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;
52 float blurred_complexity;
54 } ratecontrol_entry_t;
63 struct x264_ratecontrol_t
72 double rate_tolerance;
73 int nmb; /* number of macroblocks in a frame */
77 ratecontrol_entry_t *rce;
78 int qp; /* qp for current frame */
79 int qpm; /* qp for current macroblock */
80 float f_qpm; /* qp for current macroblock: precise float for AQ */
81 float qpa_rc; /* average of macroblocks' qp before aq */
82 float qpa_aq; /* average of macroblocks' qp after aq */
87 double buffer_fill_final; /* real buffer as of the last finished frame */
88 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
89 double buffer_rate; /* # of bits added to buffer_fill after each frame */
90 predictor_t *pred; /* predict frame size from satd */
95 double cplxr_sum; /* sum of bits*qscale/rceq */
96 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow */
97 double wanted_bits_window; /* target bitrate * window */
99 double short_term_cplxsum;
100 double short_term_cplxcount;
101 double rate_factor_constant;
106 FILE *p_stat_file_out;
107 char *psz_stat_file_tmpname;
109 int num_entries; /* number of ratecontrol_entry_ts */
110 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
112 double last_qscale_for[5]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
113 int last_non_b_pict_type;
114 double accum_p_qp; /* for determining I-frame quant */
116 double last_accum_p_norm;
117 double lmin[5]; /* min qscale by frame type */
119 double lstep; /* max change (multiply) in qscale per frame */
122 double frame_size_estimated;
123 double frame_size_planned;
124 predictor_t *row_pred;
125 predictor_t row_preds[5];
126 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
127 int bframes; /* # consecutive B-frames before this P-frame */
128 int bframe_bits; /* total cost of those frames */
136 x264_zone_t *prev_zone;
140 static int parse_zones( x264_t *h );
141 static int init_pass2(x264_t *);
142 static float rate_estimate_qscale( x264_t *h );
143 static void update_vbv( x264_t *h, int bits );
144 static void update_vbv_plan( x264_t *h );
145 static double predict_size( predictor_t *p, double q, double var );
146 static void update_predictor( predictor_t *p, double q, double var, double bits );
149 * qp = h.264's quantizer
150 * qscale = linearized quantizer = Lagrange multiplier
152 static inline double qp2qscale(double qp)
154 return 0.85 * pow(2.0, ( qp - 12.0 ) / 6.0);
156 static inline double qscale2qp(double qscale)
158 return 12.0 + 6.0 * log(qscale/0.85) / log(2.0);
161 /* Texture bitrate is not quite inversely proportional to qscale,
162 * probably due the the changing number of SKIP blocks.
163 * MV bits level off at about qp<=12, because the lambda used
164 * for motion estimation is constant there. */
165 static inline double qscale2bits(ratecontrol_entry_t *rce, double qscale)
169 return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
170 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
174 // Find the total AC energy of the block in all planes.
175 static NOINLINE int ac_energy_mb( x264_t *h, int mb_x, int mb_y, int *satd )
177 /* This function contains annoying hacks because GCC has a habit of reordering emms
178 * and putting it after floating point ops. As a result, we put the emms at the end of the
179 * function and make sure that its always called before the float math. Noinline makes
180 * sure no reordering goes on. */
181 /* FIXME: This array is larger than necessary because a bug in GCC causes an all-zero
182 * array to be placed in .bss despite .bss not being correctly aligned on some platforms (win32?) */
183 DECLARE_ALIGNED_16( static uint8_t zero[17] ) = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1};
184 unsigned int var=0, sad, i;
185 if( satd || h->param.rc.i_aq_mode == X264_AQ_GLOBAL )
190 int stride = h->fenc->i_stride[i];
191 int offset = h->mb.b_interlaced
192 ? w * (mb_x + (mb_y&~1) * stride) + (mb_y&1) * stride
193 : w * (mb_x + mb_y * stride);
194 int pix = i ? PIXEL_8x8 : PIXEL_16x16;
195 stride <<= h->mb.b_interlaced;
196 var += h->pixf.var[pix]( h->fenc->plane[i]+offset, stride, &sad );
197 // SATD to represent the block's overall complexity (bit cost) for intra encoding.
198 // exclude the DC coef, because nothing short of an actual intra prediction will estimate DC cost.
200 *satd += h->pixf.satd[pix]( zero, 0, h->fenc->plane[i]+offset, stride ) - sad/2;
202 var = X264_MAX(var,1);
204 else var = h->rc->ac_energy[h->mb.i_mb_xy];
209 static void x264_autosense_aq( x264_t *h )
214 // FIXME: Some of the SATDs might be already calculated elsewhere (ratecontrol?). Can we reuse them?
215 // FIXME: Is chroma SATD necessary?
216 for( mb_y=0; mb_y<h->sps->i_mb_height; mb_y++ )
217 for( mb_x=0; mb_x<h->sps->i_mb_width; mb_x++ )
220 int energy = ac_energy_mb( h, mb_x, mb_y, &satd );
221 h->rc->ac_energy[mb_x + mb_y * h->sps->i_mb_width] = energy;
222 /* Weight the energy value by the SATD value of the MB.
223 * This represents the fact that the more complex blocks in a frame should
224 * be weighted more when calculating the optimal threshold. This also helps
225 * diminish the negative effect of large numbers of simple blocks in a frame,
226 * such as in the case of a letterboxed film. */
227 total += logf(energy) * satd;
231 /* Calculate and store the threshold. */
232 h->rc->aq_threshold = n ? total/n : 15;
235 /*****************************************************************************
236 * x264_adaptive_quant:
237 * adjust macroblock QP based on variance (AC energy) of the MB.
238 * high variance = higher QP
239 * low variance = lower QP
240 * This generally increases SSIM and lowers PSNR.
