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;
207 /*****************************************************************************
208 * x264_adaptive_quant:
209 * adjust macroblock QP based on variance (AC energy) of the MB.
210 * high variance = higher QP
211 * low variance = lower QP
212 * This generally increases SSIM and lowers PSNR.
213 *****************************************************************************/
214 void x264_adaptive_quant( x264_t *h )
219 qp_adj = h->fenc->f_qp_offset[h->mb.i_mb_x + h->mb.i_mb_y*h->mb.i_mb_stride];
220 h->mb.i_qp = x264_clip3( qp + qp_adj + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
221 /* If the QP of this MB is within 1 of the previous MB, code the same QP as the previous MB,
222 * to lower the bit cost of the qp_delta. */
223 if( abs(h->mb.i_qp - h->mb.i_last_qp) == 1 )
224 h->mb.i_qp = h->mb.i_last_qp;
225 h->mb.i_chroma_qp = h->chroma_qp_table[h->mb.i_qp];
228 int x264_ratecontrol_new( x264_t *h )
230 x264_ratecontrol_t *rc;
235 rc = h->rc = x264_malloc( h->param.i_threads * sizeof(x264_ratecontrol_t) );
236 memset( rc, 0, h->param.i_threads * sizeof(x264_ratecontrol_t) );
238 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
239 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
241 /* FIXME: use integers */
242 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
243 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
247 rc->bitrate = h->param.rc.i_bitrate * 1000.;
248 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
249 rc->nmb = h->mb.i_mb_count;
250 rc->last_non_b_pict_type = -1;
253 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
255 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
258 if( h->param.rc.i_vbv_buffer_size )
260 if( h->param.rc.i_rc_method == X264_RC_CQP )
261 x264_log(h, X264_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
262 else if( h->param.rc.i_vbv_max_bitrate == 0 )
264 x264_log( h, X264_LOG_DEBUG, "VBV maxrate unspecified, assuming CBR\n" );
265 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
268 if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate &&
269 h->param.rc.i_vbv_max_bitrate > 0)
270 x264_log(h, X264_LOG_WARNING, "max bitrate less than average bitrate, ignored.\n");
271 else if( h->param.rc.i_vbv_max_bitrate > 0 &&
272 h->param.rc.i_vbv_buffer_size > 0 )
274 if( h->param.rc.i_vbv_buffer_size < 3 * h->param.rc.i_vbv_max_bitrate / rc->fps )
276 h->param.rc.i_vbv_buffer_size = 3 * h->param.rc.i_vbv_max_bitrate / rc->fps;
277 x264_log( h, X264_LOG_WARNING, "VBV buffer size too small, using %d kbit\n",
278 h->param.rc.i_vbv_buffer_size );
280 if( h->param.rc.f_vbv_buffer_init > 1. )
281 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 );
282 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000. / rc->fps;
283 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000.;
284 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
285 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
286 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
288 rc->b_vbv_min_rate = !rc->b_2pass
289 && h->param.rc.i_rc_method == X264_RC_ABR
290 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
292 else if( h->param.rc.i_vbv_max_bitrate )
294 x264_log(h, X264_LOG_WARNING, "VBV maxrate specified, but no bufsize.\n");
295 h->param.rc.i_vbv_max_bitrate = 0;
297 if(rc->rate_tolerance < 0.01)
299 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
300 rc->rate_tolerance = 0.01;
303 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
307 /* FIXME ABR_INIT_QP is actually used only in CRF */
308 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
309 rc->accum_p_norm = .01;
310 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
311 /* estimated ratio that produces a reasonable QP for the first I-frame */
312 rc->cplxr_sum = .01 * pow( 7.0e5, h->param.rc.f_qcompress ) * pow( h->mb.i_mb_count, 0.5 );
313 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
314 rc->last_non_b_pict_type = SLICE_TYPE_I;
317 if( h->param.rc.i_rc_method == X264_RC_CRF )
319 /* arbitrary rescaling to make CRF somewhat similar to QP */
320 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
321 rc->rate_factor_constant = pow( base_cplx, 1 - h->param.rc.f_qcompress )
322 / qp2qscale( h->param.rc.f_rf_constant );
325 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
326 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
