc->last_non_b_pict_type = pict_type;
1109 if(pict_type==SLICE_TYPE_I)
1111 rcc->last_accum_p_norm = rcc->accum_p_norm;
1112 rcc->accum_p_norm = 0;
1113 rcc->accum_p_qp = 0;
1115 if(pict_type==SLICE_TYPE_P)
1117 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1118 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1119 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1124 static double predict_size( predictor_t *p, double q, double var )
1126 return p->coeff*var / (q*p->count);
1129 static void update_predictor( predictor_t *p, double q, double var, double bits )
1133 p->count *= p->decay;
1134 p->coeff *= p->decay;
1136 p->coeff += bits*q / var;
1139 // update VBV after encoding a frame
1140 static void update_vbv( x264_t *h, int bits )
1142 x264_ratecontrol_t *rcc = h->rc;
1143 x264_ratecontrol_t *rct = h->thread[0]->rc;
1145 if( rcc->last_satd >= h->mb.i_mb_count )
1146 update_predictor( &rct->pred[h->sh.i_type], qp2qscale(rcc->qpa_rc), rcc->last_satd, bits );
1151 rct->buffer_fill_final += rct->buffer_rate - bits;
1152 if( rct->buffer_fill_final < 0 && !rct->b_2pass )
1153 x264_log( h, X264_LOG_WARNING, "VBV underflow (%.0f bits)\n", rct->buffer_fill_final );
1154 rct->buffer_fill_final = x264_clip3f( rct->buffer_fill_final, 0, rct->buffer_size );
1157 // provisionally update VBV according to the planned size of all frames currently in progress
1158 static void update_vbv_plan( x264_t *h )
1160 x264_ratecontrol_t *rcc = h->rc;
1161 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final;
1162 if( h->param.i_threads > 1 )
1164 int j = h->rc - h->thread[0]->rc;
1166 for( i=1; i<h->param.i_threads; i++ )
1168 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1169 double bits = t->rc->frame_size_planned;
1170 if( !t->b_thread_active )
1172 rcc->buffer_fill += rcc->buffer_rate - bits;
1173 rcc->buffer_fill = x264_clip3( rcc->buffer_fill, 0, rcc->buffer_size );
1178 // apply VBV constraints and clip qscale to between lmin and lmax
1179 static double clip_qscale( x264_t *h, int pict_type, double q )
1181 x264_ratecontrol_t *rcc = h->rc;
1182 double lmin = rcc->lmin[pict_type];
1183 double lmax = rcc->lmax[pict_type];
1186 /* B-frames are not directly subject to VBV,
1187 * since they are controlled by the P-frames' QPs.
1188 * FIXME: in 2pass we could modify previous frames' QP too,
1189 * instead of waiting for the buffer to fill */
1191 ( pict_type == SLICE_TYPE_P ||
1192 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) )
1194 if( rcc->buffer_fill/rcc->buffer_size < 0.5 )
1195 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1198 if( rcc->b_vbv && rcc->last_satd > 0 )
1200 /* Now a hard threshold to make sure the frame fits in VBV.
1201 * This one is mostly for I-frames. */
1202 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1204 if( bits > rcc->buffer_fill/2 )
1205 qf = x264_clip3f( rcc->buffer_fill/(2*bits), 0.2, 1.0 );
1208 if( bits < rcc->buffer_rate/2 )
1209 q *= bits*2/rcc->buffer_rate;
1210 q = X264_MAX( q0, q );
1212 /* Check B-frame complexity, and use up any bits that would
1213 * overflow before the next P-frame. */
1214 if( h->sh.i_type == SLICE_TYPE_P )
1216 int nb = rcc->bframes;
1217 double pbbits = bits;
1218 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1221 if( bbits > rcc->buffer_rate )
1223 pbbits += nb * bbits;
1225 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1226 if( pbbits < space )
1228 q *= X264_MAX( pbbits / space,
1229 bits / (0.5 * rcc->buffer_size) );
1231 q = X264_MAX( q0-5, q );
1234 if( !rcc->b_vbv_min_rate )
1235 q = X264_MAX( q0, q );
1240 else if(rcc->b_2pass)
1242 double min2 = log(lmin);
1243 double max2 = log(lmax);
1244 q = (log(q) - min2)/(max2-min2) - 0.5;
1245 q = 1.0/(1.0 + exp(-4*q));
1246 q = q*(max2-min2) + min2;
1250 return x264_clip3f(q, lmin, lmax);
1253 // update qscale for 1 frame based on actual bits used so far
1254 static float rate_estimate_qscale( x264_t *h )
1257 x264_ratecontrol_t *rcc = h->rc;
1258 ratecontrol_entry_t rce;
1259 int pict_type = h->sh.i_type;
1260 double lmin = rcc->lmin[pict_type];
1261 double lmax = rcc->lmax[pict_type];
1262 int64_t total_bits = 8*(h->stat.i_slice_size[SLICE_TYPE_I]
1263 + h->stat.i_slice_size[SLICE_TYPE_P]
1264 + h->stat.i_slice_size[SLICE_TYPE_B]);
1269 if(pict_type != rce.pict_type)
1271 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1272 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1276 if( pict_type == SLICE_TYPE_B )