241 *****************************************************************************/
242 void x264_adaptive_quant( x264_t *h )
244 int energy = ac_energy_mb( h, h->mb.i_mb_x, h->mb.i_mb_y, NULL );
245 /* Adjust the QP based on the AC energy of the macroblock. */
246 float qp = h->rc->f_qpm;
247 float qp_adj = 1.5 * (logf(energy) - h->rc->aq_threshold);
248 if( h->param.rc.i_aq_mode == X264_AQ_LOCAL )
249 qp_adj = x264_clip3f( qp_adj, -5, 5 );
250 h->mb.i_qp = x264_clip3( qp + qp_adj * h->param.rc.f_aq_strength + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
251 /* If the QP of this MB is within 1 of the previous MB, code the same QP as the previous MB,
252 * to lower the bit cost of the qp_delta. */
253 if( abs(h->mb.i_qp - h->mb.i_last_qp) == 1 )
254 h->mb.i_qp = h->mb.i_last_qp;
255 h->mb.i_chroma_qp = h->chroma_qp_table[h->mb.i_qp];
258 int x264_ratecontrol_new( x264_t *h )
260 x264_ratecontrol_t *rc;
265 rc = h->rc = x264_malloc( h->param.i_threads * sizeof(x264_ratecontrol_t) );
266 memset( rc, 0, h->param.i_threads * sizeof(x264_ratecontrol_t) );
268 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
269 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
271 /* FIXME: use integers */
272 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
273 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
277 rc->bitrate = h->param.rc.i_bitrate * 1000.;
278 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
279 rc->nmb = h->mb.i_mb_count;
280 rc->last_non_b_pict_type = -1;
283 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
285 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
288 if( h->param.rc.i_vbv_buffer_size )
290 if( h->param.rc.i_rc_method == X264_RC_CQP )
291 x264_log(h, X264_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
292 else if( h->param.rc.i_vbv_max_bitrate == 0 )
294 x264_log( h, X264_LOG_DEBUG, "VBV maxrate unspecified, assuming CBR\n" );
295 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
298 if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate &&
299 h->param.rc.i_vbv_max_bitrate > 0)
300 x264_log(h, X264_LOG_WARNING, "max bitrate less than average bitrate, ignored.\n");
301 else if( h->param.rc.i_vbv_max_bitrate > 0 &&
302 h->param.rc.i_vbv_buffer_size > 0 )
304 if( h->param.rc.i_vbv_buffer_size < 3 * h->param.rc.i_vbv_max_bitrate / rc->fps )
306 h->param.rc.i_vbv_buffer_size = 3 * h->param.rc.i_vbv_max_bitrate / rc->fps;
307 x264_log( h, X264_LOG_WARNING, "VBV buffer size too small, using %d kbit\n",
308 h->param.rc.i_vbv_buffer_size );
310 if( h->param.rc.f_vbv_buffer_init > 1. )
311 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 );
312 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000. / rc->fps;
313 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000.;
314 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
315 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
316 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
318 rc->b_vbv_min_rate = !rc->b_2pass
319 && h->param.rc.i_rc_method == X264_RC_ABR
320 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
322 else if( h->param.rc.i_vbv_max_bitrate )
324 x264_log(h, X264_LOG_WARNING, "VBV maxrate specified, but no bufsize.\n");
325 h->param.rc.i_vbv_max_bitrate = 0;
327 if(rc->rate_tolerance < 0.01)
329 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
330 rc->rate_tolerance = 0.01;
333 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
337 /* FIXME ABR_INIT_QP is actually used only in CRF */
338 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
339 rc->accum_p_norm = .01;
340 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
341 /* estimated ratio that produces a reasonable QP for the first I-frame */
342 rc->cplxr_sum = .01 * pow( 7.0e5, h->param.rc.f_qcompress ) * pow( h->mb.i_mb_count, 0.5 );
343 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
344 rc->last_non_b_pict_type = SLICE_TYPE_I;
347 if( h->param.rc.i_rc_method == X264_RC_CRF )
349 /* arbitrary rescaling to make CRF somewhat similar to QP */
350 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
351 rc->rate_factor_constant = pow( base_cplx, 1 - h->param.rc.f_qcompress )
352 / qp2qscale( h->param.rc.f_rf_constant );
355 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
356 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
357 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
358 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
359 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
361 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
362 rc->last_qscale = qp2qscale(26);
363 rc->pred = x264_malloc( 5*sizeof(predictor_t) );
364 rc->pred_b_from_p = x264_malloc( sizeof(predictor_t) );
365 for( i = 0; i < 5; i++ )
367 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
368 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
369 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
370 rc->pred[i].coeff= 2.0;
371 rc->pred[i].count= 1.0;
372 rc->pred[i].decay= 0.5;
373 rc->row_preds[i].coeff= .25;
374 rc->row_preds[i].count= 1.0;
375 rc->row_preds[i].decay= 0.5;
377 *rc->pred_b_from_p = rc->pred[0];
379 if( parse_zones( h ) < 0 )
381 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
385 /* Load stat file and init 2pass algo */
386 if( h->param.rc.b_stat_read )
388 char *p, *stats_in, *stats_buf;
390 /* read 1st pass stats */
391 assert( h->param.rc.psz_stat_in );
392 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
395 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
399 /* check whether 1st pass options were compatible with current options */
400 if( !strncmp( stats_buf, "#options:", 9 ) )
403 char *opts = stats_buf;
404 stats_in = strchr( stats_buf, '\n' );
410 if( ( p = strstr( opts, "bframes=" ) ) && sscanf( p, "bframes=%d", &i )
411 && h->param.i_bframe != i )
