1 /***************************************************-*- coding: iso-8859-1 -*-
2 * ratecontrol.c: h264 encoder library (Rate Control)
3 *****************************************************************************
4 * Copyright (C) 2005-2008 x264 project
6 * Authors: Loren Merritt <lorenm@u.washington.edu>
7 * Michael Niedermayer <michaelni@gmx.at>
8 * Gabriel Bouvigne <gabriel.bouvigne@joost.com>
9 * Fiona Glaser <fiona@x264.com>
10 * Måns Rullgård <mru@mru.ath.cx>
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License as published by
14 * the Free Software Foundation; either version 2 of the License, or
15 * (at your option) any later version.
17 * This program is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 * GNU General Public License for more details.
22 * You should have received a copy of the GNU General Public License
23 * along with this program; if not, write to the Free Software
24 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
25 *****************************************************************************/
27 #define _ISOC99_SOURCE
28 #undef NDEBUG // always check asserts, the speed effect is far too small to disable them
33 #include "common/common.h"
34 #include "common/cpu.h"
35 #include "ratecontrol.h"
45 uint64_t expected_bits; /*total expected bits up to the current frame (current one excluded)*/
52 float blurred_complexity;
54 } ratecontrol_entry_t;
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, only includes finished frames */
97 double wanted_bits_window; /* target bitrate * window */
99 double short_term_cplxsum;
100 double short_term_cplxcount;
101 double rate_factor_constant;
106 FILE *p_stat_file_out;
107 char *psz_stat_file_tmpname;
109 int num_entries; /* number of ratecontrol_entry_ts */
110 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
112 double last_qscale_for[5]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
113 int last_non_b_pict_type;
114 double accum_p_qp; /* for determining I-frame quant */
116 double last_accum_p_norm;
117 double lmin[5]; /* min qscale by frame type */
119 double lstep; /* max change (multiply) in qscale per frame */
122 double frame_size_estimated;
123 double frame_size_planned;
124 predictor_t *row_pred;
125 predictor_t row_preds[5];
126 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
127 int bframes; /* # consecutive B-frames before this P-frame */
128 int bframe_bits; /* total cost of those frames */
132 x264_zone_t *prev_zone;
136 static int parse_zones( x264_t *h );
137 static int init_pass2(x264_t *);
138 static float rate_estimate_qscale( x264_t *h );
139 static void update_vbv( x264_t *h, int bits );
140 static void update_vbv_plan( x264_t *h );
141 static double predict_size( predictor_t *p, double q, double var );
142 static void update_predictor( predictor_t *p, double q, double var, double bits );
145 * qp = h.264's quantizer
146 * qscale = linearized quantizer = Lagrange multiplier
148 static inline double qp2qscale(double qp)
150 return 0.85 * pow(2.0, ( qp - 12.0 ) / 6.0);
152 static inline double qscale2qp(double qscale)
154 return 12.0 + 6.0 * log(qscale/0.85) / log(2.0);
157 /* Texture bitrate is not quite inversely proportional to qscale,
158 * probably due the the changing number of SKIP blocks.
159 * MV bits level off at about qp<=12, because the lambda used
160 * for motion estimation is constant there. */
161 static inline double qscale2bits(ratecontrol_entry_t *rce, double qscale)
165 return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
166 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
170 // Find the total AC energy of the block in all planes.
171 static NOINLINE int ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
173 /* This function contains annoying hacks because GCC has a habit of reordering emms
174 * and putting it after floating point ops. As a result, we put the emms at the end of the
175 * function and make sure that its always called before the float math. Noinline makes
176 * sure no reordering goes on. */
177 unsigned int var=0, sad, i;
181 int stride = frame->i_stride[i];
182 int offset = h->mb.b_interlaced
183 ? w * (mb_x + (mb_y&~1) * stride) + (mb_y&1) * stride
184 : w * (mb_x + mb_y * stride);
185 int pix = i ? PIXEL_8x8 : PIXEL_16x16;
186 stride <<= h->mb.b_interlaced;
187 var += h->pixf.var[pix]( frame->plane[i]+offset, stride, &sad );
189 var = X264_MAX(var,1);
194 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame )
197 for( mb_y=0; mb_y<h->sps->i_mb_height; mb_y++ )
198 for( mb_x=0; mb_x<h->sps->i_mb_width; mb_x++ )
200 int energy = ac_energy_mb( h, mb_x, mb_y, frame );
201 /* 10 constant chosen to result in approximately the same overall bitrate as without AQ. */
202 float qp_adj = h->param.rc.f_aq_strength * 1.5 * (logf(energy) - 10.0);
203 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
204 if( h->frames.b_have_lowres )
205 frame->i_inv_qscale_factor[mb_x+mb_y*h->mb.i_mb_stride] = FIX8(pow(2.0,-qp_adj/6.0));
209 /*****************************************************************************
210 * x264_adaptive_quant:
211 * adjust macroblock QP based on variance (AC energy) of the MB.
212 * high variance = higher QP
213 * low variance = lower QP
214 * This generally increases SSIM and lowers PSNR.