1278 /* B-frames don't have independent ratecontrol, but rather get the
1279 * average QP of the two adjacent P-frames + an offset */
1281 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1282 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1283 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1284 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1285 float q0 = h->fref0[0]->f_qp_avg_rc;
1286 float q1 = h->fref1[0]->f_qp_avg_rc;
1288 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1289 q0 -= rcc->pb_offset/2;
1290 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1291 q1 -= rcc->pb_offset/2;
1294 q = (q0 + q1) / 2 + rcc->ip_offset;
1300 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1302 if(h->fenc->b_kept_as_ref)
1303 q += rcc->pb_offset/2;
1305 q += rcc->pb_offset;
1307 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1309 return qp2qscale(q);
1313 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1316 //FIXME adjust abr_buffer based on distance to the end of the video
1317 int64_t diff = total_bits - (int64_t)rce.expected_bits;
1319 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1320 if( h->fenc->i_frame > 30 )
1322 /* Adjust quant based on the difference between
1323 * achieved and expected bitrate so far */
1324 double time = (double)h->fenc->i_frame / rcc->num_entries;
1325 double w = x264_clip3f( time*100, 0.0, 1.0 );
1326 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1328 q = x264_clip3f( q, lmin, lmax );
1330 else /* 1pass ABR */
1332 /* Calculate the quantizer which would have produced the desired
1333 * average bitrate if it had been applied to all frames so far.
1334 * Then modulate that quant based on the current frame's complexity
1335 * relative to the average complexity so far (using the 2pass RCEQ).
1336 * Then bias the quant up or down if total size so far was far from
1338 * Result: Depending on the value of rate_tolerance, there is a
1339 * tradeoff between quality and bitrate precision. But at large
1340 * tolerances, the bit distribution approaches that of 2pass. */
1342 double wanted_bits, overflow=1, lmin, lmax;
1344 rcc->last_satd = x264_rc_analyse_slice( h );
1345 rcc->short_term_cplxsum *= 0.5;
1346 rcc->short_term_cplxcount *= 0.5;
1347 rcc->short_term_cplxsum += rcc->last_satd;
1348 rcc->short_term_cplxcount ++;
1350 rce.p_tex_bits = rcc->last_satd;
1351 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1354 rce.p_count = rcc->nmb;
1358 rce.pict_type = pict_type;
1360 if( h->param.rc.i_rc_method == X264_RC_CRF )
1362 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1366 int i_frame_done = h->fenc->i_frame + 1 - h->param.i_threads;
1368 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1370 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
1371 wanted_bits = i_frame_done * rcc->bitrate / rcc->fps;
1372 if( wanted_bits > 0 )
1374 abr_buffer *= X264_MAX( 1, sqrt(i_frame_done/25) );
1375 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1380 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1381 /* should test _next_ pict type, but that isn't decided yet */
1382 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1384 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1385 q /= fabs( h->param.rc.f_ip_factor );
1387 else if( h->i_frame > 0 )
1389 /* Asymmetric clipping, because symmetric would prevent
1390 * overflow control in areas of rapidly oscillating complexity */
1391 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1392 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1393 if( overflow > 1.1 && h->i_frame > 3 )
1395 else if( overflow < 0.9 )
1398 q = x264_clip3f(q, lmin, lmax);
1400 else if( h->param.rc.i_rc_method == X264_RC_CRF )
1402 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1405 //FIXME use get_diff_limited_q() ?
1406 q = clip_qscale( h, pict_type, q );
1409 rcc->last_qscale_for[pict_type] =
1410 rcc->last_qscale = q;
1412 if( !rcc->b_2pass && h->fenc->i_frame == 0 )
1413 rcc->last_qscale_for[SLICE_TYPE_P] = q;
1415 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1420 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
1424 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
1425 /* these vars are updated in x264_ratecontrol_start()
1426 * so copy them from the context that most recently started (prev)
1427 * to the context that's about to start (cur).
