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 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 );
274 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame )
276 x264_ratecontrol_t *rc = h->rc;
277 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
280 if( i_type_actual != SLICE_TYPE_B )
284 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
287 if( i_type != i_type_actual )
289 x264_log(h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type,i_type_actual);
293 if( fread( rc->qp_buffer, sizeof(uint16_t), h->mb.i_mb_count, rc->p_mbtree_stat_file_in ) != h->mb.i_mb_count )
296 for( i = 0; i < h->mb.i_mb_count; i++ )
297 frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[i] )) * (1/256.0);
300 x264_adaptive_quant_frame( h, frame );
303 x264_log(h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n");
307 int x264_reference_build_list_optimal( x264_t *h )
309 ratecontrol_entry_t *rce = h->rc->rce;
310 x264_frame_t *frames[16];
313 if( rce->refs != h->i_ref0 )
316 memcpy( frames, h->fref0, sizeof(frames) );
318 /* For now don't reorder ref 0; it seems to lower quality
319 in most cases due to skips. */
320 for( ref = 1; ref < h->i_ref0; ref++ )
324 for( i = 1; i < h->i_ref0; i++ )
325 /* Favor lower POC as a tiebreaker. */
326 COPY2_IF_GT( max, rce->refcount[i], bestref, i );
327 rce->refcount[bestref] = -1;
328 h->fref0[ref] = frames[bestref];
334 static char *x264_strcat_filename( char *input, char *suffix )
336 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
339 strcpy( output, input );
340 strcat( output, suffix );
344 int x264_ratecontrol_new( x264_t *h )
346 x264_ratecontrol_t *rc;
351 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
354 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
355 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
357 /* FIXME: use integers */
358 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
359 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
363 if( h->param.rc.b_mb_tree )
365 h->param.rc.f_pb_factor = 1;
369 rc->qcompress = h->param.rc.f_qcompress;
371 rc->bitrate = h->param.rc.i_bitrate * 1000.;
372 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
373 rc->nmb = h->mb.i_mb_count;
374 rc->last_non_b_pict_type = -1;
377 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
379 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
382 if( h->param.rc.i_vbv_buffer_size )
384 if( h->param.rc.i_rc_method == X264_RC_CQP )
386 x264_log(h, X264_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
387 h->param.rc.i_vbv_max_bitrate = 0;
388 h->param.rc.i_vbv_buffer_size = 0;
390 else if( h->param.rc.i_vbv_max_bitrate == 0 )
392 x264_log( h, X264_LOG_DEBUG, "VBV maxrate unspecified, assuming CBR\n" );
393 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
396 if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate &&
397 h->param.rc.i_vbv_max_bitrate > 0)
398 x264_log(h, X264_LOG_WARNING, "max bitrate less than average bitrate, ignored.\n");
399 else if( h->param.rc.i_vbv_max_bitrate > 0 &&
400 h->param.rc.i_vbv_buffer_size > 0 )
402 if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
404 h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
405 x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
406 h->param.rc.i_vbv_buffer_size );
408 if( h->param.rc.f_vbv_buffer_init > 1. )
409 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 );
410 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000. / rc->fps;
411 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000.;
412 rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
413 h->param.rc.f_vbv_buffer_init = X264_MAX( h->param.rc.f_vbv_buffer_init, rc->buffer_rate / rc->buffer_size );
414 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
415 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
416 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
418 rc->b_vbv_min_rate = !rc->b_2pass
419 && h->param.rc.i_rc_method == X264_RC_ABR
420 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
422 else if( h->param.rc.i_vbv_max_bitrate )
424 x264_log(h, X264_LOG_WARNING, "VBV maxrate specified, but no bufsize.\n");
425 h->param.rc.i_vbv_max_bitrate = 0;
427 if(rc->rate_tolerance < 0.01)
429 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
430 rc->rate_tolerance = 0.01;
433 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
437 /* FIXME ABR_INIT_QP is actually used only in CRF */
438 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
439 rc->accum_p_norm = .01;
440 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
441 /* estimated ratio that produces a reasonable QP for the first I-frame */
442 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
443 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
444 rc->last_non_b_pict_type = SLICE_TYPE_I;
447 if( h->param.rc.i_rc_method == X264_RC_CRF )
449 /* Arbitrary rescaling to make CRF somewhat similar to QP.