327 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
328 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
329 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
331 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
332 rc->last_qscale = qp2qscale(26);
333 rc->pred = x264_malloc( 5*sizeof(predictor_t) );
334 rc->pred_b_from_p = x264_malloc( sizeof(predictor_t) );
335 for( i = 0; i < 5; i++ )
337 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
338 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
339 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
340 rc->pred[i].coeff= 2.0;
341 rc->pred[i].count= 1.0;
342 rc->pred[i].decay= 0.5;
343 rc->row_preds[i].coeff= .25;
344 rc->row_preds[i].count= 1.0;
345 rc->row_preds[i].decay= 0.5;
347 *rc->pred_b_from_p = rc->pred[0];
349 if( parse_zones( h ) < 0 )
351 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
355 /* Load stat file and init 2pass algo */
356 if( h->param.rc.b_stat_read )
358 char *p, *stats_in, *stats_buf;
360 /* read 1st pass stats */
361 assert( h->param.rc.psz_stat_in );
362 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
365 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
369 /* check whether 1st pass options were compatible with current options */
370 if( !strncmp( stats_buf, "#options:", 9 ) )
373 char *opts = stats_buf;
374 stats_in = strchr( stats_buf, '\n' );
380 if( ( p = strstr( opts, "bframes=" ) ) && sscanf( p, "bframes=%d", &i )
381 && h->param.i_bframe != i )
383 x264_log( h, X264_LOG_ERROR, "different number of B-frames than 1st pass (%d vs %d)\n",
384 h->param.i_bframe, i );
388 /* since B-adapt doesn't (yet) take into account B-pyramid,
389 * the converse is not a problem */
390 if( strstr( opts, "b_pyramid=1" ) && !h->param.b_bframe_pyramid )
391 x264_log( h, X264_LOG_WARNING, "1st pass used B-pyramid, 2nd doesn't\n" );
393 if( ( p = strstr( opts, "keyint=" ) ) && sscanf( p, "keyint=%d", &i )
394 && h->param.i_keyint_max != i )
395 x264_log( h, X264_LOG_WARNING, "different keyint than 1st pass (%d vs %d)\n",
396 h->param.i_keyint_max, i );
398 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
399 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
401 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
402 h->param.i_bframe_adaptive = i;
403 else if( h->param.i_bframe )
405 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
409 if( ( p = strstr( opts, "scenecut=" ) ) && sscanf( p, "scenecut=%d", &i ) && i >= -1 && i <= 100 )
411 h->param.i_scenecut_threshold = i;
412 h->param.b_pre_scenecut = !!strstr( p, "(pre)" );
416 x264_log( h, X264_LOG_ERROR, "scenecut method specified in stats file not valid\n" );
421 /* find number of pics */
424 p = strchr(p+1, ';');
427 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
432 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
434 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
435 h->param.i_frame_total, rc->num_entries );
437 if( h->param.i_frame_total > rc->num_entries + h->param.i_bframe )
439 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
440 h->param.i_frame_total, rc->num_entries );
444 /* FIXME: ugly padding because VfW drops delayed B-frames */
445 rc->num_entries += h->param.i_bframe;
447 rc->entry = (ratecontrol_entry_t*) x264_malloc(rc->num_entries * sizeof(ratecontrol_entry_t));
448 memset(rc->entry, 0, rc->num_entries * sizeof(ratecontrol_entry_t));
450 /* init all to skipped p frames */
451 for(i=0; i<rc->num_entries; i++)
453 ratecontrol_entry_t *rce = &rc->entry[i];
454 rce->pict_type = SLICE_TYPE_P;
455 rce->qscale = rce->new_qscale = qp2qscale(20);
456 rce->misc_bits = rc->nmb + 10;
462 for(i=0; i < rc->num_entries - h->param.i_bframe; i++)
464 ratecontrol_entry_t *rce;
471 next= strchr(p, ';');
474 (*next)=0; //sscanf is unbelievably slow on long strings
477 e = sscanf(p, " in:%d ", &frame_number);
479 if(frame_number < 0 || frame_number >= rc->num_entries)
481 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
484 rce = &rc->entry[frame_number];
485 rce->direct_mode = 0;
487 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",
488 &pict_type, &qp, &rce->tex_bits,
489 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
490 &rce->s_count, &rce->direct_mode);
494 case 'I': rce->kept_as_ref = 1;
495 case 'i': rce->pict_type = SLICE_TYPE_I; break;
496 case 'P': rce->pict_type = SLICE_TYPE_P; break;
497 case 'B': rce->kept_as_ref = 1;
498 case 'b': rce->pict_type = SLICE_TYPE_B; break;
