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
33 #include "common/common.h"
34 #include "common/cpu.h"
35 #include "ratecontrol.h"
45 uint64_t expected_bits; /*total expected bits up to the current frame (current one excluded)*/
52 float blurred_complexity;
54 } ratecontrol_entry_t;
64 struct x264_ratecontrol_t
73 double rate_tolerance;
75 int nmb; /* number of macroblocks in a frame */
79 ratecontrol_entry_t *rce;
80 int qp; /* qp for current frame */
81 int qpm; /* qp for current macroblock */
82 float f_qpm; /* qp for current macroblock: precise float for AQ */
83 float qpa_rc; /* average of macroblocks' qp before aq */
84 float qpa_aq; /* average of macroblocks' qp after aq */
85 float qp_novbv; /* QP for the current frame if 1-pass VBV was disabled. */
90 double buffer_fill_final; /* real buffer as of the last finished frame */
91 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
92 double buffer_rate; /* # of bits added to buffer_fill after each frame */
93 predictor_t *pred; /* predict frame size from satd */
98 double cplxr_sum; /* sum of bits*qscale/rceq */
99 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow, only includes finished frames */
100 double wanted_bits_window; /* target bitrate * window */
102 double short_term_cplxsum;
103 double short_term_cplxcount;
104 double rate_factor_constant;
109 FILE *p_stat_file_out;
110 char *psz_stat_file_tmpname;
111 FILE *p_mbtree_stat_file_out;
112 char *psz_mbtree_stat_file_tmpname;
113 char *psz_mbtree_stat_file_name;
114 FILE *p_mbtree_stat_file_in;
116 int num_entries; /* number of ratecontrol_entry_ts */
117 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
119 double last_qscale_for[5]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
120 int last_non_b_pict_type;
121 double accum_p_qp; /* for determining I-frame quant */
123 double last_accum_p_norm;
124 double lmin[5]; /* min qscale by frame type */
126 double lstep; /* max change (multiply) in qscale per frame */
127 uint16_t *qp_buffer; /* Global buffer for converting MB-tree quantizer data. */
130 double frame_size_estimated;
131 double frame_size_planned;
132 predictor_t *row_pred;
133 predictor_t row_preds[5];
134 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
135 int bframes; /* # consecutive B-frames before this P-frame */
136 int bframe_bits; /* total cost of those frames */
140 x264_zone_t *prev_zone;
144 static int parse_zones( x264_t *h );
145 static int init_pass2(x264_t *);
146 static float rate_estimate_qscale( x264_t *h );
147 static void update_vbv( x264_t *h, int bits );
148 static void update_vbv_plan( x264_t *h );
149 static double predict_size( predictor_t *p, double q, double var );
150 static void update_predictor( predictor_t *p, double q, double var, double bits );
153 * qp = h.264's quantizer
154 * qscale = linearized quantizer = Lagrange multiplier
156 static inline double qp2qscale(double qp)
158 return 0.85 * pow(2.0, ( qp - 12.0 ) / 6.0);
160 static inline double qscale2qp(double qscale)
162 return 12.0 + 6.0 * log(qscale/0.85) / log(2.0);
165 /* Texture bitrate is not quite inversely proportional to qscale,
166 * probably due the the changing number of SKIP blocks.
167 * MV bits level off at about qp<=12, because the lambda used
168 * for motion estimation is constant there. */
169 static inline double qscale2bits(ratecontrol_entry_t *rce, double qscale)
173 return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
174 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
178 // Find the total AC energy of the block in all planes.
179 static NOINLINE uint32_t ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
181 /* This function contains annoying hacks because GCC has a habit of reordering emms
182 * and putting it after floating point ops. As a result, we put the emms at the end of the
183 * function and make sure that its always called before the float math. Noinline makes
184 * sure no reordering goes on. */
186 for( i = 0; i < 3; i++ )
189 int stride = frame->i_stride[i];
190 int offset = h->mb.b_interlaced
191 ? w * (mb_x + (mb_y&~1) * stride) + (mb_y&1) * stride
192 : w * (mb_x + mb_y * stride);
193 int pix = i ? PIXEL_8x8 : PIXEL_16x16;
194 stride <<= h->mb.b_interlaced;
195 var += h->pixf.var[pix]( frame->plane[i]+offset, stride );
201 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame )
203 /* constants chosen to result in approximately the same overall bitrate as without AQ.
204 * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
208 /* Need to init it anyways for MB tree. */
209 if( h->param.rc.f_aq_strength == 0 )
212 memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
213 memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
214 if( h->frames.b_have_lowres )
215 for( mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
216 frame->i_inv_qscale_factor[mb_xy] = 256;
220 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
222 for( mb_y = 0; mb_y < h->sps->i_mb_height; mb_y++ )
223 for( mb_x = 0; mb_x < h->sps->i_mb_width; mb_x++ )
225 uint32_t energy = ac_energy_mb( h, mb_x, mb_y, frame );
226 float qp_adj = x264_log2( energy + 2 );
228 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
231 avg_adj /= h->mb.i_mb_count;
232 strength = h->param.rc.f_aq_strength * avg_adj * (1.f / 6000.f);
235 strength = h->param.rc.f_aq_strength * 1.0397f;
237 for( mb_y = 0; mb_y < h->sps->i_mb_height; mb_y++ )
238 for( mb_x = 0; mb_x < h->sps->i_mb_width; mb_x++ )
241 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
243 qp_adj = frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride];
244 qp_adj = strength * (qp_adj - avg_adj);
248 uint32_t energy = ac_energy_mb( h, mb_x, mb_y, frame );
249 qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - 14.427f);
251 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] =
252 frame->f_qp_offset_aq[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
253 if( h->frames.b_have_lowres )
254 frame->i_inv_qscale_factor[mb_x + mb_y*h->mb.i_mb_stride] = x264_exp2fix8(qp_adj);
259 /*****************************************************************************
260 * x264_adaptive_quant:
261 * adjust macroblock QP based on variance (AC energy) of the MB.
262 * high variance = higher QP
263 * low variance = lower QP
264 * This generally increases SSIM and lowers PSNR.
