1 /***************************************************-*- coding: iso-8859-1 -*-
2 * ratecontrol.c: h264 encoder library (Rate Control)
3 *****************************************************************************
4 * Copyright (C) 2005 x264 project
5 * $Id: ratecontrol.c,v 1.1 2004/06/03 19:27:08 fenrir Exp $
7 * Authors: Loren Merritt <lorenm@u.washington.edu>
8 * Michael Niedermayer <michaelni@gmx.at>
9 * Måns Rullgård <mru@mru.ath.cx>
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2 of the License, or
14 * (at your option) any later version.
16 * This program is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with this program; if not, write to the Free Software
23 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111, USA.
24 *****************************************************************************/
26 #define _ISOC99_SOURCE
27 #undef NDEBUG // always check asserts, the speed effect is far too small to disable them
32 #include "common/common.h"
33 #include "common/cpu.h"
34 #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 */
120 double i_cplx_sum[5]; /* estimated total texture bits in intra MBs at qscale=1 */
121 double p_cplx_sum[5];
122 double mv_bits_sum[5];
123 int frame_count[5]; /* number of frames of each type */
126 double frame_size_estimated;
127 double frame_size_planned;
128 predictor_t *row_pred;
129 predictor_t row_preds[5];
130 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
131 int bframes; /* # consecutive B-frames before this P-frame */
132 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 );
151 int x264_rc_analyse_slice( x264_t *h );
154 * qp = h.264's quantizer
155 * qscale = linearized quantizer = Lagrange multiplier
157 static inline double qp2qscale(double qp)
159 return 0.85 * pow(2.0, ( qp - 12.0 ) / 6.0);
161 static inline double qscale2qp(double qscale)
163 return 12.0 + 6.0 * log(qscale/0.85) / log(2.0);
166 /* Texture bitrate is not quite inversely proportional to qscale,
167 * probably due the the changing number of SKIP blocks.
168 * MV bits level off at about qp<=12, because the lambda used
169 * for motion estimation is constant there. */
170 static inline double qscale2bits(ratecontrol_entry_t *rce, double qscale)
174 return (rce->i_tex_bits + rce->p_tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
175 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
179 // Find the total AC energy of the block in all planes.
180 static NOINLINE int ac_energy_mb( x264_t *h, int mb_x, int mb_y, int *satd )
182 /* This function contains annoying hacks because GCC has a habit of reordering emms
183 * and putting it after floating point ops. As a result, we put the emms at the end of the
184 * function and make sure that its always called before the float math. Noinline makes
185 * sure no reordering goes on. */
186 DECLARE_ALIGNED_16( static uint8_t flat[16] ) = {128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128};
187 unsigned int var=0, sad, ssd, i;
188 if( satd || h->param.rc.i_aq_mode == X264_AQ_GLOBAL )
193 int stride = h->fenc->i_stride[i];
194 int offset = h->mb.b_interlaced
195 ? w * (mb_x + (mb_y&~1) * stride) + (mb_y&1) * stride
196 : w * (mb_x + mb_y * stride);
197 int pix = i ? PIXEL_8x8 : PIXEL_16x16;
198 stride <<= h->mb.b_interlaced;
199 sad = h->pixf.sad[pix]( flat, 0, h->fenc->plane[i]+offset, stride );
200 ssd = h->pixf.ssd[pix]( flat, 0, h->fenc->plane[i]+offset, stride );
201 var += ssd - (sad * sad >> (i?6:8));
202 // SATD to represent the block's overall complexity (bit cost) for intra encoding.
203 // exclude the DC coef, because nothing short of an actual intra prediction will estimate DC cost.
205 *satd += h->pixf.satd[pix]( flat, 0, h->fenc->plane[i]+offset, stride ) - sad/2;
207 var = X264_MAX(var,1);
209 else var = h->rc->ac_energy[h->mb.i_mb_xy];
214 void x264_autosense_aq( x264_t *h )
219 // FIXME: Some of the SATDs might be already calculated elsewhere (ratecontrol?). Can we reuse them?
220 // FIXME: Is chroma SATD necessary?
221 for( mb_y=0; mb_y<h->sps->i_mb_height; mb_y++ )
222 for( mb_x=0; mb_x<h->sps->i_mb_width; mb_x++ )
225 int energy = ac_energy_mb( h, mb_x, mb_y, &satd );
226 h->rc->ac_energy[mb_x + mb_y * h->sps->i_mb_width] = energy;
227 /* Weight the energy value by the SATD value of the MB.
228 * This represents the fact that the more complex blocks in a frame should
229 * be weighted more when calculating the optimal threshold. This also helps
230 * diminish the negative effect of large numbers of simple blocks in a frame,
231 * such as in the case of a letterboxed film. */
232 total += logf(energy) * satd;
236 /* Calculate and store the threshold. */
237 h->rc->aq_threshold = n ? total/n : 15;
240 /*****************************************************************************
241 * x264_adaptive_quant:
242 * adjust macroblock QP based on variance (AC energy) of the MB.
243 * high variance = higher QP
244 * low variance = lower QP
245 * This generally increases SSIM and lowers PSNR.