215 *****************************************************************************/
216 void x264_adaptive_quant( x264_t *h )
221 qp_adj = h->fenc->f_qp_offset[h->mb.i_mb_x + h->mb.i_mb_y*h->mb.i_mb_stride];
222 h->mb.i_qp = x264_clip3( qp + qp_adj + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
223 /* If the QP of this MB is within 1 of the previous MB, code the same QP as the previous MB,
224 * to lower the bit cost of the qp_delta. */
225 if( abs(h->mb.i_qp - h->mb.i_last_qp) == 1 )
226 h->mb.i_qp = h->mb.i_last_qp;
227 h->mb.i_chroma_qp = h->chroma_qp_table[h->mb.i_qp];
230 int x264_ratecontrol_new( x264_t *h )
232 x264_ratecontrol_t *rc;
237 rc = h->rc = x264_malloc( h->param.i_threads * sizeof(x264_ratecontrol_t) );
238 memset( rc, 0, h->param.i_threads * sizeof(x264_ratecontrol_t) );
240 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
241 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
243 /* FIXME: use integers */
244 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
245 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
249 rc->bitrate = h->param.rc.i_bitrate * 1000.;
250 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
251 rc->nmb = h->mb.i_mb_count;
252 rc->last_non_b_pict_type = -1;
255 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
257 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
260 if( h->param.rc.i_vbv_buffer_size )
262 if( h->param.rc.i_rc_method == X264_RC_CQP )
264 x264_log(h, X264_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
265 h->param.rc.i_vbv_max_bitrate = 0;
266 h->param.rc.i_vbv_buffer_size = 0;
268 else if( h->param.rc.i_vbv_max_bitrate == 0 )
270 x264_log( h, X264_LOG_DEBUG, "VBV maxrate unspecified, assuming CBR\n" );
271 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
274 if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate &&
275 h->param.rc.i_vbv_max_bitrate > 0)
276 x264_log(h, X264_LOG_WARNING, "max bitrate less than average bitrate, ignored.\n");
277 else if( h->param.rc.i_vbv_max_bitrate > 0 &&
278 h->param.rc.i_vbv_buffer_size > 0 )
280 if( h->param.rc.i_vbv_buffer_size < 3 * h->param.rc.i_vbv_max_bitrate / rc->fps )
282 h->param.rc.i_vbv_buffer_size = 3 * h->param.rc.i_vbv_max_bitrate / rc->fps;
283 x264_log( h, X264_LOG_WARNING, "VBV buffer size too small, using %d kbit\n",
284 h->param.rc.i_vbv_buffer_size );
286 if( h->param.rc.f_vbv_buffer_init > 1. )
287 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 );
288 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000. / rc->fps;
289 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000.;
290 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
291 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
292 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
294 rc->b_vbv_min_rate = !rc->b_2pass
295 && h->param.rc.i_rc_method == X264_RC_ABR
296 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
298 else if( h->param.rc.i_vbv_max_bitrate )
300 x264_log(h, X264_LOG_WARNING, "VBV maxrate specified, but no bufsize.\n");
301 h->param.rc.i_vbv_max_bitrate = 0;
303 if(rc->rate_tolerance < 0.01)
305 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
306 rc->rate_tolerance = 0.01;
309 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
313 /* FIXME ABR_INIT_QP is actually used only in CRF */
314 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
315 rc->accum_p_norm = .01;
316 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
317 /* estimated ratio that produces a reasonable QP for the first I-frame */
318 rc->cplxr_sum = .01 * pow( 7.0e5, h->param.rc.f_qcompress ) * pow( h->mb.i_mb_count, 0.5 );
319 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
320 rc->last_non_b_pict_type = SLICE_TYPE_I;
323 if( h->param.rc.i_rc_method == X264_RC_CRF )
325 /* arbitrary rescaling to make CRF somewhat similar to QP */
326 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
327 rc->rate_factor_constant = pow( base_cplx, 1 - h->param.rc.f_qcompress )
328 / qp2qscale( h->param.rc.f_rf_constant );
331 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
332 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
333 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
334 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
335 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
337 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
338 rc->last_qscale = qp2qscale(26);
339 rc->pred = x264_malloc( 5*sizeof(predictor_t) );
340 rc->pred_b_from_p = x264_malloc( sizeof(predictor_t) );
341 for( i = 0; i < 5; i++ )
343 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
344 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
345 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
346 rc->pred[i].coeff= 2.0;
347 rc->pred[i].count= 1.0;
348 rc->pred[i].decay= 0.5;
349 rc->row_preds[i].coeff= .25;
350 rc->row_preds[i].count= 1.0;
351 rc->row_preds[i].decay= 0.5;
353 *rc->pred_b_from_p = rc->pred[0];
355 if( parse_zones( h ) < 0 )
357 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
361 /* Load stat file and init 2pass algo */
362 if( h->param.rc.b_stat_read )
364 char *p, *stats_in, *stats_buf;
366 /* read 1st pass stats */
367 assert( h->param.rc.psz_stat_in );
368 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
371 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
375 /* check whether 1st pass options were compatible with current options */
376 if( !strncmp( stats_buf, "#options:", 9 ) )
379 char *opts = stats_buf;
380 stats_in = strchr( stats_buf, '\n' );
386 if( ( p = strstr( opts, "bframes=" ) ) && sscanf( p, "bframes=%d", &i )
387 && h->param.i_bframe != i )
389 x264_log( h, X264_LOG_ERROR, "different number of B-frames than 1st pass (%d vs %d)\n",
390 h->param.i_bframe, i );
394 /* since B-adapt doesn't (yet) take into account B-pyramid,
395 * the converse is not a problem */
396 if( strstr( opts, "b_pyramid=1" ) && !h->param.b_bframe_pyramid )
397 x264_log( h, X264_LOG_WARNING, "1st pass used B-pyramid, 2nd doesn't\n" );
399 if( ( p = strstr( opts, "keyint=" ) ) && sscanf( p, "keyint=%d", &i )
400 && h->param.i_keyint_max != i )
401 x264_log( h, X264_LOG_WARNING, "different keyint than 1st pass (%d vs %d)\n",
402 h->param.i_keyint_max, i );
404 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
405 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
407 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
408 h->param.i_bframe_adaptive = i;
409 else if( h->param.i_bframe )
411 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
415 if( ( p = strstr( opts, "scenecut=" ) ) && sscanf( p, "scenecut=%d", &i ) && i >= -1 && i <= 100 )
417 h->param.i_scenecut_threshold = i;