265 *****************************************************************************/
266 void x264_adaptive_quant( x264_t *h )
269 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 );
272 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame )
274 x264_ratecontrol_t *rc = h->rc;
275 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
278 if( i_type_actual != SLICE_TYPE_B )
282 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
285 if( i_type != i_type_actual )
287 x264_log(h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type,i_type_actual);
291 if( fread( rc->qp_buffer, sizeof(uint16_t), h->mb.i_mb_count, rc->p_mbtree_stat_file_in ) != h->mb.i_mb_count )
294 for( i = 0; i < h->mb.i_mb_count; i++ )
295 frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[i] )) * (1/256.0);
298 x264_adaptive_quant_frame( h, frame );
301 x264_log(h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n");
305 static char *x264_strcat_filename( char *input, char *suffix )
307 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
310 strcpy( output, input );
311 strcat( output, suffix );
315 int x264_ratecontrol_new( x264_t *h )
317 x264_ratecontrol_t *rc;
322 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
325 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
326 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
328 /* FIXME: use integers */
329 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
330 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
334 if( h->param.rc.b_mb_tree )
336 h->param.rc.f_pb_factor = 1;
340 rc->qcompress = h->param.rc.f_qcompress;
342 rc->bitrate = h->param.rc.i_bitrate * 1000.;
343 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
344 rc->nmb = h->mb.i_mb_count;
345 rc->last_non_b_pict_type = -1;
348 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
350 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
353 if( h->param.rc.i_vbv_buffer_size )
355 if( h->param.rc.i_rc_method == X264_RC_CQP )
357 x264_log(h, X264_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
358 h->param.rc.i_vbv_max_bitrate = 0;
359 h->param.rc.i_vbv_buffer_size = 0;
361 else if( h->param.rc.i_vbv_max_bitrate == 0 )
363 x264_log( h, X264_LOG_DEBUG, "VBV maxrate unspecified, assuming CBR\n" );
364 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
367 if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate &&
368 h->param.rc.i_vbv_max_bitrate > 0)
369 x264_log(h, X264_LOG_WARNING, "max bitrate less than average bitrate, ignored.\n");
370 else if( h->param.rc.i_vbv_max_bitrate > 0 &&
371 h->param.rc.i_vbv_buffer_size > 0 )
373 if( h->param.rc.i_vbv_buffer_size < 3 * h->param.rc.i_vbv_max_bitrate / rc->fps )
375 h->param.rc.i_vbv_buffer_size = 3 * h->param.rc.i_vbv_max_bitrate / rc->fps;
376 x264_log( h, X264_LOG_WARNING, "VBV buffer size too small, using %d kbit\n",
377 h->param.rc.i_vbv_buffer_size );
379 if( h->param.rc.f_vbv_buffer_init > 1. )
380 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 );
381 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000. / rc->fps;
382 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000.;
383 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
384 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
385 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
387 rc->b_vbv_min_rate = !rc->b_2pass
388 && h->param.rc.i_rc_method == X264_RC_ABR
389 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
391 else if( h->param.rc.i_vbv_max_bitrate )
393 x264_log(h, X264_LOG_WARNING, "VBV maxrate specified, but no bufsize.\n");
394 h->param.rc.i_vbv_max_bitrate = 0;
396 if(rc->rate_tolerance < 0.01)
398 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
399 rc->rate_tolerance = 0.01;
402 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
406 /* FIXME ABR_INIT_QP is actually used only in CRF */
407 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
408 rc->accum_p_norm = .01;
409 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
410 /* estimated ratio that produces a reasonable QP for the first I-frame */
411 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
412 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
413 rc->last_non_b_pict_type = SLICE_TYPE_I;
416 if( h->param.rc.i_rc_method == X264_RC_CRF )
418 /* Arbitrary rescaling to make CRF somewhat similar to QP.
419 * Try to compensate for MB-tree's effects as well. */
420 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
421 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
422 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
423 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset );
426 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
427 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
428 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
429 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
430 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
432 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
433 rc->last_qscale = qp2qscale(26);
434 CHECKED_MALLOC( rc->pred, 5*sizeof(predictor_t) );
435 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
436 for( i = 0; i < 5; i++ )
438 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
439 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
440 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
441 rc->pred[i].coeff= 2.0;
442 rc->pred[i].count= 1.0;
443 rc->pred[i].decay= 0.5;
444 rc->pred[i].offset= 0.0;
445 rc->row_preds[i].coeff= .25;
446 rc->row_preds[i].count= 1.0;
447 rc->row_preds[i].decay= 0.5;
448 rc->row_preds[i].offset= 0.0;
450 *rc->pred_b_from_p = rc->pred[0];
452 if( parse_zones( h ) < 0 )
454 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
458 /* Load stat file and init 2pass algo */
459 if( h->param.rc.b_stat_read )
461 char *p, *stats_in, *stats_buf;
463 /* read 1st pass stats */
464 assert( h->param.rc.psz_stat_in );
465 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
468 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
471 if( h->param.rc.b_mb_tree )
473 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
474 if( !mbtree_stats_in )
476 rc->p_mbtree_stat_file_in = fopen( mbtree_stats_in, "rb" );
477 x264_free( mbtree_stats_in );
478 if( !rc->p_mbtree_stat_file_in )
480 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
485 /* check whether 1st pass options were compatible with current options */
486 if( !strncmp( stats_buf, "#options:", 9 ) )
489 char *opts = stats_buf;
490 stats_in = strchr( stats_buf, '\n' );
496 if( ( p = strstr( opts, "bframes=" ) ) && sscanf( p, "bframes=%d", &i )
497 && h->param.i_bframe != i )
499 x264_log( h, X264_LOG_ERROR, "different number of B-frames than 1st pass (%d vs %d)\n",
500 h->param.i_bframe, i );
504 /* since B-adapt doesn't (yet) take into account B-pyramid,
505 * the converse is not a problem */
506 if( strstr( opts, "b_pyramid=1" ) && !h->param.b_bframe_pyramid )
507 x264_log( h, X264_LOG_WARNING, "1st pass used B-pyramid, 2nd doesn't\n" );
509 if( ( p = strstr( opts, "keyint=" ) ) && sscanf( p, "keyint=%d", &i )
510 && h->param.i_keyint_max != i )
511 x264_log( h, X264_LOG_WARNING, "different keyint than 1st pass (%d vs %d)\n",
512 h->param.i_keyint_max, i );
514 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
515 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