450 * Try to compensate for MB-tree's effects as well. */
451 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
452 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
453 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
454 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset );
457 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
458 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
459 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
460 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
461 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
463 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
464 rc->last_qscale = qp2qscale(26);
465 CHECKED_MALLOC( rc->pred, 5*sizeof(predictor_t) );
466 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
467 for( i = 0; i < 5; i++ )
469 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
470 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
471 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
472 rc->pred[i].coeff= 2.0;
473 rc->pred[i].count= 1.0;
474 rc->pred[i].decay= 0.5;
475 rc->pred[i].offset= 0.0;
476 for( j = 0; j < 2; j++ )
478 rc->row_preds[i][j].coeff= .25;
479 rc->row_preds[i][j].count= 1.0;
480 rc->row_preds[i][j].decay= 0.5;
481 rc->row_preds[i][j].offset= 0.0;
484 *rc->pred_b_from_p = rc->pred[0];
486 if( parse_zones( h ) < 0 )
488 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
492 /* Load stat file and init 2pass algo */
493 if( h->param.rc.b_stat_read )
495 char *p, *stats_in, *stats_buf;
497 /* read 1st pass stats */
498 assert( h->param.rc.psz_stat_in );
499 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
502 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
505 if( h->param.rc.b_mb_tree )
507 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
508 if( !mbtree_stats_in )
510 rc->p_mbtree_stat_file_in = fopen( mbtree_stats_in, "rb" );
511 x264_free( mbtree_stats_in );
512 if( !rc->p_mbtree_stat_file_in )
514 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
519 /* check whether 1st pass options were compatible with current options */
520 if( !strncmp( stats_buf, "#options:", 9 ) )
523 char *opts = stats_buf;
524 stats_in = strchr( stats_buf, '\n' );
530 if( ( p = strstr( opts, "bframes=" ) ) && sscanf( p, "bframes=%d", &i )
531 && h->param.i_bframe != i )
533 x264_log( h, X264_LOG_ERROR, "different number of B-frames than 1st pass (%d vs %d)\n",
534 h->param.i_bframe, i );
538 /* since B-adapt doesn't (yet) take into account B-pyramid,
539 * the converse is not a problem */
540 if( strstr( opts, "b_pyramid=1" ) && !h->param.b_bframe_pyramid )
541 x264_log( h, X264_LOG_WARNING, "1st pass used B-pyramid, 2nd doesn't\n" );
543 if( ( p = strstr( opts, "keyint=" ) ) && sscanf( p, "keyint=%d", &i )
544 && h->param.i_keyint_max != i )
545 x264_log( h, X264_LOG_WARNING, "different keyint than 1st pass (%d vs %d)\n",
546 h->param.i_keyint_max, i );
548 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
549 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
551 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
553 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
554 h->mb.b_direct_auto_write = 1;
557 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
558 h->param.i_bframe_adaptive = i;
559 else if( h->param.i_bframe )
561 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
565 if( h->param.rc.b_mb_tree && ( p = strstr( opts, "rc_lookahead=" ) ) && sscanf( p, "rc_lookahead=%d", &i ) )
566 h->param.rc.i_lookahead = i;
569 /* find number of pics */
572 p = strchr(p+1, ';');
575 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
580 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
582 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
583 h->param.i_frame_total, rc->num_entries );
585 if( h->param.i_frame_total > rc->num_entries )
587 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
588 h->param.i_frame_total, rc->num_entries );
592 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
594 /* init all to skipped p frames */
595 for(i=0; i<rc->num_entries; i++)
597 ratecontrol_entry_t *rce = &rc->entry[i];
598 rce->pict_type = SLICE_TYPE_P;
599 rce->qscale = rce->new_qscale = qp2qscale(20);
600 rce->misc_bits = rc->nmb + 10;
606 for(i=0; i < rc->num_entries; i++)
608 ratecontrol_entry_t *rce;
616 next= strchr(p, ';');
619 (*next)=0; //sscanf is unbelievably slow on long strings
622 e = sscanf(p, " in:%d ", &frame_number);
624 if(frame_number < 0 || frame_number >= rc->num_entries)
626 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
629 rce = &rc->entry[frame_number];
630 rce->direct_mode = 0;
632 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",
633 &pict_type, &qp, &rce->tex_bits,
634 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
635 &rce->s_count, &rce->direct_mode);
637 p = strstr( p, "ref:" );
641 for( ref = 0; ref < 16; ref++ )
643 if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
645 p = strchr( p+1, ' ' );
653 case 'I': rce->kept_as_ref = 1;
654 case 'i': rce->pict_type = SLICE_TYPE_I; break;
655 case 'P': rce->pict_type = SLICE_TYPE_P; break;
656 case 'B': rce->kept_as_ref = 1;
657 case 'b': rce->pict_type = SLICE_TYPE_B; break;
658 default: e = -1; break;
663 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
666 rce->qscale = qp2qscale(qp);
670 x264_free(stats_buf);
672 if(h->param.rc.i_rc_method == X264_RC_ABR)
674 if(init_pass2(h) < 0) return -1;
675 } /* else we're using constant quant, so no need to run the bitrate allocation */
678 /* Open output file */
679 /* If input and output files are the same, output to a temp file
680 * and move it to the real name only when it's complete */
681 if( h->param.rc.b_stat_write )
684 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
685 if( !rc->psz_stat_file_tmpname )
688 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
689 if( rc->p_stat_file_out == NULL )
691 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
695 p = x264_param2string( &h->param, 1 );
697 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
699 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
701 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
702 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
703 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
706 rc->p_mbtree_stat_file_out = fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
707 if( rc->p_mbtree_stat_file_out == NULL )
709 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
715 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
716 CHECKED_MALLOC( rc->qp_buffer, h->mb.i_mb_count * sizeof(uint16_t) );
718 for( i=0; i<h->param.i_threads; i++ )
720 h->thread[i]->rc = rc+i;
724 memcpy( &h->thread[i]->param, &h->param, sizeof(x264_param_t) );
725 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
734 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
737 char *tok, UNUSED *saveptr;
739 z->f_bitrate_factor = 1;
740 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
742 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
744 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
748 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
754 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
755 memcpy( z->param, &h->param, sizeof(x264_param_t) );
756 z->param->param_free = x264_free;
757 while( (tok = strtok_r( p, ",", &saveptr )) )
759 char *val = strchr( tok, '=' );
765 if( x264_param_parse( z->param, tok, val ) )
767 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
777 static int parse_zones( x264_t *h )
779 x264_ratecontrol_t *rc = h->rc;
781 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