246 *****************************************************************************/
247 void x264_adaptive_quant( x264_t *h )
249 int energy = ac_energy_mb( h, h->mb.i_mb_x, h->mb.i_mb_y, NULL );
250 /* Adjust the QP based on the AC energy of the macroblock. */
251 float qp = h->rc->f_qpm;
252 float qp_adj = 1.5 * (logf(energy) - h->rc->aq_threshold);
253 if( h->param.rc.i_aq_mode == X264_AQ_LOCAL )
254 qp_adj = x264_clip3f( qp_adj, -5, 5 );
255 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 );
256 /* If the QP of this MB is within 1 of the previous MB, code the same QP as the previous MB,
257 * to lower the bit cost of the qp_delta. */
258 if( abs(h->mb.i_qp - h->mb.i_last_qp) == 1 )
259 h->mb.i_qp = h->mb.i_last_qp;
260 h->mb.i_chroma_qp = i_chroma_qp_table[x264_clip3( h->mb.i_qp + h->pps->i_chroma_qp_index_offset, 0, 51 )];
263 int x264_ratecontrol_new( x264_t *h )
265 x264_ratecontrol_t *rc;
270 rc = h->rc = x264_malloc( h->param.i_threads * sizeof(x264_ratecontrol_t) );
271 memset( rc, 0, h->param.i_threads * sizeof(x264_ratecontrol_t) );
273 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
274 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
276 /* FIXME: use integers */
277 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
278 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
282 rc->bitrate = h->param.rc.i_bitrate * 1000.;
283 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
284 rc->nmb = h->mb.i_mb_count;
285 rc->last_non_b_pict_type = -1;
288 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
290 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
293 if( h->param.rc.i_vbv_buffer_size )
295 if( h->param.rc.i_rc_method == X264_RC_CQP )
296 x264_log(h, X264_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
297 else if( h->param.rc.i_vbv_max_bitrate == 0 )
299 x264_log( h, X264_LOG_DEBUG, "VBV maxrate unspecified, assuming CBR\n" );
300 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
303 if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate &&
304 h->param.rc.i_vbv_max_bitrate > 0)
305 x264_log(h, X264_LOG_WARNING, "max bitrate less than average bitrate, ignored.\n");
306 else if( h->param.rc.i_vbv_max_bitrate > 0 &&
307 h->param.rc.i_vbv_buffer_size > 0 )
309 if( h->param.rc.i_vbv_buffer_size < 3 * h->param.rc.i_vbv_max_bitrate / rc->fps )
311 h->param.rc.i_vbv_buffer_size = 3 * h->param.rc.i_vbv_max_bitrate / rc->fps;
312 x264_log( h, X264_LOG_WARNING, "VBV buffer size too small, using %d kbit\n",
313 h->param.rc.i_vbv_buffer_size );
315 if( h->param.rc.f_vbv_buffer_init > 1. )
316 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 );
317 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000. / rc->fps;
318 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000.;
319 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
320 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
321 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
323 rc->b_vbv_min_rate = !rc->b_2pass
324 && h->param.rc.i_rc_method == X264_RC_ABR
325 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
327 else if( h->param.rc.i_vbv_max_bitrate )
329 x264_log(h, X264_LOG_WARNING, "VBV maxrate specified, but no bufsize.\n");
330 h->param.rc.i_vbv_max_bitrate = 0;
332 if(rc->rate_tolerance < 0.01)
334 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
335 rc->rate_tolerance = 0.01;
338 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
342 /* FIXME ABR_INIT_QP is actually used only in CRF */
343 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
344 rc->accum_p_norm = .01;
345 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
346 /* estimated ratio that produces a reasonable QP for the first I-frame */
347 rc->cplxr_sum = .01 * pow( 7.0e5, h->param.rc.f_qcompress ) * pow( h->mb.i_mb_count, 0.5 );
348 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
349 rc->last_non_b_pict_type = SLICE_TYPE_I;
352 if( h->param.rc.i_rc_method == X264_RC_CRF )
354 /* arbitrary rescaling to make CRF somewhat similar to QP */
355 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
356 rc->rate_factor_constant = pow( base_cplx, 1 - h->param.rc.f_qcompress )
357 / qp2qscale( h->param.rc.f_rf_constant );
360 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
361 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
362 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
363 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
364 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
366 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
367 rc->last_qscale = qp2qscale(26);
368 rc->pred = x264_malloc( 5*sizeof(predictor_t) );
369 rc->pred_b_from_p = x264_malloc( sizeof(predictor_t) );
370 for( i = 0; i < 5; i++ )
372 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
373 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
374 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
375 rc->pred[i].coeff= 2.0;
376 rc->pred[i].count= 1.0;
377 rc->pred[i].decay= 0.5;
378 rc->row_preds[i].coeff= .25;
379 rc->row_preds[i].count= 1.0;
380 rc->row_preds[i].decay= 0.5;
382 *rc->pred_b_from_p = rc->pred[0];
384 if( parse_zones( h ) < 0 )
386 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
390 /* Load stat file and init 2pass algo */
391 if( h->param.rc.b_stat_read )
393 char *p, *stats_in, *stats_buf;
395 /* read 1st pass stats */
396 assert( h->param.rc.psz_stat_in );
397 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
400 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
404 /* check whether 1st pass options were compatible with current options */
405 if( !strncmp( stats_buf, "#options:", 9 ) )
408 char *opts = stats_buf;
409 stats_in = strchr( stats_buf, '\n' );
415 if( ( p = strstr( opts, "bframes=" ) ) && sscanf( p, "bframes=%d", &i )
416 && h->param.i_bframe != i )
418 x264_log( h, X264_LOG_ERROR, "different number of B-frames than 1st pass (%d vs %d)\n",
419 h->param.i_bframe, i );
423 /* since B-adapt doesn't (yet) take into account B-pyramid,
424 * the converse is not a problem */
425 if( strstr( opts, "b_pyramid=1" ) && !h->param.b_bframe_pyramid )
426 x264_log( h, X264_LOG_WARNING, "1st pass used B-pyramid, 2nd doesn't\n" );
428 if( ( p = strstr( opts, "keyint=" ) ) && sscanf( p, "keyint=%d", &i )
429 && h->param.i_keyint_max != i )
430 x264_log( h, X264_LOG_WARNING, "different keyint than 1st pass (%d vs %d)\n",
431 h->param.i_keyint_max, i );
433 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
434 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
437 /* find number of pics */
440 p = strchr(p+1, ';');
443 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
448 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
