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"
46 uint64_t expected_bits;
53 float blurred_complexity;
55 } ratecontrol_entry_t;
64 struct x264_ratecontrol_t
73 double rate_tolerance;
74 int nmb; /* number of macroblocks in a frame */
78 ratecontrol_entry_t *rce;
79 int qp; /* qp for current frame */
80 int qpm; /* qp for current macroblock */
81 float f_qpm; /* qp for current macroblock: precise float for AQ */
82 float qpa_rc; /* average of macroblocks' qp before aq */
83 float qpa_aq; /* average of macroblocks' qp after aq */
88 double buffer_fill_final; /* real buffer as of the last finished frame */
89 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
90 double buffer_rate; /* # of bits added to buffer_fill after each frame */
91 predictor_t *pred; /* predict frame size from satd */
96 double cplxr_sum; /* sum of bits*qscale/rceq */
97 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow */
98 double wanted_bits_window; /* target bitrate * window */
100 double short_term_cplxsum;
101 double short_term_cplxcount;
102 double rate_factor_constant;
107 FILE *p_stat_file_out;
108 char *psz_stat_file_tmpname;
110 int num_entries; /* number of ratecontrol_entry_ts */
111 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
113 double last_qscale_for[5]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
114 int last_non_b_pict_type;
115 double accum_p_qp; /* for determining I-frame quant */
117 double last_accum_p_norm;
118 double lmin[5]; /* min qscale by frame type */
120 double lstep; /* max change (multiply) in qscale per frame */
121 double i_cplx_sum[5]; /* estimated total texture bits in intra MBs at qscale=1 */
122 double p_cplx_sum[5];
123 double mv_bits_sum[5];
124 int frame_count[5]; /* number of frames of each type */
127 double frame_size_estimated;
128 double frame_size_planned;
129 predictor_t *row_pred;
130 predictor_t row_preds[5];
131 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
132 int bframes; /* # consecutive B-frames before this P-frame */
133 int bframe_bits; /* total cost of those frames */
141 x264_zone_t *prev_zone;
145 static int parse_zones( x264_t *h );
146 static int init_pass2(x264_t *);
147 static float rate_estimate_qscale( x264_t *h );
148 static void update_vbv( x264_t *h, int bits );
149 static void update_vbv_plan( x264_t *h );
150 static double predict_size( predictor_t *p, double q, double var );
151 static void update_predictor( predictor_t *p, double q, double var, double bits );
152 int x264_rc_analyse_slice( x264_t *h );
155 * qp = h.264's quantizer
156 * qscale = linearized quantizer = Lagrange multiplier
158 static inline double qp2qscale(double qp)
160 return 0.85 * pow(2.0, ( qp - 12.0 ) / 6.0);
162 static inline double qscale2qp(double qscale)
164 return 12.0 + 6.0 * log(qscale/0.85) / log(2.0);
167 /* Texture bitrate is not quite inversely proportional to qscale,
168 * probably due the the changing number of SKIP blocks.
169 * MV bits level off at about qp<=12, because the lambda used
170 * for motion estimation is constant there. */
171 static inline double qscale2bits(ratecontrol_entry_t *rce, double qscale)
175 return (rce->i_tex_bits + rce->p_tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
176 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
180 // Find the total AC energy of the block in all planes.
181 static NOINLINE int ac_energy_mb( x264_t *h, int mb_x, int mb_y, int *satd )
183 /* This function contains annoying hacks because GCC has a habit of reordering emms
184 * and putting it after floating point ops. As a result, we put the emms at the end of the
185 * function and make sure that its always called before the float math. Noinline makes
186 * sure no reordering goes on. */
187 /* FIXME: This array is larger than necessary because a bug in GCC causes an all-zero
188 * array to be placed in .bss despite .bss not being correctly aligned on some platforms (win32?) */
189 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};
190 unsigned int var=0, sad, ssd, i;
191 if( satd || h->param.rc.i_aq_mode == X264_AQ_GLOBAL )
196 int stride = h->fenc->i_stride[i];
197 int offset = h->mb.b_interlaced
198 ? w * (mb_x + (mb_y&~1) * stride) + (mb_y&1) * stride
199 : w * (mb_x + mb_y * stride);
200 int pix = i ? PIXEL_8x8 : PIXEL_16x16;
201 stride <<= h->mb.b_interlaced;
202 sad = h->pixf.sad[pix]( zero, 0, h->fenc->plane[i]+offset, stride );
203 ssd = h->pixf.ssd[pix]( zero, 0, h->fenc->plane[i]+offset, stride );
204 var += ssd - (sad * sad >> (i?6:8));
205 // SATD to represent the block's overall complexity (bit cost) for intra encoding.
206 // exclude the DC coef, because nothing short of an actual intra prediction will estimate DC cost.
208 *satd += h->pixf.satd[pix]( zero, 0, h->fenc->plane[i]+offset, stride ) - sad/2;
210 var = X264_MAX(var,1);
212 else var = h->rc->ac_energy[h->mb.i_mb_xy];
217 void x264_autosense_aq( x264_t *h )
222 // FIXME: Some of the SATDs might be already calculated elsewhere (ratecontrol?). Can we reuse them?
223 // FIXME: Is chroma SATD necessary?
224 for( mb_y=0; mb_y<h->sps->i_mb_height; mb_y++ )
225 for( mb_x=0; mb_x<h->sps->i_mb_width; mb_x++ )
228 int energy = ac_energy_mb( h, mb_x, mb_y, &satd );
229 h->rc->ac_energy[mb_x + mb_y * h->sps->i_mb_width] = energy;
230 /* Weight the energy value by the SATD value of the MB.
231 * This represents the fact that the more complex blocks in a frame should
232 * be weighted more when calculating the optimal threshold. This also helps
233 * diminish the negative effect of large numbers of simple blocks in a frame,
234 * such as in the case of a letterboxed film. */
235 total += logf(energy) * satd;
239 /* Calculate and store the threshold. */
240 h->rc->aq_threshold = n ? total/n : 15;
243 /*****************************************************************************
244 * x264_adaptive_quant:
245 * adjust macroblock QP based on variance (AC energy) of the MB.
246 * high variance = higher QP
247 * low variance = lower QP
248 * This generally increases SSIM and lowers PSNR.