418 h->param.b_pre_scenecut = !!strstr( p, "(pre)" );
422 x264_log( h, X264_LOG_ERROR, "scenecut method specified in stats file not valid\n" );
427 /* find number of pics */
430 p = strchr(p+1, ';');
433 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
438 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
440 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
441 h->param.i_frame_total, rc->num_entries );
443 if( h->param.i_frame_total > rc->num_entries )
445 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
446 h->param.i_frame_total, rc->num_entries );
450 rc->entry = (ratecontrol_entry_t*) x264_malloc(rc->num_entries * sizeof(ratecontrol_entry_t));
451 memset(rc->entry, 0, rc->num_entries * sizeof(ratecontrol_entry_t));
453 /* init all to skipped p frames */
454 for(i=0; i<rc->num_entries; i++)
456 ratecontrol_entry_t *rce = &rc->entry[i];
457 rce->pict_type = SLICE_TYPE_P;
458 rce->qscale = rce->new_qscale = qp2qscale(20);
459 rce->misc_bits = rc->nmb + 10;
465 for(i=0; i < rc->num_entries; i++)
467 ratecontrol_entry_t *rce;
474 next= strchr(p, ';');
477 (*next)=0; //sscanf is unbelievably slow on long strings
480 e = sscanf(p, " in:%d ", &frame_number);
482 if(frame_number < 0 || frame_number >= rc->num_entries)
484 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
487 rce = &rc->entry[frame_number];
488 rce->direct_mode = 0;
490 e += sscanf(p, " in:%*d out:%*d type:%c q:%f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c",
491 &pict_type, &qp, &rce->tex_bits,
492 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
493 &rce->s_count, &rce->direct_mode);
497 case 'I': rce->kept_as_ref = 1;
498 case 'i': rce->pict_type = SLICE_TYPE_I; break;
499 case 'P': rce->pict_type = SLICE_TYPE_P; break;
500 case 'B': rce->kept_as_ref = 1;
501 case 'b': rce->pict_type = SLICE_TYPE_B; break;
502 default: e = -1; break;
506 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
509 rce->qscale = qp2qscale(qp);
513 x264_free(stats_buf);
515 if(h->param.rc.i_rc_method == X264_RC_ABR)
517 if(init_pass2(h) < 0) return -1;
518 } /* else we're using constant quant, so no need to run the bitrate allocation */
521 /* Open output file */
522 /* If input and output files are the same, output to a temp file
523 * and move it to the real name only when it's complete */
524 if( h->param.rc.b_stat_write )
528 rc->psz_stat_file_tmpname = x264_malloc( strlen(h->param.rc.psz_stat_out) + 6 );
529 strcpy( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
530 strcat( rc->psz_stat_file_tmpname, ".temp" );
532 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
533 if( rc->p_stat_file_out == NULL )
535 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
539 p = x264_param2string( &h->param, 1 );
540 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
544 for( i=0; i<h->param.i_threads; i++ )
546 h->thread[i]->rc = rc+i;
550 memcpy( &h->thread[i]->param, &h->param, sizeof( x264_param_t ) );
551 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
558 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
561 char *tok, UNUSED *saveptr;
563 z->f_bitrate_factor = 1;
564 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
566 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
568 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
572 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
578 z->param = malloc( sizeof(x264_param_t) );
579 memcpy( z->param, &h->param, sizeof(x264_param_t) );
580 while( (tok = strtok_r( p, ",", &saveptr )) )
582 char *val = strchr( tok, '=' );
588 if( x264_param_parse( z->param, tok, val ) )
590 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
598 static int parse_zones( x264_t *h )
600 x264_ratecontrol_t *rc = h->rc;
602 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
604 char *p, *tok, UNUSED *saveptr;
605 char *psz_zones = x264_malloc( strlen(h->param.rc.psz_zones)+1 );
606 strcpy( psz_zones, h->param.rc.psz_zones );
607 h->param.rc.i_zones = 1;
608 for( p = psz_zones; *p; p++ )
609 h->param.rc.i_zones += (*p == '/');
610 h->param.rc.zones = x264_malloc( h->param.rc.i_zones * sizeof(x264_zone_t) );
612 for( i = 0; i < h->param.rc.i_zones; i++ )
614 tok = strtok_r( p, "/", &saveptr );
615 if( !tok || parse_zone( h, &h->param.rc.zones[i], tok ) )
619 x264_free( psz_zones );
622 if( h->param.rc.i_zones > 0 )
624 for( i = 0; i < h->param.rc.i_zones; i++ )
626 x264_zone_t z = h->param.rc.zones[i];
627 if( z.i_start < 0 || z.i_start > z.i_end )
629 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
630 z.i_start, z.i_end );
633 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
635 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
636 z.f_bitrate_factor );
641 rc->i_zones = h->param.rc.i_zones + 1;
642 rc->zones = x264_malloc( rc->i_zones * sizeof(x264_zone_t) );
643 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
645 // default zone to fall back to if none of the others match
646 rc->zones[0].i_start = 0;
647 rc->zones[0].i_end = INT_MAX;
648 rc->zones[0].b_force_qp = 0;
649 rc->zones[0].f_bitrate_factor = 1;
650 rc->zones[0].param = x264_malloc( sizeof(x264_param_t) );
651 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
652 for( i = 1; i < rc->i_zones; i++ )
654 if( !rc->zones[i].param )
655 rc->zones[i].param = rc->zones[0].param;
662 static x264_zone_t *get_zone( x264_t *h, int frame_num )
665 for( i = h->rc->i_zones-1; i >= 0; i-- )
667 x264_zone_t *z = &h->rc->zones[i];
668 if( frame_num >= z->i_start && frame_num <= z->i_end )
674 void x264_ratecontrol_summary( x264_t *h )
676 x264_ratecontrol_t *rc = h->rc;
677 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
679 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
680 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
681 qscale2qp( pow( base_cplx, 1 - h->param.rc.f_qcompress )
682 * rc->cplxr_sum / rc->wanted_bits_window ) );
686 void x264_ratecontrol_delete( x264_t *h )
688 x264_ratecontrol_t *rc = h->rc;
691 if( rc->p_stat_file_out )
693 fclose( rc->p_stat_file_out );
694 if( h->i_frame >= rc->num_entries )
695 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
697 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
698 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
700 x264_free( rc->psz_stat_file_tmpname );
702 x264_free( rc->pred );
703 x264_free( rc->pred_b_from_p );
704 x264_free( rc->entry );
707 x264_free( rc->zones[0].param );
708 if( h->param.rc.psz_zones )
709 for( i=1; i<rc->i_zones; i++ )
710 if( rc->zones[i].param != rc->zones[0].param )
711 x264_free( rc->zones[i].param );
712 x264_free( rc->zones );
717 void x264_ratecontrol_set_estimated_size( x264_t *h, int bits )
