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 */
150 static int parse_zones( x264_t *h );
151 static int init_pass2(x264_t *);
152 static float rate_estimate_qscale( x264_t *h );
153 static void update_vbv( x264_t *h, int bits );
154 static void update_vbv_plan( x264_t *h );
155 static double predict_size( predictor_t *p, double q, double var );
156 static void update_predictor( predictor_t *p, double q, double var, double bits );
157 int x264_rc_analyse_slice( x264_t *h );
160 * qp = h.264's quantizer
161 * qscale = linearized quantizer = Lagrange multiplier
163 static inline double qp2qscale(double qp)
165 return 0.85 * pow(2.0, ( qp - 12.0 ) / 6.0);
167 static inline double qscale2qp(double qscale)
169 return 12.0 + 6.0 * log(qscale/0.85) / log(2.0);
172 /* Texture bitrate is not quite inversely proportional to qscale,
173 * probably due the the changing number of SKIP blocks.
174 * MV bits level off at about qp<=12, because the lambda used
175 * for motion estimation is constant there. */
176 static inline double qscale2bits(ratecontrol_entry_t *rce, double qscale)
180 return (rce->i_tex_bits + rce->p_tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
181 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
186 int x264_ratecontrol_new( x264_t *h )
188 x264_ratecontrol_t *rc;
191 x264_cpu_restore( h->param.cpu );
193 rc = h->rc = x264_malloc( h->param.i_threads * sizeof(x264_ratecontrol_t) );
194 memset( rc, 0, h->param.i_threads * sizeof(x264_ratecontrol_t) );
196 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
197 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
199 /* FIXME: use integers */
200 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
201 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
205 rc->bitrate = h->param.rc.i_bitrate * 1000;
206 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
207 rc->nmb = h->mb.i_mb_count;
208 rc->last_non_b_pict_type = -1;
211 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
213 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
216 if( h->param.rc.i_vbv_buffer_size )
218 if( h->param.rc.i_rc_method == X264_RC_CQP )
219 x264_log(h, X264_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
220 else if( h->param.rc.i_vbv_max_bitrate == 0 )
222 x264_log( h, X264_LOG_DEBUG, "VBV maxrate unspecified, assuming CBR\n" );
223 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
226 if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate &&
227 h->param.rc.i_vbv_max_bitrate > 0)
228 x264_log(h, X264_LOG_WARNING, "max bitrate less than average bitrate, ignored.\n");
229 else if( h->param.rc.i_vbv_max_bitrate > 0 &&
230 h->param.rc.i_vbv_buffer_size > 0 )
232 if( h->param.rc.i_vbv_buffer_size < 3 * h->param.rc.i_vbv_max_bitrate / rc->fps )
234 h->param.rc.i_vbv_buffer_size = 3 * h->param.rc.i_vbv_max_bitrate / rc->fps;
235 x264_log( h, X264_LOG_WARNING, "VBV buffer size too small, using %d kbit\n",
236 h->param.rc.i_vbv_buffer_size );
238 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000 / rc->fps;
239 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000;
240 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
241 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
242 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
244 rc->b_vbv_min_rate = !rc->b_2pass
245 && h->param.rc.i_rc_method == X264_RC_ABR
246 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
248 else if( h->param.rc.i_vbv_max_bitrate )
250 x264_log(h, X264_LOG_WARNING, "VBV maxrate specified, but no bufsize.\n");
251 h->param.rc.i_vbv_max_bitrate = 0;
253 if(rc->rate_tolerance < 0.01) {
254 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
255 rc->rate_tolerance = 0.01;
258 h->mb.b_variable_qp = rc->b_vbv && !rc->b_2pass;
262 /* FIXME ABR_INIT_QP is actually used only in CRF */
263 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
264 rc->accum_p_norm = .01;
265 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
266 /* estimated ratio that produces a reasonable QP for the first I-frame */
267 rc->cplxr_sum = .01 * pow( 7.0e5, h->param.rc.f_qcompress ) * pow( h->mb.i_mb_count, 0.5 );
268 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
269 rc->last_non_b_pict_type = SLICE_TYPE_I;
272 if( h->param.rc.i_rc_method == X264_RC_CRF )
274 /* arbitrary rescaling to make CRF somewhat similar to QP */
275 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
276 rc->rate_factor_constant = pow( base_cplx, 1 - h->param.rc.f_qcompress )
277 / qp2qscale( h->param.rc.f_rf_constant );
280 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
281 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
282 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
283 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
