1 /***************************************************-*- coding: iso-8859-1 -*-
2 * ratecontrol.c: h264 encoder library (Rate Control)
3 *****************************************************************************
4 * Copyright (C) 2005-2008 x264 project
6 * Authors: Loren Merritt <lorenm@u.washington.edu>
7 * Michael Niedermayer <michaelni@gmx.at>
8 * Gabriel Bouvigne <gabriel.bouvigne@joost.com>
9 * Fiona Glaser <fiona@x264.com>
10 * Måns Rullgård <mru@mru.ath.cx>
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License as published by
14 * the Free Software Foundation; either version 2 of the License, or
15 * (at your option) any later version.
17 * This program is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 * GNU General Public License for more details.
22 * You should have received a copy of the GNU General Public License
23 * along with this program; if not, write to the Free Software
24 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
25 *****************************************************************************/
27 #define _ISOC99_SOURCE
28 #undef NDEBUG // always check asserts, the speed effect is far too small to disable them
31 #include "common/common.h"
32 #include "common/cpu.h"
33 #include "ratecontrol.h"
44 uint64_t expected_bits; /*total expected bits up to the current frame (current one excluded)*/
51 float blurred_complexity;
55 } ratecontrol_entry_t;
65 struct x264_ratecontrol_t
74 double rate_tolerance;
76 int nmb; /* number of macroblocks in a frame */
80 ratecontrol_entry_t *rce;
81 int qp; /* qp for current frame */
82 int qpm; /* qp for current macroblock */
83 float f_qpm; /* qp for current macroblock: precise float for AQ */
84 float qpa_rc; /* average of macroblocks' qp before aq */
85 float qpa_aq; /* average of macroblocks' qp after aq */
86 float qp_novbv; /* QP for the current frame if 1-pass VBV was disabled. */
91 double buffer_fill_final; /* real buffer as of the last finished frame */
92 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
93 double buffer_rate; /* # of bits added to buffer_fill after each frame */
94 predictor_t *pred; /* predict frame size from satd */
100 double cplxr_sum; /* sum of bits*qscale/rceq */
101 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow, only includes finished frames */
102 double wanted_bits_window; /* target bitrate * window */
104 double short_term_cplxsum;
105 double short_term_cplxcount;
106 double rate_factor_constant;
111 FILE *p_stat_file_out;
112 char *psz_stat_file_tmpname;
113 FILE *p_mbtree_stat_file_out;
114 char *psz_mbtree_stat_file_tmpname;
115 char *psz_mbtree_stat_file_name;
116 FILE *p_mbtree_stat_file_in;
118 int num_entries; /* number of ratecontrol_entry_ts */
119 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
121 double last_qscale_for[5]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
122 int last_non_b_pict_type;
123 double accum_p_qp; /* for determining I-frame quant */
125 double last_accum_p_norm;
126 double lmin[5]; /* min qscale by frame type */
128 double lstep; /* max change (multiply) in qscale per frame */
129 uint16_t *qp_buffer; /* Global buffer for converting MB-tree quantizer data. */
132 double frame_size_estimated;
133 double frame_size_planned;
134 predictor_t (*row_pred)[2];
135 predictor_t row_preds[5][2];
136 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
137 int bframes; /* # consecutive B-frames before this P-frame */
138 int bframe_bits; /* total cost of those frames */
142 x264_zone_t *prev_zone;
146 static int parse_zones( x264_t *h );
147 static int init_pass2(x264_t *);
148 static float rate_estimate_qscale( x264_t *h );
149 static void update_vbv( x264_t *h, int bits );
150 static void update_vbv_plan( x264_t *h, int overhead );
151 static double predict_size( predictor_t *p, double q, double var );
152 static void update_predictor( predictor_t *p, double q, double var, double bits );
155 * qp = h.264's quantizer
156 * qscale = linearized quantizer = Lagrange multiplier
158 static inline double qp2qscale(double qp)
160 return 0.85 * pow(2.0, ( qp - 12.0 ) / 6.0);
162 static inline double qscale2qp(double qscale)
164 return 12.0 + 6.0 * log(qscale/0.85) / log(2.0);
167 /* Texture bitrate is not quite inversely proportional to qscale,
168 * probably due the the changing number of SKIP blocks.
169 * MV bits level off at about qp<=12, because the lambda used
170 * for motion estimation is constant there. */
171 static inline double qscale2bits(ratecontrol_entry_t *rce, double qscale)
175 return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
176 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
180 // Find the total AC energy of the block in all planes.
181 static NOINLINE uint32_t ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
183 /* This function contains annoying hacks because GCC has a habit of reordering emms
184 * and putting it after floating point ops. As a result, we put the emms at the end of the
185 * function and make sure that its always called before the float math. Noinline makes
186 * sure no reordering goes on. */
188 for( i = 0; i < 3; i++ )
191 int stride = frame->i_stride[i];
192 int offset = h->mb.b_interlaced
193 ? w * (mb_x + (mb_y&~1) * stride) + (mb_y&1) * stride
194 : w * (mb_x + mb_y * stride);
195 int pix = i ? PIXEL_8x8 : PIXEL_16x16;
196 stride <<= h->mb.b_interlaced;
197 var += h->pixf.var[pix]( frame->plane[i]+offset, stride );
203 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame )
205 /* constants chosen to result in approximately the same overall bitrate as without AQ.
206 * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
210 /* Need to init it anyways for MB tree. */
211 if( h->param.rc.f_aq_strength == 0 )
214 memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
215 memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
216 if( h->frames.b_have_lowres )
217 for( mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
218 frame->i_inv_qscale_factor[mb_xy] = 256;
222 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
224 for( mb_y = 0; mb_y < h->sps->i_mb_height; mb_y++ )
225 for( mb_x = 0; mb_x < h->sps->i_mb_width; mb_x++ )
227 uint32_t energy = ac_energy_mb( h, mb_x, mb_y, frame );
228 float qp_adj = x264_log2( energy + 2 );
230 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
233 avg_adj /= h->mb.i_mb_count;
234 strength = h->param.rc.f_aq_strength * avg_adj * (1.f / 6000.f);
237 strength = h->param.rc.f_aq_strength * 1.0397f;
239 for( mb_y = 0; mb_y < h->sps->i_mb_height; mb_y++ )
240 for( mb_x = 0; mb_x < h->sps->i_mb_width; mb_x++ )
243 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
245 qp_adj = frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride];
246 qp_adj = strength * (qp_adj - avg_adj);
250 uint32_t energy = ac_energy_mb( h, mb_x, mb_y, frame );
251 qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - 14.427f);
253 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] =
254 frame->f_qp_offset_aq[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
255 if( h->frames.b_have_lowres )
256 frame->i_inv_qscale_factor[mb_x + mb_y*h->mb.i_mb_stride] = x264_exp2fix8(qp_adj);
261 /*****************************************************************************
262 * x264_adaptive_quant:
263 * adjust macroblock QP based on variance (AC energy) of the MB.