719 x264_pthread_mutex_lock( &h->fenc->mutex );
720 h->rc->frame_size_estimated = bits;
721 x264_pthread_mutex_unlock( &h->fenc->mutex );
724 int x264_ratecontrol_get_estimated_size( x264_t const *h)
727 x264_pthread_mutex_lock( &h->fenc->mutex );
728 size = h->rc->frame_size_estimated;
729 x264_pthread_mutex_unlock( &h->fenc->mutex );
733 static void accum_p_qp_update( x264_t *h, float qp )
735 x264_ratecontrol_t *rc = h->rc;
736 rc->accum_p_qp *= .95;
737 rc->accum_p_norm *= .95;
738 rc->accum_p_norm += 1;
739 if( h->sh.i_type == SLICE_TYPE_I )
740 rc->accum_p_qp += qp + rc->ip_offset;
742 rc->accum_p_qp += qp;
745 /* Before encoding a frame, choose a QP for it */
746 void x264_ratecontrol_start( x264_t *h, int i_force_qp )
748 x264_ratecontrol_t *rc = h->rc;
749 ratecontrol_entry_t *rce = NULL;
750 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
755 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
756 x264_encoder_reconfig( h, zone->param );
757 rc->prev_zone = zone;
759 rc->qp_force = i_force_qp;
761 if( h->param.rc.b_stat_read )
763 int frame = h->fenc->i_frame;
764 assert( frame >= 0 && frame < rc->num_entries );
765 rce = h->rc->rce = &h->rc->entry[frame];
767 if( h->sh.i_type == SLICE_TYPE_B
768 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
770 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
771 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
777 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
778 rc->row_pred = &rc->row_preds[h->sh.i_type];
779 update_vbv_plan( h );
782 if( h->sh.i_type != SLICE_TYPE_B )
785 while( h->frames.current[rc->bframes] && IS_X264_TYPE_B(h->frames.current[rc->bframes]->i_type) )
795 q = qscale2qp( rate_estimate_qscale( h ) );
797 else if( rc->b_2pass )
799 rce->new_qscale = rate_estimate_qscale( h );
800 q = qscale2qp( rce->new_qscale );
804 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
805 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
807 q = rc->qp_constant[ h->sh.i_type ];
811 if( zone->b_force_qp )
812 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
814 q -= 6*log(zone->f_bitrate_factor)/log(2);
820 h->fdec->f_qp_avg_rc =
821 h->fdec->f_qp_avg_aq =
823 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
826 rce->new_qp = rc->qp;
828 /* accum_p_qp needs to be here so that future frames can benefit from the
829 * data before this frame is done. but this only works because threading
830 * guarantees to not re-encode any frames. so the non-threaded case does
831 * accum_p_qp later. */
832 if( h->param.i_threads > 1 )
833 accum_p_qp_update( h, rc->qp );
835 if( h->sh.i_type != SLICE_TYPE_B )
836 rc->last_non_b_pict_type = h->sh.i_type;
839 static double predict_row_size( x264_t *h, int y, int qp )
841 /* average between two predictors:
842 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
843 x264_ratecontrol_t *rc = h->rc;
844 double pred_s = predict_size( rc->row_pred, qp2qscale(qp), h->fdec->i_row_satd[y] );
846 if( h->sh.i_type != SLICE_TYPE_I
847 && h->fref0[0]->i_type == h->fdec->i_type
848 && h->fref0[0]->i_row_satd[y] > 0
849 && (abs(h->fref0[0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
851 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
852 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
857 return (pred_s + pred_t) / 2;
860 static double row_bits_so_far( x264_t *h, int y )
864 for( i = 0; i <= y; i++ )
865 bits += h->fdec->i_row_bits[i];
869 static double predict_row_size_sum( x264_t *h, int y, int qp )
872 double bits = row_bits_so_far(h, y);
873 for( i = y+1; i < h->sps->i_mb_height; i++ )
874 bits += predict_row_size( h, i, qp );
879 void x264_ratecontrol_mb( x264_t *h, int bits )
881 x264_ratecontrol_t *rc = h->rc;
882 const int y = h->mb.i_mb_y;
886 h->fdec->i_row_bits[y] += bits;
887 rc->qpa_rc += rc->f_qpm;
888 rc->qpa_aq += h->mb.i_qp;
890 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !rc->b_vbv)
893 h->fdec->i_row_qp[y] = rc->qpm;
895 if( h->sh.i_type == SLICE_TYPE_B )
897 /* B-frames shouldn't use lower QP than their reference frames.
898 * This code is a bit overzealous in limiting B-frame quantizers, but it helps avoid
899 * underflows due to the fact that B-frames are not explicitly covered by VBV. */
900 if( y < h->sps->i_mb_height-1 )
903 int avg_qp = X264_MAX(h->fref0[0]->i_row_qp[y+1], h->fref1[0]->i_row_qp[y+1])
904 + rc->pb_offset * ((h->fenc->i_type == X264_TYPE_BREF) ? 0.5 : 1);
905 rc->qpm = X264_MIN(X264_MAX( rc->qp, avg_qp), 51); //avg_qp could go higher than 51 due to pb_offset
906 i_estimated = row_bits_so_far(h, y); //FIXME: compute full estimated size
907 if (i_estimated > h->rc->frame_size_planned)
908 x264_ratecontrol_set_estimated_size(h, i_estimated);
913 update_predictor( rc->row_pred, qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
915 /* tweak quality based on difference from predicted size */
916 if( y < h->sps->i_mb_height-1 && h->stat.i_slice_count[h->sh.i_type] > 0 )
918 int prev_row_qp = h->fdec->i_row_qp[y];
919 int b0 = predict_row_size_sum( h, y, rc->qpm );
921 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
922 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
923 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
927 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
928 /* area at the top of the frame was measured inaccurately. */
929 if(row_bits_so_far(h,y) < 0.05 * rc->frame_size_planned)
932 headroom = buffer_left_planned/rc->buffer_size;
933 if(h->sh.i_type != SLICE_TYPE_I)
937 if( !rc->b_vbv_min_rate )
938 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
940 while( rc->qpm < i_qp_max
941 && (b1 > rc->frame_size_planned * rc_tol
942 || (rc->buffer_fill - b1 < buffer_left_planned * 0.5)))
945 b1 = predict_row_size_sum( h, y, rc->qpm );
948 /* avoid VBV underflow */
949 while( (rc->qpm < h->param.rc.i_qp_max)
950 && (rc->buffer_fill - b1 < rc->buffer_size * 0.005))
953 b1 = predict_row_size_sum( h, y, rc->qpm );
956 while( rc->qpm > i_qp_min
957 && rc->qpm > h->fdec->i_row_qp[0]
958 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
959 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
962 b1 = predict_row_size_sum( h, y, rc->qpm );
964 x264_ratecontrol_set_estimated_size(h, b1);
967 /* loses the fractional part of the frame-wise qp */
971 int x264_ratecontrol_qp( x264_t *h )
976 /* In 2pass, force the same frame types as in the 1st pass */
977 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
979 x264_ratecontrol_t *rc = h->rc;
980 if( h->param.rc.b_stat_read )
982 if( frame_num >= rc->num_entries )
984 /* We could try to initialize everything required for ABR and
985 * adaptive B-frames, but that would be complicated.