499 default: e = -1; break;
503 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
506 rce->qscale = qp2qscale(qp);
510 x264_free(stats_buf);
512 if(h->param.rc.i_rc_method == X264_RC_ABR)
514 if(init_pass2(h) < 0) return -1;
515 } /* else we're using constant quant, so no need to run the bitrate allocation */
518 /* Open output file */
519 /* If input and output files are the same, output to a temp file
520 * and move it to the real name only when it's complete */
521 if( h->param.rc.b_stat_write )
525 rc->psz_stat_file_tmpname = x264_malloc( strlen(h->param.rc.psz_stat_out) + 6 );
526 strcpy( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
527 strcat( rc->psz_stat_file_tmpname, ".temp" );
529 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
530 if( rc->p_stat_file_out == NULL )
532 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
536 p = x264_param2string( &h->param, 1 );
537 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
541 for( i=0; i<h->param.i_threads; i++ )
543 h->thread[i]->rc = rc+i;
547 memcpy( &h->thread[i]->param, &h->param, sizeof( x264_param_t ) );
548 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
555 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
560 z->f_bitrate_factor = 1;
561 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
563 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
565 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
569 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
575 z->param = malloc( sizeof(x264_param_t) );
576 memcpy( z->param, &h->param, sizeof(x264_param_t) );
577 while( (tok = strtok_r( p, ",", &saveptr )) )
579 char *val = strchr( tok, '=' );
585 if( x264_param_parse( z->param, tok, val ) )
587 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
595 static int parse_zones( x264_t *h )
597 x264_ratecontrol_t *rc = h->rc;
599 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
601 char *p, *tok, *saveptr;
602 char *psz_zones = x264_malloc( strlen(h->param.rc.psz_zones)+1 );
603 strcpy( psz_zones, h->param.rc.psz_zones );
604 h->param.rc.i_zones = 1;
605 for( p = psz_zones; *p; p++ )
606 h->param.rc.i_zones += (*p == '/');
607 h->param.rc.zones = x264_malloc( h->param.rc.i_zones * sizeof(x264_zone_t) );
609 for( i = 0; i < h->param.rc.i_zones; i++ )
611 tok = strtok_r( p, "/", &saveptr );
612 if( !tok || parse_zone( h, &h->param.rc.zones[i], tok ) )
616 x264_free( psz_zones );
619 if( h->param.rc.i_zones > 0 )
621 for( i = 0; i < h->param.rc.i_zones; i++ )
623 x264_zone_t z = h->param.rc.zones[i];
624 if( z.i_start < 0 || z.i_start > z.i_end )
626 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
627 z.i_start, z.i_end );
630 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
632 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
633 z.f_bitrate_factor );
638 rc->i_zones = h->param.rc.i_zones + 1;
639 rc->zones = x264_malloc( rc->i_zones * sizeof(x264_zone_t) );
640 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
642 // default zone to fall back to if none of the others match
643 rc->zones[0].i_start = 0;
644 rc->zones[0].i_end = INT_MAX;
645 rc->zones[0].b_force_qp = 0;
646 rc->zones[0].f_bitrate_factor = 1;
647 rc->zones[0].param = x264_malloc( sizeof(x264_param_t) );
648 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
649 for( i = 1; i < rc->i_zones; i++ )
651 if( !rc->zones[i].param )
652 rc->zones[i].param = rc->zones[0].param;
659 static x264_zone_t *get_zone( x264_t *h, int frame_num )
662 for( i = h->rc->i_zones-1; i >= 0; i-- )
664 x264_zone_t *z = &h->rc->zones[i];
665 if( frame_num >= z->i_start && frame_num <= z->i_end )
671 void x264_ratecontrol_summary( x264_t *h )
673 x264_ratecontrol_t *rc = h->rc;
674 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
676 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
677 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
678 qscale2qp( pow( base_cplx, 1 - h->param.rc.f_qcompress )
679 * rc->cplxr_sum / rc->wanted_bits_window ) );
683 void x264_ratecontrol_delete( x264_t *h )
685 x264_ratecontrol_t *rc = h->rc;
688 if( rc->p_stat_file_out )
690 fclose( rc->p_stat_file_out );
691 if( h->i_frame >= rc->num_entries - h->param.i_bframe )
692 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
694 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
695 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
697 x264_free( rc->psz_stat_file_tmpname );
