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"
43 uint64_t expected_bits; /*total expected bits up to the current frame (current one excluded)*/
50 float blurred_complexity;
52 } ratecontrol_entry_t;
62 struct x264_ratecontrol_t
71 double rate_tolerance;
73 int nmb; /* number of macroblocks in a frame */
77 ratecontrol_entry_t *rce;
78 int qp; /* qp for current frame */
79 int qpm; /* qp for current macroblock */
80 float f_qpm; /* qp for current macroblock: precise float for AQ */
81 float qpa_rc; /* average of macroblocks' qp before aq */
82 float qpa_aq; /* average of macroblocks' qp after aq */
83 float qp_novbv; /* QP for the current frame if 1-pass VBV was disabled. */
88 double buffer_fill_final; /* real buffer as of the last finished frame */
89 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
90 double buffer_rate; /* # of bits added to buffer_fill after each frame */
91 predictor_t *pred; /* predict frame size from satd */
96 double cplxr_sum; /* sum of bits*qscale/rceq */
97 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow, only includes finished frames */
98 double wanted_bits_window; /* target bitrate * window */
100 double short_term_cplxsum;
101 double short_term_cplxcount;
102 double rate_factor_constant;
107 FILE *p_stat_file_out;
108 char *psz_stat_file_tmpname;
109 FILE *p_mbtree_stat_file_out;
110 char *psz_mbtree_stat_file_tmpname;
111 char *psz_mbtree_stat_file_name;
112 FILE *p_mbtree_stat_file_in;
114 int num_entries; /* number of ratecontrol_entry_ts */
115 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
117 double last_qscale_for[5]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
118 int last_non_b_pict_type;
119 double accum_p_qp; /* for determining I-frame quant */
121 double last_accum_p_norm;
122 double lmin[5]; /* min qscale by frame type */
124 double lstep; /* max change (multiply) in qscale per frame */
125 uint16_t *qp_buffer; /* Global buffer for converting MB-tree quantizer data. */
128 double frame_size_estimated;
129 double frame_size_planned;
130 predictor_t *row_pred;
131 predictor_t row_preds[5];
132 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
133 int bframes; /* # consecutive B-frames before this P-frame */
134 int bframe_bits; /* total cost of those frames */
138 x264_zone_t *prev_zone;
142 static int parse_zones( x264_t *h );
143 static int init_pass2(x264_t *);
144 static float rate_estimate_qscale( x264_t *h );
145 static void update_vbv( x264_t *h, int bits );
146 static void update_vbv_plan( x264_t *h );
147 static double predict_size( predictor_t *p, double q, double var );
148 static void update_predictor( predictor_t *p, double q, double var, double bits );
151 * qp = h.264's quantizer
152 * qscale = linearized quantizer = Lagrange multiplier
154 static inline double qp2qscale(double qp)
156 return 0.85 * pow(2.0, ( qp - 12.0 ) / 6.0);
158 static inline double qscale2qp(double qscale)
160 return 12.0 + 6.0 * log(qscale/0.85) / log(2.0);
163 /* Texture bitrate is not quite inversely proportional to qscale,
164 * probably due the the changing number of SKIP blocks.
165 * MV bits level off at about qp<=12, because the lambda used
166 * for motion estimation is constant there. */
167 static inline double qscale2bits(ratecontrol_entry_t *rce, double qscale)
171 return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
172 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
176 // Find the total AC energy of the block in all planes.
177 static NOINLINE uint32_t ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
179 /* This function contains annoying hacks because GCC has a habit of reordering emms
180 * and putting it after floating point ops. As a result, we put the emms at the end of the
181 * function and make sure that its always called before the float math. Noinline makes
182 * sure no reordering goes on. */
184 for( i = 0; i < 3; i++ )
187 int stride = frame->i_stride[i];
188 int offset = h->mb.b_interlaced
189 ? w * (mb_x + (mb_y&~1) * stride) + (mb_y&1) * stride
190 : w * (mb_x + mb_y * stride);
191 int pix = i ? PIXEL_8x8 : PIXEL_16x16;
192 stride <<= h->mb.b_interlaced;
193 var += h->pixf.var[pix]( frame->plane[i]+offset, stride );
199 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame )
201 /* constants chosen to result in approximately the same overall bitrate as without AQ.
202 * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
206 /* Need to init it anyways for MB tree. */
207 if( h->param.rc.f_aq_strength == 0 )
210 memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
211 memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
212 if( h->frames.b_have_lowres )
213 for( mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
214 frame->i_inv_qscale_factor[mb_xy] = 256;
218 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
220 for( mb_y = 0; mb_y < h->sps->i_mb_height; mb_y++ )
221 for( mb_x = 0; mb_x < h->sps->i_mb_width; mb_x++ )
223 uint32_t energy = ac_energy_mb( h, mb_x, mb_y, frame );
224 float qp_adj = x264_log2( energy + 2 );
226 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
229 avg_adj /= h->mb.i_mb_count;
230 strength = h->param.rc.f_aq_strength * avg_adj * (1.f / 6000.f);
233 strength = h->param.rc.f_aq_strength * 1.0397f;
235 for( mb_y = 0; mb_y < h->sps->i_mb_height; mb_y++ )
236 for( mb_x = 0; mb_x < h->sps->i_mb_width; mb_x++ )
239 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
241 qp_adj = frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride];
242 qp_adj = strength * (qp_adj - avg_adj);
246 uint32_t energy = ac_energy_mb( h, mb_x, mb_y, frame );
247 qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - 14.427f);
249 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] =
250 frame->f_qp_offset_aq[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
251 if( h->frames.b_have_lowres )
252 frame->i_inv_qscale_factor[mb_x + mb_y*h->mb.i_mb_stride] = x264_exp2fix8(qp_adj);
257 /*****************************************************************************
258 * x264_adaptive_quant:
259 * adjust macroblock QP based on variance (AC energy) of the MB.
260 * high variance = higher QP
261 * low variance = lower QP
262 * This generally increases SSIM and lowers PSNR.