517 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
519 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
520 h->mb.b_direct_auto_write = 1;
523 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
524 h->param.i_bframe_adaptive = i;
525 else if( h->param.i_bframe )
527 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
531 if( h->param.rc.b_mb_tree && ( p = strstr( opts, "rc_lookahead=" ) ) && sscanf( p, "rc_lookahead=%d", &i ) )
532 h->param.rc.i_lookahead = i;
535 /* find number of pics */
538 p = strchr(p+1, ';');
541 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
546 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
548 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
549 h->param.i_frame_total, rc->num_entries );
551 if( h->param.i_frame_total > rc->num_entries )
553 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
554 h->param.i_frame_total, rc->num_entries );
558 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
560 /* init all to skipped p frames */
561 for(i=0; i<rc->num_entries; i++)
563 ratecontrol_entry_t *rce = &rc->entry[i];
564 rce->pict_type = SLICE_TYPE_P;
565 rce->qscale = rce->new_qscale = qp2qscale(20);
566 rce->misc_bits = rc->nmb + 10;
572 for(i=0; i < rc->num_entries; i++)
574 ratecontrol_entry_t *rce;
581 next= strchr(p, ';');
584 (*next)=0; //sscanf is unbelievably slow on long strings
587 e = sscanf(p, " in:%d ", &frame_number);
589 if(frame_number < 0 || frame_number >= rc->num_entries)
591 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
594 rce = &rc->entry[frame_number];
595 rce->direct_mode = 0;
597 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",
598 &pict_type, &qp, &rce->tex_bits,
599 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
600 &rce->s_count, &rce->direct_mode);
604 case 'I': rce->kept_as_ref = 1;
605 case 'i': rce->pict_type = SLICE_TYPE_I; break;
606 case 'P': rce->pict_type = SLICE_TYPE_P; break;
607 case 'B': rce->kept_as_ref = 1;
608 case 'b': rce->pict_type = SLICE_TYPE_B; break;
609 default: e = -1; break;
613 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
616 rce->qscale = qp2qscale(qp);
620 x264_free(stats_buf);
622 if(h->param.rc.i_rc_method == X264_RC_ABR)
624 if(init_pass2(h) < 0) return -1;
625 } /* else we're using constant quant, so no need to run the bitrate allocation */
628 /* Open output file */
629 /* If input and output files are the same, output to a temp file
630 * and move it to the real name only when it's complete */
631 if( h->param.rc.b_stat_write )
634 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
635 if( !rc->psz_stat_file_tmpname )
638 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
639 if( rc->p_stat_file_out == NULL )
641 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
645 p = x264_param2string( &h->param, 1 );
647 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
649 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
651 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
652 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
653 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
656 rc->p_mbtree_stat_file_out = fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
657 if( rc->p_mbtree_stat_file_out == NULL )
659 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
665 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
666 CHECKED_MALLOC( rc->qp_buffer, h->mb.i_mb_count * sizeof(uint16_t) );
668 for( i=0; i<h->param.i_threads; i++ )
670 h->thread[i]->rc = rc+i;
674 memcpy( &h->thread[i]->param, &h->param, sizeof(x264_param_t) );
675 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
684 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
687 char *tok, UNUSED *saveptr;
689 z->f_bitrate_factor = 1;
690 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
692 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
694 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
698 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
704 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
705 memcpy( z->param, &h->param, sizeof(x264_param_t) );
706 z->param->param_free = x264_free;
707 while( (tok = strtok_r( p, ",", &saveptr )) )
709 char *val = strchr( tok, '=' );
715 if( x264_param_parse( z->param, tok, val ) )
717 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
727 static int parse_zones( x264_t *h )
729 x264_ratecontrol_t *rc = h->rc;
731 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
733 char *psz_zones, *p, *tok, UNUSED *saveptr;
734 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
735 strcpy( psz_zones, h->param.rc.psz_zones );
736 h->param.rc.i_zones = 1;
737 for( p = psz_zones; *p; p++ )
738 h->param.rc.i_zones += (*p == '/');
739 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
741 for( i = 0; i < h->param.rc.i_zones; i++ )
743 tok = strtok_r( p, "/", &saveptr );
744 if( !tok || parse_zone( h, &h->param.rc.zones[i], tok ) )
748 x264_free( psz_zones );
751 if( h->param.rc.i_zones > 0 )
753 for( i = 0; i < h->param.rc.i_zones; i++ )
755 x264_zone_t z = h->param.rc.zones[i];
756 if( z.i_start < 0 || z.i_start > z.i_end )
758 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
759 z.i_start, z.i_end );
762 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
764 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
765 z.f_bitrate_factor );
770 rc->i_zones = h->param.rc.i_zones + 1;
771 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
772 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
774 // default zone to fall back to if none of the others match
775 rc->zones[0].i_start = 0;
776 rc->zones[0].i_end = INT_MAX;
777 rc->zones[0].b_force_qp = 0;
778 rc->zones[0].f_bitrate_factor = 1;
779 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
780 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
781 for( i = 1; i < rc->i_zones; i++ )
783 if( !rc->zones[i].param )
784 rc->zones[i].param = rc->zones[0].param;
793 static x264_zone_t *get_zone( x264_t *h, int frame_num )
796 for( i = h->rc->i_zones-1; i >= 0; i-- )
798 x264_zone_t *z = &h->rc->zones[i];
799 if( frame_num >= z->i_start && frame_num <= z->i_end )
805 void x264_ratecontrol_summary( x264_t *h )
807 x264_ratecontrol_t *rc = h->rc;
808 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
810 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
811 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*12.5 : 0;
812 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
813 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
814 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset );
818 void x264_ratecontrol_delete( x264_t *h )
820 x264_ratecontrol_t *rc = h->rc;
823 if( rc->p_stat_file_out )
825 fclose( rc->p_stat_file_out );
826 if( h->i_frame >= rc->num_entries )
827 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
829 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
830 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
832 x264_free( rc->psz_stat_file_tmpname );
834 if( rc->p_mbtree_stat_file_out )
836 fclose( rc->p_mbtree_stat_file_out );
837 if( h->i_frame >= rc->num_entries )
838 if( rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
840 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
841 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
843 x264_free( rc->psz_mbtree_stat_file_tmpname );