413 x264_log( h, X264_LOG_ERROR, "different number of B-frames than 1st pass (%d vs %d)\n",
414 h->param.i_bframe, i );
418 /* since B-adapt doesn't (yet) take into account B-pyramid,
419 * the converse is not a problem */
420 if( strstr( opts, "b_pyramid=1" ) && !h->param.b_bframe_pyramid )
421 x264_log( h, X264_LOG_WARNING, "1st pass used B-pyramid, 2nd doesn't\n" );
423 if( ( p = strstr( opts, "keyint=" ) ) && sscanf( p, "keyint=%d", &i )
424 && h->param.i_keyint_max != i )
425 x264_log( h, X264_LOG_WARNING, "different keyint than 1st pass (%d vs %d)\n",
426 h->param.i_keyint_max, i );
428 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
429 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
432 /* find number of pics */
435 p = strchr(p+1, ';');
438 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
443 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
445 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
446 h->param.i_frame_total, rc->num_entries );
448 if( h->param.i_frame_total > rc->num_entries + h->param.i_bframe )
450 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
451 h->param.i_frame_total, rc->num_entries );
455 /* FIXME: ugly padding because VfW drops delayed B-frames */
456 rc->num_entries += h->param.i_bframe;
458 rc->entry = (ratecontrol_entry_t*) x264_malloc(rc->num_entries * sizeof(ratecontrol_entry_t));
459 memset(rc->entry, 0, rc->num_entries * sizeof(ratecontrol_entry_t));
461 /* init all to skipped p frames */
462 for(i=0; i<rc->num_entries; i++)
464 ratecontrol_entry_t *rce = &rc->entry[i];
465 rce->pict_type = SLICE_TYPE_P;
466 rce->qscale = rce->new_qscale = qp2qscale(20);
467 rce->misc_bits = rc->nmb + 10;
473 for(i=0; i < rc->num_entries - h->param.i_bframe; i++)
475 ratecontrol_entry_t *rce;
482 next= strchr(p, ';');
485 (*next)=0; //sscanf is unbelievably slow on long strings
488 e = sscanf(p, " in:%d ", &frame_number);
490 if(frame_number < 0 || frame_number >= rc->num_entries)
492 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
495 rce = &rc->entry[frame_number];
496 rce->direct_mode = 0;
498 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",
499 &pict_type, &qp, &rce->tex_bits,
500 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
501 &rce->s_count, &rce->direct_mode);
505 case 'I': rce->kept_as_ref = 1;
506 case 'i': rce->pict_type = SLICE_TYPE_I; break;
507 case 'P': rce->pict_type = SLICE_TYPE_P; break;
508 case 'B': rce->kept_as_ref = 1;
509 case 'b': rce->pict_type = SLICE_TYPE_B; break;
510 default: e = -1; break;
514 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
517 rce->qscale = qp2qscale(qp);
521 x264_free(stats_buf);
523 if(h->param.rc.i_rc_method == X264_RC_ABR)
525 if(init_pass2(h) < 0) return -1;
526 } /* else we're using constant quant, so no need to run the bitrate allocation */
529 /* Open output file */
530 /* If input and output files are the same, output to a temp file
531 * and move it to the real name only when it's complete */
532 if( h->param.rc.b_stat_write )
536 rc->psz_stat_file_tmpname = x264_malloc( strlen(h->param.rc.psz_stat_out) + 6 );
537 strcpy( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
538 strcat( rc->psz_stat_file_tmpname, ".temp" );
540 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
541 if( rc->p_stat_file_out == NULL )
543 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
547 p = x264_param2string( &h->param, 1 );
548 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
552 for( i=0; i<h->param.i_threads; i++ )
554 h->thread[i]->rc = rc+i;
557 if( h->param.rc.i_aq_mode == X264_AQ_LOCAL )
558 rc[i].ac_energy = x264_malloc( h->mb.i_mb_count * sizeof(int) );
564 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
569 z->f_bitrate_factor = 1;
570 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
572 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
574 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
578 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
584 z->param = malloc( sizeof(x264_param_t) );
585 memcpy( z->param, &h->param, sizeof(x264_param_t) );
586 while( (tok = strtok_r( p, ",", &saveptr )) )
588 char *val = strchr( tok, '=' );
594 if( x264_param_parse( z->param, tok, val ) )
596 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
604 static int parse_zones( x264_t *h )
606 x264_ratecontrol_t *rc = h->rc;
608 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
610 char *p, *tok, *saveptr;
611 char *psz_zones = x264_malloc( strlen(h->param.rc.psz_zones)+1 );
612 strcpy( psz_zones, h->param.rc.psz_zones );
613 h->param.rc.i_zones = 1;
614 for( p = psz_zones; *p; p++ )
615 h->param.rc.i_zones += (*p == '/');
616 h->param.rc.zones = x264_malloc( h->param.rc.i_zones * sizeof(x264_zone_t) );
618 for( i = 0; i < h->param.rc.i_zones; i++ )
620 tok = strtok_r( p, "/", &saveptr );
621 if( !tok || parse_zone( h, &h->param.rc.zones[i], tok ) )
625 x264_free( psz_zones );
628 if( h->param.rc.i_zones > 0 )
630 for( i = 0; i < h->param.rc.i_zones; i++ )
632 x264_zone_t z = h->param.rc.zones[i];
633 if( z.i_start < 0 || z.i_start > z.i_end )
635 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
636 z.i_start, z.i_end );
639 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
641 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
642 z.f_bitrate_factor );
647 rc->i_zones = h->param.rc.i_zones + 1;
648 rc->zones = x264_malloc( rc->i_zones * sizeof(x264_zone_t) );
649 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
651 // default zone to fall back to if none of the others match
652 rc->zones[0].i_start = 0;
653 rc->zones[0].i_end = INT_MAX;
654 rc->zones[0].b_force_qp = 0;
655 rc->zones[0].f_bitrate_factor = 1;
656 rc->zones[0].param = x264_malloc( sizeof(x264_param_t) );
657 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
658 for( i = 1; i < rc->i_zones; i++ )
660 if( !rc->zones[i].param )
661 rc->zones[i].param = rc->zones[0].param;