264 * high variance = higher QP
265 * low variance = lower QP
266 * This generally increases SSIM and lowers PSNR.
267 *****************************************************************************/
268 void x264_adaptive_quant( x264_t *h )
271 /* MB-tree currently doesn't adjust quantizers in B-frames. */
272 float qp_offset = h->sh.i_type == SLICE_TYPE_B ? h->fenc->f_qp_offset_aq[h->mb.i_mb_xy] : h->fenc->f_qp_offset[h->mb.i_mb_xy];
273 h->mb.i_qp = x264_clip3( h->rc->f_qpm + qp_offset + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
276 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame )
278 x264_ratecontrol_t *rc = h->rc;
279 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
282 if( i_type_actual != SLICE_TYPE_B )
286 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
289 if( i_type != i_type_actual )
291 x264_log(h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type,i_type_actual);
295 if( fread( rc->qp_buffer, sizeof(uint16_t), h->mb.i_mb_count, rc->p_mbtree_stat_file_in ) != h->mb.i_mb_count )
298 for( i = 0; i < h->mb.i_mb_count; i++ )
300 frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[i] )) * (1/256.0);
301 if( h->frames.b_have_lowres )
302 frame->i_inv_qscale_factor[i] = x264_exp2fix8(frame->f_qp_offset[i]);
306 x264_adaptive_quant_frame( h, frame );
309 x264_log(h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n");
313 int x264_reference_build_list_optimal( x264_t *h )
315 ratecontrol_entry_t *rce = h->rc->rce;
316 x264_frame_t *frames[16];
319 if( rce->refs != h->i_ref0 )
322 memcpy( frames, h->fref0, sizeof(frames) );
324 /* For now don't reorder ref 0; it seems to lower quality
325 in most cases due to skips. */
326 for( ref = 1; ref < h->i_ref0; ref++ )
330 for( i = 1; i < h->i_ref0; i++ )
331 /* Favor lower POC as a tiebreaker. */
332 COPY2_IF_GT( max, rce->refcount[i], bestref, i );
333 rce->refcount[bestref] = -1;
334 h->fref0[ref] = frames[bestref];
340 static char *x264_strcat_filename( char *input, char *suffix )
342 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
345 strcpy( output, input );
346 strcat( output, suffix );
350 int x264_ratecontrol_new( x264_t *h )
352 x264_ratecontrol_t *rc;
357 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
360 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
361 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
363 /* FIXME: use integers */
364 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
365 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
369 if( h->param.rc.b_mb_tree )
371 h->param.rc.f_pb_factor = 1;
375 rc->qcompress = h->param.rc.f_qcompress;
377 rc->bitrate = h->param.rc.i_bitrate * 1000.;
378 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
379 rc->nmb = h->mb.i_mb_count;
380 rc->last_non_b_pict_type = -1;
383 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
385 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
388 if( h->param.rc.i_vbv_buffer_size )
390 if( h->param.rc.i_rc_method == X264_RC_CQP )
392 x264_log(h, X264_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
393 h->param.rc.i_vbv_max_bitrate = 0;
394 h->param.rc.i_vbv_buffer_size = 0;
396 else if( h->param.rc.i_vbv_max_bitrate == 0 )
398 x264_log( h, X264_LOG_DEBUG, "VBV maxrate unspecified, assuming CBR\n" );
399 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
402 if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate &&
403 h->param.rc.i_vbv_max_bitrate > 0)
404 x264_log(h, X264_LOG_WARNING, "max bitrate less than average bitrate, ignored.\n");
405 else if( h->param.rc.i_vbv_max_bitrate > 0 &&
406 h->param.rc.i_vbv_buffer_size > 0 )
408 if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
410 h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
411 x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
412 h->param.rc.i_vbv_buffer_size );
414 if( h->param.rc.f_vbv_buffer_init > 1. )
415 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 );
416 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000. / rc->fps;
417 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000.;
418 rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
419 h->param.rc.f_vbv_buffer_init = X264_MAX( h->param.rc.f_vbv_buffer_init, rc->buffer_rate / rc->buffer_size );
420 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
421 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
422 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
424 rc->b_vbv_min_rate = !rc->b_2pass
425 && h->param.rc.i_rc_method == X264_RC_ABR
426 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
428 else if( h->param.rc.i_vbv_max_bitrate )
430 x264_log(h, X264_LOG_WARNING, "VBV maxrate specified, but no bufsize.\n");
431 h->param.rc.i_vbv_max_bitrate = 0;
433 if(rc->rate_tolerance < 0.01)
435 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
436 rc->rate_tolerance = 0.01;
439 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
443 /* FIXME ABR_INIT_QP is actually used only in CRF */
444 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
445 rc->accum_p_norm = .01;
446 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
447 /* estimated ratio that produces a reasonable QP for the first I-frame */
448 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
449 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
450 rc->last_non_b_pict_type = SLICE_TYPE_I;
453 if( h->param.rc.i_rc_method == X264_RC_CRF )
455 /* Arbitrary rescaling to make CRF somewhat similar to QP.