783 char *psz_zones, *p, *tok, UNUSED *saveptr;
784 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
785 strcpy( psz_zones, h->param.rc.psz_zones );
786 h->param.rc.i_zones = 1;
787 for( p = psz_zones; *p; p++ )
788 h->param.rc.i_zones += (*p == '/');
789 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
791 for( i = 0; i < h->param.rc.i_zones; i++ )
793 tok = strtok_r( p, "/", &saveptr );
794 if( !tok || parse_zone( h, &h->param.rc.zones[i], tok ) )
798 x264_free( psz_zones );
801 if( h->param.rc.i_zones > 0 )
803 for( i = 0; i < h->param.rc.i_zones; i++ )
805 x264_zone_t z = h->param.rc.zones[i];
806 if( z.i_start < 0 || z.i_start > z.i_end )
808 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
809 z.i_start, z.i_end );
812 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
814 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
815 z.f_bitrate_factor );
820 rc->i_zones = h->param.rc.i_zones + 1;
821 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
822 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
824 // default zone to fall back to if none of the others match
825 rc->zones[0].i_start = 0;
826 rc->zones[0].i_end = INT_MAX;
827 rc->zones[0].b_force_qp = 0;
828 rc->zones[0].f_bitrate_factor = 1;
829 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
830 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
831 for( i = 1; i < rc->i_zones; i++ )
833 if( !rc->zones[i].param )
834 rc->zones[i].param = rc->zones[0].param;
843 static x264_zone_t *get_zone( x264_t *h, int frame_num )
846 for( i = h->rc->i_zones-1; i >= 0; i-- )
848 x264_zone_t *z = &h->rc->zones[i];
849 if( frame_num >= z->i_start && frame_num <= z->i_end )
855 void x264_ratecontrol_summary( x264_t *h )
857 x264_ratecontrol_t *rc = h->rc;
858 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
860 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
861 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
862 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
863 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
864 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset );
868 void x264_ratecontrol_delete( x264_t *h )
870 x264_ratecontrol_t *rc = h->rc;
873 if( rc->p_stat_file_out )
875 fclose( rc->p_stat_file_out );
876 if( h->i_frame >= rc->num_entries )
877 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
879 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
880 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
882 x264_free( rc->psz_stat_file_tmpname );
884 if( rc->p_mbtree_stat_file_out )
886 fclose( rc->p_mbtree_stat_file_out );
887 if( h->i_frame >= rc->num_entries )
888 if( rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
890 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
891 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
893 x264_free( rc->psz_mbtree_stat_file_tmpname );
894 x264_free( rc->psz_mbtree_stat_file_name );
896 if( rc->p_mbtree_stat_file_in )
897 fclose( rc->p_mbtree_stat_file_in );
898 x264_free( rc->pred );
899 x264_free( rc->pred_b_from_p );
900 x264_free( rc->entry );
901 x264_free( rc->qp_buffer );
904 x264_free( rc->zones[0].param );
905 for( i=1; i<rc->i_zones; i++ )
906 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
907 rc->zones[i].param->param_free( rc->zones[i].param );
908 x264_free( rc->zones );
913 void x264_ratecontrol_set_estimated_size( x264_t *h, int bits )
915 x264_pthread_mutex_lock( &h->fenc->mutex );
916 h->rc->frame_size_estimated = bits;
917 x264_pthread_mutex_unlock( &h->fenc->mutex );
920 int x264_ratecontrol_get_estimated_size( x264_t const *h)
923 x264_pthread_mutex_lock( &h->fenc->mutex );
924 size = h->rc->frame_size_estimated;
925 x264_pthread_mutex_unlock( &h->fenc->mutex );
929 static void accum_p_qp_update( x264_t *h, float qp )
931 x264_ratecontrol_t *rc = h->rc;
932 rc->accum_p_qp *= .95;
933 rc->accum_p_norm *= .95;
934 rc->accum_p_norm += 1;
935 if( h->sh.i_type == SLICE_TYPE_I )
936 rc->accum_p_qp += qp + rc->ip_offset;
938 rc->accum_p_qp += qp;
941 /* Before encoding a frame, choose a QP for it */
942 void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
944 x264_ratecontrol_t *rc = h->rc;
945 ratecontrol_entry_t *rce = NULL;
946 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
951 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
952 x264_encoder_reconfig( h, zone->param );
953 rc->prev_zone = zone;
955 rc->qp_force = i_force_qp;
957 if( h->param.rc.b_stat_read )
959 int frame = h->fenc->i_frame;
960 assert( frame >= 0 && frame < rc->num_entries );
961 rce = h->rc->rce = &h->rc->entry[frame];
963 if( h->sh.i_type == SLICE_TYPE_B
964 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
966 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
967 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
973 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
974 rc->row_pred = &rc->row_preds[h->sh.i_type];
975 update_vbv_plan( h, overhead );
978 if( h->sh.i_type != SLICE_TYPE_B )
979 rc->bframes = h->fenc->i_bframes;
987 q = qscale2qp( rate_estimate_qscale( h ) );
989 else if( rc->b_2pass )
991 rce->new_qscale = rate_estimate_qscale( h );
992 q = qscale2qp( rce->new_qscale );
996 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
997 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
999 q = rc->qp_constant[ h->sh.i_type ];
1003 if( zone->b_force_qp )
1004 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1006 q -= 6*log(zone->f_bitrate_factor)/log(2);
1010 q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1014 h->fdec->f_qp_avg_rc =
1015 h->fdec->f_qp_avg_aq =
1017 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
1020 rce->new_qp = rc->qp;
1022 accum_p_qp_update( h, rc->qp );
1024 if( h->sh.i_type != SLICE_TYPE_B )
1025 rc->last_non_b_pict_type = h->sh.i_type;
1028 static double predict_row_size( x264_t *h, int y, int qp )
1030 /* average between two predictors:
1031 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1032 x264_ratecontrol_t *rc = h->rc;
1033 double pred_s = predict_size( rc->row_pred[0], qp2qscale(qp), h->fdec->i_row_satd[y] );
1035 if( h->sh.i_type == SLICE_TYPE_I || qp >= h->fref0[0]->i_row_qp[y] )
1037 if( h->sh.i_type != SLICE_TYPE_I
1038 && h->fref0[0]->i_type == h->fdec->i_type
1039 && h->fref0[0]->i_row_satd[y] > 0
1040 && (abs(h->fref0[0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1042 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
1043 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
1047 return (pred_s + pred_t) / 2;
1049 /* Our QP is lower than the reference! */
1052 double newq = qp2qscale(qp);
1053 double oldq = qp2qscale(h->fref0[0]->i_row_qp[y]);
1054 double pred_intra = predict_size( rc->row_pred[1], (1 - newq / oldq) * newq, h->fdec->i_row_satds[0][0][y] );
1055 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1056 return pred_intra + pred_s;
1060 static double row_bits_so_far( x264_t *h, int y )
1064 for( i = 0; i <= y; i++ )
1065 bits += h->fdec->i_row_bits[i];
1069 static double predict_row_size_sum( x264_t *h, int y, int qp )
1072 double bits = row_bits_so_far(h, y);
1073 for( i = y+1; i < h->sps->i_mb_height; i++ )
1074 bits += predict_row_size( h, i, qp );
1079 void x264_ratecontrol_mb( x264_t *h, int bits )
1081 x264_ratecontrol_t *rc = h->rc;
1082 const int y = h->mb.i_mb_y;
1086 h->fdec->i_row_bits[y] += bits;
1087 rc->qpa_rc += rc->f_qpm;
1088 rc->qpa_aq += h->mb.i_qp;
1090 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !rc->b_vbv )
1093 h->fdec->i_row_qp[y] = rc->qpm;
1095 if( h->sh.i_type == SLICE_TYPE_B )
1097 /* B-frames shouldn't use lower QP than their reference frames.