284 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
286 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
287 rc->last_qscale = qp2qscale(26);
288 rc->pred = x264_malloc( 5*sizeof(predictor_t) );
289 rc->pred_b_from_p = x264_malloc( sizeof(predictor_t) );
290 for( i = 0; i < 5; i++ )
292 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
293 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
294 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
295 rc->pred[i].coeff= 2.0;
296 rc->pred[i].count= 1.0;
297 rc->pred[i].decay= 0.5;
298 rc->row_preds[i].coeff= .25;
299 rc->row_preds[i].count= 1.0;
300 rc->row_preds[i].decay= 0.5;
302 *rc->pred_b_from_p = rc->pred[0];
304 if( parse_zones( h ) < 0 )
307 /* Load stat file and init 2pass algo */
308 if( h->param.rc.b_stat_read )
310 char *p, *stats_in, *stats_buf;
312 /* read 1st pass stats */
313 assert( h->param.rc.psz_stat_in );
314 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
317 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
321 /* check whether 1st pass options were compatible with current options */
322 if( !strncmp( stats_buf, "#options:", 9 ) )
325 char *opts = stats_buf;
326 stats_in = strchr( stats_buf, '\n' );
332 if( ( p = strstr( opts, "bframes=" ) ) && sscanf( p, "bframes=%d", &i )
333 && h->param.i_bframe != i )
335 x264_log( h, X264_LOG_ERROR, "different number of B-frames than 1st pass (%d vs %d)\n",
336 h->param.i_bframe, i );
340 /* since B-adapt doesn't (yet) take into account B-pyramid,
341 * the converse is not a problem */
342 if( strstr( opts, "b_pyramid=1" ) && !h->param.b_bframe_pyramid )
343 x264_log( h, X264_LOG_WARNING, "1st pass used B-pyramid, 2nd doesn't\n" );
345 if( ( p = strstr( opts, "keyint=" ) ) && sscanf( p, "keyint=%d", &i )
346 && h->param.i_keyint_max != i )
347 x264_log( h, X264_LOG_WARNING, "different keyint than 1st pass (%d vs %d)\n",
348 h->param.i_keyint_max, i );
350 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
351 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
354 /* find number of pics */
357 p = strchr(p+1, ';');
360 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
365 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
367 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
368 h->param.i_frame_total, rc->num_entries );
370 if( h->param.i_frame_total > rc->num_entries + h->param.i_bframe )
372 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
373 h->param.i_frame_total, rc->num_entries );
377 /* FIXME: ugly padding because VfW drops delayed B-frames */
378 rc->num_entries += h->param.i_bframe;
380 rc->entry = (ratecontrol_entry_t*) x264_malloc(rc->num_entries * sizeof(ratecontrol_entry_t));
381 memset(rc->entry, 0, rc->num_entries * sizeof(ratecontrol_entry_t));
383 /* init all to skipped p frames */
384 for(i=0; i<rc->num_entries; i++){
385 ratecontrol_entry_t *rce = &rc->entry[i];
386 rce->pict_type = SLICE_TYPE_P;
387 rce->qscale = rce->new_qscale = qp2qscale(20);
388 rce->misc_bits = rc->nmb + 10;
394 for(i=0; i < rc->num_entries - h->param.i_bframe; i++){
395 ratecontrol_entry_t *rce;
402 next= strchr(p, ';');
404 (*next)=0; //sscanf is unbelievably slow on looong strings
407 e = sscanf(p, " in:%d ", &frame_number);
409 if(frame_number < 0 || frame_number >= rc->num_entries)
411 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
414 rce = &rc->entry[frame_number];
415 rce->direct_mode = 0;
417 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",
418 &pict_type, &qp, &rce->i_tex_bits, &rce->p_tex_bits,
419 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
420 &rce->s_count, &rce->direct_mode);
423 case 'I': rce->kept_as_ref = 1;
424 case 'i': rce->pict_type = SLICE_TYPE_I; break;
425 case 'P': rce->pict_type = SLICE_TYPE_P; break;
426 case 'B': rce->kept_as_ref = 1;
427 case 'b': rce->pict_type = SLICE_TYPE_B; break;
428 default: e = -1; break;
431 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
434 rce->qscale = qp2qscale(qp);
438 x264_free(stats_buf);
440 if(h->param.rc.i_rc_method == X264_RC_ABR)
442 if(init_pass2(h) < 0) return -1;
443 } /* else we're using constant quant, so no need to run the bitrate allocation */
446 /* Open output file */
447 /* If input and output files are the same, output to a temp file
448 * and move it to the real name only when it's complete */
449 if( h->param.rc.b_stat_write )
453 rc->psz_stat_file_tmpname = x264_malloc( strlen(h->param.rc.psz_stat_out) + 6 );
454 strcpy( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
455 strcat( rc->psz_stat_file_tmpname, ".temp" );
457 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
458 if( rc->p_stat_file_out == NULL )
460 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