249 *****************************************************************************/
250 void x264_adaptive_quant( x264_t *h )
252 int energy = ac_energy_mb( h, h->mb.i_mb_x, h->mb.i_mb_y, NULL );
253 /* Adjust the QP based on the AC energy of the macroblock. */
254 float qp = h->rc->f_qpm;
255 float qp_adj = 1.5 * (logf(energy) - h->rc->aq_threshold);
256 if( h->param.rc.i_aq_mode == X264_AQ_LOCAL )
257 qp_adj = x264_clip3f( qp_adj, -5, 5 );
258 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 );
259 /* If the QP of this MB is within 1 of the previous MB, code the same QP as the previous MB,
260 * to lower the bit cost of the qp_delta. */
261 if( abs(h->mb.i_qp - h->mb.i_last_qp) == 1 )
262 h->mb.i_qp = h->mb.i_last_qp;
263 h->mb.i_chroma_qp = i_chroma_qp_table[x264_clip3( h->mb.i_qp + h->pps->i_chroma_qp_index_offset, 0, 51 )];
266 int x264_ratecontrol_new( x264_t *h )
268 x264_ratecontrol_t *rc;
273 rc = h->rc = x264_malloc( h->param.i_threads * sizeof(x264_ratecontrol_t) );
274 memset( rc, 0, h->param.i_threads * sizeof(x264_ratecontrol_t) );
276 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
277 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
279 /* FIXME: use integers */
280 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
281 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
285 rc->bitrate = h->param.rc.i_bitrate * 1000.;
286 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
287 rc->nmb = h->mb.i_mb_count;
288 rc->last_non_b_pict_type = -1;
291 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
293 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
296 if( h->param.rc.i_vbv_buffer_size )
298 if( h->param.rc.i_rc_method == X264_RC_CQP )
299 x264_log(h, X264_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
300 else if( h->param.rc.i_vbv_max_bitrate == 0 )
302 x264_log( h, X264_LOG_DEBUG, "VBV maxrate unspecified, assuming CBR\n" );
303 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
306 if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate &&
307 h->param.rc.i_vbv_max_bitrate > 0)
308 x264_log(h, X264_LOG_WARNING, "max bitrate less than average bitrate, ignored.\n");
309 else if( h->param.rc.i_vbv_max_bitrate > 0 &&
310 h->param.rc.i_vbv_buffer_size > 0 )
312 if( h->param.rc.i_vbv_buffer_size < 3 * h->param.rc.i_vbv_max_bitrate / rc->fps )
314 h->param.rc.i_vbv_buffer_size = 3 * h->param.rc.i_vbv_max_bitrate / rc->fps;
315 x264_log( h, X264_LOG_WARNING, "VBV buffer size too small, using %d kbit\n",
316 h->param.rc.i_vbv_buffer_size );
318 if( h->param.rc.f_vbv_buffer_init > 1. )
319 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 );
320 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000. / rc->fps;
321 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000.;
322 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
323 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
324 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
326 rc->b_vbv_min_rate = !rc->b_2pass
327 && h->param.rc.i_rc_method == X264_RC_ABR
328 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
330 else if( h->param.rc.i_vbv_max_bitrate )
332 x264_log(h, X264_LOG_WARNING, "VBV maxrate specified, but no bufsize.\n");
333 h->param.rc.i_vbv_max_bitrate = 0;
335 if(rc->rate_tolerance < 0.01)
337 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
338 rc->rate_tolerance = 0.01;
341 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
345 /* FIXME ABR_INIT_QP is actually used only in CRF */
346 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
347 rc->accum_p_norm = .01;
348 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
349 /* estimated ratio that produces a reasonable QP for the first I-frame */
350 rc->cplxr_sum = .01 * pow( 7.0e5, h->param.rc.f_qcompress ) * pow( h->mb.i_mb_count, 0.5 );
351 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
352 rc->last_non_b_pict_type = SLICE_TYPE_I;
355 if( h->param.rc.i_rc_method == X264_RC_CRF )
357 /* arbitrary rescaling to make CRF somewhat similar to QP */
358 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
359 rc->rate_factor_constant = pow( base_cplx, 1 - h->param.rc.f_qcompress )
360 / qp2qscale( h->param.rc.f_rf_constant );
363 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
364 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
365 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
366 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
367 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
369 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
370 rc->last_qscale = qp2qscale(26);
371 rc->pred = x264_malloc( 5*sizeof(predictor_t) );
372 rc->pred_b_from_p = x264_malloc( sizeof(predictor_t) );
373 for( i = 0; i < 5; i++ )
375 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
376 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
377 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
378 rc->pred[i].coeff= 2.0;
379 rc->pred[i].count= 1.0;
380 rc->pred[i].decay= 0.5;
381 rc->row_preds[i].coeff= .25;
382 rc->row_preds[i].count= 1.0;
383 rc->row_preds[i].decay= 0.5;
385 *rc->pred_b_from_p = rc->pred[0];
387 if( parse_zones( h ) < 0 )
389 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
393 /* Load stat file and init 2pass algo */
394 if( h->param.rc.b_stat_read )
396 char *p, *stats_in, *stats_buf;
398 /* read 1st pass stats */
399 assert( h->param.rc.psz_stat_in );
400 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
403 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
407 /* check whether 1st pass options were compatible with current options */
408 if( !strncmp( stats_buf, "#options:", 9 ) )
411 char *opts = stats_buf;
412 stats_in = strchr( stats_buf, '\n' );
418 if( ( p = strstr( opts, "bframes=" ) ) && sscanf( p, "bframes=%d", &i )
419 && h->param.i_bframe != i )
421 x264_log( h, X264_LOG_ERROR, "different number of B-frames than 1st pass (%d vs %d)\n",
422 h->param.i_bframe, i );
426 /* since B-adapt doesn't (yet) take into account B-pyramid,
427 * the converse is not a problem */
428 if( strstr( opts, "b_pyramid=1" ) && !h->param.b_bframe_pyramid )
429 x264_log( h, X264_LOG_WARNING, "1st pass used B-pyramid, 2nd doesn't\n" );
431 if( ( p = strstr( opts, "keyint=" ) ) && sscanf( p, "keyint=%d", &i )
432 && h->param.i_keyint_max != i )
433 x264_log( h, X264_LOG_WARNING, "different keyint than 1st pass (%d vs %d)\n",
434 h->param.i_keyint_max, i );
436 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
437 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
440 /* find number of pics */
443 p = strchr(p+1, ';');
446 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
451 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