986 * So just calculate the average QP used so far. */
989 h->param.rc.i_qp_constant = (h->stat.i_slice_count[SLICE_TYPE_P] == 0) ? 24
990 : 1 + h->stat.f_slice_qp[SLICE_TYPE_P] / h->stat.i_slice_count[SLICE_TYPE_P];
991 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
992 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 );
993 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 );
995 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
996 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
997 if( h->param.i_bframe_adaptive )
998 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1000 for( i = 0; i < h->param.i_threads; i++ )
1002 h->thread[i]->rc->b_abr = 0;
1003 h->thread[i]->rc->b_2pass = 0;
1004 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1005 h->thread[i]->param.rc.b_stat_read = 0;
1006 h->thread[i]->param.i_bframe_adaptive = 0;
1007 h->thread[i]->param.b_pre_scenecut = 0;
1008 h->thread[i]->param.i_scenecut_threshold = -1;
1009 if( h->thread[i]->param.i_bframe > 1 )
1010 h->thread[i]->param.i_bframe = 1;
1012 return X264_TYPE_AUTO;
1014 switch( rc->entry[frame_num].pict_type )
1017 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I;
1020 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B;
1029 return X264_TYPE_AUTO;
1033 /* After encoding one frame, save stats and update ratecontrol state */
1034 void x264_ratecontrol_end( x264_t *h, int bits )
1036 x264_ratecontrol_t *rc = h->rc;
1037 const int *mbs = h->stat.frame.i_mb_count;
1042 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1043 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1044 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1045 for( i = B_DIRECT; i < B_8x8; i++ )
1046 h->stat.frame.i_mb_count_p += mbs[i];
1048 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1049 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1051 if( h->param.rc.b_stat_write )
1053 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1054 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1055 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1056 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1057 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1058 char c_direct = h->mb.b_direct_auto_write ?
1059 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1060 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1062 fprintf( rc->p_stat_file_out,
1063 "in:%d out:%d type:%c q:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c;\n",
1064 h->fenc->i_frame, h->i_frame,
1066 h->stat.frame.i_tex_bits,
1067 h->stat.frame.i_mv_bits,
1068 h->stat.frame.i_misc_bits,
1069 h->stat.frame.i_mb_count_i,
1070 h->stat.frame.i_mb_count_p,
1071 h->stat.frame.i_mb_count_skip,
1077 if( h->sh.i_type != SLICE_TYPE_B )
1078 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / rc->last_rceq;
1081 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1082 * Not perfectly accurate with B-refs, but good enough. */
1083 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
1085 rc->cplxr_sum *= rc->cbr_decay;
1086 rc->wanted_bits_window += rc->bitrate / rc->fps;
1087 rc->wanted_bits_window *= rc->cbr_decay;
1089 if( h->param.i_threads == 1 )
1090 accum_p_qp_update( h, rc->qpa_rc );
1095 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
1098 if( h->mb.b_variable_qp )
1100 if( h->sh.i_type == SLICE_TYPE_B )
1102 rc->bframe_bits += bits;
1103 if( !h->frames.current[0] || !IS_X264_TYPE_B(h->frames.current[0]->i_type) )
1105 update_predictor( rc->pred_b_from_p, qp2qscale(rc->qpa_rc),
1106 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1107 rc->bframe_bits = 0;
1112 update_vbv( h, bits );
1115 /****************************************************************************
1117 ***************************************************************************/
1120 * modify the bitrate curve from pass1 for one frame
1122 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1124 x264_ratecontrol_t *rcc= h->rc;
1126 x264_zone_t *zone = get_zone( h, frame_num );
1128 q = pow( rce->blurred_complexity, 1 - h->param.rc.f_qcompress );
1130 // avoid NaN's in the rc_eq
1131 if(!isfinite(q) || rce->tex_bits + rce->mv_bits == 0)
1132 q = rcc->last_qscale;
1137 rcc->last_qscale = q;
1142 if( zone->b_force_qp )
1143 q = qp2qscale(zone->i_qp);
1145 q /= zone->f_bitrate_factor;
1151 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1153 x264_ratecontrol_t *rcc = h->rc;
1154 const int pict_type = rce->pict_type;
1156 // force I/B quants as a function of P quants
1157 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1158 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1159 if( pict_type == SLICE_TYPE_I )
1162 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1163 double ip_factor = fabs( h->param.rc.f_ip_factor );
1164 /* don't apply ip_factor if the following frame is also I */
1165 if( rcc->accum_p_norm <= 0 )
1167 else if( h->param.rc.f_ip_factor < 0 )
1169 else if( rcc->accum_p_norm >= 1 )
1172 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1174 else if( pict_type == SLICE_TYPE_B )
1176 if( h->param.rc.f_pb_factor > 0 )
1178 if( !rce->kept_as_ref )
1179 q *= fabs( h->param.rc.f_pb_factor );
1181 else if( pict_type == SLICE_TYPE_P
1182 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1183 && rce->tex_bits == 0 )
1188 /* last qscale / qdiff stuff */
1189 if(rcc->last_non_b_pict_type==pict_type
1190 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1192 double last_q = rcc->last_qscale_for[pict_type];
1193 double max_qscale = last_q * rcc->lstep;
1194 double min_qscale = last_q / rcc->lstep;
1196 if (q > max_qscale) q = max_qscale;
1197 else if(q < min_qscale) q = min_qscale;
1200 rcc->last_qscale_for[pict_type] = q;
1201 if(pict_type!=SLICE_TYPE_B)
1202 rcc->last_non_b_pict_type = pict_type;
1203 if(pict_type==SLICE_TYPE_I)
1205 rcc->last_accum_p_norm = rcc->accum_p_norm;
1206 rcc->accum_p_norm = 0;
1207 rcc->accum_p_qp = 0;
1209 if(pict_type==SLICE_TYPE_P)
1211 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1212 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1213 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1218 static double predict_size( predictor_t *p, double q, double var )
1220 return p->coeff*var / (q*p->count);
1223 static void update_predictor( predictor_t *p, double q, double var, double bits )
1227 p->count *= p->decay;
1228 p->coeff *= p->decay;
1230 p->coeff += bits*q / var;
1233 // update VBV after encoding a frame
1234 static void update_vbv( x264_t *h, int bits )
1236 x264_ratecontrol_t *rcc = h->rc;
1237 x264_ratecontrol_t *rct = h->thread[0]->rc;
1239 if( rcc->last_satd >= h->mb.i_mb_count )