668 static x264_zone_t *get_zone( x264_t *h, int frame_num )
671 for( i = h->rc->i_zones-1; i >= 0; i-- )
673 x264_zone_t *z = &h->rc->zones[i];
674 if( frame_num >= z->i_start && frame_num <= z->i_end )
680 void x264_ratecontrol_summary( x264_t *h )
682 x264_ratecontrol_t *rc = h->rc;
683 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
685 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
686 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
687 qscale2qp( pow( base_cplx, 1 - h->param.rc.f_qcompress )
688 * rc->cplxr_sum / rc->wanted_bits_window ) );
692 void x264_ratecontrol_delete( x264_t *h )
694 x264_ratecontrol_t *rc = h->rc;
697 if( rc->p_stat_file_out )
699 fclose( rc->p_stat_file_out );
700 if( h->i_frame >= rc->num_entries - h->param.i_bframe )
701 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
703 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
704 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
706 x264_free( rc->psz_stat_file_tmpname );
708 x264_free( rc->pred );
709 x264_free( rc->pred_b_from_p );
710 x264_free( rc->entry );
713 x264_free( rc->zones[0].param );
714 if( h->param.rc.psz_zones )
715 for( i=1; i<rc->i_zones; i++ )
716 if( rc->zones[i].param != rc->zones[0].param )
717 x264_free( rc->zones[i].param );
718 x264_free( rc->zones );
720 for( i=0; i<h->param.i_threads; i++ )
721 x264_free( rc[i].ac_energy );
725 void x264_ratecontrol_set_estimated_size( x264_t *h, int bits )
727 x264_pthread_mutex_lock( &h->fenc->mutex );
728 h->rc->frame_size_estimated = bits;
729 x264_pthread_mutex_unlock( &h->fenc->mutex );
732 int x264_ratecontrol_get_estimated_size( x264_t const *h)
735 x264_pthread_mutex_lock( &h->fenc->mutex );
736 size = h->rc->frame_size_estimated;
737 x264_pthread_mutex_unlock( &h->fenc->mutex );
741 static void accum_p_qp_update( x264_t *h, float qp )
743 x264_ratecontrol_t *rc = h->rc;
744 rc->accum_p_qp *= .95;
745 rc->accum_p_norm *= .95;
746 rc->accum_p_norm += 1;
747 if( h->sh.i_type == SLICE_TYPE_I )
748 rc->accum_p_qp += qp + rc->ip_offset;
750 rc->accum_p_qp += qp;
753 /* Before encoding a frame, choose a QP for it */
754 void x264_ratecontrol_start( x264_t *h, int i_force_qp )
756 x264_ratecontrol_t *rc = h->rc;
757 ratecontrol_entry_t *rce = NULL;
758 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
763 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
764 x264_encoder_reconfig( h, zone->param );
765 rc->prev_zone = zone;
767 rc->qp_force = i_force_qp;
769 if( h->param.rc.b_stat_read )
771 int frame = h->fenc->i_frame;
772 assert( frame >= 0 && frame < rc->num_entries );
773 rce = h->rc->rce = &h->rc->entry[frame];
775 if( h->sh.i_type == SLICE_TYPE_B
776 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
778 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
779 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
785 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
786 rc->row_pred = &rc->row_preds[h->sh.i_type];
787 update_vbv_plan( h );
790 if( h->sh.i_type != SLICE_TYPE_B )
793 while( h->frames.current[rc->bframes] && IS_X264_TYPE_B(h->frames.current[rc->bframes]->i_type) )
803 q = qscale2qp( rate_estimate_qscale( h ) );
805 else if( rc->b_2pass )
807 rce->new_qscale = rate_estimate_qscale( h );
808 q = qscale2qp( rce->new_qscale );
812 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
813 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
815 q = rc->qp_constant[ h->sh.i_type ];
819 if( zone->b_force_qp )
820 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
822 q -= 6*log(zone->f_bitrate_factor)/log(2);
828 h->fdec->f_qp_avg_rc =
829 h->fdec->f_qp_avg_aq =
831 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
834 rce->new_qp = rc->qp;
836 /* accum_p_qp needs to be here so that future frames can benefit from the
837 * data before this frame is done. but this only works because threading
838 * guarantees to not re-encode any frames. so the non-threaded case does
839 * accum_p_qp later. */
840 if( h->param.i_threads > 1 )
841 accum_p_qp_update( h, rc->qp );
843 if( h->sh.i_type != SLICE_TYPE_B )
844 rc->last_non_b_pict_type = h->sh.i_type;
846 /* Adaptive AQ thresholding algorithm. */
847 if( h->param.rc.i_aq_mode == X264_AQ_GLOBAL )
848 /* Arbitrary value for "center" of the AQ curve.
849 * Chosen so that any given value of CRF has on average similar bitrate with and without AQ. */
850 h->rc->aq_threshold = logf(5000);
851 else if( h->param.rc.i_aq_mode == X264_AQ_LOCAL )
852 x264_autosense_aq(h);
855 static double predict_row_size( x264_t *h, int y, int qp )
857 /* average between two predictors:
858 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
859 x264_ratecontrol_t *rc = h->rc;
860 double pred_s = predict_size( rc->row_pred, qp2qscale(qp), h->fdec->i_row_satd[y] );
862 if( h->sh.i_type != SLICE_TYPE_I
863 && h->fref0[0]->i_type == h->fdec->i_type
864 && h->fref0[0]->i_row_satd[y] > 0
865 && (abs(h->fref0[0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
867 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
868 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
873 return (pred_s + pred_t) / 2;
876 static double row_bits_so_far( x264_t *h, int y )
880 for( i = 0; i <= y; i++ )
881 bits += h->fdec->i_row_bits[i];
885 static double predict_row_size_sum( x264_t *h, int y, int qp )
888 double bits = row_bits_so_far(h, y);
889 for( i = y+1; i < h->sps->i_mb_height; i++ )
890 bits += predict_row_size( h, i, qp );
895 void x264_ratecontrol_mb( x264_t *h, int bits )
897 x264_ratecontrol_t *rc = h->rc;
898 const int y = h->mb.i_mb_y;
902 h->fdec->i_row_bits[y] += bits;
903 rc->qpa_rc += rc->f_qpm;
904 rc->qpa_aq += h->mb.i_qp;
906 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !rc->b_vbv)
909 h->fdec->i_row_qp[y] = rc->qpm;
911 if( h->sh.i_type == SLICE_TYPE_B )
913 /* B-frames shouldn't use lower QP than their reference frames.