1098 * This code is a bit overzealous in limiting B-frame quantizers, but it helps avoid
1099 * underflows due to the fact that B-frames are not explicitly covered by VBV. */
1100 if( y < h->sps->i_mb_height-1 )
1103 int avg_qp = X264_MIN(h->fref0[0]->i_row_qp[y+1], h->fref1[0]->i_row_qp[y+1])
1104 + rc->pb_offset * ((h->fenc->i_type == X264_TYPE_BREF) ? 0.5 : 1);
1105 rc->qpm = X264_MIN(X264_MAX( rc->qp, avg_qp), 51); //avg_qp could go higher than 51 due to pb_offset
1106 i_estimated = row_bits_so_far(h, y); //FIXME: compute full estimated size
1107 if (i_estimated > h->rc->frame_size_planned)
1108 x264_ratecontrol_set_estimated_size(h, i_estimated);
1113 update_predictor( rc->row_pred[0], qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1114 if( h->sh.i_type != SLICE_TYPE_I && rc->qpm < h->fref0[0]->i_row_qp[y] )
1116 double newq = qp2qscale(rc->qpm);
1117 double oldq = qp2qscale(h->fref0[0]->i_row_qp[y]);
1118 update_predictor( rc->row_pred[1], (1 - newq / oldq) * newq, h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1121 /* tweak quality based on difference from predicted size */
1122 if( y < h->sps->i_mb_height-1 )// && h->stat.i_frame_count[h->sh.i_type] > 0 )
1124 int prev_row_qp = h->fdec->i_row_qp[y];
1125 int b0 = predict_row_size_sum( h, y, rc->qpm );
1127 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
1128 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1129 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1131 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits.
1132 * In 2-pass mode we can be more trusting of the planned frame sizes, since they were decided
1133 * by actual encoding instead of SATD prediction. */
1134 float rc_tol = h->param.rc.b_stat_read ? (buffer_left_planned / rc->buffer_size) * rc->frame_size_planned
1135 : (buffer_left_planned / h->param.i_threads);
1136 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1137 /* area at the top of the frame was measured inaccurately. */
1138 if( row_bits_so_far(h,y) < 0.05 * rc->frame_size_planned )
1141 if(h->sh.i_type != SLICE_TYPE_I)
1144 if( !rc->b_vbv_min_rate )
1145 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
1147 while( rc->qpm < i_qp_max
1148 && ((b1 > rc->frame_size_planned + rc_tol) ||
1149 (rc->buffer_fill - b1 < buffer_left_planned * 0.5) ||
1150 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1153 b1 = predict_row_size_sum( h, y, rc->qpm );
1156 while( rc->qpm > i_qp_min
1157 && (rc->qpm > h->fdec->i_row_qp[0] || rc->single_frame_vbv)
1158 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
1159 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
1162 b1 = predict_row_size_sum( h, y, rc->qpm );
1165 /* avoid VBV underflow */
1166 while( (rc->qpm < h->param.rc.i_qp_max)
1167 && (rc->buffer_fill - b1 < rc->buffer_rate * 0.05 ) )
1170 b1 = predict_row_size_sum( h, y, rc->qpm );
1173 x264_ratecontrol_set_estimated_size(h, b1);
1176 /* loses the fractional part of the frame-wise qp */
1177 rc->f_qpm = rc->qpm;
1180 int x264_ratecontrol_qp( x264_t *h )
1185 /* In 2pass, force the same frame types as in the 1st pass */
1186 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1188 x264_ratecontrol_t *rc = h->rc;
1189 if( h->param.rc.b_stat_read )
1191 if( frame_num >= rc->num_entries )
1193 /* We could try to initialize everything required for ABR and
1194 * adaptive B-frames, but that would be complicated.
1195 * So just calculate the average QP used so far. */
1198 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24
1199 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1200 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
1201 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 );
1202 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 );
1204 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1205 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1206 if( h->param.i_bframe_adaptive )
1207 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1209 for( i = 0; i < h->param.i_threads; i++ )
1211 h->thread[i]->rc->b_abr = 0;
1212 h->thread[i]->rc->b_2pass = 0;
1213 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1214 h->thread[i]->param.rc.b_stat_read = 0;
1215 h->thread[i]->param.i_bframe_adaptive = 0;
1216 h->thread[i]->param.i_scenecut_threshold = 0;
1217 if( h->thread[i]->param.i_bframe > 1 )
1218 h->thread[i]->param.i_bframe = 1;
1220 return X264_TYPE_AUTO;
1222 switch( rc->entry[frame_num].pict_type )
1225 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I;
1228 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B;
1237 return X264_TYPE_AUTO;
1241 /* After encoding one frame, save stats and update ratecontrol state */
1242 int x264_ratecontrol_end( x264_t *h, int bits )
1244 x264_ratecontrol_t *rc = h->rc;
1245 const int *mbs = h->stat.frame.i_mb_count;
1250 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1251 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1252 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1253 for( i = B_DIRECT; i < B_8x8; i++ )
1254 h->stat.frame.i_mb_count_p += mbs[i];
1256 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1257 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1259 if( h->param.rc.b_stat_write )
1261 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1262 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1263 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1264 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1265 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1266 char c_direct = h->mb.b_direct_auto_write ?