456 * Try to compensate for MB-tree's effects as well. */
457 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
458 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
459 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
460 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset );
463 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
464 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
465 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
466 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
467 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
469 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
470 rc->last_qscale = qp2qscale(26);
471 CHECKED_MALLOC( rc->pred, 5*sizeof(predictor_t) );
472 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
473 for( i = 0; i < 5; i++ )
475 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
476 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
477 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
478 rc->pred[i].coeff= 2.0;
479 rc->pred[i].count= 1.0;
480 rc->pred[i].decay= 0.5;
481 rc->pred[i].offset= 0.0;
482 for( j = 0; j < 2; j++ )
484 rc->row_preds[i][j].coeff= .25;
485 rc->row_preds[i][j].count= 1.0;
486 rc->row_preds[i][j].decay= 0.5;
487 rc->row_preds[i][j].offset= 0.0;
490 *rc->pred_b_from_p = rc->pred[0];
492 if( parse_zones( h ) < 0 )
494 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
498 /* Load stat file and init 2pass algo */
499 if( h->param.rc.b_stat_read )
501 char *p, *stats_in, *stats_buf;
503 /* read 1st pass stats */
504 assert( h->param.rc.psz_stat_in );
505 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
508 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
511 if( h->param.rc.b_mb_tree )
513 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
514 if( !mbtree_stats_in )
516 rc->p_mbtree_stat_file_in = fopen( mbtree_stats_in, "rb" );
517 x264_free( mbtree_stats_in );
518 if( !rc->p_mbtree_stat_file_in )
520 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
525 /* check whether 1st pass options were compatible with current options */
526 if( !strncmp( stats_buf, "#options:", 9 ) )
529 char *opts = stats_buf;
530 stats_in = strchr( stats_buf, '\n' );
536 if( ( p = strstr( opts, "bframes=" ) ) && sscanf( p, "bframes=%d", &i )
537 && h->param.i_bframe != i )
539 x264_log( h, X264_LOG_ERROR, "different number of B-frames than 1st pass (%d vs %d)\n",
540 h->param.i_bframe, i );
544 /* since B-adapt doesn't (yet) take into account B-pyramid,
545 * the converse is not a problem */
546 if( h->param.i_bframe )
549 sprintf( buf, "b_pyramid=%d", h->param.i_bframe_pyramid );
550 if( !strstr( opts, buf ) )
551 x264_log( h, X264_LOG_WARNING, "different B-pyramid setting than 1st pass\n" );
554 if( ( p = strstr( opts, "keyint=" ) ) && sscanf( p, "keyint=%d", &i )
555 && h->param.i_keyint_max != i )
556 x264_log( h, X264_LOG_WARNING, "different keyint than 1st pass (%d vs %d)\n",
557 h->param.i_keyint_max, i );
559 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
560 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
562 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
564 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
565 h->mb.b_direct_auto_write = 1;
568 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
569 h->param.i_bframe_adaptive = i;
570 else if( h->param.i_bframe )
572 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
576 if( h->param.rc.b_mb_tree && ( p = strstr( opts, "rc_lookahead=" ) ) && sscanf( p, "rc_lookahead=%d", &i ) )
577 h->param.rc.i_lookahead = i;
580 /* find number of pics */
583 p = strchr(p+1, ';');
586 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
591 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
593 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
594 h->param.i_frame_total, rc->num_entries );
596 if( h->param.i_frame_total > rc->num_entries )
598 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
599 h->param.i_frame_total, rc->num_entries );
603 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
605 /* init all to skipped p frames */
606 for(i=0; i<rc->num_entries; i++)
608 ratecontrol_entry_t *rce = &rc->entry[i];
609 rce->pict_type = SLICE_TYPE_P;
610 rce->qscale = rce->new_qscale = qp2qscale(20);
611 rce->misc_bits = rc->nmb + 10;
617 for(i=0; i < rc->num_entries; i++)
619 ratecontrol_entry_t *rce;
627 next= strchr(p, ';');
630 (*next)=0; //sscanf is unbelievably slow on long strings
633 e = sscanf(p, " in:%d ", &frame_number);
635 if(frame_number < 0 || frame_number >= rc->num_entries)
637 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
640 rce = &rc->entry[frame_number];
641 rce->direct_mode = 0;
643 e += sscanf(p, " in:%*d out:%*d type:%c q:%f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c",
644 &pict_type, &qp, &rce->tex_bits,
645 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
646 &rce->s_count, &rce->direct_mode);
648 p = strstr( p, "ref:" );
652 for( ref = 0; ref < 16; ref++ )
654 if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
656 p = strchr( p+1, ' ' );
664 case 'I': rce->kept_as_ref = 1;
665 case 'i': rce->pict_type = SLICE_TYPE_I; break;
666 case 'P': rce->pict_type = SLICE_TYPE_P; break;
667 case 'B': rce->kept_as_ref = 1;
668 case 'b': rce->pict_type = SLICE_TYPE_B; break;
669 default: e = -1; break;
674 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
677 rce->qscale = qp2qscale(qp);
681 x264_free(stats_buf);
683 if(h->param.rc.i_rc_method == X264_RC_ABR)
685 if(init_pass2(h) < 0) return -1;
686 } /* else we're using constant quant, so no need to run the bitrate allocation */
689 /* Open output file */
690 /* If input and output files are the same, output to a temp file
691 * and move it to the real name only when it's complete */
692 if( h->param.rc.b_stat_write )
695 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
696 if( !rc->psz_stat_file_tmpname )
699 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
700 if( rc->p_stat_file_out == NULL )
702 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
706 p = x264_param2string( &h->param, 1 );
708 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
710 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
712 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
713 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
714 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
717 rc->p_mbtree_stat_file_out = fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
718 if( rc->p_mbtree_stat_file_out == NULL )
720 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
726 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
727 CHECKED_MALLOC( rc->qp_buffer, h->mb.i_mb_count * sizeof(uint16_t) );
729 for( i=0; i<h->param.i_threads; i++ )
731 h->thread[i]->rc = rc+i;
735 h->thread[i]->param = h->param;
736 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
745 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
748 char *tok, UNUSED *saveptr;
750 z->f_bitrate_factor = 1;
751 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
753 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
755 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
759 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
765 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
766 memcpy( z->param, &h->param, sizeof(x264_param_t) );
767 z->param->param_free = x264_free;
768 while( (tok = strtok_r( p, ",", &saveptr )) )
770 char *val = strchr( tok, '=' );
776 if( x264_param_parse( z->param, tok, val ) )
778 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
788 static int parse_zones( x264_t *h )
790 x264_ratecontrol_t *rc = h->rc;
792 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
794 char *psz_zones, *p, *tok, UNUSED *saveptr;
795 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
796 strcpy( psz_zones, h->param.rc.psz_zones );
797 h->param.rc.i_zones = 1;
798 for( p = psz_zones; *p; p++ )
799 h->param.rc.i_zones += (*p == '/');
800 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
802 for( i = 0; i < h->param.rc.i_zones; i++ )
804 tok = strtok_r( p, "/", &saveptr );
805 if( !tok || parse_zone( h, &h->param.rc.zones[i], tok ) )
809 x264_free( psz_zones );
812 if( h->param.rc.i_zones > 0 )
814 for( i = 0; i < h->param.rc.i_zones; i++ )
816 x264_zone_t z = h->param.rc.zones[i];
817 if( z.i_start < 0 || z.i_start > z.i_end )
819 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
820 z.i_start, z.i_end );
823 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
825 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
826 z.f_bitrate_factor );
831 rc->i_zones = h->param.rc.i_zones + 1;
832 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
833 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
835 // default zone to fall back to if none of the others match
836 rc->zones[0].i_start = 0;
837 rc->zones[0].i_end = INT_MAX;
838 rc->zones[0].b_force_qp = 0;
839 rc->zones[0].f_bitrate_factor = 1;