263 *****************************************************************************/
264 void x264_adaptive_quant( x264_t *h )
267 h->mb.i_qp = x264_clip3( h->rc->f_qpm + h->fenc->f_qp_offset[h->mb.i_mb_xy] + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
270 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame )
272 x264_ratecontrol_t *rc = h->rc;
273 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
276 if( i_type_actual != SLICE_TYPE_B )
280 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
283 if( i_type != i_type_actual )
285 x264_log(h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type,i_type_actual);
289 if( fread( rc->qp_buffer, sizeof(uint16_t), h->mb.i_mb_count, rc->p_mbtree_stat_file_in ) != h->mb.i_mb_count )
292 for( i = 0; i < h->mb.i_mb_count; i++ )
293 frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[i] )) * (1/256.0);
296 x264_adaptive_quant_frame( h, frame );
299 x264_log(h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n");
303 static char *x264_strcat_filename( char *input, char *suffix )
305 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
308 strcpy( output, input );
309 strcat( output, suffix );
313 int x264_ratecontrol_new( x264_t *h )
315 x264_ratecontrol_t *rc;
320 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
323 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
324 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
326 /* FIXME: use integers */
327 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
328 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
332 if( h->param.rc.b_mb_tree )
334 h->param.rc.f_pb_factor = 1;
338 rc->qcompress = h->param.rc.f_qcompress;
340 rc->bitrate = h->param.rc.i_bitrate * 1000.;
341 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
342 rc->nmb = h->mb.i_mb_count;
343 rc->last_non_b_pict_type = -1;
346 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
348 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
351 if( h->param.rc.i_vbv_buffer_size )
353 if( h->param.rc.i_rc_method == X264_RC_CQP )
355 x264_log(h, X264_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
356 h->param.rc.i_vbv_max_bitrate = 0;
357 h->param.rc.i_vbv_buffer_size = 0;
359 else if( h->param.rc.i_vbv_max_bitrate == 0 )
361 x264_log( h, X264_LOG_DEBUG, "VBV maxrate unspecified, assuming CBR\n" );
362 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
365 if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate &&
366 h->param.rc.i_vbv_max_bitrate > 0)
367 x264_log(h, X264_LOG_WARNING, "max bitrate less than average bitrate, ignored.\n");
368 else if( h->param.rc.i_vbv_max_bitrate > 0 &&
369 h->param.rc.i_vbv_buffer_size > 0 )
371 if( h->param.rc.i_vbv_buffer_size < 3 * h->param.rc.i_vbv_max_bitrate / rc->fps )
373 h->param.rc.i_vbv_buffer_size = 3 * h->param.rc.i_vbv_max_bitrate / rc->fps;
374 x264_log( h, X264_LOG_WARNING, "VBV buffer size too small, using %d kbit\n",
375 h->param.rc.i_vbv_buffer_size );
377 if( h->param.rc.f_vbv_buffer_init > 1. )
378 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 );
379 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000. / rc->fps;
380 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000.;
381 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
382 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
383 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
385 rc->b_vbv_min_rate = !rc->b_2pass
386 && h->param.rc.i_rc_method == X264_RC_ABR
387 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
389 else if( h->param.rc.i_vbv_max_bitrate )
391 x264_log(h, X264_LOG_WARNING, "VBV maxrate specified, but no bufsize.\n");
392 h->param.rc.i_vbv_max_bitrate = 0;
394 if(rc->rate_tolerance < 0.01)
396 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
397 rc->rate_tolerance = 0.01;
400 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
404 /* FIXME ABR_INIT_QP is actually used only in CRF */
405 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
406 rc->accum_p_norm = .01;
407 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
408 /* estimated ratio that produces a reasonable QP for the first I-frame */
409 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
410 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
411 rc->last_non_b_pict_type = SLICE_TYPE_I;
414 if( h->param.rc.i_rc_method == X264_RC_CRF )
416 /* Arbitrary rescaling to make CRF somewhat similar to QP.
417 * Try to compensate for MB-tree's effects as well. */
418 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
419 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
420 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
421 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset );
424 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
425 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
426 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
427 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
428 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
430 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
431 rc->last_qscale = qp2qscale(26);
432 CHECKED_MALLOC( rc->pred, 5*sizeof(predictor_t) );
433 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
434 for( i = 0; i < 5; i++ )
436 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
437 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
438 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
439 rc->pred[i].coeff= 2.0;
440 rc->pred[i].count= 1.0;
441 rc->pred[i].decay= 0.5;
442 rc->pred[i].offset= 0.0;
443 rc->row_preds[i].coeff= .25;
444 rc->row_preds[i].count= 1.0;
445 rc->row_preds[i].decay= 0.5;
446 rc->row_preds[i].offset= 0.0;
448 *rc->pred_b_from_p = rc->pred[0];
450 if( parse_zones( h ) < 0 )
452 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
456 /* Load stat file and init 2pass algo */
457 if( h->param.rc.b_stat_read )
459 char *p, *stats_in, *stats_buf;
461 /* read 1st pass stats */
462 assert( h->param.rc.psz_stat_in );
463 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
466 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
469 if( h->param.rc.b_mb_tree )
471 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
472 if( !mbtree_stats_in )
474 rc->p_mbtree_stat_file_in = fopen( mbtree_stats_in, "rb" );
475 x264_free( mbtree_stats_in );
476 if( !rc->p_mbtree_stat_file_in )
478 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
483 /* check whether 1st pass options were compatible with current options */
484 if( !strncmp( stats_buf, "#options:", 9 ) )
487 char *opts = stats_buf;
488 stats_in = strchr( stats_buf, '\n' );
494 if( ( p = strstr( opts, "bframes=" ) ) && sscanf( p, "bframes=%d", &i )
495 && h->param.i_bframe != i )
497 x264_log( h, X264_LOG_ERROR, "different number of B-frames than 1st pass (%d vs %d)\n",
498 h->param.i_bframe, i );
502 /* since B-adapt doesn't (yet) take into account B-pyramid,
503 * the converse is not a problem */
504 if( strstr( opts, "b_pyramid=1" ) && !h->param.b_bframe_pyramid )
505 x264_log( h, X264_LOG_WARNING, "1st pass used B-pyramid, 2nd doesn't\n" );
507 if( ( p = strstr( opts, "keyint=" ) ) && sscanf( p, "keyint=%d", &i )
508 && h->param.i_keyint_max != i )
509 x264_log( h, X264_LOG_WARNING, "different keyint than 1st pass (%d vs %d)\n",
510 h->param.i_keyint_max, i );
512 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