450 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
451 h->param.i_frame_total, rc->num_entries );
453 if( h->param.i_frame_total > rc->num_entries + h->param.i_bframe )
455 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
456 h->param.i_frame_total, rc->num_entries );
460 /* FIXME: ugly padding because VfW drops delayed B-frames */
461 rc->num_entries += h->param.i_bframe;
463 rc->entry = (ratecontrol_entry_t*) x264_malloc(rc->num_entries * sizeof(ratecontrol_entry_t));
464 memset(rc->entry, 0, rc->num_entries * sizeof(ratecontrol_entry_t));
466 /* init all to skipped p frames */
467 for(i=0; i<rc->num_entries; i++)
469 ratecontrol_entry_t *rce = &rc->entry[i];
470 rce->pict_type = SLICE_TYPE_P;
471 rce->qscale = rce->new_qscale = qp2qscale(20);
472 rce->misc_bits = rc->nmb + 10;
478 for(i=0; i < rc->num_entries - h->param.i_bframe; i++)
480 ratecontrol_entry_t *rce;
487 next= strchr(p, ';');
490 (*next)=0; //sscanf is unbelievably slow on long strings
493 e = sscanf(p, " in:%d ", &frame_number);
495 if(frame_number < 0 || frame_number >= rc->num_entries)
497 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
500 rce = &rc->entry[frame_number];
501 rce->direct_mode = 0;
503 e += sscanf(p, " in:%*d out:%*d type:%c q:%f itex:%d ptex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c",
504 &pict_type, &qp, &rce->i_tex_bits, &rce->p_tex_bits,
505 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
506 &rce->s_count, &rce->direct_mode);
510 case 'I': rce->kept_as_ref = 1;
511 case 'i': rce->pict_type = SLICE_TYPE_I; break;
512 case 'P': rce->pict_type = SLICE_TYPE_P; break;
513 case 'B': rce->kept_as_ref = 1;
514 case 'b': rce->pict_type = SLICE_TYPE_B; break;
515 default: e = -1; break;
519 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
522 rce->qscale = qp2qscale(qp);
526 x264_free(stats_buf);
528 if(h->param.rc.i_rc_method == X264_RC_ABR)
530 if(init_pass2(h) < 0) return -1;
531 } /* else we're using constant quant, so no need to run the bitrate allocation */
534 /* Open output file */
535 /* If input and output files are the same, output to a temp file
536 * and move it to the real name only when it's complete */
537 if( h->param.rc.b_stat_write )
541 rc->psz_stat_file_tmpname = x264_malloc( strlen(h->param.rc.psz_stat_out) + 6 );
542 strcpy( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
543 strcat( rc->psz_stat_file_tmpname, ".temp" );
545 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
546 if( rc->p_stat_file_out == NULL )
548 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
552 p = x264_param2string( &h->param, 1 );
553 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
557 for( i=0; i<h->param.i_threads; i++ )
559 h->thread[i]->rc = rc+i;
562 if( h->param.rc.i_aq_mode == X264_AQ_LOCAL )
563 rc[i].ac_energy = x264_malloc( h->mb.i_mb_count * sizeof(int) );
569 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
574 z->f_bitrate_factor = 1;
575 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
577 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
579 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
583 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
589 z->param = malloc( sizeof(x264_param_t) );
590 memcpy( z->param, &h->param, sizeof(x264_param_t) );
591 while( (tok = strtok_r( p, ",", &saveptr )) )
593 char *val = strchr( tok, '=' );
599 if( x264_param_parse( z->param, tok, val ) )
601 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
609 static int parse_zones( x264_t *h )
611 x264_ratecontrol_t *rc = h->rc;
613 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
615 char *p, *tok, *saveptr;
616 char *psz_zones = x264_malloc( strlen(h->param.rc.psz_zones)+1 );
617 strcpy( psz_zones, h->param.rc.psz_zones );
618 h->param.rc.i_zones = 1;
619 for( p = psz_zones; *p; p++ )
620 h->param.rc.i_zones += (*p == '/');
621 h->param.rc.zones = x264_malloc( h->param.rc.i_zones * sizeof(x264_zone_t) );
623 for( i = 0; i < h->param.rc.i_zones; i++ )
625 tok = strtok_r( p, "/", &saveptr );
626 if( !tok || parse_zone( h, &h->param.rc.zones[i], tok ) )
630 x264_free( psz_zones );
633 if( h->param.rc.i_zones > 0 )
635 for( i = 0; i < h->param.rc.i_zones; i++ )
637 x264_zone_t z = h->param.rc.zones[i];
638 if( z.i_start < 0 || z.i_start > z.i_end )
640 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
641 z.i_start, z.i_end );
644 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
646 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
647 z.f_bitrate_factor );
652 rc->i_zones = h->param.rc.i_zones + 1;
653 rc->zones = x264_malloc( rc->i_zones * sizeof(x264_zone_t) );
654 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
656 // default zone to fall back to if none of the others match
657 rc->zones[0].i_start = 0;
658 rc->zones[0].i_end = INT_MAX;
659 rc->zones[0].b_force_qp = 0;
660 rc->zones[0].f_bitrate_factor = 1;
661 rc->zones[0].param = x264_malloc( sizeof(x264_param_t) );
662 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
663 for( i = 1; i < rc->i_zones; i++ )
665 if( !rc->zones[i].param )
666 rc->zones[i].param = rc->zones[0].param;
673 x264_zone_t *get_zone( x264_t *h, int frame_num )
676 for( i = h->rc->i_zones-1; i >= 0; i-- )
678 x264_zone_t *z = &h->rc->zones[i];
679 if( frame_num >= z->i_start && frame_num <= z->i_end )
685 void x264_ratecontrol_summary( x264_t *h )
687 x264_ratecontrol_t *rc = h->rc;
688 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
690 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
691 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
692 qscale2qp( pow( base_cplx, 1 - h->param.rc.f_qcompress )
693 * rc->cplxr_sum / rc->wanted_bits_window ) );
697 void x264_ratecontrol_delete( x264_t *h )
699 x264_ratecontrol_t *rc = h->rc;
702 if( rc->p_stat_file_out )
704 fclose( rc->p_stat_file_out );
705 if( h->i_frame >= rc->num_entries - h->param.i_bframe )
706 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
708 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
709 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
711 x264_free( rc->psz_stat_file_tmpname );
713 x264_free( rc->pred );
714 x264_free( rc->pred_b_from_p );
715 x264_free( rc->entry );
718 x264_free( rc->zones[0].param );
719 if( h->param.rc.psz_zones )
720 for( i=1; i<rc->i_zones; i++ )
721 if( rc->zones[i].param != rc->zones[0].param )
722 x264_free( rc->zones[i].param );
723 x264_free( rc->zones );
725 for( i=0; i<h->param.i_threads; i++ )
726 x264_free( rc[i].ac_energy );
730 void x264_ratecontrol_set_estimated_size( x264_t *h, int bits )
732 h->rc->frame_size_estimated = bits;
735 int x264_ratecontrol_get_estimated_size( x264_t const *h)