840 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
841 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
842 for( i = 1; i < rc->i_zones; i++ )
844 if( !rc->zones[i].param )
845 rc->zones[i].param = rc->zones[0].param;
854 static x264_zone_t *get_zone( x264_t *h, int frame_num )
857 for( i = h->rc->i_zones-1; i >= 0; i-- )
859 x264_zone_t *z = &h->rc->zones[i];
860 if( frame_num >= z->i_start && frame_num <= z->i_end )
866 void x264_ratecontrol_summary( x264_t *h )
868 x264_ratecontrol_t *rc = h->rc;
869 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
871 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
872 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
873 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
874 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
875 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset );
879 void x264_ratecontrol_delete( x264_t *h )
881 x264_ratecontrol_t *rc = h->rc;
884 if( rc->p_stat_file_out )
886 fclose( rc->p_stat_file_out );
887 if( h->i_frame >= rc->num_entries )
888 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
890 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
891 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
893 x264_free( rc->psz_stat_file_tmpname );
895 if( rc->p_mbtree_stat_file_out )
897 fclose( rc->p_mbtree_stat_file_out );
898 if( h->i_frame >= rc->num_entries )
899 if( rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
901 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
902 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
904 x264_free( rc->psz_mbtree_stat_file_tmpname );
905 x264_free( rc->psz_mbtree_stat_file_name );
907 if( rc->p_mbtree_stat_file_in )
908 fclose( rc->p_mbtree_stat_file_in );
909 x264_free( rc->pred );
910 x264_free( rc->pred_b_from_p );
911 x264_free( rc->entry );
912 x264_free( rc->qp_buffer );
915 x264_free( rc->zones[0].param );
916 for( i=1; i<rc->i_zones; i++ )
917 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
918 rc->zones[i].param->param_free( rc->zones[i].param );
919 x264_free( rc->zones );
924 void x264_ratecontrol_set_estimated_size( x264_t *h, int bits )
926 x264_pthread_mutex_lock( &h->fenc->mutex );
927 h->rc->frame_size_estimated = bits;
928 x264_pthread_mutex_unlock( &h->fenc->mutex );
931 int x264_ratecontrol_get_estimated_size( x264_t const *h)
934 x264_pthread_mutex_lock( &h->fenc->mutex );
935 size = h->rc->frame_size_estimated;
936 x264_pthread_mutex_unlock( &h->fenc->mutex );
940 static void accum_p_qp_update( x264_t *h, float qp )
942 x264_ratecontrol_t *rc = h->rc;
943 rc->accum_p_qp *= .95;
944 rc->accum_p_norm *= .95;
945 rc->accum_p_norm += 1;
946 if( h->sh.i_type == SLICE_TYPE_I )
947 rc->accum_p_qp += qp + rc->ip_offset;
949 rc->accum_p_qp += qp;
952 /* Before encoding a frame, choose a QP for it */
953 void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
955 x264_ratecontrol_t *rc = h->rc;
956 ratecontrol_entry_t *rce = NULL;
957 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
962 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
963 x264_encoder_reconfig( h, zone->param );
964 rc->prev_zone = zone;
966 rc->qp_force = i_force_qp;
968 if( h->param.rc.b_stat_read )
970 int frame = h->fenc->i_frame;
971 assert( frame >= 0 && frame < rc->num_entries );
972 rce = h->rc->rce = &h->rc->entry[frame];
974 if( h->sh.i_type == SLICE_TYPE_B
975 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
977 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
978 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
984 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
985 rc->row_pred = &rc->row_preds[h->sh.i_type];
986 update_vbv_plan( h, overhead );
989 if( h->sh.i_type != SLICE_TYPE_B )
990 rc->bframes = h->fenc->i_bframes;
998 q = qscale2qp( rate_estimate_qscale( h ) );
1000 else if( rc->b_2pass )
1002 rce->new_qscale = rate_estimate_qscale( h );
1003 q = qscale2qp( rce->new_qscale );
1007 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
1008 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
1010 q = rc->qp_constant[ h->sh.i_type ];
1014 if( zone->b_force_qp )
1015 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1017 q -= 6*log(zone->f_bitrate_factor)/log(2);
1021 q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1025 h->fdec->f_qp_avg_rc =
1026 h->fdec->f_qp_avg_aq =
1028 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
1031 rce->new_qp = rc->qp;
1033 accum_p_qp_update( h, rc->qp );
1035 if( h->sh.i_type != SLICE_TYPE_B )
1036 rc->last_non_b_pict_type = h->sh.i_type;
1039 static double predict_row_size( x264_t *h, int y, int qp )
1041 /* average between two predictors:
1042 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1043 x264_ratecontrol_t *rc = h->rc;
1044 double pred_s = predict_size( rc->row_pred[0], qp2qscale(qp), h->fdec->i_row_satd[y] );
1046 if( h->sh.i_type == SLICE_TYPE_I || qp >= h->fref0[0]->i_row_qp[y] )
1048 if( h->sh.i_type == SLICE_TYPE_P
1049 && h->fref0[0]->i_type == h->fdec->i_type
1050 && h->fref0[0]->i_row_satd[y] > 0
1051 && (abs(h->fref0[0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1053 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
1054 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
1058 return (pred_s + pred_t) / 2;
1060 /* Our QP is lower than the reference! */
1063 double newq = qp2qscale(qp);
1064 double oldq = qp2qscale(h->fref0[0]->i_row_qp[y]);
1065 double pred_intra = predict_size( rc->row_pred[1], (1 - newq / oldq) * newq, h->fdec->i_row_satds[0][0][y] );
1066 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1067 return pred_intra + pred_s;
1071 static double row_bits_so_far( x264_t *h, int y )
1075 for( i = 0; i <= y; i++ )
1076 bits += h->fdec->i_row_bits[i];
1080 static double predict_row_size_sum( x264_t *h, int y, int qp )
1083 double bits = row_bits_so_far(h, y);
1084 for( i = y+1; i < h->sps->i_mb_height; i++ )
1085 bits += predict_row_size( h, i, qp );
1090 void x264_ratecontrol_mb( x264_t *h, int bits )
1092 x264_ratecontrol_t *rc = h->rc;
1093 const int y = h->mb.i_mb_y;
1097 h->fdec->i_row_bits[y] += bits;
1098 rc->qpa_rc += rc->f_qpm;
1099 rc->qpa_aq += h->mb.i_qp;
1101 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !rc->b_vbv )
1104 h->fdec->i_row_qp[y] = rc->qpm;
1106 update_predictor( rc->row_pred[0], qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1107 if( h->sh.i_type == SLICE_TYPE_P && rc->qpm < h->fref0[0]->i_row_qp[y] )
1109 double newq = qp2qscale(rc->qpm);
1110 double oldq = qp2qscale(h->fref0[0]->i_row_qp[y]);
1111 update_predictor( rc->row_pred[1], (1 - newq / oldq) * newq, h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1114 /* tweak quality based on difference from predicted size */
1115 if( y < h->sps->i_mb_height-1 )
1117 int prev_row_qp = h->fdec->i_row_qp[y];
1118 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
1119 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1121 /* B-frames shouldn't use lower QP than their reference frames. */
1122 if( h->sh.i_type == SLICE_TYPE_B )
1124 i_qp_min = X264_MAX( i_qp_min, X264_MAX( h->fref0[0]->i_row_qp[y+1], h->fref1[0]->i_row_qp[y+1] ) );
1125 rc->qpm = X264_MAX( rc->qpm, i_qp_min );
1128 int b0 = predict_row_size_sum( h, y, rc->qpm );
1130 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1132 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1133 float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
1135 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1136 /* area at the top of the frame was measured inaccurately. */
1137 if( row_bits_so_far(h,y) < 0.05 * rc->frame_size_planned )
1140 if( h->sh.i_type != SLICE_TYPE_I )
1143 if( !rc->b_vbv_min_rate )
1144 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
1146 while( rc->qpm < i_qp_max
1147 && ((b1 > rc->frame_size_planned + rc_tol) ||
1148 (rc->buffer_fill - b1 < buffer_left_planned * 0.5) ||
1149 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1152 b1 = predict_row_size_sum( h, y, rc->qpm );
1155 while( rc->qpm > i_qp_min
1156 && (rc->qpm > h->fdec->i_row_qp[0] || rc->single_frame_vbv)
1157 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
1158 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
1161 b1 = predict_row_size_sum( h, y, rc->qpm );
1164 /* avoid VBV underflow */
1165 while( (rc->qpm < h->param.rc.i_qp_max)
1166 && (rc->buffer_fill - b1 < rc->buffer_rate * 0.05 ) )
1169 b1 = predict_row_size_sum( h, y, rc->qpm );
1172 x264_ratecontrol_set_estimated_size(h, b1);
1175 /* loses the fractional part of the frame-wise qp */
1176 rc->f_qpm = rc->qpm;
1179 int x264_ratecontrol_qp( x264_t *h )
1184 /* In 2pass, force the same frame types as in the 1st pass */
1185 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1187 x264_ratecontrol_t *rc = h->rc;
1188 if( h->param.rc.b_stat_read )
1190 if( frame_num >= rc->num_entries )
1192 /* We could try to initialize everything required for ABR and
1193 * adaptive B-frames, but that would be complicated.