464 p = x264_param2string( &h->param, 1 );
465 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
469 for( i=1; i<h->param.i_threads; i++ )
471 h->thread[i]->rc = rc+i;
478 static int parse_zones( x264_t *h )
480 x264_ratecontrol_t *rc = h->rc;
482 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
485 h->param.rc.i_zones = 1;
486 for( p = h->param.rc.psz_zones; *p; p++ )
487 h->param.rc.i_zones += (*p == '/');
488 h->param.rc.zones = x264_malloc( h->param.rc.i_zones * sizeof(x264_zone_t) );
489 p = h->param.rc.psz_zones;
490 for( i = 0; i < h->param.rc.i_zones; i++)
492 x264_zone_t *z = &h->param.rc.zones[i];
493 if( 3 == sscanf(p, "%u,%u,q=%u", &z->i_start, &z->i_end, &z->i_qp) )
495 else if( 3 == sscanf(p, "%u,%u,b=%f", &z->i_start, &z->i_end, &z->f_bitrate_factor) )
499 char *slash = strchr(p, '/');
500 if(slash) *slash = '\0';
501 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
504 p = strchr(p, '/') + 1;
508 if( h->param.rc.i_zones > 0 )
510 for( i = 0; i < h->param.rc.i_zones; i++ )
512 x264_zone_t z = h->param.rc.zones[i];
513 if( z.i_start < 0 || z.i_start > z.i_end )
515 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
516 z.i_start, z.i_end );
519 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
521 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
522 z.f_bitrate_factor );
527 rc->i_zones = h->param.rc.i_zones;
528 rc->zones = x264_malloc( rc->i_zones * sizeof(x264_zone_t) );
529 memcpy( rc->zones, h->param.rc.zones, rc->i_zones * sizeof(x264_zone_t) );
535 x264_zone_t *get_zone( x264_t *h, int frame_num )
538 for( i = h->rc->i_zones-1; i >= 0; i-- )
540 x264_zone_t *z = &h->rc->zones[i];
541 if( frame_num >= z->i_start && frame_num <= z->i_end )
547 void x264_ratecontrol_summary( x264_t *h )
549 x264_ratecontrol_t *rc = h->rc;
550 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
552 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
553 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
554 qscale2qp( pow( base_cplx, 1 - h->param.rc.f_qcompress )
555 * rc->cplxr_sum / rc->wanted_bits_window ) );
559 void x264_ratecontrol_delete( x264_t *h )
561 x264_ratecontrol_t *rc = h->rc;
563 if( rc->p_stat_file_out )
565 fclose( rc->p_stat_file_out );
566 if( h->i_frame >= rc->num_entries - h->param.i_bframe )
567 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
569 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
570 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
572 x264_free( rc->psz_stat_file_tmpname );
574 x264_free( rc->pred );
575 x264_free( rc->pred_b_from_p );
576 x264_free( rc->entry );
577 x264_free( rc->zones );
581 static void accum_p_qp_update( x264_t *h, float qp )
583 x264_ratecontrol_t *rc = h->rc;
584 rc->accum_p_qp *= .95;
585 rc->accum_p_norm *= .95;
586 rc->accum_p_norm += 1;
587 if( h->sh.i_type == SLICE_TYPE_I )
588 rc->accum_p_qp += qp + rc->ip_offset;
590 rc->accum_p_qp += qp;
593 /* Before encoding a frame, choose a QP for it */
594 void x264_ratecontrol_start( x264_t *h, int i_force_qp )
596 x264_ratecontrol_t *rc = h->rc;
597 ratecontrol_entry_t *rce = NULL;
600 x264_cpu_restore( h->param.cpu );
602 rc->qp_force = i_force_qp;
604 if( h->param.rc.b_stat_read )
606 int frame = h->fenc->i_frame;
607 assert( frame >= 0 && frame < rc->num_entries );
608 rce = h->rc->rce = &h->rc->entry[frame];
610 if( h->sh.i_type == SLICE_TYPE_B
611 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
613 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
614 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
620 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
621 rc->row_pred = &rc->row_preds[h->sh.i_type];
622 update_vbv_plan( h );
625 if( h->sh.i_type != SLICE_TYPE_B )
628 while( h->frames.current[rc->bframes] && IS_X264_TYPE_B(h->frames.current[rc->bframes]->i_type) )
640 q = qscale2qp( rate_estimate_qscale( h ) );
642 else if( rc->b_2pass )
644 rce->new_qscale = rate_estimate_qscale( h );
645 q = qscale2qp( rce->new_qscale );
649 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
650 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
651 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
653 q = rc->qp_constant[ h->sh.i_type ];
657 if( zone->b_force_qp )
658 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
660 q -= 6*log(zone->f_bitrate_factor)/log(2);
666 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
668 rce->new_qp = rc->qp;
670 /* accum_p_qp needs to be here so that future frames can benefit from the
671 * data before this frame is done. but this only works because threading
672 * guarantees to not re-encode any frames. so the non-threaded case does
673 * accum_p_qp later. */
674 if( h->param.i_threads > 1 )
675 accum_p_qp_update( h, rc->qp );