699 x264_free( rc->pred );
700 x264_free( rc->pred_b_from_p );
701 x264_free( rc->entry );
704 x264_free( rc->zones[0].param );
705 if( h->param.rc.psz_zones )
706 for( i=1; i<rc->i_zones; i++ )
707 if( rc->zones[i].param != rc->zones[0].param )
708 x264_free( rc->zones[i].param );
709 x264_free( rc->zones );
714 void x264_ratecontrol_set_estimated_size( x264_t *h, int bits )
716 x264_pthread_mutex_lock( &h->fenc->mutex );
717 h->rc->frame_size_estimated = bits;
718 x264_pthread_mutex_unlock( &h->fenc->mutex );
721 int x264_ratecontrol_get_estimated_size( x264_t const *h)
724 x264_pthread_mutex_lock( &h->fenc->mutex );
725 size = h->rc->frame_size_estimated;
726 x264_pthread_mutex_unlock( &h->fenc->mutex );
730 static void accum_p_qp_update( x264_t *h, float qp )
732 x264_ratecontrol_t *rc = h->rc;
733 rc->accum_p_qp *= .95;
734 rc->accum_p_norm *= .95;
735 rc->accum_p_norm += 1;
736 if( h->sh.i_type == SLICE_TYPE_I )
737 rc->accum_p_qp += qp + rc->ip_offset;
739 rc->accum_p_qp += qp;
742 /* Before encoding a frame, choose a QP for it */
743 void x264_ratecontrol_start( x264_t *h, int i_force_qp )
745 x264_ratecontrol_t *rc = h->rc;
746 ratecontrol_entry_t *rce = NULL;
747 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
752 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
753 x264_encoder_reconfig( h, zone->param );
754 rc->prev_zone = zone;
756 rc->qp_force = i_force_qp;
758 if( h->param.rc.b_stat_read )
760 int frame = h->fenc->i_frame;
761 assert( frame >= 0 && frame < rc->num_entries );
762 rce = h->rc->rce = &h->rc->entry[frame];
764 if( h->sh.i_type == SLICE_TYPE_B
765 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
767 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
768 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
774 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
775 rc->row_pred = &rc->row_preds[h->sh.i_type];
776 update_vbv_plan( h );
779 if( h->sh.i_type != SLICE_TYPE_B )
782 while( h->frames.current[rc->bframes] && IS_X264_TYPE_B(h->frames.current[rc->bframes]->i_type) )
792 q = qscale2qp( rate_estimate_qscale( h ) );
794 else if( rc->b_2pass )
796 rce->new_qscale = rate_estimate_qscale( h );
797 q = qscale2qp( rce->new_qscale );
801 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
802 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
804 q = rc->qp_constant[ h->sh.i_type ];
808 if( zone->b_force_qp )
809 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
811 q -= 6*log(zone->f_bitrate_factor)/log(2);
817 h->fdec->f_qp_avg_rc =
818 h->fdec->f_qp_avg_aq =
820 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
823 rce->new_qp = rc->qp;
825 /* accum_p_qp needs to be here so that future frames can benefit from the
826 * data before this frame is done. but this only works because threading
827 * guarantees to not re-encode any frames. so the non-threaded case does
828 * accum_p_qp later. */
829 if( h->param.i_threads > 1 )
830 accum_p_qp_update( h, rc->qp );
832 if( h->sh.i_type != SLICE_TYPE_B )
833 rc->last_non_b_pict_type = h->sh.i_type;
836 static double predict_row_size( x264_t *h, int y, int qp )
838 /* average between two predictors:
839 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
840 x264_ratecontrol_t *rc = h->rc;
841 double pred_s = predict_size( rc->row_pred, qp2qscale(qp), h->fdec->i_row_satd[y] );
843 if( h->sh.i_type != SLICE_TYPE_I
844 && h->fref0[0]->i_type == h->fdec->i_type
845 && h->fref0[0]->i_row_satd[y] > 0
846 && (abs(h->fref0[0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
848 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
849 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
854 return (pred_s + pred_t) / 2;
857 static double row_bits_so_far( x264_t *h, int y )
861 for( i = 0; i <= y; i++ )
862 bits += h->fdec->i_row_bits[i];
866 static double predict_row_size_sum( x264_t *h, int y, int qp )
869 double bits = row_bits_so_far(h, y);
870 for( i = y+1; i < h->sps->i_mb_height; i++ )
871 bits += predict_row_size( h, i, qp );
876 void x264_ratecontrol_mb( x264_t *h, int bits )
878 x264_ratecontrol_t *rc = h->rc;
879 const int y = h->mb.i_mb_y;
883 h->fdec->i_row_bits[y] += bits;
884 rc->qpa_rc += rc->f_qpm;
885 rc->qpa_aq += h->mb.i_qp;
887 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !rc->b_vbv)
890 h->fdec->i_row_qp[y] = rc->qpm;
892 if( h->sh.i_type == SLICE_TYPE_B )
894 /* B-frames shouldn't use lower QP than their reference frames.