1194 * So just calculate the average QP used so far. */
1197 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24
1198 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1199 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
1200 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 );
1201 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 );
1203 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1204 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1205 if( h->param.i_bframe_adaptive )
1206 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1208 for( i = 0; i < h->param.i_threads; i++ )
1210 h->thread[i]->rc->b_abr = 0;
1211 h->thread[i]->rc->b_2pass = 0;
1212 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1213 h->thread[i]->param.rc.b_stat_read = 0;
1214 h->thread[i]->param.i_bframe_adaptive = 0;
1215 h->thread[i]->param.i_scenecut_threshold = 0;
1216 if( h->thread[i]->param.i_bframe > 1 )
1217 h->thread[i]->param.i_bframe = 1;
1219 return X264_TYPE_AUTO;
1221 switch( rc->entry[frame_num].pict_type )
1224 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I;
1227 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B;
1236 return X264_TYPE_AUTO;
1240 /* After encoding one frame, save stats and update ratecontrol state */
1241 int x264_ratecontrol_end( x264_t *h, int bits )
1243 x264_ratecontrol_t *rc = h->rc;
1244 const int *mbs = h->stat.frame.i_mb_count;
1249 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1250 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1251 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1252 for( i = B_DIRECT; i < B_8x8; i++ )
1253 h->stat.frame.i_mb_count_p += mbs[i];
1255 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1256 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1258 if( h->param.rc.b_stat_write )
1260 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1261 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1262 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1263 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1264 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1265 char c_direct = h->mb.b_direct_auto_write ?
1266 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1267 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1269 if( fprintf( rc->p_stat_file_out,
1270 "in:%d out:%d type:%c q:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c ref:",
1271 h->fenc->i_frame, h->i_frame,
1273 h->stat.frame.i_tex_bits,
1274 h->stat.frame.i_mv_bits,
1275 h->stat.frame.i_misc_bits,
1276 h->stat.frame.i_mb_count_i,
1277 h->stat.frame.i_mb_count_p,
1278 h->stat.frame.i_mb_count_skip,
1282 for( i = 0; i < h->i_ref0; i++ )
1284 int refcount = h->param.b_interlaced ? h->stat.frame.i_mb_count_ref[0][i*2]
1285 + h->stat.frame.i_mb_count_ref[0][i*2+1] :
1286 h->stat.frame.i_mb_count_ref[0][i];
1287 if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1291 if( fprintf( rc->p_stat_file_out, ";\n" ) < 0 )
1294 /* Don't re-write the data in multi-pass mode. */
1295 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1297 uint8_t i_type = h->sh.i_type;
1299 /* Values are stored as big-endian FIX8.8 */
1300 for( i = 0; i < h->mb.i_mb_count; i++ )
1301 rc->qp_buffer[i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1302 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1304 if( fwrite( rc->qp_buffer, sizeof(uint16_t), h->mb.i_mb_count, rc->p_mbtree_stat_file_out ) < h->mb.i_mb_count )
1311 if( h->sh.i_type != SLICE_TYPE_B )
1312 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / rc->last_rceq;
1315 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1316 * Not perfectly accurate with B-refs, but good enough. */
1317 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
1319 rc->cplxr_sum *= rc->cbr_decay;
1320 rc->wanted_bits_window += rc->bitrate / rc->fps;
1321 rc->wanted_bits_window *= rc->cbr_decay;
1326 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
1329 if( h->mb.b_variable_qp )
1331 if( h->sh.i_type == SLICE_TYPE_B )
1333 rc->bframe_bits += bits;
1334 if( h->fenc->b_last_minigop_bframe )
1336 update_predictor( rc->pred_b_from_p, qp2qscale(rc->qpa_rc),
1337 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1338 rc->bframe_bits = 0;
1343 update_vbv( h, bits );
1346 x264_log(h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n");
1350 /****************************************************************************
1352 ***************************************************************************/
1355 * modify the bitrate curve from pass1 for one frame
1357 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1359 x264_ratecontrol_t *rcc= h->rc;
1361 x264_zone_t *zone = get_zone( h, frame_num );
1363 q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1365 // avoid NaN's in the rc_eq
1366 if(!isfinite(q) || rce->tex_bits + rce->mv_bits == 0)
1367 q = rcc->last_qscale_for[rce->pict_type];
1372 rcc->last_qscale = q;
1377 if( zone->b_force_qp )
1378 q = qp2qscale(zone->i_qp);
1380 q /= zone->f_bitrate_factor;
1386 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1388 x264_ratecontrol_t *rcc = h->rc;
1389 const int pict_type = rce->pict_type;
1391 // force I/B quants as a function of P quants
1392 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1393 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1394 if( pict_type == SLICE_TYPE_I )
1397 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1398 double ip_factor = fabs( h->param.rc.f_ip_factor );
1399 /* don't apply ip_factor if the following frame is also I */
1400 if( rcc->accum_p_norm <= 0 )
1402 else if( h->param.rc.f_ip_factor < 0 )
1404 else if( rcc->accum_p_norm >= 1 )
1407 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1409 else if( pict_type == SLICE_TYPE_B )
1411 if( h->param.rc.f_pb_factor > 0 )
1413 if( !rce->kept_as_ref )
1414 q *= fabs( h->param.rc.f_pb_factor );
1416 else if( pict_type == SLICE_TYPE_P
1417 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1418 && rce->tex_bits == 0 )
1423 /* last qscale / qdiff stuff */
1424 if(rcc->last_non_b_pict_type==pict_type
1425 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1427 double last_q = rcc->last_qscale_for[pict_type];
1428 double max_qscale = last_q * rcc->lstep;
1429 double min_qscale = last_q / rcc->lstep;
1431 if (q > max_qscale) q = max_qscale;
1432 else if(q < min_qscale) q = min_qscale;
1435 rcc->last_qscale_for[pict_type] = q;
1436 if(pict_type!=SLICE_TYPE_B)
1437 rcc->last_non_b_pict_type = pict_type;
1438 if(pict_type==SLICE_TYPE_I)
1440 rcc->last_accum_p_norm = rcc->accum_p_norm;
1441 rcc->accum_p_norm = 0;
1442 rcc->accum_p_qp = 0;
1444 if(pict_type==SLICE_TYPE_P)
1446 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1447 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1448 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1453 static double predict_size( predictor_t *p, double q, double var )
1455 return (p->coeff*var + p->offset) / (q*p->count);
1458 static void update_predictor( predictor_t *p, double q, double var, double bits )
1460 const double range = 1.5;
1463 double old_coeff = p->coeff / p->count;
1464 double new_coeff = bits*q / var;
1465 double new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
1466 double new_offset = bits*q - new_coeff_clipped * var;
1467 if( new_offset >= 0 )