737 return h->rc->frame_size_estimated;
740 static void accum_p_qp_update( x264_t *h, float qp )
742 x264_ratecontrol_t *rc = h->rc;
743 rc->accum_p_qp *= .95;
744 rc->accum_p_norm *= .95;
745 rc->accum_p_norm += 1;
746 if( h->sh.i_type == SLICE_TYPE_I )
747 rc->accum_p_qp += qp + rc->ip_offset;
749 rc->accum_p_qp += qp;
752 /* Before encoding a frame, choose a QP for it */
753 void x264_ratecontrol_start( x264_t *h, int i_force_qp )
755 x264_ratecontrol_t *rc = h->rc;
756 ratecontrol_entry_t *rce = NULL;
757 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
762 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
763 x264_encoder_reconfig( h, zone->param );
764 rc->prev_zone = zone;
766 rc->qp_force = i_force_qp;
768 if( h->param.rc.b_stat_read )
770 int frame = h->fenc->i_frame;
771 assert( frame >= 0 && frame < rc->num_entries );
772 rce = h->rc->rce = &h->rc->entry[frame];
774 if( h->sh.i_type == SLICE_TYPE_B
775 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
777 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
778 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
784 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
785 rc->row_pred = &rc->row_preds[h->sh.i_type];
786 update_vbv_plan( h );
789 if( h->sh.i_type != SLICE_TYPE_B )
792 while( h->frames.current[rc->bframes] && IS_X264_TYPE_B(h->frames.current[rc->bframes]->i_type) )
802 q = qscale2qp( rate_estimate_qscale( h ) );
804 else if( rc->b_2pass )
806 rce->new_qscale = rate_estimate_qscale( h );
807 q = qscale2qp( rce->new_qscale );
811 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
812 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
814 q = rc->qp_constant[ h->sh.i_type ];
818 if( zone->b_force_qp )
819 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
821 q -= 6*log(zone->f_bitrate_factor)/log(2);
827 h->fdec->f_qp_avg_rc =
828 h->fdec->f_qp_avg_aq =
830 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
833 rce->new_qp = rc->qp;
835 /* accum_p_qp needs to be here so that future frames can benefit from the
836 * data before this frame is done. but this only works because threading
837 * guarantees to not re-encode any frames. so the non-threaded case does
838 * accum_p_qp later. */
839 if( h->param.i_threads > 1 )
840 accum_p_qp_update( h, rc->qp );
842 if( h->sh.i_type != SLICE_TYPE_B )
843 rc->last_non_b_pict_type = h->sh.i_type;
845 /* Adaptive AQ thresholding algorithm. */
846 if( h->param.rc.i_aq_mode == X264_AQ_GLOBAL )
847 /* Arbitrary value for "center" of the AQ curve.
848 * Chosen so that any given value of CRF has on average similar bitrate with and without AQ. */
849 h->rc->aq_threshold = logf(5000);
850 else if( h->param.rc.i_aq_mode == X264_AQ_LOCAL )
851 x264_autosense_aq(h);
854 double predict_row_size( x264_t *h, int y, int qp )
856 /* average between two predictors:
857 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
858 x264_ratecontrol_t *rc = h->rc;
859 double pred_s = predict_size( rc->row_pred, qp2qscale(qp), h->fdec->i_row_satd[y] );
861 if( h->sh.i_type != SLICE_TYPE_I
862 && h->fref0[0]->i_type == h->fdec->i_type
863 && h->fref0[0]->i_row_satd[y] > 0 )
865 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
866 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
871 return (pred_s + pred_t) / 2;
874 double row_bits_so_far( x264_t *h, int y )
878 for( i = 0; i <= y; i++ )
879 bits += h->fdec->i_row_bits[i];
883 double predict_row_size_sum( x264_t *h, int y, int qp )
886 double bits = row_bits_so_far(h, y);
887 for( i = y+1; i < h->sps->i_mb_height; i++ )
888 bits += predict_row_size( h, i, qp );
893 void x264_ratecontrol_mb( x264_t *h, int bits )
895 x264_ratecontrol_t *rc = h->rc;
896 const int y = h->mb.i_mb_y;
900 h->fdec->i_row_bits[y] += bits;
901 rc->qpa_rc += rc->f_qpm;
902 rc->qpa_aq += h->mb.i_qp;
904 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !rc->b_vbv)
907 h->fdec->i_row_qp[y] = rc->qpm;
909 if( h->sh.i_type == SLICE_TYPE_B )
911 /* B-frames shouldn't use lower QP than their reference frames.
912 * This code is a bit overzealous in limiting B-frame quantizers, but it helps avoid
913 * underflows due to the fact that B-frames are not explicitly covered by VBV. */
914 if( y < h->sps->i_mb_height-1 )
917 int avg_qp = X264_MAX(h->fref0[0]->i_row_qp[y+1], h->fref1[0]->i_row_qp[y+1])
918 + rc->pb_offset * ((h->fenc->i_type == X264_TYPE_BREF) ? 0.5 : 1);
919 rc->qpm = X264_MIN(X264_MAX( rc->qp, avg_qp), 51); //avg_qp could go higher than 51 due to pb_offset
920 i_estimated = row_bits_so_far(h, y); //FIXME: compute full estimated size
921 if (i_estimated > h->rc->frame_size_planned)
922 x264_frame_size_estimated_set(h, i_estimated);
927 update_predictor( rc->row_pred, qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
929 /* tweak quality based on difference from predicted size */
930 if( y < h->sps->i_mb_height-1 && h->stat.i_slice_count[h->sh.i_type] > 0 )
932 int prev_row_qp = h->fdec->i_row_qp[y];
933 int b0 = predict_row_size_sum( h, y, rc->qpm );
935 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
936 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
937 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
941 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
942 /* area at the top of the frame was measured inaccurately. */
943 if(row_bits_so_far(h,y) < 0.05 * rc->frame_size_planned)
946 headroom = buffer_left_planned/rc->buffer_size;
947 if(h->sh.i_type != SLICE_TYPE_I)
951 if( !rc->b_vbv_min_rate )
952 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
954 while( rc->qpm < i_qp_max
955 && (b1 > rc->frame_size_planned * rc_tol
956 || (rc->buffer_fill - b1 < buffer_left_planned * 0.5)))
959 b1 = predict_row_size_sum( h, y, rc->qpm );
962 /* avoid VBV underflow */
963 while( (rc->qpm < h->param.rc.i_qp_max)
964 && (rc->buffer_fill - b1 < rc->buffer_size * 0.005))
967 b1 = predict_row_size_sum( h, y, rc->qpm );
970 while( rc->qpm > i_qp_min
971 && ((buffer_left_planned > rc->buffer_size * 0.4) || rc->qpm > h->fdec->i_row_qp[0])
972 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
973 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
976 b1 = predict_row_size_sum( h, y, rc->qpm );
978 x264_frame_size_estimated_set(h, b1);
981 /* loses the fractional part of the frame-wise qp */
985 int x264_ratecontrol_qp( x264_t *h )
990 /* In 2pass, force the same frame types as in the 1st pass */
991 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
993 x264_ratecontrol_t *rc = h->rc;
994 if( h->param.rc.b_stat_read )
996 if( frame_num >= rc->num_entries )
998 /* We could try to initialize everything required for ABR and
999 * adaptive B-frames, but that would be complicated.