453 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
454 h->param.i_frame_total, rc->num_entries );
456 if( h->param.i_frame_total > rc->num_entries + h->param.i_bframe )
458 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
459 h->param.i_frame_total, rc->num_entries );
463 /* FIXME: ugly padding because VfW drops delayed B-frames */
464 rc->num_entries += h->param.i_bframe;
466 rc->entry = (ratecontrol_entry_t*) x264_malloc(rc->num_entries * sizeof(ratecontrol_entry_t));
467 memset(rc->entry, 0, rc->num_entries * sizeof(ratecontrol_entry_t));
469 /* init all to skipped p frames */
470 for(i=0; i<rc->num_entries; i++)
472 ratecontrol_entry_t *rce = &rc->entry[i];
473 rce->pict_type = SLICE_TYPE_P;
474 rce->qscale = rce->new_qscale = qp2qscale(20);
475 rce->misc_bits = rc->nmb + 10;
481 for(i=0; i < rc->num_entries - h->param.i_bframe; i++)
483 ratecontrol_entry_t *rce;
490 next= strchr(p, ';');
493 (*next)=0; //sscanf is unbelievably slow on long strings
496 e = sscanf(p, " in:%d ", &frame_number);
498 if(frame_number < 0 || frame_number >= rc->num_entries)
500 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
503 rce = &rc->entry[frame_number];
504 rce->direct_mode = 0;
506 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",
507 &pict_type, &qp, &rce->i_tex_bits, &rce->p_tex_bits,
508 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
509 &rce->s_count, &rce->direct_mode);
513 case 'I': rce->kept_as_ref = 1;
514 case 'i': rce->pict_type = SLICE_TYPE_I; break;
515 case 'P': rce->pict_type = SLICE_TYPE_P; break;
516 case 'B': rce->kept_as_ref = 1;
517 case 'b': rce->pict_type = SLICE_TYPE_B; break;
518 default: e = -1; break;
522 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
525 rce->qscale = qp2qscale(qp);
529 x264_free(stats_buf);
531 if(h->param.rc.i_rc_method == X264_RC_ABR)
533 if(init_pass2(h) < 0) return -1;
534 } /* else we're using constant quant, so no need to run the bitrate allocation */
537 /* Open output file */
538 /* If input and output files are the same, output to a temp file
539 * and move it to the real name only when it's complete */
540 if( h->param.rc.b_stat_write )
544 rc->psz_stat_file_tmpname = x264_malloc( strlen(h->param.rc.psz_stat_out) + 6 );
545 strcpy( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
546 strcat( rc->psz_stat_file_tmpname, ".temp" );
548 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
549 if( rc->p_stat_file_out == NULL )
551 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
555 p = x264_param2string( &h->param, 1 );
556 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
560 for( i=0; i<h->param.i_threads; i++ )
562 h->thread[i]->rc = rc+i;
565 if( h->param.rc.i_aq_mode == X264_AQ_LOCAL )
566 rc[i].ac_energy = x264_malloc( h->mb.i_mb_count * sizeof(int) );
572 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
577 z->f_bitrate_factor = 1;
578 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
580 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
582 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
586 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
592 z->param = malloc( sizeof(x264_param_t) );
593 memcpy( z->param, &h->param, sizeof(x264_param_t) );
594 while( (tok = strtok_r( p, ",", &saveptr )) )
596 char *val = strchr( tok, '=' );
602 if( x264_param_parse( z->param, tok, val ) )
604 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
612 static int parse_zones( x264_t *h )
614 x264_ratecontrol_t *rc = h->rc;
616 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
618 char *p, *tok, *saveptr;
619 char *psz_zones = x264_malloc( strlen(h->param.rc.psz_zones)+1 );
620 strcpy( psz_zones, h->param.rc.psz_zones );
621 h->param.rc.i_zones = 1;
622 for( p = psz_zones; *p; p++ )
623 h->param.rc.i_zones += (*p == '/');
624 h->param.rc.zones = x264_malloc( h->param.rc.i_zones * sizeof(x264_zone_t) );
626 for( i = 0; i < h->param.rc.i_zones; i++ )
628 tok = strtok_r( p, "/", &saveptr );
629 if( !tok || parse_zone( h, &h->param.rc.zones[i], tok ) )
633 x264_free( psz_zones );
636 if( h->param.rc.i_zones > 0 )
638 for( i = 0; i < h->param.rc.i_zones; i++ )
640 x264_zone_t z = h->param.rc.zones[i];
641 if( z.i_start < 0 || z.i_start > z.i_end )
643 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
644 z.i_start, z.i_end );
647 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
649 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
650 z.f_bitrate_factor );
655 rc->i_zones = h->param.rc.i_zones + 1;
656 rc->zones = x264_malloc( rc->i_zones * sizeof(x264_zone_t) );
657 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
659 // default zone to fall back to if none of the others match
660 rc->zones[0].i_start = 0;
661 rc->zones[0].i_end = INT_MAX;
662 rc->zones[0].b_force_qp = 0;
663 rc->zones[0].f_bitrate_factor = 1;
664 rc->zones[0].param = x264_malloc( sizeof(x264_param_t) );
665 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
666 for( i = 1; i < rc->i_zones; i++ )
668 if( !rc->zones[i].param )
669 rc->zones[i].param = rc->zones[0].param;
676 x264_zone_t *get_zone( x264_t *h, int frame_num )
679 for( i = h->rc->i_zones-1; i >= 0; i-- )
681 x264_zone_t *z = &h->rc->zones[i];
682 if( frame_num >= z->i_start && frame_num <= z->i_end )
688 void x264_ratecontrol_summary( x264_t *h )
690 x264_ratecontrol_t *rc = h->rc;
691 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
693 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
694 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
695 qscale2qp( pow( base_cplx, 1 - h->param.rc.f_qcompress )
696 * rc->cplxr_sum / rc->wanted_bits_window ) );
700 void x264_ratecontrol_delete( x264_t *h )
702 x264_ratecontrol_t *rc = h->rc;
705 if( rc->p_stat_file_out )
707 fclose( rc->p_stat_file_out );
708 if( h->i_frame >= rc->num_entries - h->param.i_bframe )
709 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
711 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
712 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
714 x264_free( rc->psz_stat_file_tmpname );
716 x264_free( rc->pred );
717 x264_free( rc->pred_b_from_p );
718 x264_free( rc->entry );
721 x264_free( rc->zones[0].param );
722 if( h->param.rc.psz_zones )
723 for( i=1; i<rc->i_zones; i++ )
724 if( rc->zones[i].param != rc->zones[0].param )
725 x264_free( rc->zones[i].param );
726 x264_free( rc->zones );
728 for( i=0; i<h->param.i_threads; i++ )
729 x264_free( rc[i].ac_energy );
733 void x264_ratecontrol_set_estimated_size( x264_t *h, int bits )
735 x264_pthread_mutex_lock( &h->fenc->mutex );
736 h->rc->frame_size_estimated = bits;
737 x264_pthread_mutex_unlock( &h->fenc->mutex );
740 int x264_ratecontrol_get_estimated_size( x264_t const *h)
743 x264_pthread_mutex_lock( &h->fenc->mutex );