895 * This code is a bit overzealous in limiting B-frame quantizers, but it helps avoid
896 * underflows due to the fact that B-frames are not explicitly covered by VBV. */
897 if( y < h->sps->i_mb_height-1 )
900 int avg_qp = X264_MAX(h->fref0[0]->i_row_qp[y+1], h->fref1[0]->i_row_qp[y+1])
901 + rc->pb_offset * ((h->fenc->i_type == X264_TYPE_BREF) ? 0.5 : 1);
902 rc->qpm = X264_MIN(X264_MAX( rc->qp, avg_qp), 51); //avg_qp could go higher than 51 due to pb_offset
903 i_estimated = row_bits_so_far(h, y); //FIXME: compute full estimated size
904 if (i_estimated > h->rc->frame_size_planned)
905 x264_ratecontrol_set_estimated_size(h, i_estimated);
910 update_predictor( rc->row_pred, qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
912 /* tweak quality based on difference from predicted size */
913 if( y < h->sps->i_mb_height-1 && h->stat.i_slice_count[h->sh.i_type] > 0 )
915 int prev_row_qp = h->fdec->i_row_qp[y];
916 int b0 = predict_row_size_sum( h, y, rc->qpm );
918 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
919 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
920 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
924 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
925 /* area at the top of the frame was measured inaccurately. */
926 if(row_bits_so_far(h,y) < 0.05 * rc->frame_size_planned)
929 headroom = buffer_left_planned/rc->buffer_size;
930 if(h->sh.i_type != SLICE_TYPE_I)
934 if( !rc->b_vbv_min_rate )
935 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
937 while( rc->qpm < i_qp_max
938 && (b1 > rc->frame_size_planned * rc_tol
939 || (rc->buffer_fill - b1 < buffer_left_planned * 0.5)))
942 b1 = predict_row_size_sum( h, y, rc->qpm );
945 /* avoid VBV underflow */
946 while( (rc->qpm < h->param.rc.i_qp_max)
947 && (rc->buffer_fill - b1 < rc->buffer_size * 0.005))
950 b1 = predict_row_size_sum( h, y, rc->qpm );
953 while( rc->qpm > i_qp_min
954 && rc->qpm > h->fdec->i_row_qp[0]
955 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
956 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
959 b1 = predict_row_size_sum( h, y, rc->qpm );
961 x264_ratecontrol_set_estimated_size(h, b1);
964 /* loses the fractional part of the frame-wise qp */
968 int x264_ratecontrol_qp( x264_t *h )
973 /* In 2pass, force the same frame types as in the 1st pass */
974 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
976 x264_ratecontrol_t *rc = h->rc;
977 if( h->param.rc.b_stat_read )
979 if( frame_num >= rc->num_entries )
981 /* We could try to initialize everything required for ABR and
982 * adaptive B-frames, but that would be complicated.
983 * So just calculate the average QP used so far. */
985 h->param.rc.i_qp_constant = (h->stat.i_slice_count[SLICE_TYPE_P] == 0) ? 24
986 : 1 + h->stat.f_slice_qp[SLICE_TYPE_P] / h->stat.i_slice_count[SLICE_TYPE_P];
987 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
988 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 );
989 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 );
991 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
992 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
993 if( h->param.i_bframe_adaptive )
994 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
998 h->param.rc.i_rc_method = X264_RC_CQP;
999 h->param.rc.b_stat_read = 0;
1000 h->param.i_bframe_adaptive = 0;
1001 if( h->param.i_bframe > 1 )
1002 h->param.i_bframe = 1;
1005 switch( rc->entry[frame_num].pict_type )
1008 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I;
1011 return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B;
1020 return X264_TYPE_AUTO;
1024 /* After encoding one frame, save stats and update ratecontrol state */
1025 void x264_ratecontrol_end( x264_t *h, int bits )
1027 x264_ratecontrol_t *rc = h->rc;
1028 const int *mbs = h->stat.frame.i_mb_count;
1033 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1034 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1035 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1036 for( i = B_DIRECT; i < B_8x8; i++ )
1037 h->stat.frame.i_mb_count_p += mbs[i];
1039 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1040 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1042 if( h->param.rc.b_stat_write )
1044 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1045 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1046 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1047 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1048 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1049 char c_direct = h->mb.b_direct_auto_write ?