844 x264_free( rc->psz_mbtree_stat_file_name );
846 x264_free( rc->pred );
847 x264_free( rc->pred_b_from_p );
848 x264_free( rc->entry );
849 x264_free( rc->qp_buffer );
852 x264_free( rc->zones[0].param );
853 for( i=1; i<rc->i_zones; i++ )
854 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
855 rc->zones[i].param->param_free( rc->zones[i].param );
856 x264_free( rc->zones );
861 void x264_ratecontrol_set_estimated_size( x264_t *h, int bits )
863 x264_pthread_mutex_lock( &h->fenc->mutex );
864 h->rc->frame_size_estimated = bits;
865 x264_pthread_mutex_unlock( &h->fenc->mutex );
868 int x264_ratecontrol_get_estimated_size( x264_t const *h)
871 x264_pthread_mutex_lock( &h->fenc->mutex );
872 size = h->rc->frame_size_estimated;
873 x264_pthread_mutex_unlock( &h->fenc->mutex );
877 static void accum_p_qp_update( x264_t *h, float qp )
879 x264_ratecontrol_t *rc = h->rc;
880 rc->accum_p_qp *= .95;
881 rc->accum_p_norm *= .95;
882 rc->accum_p_norm += 1;
883 if( h->sh.i_type == SLICE_TYPE_I )
884 rc->accum_p_qp += qp + rc->ip_offset;
886 rc->accum_p_qp += qp;
889 /* Before encoding a frame, choose a QP for it */
890 void x264_ratecontrol_start( x264_t *h, int i_force_qp )
892 x264_ratecontrol_t *rc = h->rc;
893 ratecontrol_entry_t *rce = NULL;
894 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
899 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
900 x264_encoder_reconfig( h, zone->param );
901 rc->prev_zone = zone;
903 rc->qp_force = i_force_qp;
905 if( h->param.rc.b_stat_read )
907 int frame = h->fenc->i_frame;
908 assert( frame >= 0 && frame < rc->num_entries );
909 rce = h->rc->rce = &h->rc->entry[frame];
911 if( h->sh.i_type == SLICE_TYPE_B
912 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
914 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
915 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
921 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
922 rc->row_pred = &rc->row_preds[h->sh.i_type];
923 update_vbv_plan( h );
926 if( h->sh.i_type != SLICE_TYPE_B )
929 while( h->frames.current[rc->bframes] && IS_X264_TYPE_B(h->frames.current[rc->bframes]->i_type) )
939 q = qscale2qp( rate_estimate_qscale( h ) );
941 else if( rc->b_2pass )
943 rce->new_qscale = rate_estimate_qscale( h );
944 q = qscale2qp( rce->new_qscale );
948 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
949 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
951 q = rc->qp_constant[ h->sh.i_type ];
955 if( zone->b_force_qp )
956 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
958 q -= 6*log(zone->f_bitrate_factor)/log(2);
964 h->fdec->f_qp_avg_rc =
965 h->fdec->f_qp_avg_aq =
967 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
970 rce->new_qp = rc->qp;
972 accum_p_qp_update( h, rc->qp );
974 if( h->sh.i_type != SLICE_TYPE_B )
975 rc->last_non_b_pict_type = h->sh.i_type;
978 static double predict_row_size( x264_t *h, int y, int qp )
980 /* average between two predictors:
981 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
982 x264_ratecontrol_t *rc = h->rc;
983 double pred_s = predict_size( rc->row_pred, qp2qscale(qp), h->fdec->i_row_satd[y] );
985 if( h->sh.i_type != SLICE_TYPE_I
986 && h->fref0[0]->i_type == h->fdec->i_type
987 && h->fref0[0]->i_row_satd[y] > 0
988 && (abs(h->fref0[0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
990 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
991 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
996 return (pred_s + pred_t) / 2;
999 static double row_bits_so_far( x264_t *h, int y )
1003 for( i = 0; i <= y; i++ )
1004 bits += h->fdec->i_row_bits[i];
1008 static double predict_row_size_sum( x264_t *h, int y, int qp )
1011 double bits = row_bits_so_far(h, y);
1012 for( i = y+1; i < h->sps->i_mb_height; i++ )
1013 bits += predict_row_size( h, i, qp );
1018 void x264_ratecontrol_mb( x264_t *h, int bits )
1020 x264_ratecontrol_t *rc = h->rc;
1021 const int y = h->mb.i_mb_y;
1025 h->fdec->i_row_bits[y] += bits;
1026 rc->qpa_rc += rc->f_qpm;
1027 rc->qpa_aq += h->mb.i_qp;
1029 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !rc->b_vbv )
1032 h->fdec->i_row_qp[y] = rc->qpm;
1034 if( h->sh.i_type == SLICE_TYPE_B )
1036 /* B-frames shouldn't use lower QP than their reference frames.
1037 * This code is a bit overzealous in limiting B-frame quantizers, but it helps avoid
1038 * underflows due to the fact that B-frames are not explicitly covered by VBV. */
1039 if( y < h->sps->i_mb_height-1 )
1042 int avg_qp = X264_MIN(h->fref0[0]->i_row_qp[y+1], h->fref1[0]->i_row_qp[y+1])
1043 + rc->pb_offset * ((h->fenc->i_type == X264_TYPE_BREF) ? 0.5 : 1);
1044 rc->qpm = X264_MIN(X264_MAX( rc->qp, avg_qp), 51); //avg_qp could go higher than 51 due to pb_offset
1045 i_estimated = row_bits_so_far(h, y); //FIXME: compute full estimated size
1046 if (i_estimated > h->rc->frame_size_planned)
1047 x264_ratecontrol_set_estimated_size(h, i_estimated);
1052 update_predictor( rc->row_pred, qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1054 /* tweak quality based on difference from predicted size */
1055 if( y < h->sps->i_mb_height-1 && h->stat.i_slice_count[h->sh.i_type] > 0 )
1057 int prev_row_qp = h->fdec->i_row_qp[y];
1058 int b0 = predict_row_size_sum( h, y, rc->qpm );
1060 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
1061 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1062 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1063 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits.
1064 * In 2-pass mode we can be more trusting of the planned frame sizes, since they were decided
1065 * by actual encoding instead of SATD prediction. */
1066 float rc_tol = h->param.rc.b_stat_read ? (buffer_left_planned / rc->buffer_size) * rc->frame_size_planned
1067 : (buffer_left_planned / h->param.i_threads);
1069 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1070 /* area at the top of the frame was measured inaccurately. */
1071 if(row_bits_so_far(h,y) < 0.05 * rc->frame_size_planned)
1074 if(h->sh.i_type != SLICE_TYPE_I)
1077 if( !rc->b_vbv_min_rate )
1078 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
1080 while( rc->qpm < i_qp_max
1081 && ((b1 > rc->frame_size_planned + rc_tol) ||
1082 (rc->buffer_fill - b1 < buffer_left_planned * 0.5) ||
1083 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1086 b1 = predict_row_size_sum( h, y, rc->qpm );
1089 /* avoid VBV underflow */
1090 while( (rc->qpm < h->param.rc.i_qp_max)
1091 && (rc->buffer_fill - b1 < rc->buffer_size * 0.005))
1094 b1 = predict_row_size_sum( h, y, rc->qpm );
1097 while( rc->qpm > i_qp_min
1098 && rc->qpm > h->fdec->i_row_qp[0]
1099 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
1100 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
1103 b1 = predict_row_size_sum( h, y, rc->qpm );
1105 x264_ratecontrol_set_estimated_size(h, b1);
1108 /* loses the fractional part of the frame-wise qp */
1109 rc->f_qpm = rc->qpm;
1112 int x264_ratecontrol_qp( x264_t *h )
1117 /* In 2pass, force the same frame types as in the 1st pass */
1118 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1120 x264_ratecontrol_t *rc = h->rc;
1121 if( h->param.rc.b_stat_read )
1123 if( frame_num >= rc->num_entries )
1125 /* We could try to initialize everything required for ABR and
1126 * adaptive B-frames, but that would be complicated.