1468 new_coeff = new_coeff_clipped;
1471 p->count *= p->decay;
1472 p->coeff *= p->decay;
1473 p->offset *= p->decay;
1475 p->coeff += new_coeff;
1476 p->offset += new_offset;
1479 // update VBV after encoding a frame
1480 static void update_vbv( x264_t *h, int bits )
1482 x264_ratecontrol_t *rcc = h->rc;
1483 x264_ratecontrol_t *rct = h->thread[0]->rc;
1485 if( rcc->last_satd >= h->mb.i_mb_count )
1486 update_predictor( &rct->pred[h->sh.i_type], qp2qscale(rcc->qpa_rc), rcc->last_satd, bits );
1491 rct->buffer_fill_final -= bits;
1492 if( rct->buffer_fill_final < 0 )
1493 x264_log( h, X264_LOG_WARNING, "VBV underflow (frame %d, %.0f bits)\n", h->i_frame, rct->buffer_fill_final );
1494 rct->buffer_fill_final = X264_MAX( rct->buffer_fill_final, 0 );
1495 rct->buffer_fill_final += rct->buffer_rate;
1496 rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, rct->buffer_size );
1499 // provisionally update VBV according to the planned size of all frames currently in progress
1500 static void update_vbv_plan( x264_t *h, int overhead )
1502 x264_ratecontrol_t *rcc = h->rc;
1503 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final - overhead;
1504 if( h->param.i_threads > 1 )
1506 int j = h->rc - h->thread[0]->rc;
1508 for( i=1; i<h->param.i_threads; i++ )
1510 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1511 double bits = t->rc->frame_size_planned;
1512 if( !t->b_thread_active )
1514 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1515 rcc->buffer_fill -= bits;
1516 rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
1517 rcc->buffer_fill += rcc->buffer_rate;
1518 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1523 // apply VBV constraints and clip qscale to between lmin and lmax
1524 static double clip_qscale( x264_t *h, int pict_type, double q )
1526 x264_ratecontrol_t *rcc = h->rc;
1527 double lmin = rcc->lmin[pict_type];
1528 double lmax = rcc->lmax[pict_type];
1531 /* B-frames are not directly subject to VBV,
1532 * since they are controlled by the P-frames' QPs. */
1534 if( rcc->b_vbv && rcc->last_satd > 0 )
1536 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
1537 * the lookahead overflow and such that the buffer is in a reasonable state
1538 * by the end of the lookahead. */
1539 if( h->param.rc.i_lookahead )
1541 int j, iterations, terminate = 0;
1543 /* Avoid an infinite loop. */
1544 for( iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
1547 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1548 double buffer_fill_cur = rcc->buffer_fill - cur_bits + rcc->buffer_rate;
1550 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
1551 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
1552 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
1554 /* Loop over the planned future frames. */
1555 for( j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
1557 int i_type = h->fenc->i_planned_type[j];
1558 int i_satd = h->fenc->i_planned_satd[j];
1559 if( i_type == X264_TYPE_AUTO )
1561 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
1562 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
1563 buffer_fill_cur = buffer_fill_cur - cur_bits + rcc->buffer_rate;
1565 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
1566 target_fill = X264_MIN( rcc->buffer_fill + j * rcc->buffer_rate * 0.5, rcc->buffer_size * 0.5 );
1567 if( buffer_fill_cur < target_fill )
1573 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
1574 target_fill = x264_clip3f( rcc->buffer_fill - j * rcc->buffer_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
1575 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
1584 /* Fallback to old purely-reactive algorithm: no lookahead. */
1587 if( ( pict_type == SLICE_TYPE_P ||
1588 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
1589 rcc->buffer_fill/rcc->buffer_size < 0.5 )
1591 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1594 /* Now a hard threshold to make sure the frame fits in VBV.
1595 * This one is mostly for I-frames. */
1596 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1598 /* For small VBVs, allow the frame to use up the entire VBV. */
1599 double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
1600 /* For single-frame VBVs, request that the frame use up the entire VBV. */
1601 double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
1603 if( bits > rcc->buffer_fill/max_fill_factor )
1604 qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
1607 if( bits < rcc->buffer_rate/min_fill_factor )
1608 q *= bits*min_fill_factor/rcc->buffer_rate;
1609 q = X264_MAX( q0, q );
1612 /* Check B-frame complexity, and use up any bits that would
1613 * overflow before the next P-frame. */
1614 if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
1616 int nb = rcc->bframes;
1617 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1618 double pbbits = bits;
1619 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1621 if( bbits > rcc->buffer_rate )
1623 pbbits += nb * bbits;
1625 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1626 if( pbbits < space )
1628 q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
1630 q = X264_MAX( q0-5, q );
1633 if( !rcc->b_vbv_min_rate )
1634 q = X264_MAX( q0, q );
1639 else if(rcc->b_2pass)
1641 double min2 = log(lmin);
1642 double max2 = log(lmax);
1643 q = (log(q) - min2)/(max2-min2) - 0.5;
1644 q = 1.0/(1.0 + exp(-4*q));
1645 q = q*(max2-min2) + min2;
1649 return x264_clip3f(q, lmin, lmax);
1652 // update qscale for 1 frame based on actual bits used so far
1653 static float rate_estimate_qscale( x264_t *h )
1656 x264_ratecontrol_t *rcc = h->rc;
1657 ratecontrol_entry_t rce;
1658 int pict_type = h->sh.i_type;
1659 double lmin = rcc->lmin[pict_type];
1660 double lmax = rcc->lmax[pict_type];
1661 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
1662 + h->stat.i_frame_size[SLICE_TYPE_P]
1663 + h->stat.i_frame_size[SLICE_TYPE_B]);
1668 if(pict_type != rce.pict_type)
1670 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1671 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1675 if( pict_type == SLICE_TYPE_B )
1677 /* B-frames don't have independent ratecontrol, but rather get the
1678 * average QP of the two adjacent P-frames + an offset */
1680 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1681 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1682 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1683 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1684 float q0 = h->fref0[0]->f_qp_avg_rc;
1685 float q1 = h->fref1[0]->f_qp_avg_rc;
1687 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1688 q0 -= rcc->pb_offset/2;
1689 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1690 q1 -= rcc->pb_offset/2;
1693 q = (q0 + q1) / 2 + rcc->ip_offset;
1699 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1701 if(h->fenc->b_kept_as_ref)
1702 q += rcc->pb_offset/2;
1704 q += rcc->pb_offset;
1706 if( rcc->b_2pass && rcc->b_vbv )
1707 rcc->frame_size_planned = qscale2bits( &rce, q );
1709 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1710 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1714 rcc->last_satd = x264_rc_analyse_slice( h );
1715 return qp2qscale(q);
1719 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1723 //FIXME adjust abr_buffer based on distance to the end of the video
1725 int64_t predicted_bits = total_bits;