1240 update_predictor( &rct->pred[h->sh.i_type], qp2qscale(rcc->qpa_rc), rcc->last_satd, bits );
1245 rct->buffer_fill_final += rct->buffer_rate - bits;
1246 if( rct->buffer_fill_final < 0 )
1247 x264_log( h, X264_LOG_WARNING, "VBV underflow (%.0f bits)\n", rct->buffer_fill_final );
1248 rct->buffer_fill_final = x264_clip3f( rct->buffer_fill_final, 0, rct->buffer_size );
1251 // provisionally update VBV according to the planned size of all frames currently in progress
1252 static void update_vbv_plan( x264_t *h )
1254 x264_ratecontrol_t *rcc = h->rc;
1255 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final;
1256 if( h->param.i_threads > 1 )
1258 int j = h->rc - h->thread[0]->rc;
1260 for( i=1; i<h->param.i_threads; i++ )
1262 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1263 double bits = t->rc->frame_size_planned;
1264 if( !t->b_thread_active )
1266 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1267 rcc->buffer_fill += rcc->buffer_rate - bits;
1268 rcc->buffer_fill = x264_clip3( rcc->buffer_fill, 0, rcc->buffer_size );
1273 // apply VBV constraints and clip qscale to between lmin and lmax
1274 static double clip_qscale( x264_t *h, int pict_type, double q )
1276 x264_ratecontrol_t *rcc = h->rc;
1277 double lmin = rcc->lmin[pict_type];
1278 double lmax = rcc->lmax[pict_type];
1281 /* B-frames are not directly subject to VBV,
1282 * since they are controlled by the P-frames' QPs.
1283 * FIXME: in 2pass we could modify previous frames' QP too,
1284 * instead of waiting for the buffer to fill */
1286 ( pict_type == SLICE_TYPE_P ||
1287 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) )
1289 if( rcc->buffer_fill/rcc->buffer_size < 0.5 )
1290 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1293 if( rcc->b_vbv && rcc->last_satd > 0 )
1295 /* Now a hard threshold to make sure the frame fits in VBV.
1296 * This one is mostly for I-frames. */
1297 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1299 if( bits > rcc->buffer_fill/2 )
1300 qf = x264_clip3f( rcc->buffer_fill/(2*bits), 0.2, 1.0 );
1303 if( bits < rcc->buffer_rate/2 )
1304 q *= bits*2/rcc->buffer_rate;
1305 q = X264_MAX( q0, q );
1307 /* Check B-frame complexity, and use up any bits that would
1308 * overflow before the next P-frame. */
1309 if( h->sh.i_type == SLICE_TYPE_P )
1311 int nb = rcc->bframes;
1312 double pbbits = bits;
1313 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1316 if( bbits > rcc->buffer_rate )
1318 pbbits += nb * bbits;
1320 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1321 if( pbbits < space )
1323 q *= X264_MAX( pbbits / space,
1324 bits / (0.5 * rcc->buffer_size) );
1326 q = X264_MAX( q0-5, q );
1329 if( !rcc->b_vbv_min_rate )
1330 q = X264_MAX( q0, q );
1335 else if(rcc->b_2pass)
1337 double min2 = log(lmin);
1338 double max2 = log(lmax);
1339 q = (log(q) - min2)/(max2-min2) - 0.5;
1340 q = 1.0/(1.0 + exp(-4*q));
1341 q = q*(max2-min2) + min2;
1345 return x264_clip3f(q, lmin, lmax);
1348 // update qscale for 1 frame based on actual bits used so far
1349 static float rate_estimate_qscale( x264_t *h )
1352 x264_ratecontrol_t *rcc = h->rc;
1353 ratecontrol_entry_t rce;
1354 int pict_type = h->sh.i_type;
1355 double lmin = rcc->lmin[pict_type];
1356 double lmax = rcc->lmax[pict_type];
1357 int64_t total_bits = 8*(h->stat.i_slice_size[SLICE_TYPE_I]
1358 + h->stat.i_slice_size[SLICE_TYPE_P]
1359 + h->stat.i_slice_size[SLICE_TYPE_B]);
1364 if(pict_type != rce.pict_type)
1366 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1367 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1371 if( pict_type == SLICE_TYPE_B )
1373 /* B-frames don't have independent ratecontrol, but rather get the
1374 * average QP of the two adjacent P-frames + an offset */
1376 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1377 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1378 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1379 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1380 float q0 = h->fref0[0]->f_qp_avg_rc;
1381 float q1 = h->fref1[0]->f_qp_avg_rc;
1383 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1384 q0 -= rcc->pb_offset/2;
1385 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1386 q1 -= rcc->pb_offset/2;
1389 q = (q0 + q1) / 2 + rcc->ip_offset;
1395 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1397 if(h->fenc->b_kept_as_ref)
1398 q += rcc->pb_offset/2;
1400 q += rcc->pb_offset;
1402 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1403 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1405 return qp2qscale(q);
1409 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1413 //FIXME adjust abr_buffer based on distance to the end of the video
1415 int64_t predicted_bits = total_bits;
1419 if( h->param.i_threads > 1 )
1421 int j = h->rc - h->thread[0]->rc;
1423 for( i=1; i<h->param.i_threads; i++ )
1425 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1426 double bits = t->rc->frame_size_planned;
1427 if( !t->b_thread_active )
1429 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1430 predicted_bits += (int64_t)bits;
1436 if( h->fenc->i_frame < h->param.i_threads )
1437 predicted_bits += (int64_t)h->fenc->i_frame * rcc->bitrate / rcc->fps;
1439 predicted_bits += (int64_t)(h->param.i_threads - 1) * rcc->bitrate / rcc->fps;
1442 diff = predicted_bits - (int64_t)rce.expected_bits;
1444 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1445 if( ((h->fenc->i_frame + 1 - h->param.i_threads) >= rcc->fps) &&
1446 (rcc->expected_bits_sum > 0))
1448 /* Adjust quant based on the difference between
1449 * achieved and expected bitrate so far */
1450 double time = (double)h->fenc->i_frame / rcc->num_entries;
1451 double w = x264_clip3f( time*100, 0.0, 1.0 );
1452 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1456 /* Do not overflow vbv */
1457 double expected_size = qscale2bits(&rce, q);
1458 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1459 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
1460 double qmax = q*(2 - expected_fullness);
1461 double size_constraint = 1 + expected_fullness;
1462 qmax = X264_MAX(qmax, rce.new_qscale);
1463 if (expected_fullness < .05)
1465 qmax = X264_MIN(qmax, lmax);
1466 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
1467 ((expected_vbv < 0) && (q < lmax)))
1470 expected_size = qscale2bits(&rce, q);
1471 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1473 rcc->last_satd = x264_rc_analyse_slice( h );
1475 q = x264_clip3f( q, lmin, lmax );
1477 else /* 1pass ABR */
1479 /* Calculate the quantizer which would have produced the desired
1480 * average bitrate if it had been applied to all frames so far.