914 * This code is a bit overzealous in limiting B-frame quantizers, but it helps avoid
915 * underflows due to the fact that B-frames are not explicitly covered by VBV. */
916 if( y < h->sps->i_mb_height-1 )
919 int avg_qp = X264_MAX(h->fref0[0]->i_row_qp[y+1], h->fref1[0]->i_row_qp[y+1])
920 + rc->pb_offset * ((h->fenc->i_type == X264_TYPE_BREF) ? 0.5 : 1);
921 rc->qpm = X264_MIN(X264_MAX( rc->qp, avg_qp), 51); //avg_qp could go higher than 51 due to pb_offset
922 i_estimated = row_bits_so_far(h, y); //FIXME: compute full estimated size
923 if (i_estimated > h->rc->frame_size_planned)
924 x264_ratecontrol_set_estimated_size(h, i_estimated);
929 update_predictor( rc->row_pred, qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
931 /* tweak quality based on difference from predicted size */
932 if( y < h->sps->i_mb_height-1 && h->stat.i_slice_count[h->sh.i_type] > 0 )
934 int prev_row_qp = h->fdec->i_row_qp[y];
935 int b0 = predict_row_size_sum( h, y, rc->qpm );
937 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
938 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
939 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
943 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
944 /* area at the top of the frame was measured inaccurately. */
945 if(row_bits_so_far(h,y) < 0.05 * rc->frame_size_planned)
948 headroom = buffer_left_planned/rc->buffer_size;
949 if(h->sh.i_type != SLICE_TYPE_I)
953 if( !rc->b_vbv_min_rate )
954 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
956 while( rc->qpm < i_qp_max
957 && (b1 > rc->frame_size_planned * rc_tol
958 || (rc->buffer_fill - b1 < buffer_left_planned * 0.5)))
961 b1 = predict_row_size_sum( h, y, rc->qpm );
964 /* avoid VBV underflow */
965 while( (rc->qpm < h->param.rc.i_qp_max)
966 && (rc->buffer_fill - b1 < rc->buffer_size * 0.005))
969 b1 = predict_row_size_sum( h, y, rc->qpm );
972 while( rc->qpm > i_qp_min
973 && rc->qpm > h->fdec->i_row_qp[0]
974 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
975 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
978 b1 = predict_row_size_sum( h, y, rc->qpm );
980 x264_ratecontrol_set_estimated_size(h, b1);
983 /* loses the fractional part of the frame-wise qp */
987 int x264_ratecontrol_qp( x264_t *h )
992 /* In 2pass, force the same frame types as in the 1st pass */
993 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
995 x264_ratecontrol_t *rc = h->rc;
996 if( h->param.rc.b_stat_read )
998 if( frame_num >= rc->num_entries )
1000 /* We could try to initialize everything required for ABR and
1001 * adaptive B-frames, but that would be complicated.
1002 * So just calculate the average QP used so far. */
1004 h->param.rc.i_qp_constant = (h->stat.i_slice_count[SLICE_TYPE_P] == 0) ? 24
1005 : 1 + h->stat.f_slice_qp[SLICE_TYPE_P] / h->stat.i_slice_count[SLICE_TYPE_P];
1006 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
1007 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 );
1008 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 );
1010 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1011 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1012 if( h->param.b_bframe_adaptive )
1013 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1017 h->param.rc.i_rc_method = X264_RC_CQP;
1018 h->param.rc.b_stat_read = 0;
1019 h->param.b_bframe_adaptive = 0;
1020 if( h->param.i_bframe > 1 )
1021 h->param.i_bframe = 1;
1024 switch( rc->entry[frame_num].pict_type )
1027 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I;
1030 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B;
1039 return X264_TYPE_AUTO;
1043 /* After encoding one frame, save stats and update ratecontrol state */
1044 void x264_ratecontrol_end( x264_t *h, int bits )
1046 x264_ratecontrol_t *rc = h->rc;
1047 const int *mbs = h->stat.frame.i_mb_count;
1052 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1053 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1054 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1055 for( i = B_DIRECT; i < B_8x8; i++ )
1056 h->stat.frame.i_mb_count_p += mbs[i];
1058 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1059 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1061 if( h->param.rc.b_stat_write )
1063 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1064 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1065 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1066 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1067 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1068 char c_direct = h->mb.b_direct_auto_write ?
1069 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1070 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1072 fprintf( rc->p_stat_file_out,
1073 "in:%d out:%d type:%c q:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c;\n",
1074 h->fenc->i_frame, h->i_frame,
1076 h->stat.frame.i_tex_bits,
1077 h->stat.frame.i_mv_bits,
1078 h->stat.frame.i_misc_bits,
1079 h->stat.frame.i_mb_count_i,
1080 h->stat.frame.i_mb_count_p,
1081 h->stat.frame.i_mb_count_skip,
1087 if( h->sh.i_type != SLICE_TYPE_B )
1088 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / rc->last_rceq;
1091 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1092 * Not perfectly accurate with B-refs, but good enough. */
1093 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
1095 rc->cplxr_sum *= rc->cbr_decay;
1096 rc->wanted_bits_window += rc->bitrate / rc->fps;
1097 rc->wanted_bits_window *= rc->cbr_decay;
1099 if( h->param.i_threads == 1 )
1100 accum_p_qp_update( h, rc->qpa_rc );
1105 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
1108 if( h->mb.b_variable_qp )
1110 if( h->sh.i_type == SLICE_TYPE_B )
1112 rc->bframe_bits += bits;
1113 if( !h->frames.current[0] || !IS_X264_TYPE_B(h->frames.current[0]->i_type) )
1115 update_predictor( rc->pred_b_from_p, qp2qscale(rc->qpa_rc),
1116 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1117 rc->bframe_bits = 0;
1122 update_vbv( h, bits );
1125 /****************************************************************************
1127 ***************************************************************************/
1130 * modify the bitrate curve from pass1 for one frame
1132 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1134 x264_ratecontrol_t *rcc= h->rc;
1136 x264_zone_t *zone = get_zone( h, frame_num );
1138 q = pow( rce->blurred_complexity, 1 - h->param.rc.f_qcompress );
1140 // avoid NaN's in the rc_eq
1141 if(!isfinite(q) || rce->tex_bits + rce->mv_bits == 0)
1142 q = rcc->last_qscale;
1147 rcc->last_qscale = q;
1152 if( zone->b_force_qp )
1153 q = qp2qscale(zone->i_qp);
1155 q /= zone->f_bitrate_factor;
1161 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1163 x264_ratecontrol_t *rcc = h->rc;
1164 const int pict_type = rce->pict_type;
1166 // force I/B quants as a function of P quants
1167 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1168 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1169 if( pict_type == SLICE_TYPE_I )
1172 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1173 double ip_factor = fabs( h->param.rc.f_ip_factor );
1174 /* don't apply ip_factor if the following frame is also I */
1175 if( rcc->accum_p_norm <= 0 )
1177 else if( h->param.rc.f_ip_factor < 0 )
1179 else if( rcc->accum_p_norm >= 1 )
1182 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1184 else if( pict_type == SLICE_TYPE_B )
1186 if( h->param.rc.f_pb_factor > 0 )