1729 if( h->param.i_threads > 1 )
1731 int j = h->rc - h->thread[0]->rc;
1733 for( i=1; i<h->param.i_threads; i++ )
1735 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1736 double bits = t->rc->frame_size_planned;
1737 if( !t->b_thread_active )
1739 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1740 predicted_bits += (int64_t)bits;
1746 if( h->fenc->i_frame < h->param.i_threads )
1747 predicted_bits += (int64_t)h->fenc->i_frame * rcc->bitrate / rcc->fps;
1749 predicted_bits += (int64_t)(h->param.i_threads - 1) * rcc->bitrate / rcc->fps;
1752 diff = predicted_bits - (int64_t)rce.expected_bits;
1754 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1755 if( ((h->fenc->i_frame + 1 - h->param.i_threads) >= rcc->fps) &&
1756 (rcc->expected_bits_sum > 0))
1758 /* Adjust quant based on the difference between
1759 * achieved and expected bitrate so far */
1760 double time = (double)h->fenc->i_frame / rcc->num_entries;
1761 double w = x264_clip3f( time*100, 0.0, 1.0 );
1762 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1766 /* Do not overflow vbv */
1767 double expected_size = qscale2bits(&rce, q);
1768 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1769 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
1770 double qmax = q*(2 - expected_fullness);
1771 double size_constraint = 1 + expected_fullness;
1772 qmax = X264_MAX(qmax, rce.new_qscale);
1773 if (expected_fullness < .05)
1775 qmax = X264_MIN(qmax, lmax);
1776 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
1777 ((expected_vbv < 0) && (q < lmax)))
1780 expected_size = qscale2bits(&rce, q);
1781 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1783 rcc->last_satd = x264_rc_analyse_slice( h );
1785 q = x264_clip3f( q, lmin, lmax );
1787 else /* 1pass ABR */
1789 /* Calculate the quantizer which would have produced the desired
1790 * average bitrate if it had been applied to all frames so far.
1791 * Then modulate that quant based on the current frame's complexity
1792 * relative to the average complexity so far (using the 2pass RCEQ).
1793 * Then bias the quant up or down if total size so far was far from
1795 * Result: Depending on the value of rate_tolerance, there is a
1796 * tradeoff between quality and bitrate precision. But at large
1797 * tolerances, the bit distribution approaches that of 2pass. */
1799 double wanted_bits, overflow=1, lmin, lmax;
1801 rcc->last_satd = x264_rc_analyse_slice( h );
1802 rcc->short_term_cplxsum *= 0.5;
1803 rcc->short_term_cplxcount *= 0.5;
1804 rcc->short_term_cplxsum += rcc->last_satd;
1805 rcc->short_term_cplxcount ++;
1807 rce.tex_bits = rcc->last_satd;
1808 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1810 rce.p_count = rcc->nmb;
1814 rce.pict_type = pict_type;
1816 if( h->param.rc.i_rc_method == X264_RC_CRF )
1818 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1822 int i_frame_done = h->fenc->i_frame + 1 - h->param.i_threads;
1824 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1826 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
1827 wanted_bits = i_frame_done * rcc->bitrate / rcc->fps;
1828 if( wanted_bits > 0 )
1830 abr_buffer *= X264_MAX( 1, sqrt(i_frame_done/25) );
1831 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1836 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1837 /* should test _next_ pict type, but that isn't decided yet */
1838 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1840 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1841 q /= fabs( h->param.rc.f_ip_factor );
1843 else if( h->i_frame > 0 )
1845 /* Asymmetric clipping, because symmetric would prevent
1846 * overflow control in areas of rapidly oscillating complexity */
1847 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1848 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1849 if( overflow > 1.1 && h->i_frame > 3 )
1851 else if( overflow < 0.9 )
1854 q = x264_clip3f(q, lmin, lmax);
1856 else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
1858 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1860 rcc->qp_novbv = qscale2qp(q);
1862 //FIXME use get_diff_limited_q() ?
1863 q = clip_qscale( h, pict_type, q );
1866 rcc->last_qscale_for[pict_type] =
1867 rcc->last_qscale = q;
1869 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
1870 rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );
1872 if( rcc->b_2pass && rcc->b_vbv )
1873 rcc->frame_size_planned = qscale2bits(&rce, q);
1875 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1876 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1881 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
1885 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
1886 /* these vars are updated in x264_ratecontrol_start()
1887 * so copy them from the context that most recently started (prev)
1888 * to the context that's about to start (cur).
1894 COPY(last_qscale_for);
1895 COPY(last_non_b_pict_type);
1896 COPY(short_term_cplxsum);
1897 COPY(short_term_cplxcount);
1904 #define COPY(var) next->rc->var = cur->rc->var
1905 /* these vars are updated in x264_ratecontrol_end()
1906 * so copy them from the context that most recently ended (cur)
1907 * to the context that's about to end (next)
1910 COPY(expected_bits_sum);
1911 COPY(wanted_bits_window);
1915 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
1916 /* the rest of the variables are either constant or thread-local */
1919 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
1921 /* find an interval ending on an overflow or underflow (depending on whether
1922 * we're adding or removing bits), and starting on the earliest frame that
1923 * can influence the buffer fill of that end frame. */
1924 x264_ratecontrol_t *rcc = h->rc;
1925 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
1926 const double buffer_max = .9 * rcc->buffer_size;
1927 double fill = fills[*t0-1];
1928 double parity = over ? 1. : -1.;
1929 int i, start=-1, end=-1;
1930 for(i = *t0; i < rcc->num_entries; i++)
1932 fill += (rcc->buffer_rate - qscale2bits(&rcc->entry[i], rcc->entry[i].new_qscale)) * parity;
1933 fill = x264_clip3f(fill, 0, rcc->buffer_size);
1935 if(fill <= buffer_min || i == 0)
1941 else if(fill >= buffer_max && start >= 0)
1946 return start>=0 && end>=0;
1949 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
1951 x264_ratecontrol_t *rcc = h->rc;
1952 double qscale_orig, qscale_new;
1957 for(i = t0; i <= t1; i++)
1959 qscale_orig = rcc->entry[i].new_qscale;
1960 qscale_orig = x264_clip3f(qscale_orig, qscale_min, qscale_max);
1961 qscale_new = qscale_orig * adjustment;
1962 qscale_new = x264_clip3f(qscale_new, qscale_min, qscale_max);
1963 rcc->entry[i].new_qscale = qscale_new;
1964 adjusted = adjusted || (qscale_new != qscale_orig);
1969 static double count_expected_bits( x264_t *h )
1971 x264_ratecontrol_t *rcc = h->rc;
1972 double expected_bits = 0;
1974 for(i = 0; i < rcc->num_entries; i++)
1976 ratecontrol_entry_t *rce = &rcc->entry[i];
1977 rce->expected_bits = expected_bits;
1978 expected_bits += qscale2bits(rce, rce->new_qscale);
1980 return expected_bits;
1983 static int vbv_pass2( x264_t *h )
1985 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
1986 * frames in the interval until either buffer is full at some intermediate frame or the
1987 * last frame in the interval no longer underflows. Recompute intervals and repeat.