1481 * Then modulate that quant based on the current frame's complexity
1482 * relative to the average complexity so far (using the 2pass RCEQ).
1483 * Then bias the quant up or down if total size so far was far from
1485 * Result: Depending on the value of rate_tolerance, there is a
1486 * tradeoff between quality and bitrate precision. But at large
1487 * tolerances, the bit distribution approaches that of 2pass. */
1489 double wanted_bits, overflow=1, lmin, lmax;
1491 rcc->last_satd = x264_rc_analyse_slice( h );
1492 rcc->short_term_cplxsum *= 0.5;
1493 rcc->short_term_cplxcount *= 0.5;
1494 rcc->short_term_cplxsum += rcc->last_satd;
1495 rcc->short_term_cplxcount ++;
1497 rce.tex_bits = rcc->last_satd;
1498 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1500 rce.p_count = rcc->nmb;
1504 rce.pict_type = pict_type;
1506 if( h->param.rc.i_rc_method == X264_RC_CRF )
1508 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1512 int i_frame_done = h->fenc->i_frame + 1 - h->param.i_threads;
1514 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1516 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
1517 wanted_bits = i_frame_done * rcc->bitrate / rcc->fps;
1518 if( wanted_bits > 0 )
1520 abr_buffer *= X264_MAX( 1, sqrt(i_frame_done/25) );
1521 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1526 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1527 /* should test _next_ pict type, but that isn't decided yet */
1528 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1530 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1531 q /= fabs( h->param.rc.f_ip_factor );
1533 else if( h->i_frame > 0 )
1535 /* Asymmetric clipping, because symmetric would prevent
1536 * overflow control in areas of rapidly oscillating complexity */
1537 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1538 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1539 if( overflow > 1.1 && h->i_frame > 3 )
1541 else if( overflow < 0.9 )
1544 q = x264_clip3f(q, lmin, lmax);
1546 else if( h->param.rc.i_rc_method == X264_RC_CRF )
1548 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1551 //FIXME use get_diff_limited_q() ?
1552 q = clip_qscale( h, pict_type, q );
1555 rcc->last_qscale_for[pict_type] =
1556 rcc->last_qscale = q;
1558 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
1559 rcc->last_qscale_for[SLICE_TYPE_P] = q;
1561 if( rcc->b_2pass && rcc->b_vbv)
1562 rcc->frame_size_planned = qscale2bits(&rce, q);
1564 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1565 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1570 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
1574 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
1575 /* these vars are updated in x264_ratecontrol_start()
1576 * so copy them from the context that most recently started (prev)
1577 * to the context that's about to start (cur).
1583 COPY(last_qscale_for);
1584 COPY(last_non_b_pict_type);
1585 COPY(short_term_cplxsum);
1586 COPY(short_term_cplxcount);
1593 #define COPY(var) next->rc->var = cur->rc->var
1594 /* these vars are updated in x264_ratecontrol_end()
1595 * so copy them from the context that most recently ended (cur)
1596 * to the context that's about to end (next)
1599 COPY(expected_bits_sum);
1600 COPY(wanted_bits_window);
1604 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
1605 /* the rest of the variables are either constant or thread-local */
1608 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
1610 /* find an interval ending on an overflow or underflow (depending on whether
1611 * we're adding or removing bits), and starting on the earliest frame that
1612 * can influence the buffer fill of that end frame. */
1613 x264_ratecontrol_t *rcc = h->rc;
1614 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
1615 const double buffer_max = .9 * rcc->buffer_size;
1616 double fill = fills[*t0-1];
1617 double parity = over ? 1. : -1.;
1618 int i, start=-1, end=-1;
1619 for(i = *t0; i < rcc->num_entries; i++)
1621 fill += (rcc->buffer_rate - qscale2bits(&rcc->entry[i], rcc->entry[i].new_qscale)) * parity;
1622 fill = x264_clip3f(fill, 0, rcc->buffer_size);
1624 if(fill <= buffer_min || i == 0)
1630 else if(fill >= buffer_max && start >= 0)
1635 return start>=0 && end>=0;
1638 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
1640 x264_ratecontrol_t *rcc = h->rc;
1641 double qscale_orig, qscale_new;
1646 for(i = t0; i <= t1; i++)
1648 qscale_orig = rcc->entry[i].new_qscale;
1649 qscale_orig = x264_clip3f(qscale_orig, qscale_min, qscale_max);
1650 qscale_new = qscale_orig * adjustment;
1651 qscale_new = x264_clip3f(qscale_new, qscale_min, qscale_max);
1652 rcc->entry[i].new_qscale = qscale_new;
1653 adjusted = adjusted || (qscale_new != qscale_orig);
1658 static double count_expected_bits( x264_t *h )
1660 x264_ratecontrol_t *rcc = h->rc;
1661 double expected_bits = 0;
1663 for(i = 0; i < rcc->num_entries; i++)
1665 ratecontrol_entry_t *rce = &rcc->entry[i];
1666 rce->expected_bits = expected_bits;
1667 expected_bits += qscale2bits(rce, rce->new_qscale);
1669 return expected_bits;
1672 static void vbv_pass2( x264_t *h )
1674 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
1675 * frames in the interval until either buffer is full at some intermediate frame or the
1676 * last frame in the interval no longer underflows. Recompute intervals and repeat.