513 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
515 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
517 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
518 h->mb.b_direct_auto_write = 1;
521 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
522 h->param.i_bframe_adaptive = i;
523 else if( h->param.i_bframe )
525 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
529 if( h->param.rc.b_mb_tree && ( p = strstr( opts, "rc_lookahead=" ) ) && sscanf( p, "rc_lookahead=%d", &i ) )
530 h->param.rc.i_lookahead = i;
533 /* find number of pics */
536 p = strchr(p+1, ';');
539 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
544 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
546 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
547 h->param.i_frame_total, rc->num_entries );
549 if( h->param.i_frame_total > rc->num_entries )
551 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
552 h->param.i_frame_total, rc->num_entries );
556 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
558 /* init all to skipped p frames */
559 for(i=0; i<rc->num_entries; i++)
561 ratecontrol_entry_t *rce = &rc->entry[i];
562 rce->pict_type = SLICE_TYPE_P;
563 rce->qscale = rce->new_qscale = qp2qscale(20);
564 rce->misc_bits = rc->nmb + 10;
570 for(i=0; i < rc->num_entries; i++)
572 ratecontrol_entry_t *rce;
579 next= strchr(p, ';');
582 (*next)=0; //sscanf is unbelievably slow on long strings
585 e = sscanf(p, " in:%d ", &frame_number);
587 if(frame_number < 0 || frame_number >= rc->num_entries)
589 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
592 rce = &rc->entry[frame_number];
593 rce->direct_mode = 0;
595 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",
596 &pict_type, &qp, &rce->tex_bits,
597 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
598 &rce->s_count, &rce->direct_mode);
602 case 'I': rce->kept_as_ref = 1;
603 case 'i': rce->pict_type = SLICE_TYPE_I; break;
604 case 'P': rce->pict_type = SLICE_TYPE_P; break;
605 case 'B': rce->kept_as_ref = 1;
606 case 'b': rce->pict_type = SLICE_TYPE_B; break;
607 default: e = -1; break;
611 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
614 rce->qscale = qp2qscale(qp);
618 x264_free(stats_buf);
620 if(h->param.rc.i_rc_method == X264_RC_ABR)
622 if(init_pass2(h) < 0) return -1;
623 } /* else we're using constant quant, so no need to run the bitrate allocation */
626 /* Open output file */
627 /* If input and output files are the same, output to a temp file
628 * and move it to the real name only when it's complete */
629 if( h->param.rc.b_stat_write )
632 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
633 if( !rc->psz_stat_file_tmpname )
636 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
637 if( rc->p_stat_file_out == NULL )
639 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
643 p = x264_param2string( &h->param, 1 );
645 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
647 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
649 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
650 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
651 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
654 rc->p_mbtree_stat_file_out = fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
655 if( rc->p_mbtree_stat_file_out == NULL )
657 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
663 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
664 CHECKED_MALLOC( rc->qp_buffer, h->mb.i_mb_count * sizeof(uint16_t) );
666 for( i=0; i<h->param.i_threads; i++ )
668 h->thread[i]->rc = rc+i;
672 memcpy( &h->thread[i]->param, &h->param, sizeof(x264_param_t) );
673 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
682 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
685 char *tok, UNUSED *saveptr;
687 z->f_bitrate_factor = 1;
688 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
690 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
692 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
696 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
702 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
703 memcpy( z->param, &h->param, sizeof(x264_param_t) );
704 z->param->param_free = x264_free;
705 while( (tok = strtok_r( p, ",", &saveptr )) )
707 char *val = strchr( tok, '=' );
713 if( x264_param_parse( z->param, tok, val ) )
715 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
725 static int parse_zones( x264_t *h )
727 x264_ratecontrol_t *rc = h->rc;
729 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
731 char *psz_zones, *p, *tok, UNUSED *saveptr;
732 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
733 strcpy( psz_zones, h->param.rc.psz_zones );
734 h->param.rc.i_zones = 1;
735 for( p = psz_zones; *p; p++ )
736 h->param.rc.i_zones += (*p == '/');
737 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
739 for( i = 0; i < h->param.rc.i_zones; i++ )
741 tok = strtok_r( p, "/", &saveptr );
742 if( !tok || parse_zone( h, &h->param.rc.zones[i], tok ) )
746 x264_free( psz_zones );
749 if( h->param.rc.i_zones > 0 )
751 for( i = 0; i < h->param.rc.i_zones; i++ )
753 x264_zone_t z = h->param.rc.zones[i];
754 if( z.i_start < 0 || z.i_start > z.i_end )
756 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
757 z.i_start, z.i_end );
760 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
762 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
763 z.f_bitrate_factor );
768 rc->i_zones = h->param.rc.i_zones + 1;
769 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
770 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
772 // default zone to fall back to if none of the others match
773 rc->zones[0].i_start = 0;
774 rc->zones[0].i_end = INT_MAX;
775 rc->zones[0].b_force_qp = 0;
776 rc->zones[0].f_bitrate_factor = 1;
777 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
778 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
779 for( i = 1; i < rc->i_zones; i++ )
781 if( !rc->zones[i].param )
782 rc->zones[i].param = rc->zones[0].param;
791 static x264_zone_t *get_zone( x264_t *h, int frame_num )
794 for( i = h->rc->i_zones-1; i >= 0; i-- )
796 x264_zone_t *z = &h->rc->zones[i];
797 if( frame_num >= z->i_start && frame_num <= z->i_end )
803 void x264_ratecontrol_summary( x264_t *h )
805 x264_ratecontrol_t *rc = h->rc;
806 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
808 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
809 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*12.5 : 0;
810 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
811 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
812 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset );
816 void x264_ratecontrol_delete( x264_t *h )
818 x264_ratecontrol_t *rc = h->rc;
821 if( rc->p_stat_file_out )
823 fclose( rc->p_stat_file_out );
824 if( h->i_frame >= rc->num_entries )
825 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
827 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
828 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
830 x264_free( rc->psz_stat_file_tmpname );
832 if( rc->p_mbtree_stat_file_out )
834 fclose( rc->p_mbtree_stat_file_out );
835 if( h->i_frame >= rc->num_entries )
836 if( rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
838 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
839 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