1188 if( !rce->kept_as_ref )
1189 q *= fabs( h->param.rc.f_pb_factor );
1191 else if( pict_type == SLICE_TYPE_P
1192 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1193 && rce->tex_bits == 0 )
1198 /* last qscale / qdiff stuff */
1199 if(rcc->last_non_b_pict_type==pict_type
1200 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1202 double last_q = rcc->last_qscale_for[pict_type];
1203 double max_qscale = last_q * rcc->lstep;
1204 double min_qscale = last_q / rcc->lstep;
1206 if (q > max_qscale) q = max_qscale;
1207 else if(q < min_qscale) q = min_qscale;
1210 rcc->last_qscale_for[pict_type] = q;
1211 if(pict_type!=SLICE_TYPE_B)
1212 rcc->last_non_b_pict_type = pict_type;
1213 if(pict_type==SLICE_TYPE_I)
1215 rcc->last_accum_p_norm = rcc->accum_p_norm;
1216 rcc->accum_p_norm = 0;
1217 rcc->accum_p_qp = 0;
1219 if(pict_type==SLICE_TYPE_P)
1221 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1222 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1223 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1228 static double predict_size( predictor_t *p, double q, double var )
1230 return p->coeff*var / (q*p->count);
1233 static void update_predictor( predictor_t *p, double q, double var, double bits )
1237 p->count *= p->decay;
1238 p->coeff *= p->decay;
1240 p->coeff += bits*q / var;
1243 // update VBV after encoding a frame
1244 static void update_vbv( x264_t *h, int bits )
1246 x264_ratecontrol_t *rcc = h->rc;
1247 x264_ratecontrol_t *rct = h->thread[0]->rc;
1249 if( rcc->last_satd >= h->mb.i_mb_count )
1250 update_predictor( &rct->pred[h->sh.i_type], qp2qscale(rcc->qpa_rc), rcc->last_satd, bits );
1255 rct->buffer_fill_final += rct->buffer_rate - bits;
1256 if( rct->buffer_fill_final < 0 )
1257 x264_log( h, X264_LOG_WARNING, "VBV underflow (%.0f bits)\n", rct->buffer_fill_final );
1258 rct->buffer_fill_final = x264_clip3f( rct->buffer_fill_final, 0, rct->buffer_size );
1261 // provisionally update VBV according to the planned size of all frames currently in progress
1262 static void update_vbv_plan( x264_t *h )
1264 x264_ratecontrol_t *rcc = h->rc;
1265 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final;
1266 if( h->param.i_threads > 1 )
1268 int j = h->rc - h->thread[0]->rc;
1270 for( i=1; i<h->param.i_threads; i++ )
1272 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1273 double bits = t->rc->frame_size_planned;
1274 if( !t->b_thread_active )
1276 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1277 rcc->buffer_fill += rcc->buffer_rate - bits;
1278 rcc->buffer_fill = x264_clip3( rcc->buffer_fill, 0, rcc->buffer_size );
1283 // apply VBV constraints and clip qscale to between lmin and lmax
1284 static double clip_qscale( x264_t *h, int pict_type, double q )
1286 x264_ratecontrol_t *rcc = h->rc;
1287 double lmin = rcc->lmin[pict_type];
1288 double lmax = rcc->lmax[pict_type];
1291 /* B-frames are not directly subject to VBV,
1292 * since they are controlled by the P-frames' QPs.
1293 * FIXME: in 2pass we could modify previous frames' QP too,
1294 * instead of waiting for the buffer to fill */
1296 ( pict_type == SLICE_TYPE_P ||
1297 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) )
1299 if( rcc->buffer_fill/rcc->buffer_size < 0.5 )
1300 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1303 if( rcc->b_vbv && rcc->last_satd > 0 )
1305 /* Now a hard threshold to make sure the frame fits in VBV.
1306 * This one is mostly for I-frames. */
1307 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1309 if( bits > rcc->buffer_fill/2 )
1310 qf = x264_clip3f( rcc->buffer_fill/(2*bits), 0.2, 1.0 );
1313 if( bits < rcc->buffer_rate/2 )
1314 q *= bits*2/rcc->buffer_rate;
1315 q = X264_MAX( q0, q );
1317 /* Check B-frame complexity, and use up any bits that would
1318 * overflow before the next P-frame. */
1319 if( h->sh.i_type == SLICE_TYPE_P )
1321 int nb = rcc->bframes;
1322 double pbbits = bits;
1323 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1326 if( bbits > rcc->buffer_rate )
1328 pbbits += nb * bbits;
1330 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1331 if( pbbits < space )
1333 q *= X264_MAX( pbbits / space,
1334 bits / (0.5 * rcc->buffer_size) );
1336 q = X264_MAX( q0-5, q );
1339 if( !rcc->b_vbv_min_rate )
1340 q = X264_MAX( q0, q );
1345 else if(rcc->b_2pass)
1347 double min2 = log(lmin);
1348 double max2 = log(lmax);
1349 q = (log(q) - min2)/(max2-min2) - 0.5;
1350 q = 1.0/(1.0 + exp(-4*q));
1351 q = q*(max2-min2) + min2;
1355 return x264_clip3f(q, lmin, lmax);
1358 // update qscale for 1 frame based on actual bits used so far
1359 static float rate_estimate_qscale( x264_t *h )
1362 x264_ratecontrol_t *rcc = h->rc;
1363 ratecontrol_entry_t rce;
1364 int pict_type = h->sh.i_type;
1365 double lmin = rcc->lmin[pict_type];
1366 double lmax = rcc->lmax[pict_type];
1367 int64_t total_bits = 8*(h->stat.i_slice_size[SLICE_TYPE_I]
1368 + h->stat.i_slice_size[SLICE_TYPE_P]
1369 + h->stat.i_slice_size[SLICE_TYPE_B]);
1374 if(pict_type != rce.pict_type)
1376 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1377 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1381 if( pict_type == SLICE_TYPE_B )
1383 /* B-frames don't have independent ratecontrol, but rather get the
1384 * average QP of the two adjacent P-frames + an offset */
1386 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1387 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1388 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1389 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1390 float q0 = h->fref0[0]->f_qp_avg_rc;
1391 float q1 = h->fref1[0]->f_qp_avg_rc;
1393 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1394 q0 -= rcc->pb_offset/2;
1395 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1396 q1 -= rcc->pb_offset/2;
1399 q = (q0 + q1) / 2 + rcc->ip_offset;
1405 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1407 if(h->fenc->b_kept_as_ref)
1408 q += rcc->pb_offset/2;
1410 q += rcc->pb_offset;
1412 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1413 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1415 return qp2qscale(q);
1419 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1422 //FIXME adjust abr_buffer based on distance to the end of the video
1423 int64_t diff = total_bits - (int64_t)rce.expected_bits;
1425 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1426 if( h->fenc->i_frame > 30 )
1428 /* Adjust quant based on the difference between
1429 * achieved and expected bitrate so far */
1430 double time = (double)h->fenc->i_frame / rcc->num_entries;
1431 double w = x264_clip3f( time*100, 0.0, 1.0 );
1432 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1436 double expected_size = qscale2bits(&rce, q);
1437 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1438 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
1439 double qmax = q*(2 - expected_fullness);
1440 double size_constraint = 1 + expected_fullness;
1441 if (expected_fullness < .05)
1443 qmax = X264_MIN(qmax, lmax);
1444 while( (expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax) )
1447 expected_size = qscale2bits(&rce, q);
1448 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1450 rcc->last_satd = x264_rc_analyse_slice( h );
1452 q = x264_clip3f( q, lmin, lmax );
1454 else /* 1pass ABR */
1456 /* Calculate the quantizer which would have produced the desired
1457 * average bitrate if it had been applied to all frames so far.