1433 COPY(last_qscale_for);
1434 COPY(last_non_b_pict_type);
1435 COPY(short_term_cplxsum);
1436 COPY(short_term_cplxcount);
1443 #define COPY(var) next->rc->var = cur->rc->var
1444 /* these vars are updated in x264_ratecontrol_end()
1445 * so copy them from the context that most recently ended (cur)
1446 * to the context that's about to end (next)
1449 COPY(expected_bits_sum);
1450 COPY(wanted_bits_window);
1454 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
1455 /* the rest of the variables are either constant or thread-local */
1458 static int init_pass2( x264_t *h )
1460 x264_ratecontrol_t *rcc = h->rc;
1461 uint64_t all_const_bits = 0;
1462 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps);
1463 double rate_factor, step, step_mult;
1464 double qblur = h->param.rc.f_qblur;
1465 double cplxblur = h->param.rc.f_complexity_blur;
1466 const int filter_size = (int)(qblur*4) | 1;
1467 double expected_bits;
1468 double *qscale, *blurred_qscale;
1471 /* find total/average complexity & const_bits */
1472 for(i=0; i<rcc->num_entries; i++){
1473 ratecontrol_entry_t *rce = &rcc->entry[i];
1474 all_const_bits += rce->misc_bits;
1475 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
1476 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
1477 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits * rce->qscale;
1478 rcc->frame_count[rce->pict_type] ++;
1481 if( all_available_bits < all_const_bits)
1483 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
1484 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)));
1488 /* Blur complexities, to reduce local fluctuation of QP.
1489 * We don't blur the QPs directly, because then one very simple frame
1490 * could drag down the QP of a nearby complex frame and give it more
1491 * bits than intended. */
1492 for(i=0; i<rcc->num_entries; i++){
1493 ratecontrol_entry_t *rce = &rcc->entry[i];
1494 double weight_sum = 0;
1495 double cplx_sum = 0;
1496 double weight = 1.0;
1498 /* weighted average of cplx of future frames */
1499 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++){
1500 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
1501 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1504 weight_sum += weight;
1505 cplx_sum += weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1507 /* weighted average of cplx of past frames */
1509 for(j=0; j<=cplxblur*2 && j<=i; j++){
1510 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
1511 weight_sum += weight;
1512 cplx_sum += weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1513 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1517 rce->blurred_complexity = cplx_sum / weight_sum;
1520 qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1522 blurred_qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1524 blurred_qscale = qscale;
1526 /* Search for a factor which, when multiplied by the RCEQ values from
1527 * each frame, adds up to the desired total size.
1528 * There is no exact closed-form solution because of VBV constraints and
1529 * because qscale2bits is not invertible, but we can start with the simple
1530 * approximation of scaling the 1st pass by the ratio of bitrates.
1531 * The search range is probably overkill, but speed doesn't matter here. */
1534 for(i=0; i<rcc->num_entries; i++)
1535 expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i));
1536 step_mult = all_available_bits / expected_bits;
1539 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5){
1541 rate_factor += step;
1543 rcc->last_non_b_pict_type = -1;
1544 rcc->last_accum_p_norm = 1;
1545 rcc->accum_p_norm = 0;
1546 rcc->buffer_fill = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
1549 for(i=0; i<rcc->num_entries; i++){
1550 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
1553 /* fixed I/B qscale relative to P */
1554 for(i=rcc->num_entries-1; i>=0; i--){
1555 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
1556 assert(qscale[i] >= 0);
1560 if(filter_size > 1){
1561 assert(filter_size%2==1);
1562 for(i=0; i<rcc->num_entries; i++){
1563 ratecontrol_entry_t *rce = &rcc->entry[i];
1565 double q=0.0, sum=0.0;
1567 for(j=0; j<filter_size; j++){
1568 int index = i+j-filter_size/2;
1570 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
1571 if(index < 0 || index >= rcc->num_entries) continue;
1572 if(rce->pict_type != rcc->entry[index].pict_type) continue;
1573 q += qscale[index] * coeff;
1576 blurred_qscale[i] = q/sum;
1580 /* find expected bits */
1581 for(i=0; i<rcc->num_entries; i++){
1582 ratecontrol_entry_t *rce = &rcc->entry[i];
1584 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
1585 assert(rce->new_qscale >= 0);
1586 bits = qscale2bits(rce, rce->new_qscale);
1588 rce->expected_bits = expected_bits;
1589 expected_bits += bits;
1590 update_vbv(h, bits);
1591 rcc->buffer_fill = rcc->buffer_fill_final;
1594 //printf("expected:%llu available:%llu factor:%lf avgQ:%lf\n", (uint64_t)expected_bits, all_available_bits, rate_factor);
1595 if(expected_bits > all_available_bits) rate_factor -= step;
1600 x264_free(blurred_qscale);
1602 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
1605 for(i=0; i<rcc->num_entries; i++)
1606 avgq += rcc->entry[i].new_qscale;
1607 avgq = qscale2qp(avgq / rcc->num_entries);
1609 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
1610 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
1611 (float)h->param.rc.i_bitrate,
1612 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
1614 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
1616 if(h->param.rc.i_qp_min > 0)
1617 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
1619 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
1621 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
1623 if(h->param.rc.i_qp_max < 51)
1624 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
1626 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
1629 x264_log(h, X264_LOG_WARNING, "internal error\n");