744 size = h->rc->frame_size_estimated;
745 x264_pthread_mutex_unlock( &h->fenc->mutex );
749 static void accum_p_qp_update( x264_t *h, float qp )
751 x264_ratecontrol_t *rc = h->rc;
752 rc->accum_p_qp *= .95;
753 rc->accum_p_norm *= .95;
754 rc->accum_p_norm += 1;
755 if( h->sh.i_type == SLICE_TYPE_I )
756 rc->accum_p_qp += qp + rc->ip_offset;
758 rc->accum_p_qp += qp;
761 /* Before encoding a frame, choose a QP for it */
762 void x264_ratecontrol_start( x264_t *h, int i_force_qp )
764 x264_ratecontrol_t *rc = h->rc;
765 ratecontrol_entry_t *rce = NULL;
766 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
771 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
772 x264_encoder_reconfig( h, zone->param );
773 rc->prev_zone = zone;
775 rc->qp_force = i_force_qp;
777 if( h->param.rc.b_stat_read )
779 int frame = h->fenc->i_frame;
780 assert( frame >= 0 && frame < rc->num_entries );
781 rce = h->rc->rce = &h->rc->entry[frame];
783 if( h->sh.i_type == SLICE_TYPE_B
784 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
786 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
787 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
793 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
794 rc->row_pred = &rc->row_preds[h->sh.i_type];
795 update_vbv_plan( h );
798 if( h->sh.i_type != SLICE_TYPE_B )
801 while( h->frames.current[rc->bframes] && IS_X264_TYPE_B(h->frames.current[rc->bframes]->i_type) )
811 q = qscale2qp( rate_estimate_qscale( h ) );
813 else if( rc->b_2pass )
815 rce->new_qscale = rate_estimate_qscale( h );
816 q = qscale2qp( rce->new_qscale );
820 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
821 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
823 q = rc->qp_constant[ h->sh.i_type ];
827 if( zone->b_force_qp )
828 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
830 q -= 6*log(zone->f_bitrate_factor)/log(2);
836 h->fdec->f_qp_avg_rc =
837 h->fdec->f_qp_avg_aq =
839 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
842 rce->new_qp = rc->qp;
844 /* accum_p_qp needs to be here so that future frames can benefit from the
845 * data before this frame is done. but this only works because threading
846 * guarantees to not re-encode any frames. so the non-threaded case does
847 * accum_p_qp later. */
848 if( h->param.i_threads > 1 )
849 accum_p_qp_update( h, rc->qp );
851 if( h->sh.i_type != SLICE_TYPE_B )
852 rc->last_non_b_pict_type = h->sh.i_type;
854 /* Adaptive AQ thresholding algorithm. */
855 if( h->param.rc.i_aq_mode == X264_AQ_GLOBAL )
856 /* Arbitrary value for "center" of the AQ curve.
857 * Chosen so that any given value of CRF has on average similar bitrate with and without AQ. */
858 h->rc->aq_threshold = logf(5000);
859 else if( h->param.rc.i_aq_mode == X264_AQ_LOCAL )
860 x264_autosense_aq(h);
863 double predict_row_size( x264_t *h, int y, int qp )
865 /* average between two predictors:
866 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
867 x264_ratecontrol_t *rc = h->rc;
868 double pred_s = predict_size( rc->row_pred, qp2qscale(qp), h->fdec->i_row_satd[y] );
870 if( h->sh.i_type != SLICE_TYPE_I
871 && h->fref0[0]->i_type == h->fdec->i_type
872 && h->fref0[0]->i_row_satd[y] > 0 )
874 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
875 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
880 return (pred_s + pred_t) / 2;
883 double row_bits_so_far( x264_t *h, int y )
887 for( i = 0; i <= y; i++ )
888 bits += h->fdec->i_row_bits[i];
892 double predict_row_size_sum( x264_t *h, int y, int qp )
895 double bits = row_bits_so_far(h, y);
896 for( i = y+1; i < h->sps->i_mb_height; i++ )
897 bits += predict_row_size( h, i, qp );
902 void x264_ratecontrol_mb( x264_t *h, int bits )
904 x264_ratecontrol_t *rc = h->rc;
905 const int y = h->mb.i_mb_y;
909 h->fdec->i_row_bits[y] += bits;
910 rc->qpa_rc += rc->f_qpm;
911 rc->qpa_aq += h->mb.i_qp;
913 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !rc->b_vbv)
916 h->fdec->i_row_qp[y] = rc->qpm;
918 if( h->sh.i_type == SLICE_TYPE_B )
920 /* B-frames shouldn't use lower QP than their reference frames.
921 * This code is a bit overzealous in limiting B-frame quantizers, but it helps avoid
922 * underflows due to the fact that B-frames are not explicitly covered by VBV. */
923 if( y < h->sps->i_mb_height-1 )
926 int avg_qp = X264_MAX(h->fref0[0]->i_row_qp[y+1], h->fref1[0]->i_row_qp[y+1])
927 + rc->pb_offset * ((h->fenc->i_type == X264_TYPE_BREF) ? 0.5 : 1);
928 rc->qpm = X264_MIN(X264_MAX( rc->qp, avg_qp), 51); //avg_qp could go higher than 51 due to pb_offset
929 i_estimated = row_bits_so_far(h, y); //FIXME: compute full estimated size
930 if (i_estimated > h->rc->frame_size_planned)
931 x264_ratecontrol_set_estimated_size(h, i_estimated);
936 update_predictor( rc->row_pred, qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
938 /* tweak quality based on difference from predicted size */
939 if( y < h->sps->i_mb_height-1 && h->stat.i_slice_count[h->sh.i_type] > 0 )
941 int prev_row_qp = h->fdec->i_row_qp[y];
942 int b0 = predict_row_size_sum( h, y, rc->qpm );
944 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
945 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
946 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
950 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
951 /* area at the top of the frame was measured inaccurately. */
952 if(row_bits_so_far(h,y) < 0.05 * rc->frame_size_planned)
955 headroom = buffer_left_planned/rc->buffer_size;
956 if(h->sh.i_type != SLICE_TYPE_I)
960 if( !rc->b_vbv_min_rate )
961 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
963 while( rc->qpm < i_qp_max
964 && (b1 > rc->frame_size_planned * rc_tol
965 || (rc->buffer_fill - b1 < buffer_left_planned * 0.5)))
968 b1 = predict_row_size_sum( h, y, rc->qpm );
971 /* avoid VBV underflow */
972 while( (rc->qpm < h->param.rc.i_qp_max)
973 && (rc->buffer_fill - b1 < rc->buffer_size * 0.005))
976 b1 = predict_row_size_sum( h, y, rc->qpm );
979 while( rc->qpm > i_qp_min
980 && ((buffer_left_planned > rc->buffer_size * 0.4) || rc->qpm > h->fdec->i_row_qp[0])
981 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
982 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
985 b1 = predict_row_size_sum( h, y, rc->qpm );
987 x264_ratecontrol_set_estimated_size(h, b1);
990 /* loses the fractional part of the frame-wise qp */
994 int x264_ratecontrol_qp( x264_t *h )
999 /* In 2pass, force the same frame types as in the 1st pass */
1000 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1002 x264_ratecontrol_t *rc = h->rc;
1003 if( h->param.rc.b_stat_read )
1005 if( frame_num >= rc->num_entries )
1007 /* We could try to initialize everything required for ABR and
1008 * adaptive B-frames, but that would be complicated.