1000 * So just calculate the average QP used so far. */
1002 h->param.rc.i_qp_constant = (h->stat.i_slice_count[SLICE_TYPE_P] == 0) ? 24
1003 : 1 + h->stat.f_slice_qp[SLICE_TYPE_P] / h->stat.i_slice_count[SLICE_TYPE_P];
1004 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
1005 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 );
1006 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 );
1008 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1009 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1010 if( h->param.b_bframe_adaptive )
1011 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1015 h->param.rc.i_rc_method = X264_RC_CQP;
1016 h->param.rc.b_stat_read = 0;
1017 h->param.b_bframe_adaptive = 0;
1018 if( h->param.i_bframe > 1 )
1019 h->param.i_bframe = 1;
1022 switch( rc->entry[frame_num].pict_type )
1025 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I;
1028 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B;
1037 return X264_TYPE_AUTO;
1041 /* After encoding one frame, save stats and update ratecontrol state */
1042 void x264_ratecontrol_end( x264_t *h, int bits )
1044 x264_ratecontrol_t *rc = h->rc;
1045 const int *mbs = h->stat.frame.i_mb_count;
1050 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1051 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1052 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1053 for( i = B_DIRECT; i < B_8x8; i++ )
1054 h->stat.frame.i_mb_count_p += mbs[i];
1056 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1057 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1059 if( h->param.rc.b_stat_write )
1061 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1062 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1063 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1064 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1065 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1066 char c_direct = h->mb.b_direct_auto_write ?
1067 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1068 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1070 fprintf( rc->p_stat_file_out,
1071 "in:%d out:%d type:%c q:%.2f itex:%d ptex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c;\n",
1072 h->fenc->i_frame, h->i_frame,
1074 h->stat.frame.i_itex_bits, h->stat.frame.i_ptex_bits,
1075 h->stat.frame.i_hdr_bits, h->stat.frame.i_misc_bits,
1076 h->stat.frame.i_mb_count_i,
1077 h->stat.frame.i_mb_count_p,
1078 h->stat.frame.i_mb_count_skip,
1084 if( h->sh.i_type != SLICE_TYPE_B )
1085 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / rc->last_rceq;
1088 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1089 * Not perfectly accurate with B-refs, but good enough. */
1090 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
1092 rc->cplxr_sum *= rc->cbr_decay;
1093 rc->wanted_bits_window += rc->bitrate / rc->fps;
1094 rc->wanted_bits_window *= rc->cbr_decay;
1096 if( h->param.i_threads == 1 )
1097 accum_p_qp_update( h, rc->qpa_rc );
1102 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
1105 if( h->mb.b_variable_qp )
1107 if( h->sh.i_type == SLICE_TYPE_B )
1109 rc->bframe_bits += bits;
1110 if( !h->frames.current[0] || !IS_X264_TYPE_B(h->frames.current[0]->i_type) )
1112 update_predictor( rc->pred_b_from_p, qp2qscale(rc->qpa_rc),
1113 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1114 rc->bframe_bits = 0;
1119 update_vbv( h, bits );
1122 /****************************************************************************
1124 ***************************************************************************/
1126 double x264_eval( char *s, double *const_value, const char **const_name,
1127 double (**func1)(void *, double), const char **func1_name,
1128 double (**func2)(void *, double, double), char **func2_name,
1132 * modify the bitrate curve from pass1 for one frame
1134 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1136 x264_ratecontrol_t *rcc= h->rc;
1137 const int pict_type = rce->pict_type;
1139 x264_zone_t *zone = get_zone( h, frame_num );
1141 double const_values[]={
1142 rce->i_tex_bits * rce->qscale,
1143 rce->p_tex_bits * rce->qscale,
1144 (rce->i_tex_bits + rce->p_tex_bits) * rce->qscale,
1145 rce->mv_bits * rce->qscale,
1146 (double)rce->i_count / rcc->nmb,
1147 (double)rce->p_count / rcc->nmb,
1148 (double)rce->s_count / rcc->nmb,
1149 rce->pict_type == SLICE_TYPE_I,
1150 rce->pict_type == SLICE_TYPE_P,
1151 rce->pict_type == SLICE_TYPE_B,
1152 h->param.rc.f_qcompress,
1153 rcc->i_cplx_sum[SLICE_TYPE_I] / rcc->frame_count[SLICE_TYPE_I],
1154 rcc->i_cplx_sum[SLICE_TYPE_P] / rcc->frame_count[SLICE_TYPE_P],
1155 rcc->p_cplx_sum[SLICE_TYPE_P] / rcc->frame_count[SLICE_TYPE_P],
1156 rcc->p_cplx_sum[SLICE_TYPE_B] / rcc->frame_count[SLICE_TYPE_B],
1157 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / rcc->frame_count[pict_type],
1158 rce->blurred_complexity,
1161 static const char *const_names[]={
1181 static double (*func1[])(void *, double)={
1182 // (void *)bits2qscale,
1183 (void *)qscale2bits,
1186 static const char *func1_names[]={
1192 q = x264_eval((char*)h->param.rc.psz_rc_eq, const_values, const_names, func1, func1_names, NULL, NULL, rce);
1194 // avoid NaN's in the rc_eq
1195 if(!isfinite(q) || rce->i_tex_bits + rce->p_tex_bits + rce->mv_bits == 0)
1196 q = rcc->last_qscale;
1201 rcc->last_qscale = q;
1206 if( zone->b_force_qp )
1207 q = qp2qscale(zone->i_qp);
1209 q /= zone->f_bitrate_factor;
1215 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1217 x264_ratecontrol_t *rcc = h->rc;
1218 const int pict_type = rce->pict_type;
1220 // force I/B quants as a function of P quants
1221 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1222 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1223 if( pict_type == SLICE_TYPE_I )
1226 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1227 double ip_factor = fabs( h->param.rc.f_ip_factor );
1228 /* don't apply ip_factor if the following frame is also I */
1229 if( rcc->accum_p_norm <= 0 )
1231 else if( h->param.rc.f_ip_factor < 0 )
1233 else if( rcc->accum_p_norm >= 1 )
1236 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1238 else if( pict_type == SLICE_TYPE_B )
1240 if( h->param.rc.f_pb_factor > 0 )
1242 if( !rce->kept_as_ref )
1243 q *= fabs( h->param.rc.f_pb_factor );
1245 else if( pict_type == SLICE_TYPE_P
1246 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1247 && rce->i_tex_bits + rce->p_tex_bits == 0 )
1252 /* last qscale / qdiff stuff */
1253 if(rcc->last_non_b_pict_type==pict_type
1254 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1256 double last_q = rcc->last_qscale_for[pict_type];
1257 double max_qscale = last_q * rcc->lstep;
1258 double min_qscale = last_q / rcc->lstep;
1260 if (q > max_qscale) q = max_qscale;
1261 else if(q < min_qscale) q = min_qscale;
1264 rcc->last_qscale_for[pict_type] = q;