1127 * So just calculate the average QP used so far. */
1130 h->param.rc.i_qp_constant = (h->stat.i_slice_count[SLICE_TYPE_P] == 0) ? 24
1131 : 1 + h->stat.f_slice_qp[SLICE_TYPE_P] / h->stat.i_slice_count[SLICE_TYPE_P];
1132 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
1133 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 );
1134 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 );
1136 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1137 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1138 if( h->param.i_bframe_adaptive )
1139 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1141 for( i = 0; i < h->param.i_threads; i++ )
1143 h->thread[i]->rc->b_abr = 0;
1144 h->thread[i]->rc->b_2pass = 0;
1145 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1146 h->thread[i]->param.rc.b_stat_read = 0;
1147 h->thread[i]->param.i_bframe_adaptive = 0;
1148 h->thread[i]->param.i_scenecut_threshold = 0;
1149 if( h->thread[i]->param.i_bframe > 1 )
1150 h->thread[i]->param.i_bframe = 1;
1152 return X264_TYPE_AUTO;
1154 switch( rc->entry[frame_num].pict_type )
1157 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I;
1160 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B;
1169 return X264_TYPE_AUTO;
1173 /* After encoding one frame, save stats and update ratecontrol state */
1174 int x264_ratecontrol_end( x264_t *h, int bits )
1176 x264_ratecontrol_t *rc = h->rc;
1177 const int *mbs = h->stat.frame.i_mb_count;
1182 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1183 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1184 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1185 for( i = B_DIRECT; i < B_8x8; i++ )
1186 h->stat.frame.i_mb_count_p += mbs[i];
1188 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1189 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1191 if( h->param.rc.b_stat_write )
1193 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1194 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1195 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1196 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1197 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1198 char c_direct = h->mb.b_direct_auto_write ?
1199 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1200 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1202 if( fprintf( rc->p_stat_file_out,
1203 "in:%d out:%d type:%c q:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c;\n",
1204 h->fenc->i_frame, h->i_frame,
1206 h->stat.frame.i_tex_bits,
1207 h->stat.frame.i_mv_bits,
1208 h->stat.frame.i_misc_bits,
1209 h->stat.frame.i_mb_count_i,
1210 h->stat.frame.i_mb_count_p,
1211 h->stat.frame.i_mb_count_skip,
1215 /* Don't re-write the data in multi-pass mode. */
1216 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1218 uint8_t i_type = h->sh.i_type;
1220 /* Values are stored as big-endian FIX8.8 */
1221 for( i = 0; i < h->mb.i_mb_count; i++ )
1222 rc->qp_buffer[i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1223 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1225 if( fwrite( rc->qp_buffer, sizeof(uint16_t), h->mb.i_mb_count, rc->p_mbtree_stat_file_out ) < h->mb.i_mb_count )
1232 if( h->sh.i_type != SLICE_TYPE_B )
1233 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / rc->last_rceq;
1236 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1237 * Not perfectly accurate with B-refs, but good enough. */
1238 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
1240 rc->cplxr_sum *= rc->cbr_decay;
1241 rc->wanted_bits_window += rc->bitrate / rc->fps;
1242 rc->wanted_bits_window *= rc->cbr_decay;
1247 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
1250 if( h->mb.b_variable_qp )
1252 if( h->sh.i_type == SLICE_TYPE_B )
1254 rc->bframe_bits += bits;
1255 if( !h->frames.current[0] || !IS_X264_TYPE_B(h->frames.current[0]->i_type) )
1257 update_predictor( rc->pred_b_from_p, qp2qscale(rc->qpa_rc),
1258 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1259 rc->bframe_bits = 0;
1264 update_vbv( h, bits );
1267 x264_log(h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n");
1271 /****************************************************************************
1273 ***************************************************************************/
1276 * modify the bitrate curve from pass1 for one frame
1278 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1280 x264_ratecontrol_t *rcc= h->rc;
1282 x264_zone_t *zone = get_zone( h, frame_num );
1284 q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1286 // avoid NaN's in the rc_eq
1287 if(!isfinite(q) || rce->tex_bits + rce->mv_bits == 0)
1288 q = rcc->last_qscale;
1293 rcc->last_qscale = q;
1298 if( zone->b_force_qp )
1299 q = qp2qscale(zone->i_qp);
1301 q /= zone->f_bitrate_factor;
1307 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1309 x264_ratecontrol_t *rcc = h->rc;
1310 const int pict_type = rce->pict_type;
1312 // force I/B quants as a function of P quants
1313 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1314 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1315 if( pict_type == SLICE_TYPE_I )
1318 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1319 double ip_factor = fabs( h->param.rc.f_ip_factor );
1320 /* don't apply ip_factor if the following frame is also I */
1321 if( rcc->accum_p_norm <= 0 )
1323 else if( h->param.rc.f_ip_factor < 0 )
1325 else if( rcc->accum_p_norm >= 1 )
1328 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1330 else if( pict_type == SLICE_TYPE_B )
1332 if( h->param.rc.f_pb_factor > 0 )
1334 if( !rce->kept_as_ref )
1335 q *= fabs( h->param.rc.f_pb_factor );
1337 else if( pict_type == SLICE_TYPE_P
1338 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1339 && rce->tex_bits == 0 )
1344 /* last qscale / qdiff stuff */
1345 if(rcc->last_non_b_pict_type==pict_type
1346 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1348 double last_q = rcc->last_qscale_for[pict_type];
1349 double max_qscale = last_q * rcc->lstep;
1350 double min_qscale = last_q / rcc->lstep;
1352 if (q > max_qscale) q = max_qscale;
1353 else if(q < min_qscale) q = min_qscale;
1356 rcc->last_qscale_for[pict_type] = q;
1357 if(pict_type!=SLICE_TYPE_B)
1358 rcc->last_non_b_pict_type = pict_type;
1359 if(pict_type==SLICE_TYPE_I)
1361 rcc->last_accum_p_norm = rcc->accum_p_norm;
1362 rcc->accum_p_norm = 0;
1363 rcc->accum_p_qp = 0;
1365 if(pict_type==SLICE_TYPE_P)
1367 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1368 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1369 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1374 static double predict_size( predictor_t *p, double q, double var )
1376 return (p->coeff*var + p->offset) / (q*p->count);
1379 static void update_predictor( predictor_t *p, double q, double var, double bits )
1381 const double range = 1.5;
1384 double old_coeff = p->coeff / p->count;
1385 double new_coeff = bits*q / var;
1386 double new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
1387 double new_offset = bits*q - new_coeff_clipped * var;
1388 if( new_offset >= 0 )
1389 new_coeff = new_coeff_clipped;
1392 p->count *= p->decay;
1393 p->coeff *= p->decay;
1394 p->offset *= p->decay;
1396 p->coeff += new_coeff;
1397 p->offset += new_offset;
1400 // update VBV after encoding a frame
1401 static void update_vbv( x264_t *h, int bits )
1403 x264_ratecontrol_t *rcc = h->rc;
1404 x264_ratecontrol_t *rct = h->thread[0]->rc;
1406 if( rcc->last_satd >= h->mb.i_mb_count )
1407 update_predictor( &rct->pred[h->sh.i_type], qp2qscale(rcc->qpa_rc), rcc->last_satd, bits );
1412 rct->buffer_fill_final += rct->buffer_rate - bits;
1413 if( rct->buffer_fill_final < 0 )
1414 x264_log( h, X264_LOG_WARNING, "VBV underflow (%.0f bits)\n", rct->buffer_fill_final );
1415 rct->buffer_fill_final = x264_clip3f( rct->buffer_fill_final, 0, rct->buffer_size );