1009 * So just calculate the average QP used so far. */
1011 h->param.rc.i_qp_constant = (h->stat.i_slice_count[SLICE_TYPE_P] == 0) ? 24
1012 : 1 + h->stat.f_slice_qp[SLICE_TYPE_P] / h->stat.i_slice_count[SLICE_TYPE_P];
1013 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
1014 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 );
1015 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 );
1017 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1018 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1019 if( h->param.b_bframe_adaptive )
1020 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1024 h->param.rc.i_rc_method = X264_RC_CQP;
1025 h->param.rc.b_stat_read = 0;
1026 h->param.b_bframe_adaptive = 0;
1027 if( h->param.i_bframe > 1 )
1028 h->param.i_bframe = 1;
1031 switch( rc->entry[frame_num].pict_type )
1034 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I;
1037 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B;
1046 return X264_TYPE_AUTO;
1050 /* After encoding one frame, save stats and update ratecontrol state */
1051 void x264_ratecontrol_end( x264_t *h, int bits )
1053 x264_ratecontrol_t *rc = h->rc;
1054 const int *mbs = h->stat.frame.i_mb_count;
1059 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1060 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1061 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1062 for( i = B_DIRECT; i < B_8x8; i++ )
1063 h->stat.frame.i_mb_count_p += mbs[i];
1065 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1066 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1068 if( h->param.rc.b_stat_write )
1070 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1071 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1072 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1073 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1074 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1075 char c_direct = h->mb.b_direct_auto_write ?
1076 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1077 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1079 fprintf( rc->p_stat_file_out,
1080 "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",
1081 h->fenc->i_frame, h->i_frame,
1083 h->stat.frame.i_itex_bits, h->stat.frame.i_ptex_bits,
1084 h->stat.frame.i_hdr_bits, h->stat.frame.i_misc_bits,
1085 h->stat.frame.i_mb_count_i,
1086 h->stat.frame.i_mb_count_p,
1087 h->stat.frame.i_mb_count_skip,
1093 if( h->sh.i_type != SLICE_TYPE_B )
1094 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / rc->last_rceq;
1097 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1098 * Not perfectly accurate with B-refs, but good enough. */
1099 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
1101 rc->cplxr_sum *= rc->cbr_decay;
1102 rc->wanted_bits_window += rc->bitrate / rc->fps;
1103 rc->wanted_bits_window *= rc->cbr_decay;
1105 if( h->param.i_threads == 1 )
1106 accum_p_qp_update( h, rc->qpa_rc );
1111 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
1114 if( h->mb.b_variable_qp )
1116 if( h->sh.i_type == SLICE_TYPE_B )
1118 rc->bframe_bits += bits;
1119 if( !h->frames.current[0] || !IS_X264_TYPE_B(h->frames.current[0]->i_type) )
1121 update_predictor( rc->pred_b_from_p, qp2qscale(rc->qpa_rc),
1122 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1123 rc->bframe_bits = 0;
1128 update_vbv( h, bits );
1131 /****************************************************************************
1133 ***************************************************************************/
1135 double x264_eval( char *s, double *const_value, const char **const_name,
1136 double (**func1)(void *, double), const char **func1_name,
1137 double (**func2)(void *, double, double), char **func2_name,
1141 * modify the bitrate curve from pass1 for one frame
1143 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1145 x264_ratecontrol_t *rcc= h->rc;
1146 const int pict_type = rce->pict_type;
1148 x264_zone_t *zone = get_zone( h, frame_num );
1150 double const_values[]={
1151 rce->i_tex_bits * rce->qscale,
1152 rce->p_tex_bits * rce->qscale,
1153 (rce->i_tex_bits + rce->p_tex_bits) * rce->qscale,
1154 rce->mv_bits * rce->qscale,
1155 (double)rce->i_count / rcc->nmb,
1156 (double)rce->p_count / rcc->nmb,
1157 (double)rce->s_count / rcc->nmb,
1158 rce->pict_type == SLICE_TYPE_I,
1159 rce->pict_type == SLICE_TYPE_P,
1160 rce->pict_type == SLICE_TYPE_B,
1161 h->param.rc.f_qcompress,
1162 rcc->i_cplx_sum[SLICE_TYPE_I] / rcc->frame_count[SLICE_TYPE_I],
1163 rcc->i_cplx_sum[SLICE_TYPE_P] / rcc->frame_count[SLICE_TYPE_P],
1164 rcc->p_cplx_sum[SLICE_TYPE_P] / rcc->frame_count[SLICE_TYPE_P],
1165 rcc->p_cplx_sum[SLICE_TYPE_B] / rcc->frame_count[SLICE_TYPE_B],
1166 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / rcc->frame_count[pict_type],
1167 rce->blurred_complexity,
1170 static const char *const_names[]={
1190 static double (*func1[])(void *, double)={
1191 // (void *)bits2qscale,
1192 (void *)qscale2bits,
1195 static const char *func1_names[]={
1201 q = x264_eval((char*)h->param.rc.psz_rc_eq, const_values, const_names, func1, func1_names, NULL, NULL, rce);
1203 // avoid NaN's in the rc_eq
1204 if(!isfinite(q) || rce->i_tex_bits + rce->p_tex_bits + rce->mv_bits == 0)
1205 q = rcc->last_qscale;
1210 rcc->last_qscale = q;
1215 if( zone->b_force_qp )
1216 q = qp2qscale(zone->i_qp);
1218 q /= zone->f_bitrate_factor;
1224 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1226 x264_ratecontrol_t *rcc = h->rc;
1227 const int pict_type = rce->pict_type;
1229 // force I/B quants as a function of P quants
1230 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1231 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1232 if( pict_type == SLICE_TYPE_I )
1235 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1236 double ip_factor = fabs( h->param.rc.f_ip_factor );
1237 /* don't apply ip_factor if the following frame is also I */
1238 if( rcc->accum_p_norm <= 0 )
1240 else if( h->param.rc.f_ip_factor < 0 )
1242 else if( rcc->accum_p_norm >= 1 )
1245 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1247 else if( pict_type == SLICE_TYPE_B )
1249 if( h->param.rc.f_pb_factor > 0 )
1251 if( !rce->kept_as_ref )
1252 q *= fabs( h->param.rc.f_pb_factor );
1254 else if( pict_type == SLICE_TYPE_P
1255 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1256 && rce->i_tex_bits + rce->p_tex_bits == 0 )
1261 /* last qscale / qdiff stuff */
1262 if(rcc->last_non_b_pict_type==pict_type
1263 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1265 double last_q = rcc->last_qscale_for[pict_type];
1266 double max_qscale = last_q * rcc->lstep;
1267 double min_qscale = last_q / rcc->lstep;
1269 if (q > max_qscale) q = max_qscale;
1270 else if(q < min_qscale) q = min_qscale;
1273 rcc->last_qscale_for[pict_type] = q;