1050 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1051 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1053 fprintf( rc->p_stat_file_out,
1054 "in:%d out:%d type:%c q:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c;\n",
1055 h->fenc->i_frame, h->i_frame,
1057 h->stat.frame.i_tex_bits,
1058 h->stat.frame.i_mv_bits,
1059 h->stat.frame.i_misc_bits,
1060 h->stat.frame.i_mb_count_i,
1061 h->stat.frame.i_mb_count_p,
1062 h->stat.frame.i_mb_count_skip,
1068 if( h->sh.i_type != SLICE_TYPE_B )
1069 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / rc->last_rceq;
1072 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1073 * Not perfectly accurate with B-refs, but good enough. */
1074 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
1076 rc->cplxr_sum *= rc->cbr_decay;
1077 rc->wanted_bits_window += rc->bitrate / rc->fps;
1078 rc->wanted_bits_window *= rc->cbr_decay;
1080 if( h->param.i_threads == 1 )
1081 accum_p_qp_update( h, rc->qpa_rc );
1086 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
1089 if( h->mb.b_variable_qp )
1091 if( h->sh.i_type == SLICE_TYPE_B )
1093 rc->bframe_bits += bits;
1094 if( !h->frames.current[0] || !IS_X264_TYPE_B(h->frames.current[0]->i_type) )
1096 update_predictor( rc->pred_b_from_p, qp2qscale(rc->qpa_rc),
1097 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1098 rc->bframe_bits = 0;
1103 update_vbv( h, bits );
1106 /****************************************************************************
1108 ***************************************************************************/
1111 * modify the bitrate curve from pass1 for one frame
1113 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1115 x264_ratecontrol_t *rcc= h->rc;
1117 x264_zone_t *zone = get_zone( h, frame_num );
1119 q = pow( rce->blurred_complexity, 1 - h->param.rc.f_qcompress );
1121 // avoid NaN's in the rc_eq
1122 if(!isfinite(q) || rce->tex_bits + rce->mv_bits == 0)
1123 q = rcc->last_qscale;
1128 rcc->last_qscale = q;
1133 if( zone->b_force_qp )
1134 q = qp2qscale(zone->i_qp);
1136 q /= zone->f_bitrate_factor;
1142 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1144 x264_ratecontrol_t *rcc = h->rc;
1145 const int pict_type = rce->pict_type;
1147 // force I/B quants as a function of P quants
1148 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1149 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1150 if( pict_type == SLICE_TYPE_I )
1153 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1154 double ip_factor = fabs( h->param.rc.f_ip_factor );
1155 /* don't apply ip_factor if the following frame is also I */
1156 if( rcc->accum_p_norm <= 0 )
1158 else if( h->param.rc.f_ip_factor < 0 )
1160 else if( rcc->accum_p_norm >= 1 )
1163 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1165 else if( pict_type == SLICE_TYPE_B )
1167 if( h->param.rc.f_pb_factor > 0 )
1169 if( !rce->kept_as_ref )
1170 q *= fabs( h->param.rc.f_pb_factor );
1172 else if( pict_type == SLICE_TYPE_P
1173 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1174 && rce->tex_bits == 0 )
1179 /* last qscale / qdiff stuff */
1180 if(rcc->last_non_b_pict_type==pict_type
1181 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1183 double last_q = rcc->last_qscale_for[pict_type];
1184 double max_qscale = last_q * rcc->lstep;
1185 double min_qscale = last_q / rcc->lstep;
1187 if (q > max_qscale) q = max_qscale;
1188 else if(q < min_qscale) q = min_qscale;
1191 rcc->last_qscale_for[pict_type] = q;
1192 if(pict_type!=SLICE_TYPE_B)
1193 rcc->last_non_b_pict_type = pict_type;
1194 if(pict_type==SLICE_TYPE_I)
1196 rcc->last_accum_p_norm = rcc->accum_p_norm;
1197 rcc->accum_p_norm = 0;
1198 rcc->accum_p_qp = 0;
1200 if(pict_type==SLICE_TYPE_P)
1202 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1203 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1204 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1209 static double predict_size( predictor_t *p, double q, double var )
1211 return p->coeff*var / (q*p->count);
1214 static void update_predictor( predictor_t *p, double q, double var, double bits )