1267 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1268 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1270 if( fprintf( rc->p_stat_file_out,
1271 "in:%d out:%d type:%c q:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c ref:",
1272 h->fenc->i_frame, h->i_frame,
1274 h->stat.frame.i_tex_bits,
1275 h->stat.frame.i_mv_bits,
1276 h->stat.frame.i_misc_bits,
1277 h->stat.frame.i_mb_count_i,
1278 h->stat.frame.i_mb_count_p,
1279 h->stat.frame.i_mb_count_skip,
1283 for( i = 0; i < h->i_ref0; i++ )
1285 int refcount = h->param.b_interlaced ? h->stat.frame.i_mb_count_ref[0][i*2]
1286 + h->stat.frame.i_mb_count_ref[0][i*2+1] :
1287 h->stat.frame.i_mb_count_ref[0][i];
1288 if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1292 if( fprintf( rc->p_stat_file_out, ";\n" ) < 0 )
1295 /* Don't re-write the data in multi-pass mode. */
1296 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1298 uint8_t i_type = h->sh.i_type;
1300 /* Values are stored as big-endian FIX8.8 */
1301 for( i = 0; i < h->mb.i_mb_count; i++ )
1302 rc->qp_buffer[i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1303 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1305 if( fwrite( rc->qp_buffer, sizeof(uint16_t), h->mb.i_mb_count, rc->p_mbtree_stat_file_out ) < h->mb.i_mb_count )
1312 if( h->sh.i_type != SLICE_TYPE_B )
1313 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / rc->last_rceq;
1316 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1317 * Not perfectly accurate with B-refs, but good enough. */
1318 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
1320 rc->cplxr_sum *= rc->cbr_decay;
1321 rc->wanted_bits_window += rc->bitrate / rc->fps;
1322 rc->wanted_bits_window *= rc->cbr_decay;
1327 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
1330 if( h->mb.b_variable_qp )
1332 if( h->sh.i_type == SLICE_TYPE_B )
1334 rc->bframe_bits += bits;
1335 if( h->fenc->b_last_minigop_bframe )
1337 update_predictor( rc->pred_b_from_p, qp2qscale(rc->qpa_rc),
1338 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1339 rc->bframe_bits = 0;
1344 update_vbv( h, bits );
1347 x264_log(h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n");
1351 /****************************************************************************
1353 ***************************************************************************/
1356 * modify the bitrate curve from pass1 for one frame
1358 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1360 x264_ratecontrol_t *rcc= h->rc;
1362 x264_zone_t *zone = get_zone( h, frame_num );
1364 q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1366 // avoid NaN's in the rc_eq
1367 if(!isfinite(q) || rce->tex_bits + rce->mv_bits == 0)
1368 q = rcc->last_qscale;
1373 rcc->last_qscale = q;
1378 if( zone->b_force_qp )
1379 q = qp2qscale(zone->i_qp);
1381 q /= zone->f_bitrate_factor;
1387 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1389 x264_ratecontrol_t *rcc = h->rc;
1390 const int pict_type = rce->pict_type;
1392 // force I/B quants as a function of P quants
1393 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1394 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1395 if( pict_type == SLICE_TYPE_I )
1398 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1399 double ip_factor = fabs( h->param.rc.f_ip_factor );
1400 /* don't apply ip_factor if the following frame is also I */
1401 if( rcc->accum_p_norm <= 0 )
1403 else if( h->param.rc.f_ip_factor < 0 )
1405 else if( rcc->accum_p_norm >= 1 )
1408 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1410 else if( pict_type == SLICE_TYPE_B )
1412 if( h->param.rc.f_pb_factor > 0 )
1414 if( !rce->kept_as_ref )
1415 q *= fabs( h->param.rc.f_pb_factor );
1417 else if( pict_type == SLICE_TYPE_P
1418 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1419 && rce->tex_bits == 0 )
1424 /* last qscale / qdiff stuff */
1425 if(rcc->last_non_b_pict_type==pict_type
1426 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1428 double last_q = rcc->last_qscale_for[pict_type];
1429 double max_qscale = last_q * rcc->lstep;
1430 double min_qscale = last_q / rcc->lstep;
1432 if (q > max_qscale) q = max_qscale;
1433 else if(q < min_qscale) q = min_qscale;
1436 rcc->last_qscale_for[pict_type] = q;
1437 if(pict_type!=SLICE_TYPE_B)
1438 rcc->last_non_b_pict_type = pict_type;
1439 if(pict_type==SLICE_TYPE_I)
1441 rcc->last_accum_p_norm = rcc->accum_p_norm;
1442 rcc->accum_p_norm = 0;
1443 rcc->accum_p_qp = 0;
1445 if(pict_type==SLICE_TYPE_P)
1447 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1448 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1449 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1454 static double predict_size( predictor_t *p, double q, double var )
1456 return (p->coeff*var + p->offset) / (q*p->count);
1459 static void update_predictor( predictor_t *p, double q, double var, double bits )
1461 const double range = 1.5;
1464 double old_coeff = p->coeff / p->count;
1465 double new_coeff = bits*q / var;
1466 double new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
1467 double new_offset = bits*q - new_coeff_clipped * var;
1468 if( new_offset >= 0 )
1469 new_coeff = new_coeff_clipped;
1472 p->count *= p->decay;
1473 p->coeff *= p->decay;
1474 p->offset *= p->decay;
1476 p->coeff += new_coeff;
1477 p->offset += new_offset;
1480 // update VBV after encoding a frame
1481 static void update_vbv( x264_t *h, int bits )
1483 x264_ratecontrol_t *rcc = h->rc;
1484 x264_ratecontrol_t *rct = h->thread[0]->rc;
1486 if( rcc->last_satd >= h->mb.i_mb_count )
1487 update_predictor( &rct->pred[h->sh.i_type], qp2qscale(rcc->qpa_rc), rcc->last_satd, bits );
1492 rct->buffer_fill_final -= bits;
1493 if( rct->buffer_fill_final < 0 )
1494 x264_log( h, X264_LOG_WARNING, "VBV underflow (frame %d, %.0f bits)\n", h->i_frame, rct->buffer_fill_final );
1495 rct->buffer_fill_final = X264_MAX( rct->buffer_fill_final, 0 );
1496 rct->buffer_fill_final += rct->buffer_rate;
1497 rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, rct->buffer_size );
1500 // provisionally update VBV according to the planned size of all frames currently in progress
1501 static void update_vbv_plan( x264_t *h, int overhead )
1503 x264_ratecontrol_t *rcc = h->rc;
1504 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final - overhead;
1505 if( h->param.i_threads > 1 )
1507 int j = h->rc - h->thread[0]->rc;
1509 for( i=1; i<h->param.i_threads; i++ )
1511 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1512 double bits = t->rc->frame_size_planned;
1513 if( !t->b_thread_active )
1515 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1516 rcc->buffer_fill -= bits;
1517 rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
1518 rcc->buffer_fill += rcc->buffer_rate;
1519 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1524 // apply VBV constraints and clip qscale to between lmin and lmax
1525 static double clip_qscale( x264_t *h, int pict_type, double q )
1527 x264_ratecontrol_t *rcc = h->rc;
1528 double lmin = rcc->lmin[pict_type];
1529 double lmax = rcc->lmax[pict_type];
1532 /* B-frames are not directly subject to VBV,