1274 if(pict_type!=SLICE_TYPE_B)
1275 rcc->last_non_b_pict_type = pict_type;
1276 if(pict_type==SLICE_TYPE_I)
1278 rcc->last_accum_p_norm = rcc->accum_p_norm;
1279 rcc->accum_p_norm = 0;
1280 rcc->accum_p_qp = 0;
1282 if(pict_type==SLICE_TYPE_P)
1284 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1285 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1286 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1291 static double predict_size( predictor_t *p, double q, double var )
1293 return p->coeff*var / (q*p->count);
1296 static void update_predictor( predictor_t *p, double q, double var, double bits )
1300 p->count *= p->decay;
1301 p->coeff *= p->decay;
1303 p->coeff += bits*q / var;
1306 // update VBV after encoding a frame
1307 static void update_vbv( x264_t *h, int bits )
1309 x264_ratecontrol_t *rcc = h->rc;
1310 x264_ratecontrol_t *rct = h->thread[0]->rc;
1312 if( rcc->last_satd >= h->mb.i_mb_count )
1313 update_predictor( &rct->pred[h->sh.i_type], qp2qscale(rcc->qpa_rc), rcc->last_satd, bits );
1318 rct->buffer_fill_final += rct->buffer_rate - bits;
1319 if( rct->buffer_fill_final < 0 )
1320 x264_log( h, X264_LOG_WARNING, "VBV underflow (%.0f bits)\n", rct->buffer_fill_final );
1321 rct->buffer_fill_final = x264_clip3f( rct->buffer_fill_final, 0, rct->buffer_size );
1324 // provisionally update VBV according to the planned size of all frames currently in progress
1325 static void update_vbv_plan( x264_t *h )
1327 x264_ratecontrol_t *rcc = h->rc;
1328 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final;
1329 if( h->param.i_threads > 1 )
1331 int j = h->rc - h->thread[0]->rc;
1333 for( i=1; i<h->param.i_threads; i++ )
1335 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1336 double bits = t->rc->frame_size_planned;
1337 if( !t->b_thread_active )
1339 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1340 rcc->buffer_fill += rcc->buffer_rate - bits;
1341 rcc->buffer_fill = x264_clip3( rcc->buffer_fill, 0, rcc->buffer_size );
1346 // apply VBV constraints and clip qscale to between lmin and lmax
1347 static double clip_qscale( x264_t *h, int pict_type, double q )
1349 x264_ratecontrol_t *rcc = h->rc;
1350 double lmin = rcc->lmin[pict_type];
1351 double lmax = rcc->lmax[pict_type];
1354 /* B-frames are not directly subject to VBV,
1355 * since they are controlled by the P-frames' QPs.
1356 * FIXME: in 2pass we could modify previous frames' QP too,
1357 * instead of waiting for the buffer to fill */
1359 ( pict_type == SLICE_TYPE_P ||
1360 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) )
1362 if( rcc->buffer_fill/rcc->buffer_size < 0.5 )
1363 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1366 if( rcc->b_vbv && rcc->last_satd > 0 )
1368 /* Now a hard threshold to make sure the frame fits in VBV.
1369 * This one is mostly for I-frames. */
1370 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1372 if( bits > rcc->buffer_fill/2 )
1373 qf = x264_clip3f( rcc->buffer_fill/(2*bits), 0.2, 1.0 );
1376 if( bits < rcc->buffer_rate/2 )
1377 q *= bits*2/rcc->buffer_rate;
1378 q = X264_MAX( q0, q );
1380 /* Check B-frame complexity, and use up any bits that would
1381 * overflow before the next P-frame. */
1382 if( h->sh.i_type == SLICE_TYPE_P )
1384 int nb = rcc->bframes;
1385 double pbbits = bits;
1386 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1389 if( bbits > rcc->buffer_rate )
1391 pbbits += nb * bbits;
1393 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1394 if( pbbits < space )
1396 q *= X264_MAX( pbbits / space,
1397 bits / (0.5 * rcc->buffer_size) );
1399 q = X264_MAX( q0-5, q );
1402 if( !rcc->b_vbv_min_rate )
1403 q = X264_MAX( q0, q );
1408 else if(rcc->b_2pass)
1410 double min2 = log(lmin);
1411 double max2 = log(lmax);
1412 q = (log(q) - min2)/(max2-min2) - 0.5;
1413 q = 1.0/(1.0 + exp(-4*q));
1414 q = q*(max2-min2) + min2;
1418 return x264_clip3f(q, lmin, lmax);
1421 // update qscale for 1 frame based on actual bits used so far
1422 static float rate_estimate_qscale( x264_t *h )
1425 x264_ratecontrol_t *rcc = h->rc;
1426 ratecontrol_entry_t rce;
1427 int pict_type = h->sh.i_type;
1428 double lmin = rcc->lmin[pict_type];
1429 double lmax = rcc->lmax[pict_type];
1430 int64_t total_bits = 8*(h->stat.i_slice_size[SLICE_TYPE_I]
1431 + h->stat.i_slice_size[SLICE_TYPE_P]
1432 + h->stat.i_slice_size[SLICE_TYPE_B]);
1437 if(pict_type != rce.pict_type)
1439 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1440 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1444 if( pict_type == SLICE_TYPE_B )
1446 /* B-frames don't have independent ratecontrol, but rather get the
1447 * average QP of the two adjacent P-frames + an offset */
1449 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1450 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1451 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1452 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1453 float q0 = h->fref0[0]->f_qp_avg_rc;
1454 float q1 = h->fref1[0]->f_qp_avg_rc;
1456 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1457 q0 -= rcc->pb_offset/2;
1458 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1459 q1 -= rcc->pb_offset/2;
1462 q = (q0 + q1) / 2 + rcc->ip_offset;
1468 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1470 if(h->fenc->b_kept_as_ref)
1471 q += rcc->pb_offset/2;
1473 q += rcc->pb_offset;
1475 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1476 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1478 return qp2qscale(q);
1482 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1485 //FIXME adjust abr_buffer based on distance to the end of the video
1486 int64_t diff = total_bits - (int64_t)rce.expected_bits;
1488 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1489 if( h->fenc->i_frame > 30 )
1491 /* Adjust quant based on the difference between
1492 * achieved and expected bitrate so far */
1493 double time = (double)h->fenc->i_frame / rcc->num_entries;
1494 double w = x264_clip3f( time*100, 0.0, 1.0 );
1495 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1499 double expected_size = qscale2bits(&rce, q);
1500 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1501 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
1502 double qmax = q*(2 - expected_fullness);
1503 double size_constraint = 1 + expected_fullness;
1504 if (expected_fullness < .05)
1506 qmax = X264_MIN(qmax, lmax);
1507 while( (expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax) )
1510 expected_size = qscale2bits(&rce, q);
1511 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1513 rcc->last_satd = x264_rc_analyse_slice( h );
1515 q = x264_clip3f( q, lmin, lmax );
1517 else /* 1pass ABR */
1519 /* Calculate the quantizer which would have produced the desired
1520 * average bitrate if it had been applied to all frames so far.