1418 // provisionally update VBV according to the planned size of all frames currently in progress
1419 static void update_vbv_plan( x264_t *h )
1421 x264_ratecontrol_t *rcc = h->rc;
1422 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final;
1423 if( h->param.i_threads > 1 )
1425 int j = h->rc - h->thread[0]->rc;
1427 for( i=1; i<h->param.i_threads; i++ )
1429 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1430 double bits = t->rc->frame_size_planned;
1431 if( !t->b_thread_active )
1433 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1434 rcc->buffer_fill += rcc->buffer_rate - bits;
1435 rcc->buffer_fill = x264_clip3( rcc->buffer_fill, 0, rcc->buffer_size );
1440 // apply VBV constraints and clip qscale to between lmin and lmax
1441 static double clip_qscale( x264_t *h, int pict_type, double q )
1443 x264_ratecontrol_t *rcc = h->rc;
1444 double lmin = rcc->lmin[pict_type];
1445 double lmax = rcc->lmax[pict_type];
1448 /* B-frames are not directly subject to VBV,
1449 * since they are controlled by the P-frames' QPs. */
1451 if( rcc->b_vbv && rcc->last_satd > 0 )
1453 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
1454 * the lookahead overflow and such that the buffer is in a reasonable state
1455 * by the end of the lookahead. */
1456 if( h->param.rc.i_lookahead )
1458 int j, iterations, terminate = 0;
1460 /* Avoid an infinite loop. */
1461 for( iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
1464 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1465 double buffer_fill_cur = rcc->buffer_fill - cur_bits + rcc->buffer_rate;
1467 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
1468 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
1469 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
1471 /* Loop over the planned future frames. */
1472 for( j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
1474 int i_type = h->fenc->i_planned_type[j];
1475 int i_satd = h->fenc->i_planned_satd[j];
1476 if( i_type == X264_TYPE_AUTO )
1478 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
1479 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
1480 buffer_fill_cur = buffer_fill_cur - cur_bits + rcc->buffer_rate;
1482 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
1483 target_fill = X264_MIN( rcc->buffer_fill + j * rcc->buffer_rate * 0.5, rcc->buffer_size * 0.5 );
1484 if( buffer_fill_cur < target_fill )
1490 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
1491 target_fill = x264_clip3f( rcc->buffer_fill - j * rcc->buffer_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
1492 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
1501 /* Fallback to old purely-reactive algorithm: no lookahead. */
1504 if( ( pict_type == SLICE_TYPE_P ||
1505 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
1506 rcc->buffer_fill/rcc->buffer_size < 0.5 )
1508 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1511 /* Now a hard threshold to make sure the frame fits in VBV.
1512 * This one is mostly for I-frames. */
1513 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1515 if( bits > rcc->buffer_fill/2 )
1516 qf = x264_clip3f( rcc->buffer_fill/(2*bits), 0.2, 1.0 );
1519 if( bits < rcc->buffer_rate/2 )
1520 q *= bits*2/rcc->buffer_rate;
1521 q = X264_MAX( q0, q );
1524 /* Check B-frame complexity, and use up any bits that would
1525 * overflow before the next P-frame. */
1526 if( h->sh.i_type == SLICE_TYPE_P )
1528 int nb = rcc->bframes;
1529 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1530 double pbbits = bits;
1531 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1534 if( bbits > rcc->buffer_rate )
1536 pbbits += nb * bbits;
1538 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1539 if( pbbits < space )
1541 q *= X264_MAX( pbbits / space,
1542 bits / (0.5 * rcc->buffer_size) );
1544 q = X264_MAX( q0-5, q );
1547 if( !rcc->b_vbv_min_rate )
1548 q = X264_MAX( q0, q );
1553 else if(rcc->b_2pass)
1555 double min2 = log(lmin);
1556 double max2 = log(lmax);
1557 q = (log(q) - min2)/(max2-min2) - 0.5;
1558 q = 1.0/(1.0 + exp(-4*q));
1559 q = q*(max2-min2) + min2;
1563 return x264_clip3f(q, lmin, lmax);
1566 // update qscale for 1 frame based on actual bits used so far
1567 static float rate_estimate_qscale( x264_t *h )
1570 x264_ratecontrol_t *rcc = h->rc;
1571 ratecontrol_entry_t rce;
1572 int pict_type = h->sh.i_type;
1573 double lmin = rcc->lmin[pict_type];
1574 double lmax = rcc->lmax[pict_type];
1575 int64_t total_bits = 8*(h->stat.i_slice_size[SLICE_TYPE_I]
1576 + h->stat.i_slice_size[SLICE_TYPE_P]
1577 + h->stat.i_slice_size[SLICE_TYPE_B]);
1582 if(pict_type != rce.pict_type)
1584 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1585 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1589 if( pict_type == SLICE_TYPE_B )
1591 /* B-frames don't have independent ratecontrol, but rather get the
1592 * average QP of the two adjacent P-frames + an offset */
1594 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1595 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1596 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1597 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1598 float q0 = h->fref0[0]->f_qp_avg_rc;
1599 float q1 = h->fref1[0]->f_qp_avg_rc;
1601 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1602 q0 -= rcc->pb_offset/2;
1603 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1604 q1 -= rcc->pb_offset/2;
1607 q = (q0 + q1) / 2 + rcc->ip_offset;
1613 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1615 if(h->fenc->b_kept_as_ref)
1616 q += rcc->pb_offset/2;
1618 q += rcc->pb_offset;
1620 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1621 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1623 return qp2qscale(q);
1627 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1631 //FIXME adjust abr_buffer based on distance to the end of the video
1633 int64_t predicted_bits = total_bits;
1637 if( h->param.i_threads > 1 )
1639 int j = h->rc - h->thread[0]->rc;
1641 for( i=1; i<h->param.i_threads; i++ )
1643 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1644 double bits = t->rc->frame_size_planned;
1645 if( !t->b_thread_active )
1647 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1648 predicted_bits += (int64_t)bits;
1654 if( h->fenc->i_frame < h->param.i_threads )
1655 predicted_bits += (int64_t)h->fenc->i_frame * rcc->bitrate / rcc->fps;
1657 predicted_bits += (int64_t)(h->param.i_threads - 1) * rcc->bitrate / rcc->fps;
1660 diff = predicted_bits - (int64_t)rce.expected_bits;
1662 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1663 if( ((h->fenc->i_frame + 1 - h->param.i_threads) >= rcc->fps) &&
1664 (rcc->expected_bits_sum > 0))
1666 /* Adjust quant based on the difference between
1667 * achieved and expected bitrate so far */
1668 double time = (double)h->fenc->i_frame / rcc->num_entries;
1669 double w = x264_clip3f( time*100, 0.0, 1.0 );
1670 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1674 /* Do not overflow vbv */
1675 double expected_size = qscale2bits(&rce, q);
1676 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1677 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
1678 double qmax = q*(2 - expected_fullness);
1679 double size_constraint = 1 + expected_fullness;
1680 qmax = X264_MAX(qmax, rce.new_qscale);
1681 if (expected_fullness < .05)
1683 qmax = X264_MIN(qmax, lmax);
1684 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
1685 ((expected_vbv < 0) && (q < lmax)))
1688 expected_size = qscale2bits(&rce, q);
1689 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1691 rcc->last_satd = x264_stack_align( x264_rc_analyse_slice, h );
1693 q = x264_clip3f( q, lmin, lmax );
1695 else /* 1pass ABR */
1697 /* Calculate the quantizer which would have produced the desired
1698 * average bitrate if it had been applied to all frames so far.