1988 * Then do the converse to put bits back into overflow areas until target size is met */
1990 x264_ratecontrol_t *rcc = h->rc;
1992 double all_available_bits = h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps;
1993 double expected_bits = 0;
1995 double prev_bits = 0;
1997 double qscale_min = qp2qscale(h->param.rc.i_qp_min);
1998 double qscale_max = qp2qscale(h->param.rc.i_qp_max);
2000 int adj_min, adj_max;
2001 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2005 /* adjust overall stream size */
2009 prev_bits = expected_bits;
2011 if(expected_bits != 0)
2012 { /* not first iteration */
2013 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2014 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2018 while(adj_min && find_underflow(h, fills, &t0, &t1, 1))
2020 adj_min = fix_underflow(h, t0, t1, adjustment, qscale_min, qscale_max);
2025 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2027 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2029 while(adj_max && find_underflow(h, fills, &t0, &t1, 0))
2030 adj_max = fix_underflow(h, t0, t1, 1.001, qscale_min, qscale_max);
2032 expected_bits = count_expected_bits(h);
2033 } while((expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2036 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2038 /* store expected vbv filling values for tracking when encoding */
2039 for(i = 0; i < rcc->num_entries; i++)
2040 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
2048 static int init_pass2( x264_t *h )
2050 x264_ratecontrol_t *rcc = h->rc;
2051 uint64_t all_const_bits = 0;
2052 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps);
2053 double rate_factor, step, step_mult;
2054 double qblur = h->param.rc.f_qblur;
2055 double cplxblur = h->param.rc.f_complexity_blur;
2056 const int filter_size = (int)(qblur*4) | 1;
2057 double expected_bits;
2058 double *qscale, *blurred_qscale;
2061 /* find total/average complexity & const_bits */
2062 for(i=0; i<rcc->num_entries; i++)
2064 ratecontrol_entry_t *rce = &rcc->entry[i];
2065 all_const_bits += rce->misc_bits;
2068 if( all_available_bits < all_const_bits)
2070 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2071 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)));
2075 /* Blur complexities, to reduce local fluctuation of QP.
2076 * We don't blur the QPs directly, because then one very simple frame
2077 * could drag down the QP of a nearby complex frame and give it more
2078 * bits than intended. */
2079 for(i=0; i<rcc->num_entries; i++)
2081 ratecontrol_entry_t *rce = &rcc->entry[i];
2082 double weight_sum = 0;
2083 double cplx_sum = 0;
2084 double weight = 1.0;
2085 double gaussian_weight;
2087 /* weighted average of cplx of future frames */
2088 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++)
2090 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2091 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2094 gaussian_weight = weight * exp(-j*j/200.0);
2095 weight_sum += gaussian_weight;
2096 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2098 /* weighted average of cplx of past frames */
2100 for(j=0; j<=cplxblur*2 && j<=i; j++)
2102 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2103 gaussian_weight = weight * exp(-j*j/200.0);
2104 weight_sum += gaussian_weight;
2105 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2106 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2110 rce->blurred_complexity = cplx_sum / weight_sum;
2113 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2114 if( filter_size > 1 )
2115 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2117 blurred_qscale = qscale;
2119 /* Search for a factor which, when multiplied by the RCEQ values from
2120 * each frame, adds up to the desired total size.
2121 * There is no exact closed-form solution because of VBV constraints and
2122 * because qscale2bits is not invertible, but we can start with the simple
2123 * approximation of scaling the 1st pass by the ratio of bitrates.
2124 * The search range is probably overkill, but speed doesn't matter here. */
2127 for(i=0; i<rcc->num_entries; i++)
2128 expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i));
2129 step_mult = all_available_bits / expected_bits;
2132 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2135 rate_factor += step;
2137 rcc->last_non_b_pict_type = -1;
2138 rcc->last_accum_p_norm = 1;
2139 rcc->accum_p_norm = 0;
2142 for(i=0; i<rcc->num_entries; i++)
2144 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
2147 /* fixed I/B qscale relative to P */
2148 for(i=rcc->num_entries-1; i>=0; i--)
2150 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
2151 assert(qscale[i] >= 0);
2157 assert(filter_size%2==1);
2158 for(i=0; i<rcc->num_entries; i++)
2160 ratecontrol_entry_t *rce = &rcc->entry[i];
2162 double q=0.0, sum=0.0;
2164 for(j=0; j<filter_size; j++)
2166 int index = i+j-filter_size/2;
2168 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
2169 if(index < 0 || index >= rcc->num_entries)
2171 if(rce->pict_type != rcc->entry[index].pict_type)
2173 q += qscale[index] * coeff;
2176 blurred_qscale[i] = q/sum;
2180 /* find expected bits */
2181 for(i=0; i<rcc->num_entries; i++)
2183 ratecontrol_entry_t *rce = &rcc->entry[i];
2184 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
2185 assert(rce->new_qscale >= 0);
2186 expected_bits += qscale2bits(rce, rce->new_qscale);
2189 if(expected_bits > all_available_bits) rate_factor -= step;
2194 x264_free(blurred_qscale);
2197 if( vbv_pass2( h ) )
2199 expected_bits = count_expected_bits(h);
2201 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
2204 for(i=0; i<rcc->num_entries; i++)
2205 avgq += rcc->entry[i].new_qscale;
2206 avgq = qscale2qp(avgq / rcc->num_entries);
2208 if ((expected_bits > all_available_bits) || (!rcc->b_vbv))
2209 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
2210 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
2211 (float)h->param.rc.i_bitrate,
2212 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
2214 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
2216 if(h->param.rc.i_qp_min > 0)
2217 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
2219 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
2221 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
2223 if(h->param.rc.i_qp_max < 51)
2224 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
2226 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
2228 else if(!(rcc->b_2pass && rcc->b_vbv))
2229 x264_log(h, X264_LOG_WARNING, "internal error\n");