677 if( h->sh.i_type != SLICE_TYPE_B )
678 rc->last_non_b_pict_type = h->sh.i_type;
681 double predict_row_size( x264_t *h, int y, int qp )
683 /* average between two predictors:
684 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
685 x264_ratecontrol_t *rc = h->rc;
686 double pred_s = predict_size( rc->row_pred, qp2qscale(qp), h->fdec->i_row_satd[y] );
688 if( h->sh.i_type != SLICE_TYPE_I
689 && h->fref0[0]->i_type == h->fdec->i_type
690 && h->fref0[0]->i_row_satd[y] > 0 )
692 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
693 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
698 return (pred_s + pred_t) / 2;
701 double predict_row_size_sum( x264_t *h, int y, int qp )
705 for( i = 0; i <= y; i++ )
706 bits += h->fdec->i_row_bits[i];
707 for( i = y+1; i < h->sps->i_mb_height; i++ )
708 bits += predict_row_size( h, i, qp );
712 void x264_ratecontrol_mb( x264_t *h, int bits )
714 x264_ratecontrol_t *rc = h->rc;
715 const int y = h->mb.i_mb_y;
717 x264_cpu_restore( h->param.cpu );
719 h->fdec->i_row_bits[y] += bits;
722 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !h->mb.b_variable_qp )
725 h->fdec->i_row_qp[y] = rc->qpm;
727 if( h->sh.i_type == SLICE_TYPE_B )
729 /* B-frames shouldn't use lower QP than their reference frames */
730 if( y < h->sps->i_mb_height-1 )
732 rc->qpm = X264_MAX( rc->qp,
733 X264_MIN( h->fref0[0]->i_row_qp[y+1],
734 h->fref1[0]->i_row_qp[y+1] ));
739 update_predictor( rc->row_pred, qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
741 /* tweak quality based on difference from predicted size */
742 if( y < h->sps->i_mb_height-1 && h->stat.i_slice_count[h->sh.i_type] > 0 )
744 int prev_row_qp = h->fdec->i_row_qp[y];
745 int b0 = predict_row_size_sum( h, y, rc->qpm );
747 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
748 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
749 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
751 if( !rc->b_vbv_min_rate )
752 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
754 while( rc->qpm < i_qp_max
755 && (b1 > rc->frame_size_planned * 1.15
756 || (rc->buffer_fill - b1 < buffer_left_planned * 0.5)))
759 b1 = predict_row_size_sum( h, y, rc->qpm );
762 while( rc->qpm > i_qp_min
763 && buffer_left_planned > rc->buffer_size * 0.4
764 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
765 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
768 b1 = predict_row_size_sum( h, y, rc->qpm );
774 int x264_ratecontrol_qp( x264_t *h )
779 /* In 2pass, force the same frame types as in the 1st pass */
780 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
782 x264_ratecontrol_t *rc = h->rc;
783 if( h->param.rc.b_stat_read )
785 if( frame_num >= rc->num_entries )
787 /* We could try to initialize everything required for ABR and
788 * adaptive B-frames, but that would be complicated.
789 * So just calculate the average QP used so far. */
791 h->param.rc.i_qp_constant = (h->stat.i_slice_count[SLICE_TYPE_P] == 0) ? 24
792 : 1 + h->stat.i_slice_qp[SLICE_TYPE_P] / h->stat.i_slice_count[SLICE_TYPE_P];
793 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
794 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 );
795 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 );
797 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
798 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
799 if( h->param.b_bframe_adaptive )
800 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
804 h->param.rc.i_rc_method = X264_RC_CQP;
805 h->param.rc.b_stat_read = 0;
806 h->param.b_bframe_adaptive = 0;
807 if( h->param.i_bframe > 1 )
808 h->param.i_bframe = 1;
811 switch( rc->entry[frame_num].pict_type )
814 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I;
817 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B;
826 return X264_TYPE_AUTO;
830 /* After encoding one frame, save stats and update ratecontrol state */
831 void x264_ratecontrol_end( x264_t *h, int bits )
833 x264_ratecontrol_t *rc = h->rc;
834 const int *mbs = h->stat.frame.i_mb_count;
837 x264_cpu_restore( h->param.cpu );
839 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
840 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
841 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
842 for( i = B_DIRECT; i < B_8x8; i++ )
843 h->stat.frame.i_mb_count_p += mbs[i];
845 if( h->mb.b_variable_qp )
846 rc->qpa /= h->mb.i_mb_count;
849 h->fdec->f_qp_avg = rc->qpa;
851 if( h->param.rc.b_stat_write )
853 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
854 : h->sh.i_type==SLICE_TYPE_P ? 'P'
855 : h->fenc->b_kept_as_ref ? 'B' : 'b';
856 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
857 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
858 char c_direct = h->mb.b_direct_auto_write ?