1699 * Then modulate that quant based on the current frame's complexity
1700 * relative to the average complexity so far (using the 2pass RCEQ).
1701 * Then bias the quant up or down if total size so far was far from
1703 * Result: Depending on the value of rate_tolerance, there is a
1704 * tradeoff between quality and bitrate precision. But at large
1705 * tolerances, the bit distribution approaches that of 2pass. */
1707 double wanted_bits, overflow=1, lmin, lmax;
1709 rcc->last_satd = x264_stack_align( x264_rc_analyse_slice, h );
1710 rcc->short_term_cplxsum *= 0.5;
1711 rcc->short_term_cplxcount *= 0.5;
1712 rcc->short_term_cplxsum += rcc->last_satd;
1713 rcc->short_term_cplxcount ++;
1715 rce.tex_bits = rcc->last_satd;
1716 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1718 rce.p_count = rcc->nmb;
1722 rce.pict_type = pict_type;
1724 if( h->param.rc.i_rc_method == X264_RC_CRF )
1726 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1730 int i_frame_done = h->fenc->i_frame + 1 - h->param.i_threads;
1732 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1734 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
1735 wanted_bits = i_frame_done * rcc->bitrate / rcc->fps;
1736 if( wanted_bits > 0 )
1738 abr_buffer *= X264_MAX( 1, sqrt(i_frame_done/25) );
1739 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1744 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1745 /* should test _next_ pict type, but that isn't decided yet */
1746 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1748 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1749 q /= fabs( h->param.rc.f_ip_factor );
1751 else if( h->i_frame > 0 )
1753 /* Asymmetric clipping, because symmetric would prevent
1754 * overflow control in areas of rapidly oscillating complexity */
1755 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1756 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1757 if( overflow > 1.1 && h->i_frame > 3 )
1759 else if( overflow < 0.9 )
1762 q = x264_clip3f(q, lmin, lmax);
1764 else if( h->param.rc.i_rc_method == X264_RC_CRF )
1766 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1768 rcc->qp_novbv = qscale2qp(q);
1770 //FIXME use get_diff_limited_q() ?
1771 q = clip_qscale( h, pict_type, q );
1774 rcc->last_qscale_for[pict_type] =
1775 rcc->last_qscale = q;
1777 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
1778 rcc->last_qscale_for[SLICE_TYPE_P] = q;
1780 if( rcc->b_2pass && rcc->b_vbv )
1781 rcc->frame_size_planned = qscale2bits(&rce, q);
1783 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1784 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1789 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
1793 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
1794 /* these vars are updated in x264_ratecontrol_start()
1795 * so copy them from the context that most recently started (prev)
1796 * to the context that's about to start (cur).
1802 COPY(last_qscale_for);
1803 COPY(last_non_b_pict_type);
1804 COPY(short_term_cplxsum);
1805 COPY(short_term_cplxcount);
1812 #define COPY(var) next->rc->var = cur->rc->var
1813 /* these vars are updated in x264_ratecontrol_end()
1814 * so copy them from the context that most recently ended (cur)
1815 * to the context that's about to end (next)
1818 COPY(expected_bits_sum);
1819 COPY(wanted_bits_window);
1823 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
1824 /* the rest of the variables are either constant or thread-local */
1827 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
1829 /* find an interval ending on an overflow or underflow (depending on whether
1830 * we're adding or removing bits), and starting on the earliest frame that
1831 * can influence the buffer fill of that end frame. */
1832 x264_ratecontrol_t *rcc = h->rc;
1833 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
1834 const double buffer_max = .9 * rcc->buffer_size;
1835 double fill = fills[*t0-1];
1836 double parity = over ? 1. : -1.;
1837 int i, start=-1, end=-1;
1838 for(i = *t0; i < rcc->num_entries; i++)
1840 fill += (rcc->buffer_rate - qscale2bits(&rcc->entry[i], rcc->entry[i].new_qscale)) * parity;
1841 fill = x264_clip3f(fill, 0, rcc->buffer_size);
1843 if(fill <= buffer_min || i == 0)
1849 else if(fill >= buffer_max && start >= 0)
1854 return start>=0 && end>=0;
1857 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
1859 x264_ratecontrol_t *rcc = h->rc;
1860 double qscale_orig, qscale_new;
1865 for(i = t0; i <= t1; i++)
1867 qscale_orig = rcc->entry[i].new_qscale;
1868 qscale_orig = x264_clip3f(qscale_orig, qscale_min, qscale_max);
1869 qscale_new = qscale_orig * adjustment;
1870 qscale_new = x264_clip3f(qscale_new, qscale_min, qscale_max);
1871 rcc->entry[i].new_qscale = qscale_new;
1872 adjusted = adjusted || (qscale_new != qscale_orig);
1877 static double count_expected_bits( x264_t *h )
1879 x264_ratecontrol_t *rcc = h->rc;
1880 double expected_bits = 0;
1882 for(i = 0; i < rcc->num_entries; i++)
1884 ratecontrol_entry_t *rce = &rcc->entry[i];
1885 rce->expected_bits = expected_bits;
1886 expected_bits += qscale2bits(rce, rce->new_qscale);
1888 return expected_bits;
1891 static int vbv_pass2( x264_t *h )
1893 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
1894 * frames in the interval until either buffer is full at some intermediate frame or the
1895 * last frame in the interval no longer underflows. Recompute intervals and repeat.