1458 * Then modulate that quant based on the current frame's complexity
1459 * relative to the average complexity so far (using the 2pass RCEQ).
1460 * Then bias the quant up or down if total size so far was far from
1462 * Result: Depending on the value of rate_tolerance, there is a
1463 * tradeoff between quality and bitrate precision. But at large
1464 * tolerances, the bit distribution approaches that of 2pass. */
1466 double wanted_bits, overflow=1, lmin, lmax;
1468 rcc->last_satd = x264_rc_analyse_slice( h );
1469 rcc->short_term_cplxsum *= 0.5;
1470 rcc->short_term_cplxcount *= 0.5;
1471 rcc->short_term_cplxsum += rcc->last_satd;
1472 rcc->short_term_cplxcount ++;
1474 rce.tex_bits = rcc->last_satd;
1475 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1477 rce.p_count = rcc->nmb;
1481 rce.pict_type = pict_type;
1483 if( h->param.rc.i_rc_method == X264_RC_CRF )
1485 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1489 int i_frame_done = h->fenc->i_frame + 1 - h->param.i_threads;
1491 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1493 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
1494 wanted_bits = i_frame_done * rcc->bitrate / rcc->fps;
1495 if( wanted_bits > 0 )
1497 abr_buffer *= X264_MAX( 1, sqrt(i_frame_done/25) );
1498 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1503 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1504 /* should test _next_ pict type, but that isn't decided yet */
1505 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1507 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1508 q /= fabs( h->param.rc.f_ip_factor );
1510 else if( h->i_frame > 0 )
1512 /* Asymmetric clipping, because symmetric would prevent
1513 * overflow control in areas of rapidly oscillating complexity */
1514 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1515 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1516 if( overflow > 1.1 && h->i_frame > 3 )
1518 else if( overflow < 0.9 )
1521 q = x264_clip3f(q, lmin, lmax);
1523 else if( h->param.rc.i_rc_method == X264_RC_CRF )
1525 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1528 //FIXME use get_diff_limited_q() ?
1529 q = clip_qscale( h, pict_type, q );
1532 rcc->last_qscale_for[pict_type] =
1533 rcc->last_qscale = q;
1535 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
1536 rcc->last_qscale_for[SLICE_TYPE_P] = q;
1538 if( rcc->b_2pass && rcc->b_vbv)
1539 rcc->frame_size_planned = qscale2bits(&rce, q);
1541 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1542 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1547 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
1551 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
1552 /* these vars are updated in x264_ratecontrol_start()
1553 * so copy them from the context that most recently started (prev)
1554 * to the context that's about to start (cur).
1560 COPY(last_qscale_for);
1561 COPY(last_non_b_pict_type);
1562 COPY(short_term_cplxsum);
1563 COPY(short_term_cplxcount);
1570 #define COPY(var) next->rc->var = cur->rc->var
1571 /* these vars are updated in x264_ratecontrol_end()
1572 * so copy them from the context that most recently ended (cur)
1573 * to the context that's about to end (next)
1576 COPY(expected_bits_sum);
1577 COPY(wanted_bits_window);
1581 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
1582 /* the rest of the variables are either constant or thread-local */
1585 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
1587 /* find an interval ending on an overflow or underflow (depending on whether
1588 * we're adding or removing bits), and starting on the earliest frame that
1589 * can influence the buffer fill of that end frame. */
1590 x264_ratecontrol_t *rcc = h->rc;
1591 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
1592 const double buffer_max = .9 * rcc->buffer_size;
1593 double fill = fills[*t0-1];
1594 double parity = over ? 1. : -1.;
1595 int i, start=-1, end=-1;
1596 for(i = *t0; i < rcc->num_entries; i++)
1598 fill += (rcc->buffer_rate - qscale2bits(&rcc->entry[i], rcc->entry[i].new_qscale)) * parity;
1599 fill = x264_clip3f(fill, 0, rcc->buffer_size);
1601 if(fill <= buffer_min || i == 0)
1607 else if(fill >= buffer_max && start >= 0)
1612 return start>=0 && end>=0;
1615 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
1617 x264_ratecontrol_t *rcc = h->rc;
1618 double qscale_orig, qscale_new;
1623 for(i = t0; i <= t1; i++)
1625 qscale_orig = rcc->entry[i].new_qscale;
1626 qscale_orig = x264_clip3f(qscale_orig, qscale_min, qscale_max);
1627 qscale_new = qscale_orig * adjustment;
1628 qscale_new = x264_clip3f(qscale_new, qscale_min, qscale_max);
1629 rcc->entry[i].new_qscale = qscale_new;
1630 adjusted = adjusted || (qscale_new != qscale_orig);
1635 static double count_expected_bits( x264_t *h )
1637 x264_ratecontrol_t *rcc = h->rc;
1638 double expected_bits = 0;
1640 for(i = 0; i < rcc->num_entries; i++)
1642 ratecontrol_entry_t *rce = &rcc->entry[i];
1643 rce->expected_bits = expected_bits;
1644 expected_bits += qscale2bits(rce, rce->new_qscale);
1646 return expected_bits;
1649 static void vbv_pass2( x264_t *h )
1651 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
1652 * frames in the interval until either buffer is full at some intermediate frame or the
1653 * last frame in the interval no longer underflows. Recompute intervals and repeat.