1265 if(pict_type!=SLICE_TYPE_B)
1266 rcc->last_non_b_pict_type = pict_type;
1267 if(pict_type==SLICE_TYPE_I)
1269 rcc->last_accum_p_norm = rcc->accum_p_norm;
1270 rcc->accum_p_norm = 0;
1271 rcc->accum_p_qp = 0;
1273 if(pict_type==SLICE_TYPE_P)
1275 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1276 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1277 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1282 static double predict_size( predictor_t *p, double q, double var )
1284 return p->coeff*var / (q*p->count);
1287 static void update_predictor( predictor_t *p, double q, double var, double bits )
1291 p->count *= p->decay;
1292 p->coeff *= p->decay;
1294 p->coeff += bits*q / var;
1297 // update VBV after encoding a frame
1298 static void update_vbv( x264_t *h, int bits )
1300 x264_ratecontrol_t *rcc = h->rc;
1301 x264_ratecontrol_t *rct = h->thread[0]->rc;
1303 if( rcc->last_satd >= h->mb.i_mb_count )
1304 update_predictor( &rct->pred[h->sh.i_type], qp2qscale(rcc->qpa_rc), rcc->last_satd, bits );
1309 rct->buffer_fill_final += rct->buffer_rate - bits;
1310 if( rct->buffer_fill_final < 0 )
1311 x264_log( h, X264_LOG_WARNING, "VBV underflow (%.0f bits)\n", rct->buffer_fill_final );
1312 rct->buffer_fill_final = x264_clip3f( rct->buffer_fill_final, 0, rct->buffer_size );
1315 // provisionally update VBV according to the planned size of all frames currently in progress
1316 static void update_vbv_plan( x264_t *h )
1318 x264_ratecontrol_t *rcc = h->rc;
1319 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final;
1320 if( h->param.i_threads > 1 )
1322 int j = h->rc - h->thread[0]->rc;
1324 for( i=1; i<h->param.i_threads; i++ )
1326 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1327 double bits = t->rc->frame_size_planned;
1328 if( !t->b_thread_active )
1330 bits = X264_MAX(bits, x264_frame_size_estimated_get(t));
1331 rcc->buffer_fill += rcc->buffer_rate - bits;
1332 rcc->buffer_fill = x264_clip3( rcc->buffer_fill, 0, rcc->buffer_size );
1337 // apply VBV constraints and clip qscale to between lmin and lmax
1338 static double clip_qscale( x264_t *h, int pict_type, double q )
1340 x264_ratecontrol_t *rcc = h->rc;
1341 double lmin = rcc->lmin[pict_type];
1342 double lmax = rcc->lmax[pict_type];
1345 /* B-frames are not directly subject to VBV,
1346 * since they are controlled by the P-frames' QPs.
1347 * FIXME: in 2pass we could modify previous frames' QP too,
1348 * instead of waiting for the buffer to fill */
1350 ( pict_type == SLICE_TYPE_P ||
1351 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) )
1353 if( rcc->buffer_fill/rcc->buffer_size < 0.5 )
1354 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1357 if( rcc->b_vbv && rcc->last_satd > 0 )
1359 /* Now a hard threshold to make sure the frame fits in VBV.
1360 * This one is mostly for I-frames. */
1361 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1363 if( bits > rcc->buffer_fill/2 )
1364 qf = x264_clip3f( rcc->buffer_fill/(2*bits), 0.2, 1.0 );
1367 if( bits < rcc->buffer_rate/2 )
1368 q *= bits*2/rcc->buffer_rate;
1369 q = X264_MAX( q0, q );
1371 /* Check B-frame complexity, and use up any bits that would
1372 * overflow before the next P-frame. */
1373 if( h->sh.i_type == SLICE_TYPE_P )
1375 int nb = rcc->bframes;
1376 double pbbits = bits;
1377 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1380 if( bbits > rcc->buffer_rate )
1382 pbbits += nb * bbits;
1384 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1385 if( pbbits < space )
1387 q *= X264_MAX( pbbits / space,
1388 bits / (0.5 * rcc->buffer_size) );
1390 q = X264_MAX( q0-5, q );
1393 if( !rcc->b_vbv_min_rate )
1394 q = X264_MAX( q0, q );
1399 else if(rcc->b_2pass)
1401 double min2 = log(lmin);
1402 double max2 = log(lmax);
1403 q = (log(q) - min2)/(max2-min2) - 0.5;
1404 q = 1.0/(1.0 + exp(-4*q));
1405 q = q*(max2-min2) + min2;
1409 return x264_clip3f(q, lmin, lmax);
1412 // update qscale for 1 frame based on actual bits used so far
1413 static float rate_estimate_qscale( x264_t *h )
1416 x264_ratecontrol_t *rcc = h->rc;
1417 ratecontrol_entry_t rce;
1418 int pict_type = h->sh.i_type;
1419 double lmin = rcc->lmin[pict_type];
1420 double lmax = rcc->lmax[pict_type];
1421 int64_t total_bits = 8*(h->stat.i_slice_size[SLICE_TYPE_I]
1422 + h->stat.i_slice_size[SLICE_TYPE_P]
1423 + h->stat.i_slice_size[SLICE_TYPE_B]);
1428 if(pict_type != rce.pict_type)
1430 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1431 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1435 if( pict_type == SLICE_TYPE_B )
1437 /* B-frames don't have independent ratecontrol, but rather get the
1438 * average QP of the two adjacent P-frames + an offset */
1440 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1441 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1442 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1443 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1444 float q0 = h->fref0[0]->f_qp_avg_rc;
1445 float q1 = h->fref1[0]->f_qp_avg_rc;
1447 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1448 q0 -= rcc->pb_offset/2;
1449 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1450 q1 -= rcc->pb_offset/2;
1453 q = (q0 + q1) / 2 + rcc->ip_offset;
1459 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1461 if(h->fenc->b_kept_as_ref)
1462 q += rcc->pb_offset/2;
1464 q += rcc->pb_offset;
1466 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1467 x264_frame_size_estimated_set(h, rcc->frame_size_planned);
1469 return qp2qscale(q);
1473 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1476 //FIXME adjust abr_buffer based on distance to the end of the video
1477 int64_t diff = total_bits - (int64_t)rce.expected_bits;
1479 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1480 if( h->fenc->i_frame > 30 )
1482 /* Adjust quant based on the difference between
1483 * achieved and expected bitrate so far */
1484 double time = (double)h->fenc->i_frame / rcc->num_entries;
1485 double w = x264_clip3f( time*100, 0.0, 1.0 );
1486 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1490 double expected_size = qscale2bits(&rce, q);
1491 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1492 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
1493 double qmax = q*(2 - expected_fullness);
1494 double size_constraint = 1 + expected_fullness;
1495 if (expected_fullness < .05)
1497 qmax = X264_MIN(qmax, lmax);
1498 while( (expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax) )
1501 expected_size = qscale2bits(&rce, q);
1502 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1504 rcc->last_satd = x264_rc_analyse_slice( h );
1506 q = x264_clip3f( q, lmin, lmax );
1508 else /* 1pass ABR */
1510 /* Calculate the quantizer which would have produced the desired
1511 * average bitrate if it had been applied to all frames so far.