1218 p->count *= p->decay;
1219 p->coeff *= p->decay;
1221 p->coeff += bits*q / var;
1224 // update VBV after encoding a frame
1225 static void update_vbv( x264_t *h, int bits )
1227 x264_ratecontrol_t *rcc = h->rc;
1228 x264_ratecontrol_t *rct = h->thread[0]->rc;
1230 if( rcc->last_satd >= h->mb.i_mb_count )
1231 update_predictor( &rct->pred[h->sh.i_type], qp2qscale(rcc->qpa_rc), rcc->last_satd, bits );
1236 rct->buffer_fill_final += rct->buffer_rate - bits;
1237 if( rct->buffer_fill_final < 0 )
1238 x264_log( h, X264_LOG_WARNING, "VBV underflow (%.0f bits)\n", rct->buffer_fill_final );
1239 rct->buffer_fill_final = x264_clip3f( rct->buffer_fill_final, 0, rct->buffer_size );
1242 // provisionally update VBV according to the planned size of all frames currently in progress
1243 static void update_vbv_plan( x264_t *h )
1245 x264_ratecontrol_t *rcc = h->rc;
1246 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final;
1247 if( h->param.i_threads > 1 )
1249 int j = h->rc - h->thread[0]->rc;
1251 for( i=1; i<h->param.i_threads; i++ )
1253 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1254 double bits = t->rc->frame_size_planned;
1255 if( !t->b_thread_active )
1257 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1258 rcc->buffer_fill += rcc->buffer_rate - bits;
1259 rcc->buffer_fill = x264_clip3( rcc->buffer_fill, 0, rcc->buffer_size );
1264 // apply VBV constraints and clip qscale to between lmin and lmax
1265 static double clip_qscale( x264_t *h, int pict_type, double q )
1267 x264_ratecontrol_t *rcc = h->rc;
1268 double lmin = rcc->lmin[pict_type];
1269 double lmax = rcc->lmax[pict_type];
1272 /* B-frames are not directly subject to VBV,
1273 * since they are controlled by the P-frames' QPs.
1274 * FIXME: in 2pass we could modify previous frames' QP too,
1275 * instead of waiting for the buffer to fill */
1277 ( pict_type == SLICE_TYPE_P ||
1278 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) )
1280 if( rcc->buffer_fill/rcc->buffer_size < 0.5 )
1281 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1284 if( rcc->b_vbv && rcc->last_satd > 0 )
1286 /* Now a hard threshold to make sure the frame fits in VBV.
1287 * This one is mostly for I-frames. */
1288 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1290 if( bits > rcc->buffer_fill/2 )
1291 qf = x264_clip3f( rcc->buffer_fill/(2*bits), 0.2, 1.0 );
1294 if( bits < rcc->buffer_rate/2 )
1295 q *= bits*2/rcc->buffer_rate;
1296 q = X264_MAX( q0, q );
1298 /* Check B-frame complexity, and use up any bits that would
1299 * overflow before the next P-frame. */
1300 if( h->sh.i_type == SLICE_TYPE_P )
1302 int nb = rcc->bframes;
1303 double pbbits = bits;
1304 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1307 if( bbits > rcc->buffer_rate )
1309 pbbits += nb * bbits;
1311 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1312 if( pbbits < space )
1314 q *= X264_MAX( pbbits / space,
1315 bits / (0.5 * rcc->buffer_size) );
1317 q = X264_MAX( q0-5, q );
1320 if( !rcc->b_vbv_min_rate )
1321 q = X264_MAX( q0, q );
1326 else if(rcc->b_2pass)
1328 double min2 = log(lmin);
1329 double max2 = log(lmax);
1330 q = (log(q) - min2)/(max2-min2) - 0.5;
1331 q = 1.0/(1.0 + exp(-4*q));
1332 q = q*(max2-min2) + min2;
1336 return x264_clip3f(q, lmin, lmax);
1339 // update qscale for 1 frame based on actual bits used so far
1340 static float rate_estimate_qscale( x264_t *h )
1343 x264_ratecontrol_t *rcc = h->rc;
1344 ratecontrol_entry_t rce;
1345 int pict_type = h->sh.i_type;
1346 double lmin = rcc->lmin[pict_type];
1347 double lmax = rcc->lmax[pict_type];
1348 int64_t total_bits = 8*(h->stat.i_slice_size[SLICE_TYPE_I]
1349 + h->stat.i_slice_size[SLICE_TYPE_P]
1350 + h->stat.i_slice_size[SLICE_TYPE_B]);
1355 if(pict_type != rce.pict_type)
1357 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1358 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1362 if( pict_type == SLICE_TYPE_B )
1364 /* B-frames don't have independent ratecontrol, but rather get the
1365 * average QP of the two adjacent P-frames + an offset */