1533 * since they are controlled by the P-frames' QPs. */
1535 if( rcc->b_vbv && rcc->last_satd > 0 )
1537 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
1538 * the lookahead overflow and such that the buffer is in a reasonable state
1539 * by the end of the lookahead. */
1540 if( h->param.rc.i_lookahead )
1542 int j, iterations, terminate = 0;
1544 /* Avoid an infinite loop. */
1545 for( iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
1548 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1549 double buffer_fill_cur = rcc->buffer_fill - cur_bits + rcc->buffer_rate;
1551 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
1552 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
1553 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
1555 /* Loop over the planned future frames. */
1556 for( j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
1558 int i_type = h->fenc->i_planned_type[j];
1559 int i_satd = h->fenc->i_planned_satd[j];
1560 if( i_type == X264_TYPE_AUTO )
1562 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
1563 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
1564 buffer_fill_cur = buffer_fill_cur - cur_bits + rcc->buffer_rate;
1566 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
1567 target_fill = X264_MIN( rcc->buffer_fill + j * rcc->buffer_rate * 0.5, rcc->buffer_size * 0.5 );
1568 if( buffer_fill_cur < target_fill )
1574 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
1575 target_fill = x264_clip3f( rcc->buffer_fill - j * rcc->buffer_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
1576 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
1585 /* Fallback to old purely-reactive algorithm: no lookahead. */
1588 if( ( pict_type == SLICE_TYPE_P ||
1589 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
1590 rcc->buffer_fill/rcc->buffer_size < 0.5 )
1592 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1595 /* Now a hard threshold to make sure the frame fits in VBV.
1596 * This one is mostly for I-frames. */
1597 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1599 /* For small VBVs, allow the frame to use up the entire VBV. */
1600 double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
1601 /* For single-frame VBVs, request that the frame use up the entire VBV. */
1602 double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
1604 if( bits > rcc->buffer_fill/max_fill_factor )
1605 qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
1608 if( bits < rcc->buffer_rate/min_fill_factor )
1609 q *= bits*min_fill_factor/rcc->buffer_rate;
1610 q = X264_MAX( q0, q );
1613 /* Check B-frame complexity, and use up any bits that would
1614 * overflow before the next P-frame. */
1615 if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
1617 int nb = rcc->bframes;
1618 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1619 double pbbits = bits;
1620 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1622 if( bbits > rcc->buffer_rate )
1624 pbbits += nb * bbits;
1626 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1627 if( pbbits < space )
1629 q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
1631 q = X264_MAX( q0-5, q );
1634 if( !rcc->b_vbv_min_rate )
1635 q = X264_MAX( q0, q );
1640 else if(rcc->b_2pass)
1642 double min2 = log(lmin);
1643 double max2 = log(lmax);
1644 q = (log(q) - min2)/(max2-min2) - 0.5;
1645 q = 1.0/(1.0 + exp(-4*q));
1646 q = q*(max2-min2) + min2;
1650 return x264_clip3f(q, lmin, lmax);
1653 // update qscale for 1 frame based on actual bits used so far
1654 static float rate_estimate_qscale( x264_t *h )
1657 x264_ratecontrol_t *rcc = h->rc;
1658 ratecontrol_entry_t rce;
1659 int pict_type = h->sh.i_type;
1660 double lmin = rcc->lmin[pict_type];
1661 double lmax = rcc->lmax[pict_type];
1662 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
1663 + h->stat.i_frame_size[SLICE_TYPE_P]
1664 + h->stat.i_frame_size[SLICE_TYPE_B]);
1669 if(pict_type != rce.pict_type)
1671 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1672 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1676 if( pict_type == SLICE_TYPE_B )
1678 /* B-frames don't have independent ratecontrol, but rather get the
1679 * average QP of the two adjacent P-frames + an offset */
1681 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1682 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1683 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1684 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1685 float q0 = h->fref0[0]->f_qp_avg_rc;
1686 float q1 = h->fref1[0]->f_qp_avg_rc;
1688 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1689 q0 -= rcc->pb_offset/2;
1690 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1691 q1 -= rcc->pb_offset/2;
1694 q = (q0 + q1) / 2 + rcc->ip_offset;
1700 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1702 if(h->fenc->b_kept_as_ref)
1703 q += rcc->pb_offset/2;
1705 q += rcc->pb_offset;
1707 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1708 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1710 return qp2qscale(q);
1714 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1718 //FIXME adjust abr_buffer based on distance to the end of the video
1720 int64_t predicted_bits = total_bits;
1724 if( h->param.i_threads > 1 )
1726 int j = h->rc - h->thread[0]->rc;
1728 for( i=1; i<h->param.i_threads; i++ )
1730 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1731 double bits = t->rc->frame_size_planned;
1732 if( !t->b_thread_active )
1734 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1735 predicted_bits += (int64_t)bits;
1741 if( h->fenc->i_frame < h->param.i_threads )
1742 predicted_bits += (int64_t)h->fenc->i_frame * rcc->bitrate / rcc->fps;
1744 predicted_bits += (int64_t)(h->param.i_threads - 1) * rcc->bitrate / rcc->fps;
1747 diff = predicted_bits - (int64_t)rce.expected_bits;
1749 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1750 if( ((h->fenc->i_frame + 1 - h->param.i_threads) >= rcc->fps) &&
1751 (rcc->expected_bits_sum > 0))
1753 /* Adjust quant based on the difference between
1754 * achieved and expected bitrate so far */
1755 double time = (double)h->fenc->i_frame / rcc->num_entries;
1756 double w = x264_clip3f( time*100, 0.0, 1.0 );
1757 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1761 /* Do not overflow vbv */
1762 double expected_size = qscale2bits(&rce, q);
1763 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1764 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
1765 double qmax = q*(2 - expected_fullness);
1766 double size_constraint = 1 + expected_fullness;
1767 qmax = X264_MAX(qmax, rce.new_qscale);
1768 if (expected_fullness < .05)
1770 qmax = X264_MIN(qmax, lmax);
1771 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
1772 ((expected_vbv < 0) && (q < lmax)))
1775 expected_size = qscale2bits(&rce, q);
1776 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1778 rcc->last_satd = x264_stack_align( x264_rc_analyse_slice, h );
1780 q = x264_clip3f( q, lmin, lmax );
1782 else /* 1pass ABR */
1784 /* Calculate the quantizer which would have produced the desired
1785 * average bitrate if it had been applied to all frames so far.