859 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
860 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
862 fprintf( rc->p_stat_file_out,
863 "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",
864 h->fenc->i_frame, h->i_frame,
866 h->stat.frame.i_itex_bits, h->stat.frame.i_ptex_bits,
867 h->stat.frame.i_hdr_bits, h->stat.frame.i_misc_bits,
868 h->stat.frame.i_mb_count_i,
869 h->stat.frame.i_mb_count_p,
870 h->stat.frame.i_mb_count_skip,
876 if( h->sh.i_type != SLICE_TYPE_B )
877 rc->cplxr_sum += bits * qp2qscale(rc->qpa) / rc->last_rceq;
880 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
881 * Not perfectly accurate with B-refs, but good enough. */
882 rc->cplxr_sum += bits * qp2qscale(rc->qpa) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
884 rc->cplxr_sum *= rc->cbr_decay;
885 rc->wanted_bits_window += rc->bitrate / rc->fps;
886 rc->wanted_bits_window *= rc->cbr_decay;
888 if( h->param.i_threads == 1 )
889 accum_p_qp_update( h, rc->qpa );
894 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
897 if( h->mb.b_variable_qp )
899 if( h->sh.i_type == SLICE_TYPE_B )
901 rc->bframe_bits += bits;
902 if( !h->frames.current[0] || !IS_X264_TYPE_B(h->frames.current[0]->i_type) )
904 update_predictor( rc->pred_b_from_p, qp2qscale(rc->qpa),
905 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
911 update_vbv( h, bits );
914 /****************************************************************************
916 ***************************************************************************/
918 double x264_eval( char *s, double *const_value, const char **const_name,
919 double (**func1)(void *, double), const char **func1_name,
920 double (**func2)(void *, double, double), char **func2_name,
924 * modify the bitrate curve from pass1 for one frame
926 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
928 x264_ratecontrol_t *rcc= h->rc;
929 const int pict_type = rce->pict_type;
931 x264_zone_t *zone = get_zone( h, frame_num );
933 double const_values[]={
934 rce->i_tex_bits * rce->qscale,
935 rce->p_tex_bits * rce->qscale,
936 (rce->i_tex_bits + rce->p_tex_bits) * rce->qscale,
937 rce->mv_bits * rce->qscale,
938 (double)rce->i_count / rcc->nmb,
939 (double)rce->p_count / rcc->nmb,
940 (double)rce->s_count / rcc->nmb,
941 rce->pict_type == SLICE_TYPE_I,
942 rce->pict_type == SLICE_TYPE_P,
943 rce->pict_type == SLICE_TYPE_B,
944 h->param.rc.f_qcompress,
945 rcc->i_cplx_sum[SLICE_TYPE_I] / rcc->frame_count[SLICE_TYPE_I],
946 rcc->i_cplx_sum[SLICE_TYPE_P] / rcc->frame_count[SLICE_TYPE_P],
947 rcc->p_cplx_sum[SLICE_TYPE_P] / rcc->frame_count[SLICE_TYPE_P],
948 rcc->p_cplx_sum[SLICE_TYPE_B] / rcc->frame_count[SLICE_TYPE_B],
949 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / rcc->frame_count[pict_type],
950 rce->blurred_complexity,
953 static const char *const_names[]={
973 static double (*func1[])(void *, double)={
974 // (void *)bits2qscale,
978 static const char *func1_names[]={
984 q = x264_eval((char*)h->param.rc.psz_rc_eq, const_values, const_names, func1, func1_names, NULL, NULL, rce);
986 // avoid NaN's in the rc_eq
987 if(!isfinite(q) || rce->i_tex_bits + rce->p_tex_bits + rce->mv_bits == 0)
988 q = rcc->last_qscale;
992 rcc->last_qscale = q;
997 if( zone->b_force_qp )
998 q = qp2qscale(zone->i_qp);
1000 q /= zone->f_bitrate_factor;
1006 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1008 x264_ratecontrol_t *rcc = h->rc;
1009 const int pict_type = rce->pict_type;
1011 // force I/B quants as a function of P quants
1012 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1013 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1014 if( pict_type == SLICE_TYPE_I )
1017 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1018 double ip_factor = fabs( h->param.rc.f_ip_factor );
1019 /* don't apply ip_factor if the following frame is also I */
1020 if( rcc->accum_p_norm <= 0 )
1022 else if( h->param.rc.f_ip_factor < 0 )
1024 else if( rcc->accum_p_norm >= 1 )
1027 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1029 else if( pict_type == SLICE_TYPE_B )
1031 if( h->param.rc.f_pb_factor > 0 )
1033 if( !rce->kept_as_ref )
1034 q *= fabs( h->param.rc.f_pb_factor );
1036 else if( pict_type == SLICE_TYPE_P
1037 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1038 && rce->i_tex_bits + rce->p_tex_bits == 0 )
1043 /* last qscale / qdiff stuff */
1044 if(rcc->last_non_b_pict_type==pict_type
1045 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1047 double last_q = rcc->last_qscale_for[pict_type];