1654 * Then do the converse to put bits back into overflow areas until target size is met */
1656 x264_ratecontrol_t *rcc = h->rc;
1657 double *fills = x264_malloc((rcc->num_entries+1)*sizeof(double));
1658 double all_available_bits = h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps;
1659 double expected_bits = 0;
1661 double prev_bits = 0;
1663 double qscale_min = qp2qscale(h->param.rc.i_qp_min);
1664 double qscale_max = qp2qscale(h->param.rc.i_qp_max);
1666 int adj_min, adj_max;
1670 /* adjust overall stream size */
1674 prev_bits = expected_bits;
1676 if(expected_bits != 0)
1677 { /* not first iteration */
1678 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
1679 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
1683 while(adj_min && find_underflow(h, fills, &t0, &t1, 1))
1685 adj_min = fix_underflow(h, t0, t1, adjustment, qscale_min, qscale_max);
1690 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
1692 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
1694 while(adj_max && find_underflow(h, fills, &t0, &t1, 0))
1695 adj_max = fix_underflow(h, t0, t1, 1.001, qscale_min, qscale_max);
1697 expected_bits = count_expected_bits(h);
1698 } while(expected_bits < .995 * all_available_bits && expected_bits > prev_bits);
1701 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
1703 /* store expected vbv filling values for tracking when encoding */
1704 for(i = 0; i < rcc->num_entries; i++)
1705 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
1710 static int init_pass2( x264_t *h )
1712 x264_ratecontrol_t *rcc = h->rc;
1713 uint64_t all_const_bits = 0;
1714 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps);
1715 double rate_factor, step, step_mult;
1716 double qblur = h->param.rc.f_qblur;
1717 double cplxblur = h->param.rc.f_complexity_blur;
1718 const int filter_size = (int)(qblur*4) | 1;
1719 double expected_bits;
1720 double *qscale, *blurred_qscale;
1723 /* find total/average complexity & const_bits */
1724 for(i=0; i<rcc->num_entries; i++)
1726 ratecontrol_entry_t *rce = &rcc->entry[i];
1727 all_const_bits += rce->misc_bits;
1730 if( all_available_bits < all_const_bits)
1732 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
1733 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)));
1737 /* Blur complexities, to reduce local fluctuation of QP.
1738 * We don't blur the QPs directly, because then one very simple frame
1739 * could drag down the QP of a nearby complex frame and give it more
1740 * bits than intended. */
1741 for(i=0; i<rcc->num_entries; i++)
1743 ratecontrol_entry_t *rce = &rcc->entry[i];
1744 double weight_sum = 0;
1745 double cplx_sum = 0;
1746 double weight = 1.0;
1747 double gaussian_weight;
1749 /* weighted average of cplx of future frames */
1750 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++)
1752 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
1753 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1756 gaussian_weight = weight * exp(-j*j/200.0);
1757 weight_sum += gaussian_weight;
1758 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1760 /* weighted average of cplx of past frames */
1762 for(j=0; j<=cplxblur*2 && j<=i; j++)
1764 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
1765 gaussian_weight = weight * exp(-j*j/200.0);
1766 weight_sum += gaussian_weight;
1767 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1768 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1772 rce->blurred_complexity = cplx_sum / weight_sum;
1775 qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1777 blurred_qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1779 blurred_qscale = qscale;
1781 /* Search for a factor which, when multiplied by the RCEQ values from
1782 * each frame, adds up to the desired total size.
1783 * There is no exact closed-form solution because of VBV constraints and
1784 * because qscale2bits is not invertible, but we can start with the simple
1785 * approximation of scaling the 1st pass by the ratio of bitrates.
1786 * The search range is probably overkill, but speed doesn't matter here. */
1789 for(i=0; i<rcc->num_entries; i++)
1790 expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i));
1791 step_mult = all_available_bits / expected_bits;
1794 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
1797 rate_factor += step;
1799 rcc->last_non_b_pict_type = -1;
1800 rcc->last_accum_p_norm = 1;
1801 rcc->accum_p_norm = 0;
1804 for(i=0; i<rcc->num_entries; i++)
1806 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
1809 /* fixed I/B qscale relative to P */
1810 for(i=rcc->num_entries-1; i>=0; i--)
1812 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
1813 assert(qscale[i] >= 0);
1819 assert(filter_size%2==1);
1820 for(i=0; i<rcc->num_entries; i++)
1822 ratecontrol_entry_t *rce = &rcc->entry[i];
1824 double q=0.0, sum=0.0;
1826 for(j=0; j<filter_size; j++)
1828 int index = i+j-filter_size/2;
1830 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
1831 if(index < 0 || index >= rcc->num_entries)
1833 if(rce->pict_type != rcc->entry[index].pict_type)
1835 q += qscale[index] * coeff;
1838 blurred_qscale[i] = q/sum;
1842 /* find expected bits */
1843 for(i=0; i<rcc->num_entries; i++)
1845 ratecontrol_entry_t *rce = &rcc->entry[i];
1846 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
1847 assert(rce->new_qscale >= 0);
1848 expected_bits += qscale2bits(rce, rce->new_qscale);
1851 if(expected_bits > all_available_bits) rate_factor -= step;
1856 x264_free(blurred_qscale);
1860 expected_bits = count_expected_bits(h);
1862 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
1865 for(i=0; i<rcc->num_entries; i++)
1866 avgq += rcc->entry[i].new_qscale;
1867 avgq = qscale2qp(avgq / rcc->num_entries);
1869 if ((expected_bits > all_available_bits) || (!rcc->b_vbv))
1870 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
1871 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
1872 (float)h->param.rc.i_bitrate,
1873 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
1875 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
1877 if(h->param.rc.i_qp_min > 0)
1878 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
1880 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
1882 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
1884 if(h->param.rc.i_qp_max < 51)
1885 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
1887 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
1889 else if(!(rcc->b_2pass && rcc->b_vbv))
1890 x264_log(h, X264_LOG_WARNING, "internal error\n");