1896 * Then do the converse to put bits back into overflow areas until target size is met */
1898 x264_ratecontrol_t *rcc = h->rc;
1900 double all_available_bits = h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps;
1901 double expected_bits = 0;
1903 double prev_bits = 0;
1905 double qscale_min = qp2qscale(h->param.rc.i_qp_min);
1906 double qscale_max = qp2qscale(h->param.rc.i_qp_max);
1908 int adj_min, adj_max;
1909 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
1913 /* adjust overall stream size */
1917 prev_bits = expected_bits;
1919 if(expected_bits != 0)
1920 { /* not first iteration */
1921 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
1922 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
1926 while(adj_min && find_underflow(h, fills, &t0, &t1, 1))
1928 adj_min = fix_underflow(h, t0, t1, adjustment, qscale_min, qscale_max);
1933 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
1935 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
1937 while(adj_max && find_underflow(h, fills, &t0, &t1, 0))
1938 adj_max = fix_underflow(h, t0, t1, 1.001, qscale_min, qscale_max);
1940 expected_bits = count_expected_bits(h);
1941 } while((expected_bits < .995*all_available_bits) && ((int)(expected_bits+.5) > (int)(prev_bits+.5)) );
1944 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
1946 /* store expected vbv filling values for tracking when encoding */
1947 for(i = 0; i < rcc->num_entries; i++)
1948 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
1956 static int init_pass2( x264_t *h )
1958 x264_ratecontrol_t *rcc = h->rc;
1959 uint64_t all_const_bits = 0;
1960 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps);
1961 double rate_factor, step, step_mult;
1962 double qblur = h->param.rc.f_qblur;
1963 double cplxblur = h->param.rc.f_complexity_blur;
1964 const int filter_size = (int)(qblur*4) | 1;
1965 double expected_bits;
1966 double *qscale, *blurred_qscale;
1969 /* find total/average complexity & const_bits */
1970 for(i=0; i<rcc->num_entries; i++)
1972 ratecontrol_entry_t *rce = &rcc->entry[i];
1973 all_const_bits += rce->misc_bits;
1976 if( all_available_bits < all_const_bits)
1978 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
1979 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)));
1983 /* Blur complexities, to reduce local fluctuation of QP.
1984 * We don't blur the QPs directly, because then one very simple frame
1985 * could drag down the QP of a nearby complex frame and give it more
1986 * bits than intended. */
1987 for(i=0; i<rcc->num_entries; i++)
1989 ratecontrol_entry_t *rce = &rcc->entry[i];
1990 double weight_sum = 0;
1991 double cplx_sum = 0;
1992 double weight = 1.0;
1993 double gaussian_weight;
1995 /* weighted average of cplx of future frames */
1996 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++)
1998 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
1999 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2002 gaussian_weight = weight * exp(-j*j/200.0);
2003 weight_sum += gaussian_weight;
2004 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2006 /* weighted average of cplx of past frames */
2008 for(j=0; j<=cplxblur*2 && j<=i; j++)
2010 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2011 gaussian_weight = weight * exp(-j*j/200.0);
2012 weight_sum += gaussian_weight;
2013 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2014 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2018 rce->blurred_complexity = cplx_sum / weight_sum;
2021 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2022 if( filter_size > 1 )
2023 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2025 blurred_qscale = qscale;
2027 /* Search for a factor which, when multiplied by the RCEQ values from
2028 * each frame, adds up to the desired total size.
2029 * There is no exact closed-form solution because of VBV constraints and
2030 * because qscale2bits is not invertible, but we can start with the simple
2031 * approximation of scaling the 1st pass by the ratio of bitrates.
2032 * The search range is probably overkill, but speed doesn't matter here. */
2035 for(i=0; i<rcc->num_entries; i++)
2036 expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i));
2037 step_mult = all_available_bits / expected_bits;
2040 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2043 rate_factor += step;
2045 rcc->last_non_b_pict_type = -1;
2046 rcc->last_accum_p_norm = 1;
2047 rcc->accum_p_norm = 0;
2050 for(i=0; i<rcc->num_entries; i++)
2052 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
2055 /* fixed I/B qscale relative to P */
2056 for(i=rcc->num_entries-1; i>=0; i--)
2058 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
2059 assert(qscale[i] >= 0);
2065 assert(filter_size%2==1);
2066 for(i=0; i<rcc->num_entries; i++)
2068 ratecontrol_entry_t *rce = &rcc->entry[i];
2070 double q=0.0, sum=0.0;
2072 for(j=0; j<filter_size; j++)
2074 int index = i+j-filter_size/2;
2076 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
2077 if(index < 0 || index >= rcc->num_entries)
2079 if(rce->pict_type != rcc->entry[index].pict_type)
2081 q += qscale[index] * coeff;
2084 blurred_qscale[i] = q/sum;
2088 /* find expected bits */
2089 for(i=0; i<rcc->num_entries; i++)
2091 ratecontrol_entry_t *rce = &rcc->entry[i];
2092 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
2093 assert(rce->new_qscale >= 0);
2094 expected_bits += qscale2bits(rce, rce->new_qscale);
2097 if(expected_bits > all_available_bits) rate_factor -= step;
2102 x264_free(blurred_qscale);
2105 if( vbv_pass2( h ) )
2107 expected_bits = count_expected_bits(h);
2109 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
2112 for(i=0; i<rcc->num_entries; i++)
2113 avgq += rcc->entry[i].new_qscale;
2114 avgq = qscale2qp(avgq / rcc->num_entries);
2116 if ((expected_bits > all_available_bits) || (!rcc->b_vbv))
2117 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
2118 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
2119 (float)h->param.rc.i_bitrate,
2120 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
2122 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
2124 if(h->param.rc.i_qp_min > 0)
2125 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
2127 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
2129 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
2131 if(h->param.rc.i_qp_max < 51)
2132 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
2134 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
2136 else if(!(rcc->b_2pass && rcc->b_vbv))
2137 x264_log(h, X264_LOG_WARNING, "internal error\n");