1048 double max_qscale = last_q * rcc->lstep;
1049 double min_qscale = last_q / rcc->lstep;
1051 if (q > max_qscale) q = max_qscale;
1052 else if(q < min_qscale) q = min_qscale;
1055 rcc->last_qscale_for[pict_type] = q;
1056 if(pict_type!=SLICE_TYPE_B)
1057 rcc->last_non_b_pict_type = pict_type;
1058 if(pict_type==SLICE_TYPE_I)
1060 rcc->last_accum_p_norm = rcc->accum_p_norm;
1061 rcc->accum_p_norm = 0;
1062 rcc->accum_p_qp = 0;
1064 if(pict_type==SLICE_TYPE_P)
1066 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1067 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1068 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1073 static double predict_size( predictor_t *p, double q, double var )
1075 return p->coeff*var / (q*p->count);
1078 static void update_predictor( predictor_t *p, double q, double var, double bits )
1082 p->count *= p->decay;
1083 p->coeff *= p->decay;
1085 p->coeff += bits*q / var;
1088 // update VBV after encoding a frame
1089 static void update_vbv( x264_t *h, int bits )
1091 x264_ratecontrol_t *rcc = h->rc;
1092 x264_ratecontrol_t *rct = h->thread[0]->rc;
1094 if( rcc->last_satd >= h->mb.i_mb_count )
1095 update_predictor( &rct->pred[h->sh.i_type], qp2qscale(rcc->qpa), rcc->last_satd, bits );
1100 rct->buffer_fill_final += rct->buffer_rate - bits;
1101 if( rct->buffer_fill_final < 0 && !rct->b_2pass )
1102 x264_log( h, X264_LOG_WARNING, "VBV underflow (%.0f bits)\n", rct->buffer_fill_final );
1103 rct->buffer_fill_final = x264_clip3( rct->buffer_fill_final, 0, rct->buffer_size );
1106 // provisionally update VBV according to the planned size of all frames currently in progress
1107 static void update_vbv_plan( x264_t *h )
1109 x264_ratecontrol_t *rcc = h->rc;
1110 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final;
1111 if( h->param.i_threads > 1 )
1113 int j = h->rc - h->thread[0]->rc;
1115 for( i=1; i<h->param.i_threads; i++ )
1117 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1118 double bits = t->rc->frame_size_planned;
1119 if( !t->b_thread_active )
1121 rcc->buffer_fill += rcc->buffer_rate - bits;
1122 rcc->buffer_fill = x264_clip3( rcc->buffer_fill, 0, rcc->buffer_size );
1127 // apply VBV constraints and clip qscale to between lmin and lmax
1128 static double clip_qscale( x264_t *h, int pict_type, double q )
1130 x264_ratecontrol_t *rcc = h->rc;
1131 double lmin = rcc->lmin[pict_type];
1132 double lmax = rcc->lmax[pict_type];
1135 /* B-frames are not directly subject to VBV,
1136 * since they are controlled by the P-frames' QPs.
1137 * FIXME: in 2pass we could modify previous frames' QP too,
1138 * instead of waiting for the buffer to fill */
1140 ( pict_type == SLICE_TYPE_P ||
1141 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) )
1143 if( rcc->buffer_fill/rcc->buffer_size < 0.5 )
1144 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1147 if( rcc->b_vbv && rcc->last_satd > 0 )
1149 /* Now a hard threshold to make sure the frame fits in VBV.
1150 * This one is mostly for I-frames. */
1151 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1153 if( bits > rcc->buffer_fill/2 )
1154 qf = x264_clip3f( rcc->buffer_fill/(2*bits), 0.2, 1.0 );
1157 if( bits < rcc->buffer_rate/2 )
1158 q *= bits*2/rcc->buffer_rate;
1159 q = X264_MAX( q0, q );
1161 /* Check B-frame complexity, and use up any bits that would
1162 * overflow before the next P-frame. */
1163 if( h->sh.i_type == SLICE_TYPE_P )
1165 int nb = rcc->bframes;
1166 double pbbits = bits;
1167 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1170 if( bbits > rcc->buffer_rate )
1172 pbbits += nb * bbits;
1174 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1175 if( pbbits < space )
1177 q *= X264_MAX( pbbits / space,
1178 bits / (0.5 * rcc->buffer_size) );
1180 q = X264_MAX( q0-5, q );
1183 if( !rcc->b_vbv_min_rate )
1184 q = X264_MAX( q0, q );
1189 else if(rcc->b_2pass)
1191 double min2 = log(lmin);
1192 double max2 = log(lmax);
1193 q = (log(q) - min2)/(max2-min2) - 0.5;
1194 q = 1.0/(1.0 + exp(-4*q));
1195 q = q*(max2-min2) + min2;
1199 return x264_clip3f(q, lmin, lmax);
1202 // update qscale for 1 frame based on actual bits used so far
1203 static float rate_estimate_qscale( x264_t *h )
1206 x264_ratecontrol_t *rcc = h->rc;
1207 ratecontrol_entry_t rce;
1208 int pict_type = h->sh.i_type;
1209 double lmin = rcc->lmin[pict_type];
1210 double lmax = rcc->lmax[pict_type];
1211 int64_t total_bits = 8*(h->stat.i_slice_size[SLICE_TYPE_I]
1212 + h->stat.i_slice_size[SLICE_TYPE_P]
1213 + h->stat.i_slice_size[SLICE_TYPE_B]);
1218 if(pict_type != rce.pict_type)