841 x264_free( rc->psz_mbtree_stat_file_tmpname );
842 x264_free( rc->psz_mbtree_stat_file_name );
844 x264_free( rc->pred );
845 x264_free( rc->pred_b_from_p );
846 x264_free( rc->entry );
847 x264_free( rc->qp_buffer );
850 x264_free( rc->zones[0].param );
851 for( i=1; i<rc->i_zones; i++ )
852 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
853 rc->zones[i].param->param_free( rc->zones[i].param );
854 x264_free( rc->zones );
859 void x264_ratecontrol_set_estimated_size( x264_t *h, int bits )
861 x264_pthread_mutex_lock( &h->fenc->mutex );
862 h->rc->frame_size_estimated = bits;
863 x264_pthread_mutex_unlock( &h->fenc->mutex );
866 int x264_ratecontrol_get_estimated_size( x264_t const *h)
869 x264_pthread_mutex_lock( &h->fenc->mutex );
870 size = h->rc->frame_size_estimated;
871 x264_pthread_mutex_unlock( &h->fenc->mutex );
875 static void accum_p_qp_update( x264_t *h, float qp )
877 x264_ratecontrol_t *rc = h->rc;
878 rc->accum_p_qp *= .95;
879 rc->accum_p_norm *= .95;
880 rc->accum_p_norm += 1;
881 if( h->sh.i_type == SLICE_TYPE_I )
882 rc->accum_p_qp += qp + rc->ip_offset;
884 rc->accum_p_qp += qp;
887 /* Before encoding a frame, choose a QP for it */
888 void x264_ratecontrol_start( x264_t *h, int i_force_qp )
890 x264_ratecontrol_t *rc = h->rc;
891 ratecontrol_entry_t *rce = NULL;
892 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
897 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
898 x264_encoder_reconfig( h, zone->param );
899 rc->prev_zone = zone;
901 rc->qp_force = i_force_qp;
903 if( h->param.rc.b_stat_read )
905 int frame = h->fenc->i_frame;
906 assert( frame >= 0 && frame < rc->num_entries );
907 rce = h->rc->rce = &h->rc->entry[frame];
909 if( h->sh.i_type == SLICE_TYPE_B
910 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
912 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
913 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
919 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
920 rc->row_pred = &rc->row_preds[h->sh.i_type];
921 update_vbv_plan( h );
924 if( h->sh.i_type != SLICE_TYPE_B )
927 while( h->frames.current[rc->bframes] && IS_X264_TYPE_B(h->frames.current[rc->bframes]->i_type) )
937 q = qscale2qp( rate_estimate_qscale( h ) );
939 else if( rc->b_2pass )
941 rce->new_qscale = rate_estimate_qscale( h );
942 q = qscale2qp( rce->new_qscale );
946 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
947 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
949 q = rc->qp_constant[ h->sh.i_type ];
953 if( zone->b_force_qp )
954 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
956 q -= 6*log(zone->f_bitrate_factor)/log(2);
962 h->fdec->f_qp_avg_rc =
963 h->fdec->f_qp_avg_aq =
965 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
968 rce->new_qp = rc->qp;
970 accum_p_qp_update( h, rc->qp );
972 if( h->sh.i_type != SLICE_TYPE_B )
973 rc->last_non_b_pict_type = h->sh.i_type;
976 static double predict_row_size( x264_t *h, int y, int qp )
978 /* average between two predictors:
979 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
980 x264_ratecontrol_t *rc = h->rc;
981 double pred_s = predict_size( rc->row_pred, qp2qscale(qp), h->fdec->i_row_satd[y] );
983 if( h->sh.i_type != SLICE_TYPE_I
984 && h->fref0[0]->i_type == h->fdec->i_type
985 && h->fref0[0]->i_row_satd[y] > 0
986 && (abs(h->fref0[0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
988 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
989 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
994 return (pred_s + pred_t) / 2;
997 static double row_bits_so_far( x264_t *h, int y )
1001 for( i = 0; i <= y; i++ )
1002 bits += h->fdec->i_row_bits[i];
1006 static double predict_row_size_sum( x264_t *h, int y, int qp )
1009 double bits = row_bits_so_far(h, y);
1010 for( i = y+1; i < h->sps->i_mb_height; i++ )
1011 bits += predict_row_size( h, i, qp );
1016 void x264_ratecontrol_mb( x264_t *h, int bits )
1018 x264_ratecontrol_t *rc = h->rc;
1019 const int y = h->mb.i_mb_y;
1023 h->fdec->i_row_bits[y] += bits;
1024 rc->qpa_rc += rc->f_qpm;
1025 rc->qpa_aq += h->mb.i_qp;
1027 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !rc->b_vbv )
1030 h->fdec->i_row_qp[y] = rc->qpm;
1032 if( h->sh.i_type == SLICE_TYPE_B )
1034 /* B-frames shouldn't use lower QP than their reference frames.
1035 * This code is a bit overzealous in limiting B-frame quantizers, but it helps avoid
1036 * underflows due to the fact that B-frames are not explicitly covered by VBV. */
1037 if( y < h->sps->i_mb_height-1 )
1040 int avg_qp = X264_MIN(h->fref0[0]->i_row_qp[y+1], h->fref1[0]->i_row_qp[y+1])
1041 + rc->pb_offset * ((h->fenc->i_type == X264_TYPE_BREF) ? 0.5 : 1);
1042 rc->qpm = X264_MIN(X264_MAX( rc->qp, avg_qp), 51); //avg_qp could go higher than 51 due to pb_offset
1043 i_estimated = row_bits_so_far(h, y); //FIXME: compute full estimated size
1044 if (i_estimated > h->rc->frame_size_planned)
1045 x264_ratecontrol_set_estimated_size(h, i_estimated);
1050 update_predictor( rc->row_pred, qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1052 /* tweak quality based on difference from predicted size */
1053 if( y < h->sps->i_mb_height-1 && h->stat.i_frame_count[h->sh.i_type] > 0 )
1055 int prev_row_qp = h->fdec->i_row_qp[y];
1056 int b0 = predict_row_size_sum( h, y, rc->qpm );
1058 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
1059 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1060 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1061 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits.
1062 * In 2-pass mode we can be more trusting of the planned frame sizes, since they were decided
1063 * by actual encoding instead of SATD prediction. */
1064 float rc_tol = h->param.rc.b_stat_read ? (buffer_left_planned / rc->buffer_size) * rc->frame_size_planned
1065 : (buffer_left_planned / h->param.i_threads);
1067 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1068 /* area at the top of the frame was measured inaccurately. */
1069 if(row_bits_so_far(h,y) < 0.05 * rc->frame_size_planned)
1072 if(h->sh.i_type != SLICE_TYPE_I)
1075 if( !rc->b_vbv_min_rate )
1076 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
1078 while( rc->qpm < i_qp_max
1079 && ((b1 > rc->frame_size_planned + rc_tol) ||
1080 (rc->buffer_fill - b1 < buffer_left_planned * 0.5) ||
1081 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1084 b1 = predict_row_size_sum( h, y, rc->qpm );
1087 /* avoid VBV underflow */
1088 while( (rc->qpm < h->param.rc.i_qp_max)
1089 && (rc->buffer_fill - b1 < rc->buffer_size * 0.005))
1092 b1 = predict_row_size_sum( h, y, rc->qpm );
1095 while( rc->qpm > i_qp_min
1096 && rc->qpm > h->fdec->i_row_qp[0]
1097 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
1098 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
1101 b1 = predict_row_size_sum( h, y, rc->qpm );
1103 x264_ratecontrol_set_estimated_size(h, b1);
1106 /* loses the fractional part of the frame-wise qp */
1107 rc->f_qpm = rc->qpm;
1110 int x264_ratecontrol_qp( x264_t *h )
1115 /* In 2pass, force the same frame types as in the 1st pass */
1116 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1118 x264_ratecontrol_t *rc = h->rc;
1119 if( h->param.rc.b_stat_read )
1121 if( frame_num >= rc->num_entries )
1123 /* We could try to initialize everything required for ABR and
1124 * adaptive B-frames, but that would be complicated.