1512 * Then modulate that quant based on the current frame's complexity
1513 * relative to the average complexity so far (using the 2pass RCEQ).
1514 * Then bias the quant up or down if total size so far was far from
1516 * Result: Depending on the value of rate_tolerance, there is a
1517 * tradeoff between quality and bitrate precision. But at large
1518 * tolerances, the bit distribution approaches that of 2pass. */
1520 double wanted_bits, overflow=1, lmin, lmax;
1522 rcc->last_satd = x264_rc_analyse_slice( h );
1523 rcc->short_term_cplxsum *= 0.5;
1524 rcc->short_term_cplxcount *= 0.5;
1525 rcc->short_term_cplxsum += rcc->last_satd;
1526 rcc->short_term_cplxcount ++;
1528 rce.p_tex_bits = rcc->last_satd;
1529 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1532 rce.p_count = rcc->nmb;
1536 rce.pict_type = pict_type;
1538 if( h->param.rc.i_rc_method == X264_RC_CRF )
1540 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1544 int i_frame_done = h->fenc->i_frame + 1 - h->param.i_threads;
1546 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1548 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
1549 wanted_bits = i_frame_done * rcc->bitrate / rcc->fps;
1550 if( wanted_bits > 0 )
1552 abr_buffer *= X264_MAX( 1, sqrt(i_frame_done/25) );
1553 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1558 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1559 /* should test _next_ pict type, but that isn't decided yet */
1560 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1562 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1563 q /= fabs( h->param.rc.f_ip_factor );
1565 else if( h->i_frame > 0 )
1567 /* Asymmetric clipping, because symmetric would prevent
1568 * overflow control in areas of rapidly oscillating complexity */
1569 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1570 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1571 if( overflow > 1.1 && h->i_frame > 3 )
1573 else if( overflow < 0.9 )
1576 q = x264_clip3f(q, lmin, lmax);
1578 else if( h->param.rc.i_rc_method == X264_RC_CRF )
1580 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1583 //FIXME use get_diff_limited_q() ?
1584 q = clip_qscale( h, pict_type, q );
1587 rcc->last_qscale_for[pict_type] =
1588 rcc->last_qscale = q;
1590 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
1591 rcc->last_qscale_for[SLICE_TYPE_P] = q;
1593 if( rcc->b_2pass && rcc->b_vbv)
1594 rcc->frame_size_planned = qscale2bits(&rce, q);
1596 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1597 x264_frame_size_estimated_set(h, rcc->frame_size_planned);
1602 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
1606 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
1607 /* these vars are updated in x264_ratecontrol_start()
1608 * so copy them from the context that most recently started (prev)
1609 * to the context that's about to start (cur).
1615 COPY(last_qscale_for);
1616 COPY(last_non_b_pict_type);
1617 COPY(short_term_cplxsum);
1618 COPY(short_term_cplxcount);
1625 #define COPY(var) next->rc->var = cur->rc->var
1626 /* these vars are updated in x264_ratecontrol_end()
1627 * so copy them from the context that most recently ended (cur)
1628 * to the context that's about to end (next)
1631 COPY(expected_bits_sum);
1632 COPY(wanted_bits_window);
1636 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
1637 /* the rest of the variables are either constant or thread-local */
1640 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
1642 /* find an interval ending on an overflow or underflow (depending on whether
1643 * we're adding or removing bits), and starting on the earliest frame that
1644 * can influence the buffer fill of that end frame. */
1645 x264_ratecontrol_t *rcc = h->rc;
1646 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
1647 const double buffer_max = .9 * rcc->buffer_size;
1648 double fill = fills[*t0-1];
1649 double parity = over ? 1. : -1.;
1650 int i, start=-1, end=-1;
1651 for(i = *t0; i < rcc->num_entries; i++)
1653 fill += (rcc->buffer_rate - qscale2bits(&rcc->entry[i], rcc->entry[i].new_qscale)) * parity;
1654 fill = x264_clip3f(fill, 0, rcc->buffer_size);
1656 if(fill <= buffer_min || i == 0)
1662 else if(fill >= buffer_max && start >= 0)
1667 return start>=0 && end>=0;
1670 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
1672 x264_ratecontrol_t *rcc = h->rc;
1673 double qscale_orig, qscale_new;
1678 for(i = t0; i <= t1; i++)
1680 qscale_orig = rcc->entry[i].new_qscale;
1681 qscale_orig = x264_clip3f(qscale_orig, qscale_min, qscale_max);
1682 qscale_new = qscale_orig * adjustment;
1683 qscale_new = x264_clip3f(qscale_new, qscale_min, qscale_max);
1684 rcc->entry[i].new_qscale = qscale_new;
1685 adjusted = adjusted || (qscale_new != qscale_orig);
1690 static double count_expected_bits( x264_t *h )
1692 x264_ratecontrol_t *rcc = h->rc;
1693 double expected_bits = 0;
1695 for(i = 0; i < rcc->num_entries; i++)
1697 ratecontrol_entry_t *rce = &rcc->entry[i];
1698 rce->expected_bits = expected_bits;
1699 expected_bits += qscale2bits(rce, rce->new_qscale);
1701 return expected_bits;
1704 static void vbv_pass2( x264_t *h )
1706 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
1707 * frames in the interval until either buffer is full at some intermediate frame or the
1708 * last frame in the interval no longer underflows. Recompute intervals and repeat.