1786 * Then modulate that quant based on the current frame's complexity
1787 * relative to the average complexity so far (using the 2pass RCEQ).
1788 * Then bias the quant up or down if total size so far was far from
1790 * Result: Depending on the value of rate_tolerance, there is a
1791 * tradeoff between quality and bitrate precision. But at large
1792 * tolerances, the bit distribution approaches that of 2pass. */
1794 double wanted_bits, overflow=1, lmin, lmax;
1796 rcc->last_satd = x264_stack_align( x264_rc_analyse_slice, h );
1797 rcc->short_term_cplxsum *= 0.5;
1798 rcc->short_term_cplxcount *= 0.5;
1799 rcc->short_term_cplxsum += rcc->last_satd;
1800 rcc->short_term_cplxcount ++;
1802 rce.tex_bits = rcc->last_satd;
1803 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1805 rce.p_count = rcc->nmb;
1809 rce.pict_type = pict_type;
1811 if( h->param.rc.i_rc_method == X264_RC_CRF )
1813 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1817 int i_frame_done = h->fenc->i_frame + 1 - h->param.i_threads;
1819 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1821 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
1822 wanted_bits = i_frame_done * rcc->bitrate / rcc->fps;
1823 if( wanted_bits > 0 )
1825 abr_buffer *= X264_MAX( 1, sqrt(i_frame_done/25) );
1826 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1831 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1832 /* should test _next_ pict type, but that isn't decided yet */
1833 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1835 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1836 q /= fabs( h->param.rc.f_ip_factor );
1838 else if( h->i_frame > 0 )
1840 /* Asymmetric clipping, because symmetric would prevent
1841 * overflow control in areas of rapidly oscillating complexity */
1842 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1843 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1844 if( overflow > 1.1 && h->i_frame > 3 )
1846 else if( overflow < 0.9 )
1849 q = x264_clip3f(q, lmin, lmax);
1851 else if( h->param.rc.i_rc_method == X264_RC_CRF )
1853 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1855 rcc->qp_novbv = qscale2qp(q);
1857 //FIXME use get_diff_limited_q() ?
1858 q = clip_qscale( h, pict_type, q );
1861 rcc->last_qscale_for[pict_type] =
1862 rcc->last_qscale = q;
1864 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
1865 rcc->last_qscale_for[SLICE_TYPE_P] = q;
1867 if( rcc->b_2pass && rcc->b_vbv )
1868 rcc->frame_size_planned = qscale2bits(&rce, q);
1870 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1871 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1876 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
1880 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
1881 /* these vars are updated in x264_ratecontrol_start()
1882 * so copy them from the context that most recently started (prev)
1883 * to the context that's about to start (cur).
1889 COPY(last_qscale_for);
1890 COPY(last_non_b_pict_type);
1891 COPY(short_term_cplxsum);
1892 COPY(short_term_cplxcount);
1899 #define COPY(var) next->rc->var = cur->rc->var
1900 /* these vars are updated in x264_ratecontrol_end()
1901 * so copy them from the context that most recently ended (cur)
1902 * to the context that's about to end (next)
1905 COPY(expected_bits_sum);
1906 COPY(wanted_bits_window);
1910 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
1911 /* the rest of the variables are either constant or thread-local */
1914 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
1916 /* find an interval ending on an overflow or underflow (depending on whether
1917 * we're adding or removing bits), and starting on the earliest frame that
1918 * can influence the buffer fill of that end frame. */
1919 x264_ratecontrol_t *rcc = h->rc;
1920 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
1921 const double buffer_max = .9 * rcc->buffer_size;
1922 double fill = fills[*t0-1];
1923 double parity = over ? 1. : -1.;
1924 int i, start=-1, end=-1;
1925 for(i = *t0; i < rcc->num_entries; i++)
1927 fill += (rcc->buffer_rate - qscale2bits(&rcc->entry[i], rcc->entry[i].new_qscale)) * parity;
1928 fill = x264_clip3f(fill, 0, rcc->buffer_size);
1930 if(fill <= buffer_min || i == 0)
1936 else if(fill >= buffer_max && start >= 0)
1941 return start>=0 && end>=0;
1944 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
1946 x264_ratecontrol_t *rcc = h->rc;
1947 double qscale_orig, qscale_new;
1952 for(i = t0; i <= t1; i++)
1954 qscale_orig = rcc->entry[i].new_qscale;
1955 qscale_orig = x264_clip3f(qscale_orig, qscale_min, qscale_max);
1956 qscale_new = qscale_orig * adjustment;
1957 qscale_new = x264_clip3f(qscale_new, qscale_min, qscale_max);
1958 rcc->entry[i].new_qscale = qscale_new;
1959 adjusted = adjusted || (qscale_new != qscale_orig);
1964 static double count_expected_bits( x264_t *h )
1966 x264_ratecontrol_t *rcc = h->rc;
1967 double expected_bits = 0;
1969 for(i = 0; i < rcc->num_entries; i++)
1971 ratecontrol_entry_t *rce = &rcc->entry[i];
1972 rce->expected_bits = expected_bits;
1973 expected_bits += qscale2bits(rce, rce->new_qscale);
1975 return expected_bits;
1978 static int vbv_pass2( x264_t *h )
1980 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
1981 * frames in the interval until either buffer is full at some intermediate frame or the
1982 * last frame in the interval no longer underflows. Recompute intervals and repeat.