1125 * So just calculate the average QP used so far. */
1128 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24
1129 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1130 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
1131 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 );
1132 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 );
1134 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1135 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1136 if( h->param.i_bframe_adaptive )
1137 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1139 for( i = 0; i < h->param.i_threads; i++ )
1141 h->thread[i]->rc->b_abr = 0;
1142 h->thread[i]->rc->b_2pass = 0;
1143 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1144 h->thread[i]->param.rc.b_stat_read = 0;
1145 h->thread[i]->param.i_bframe_adaptive = 0;
1146 h->thread[i]->param.i_scenecut_threshold = 0;
1147 if( h->thread[i]->param.i_bframe > 1 )
1148 h->thread[i]->param.i_bframe = 1;
1150 return X264_TYPE_AUTO;
1152 switch( rc->entry[frame_num].pict_type )
1155 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I;
1158 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B;
1167 return X264_TYPE_AUTO;
1171 /* After encoding one frame, save stats and update ratecontrol state */
1172 int x264_ratecontrol_end( x264_t *h, int bits )
1174 x264_ratecontrol_t *rc = h->rc;
1175 const int *mbs = h->stat.frame.i_mb_count;
1180 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1181 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1182 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1183 for( i = B_DIRECT; i < B_8x8; i++ )
1184 h->stat.frame.i_mb_count_p += mbs[i];
1186 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1187 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1189 if( h->param.rc.b_stat_write )
1191 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1192 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1193 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1194 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1195 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1196 char c_direct = h->mb.b_direct_auto_write ?
1197 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1198 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1200 if( fprintf( rc->p_stat_file_out,
1201 "in:%d out:%d type:%c q:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c;\n",
1202 h->fenc->i_frame, h->i_frame,
1204 h->stat.frame.i_tex_bits,
1205 h->stat.frame.i_mv_bits,
1206 h->stat.frame.i_misc_bits,
1207 h->stat.frame.i_mb_count_i,
1208 h->stat.frame.i_mb_count_p,
1209 h->stat.frame.i_mb_count_skip,
1213 /* Don't re-write the data in multi-pass mode. */
1214 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1216 uint8_t i_type = h->sh.i_type;
1218 /* Values are stored as big-endian FIX8.8 */
1219 for( i = 0; i < h->mb.i_mb_count; i++ )
1220 rc->qp_buffer[i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1221 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1223 if( fwrite( rc->qp_buffer, sizeof(uint16_t), h->mb.i_mb_count, rc->p_mbtree_stat_file_out ) < h->mb.i_mb_count )
1230 if( h->sh.i_type != SLICE_TYPE_B )
1231 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / rc->last_rceq;
1234 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1235 * Not perfectly accurate with B-refs, but good enough. */
1236 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
1238 rc->cplxr_sum *= rc->cbr_decay;
1239 rc->wanted_bits_window += rc->bitrate / rc->fps;
1240 rc->wanted_bits_window *= rc->cbr_decay;
1245 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
1248 if( h->mb.b_variable_qp )
1250 if( h->sh.i_type == SLICE_TYPE_B )
1252 rc->bframe_bits += bits;
1253 if( !h->frames.current[0] || !IS_X264_TYPE_B(h->frames.current[0]->i_type) )
1255 update_predictor( rc->pred_b_from_p, qp2qscale(rc->qpa_rc),
1256 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1257 rc->bframe_bits = 0;
1262 update_vbv( h, bits );
1265 x264_log(h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n");
1269 /****************************************************************************
1271 ***************************************************************************/
1274 * modify the bitrate curve from pass1 for one frame
1276 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1278 x264_ratecontrol_t *rcc= h->rc;
1280 x264_zone_t *zone = get_zone( h, frame_num );
1282 q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1284 // avoid NaN's in the rc_eq
1285 if(!isfinite(q) || rce->tex_bits + rce->mv_bits == 0)
1286 q = rcc->last_qscale;
1291 rcc->last_qscale = q;
1296 if( zone->b_force_qp )
1297 q = qp2qscale(zone->i_qp);
1299 q /= zone->f_bitrate_factor;
1305 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1307 x264_ratecontrol_t *rcc = h->rc;
1308 const int pict_type = rce->pict_type;
1310 // force I/B quants as a function of P quants
1311 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1312 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1313 if( pict_type == SLICE_TYPE_I )
1316 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1317 double ip_factor = fabs( h->param.rc.f_ip_factor );
1318 /* don't apply ip_factor if the following frame is also I */
1319 if( rcc->accum_p_norm <= 0 )
1321 else if( h->param.rc.f_ip_factor < 0 )
1323 else if( rcc->accum_p_norm >= 1 )
1326 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1328 else if( pict_type == SLICE_TYPE_B )
1330 if( h->param.rc.f_pb_factor > 0 )
1332 if( !rce->kept_as_ref )
1333 q *= fabs( h->param.rc.f_pb_factor );
1335 else if( pict_type == SLICE_TYPE_P
1336 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1337 && rce->tex_bits == 0 )
1342 /* last qscale / qdiff stuff */
1343 if(rcc->last_non_b_pict_type==pict_type
1344 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1346 double last_q = rcc->last_qscale_for[pict_type];
1347 double max_qscale = last_q * rcc->lstep;
1348 double min_qscale = last_q / rcc->lstep;
1350 if (q > max_qscale) q = max_qscale;
1351 else if(q < min_qscale) q = min_qscale;
1354 rcc->last_qscale_for[pict_type] = q;
1355 if(pict_type!=SLICE_TYPE_B)
1356 rcc->last_non_b_pict_type = pict_type;
1357 if(pict_type==SLICE_TYPE_I)
1359 rcc->last_accum_p_norm = rcc->accum_p_norm;
1360 rcc->accum_p_norm = 0;
1361 rcc->accum_p_qp = 0;
1363 if(pict_type==SLICE_TYPE_P)
1365 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1366 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1367 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1372 static double predict_size( predictor_t *p, double q, double var )
1374 return (p->coeff*var + p->offset) / (q*p->count);
1377 static void update_predictor( predictor_t *p, double q, double var, double bits )
1379 const double range = 1.5;
1382 double old_coeff = p->coeff / p->count;
1383 double new_coeff = bits*q / var;
1384 double new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
1385 double new_offset = bits*q - new_coeff_clipped * var;
1386 if( new_offset >= 0 )
1387 new_coeff = new_coeff_clipped;
1390 p->count *= p->decay;
1391 p->coeff *= p->decay;
1392 p->offset *= p->decay;
1394 p->coeff += new_coeff;
1395 p->offset += new_offset;
1398 // update VBV after encoding a frame
1399 static void update_vbv( x264_t *h, int bits )
1401 x264_ratecontrol_t *rcc = h->rc;
1402 x264_ratecontrol_t *rct = h->thread[0]->rc;
1404 if( rcc->last_satd >= h->mb.i_mb_count )
1405 update_predictor( &rct->pred[h->sh.i_type], qp2qscale(rcc->qpa_rc), rcc->last_satd, bits );
1410 rct->buffer_fill_final += rct->buffer_rate - bits;
1411 if( rct->buffer_fill_final < 0 )
1412 x264_log( h, X264_LOG_WARNING, "VBV underflow (%.0f bits)\n", rct->buffer_fill_final );
1413 rct->buffer_fill_final = x264_clip3f( rct->buffer_fill_final, 0, rct->buffer_size );