1521 * Then modulate that quant based on the current frame's complexity
1522 * relative to the average complexity so far (using the 2pass RCEQ).
1523 * Then bias the quant up or down if total size so far was far from
1525 * Result: Depending on the value of rate_tolerance, there is a
1526 * tradeoff between quality and bitrate precision. But at large
1527 * tolerances, the bit distribution approaches that of 2pass. */
1529 double wanted_bits, overflow=1, lmin, lmax;
1531 rcc->last_satd = x264_rc_analyse_slice( h );
1532 rcc->short_term_cplxsum *= 0.5;
1533 rcc->short_term_cplxcount *= 0.5;
1534 rcc->short_term_cplxsum += rcc->last_satd;
1535 rcc->short_term_cplxcount ++;
1537 rce.p_tex_bits = rcc->last_satd;
1538 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1541 rce.p_count = rcc->nmb;
1545 rce.pict_type = pict_type;
1547 if( h->param.rc.i_rc_method == X264_RC_CRF )
1549 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1553 int i_frame_done = h->fenc->i_frame + 1 - h->param.i_threads;
1555 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1557 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
1558 wanted_bits = i_frame_done * rcc->bitrate / rcc->fps;
1559 if( wanted_bits > 0 )
1561 abr_buffer *= X264_MAX( 1, sqrt(i_frame_done/25) );
1562 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1567 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1568 /* should test _next_ pict type, but that isn't decided yet */
1569 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1571 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1572 q /= fabs( h->param.rc.f_ip_factor );
1574 else if( h->i_frame > 0 )
1576 /* Asymmetric clipping, because symmetric would prevent
1577 * overflow control in areas of rapidly oscillating complexity */
1578 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1579 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1580 if( overflow > 1.1 && h->i_frame > 3 )
1582 else if( overflow < 0.9 )
1585 q = x264_clip3f(q, lmin, lmax);
1587 else if( h->param.rc.i_rc_method == X264_RC_CRF )
1589 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1592 //FIXME use get_diff_limited_q() ?
1593 q = clip_qscale( h, pict_type, q );
1596 rcc->last_qscale_for[pict_type] =
1597 rcc->last_qscale = q;
1599 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
1600 rcc->last_qscale_for[SLICE_TYPE_P] = q;
1602 if( rcc->b_2pass && rcc->b_vbv)
1603 rcc->frame_size_planned = qscale2bits(&rce, q);
1605 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1606 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1611 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
1615 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
1616 /* these vars are updated in x264_ratecontrol_start()
1617 * so copy them from the context that most recently started (prev)
1618 * to the context that's about to start (cur).
1624 COPY(last_qscale_for);
1625 COPY(last_non_b_pict_type);
1626 COPY(short_term_cplxsum);
1627 COPY(short_term_cplxcount);
1634 #define COPY(var) next->rc->var = cur->rc->var
1635 /* these vars are updated in x264_ratecontrol_end()
1636 * so copy them from the context that most recently ended (cur)
1637 * to the context that's about to end (next)
1640 COPY(expected_bits_sum);
1641 COPY(wanted_bits_window);
1645 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
1646 /* the rest of the variables are either constant or thread-local */
1649 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
1651 /* find an interval ending on an overflow or underflow (depending on whether
1652 * we're adding or removing bits), and starting on the earliest frame that
1653 * can influence the buffer fill of that end frame. */
1654 x264_ratecontrol_t *rcc = h->rc;
1655 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
1656 const double buffer_max = .9 * rcc->buffer_size;
1657 double fill = fills[*t0-1];
1658 double parity = over ? 1. : -1.;
1659 int i, start=-1, end=-1;
1660 for(i = *t0; i < rcc->num_entries; i++)
1662 fill += (rcc->buffer_rate - qscale2bits(&rcc->entry[i], rcc->entry[i].new_qscale)) * parity;
1663 fill = x264_clip3f(fill, 0, rcc->buffer_size);
1665 if(fill <= buffer_min || i == 0)
1671 else if(fill >= buffer_max && start >= 0)
1676 return start>=0 && end>=0;
1679 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
1681 x264_ratecontrol_t *rcc = h->rc;
1682 double qscale_orig, qscale_new;
1687 for(i = t0; i <= t1; i++)
1689 qscale_orig = rcc->entry[i].new_qscale;
1690 qscale_orig = x264_clip3f(qscale_orig, qscale_min, qscale_max);
1691 qscale_new = qscale_orig * adjustment;
1692 qscale_new = x264_clip3f(qscale_new, qscale_min, qscale_max);
1693 rcc->entry[i].new_qscale = qscale_new;
1694 adjusted = adjusted || (qscale_new != qscale_orig);
1699 static double count_expected_bits( x264_t *h )
1701 x264_ratecontrol_t *rcc = h->rc;
1702 double expected_bits = 0;
1704 for(i = 0; i < rcc->num_entries; i++)
1706 ratecontrol_entry_t *rce = &rcc->entry[i];
1707 rce->expected_bits = expected_bits;
1708 expected_bits += qscale2bits(rce, rce->new_qscale);
1710 return expected_bits;
1713 static void vbv_pass2( x264_t *h )
1715 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
1716 * frames in the interval until either buffer is full at some intermediate frame or the
1717 * last frame in the interval no longer underflows. Recompute intervals and repeat.