1983 * Then do the converse to put bits back into overflow areas until target size is met */
1985 x264_ratecontrol_t *rcc = h->rc;
1987 double all_available_bits = h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps;
1988 double expected_bits = 0;
1990 double prev_bits = 0;
1992 double qscale_min = qp2qscale(h->param.rc.i_qp_min);
1993 double qscale_max = qp2qscale(h->param.rc.i_qp_max);
1995 int adj_min, adj_max;
1996 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2000 /* adjust overall stream size */
2004 prev_bits = expected_bits;
2006 if(expected_bits != 0)
2007 { /* not first iteration */
2008 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2009 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2013 while(adj_min && find_underflow(h, fills, &t0, &t1, 1))
2015 adj_min = fix_underflow(h, t0, t1, adjustment, qscale_min, qscale_max);
2020 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2022 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2024 while(adj_max && find_underflow(h, fills, &t0, &t1, 0))
2025 adj_max = fix_underflow(h, t0, t1, 1.001, qscale_min, qscale_max);
2027 expected_bits = count_expected_bits(h);
2028 } while((expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2031 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2033 /* store expected vbv filling values for tracking when encoding */
2034 for(i = 0; i < rcc->num_entries; i++)
2035 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
2043 static int init_pass2( x264_t *h )
2045 x264_ratecontrol_t *rcc = h->rc;
2046 uint64_t all_const_bits = 0;
2047 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps);
2048 double rate_factor, step, step_mult;
2049 double qblur = h->param.rc.f_qblur;
2050 double cplxblur = h->param.rc.f_complexity_blur;
2051 const int filter_size = (int)(qblur*4) | 1;
2052 double expected_bits;
2053 double *qscale, *blurred_qscale;
2056 /* find total/average complexity & const_bits */
2057 for(i=0; i<rcc->num_entries; i++)
2059 ratecontrol_entry_t *rce = &rcc->entry[i];
2060 all_const_bits += rce->misc_bits;
2063 if( all_available_bits < all_const_bits)
2065 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2066 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)));
2070 /* Blur complexities, to reduce local fluctuation of QP.
2071 * We don't blur the QPs directly, because then one very simple frame
2072 * could drag down the QP of a nearby complex frame and give it more
2073 * bits than intended. */
2074 for(i=0; i<rcc->num_entries; i++)
2076 ratecontrol_entry_t *rce = &rcc->entry[i];
2077 double weight_sum = 0;
2078 double cplx_sum = 0;
2079 double weight = 1.0;
2080 double gaussian_weight;
2082 /* weighted average of cplx of future frames */
2083 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++)
2085 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2086 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2089 gaussian_weight = weight * exp(-j*j/200.0);
2090 weight_sum += gaussian_weight;
2091 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2093 /* weighted average of cplx of past frames */
2095 for(j=0; j<=cplxblur*2 && j<=i; j++)
2097 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2098 gaussian_weight = weight * exp(-j*j/200.0);
2099 weight_sum += gaussian_weight;
2100 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2101 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2105 rce->blurred_complexity = cplx_sum / weight_sum;
2108 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2109 if( filter_size > 1 )
2110 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2112 blurred_qscale = qscale;
2114 /* Search for a factor which, when multiplied by the RCEQ values from
2115 * each frame, adds up to the desired total size.
2116 * There is no exact closed-form solution because of VBV constraints and
2117 * because qscale2bits is not invertible, but we can start with the simple
2118 * approximation of scaling the 1st pass by the ratio of bitrates.
2119 * The search range is probably overkill, but speed doesn't matter here. */
2122 for(i=0; i<rcc->num_entries; i++)
2123 expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i));
2124 step_mult = all_available_bits / expected_bits;
2127 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2130 rate_factor += step;
2132 rcc->last_non_b_pict_type = -1;
2133 rcc->last_accum_p_norm = 1;
2134 rcc->accum_p_norm = 0;
2137 for(i=0; i<rcc->num_entries; i++)
2139 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
2142 /* fixed I/B qscale relative to P */
2143 for(i=rcc->num_entries-1; i>=0; i--)
2145 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
2146 assert(qscale[i] >= 0);
2152 assert(filter_size%2==1);
2153 for(i=0; i<rcc->num_entries; i++)
2155 ratecontrol_entry_t *rce = &rcc->entry[i];
2157 double q=0.0, sum=0.0;
2159 for(j=0; j<filter_size; j++)
2161 int index = i+j-filter_size/2;
2163 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
2164 if(index < 0 || index >= rcc->num_entries)
2166 if(rce->pict_type != rcc->entry[index].pict_type)
2168 q += qscale[index] * coeff;
2171 blurred_qscale[i] = q/sum;
2175 /* find expected bits */
2176 for(i=0; i<rcc->num_entries; i++)
2178 ratecontrol_entry_t *rce = &rcc->entry[i];
2179 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
2180 assert(rce->new_qscale >= 0);
2181 expected_bits += qscale2bits(rce, rce->new_qscale);
2184 if(expected_bits > all_available_bits) rate_factor -= step;
2189 x264_free(blurred_qscale);
2192 if( vbv_pass2( h ) )
2194 expected_bits = count_expected_bits(h);
2196 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
2199 for(i=0; i<rcc->num_entries; i++)
2200 avgq += rcc->entry[i].new_qscale;
2201 avgq = qscale2qp(avgq / rcc->num_entries);
2203 if ((expected_bits > all_available_bits) || (!rcc->b_vbv))
2204 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
2205 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
2206 (float)h->param.rc.i_bitrate,
2207 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
2209 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
2211 if(h->param.rc.i_qp_min > 0)
2212 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
2214 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
2216 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
2218 if(h->param.rc.i_qp_max < 51)
2219 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
2221 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
2223 else if(!(rcc->b_2pass && rcc->b_vbv))
2224 x264_log(h, X264_LOG_WARNING, "internal error\n");