1416 // provisionally update VBV according to the planned size of all frames currently in progress
1417 static void update_vbv_plan( x264_t *h )
1419 x264_ratecontrol_t *rcc = h->rc;
1420 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final;
1421 if( h->param.i_threads > 1 )
1423 int j = h->rc - h->thread[0]->rc;
1425 for( i=1; i<h->param.i_threads; i++ )
1427 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1428 double bits = t->rc->frame_size_planned;
1429 if( !t->b_thread_active )
1431 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1432 rcc->buffer_fill += rcc->buffer_rate - bits;
1433 rcc->buffer_fill = x264_clip3( rcc->buffer_fill, 0, rcc->buffer_size );
1438 // apply VBV constraints and clip qscale to between lmin and lmax
1439 static double clip_qscale( x264_t *h, int pict_type, double q )
1441 x264_ratecontrol_t *rcc = h->rc;
1442 double lmin = rcc->lmin[pict_type];
1443 double lmax = rcc->lmax[pict_type];
1446 /* B-frames are not directly subject to VBV,
1447 * since they are controlled by the P-frames' QPs. */
1449 if( rcc->b_vbv && rcc->last_satd > 0 )
1451 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
1452 * the lookahead overflow and such that the buffer is in a reasonable state
1453 * by the end of the lookahead. */
1454 if( h->param.rc.i_lookahead )
1456 int j, iterations, terminate = 0;
1458 /* Avoid an infinite loop. */
1459 for( iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
1462 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1463 double buffer_fill_cur = rcc->buffer_fill - cur_bits + rcc->buffer_rate;
1465 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
1466 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
1467 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
1469 /* Loop over the planned future frames. */
1470 for( j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
1472 int i_type = h->fenc->i_planned_type[j];
1473 int i_satd = h->fenc->i_planned_satd[j];
1474 if( i_type == X264_TYPE_AUTO )
1476 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
1477 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
1478 buffer_fill_cur = buffer_fill_cur - cur_bits + rcc->buffer_rate;
1480 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
1481 target_fill = X264_MIN( rcc->buffer_fill + j * rcc->buffer_rate * 0.5, rcc->buffer_size * 0.5 );
1482 if( buffer_fill_cur < target_fill )
1488 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
1489 target_fill = x264_clip3f( rcc->buffer_fill - j * rcc->buffer_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
1490 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
1499 /* Fallback to old purely-reactive algorithm: no lookahead. */
1502 if( ( pict_type == SLICE_TYPE_P ||
1503 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
1504 rcc->buffer_fill/rcc->buffer_size < 0.5 )
1506 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1509 /* Now a hard threshold to make sure the frame fits in VBV.
1510 * This one is mostly for I-frames. */
1511 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1513 if( bits > rcc->buffer_fill/2 )
1514 qf = x264_clip3f( rcc->buffer_fill/(2*bits), 0.2, 1.0 );
1517 if( bits < rcc->buffer_rate/2 )
1518 q *= bits*2/rcc->buffer_rate;
1519 q = X264_MAX( q0, q );
1522 /* Check B-frame complexity, and use up any bits that would
1523 * overflow before the next P-frame. */
1524 if( h->sh.i_type == SLICE_TYPE_P )
1526 int nb = rcc->bframes;
1527 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1528 double pbbits = bits;
1529 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1532 if( bbits > rcc->buffer_rate )
1534 pbbits += nb * bbits;
1536 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1537 if( pbbits < space )
1539 q *= X264_MAX( pbbits / space,
1540 bits / (0.5 * rcc->buffer_size) );
1542 q = X264_MAX( q0-5, q );
1545 if( !rcc->b_vbv_min_rate )
1546 q = X264_MAX( q0, q );
1551 else if(rcc->b_2pass)
1553 double min2 = log(lmin);
1554 double max2 = log(lmax);
1555 q = (log(q) - min2)/(max2-min2) - 0.5;
1556 q = 1.0/(1.0 + exp(-4*q));
1557 q = q*(max2-min2) + min2;
1561 return x264_clip3f(q, lmin, lmax);
1564 // update qscale for 1 frame based on actual bits used so far
1565 static float rate_estimate_qscale( x264_t *h )
1568 x264_ratecontrol_t *rcc = h->rc;
1569 ratecontrol_entry_t rce;
1570 int pict_type = h->sh.i_type;
1571 double lmin = rcc->lmin[pict_type];
1572 double lmax = rcc->lmax[pict_type];
1573 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
1574 + h->stat.i_frame_size[SLICE_TYPE_P]
1575 + h->stat.i_frame_size[SLICE_TYPE_B]);
1580 if(pict_type != rce.pict_type)
1582 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1583 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1587 if( pict_type == SLICE_TYPE_B )
1589 /* B-frames don't have independent ratecontrol, but rather get the
1590 * average QP of the two adjacent P-frames + an offset */
1592 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1593 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1594 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1595 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1596 float q0 = h->fref0[0]->f_qp_avg_rc;
1597 float q1 = h->fref1[0]->f_qp_avg_rc;
1599 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1600 q0 -= rcc->pb_offset/2;
1601 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1602 q1 -= rcc->pb_offset/2;
1605 q = (q0 + q1) / 2 + rcc->ip_offset;
1611 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1613 if(h->fenc->b_kept_as_ref)
1614 q += rcc->pb_offset/2;
1616 q += rcc->pb_offset;
1618 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1619 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1621 return qp2qscale(q);
1625 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1629 //FIXME adjust abr_buffer based on distance to the end of the video
1631 int64_t predicted_bits = total_bits;
1635 if( h->param.i_threads > 1 )
1637 int j = h->rc - h->thread[0]->rc;
1639 for( i=1; i<h->param.i_threads; i++ )
1641 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1642 double bits = t->rc->frame_size_planned;
1643 if( !t->b_thread_active )
1645 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1646 predicted_bits += (int64_t)bits;
1652 if( h->fenc->i_frame < h->param.i_threads )
1653 predicted_bits += (int64_t)h->fenc->i_frame * rcc->bitrate / rcc->fps;
1655 predicted_bits += (int64_t)(h->param.i_threads - 1) * rcc->bitrate / rcc->fps;
1658 diff = predicted_bits - (int64_t)rce.expected_bits;
1660 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1661 if( ((h->fenc->i_frame + 1 - h->param.i_threads) >= rcc->fps) &&
1662 (rcc->expected_bits_sum > 0))
1664 /* Adjust quant based on the difference between
1665 * achieved and expected bitrate so far */
1666 double time = (double)h->fenc->i_frame / rcc->num_entries;
1667 double w = x264_clip3f( time*100, 0.0, 1.0 );
1668 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1672 /* Do not overflow vbv */
1673 double expected_size = qscale2bits(&rce, q);
1674 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1675 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
1676 double qmax = q*(2 - expected_fullness);
1677 double size_constraint = 1 + expected_fullness;
1678 qmax = X264_MAX(qmax, rce.new_qscale);
1679 if (expected_fullness < .05)
1681 qmax = X264_MIN(qmax, lmax);
1682 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
1683 ((expected_vbv < 0) && (q < lmax)))
1686 expected_size = qscale2bits(&rce, q);
1687 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1689 rcc->last_satd = x264_stack_align( x264_rc_analyse_slice, h );
1691 q = x264_clip3f( q, lmin, lmax );
1693 else /* 1pass ABR */
1695 /* Calculate the quantizer which would have produced the desired
1696 * average bitrate if it had been applied to all frames so far.