1709 * Then do the converse to put bits back into overflow areas until target size is met */
1711 x264_ratecontrol_t *rcc = h->rc;
1712 double *fills = x264_malloc((rcc->num_entries+1)*sizeof(double));
1713 double all_available_bits = h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps;
1714 double expected_bits = 0;
1716 double prev_bits = 0;
1718 double qscale_min = qp2qscale(h->param.rc.i_qp_min);
1719 double qscale_max = qp2qscale(h->param.rc.i_qp_max);
1721 int adj_min, adj_max;
1725 /* adjust overall stream size */
1729 prev_bits = expected_bits;
1731 if(expected_bits != 0)
1732 { /* not first iteration */
1733 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
1734 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
1738 while(adj_min && find_underflow(h, fills, &t0, &t1, 1))
1740 adj_min = fix_underflow(h, t0, t1, adjustment, qscale_min, qscale_max);
1745 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
1747 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
1749 while(adj_max && find_underflow(h, fills, &t0, &t1, 0))
1750 adj_max = fix_underflow(h, t0, t1, 1.001, qscale_min, qscale_max);
1752 expected_bits = count_expected_bits(h);
1753 } while(expected_bits < .995 * all_available_bits && expected_bits > prev_bits);
1756 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
1758 /* store expected vbv filling values for tracking when encoding */
1759 for(i = 0; i < rcc->num_entries; i++)
1760 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
1765 static int init_pass2( x264_t *h )
1767 x264_ratecontrol_t *rcc = h->rc;
1768 uint64_t all_const_bits = 0;
1769 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps);
1770 double rate_factor, step, step_mult;
1771 double qblur = h->param.rc.f_qblur;
1772 double cplxblur = h->param.rc.f_complexity_blur;
1773 const int filter_size = (int)(qblur*4) | 1;
1774 double expected_bits;
1775 double *qscale, *blurred_qscale;
1778 /* find total/average complexity & const_bits */
1779 for(i=0; i<rcc->num_entries; i++)
1781 ratecontrol_entry_t *rce = &rcc->entry[i];
1782 all_const_bits += rce->misc_bits;
1783 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
1784 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
1785 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits * rce->qscale;
1786 rcc->frame_count[rce->pict_type] ++;
1789 if( all_available_bits < all_const_bits)
1791 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
1792 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)));
1796 /* Blur complexities, to reduce local fluctuation of QP.
1797 * We don't blur the QPs directly, because then one very simple frame
1798 * could drag down the QP of a nearby complex frame and give it more
1799 * bits than intended. */
1800 for(i=0; i<rcc->num_entries; i++)
1802 ratecontrol_entry_t *rce = &rcc->entry[i];
1803 double weight_sum = 0;
1804 double cplx_sum = 0;
1805 double weight = 1.0;
1806 double gaussian_weight;
1808 /* weighted average of cplx of future frames */
1809 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++)
1811 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
1812 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1815 gaussian_weight = weight * exp(-j*j/200.0);
1816 weight_sum += gaussian_weight;
1817 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1819 /* weighted average of cplx of past frames */
1821 for(j=0; j<=cplxblur*2 && j<=i; j++)
1823 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
1824 gaussian_weight = weight * exp(-j*j/200.0);
1825 weight_sum += gaussian_weight;
1826 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1827 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1831 rce->blurred_complexity = cplx_sum / weight_sum;
1834 qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1836 blurred_qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1838 blurred_qscale = qscale;
1840 /* Search for a factor which, when multiplied by the RCEQ values from
1841 * each frame, adds up to the desired total size.
1842 * There is no exact closed-form solution because of VBV constraints and
1843 * because qscale2bits is not invertible, but we can start with the simple
1844 * approximation of scaling the 1st pass by the ratio of bitrates.
1845 * The search range is probably overkill, but speed doesn't matter here. */
1848 for(i=0; i<rcc->num_entries; i++)
1849 expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i));
1850 step_mult = all_available_bits / expected_bits;
1853 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
1856 rate_factor += step;
1858 rcc->last_non_b_pict_type = -1;
1859 rcc->last_accum_p_norm = 1;
1860 rcc->accum_p_norm = 0;
1863 for(i=0; i<rcc->num_entries; i++)
1865 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
1868 /* fixed I/B qscale relative to P */
1869 for(i=rcc->num_entries-1; i>=0; i--)
1871 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
1872 assert(qscale[i] >= 0);
1878 assert(filter_size%2==1);
1879 for(i=0; i<rcc->num_entries; i++)
1881 ratecontrol_entry_t *rce = &rcc->entry[i];
1883 double q=0.0, sum=0.0;
1885 for(j=0; j<filter_size; j++)
1887 int index = i+j-filter_size/2;
1889 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
1890 if(index < 0 || index >= rcc->num_entries)
1892 if(rce->pict_type != rcc->entry[index].pict_type)
1894 q += qscale[index] * coeff;
1897 blurred_qscale[i] = q/sum;
1901 /* find expected bits */
1902 for(i=0; i<rcc->num_entries; i++)
1904 ratecontrol_entry_t *rce = &rcc->entry[i];
1905 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
1906 assert(rce->new_qscale >= 0);
1907 expected_bits += qscale2bits(rce, rce->new_qscale);
1910 if(expected_bits > all_available_bits) rate_factor -= step;
1915 x264_free(blurred_qscale);
1919 expected_bits = count_expected_bits(h);
1921 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
1924 for(i=0; i<rcc->num_entries; i++)
1925 avgq += rcc->entry[i].new_qscale;
1926 avgq = qscale2qp(avgq / rcc->num_entries);
1928 if ((expected_bits > all_available_bits) || (!rcc->b_vbv))
1929 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
1930 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
1931 (float)h->param.rc.i_bitrate,
1932 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
1934 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
1936 if(h->param.rc.i_qp_min > 0)
1937 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
1939 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
1941 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
1943 if(h->param.rc.i_qp_max < 51)
1944 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
1946 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
1948 else if(!(rcc->b_2pass && rcc->b_vbv))
1949 x264_log(h, X264_LOG_WARNING, "internal error\n");