1677 * Then do the converse to put bits back into overflow areas until target size is met */
1679 x264_ratecontrol_t *rcc = h->rc;
1680 double *fills = x264_malloc((rcc->num_entries+1)*sizeof(double));
1681 double all_available_bits = h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps;
1682 double expected_bits = 0;
1684 double prev_bits = 0;
1686 double qscale_min = qp2qscale(h->param.rc.i_qp_min);
1687 double qscale_max = qp2qscale(h->param.rc.i_qp_max);
1689 int adj_min, adj_max;
1693 /* adjust overall stream size */
1697 prev_bits = expected_bits;
1699 if(expected_bits != 0)
1700 { /* not first iteration */
1701 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
1702 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
1706 while(adj_min && find_underflow(h, fills, &t0, &t1, 1))
1708 adj_min = fix_underflow(h, t0, t1, adjustment, qscale_min, qscale_max);
1713 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
1715 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
1717 while(adj_max && find_underflow(h, fills, &t0, &t1, 0))
1718 adj_max = fix_underflow(h, t0, t1, 1.001, qscale_min, qscale_max);
1720 expected_bits = count_expected_bits(h);
1721 } while((expected_bits < .995*all_available_bits) && ((int)(expected_bits+.5) > (int)(prev_bits+.5)) );
1724 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
1726 /* store expected vbv filling values for tracking when encoding */
1727 for(i = 0; i < rcc->num_entries; i++)
1728 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
1733 static int init_pass2( x264_t *h )
1735 x264_ratecontrol_t *rcc = h->rc;
1736 uint64_t all_const_bits = 0;
1737 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps);
1738 double rate_factor, step, step_mult;
1739 double qblur = h->param.rc.f_qblur;
1740 double cplxblur = h->param.rc.f_complexity_blur;
1741 const int filter_size = (int)(qblur*4) | 1;
1742 double expected_bits;
1743 double *qscale, *blurred_qscale;
1746 /* find total/average complexity & const_bits */
1747 for(i=0; i<rcc->num_entries; i++)
1749 ratecontrol_entry_t *rce = &rcc->entry[i];
1750 all_const_bits += rce->misc_bits;
1753 if( all_available_bits < all_const_bits)
1755 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
1756 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)));
1760 /* Blur complexities, to reduce local fluctuation of QP.
1761 * We don't blur the QPs directly, because then one very simple frame
1762 * could drag down the QP of a nearby complex frame and give it more
1763 * bits than intended. */
1764 for(i=0; i<rcc->num_entries; i++)
1766 ratecontrol_entry_t *rce = &rcc->entry[i];
1767 double weight_sum = 0;
1768 double cplx_sum = 0;
1769 double weight = 1.0;
1770 double gaussian_weight;
1772 /* weighted average of cplx of future frames */
1773 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++)
1775 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
1776 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1779 gaussian_weight = weight * exp(-j*j/200.0);
1780 weight_sum += gaussian_weight;
1781 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1783 /* weighted average of cplx of past frames */
1785 for(j=0; j<=cplxblur*2 && j<=i; j++)
1787 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
1788 gaussian_weight = weight * exp(-j*j/200.0);
1789 weight_sum += gaussian_weight;
1790 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1791 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1795 rce->blurred_complexity = cplx_sum / weight_sum;
1798 qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1800 blurred_qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1802 blurred_qscale = qscale;
1804 /* Search for a factor which, when multiplied by the RCEQ values from
1805 * each frame, adds up to the desired total size.
1806 * There is no exact closed-form solution because of VBV constraints and
1807 * because qscale2bits is not invertible, but we can start with the simple
1808 * approximation of scaling the 1st pass by the ratio of bitrates.
1809 * The search range is probably overkill, but speed doesn't matter here. */
1812 for(i=0; i<rcc->num_entries; i++)
1813 expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i));
1814 step_mult = all_available_bits / expected_bits;
1817 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
1820 rate_factor += step;
1822 rcc->last_non_b_pict_type = -1;
1823 rcc->last_accum_p_norm = 1;
1824 rcc->accum_p_norm = 0;
1827 for(i=0; i<rcc->num_entries; i++)
1829 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
1832 /* fixed I/B qscale relative to P */
1833 for(i=rcc->num_entries-1; i>=0; i--)
1835 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
1836 assert(qscale[i] >= 0);
1842 assert(filter_size%2==1);
1843 for(i=0; i<rcc->num_entries; i++)
1845 ratecontrol_entry_t *rce = &rcc->entry[i];
1847 double q=0.0, sum=0.0;
1849 for(j=0; j<filter_size; j++)
1851 int index = i+j-filter_size/2;
1853 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
1854 if(index < 0 || index >= rcc->num_entries)
1856 if(rce->pict_type != rcc->entry[index].pict_type)
1858 q += qscale[index] * coeff;
1861 blurred_qscale[i] = q/sum;
1865 /* find expected bits */
1866 for(i=0; i<rcc->num_entries; i++)
1868 ratecontrol_entry_t *rce = &rcc->entry[i];
1869 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
1870 assert(rce->new_qscale >= 0);
1871 expected_bits += qscale2bits(rce, rce->new_qscale);
1874 if(expected_bits > all_available_bits) rate_factor -= step;
1879 x264_free(blurred_qscale);
1883 expected_bits = count_expected_bits(h);
1885 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
1888 for(i=0; i<rcc->num_entries; i++)
1889 avgq += rcc->entry[i].new_qscale;
1890 avgq = qscale2qp(avgq / rcc->num_entries);
1892 if ((expected_bits > all_available_bits) || (!rcc->b_vbv))
1893 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
1894 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
1895 (float)h->param.rc.i_bitrate,
1896 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
1898 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
1900 if(h->param.rc.i_qp_min > 0)
1901 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
1903 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
1905 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
1907 if(h->param.rc.i_qp_max < 51)
1908 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
1910 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
1912 else if(!(rcc->b_2pass && rcc->b_vbv))
1913 x264_log(h, X264_LOG_WARNING, "internal error\n");