1697 * Then modulate that quant based on the current frame's complexity
1698 * relative to the average complexity so far (using the 2pass RCEQ).
1699 * Then bias the quant up or down if total size so far was far from
1701 * Result: Depending on the value of rate_tolerance, there is a
1702 * tradeoff between quality and bitrate precision. But at large
1703 * tolerances, the bit distribution approaches that of 2pass. */
1705 double wanted_bits, overflow=1, lmin, lmax;
1707 rcc->last_satd = x264_stack_align( x264_rc_analyse_slice, h );
1708 rcc->short_term_cplxsum *= 0.5;
1709 rcc->short_term_cplxcount *= 0.5;
1710 rcc->short_term_cplxsum += rcc->last_satd;
1711 rcc->short_term_cplxcount ++;
1713 rce.tex_bits = rcc->last_satd;
1714 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1716 rce.p_count = rcc->nmb;
1720 rce.pict_type = pict_type;
1722 if( h->param.rc.i_rc_method == X264_RC_CRF )
1724 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1728 int i_frame_done = h->fenc->i_frame + 1 - h->param.i_threads;
1730 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1732 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
1733 wanted_bits = i_frame_done * rcc->bitrate / rcc->fps;
1734 if( wanted_bits > 0 )
1736 abr_buffer *= X264_MAX( 1, sqrt(i_frame_done/25) );
1737 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1742 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1743 /* should test _next_ pict type, but that isn't decided yet */
1744 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1746 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1747 q /= fabs( h->param.rc.f_ip_factor );
1749 else if( h->i_frame > 0 )
1751 /* Asymmetric clipping, because symmetric would prevent
1752 * overflow control in areas of rapidly oscillating complexity */
1753 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1754 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1755 if( overflow > 1.1 && h->i_frame > 3 )
1757 else if( overflow < 0.9 )
1760 q = x264_clip3f(q, lmin, lmax);
1762 else if( h->param.rc.i_rc_method == X264_RC_CRF )
1764 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1766 rcc->qp_novbv = qscale2qp(q);
1768 //FIXME use get_diff_limited_q() ?
1769 q = clip_qscale( h, pict_type, q );
1772 rcc->last_qscale_for[pict_type] =
1773 rcc->last_qscale = q;
1775 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
1776 rcc->last_qscale_for[SLICE_TYPE_P] = q;
1778 if( rcc->b_2pass && rcc->b_vbv )
1779 rcc->frame_size_planned = qscale2bits(&rce, q);
1781 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1782 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1787 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
1791 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
1792 /* these vars are updated in x264_ratecontrol_start()
1793 * so copy them from the context that most recently started (prev)
1794 * to the context that's about to start (cur).
1800 COPY(last_qscale_for);
1801 COPY(last_non_b_pict_type);
1802 COPY(short_term_cplxsum);
1803 COPY(short_term_cplxcount);
1810 #define COPY(var) next->rc->var = cur->rc->var
1811 /* these vars are updated in x264_ratecontrol_end()
1812 * so copy them from the context that most recently ended (cur)
1813 * to the context that's about to end (next)
1816 COPY(expected_bits_sum);
1817 COPY(wanted_bits_window);
1821 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
1822 /* the rest of the variables are either constant or thread-local */
1825 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
1827 /* find an interval ending on an overflow or underflow (depending on whether
1828 * we're adding or removing bits), and starting on the earliest frame that
1829 * can influence the buffer fill of that end frame. */
1830 x264_ratecontrol_t *rcc = h->rc;
1831 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
1832 const double buffer_max = .9 * rcc->buffer_size;
1833 double fill = fills[*t0-1];
1834 double parity = over ? 1. : -1.;
1835 int i, start=-1, end=-1;
1836 for(i = *t0; i < rcc->num_entries; i++)
1838 fill += (rcc->buffer_rate - qscale2bits(&rcc->entry[i], rcc->entry[i].new_qscale)) * parity;
1839 fill = x264_clip3f(fill, 0, rcc->buffer_size);
1841 if(fill <= buffer_min || i == 0)
1847 else if(fill >= buffer_max && start >= 0)
1852 return start>=0 && end>=0;
1855 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
1857 x264_ratecontrol_t *rcc = h->rc;
1858 double qscale_orig, qscale_new;
1863 for(i = t0; i <= t1; i++)
1865 qscale_orig = rcc->entry[i].new_qscale;
1866 qscale_orig = x264_clip3f(qscale_orig, qscale_min, qscale_max);
1867 qscale_new = qscale_orig * adjustment;
1868 qscale_new = x264_clip3f(qscale_new, qscale_min, qscale_max);
1869 rcc->entry[i].new_qscale = qscale_new;
1870 adjusted = adjusted || (qscale_new != qscale_orig);
1875 static double count_expected_bits( x264_t *h )
1877 x264_ratecontrol_t *rcc = h->rc;
1878 double expected_bits = 0;
1880 for(i = 0; i < rcc->num_entries; i++)
1882 ratecontrol_entry_t *rce = &rcc->entry[i];
1883 rce->expected_bits = expected_bits;
1884 expected_bits += qscale2bits(rce, rce->new_qscale);
1886 return expected_bits;
1889 static int vbv_pass2( x264_t *h )
1891 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
1892 * frames in the interval until either buffer is full at some intermediate frame or the
1893 * last frame in the interval no longer underflows. Recompute intervals and repeat.