1718 * Then do the converse to put bits back into overflow areas until target size is met */
1720 x264_ratecontrol_t *rcc = h->rc;
1721 double *fills = x264_malloc((rcc->num_entries+1)*sizeof(double));
1722 double all_available_bits = h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps;
1723 double expected_bits = 0;
1725 double prev_bits = 0;
1727 double qscale_min = qp2qscale(h->param.rc.i_qp_min);
1728 double qscale_max = qp2qscale(h->param.rc.i_qp_max);
1730 int adj_min, adj_max;
1734 /* adjust overall stream size */
1738 prev_bits = expected_bits;
1740 if(expected_bits != 0)
1741 { /* not first iteration */
1742 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
1743 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
1747 while(adj_min && find_underflow(h, fills, &t0, &t1, 1))
1749 adj_min = fix_underflow(h, t0, t1, adjustment, qscale_min, qscale_max);
1754 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
1756 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
1758 while(adj_max && find_underflow(h, fills, &t0, &t1, 0))
1759 adj_max = fix_underflow(h, t0, t1, 1.001, qscale_min, qscale_max);
1761 expected_bits = count_expected_bits(h);
1762 } while(expected_bits < .995 * all_available_bits && expected_bits > prev_bits);
1765 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
1767 /* store expected vbv filling values for tracking when encoding */
1768 for(i = 0; i < rcc->num_entries; i++)
1769 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
1774 static int init_pass2( x264_t *h )
1776 x264_ratecontrol_t *rcc = h->rc;
1777 uint64_t all_const_bits = 0;
1778 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps);
1779 double rate_factor, step, step_mult;
1780 double qblur = h->param.rc.f_qblur;
1781 double cplxblur = h->param.rc.f_complexity_blur;
1782 const int filter_size = (int)(qblur*4) | 1;
1783 double expected_bits;
1784 double *qscale, *blurred_qscale;
1787 /* find total/average complexity & const_bits */
1788 for(i=0; i<rcc->num_entries; i++)
1790 ratecontrol_entry_t *rce = &rcc->entry[i];
1791 all_const_bits += rce->misc_bits;
1792 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
1793 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
1794 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits * rce->qscale;
1795 rcc->frame_count[rce->pict_type] ++;
1798 if( all_available_bits < all_const_bits)
1800 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
1801 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)));
1805 /* Blur complexities, to reduce local fluctuation of QP.
1806 * We don't blur the QPs directly, because then one very simple frame
1807 * could drag down the QP of a nearby complex frame and give it more
1808 * bits than intended. */
1809 for(i=0; i<rcc->num_entries; i++)
1811 ratecontrol_entry_t *rce = &rcc->entry[i];
1812 double weight_sum = 0;
1813 double cplx_sum = 0;
1814 double weight = 1.0;
1815 double gaussian_weight;
1817 /* weighted average of cplx of future frames */
1818 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++)
1820 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
1821 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
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);
1828 /* weighted average of cplx of past frames */
1830 for(j=0; j<=cplxblur*2 && j<=i; j++)
1832 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
1833 gaussian_weight = weight * exp(-j*j/200.0);
1834 weight_sum += gaussian_weight;
1835 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1836 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1840 rce->blurred_complexity = cplx_sum / weight_sum;
1843 qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1845 blurred_qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1847 blurred_qscale = qscale;
1849 /* Search for a factor which, when multiplied by the RCEQ values from
1850 * each frame, adds up to the desired total size.
1851 * There is no exact closed-form solution because of VBV constraints and
1852 * because qscale2bits is not invertible, but we can start with the simple
1853 * approximation of scaling the 1st pass by the ratio of bitrates.
1854 * The search range is probably overkill, but speed doesn't matter here. */
1857 for(i=0; i<rcc->num_entries; i++)
1858 expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i));
1859 step_mult = all_available_bits / expected_bits;
1862 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
1865 rate_factor += step;
1867 rcc->last_non_b_pict_type = -1;
1868 rcc->last_accum_p_norm = 1;
1869 rcc->accum_p_norm = 0;
1872 for(i=0; i<rcc->num_entries; i++)
1874 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
1877 /* fixed I/B qscale relative to P */
1878 for(i=rcc->num_entries-1; i>=0; i--)
1880 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
1881 assert(qscale[i] >= 0);
1887 assert(filter_size%2==1);
1888 for(i=0; i<rcc->num_entries; i++)
1890 ratecontrol_entry_t *rce = &rcc->entry[i];
1892 double q=0.0, sum=0.0;
1894 for(j=0; j<filter_size; j++)
1896 int index = i+j-filter_size/2;
1898 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
1899 if(index < 0 || index >= rcc->num_entries)
1901 if(rce->pict_type != rcc->entry[index].pict_type)
1903 q += qscale[index] * coeff;
1906 blurred_qscale[i] = q/sum;
1910 /* find expected bits */
1911 for(i=0; i<rcc->num_entries; i++)
1913 ratecontrol_entry_t *rce = &rcc->entry[i];
1914 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
1915 assert(rce->new_qscale >= 0);
1916 expected_bits += qscale2bits(rce, rce->new_qscale);
1919 if(expected_bits > all_available_bits) rate_factor -= step;
1924 x264_free(blurred_qscale);
1928 expected_bits = count_expected_bits(h);
1930 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
1933 for(i=0; i<rcc->num_entries; i++)
1934 avgq += rcc->entry[i].new_qscale;
1935 avgq = qscale2qp(avgq / rcc->num_entries);
1937 if ((expected_bits > all_available_bits) || (!rcc->b_vbv))
1938 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
1939 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
1940 (float)h->param.rc.i_bitrate,
1941 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
1943 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
1945 if(h->param.rc.i_qp_min > 0)
1946 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
1948 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
1950 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
1952 if(h->param.rc.i_qp_max < 51)
1953 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
1955 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
1957 else if(!(rcc->b_2pass && rcc->b_vbv))
1958 x264_log(h, X264_LOG_WARNING, "internal error\n");