1220 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1221 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1225 if( pict_type == SLICE_TYPE_B )
1227 /* B-frames don't have independent ratecontrol, but rather get the
1228 * average QP of the two adjacent P-frames + an offset */
1230 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1231 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1232 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1233 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1234 float q0 = h->fref0[0]->f_qp_avg;
1235 float q1 = h->fref1[0]->f_qp_avg;
1237 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1238 q0 -= rcc->pb_offset/2;
1239 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1240 q1 -= rcc->pb_offset/2;
1243 q = (q0 + q1) / 2 + rcc->ip_offset;
1249 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1251 if(h->fenc->b_kept_as_ref)
1252 q += rcc->pb_offset/2;
1254 q += rcc->pb_offset;
1256 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1258 return qp2qscale(q);
1262 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1265 //FIXME adjust abr_buffer based on distance to the end of the video
1266 int64_t diff = total_bits - (int64_t)rce.expected_bits;
1268 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1269 if( h->fenc->i_frame > 30 )
1271 /* Adjust quant based on the difference between
1272 * achieved and expected bitrate so far */
1273 double time = (double)h->fenc->i_frame / rcc->num_entries;
1274 double w = x264_clip3f( time*100, 0.0, 1.0 );
1275 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1277 q = x264_clip3f( q, lmin, lmax );
1279 else /* 1pass ABR */
1281 /* Calculate the quantizer which would have produced the desired
1282 * average bitrate if it had been applied to all frames so far.
1283 * Then modulate that quant based on the current frame's complexity
1284 * relative to the average complexity so far (using the 2pass RCEQ).
1285 * Then bias the quant up or down if total size so far was far from
1287 * Result: Depending on the value of rate_tolerance, there is a
1288 * tradeoff between quality and bitrate precision. But at large
1289 * tolerances, the bit distribution approaches that of 2pass. */
1291 double wanted_bits, overflow=1, lmin, lmax;
1293 rcc->last_satd = x264_rc_analyse_slice( h );
1294 rcc->short_term_cplxsum *= 0.5;
1295 rcc->short_term_cplxcount *= 0.5;
1296 rcc->short_term_cplxsum += rcc->last_satd;
1297 rcc->short_term_cplxcount ++;
1299 rce.p_tex_bits = rcc->last_satd;
1300 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1303 rce.p_count = rcc->nmb;
1307 rce.pict_type = pict_type;
1309 if( h->param.rc.i_rc_method == X264_RC_CRF )
1311 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1315 int i_frame_done = h->fenc->i_frame + 1 - h->param.i_threads;
1317 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1319 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
1320 wanted_bits = i_frame_done * rcc->bitrate / rcc->fps;
1321 if( wanted_bits > 0 )
1323 abr_buffer *= X264_MAX( 1, sqrt(i_frame_done/25) );
1324 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1329 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1330 /* should test _next_ pict type, but that isn't decided yet */
1331 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1333 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1334 q /= fabs( h->param.rc.f_ip_factor );
1336 else if( h->i_frame > 0 )
1338 /* Asymmetric clipping, because symmetric would prevent
1339 * overflow control in areas of rapidly oscillating complexity */
1340 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1341 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1342 if( overflow > 1.1 && h->i_frame > 3 )
1344 else if( overflow < 0.9 )
1347 q = x264_clip3f(q, lmin, lmax);
1349 else if( h->param.rc.i_rc_method == X264_RC_CRF )
1351 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1354 //FIXME use get_diff_limited_q() ?
1355 q = clip_qscale( h, pict_type, q );
1358 rcc->last_qscale_for[pict_type] =
1359 rcc->last_qscale = q;
1361 if( !rcc->b_2pass && h->fenc->i_frame == 0 )
1362 rcc->last_qscale_for[SLICE_TYPE_P] = q;
1364 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1369 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
1373 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
1374 /* these vars are updated in x264_ratecontrol_start()
1375 * so copy them from the context that most recently started (prev)
1376 * to the context that's about to start (cur).