1894 * Then do the converse to put bits back into overflow areas until target size is met */
1896 x264_ratecontrol_t *rcc = h->rc;
1898 double all_available_bits = h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps;
1899 double expected_bits = 0;
1901 double prev_bits = 0;
1903 double qscale_min = qp2qscale(h->param.rc.i_qp_min);
1904 double qscale_max = qp2qscale(h->param.rc.i_qp_max);
1906 int adj_min, adj_max;
1907 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
1911 /* adjust overall stream size */
1915 prev_bits = expected_bits;
1917 if(expected_bits != 0)
1918 { /* not first iteration */
1919 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
1920 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
1924 while(adj_min && find_underflow(h, fills, &t0, &t1, 1))
1926 adj_min = fix_underflow(h, t0, t1, adjustment, qscale_min, qscale_max);
1931 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
1933 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
1935 while(adj_max && find_underflow(h, fills, &t0, &t1, 0))
1936 adj_max = fix_underflow(h, t0, t1, 1.001, qscale_min, qscale_max);
1938 expected_bits = count_expected_bits(h);
1939 } while((expected_bits < .995*all_available_bits) && ((int)(expected_bits+.5) > (int)(prev_bits+.5)) );
1942 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
1944 /* store expected vbv filling values for tracking when encoding */
1945 for(i = 0; i < rcc->num_entries; i++)
1946 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
1954 static int init_pass2( x264_t *h )
1956 x264_ratecontrol_t *rcc = h->rc;
1957 uint64_t all_const_bits = 0;
1958 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps);
1959 double rate_factor, step, step_mult;
1960 double qblur = h->param.rc.f_qblur;
1961 double cplxblur = h->param.rc.f_complexity_blur;
1962 const int filter_size = (int)(qblur*4) | 1;
1963 double expected_bits;
1964 double *qscale, *blurred_qscale;
1967 /* find total/average complexity & const_bits */
1968 for(i=0; i<rcc->num_entries; i++)
1970 ratecontrol_entry_t *rce = &rcc->entry[i];
1971 all_const_bits += rce->misc_bits;
1974 if( all_available_bits < all_const_bits)
1976 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
1977 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)));
1981 /* Blur complexities, to reduce local fluctuation of QP.
1982 * We don't blur the QPs directly, because then one very simple frame
1983 * could drag down the QP of a nearby complex frame and give it more
1984 * bits than intended. */
1985 for(i=0; i<rcc->num_entries; i++)
1987 ratecontrol_entry_t *rce = &rcc->entry[i];
1988 double weight_sum = 0;
1989 double cplx_sum = 0;
1990 double weight = 1.0;
1991 double gaussian_weight;
1993 /* weighted average of cplx of future frames */
1994 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++)
1996 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
1997 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2000 gaussian_weight = weight * exp(-j*j/200.0);
2001 weight_sum += gaussian_weight;
2002 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2004 /* weighted average of cplx of past frames */
2006 for(j=0; j<=cplxblur*2 && j<=i; j++)
2008 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2009 gaussian_weight = weight * exp(-j*j/200.0);
2010 weight_sum += gaussian_weight;
2011 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2012 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2016 rce->blurred_complexity = cplx_sum / weight_sum;
2019 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2020 if( filter_size > 1 )
2021 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2023 blurred_qscale = qscale;
2025 /* Search for a factor which, when multiplied by the RCEQ values from
2026 * each frame, adds up to the desired total size.
2027 * There is no exact closed-form solution because of VBV constraints and
2028 * because qscale2bits is not invertible, but we can start with the simple
2029 * approximation of scaling the 1st pass by the ratio of bitrates.
2030 * The search range is probably overkill, but speed doesn't matter here. */
2033 for(i=0; i<rcc->num_entries; i++)
2034 expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i));
2035 step_mult = all_available_bits / expected_bits;
2038 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2041 rate_factor += step;
2043 rcc->last_non_b_pict_type = -1;
2044 rcc->last_accum_p_norm = 1;
2045 rcc->accum_p_norm = 0;
2048 for(i=0; i<rcc->num_entries; i++)
2050 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
2053 /* fixed I/B qscale relative to P */
2054 for(i=rcc->num_entries-1; i>=0; i--)
2056 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
2057 assert(qscale[i] >= 0);
2063 assert(filter_size%2==1);
2064 for(i=0; i<rcc->num_entries; i++)
2066 ratecontrol_entry_t *rce = &rcc->entry[i];
2068 double q=0.0, sum=0.0;
2070 for(j=0; j<filter_size; j++)
2072 int index = i+j-filter_size/2;
2074 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
2075 if(index < 0 || index >= rcc->num_entries)
2077 if(rce->pict_type != rcc->entry[index].pict_type)
2079 q += qscale[index] * coeff;
2082 blurred_qscale[i] = q/sum;
2086 /* find expected bits */
2087 for(i=0; i<rcc->num_entries; i++)
2089 ratecontrol_entry_t *rce = &rcc->entry[i];
2090 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
2091 assert(rce->new_qscale >= 0);
2092 expected_bits += qscale2bits(rce, rce->new_qscale);
2095 if(expected_bits > all_available_bits) rate_factor -= step;
2100 x264_free(blurred_qscale);
2103 if( vbv_pass2( h ) )
2105 expected_bits = count_expected_bits(h);
2107 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
2110 for(i=0; i<rcc->num_entries; i++)
2111 avgq += rcc->entry[i].new_qscale;
2112 avgq = qscale2qp(avgq / rcc->num_entries);
2114 if ((expected_bits > all_available_bits) || (!rcc->b_vbv))
2115 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
2116 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
2117 (float)h->param.rc.i_bitrate,
2118 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
2120 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
2122 if(h->param.rc.i_qp_min > 0)
2123 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
2125 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
2127 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
2129 if(h->param.rc.i_qp_max < 51)
2130 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
2132 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
2134 else if(!(rcc->b_2pass && rcc->b_vbv))
2135 x264_log(h, X264_LOG_WARNING, "internal error\n");