1382 COPY(last_qscale_for);
1383 COPY(last_non_b_pict_type);
1384 COPY(short_term_cplxsum);
1385 COPY(short_term_cplxcount);
1391 #define COPY(var) next->rc->var = cur->rc->var
1392 /* these vars are updated in x264_ratecontrol_end()
1393 * so copy them from the context that most recently ended (cur)
1394 * to the context that's about to end (next)
1397 COPY(expected_bits_sum);
1398 COPY(wanted_bits_window);
1402 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
1403 /* the rest of the variables are either constant or thread-local */
1406 static int init_pass2( x264_t *h )
1408 x264_ratecontrol_t *rcc = h->rc;
1409 uint64_t all_const_bits = 0;
1410 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000 * (double)rcc->num_entries / rcc->fps);
1411 double rate_factor, step, step_mult;
1412 double qblur = h->param.rc.f_qblur;
1413 double cplxblur = h->param.rc.f_complexity_blur;
1414 const int filter_size = (int)(qblur*4) | 1;
1415 double expected_bits;
1416 double *qscale, *blurred_qscale;
1419 /* find total/average complexity & const_bits */
1420 for(i=0; i<rcc->num_entries; i++){
1421 ratecontrol_entry_t *rce = &rcc->entry[i];
1422 all_const_bits += rce->misc_bits;
1423 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
1424 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
1425 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits * rce->qscale;
1426 rcc->frame_count[rce->pict_type] ++;
1429 if( all_available_bits < all_const_bits)
1431 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
1432 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000)));
1436 /* Blur complexities, to reduce local fluctuation of QP.
1437 * We don't blur the QPs directly, because then one very simple frame
1438 * could drag down the QP of a nearby complex frame and give it more
1439 * bits than intended. */
1440 for(i=0; i<rcc->num_entries; i++){
1441 ratecontrol_entry_t *rce = &rcc->entry[i];
1442 double weight_sum = 0;
1443 double cplx_sum = 0;
1444 double weight = 1.0;
1446 /* weighted average of cplx of future frames */
1447 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++){
1448 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
1449 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1452 weight_sum += weight;
1453 cplx_sum += weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1455 /* weighted average of cplx of past frames */
1457 for(j=0; j<=cplxblur*2 && j<=i; j++){
1458 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
1459 weight_sum += weight;
1460 cplx_sum += weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1461 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1465 rce->blurred_complexity = cplx_sum / weight_sum;
1468 qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1470 blurred_qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1472 blurred_qscale = qscale;
1474 /* Search for a factor which, when multiplied by the RCEQ values from
1475 * each frame, adds up to the desired total size.
1476 * There is no exact closed-form solution because of VBV constraints and
1477 * because qscale2bits is not invertible, but we can start with the simple
1478 * approximation of scaling the 1st pass by the ratio of bitrates.
1479 * The search range is probably overkill, but speed doesn't matter here. */
1482 for(i=0; i<rcc->num_entries; i++)
1483 expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i));
1484 step_mult = all_available_bits / expected_bits;
1487 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5){
1489 rate_factor += step;
1491 rcc->last_non_b_pict_type = -1;
1492 rcc->last_accum_p_norm = 1;
1493 rcc->accum_p_norm = 0;
1494 rcc->buffer_fill = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
1497 for(i=0; i<rcc->num_entries; i++){
1498 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
1501 /* fixed I/B qscale relative to P */
1502 for(i=rcc->num_entries-1; i>=0; i--){
1503 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
1504 assert(qscale[i] >= 0);
1508 if(filter_size > 1){
1509 assert(filter_size%2==1);
1510 for(i=0; i<rcc->num_entries; i++){
1511 ratecontrol_entry_t *rce = &rcc->entry[i];
1513 double q=0.0, sum=0.0;
1515 for(j=0; j<filter_size; j++){
1516 int index = i+j-filter_size/2;
1518 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
1519 if(index < 0 || index >= rcc->num_entries) continue;
1520 if(rce->pict_type != rcc->entry[index].pict_type) continue;
1521 q += qscale[index] * coeff;
1524 blurred_qscale[i] = q/sum;
1528 /* find expected bits */
1529 for(i=0; i<rcc->num_entries; i++){
1530 ratecontrol_entry_t *rce = &rcc->entry[i];
1532 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
1533 assert(rce->new_qscale >= 0);
1534 bits = qscale2bits(rce, rce->new_qscale);
1536 rce->expected_bits = expected_bits;
1537 expected_bits += bits;
1538 update_vbv(h, bits);
1539 rcc->buffer_fill = rcc->buffer_fill_final;
1542 //printf("expected:%llu available:%llu factor:%lf avgQ:%lf\n", (uint64_t)expected_bits, all_available_bits, rate_factor);
1543 if(expected_bits > all_available_bits) rate_factor -= step;
1548 x264_free(blurred_qscale);
1550 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
1553 for(i=0; i<rcc->num_entries; i++)
1554 avgq += rcc->entry[i].new_qscale;
1555 avgq = qscale2qp(avgq / rcc->num_entries);
1557 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
1558 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
1559 (float)h->param.rc.i_bitrate,
1560 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
1562 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
1564 if(h->param.rc.i_qp_min > 0)
1565 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
1567 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
1569 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
1571 if(h->param.rc.i_qp_max < 51)
1572 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
1574 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
1577 x264_log(h, X264_LOG_WARNING, "internal error\n");