1367 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1368 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1369 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1370 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1371 float q0 = h->fref0[0]->f_qp_avg_rc;
1372 float q1 = h->fref1[0]->f_qp_avg_rc;
1374 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1375 q0 -= rcc->pb_offset/2;
1376 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1377 q1 -= rcc->pb_offset/2;
1380 q = (q0 + q1) / 2 + rcc->ip_offset;
1386 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1388 if(h->fenc->b_kept_as_ref)
1389 q += rcc->pb_offset/2;
1391 q += rcc->pb_offset;
1393 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1394 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1396 return qp2qscale(q);
1400 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1404 //FIXME adjust abr_buffer based on distance to the end of the video
1406 int64_t predicted_bits = total_bits;
1410 if( h->param.i_threads > 1 )
1412 int j = h->rc - h->thread[0]->rc;
1414 for( i=1; i<h->param.i_threads; i++ )
1416 x264_t *t = h->thread[ (j+i)%h->param.i_threads ];
1417 double bits = t->rc->frame_size_planned;
1418 if( !t->b_thread_active )
1420 bits = X264_MAX(bits, x264_ratecontrol_get_estimated_size(t));
1421 predicted_bits += (int64_t)bits;
1427 if( h->fenc->i_frame < h->param.i_threads )
1428 predicted_bits += (int64_t)h->fenc->i_frame * rcc->bitrate / rcc->fps;
1430 predicted_bits += (int64_t)(h->param.i_threads - 1) * rcc->bitrate / rcc->fps;
1433 diff = predicted_bits - (int64_t)rce.expected_bits;
1435 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1436 if( ((h->fenc->i_frame + 1 - h->param.i_threads) >= rcc->fps) &&
1437 (rcc->expected_bits_sum > 0))
1439 /* Adjust quant based on the difference between
1440 * achieved and expected bitrate so far */
1441 double time = (double)h->fenc->i_frame / rcc->num_entries;
1442 double w = x264_clip3f( time*100, 0.0, 1.0 );
1443 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1447 /* Do not overflow vbv */
1448 double expected_size = qscale2bits(&rce, q);
1449 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1450 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
1451 double qmax = q*(2 - expected_fullness);
1452 double size_constraint = 1 + expected_fullness;
1453 qmax = X264_MAX(qmax, rce.new_qscale);
1454 if (expected_fullness < .05)
1456 qmax = X264_MIN(qmax, lmax);
1457 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
1458 ((expected_vbv < 0) && (q < lmax)))
1461 expected_size = qscale2bits(&rce, q);
1462 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1464 rcc->last_satd = x264_rc_analyse_slice( h );
1466 q = x264_clip3f( q, lmin, lmax );
1468 else /* 1pass ABR */
1470 /* Calculate the quantizer which would have produced the desired
1471 * average bitrate if it had been applied to all frames so far.
1472 * Then modulate that quant based on the current frame's complexity
1473 * relative to the average complexity so far (using the 2pass RCEQ).
1474 * Then bias the quant up or down if total size so far was far from
1476 * Result: Depending on the value of rate_tolerance, there is a
1477 * tradeoff between quality and bitrate precision. But at large
1478 * tolerances, the bit distribution approaches that of 2pass. */
1480 double wanted_bits, overflow=1, lmin, lmax;
1482 rcc->last_satd = x264_rc_analyse_slice( h );
1483 rcc->short_term_cplxsum *= 0.5;
1484 rcc->short_term_cplxcount *= 0.5;
1485 rcc->short_term_cplxsum += rcc->last_satd;
1486 rcc->short_term_cplxcount ++;
1488 rce.tex_bits = rcc->last_satd;
1489 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1491 rce.p_count = rcc->nmb;
1495 rce.pict_type = pict_type;
1497 if( h->param.rc.i_rc_method == X264_RC_CRF )
1499 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1503 int i_frame_done = h->fenc->i_frame + 1 - h->param.i_threads;
1505 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1507 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
1508 wanted_bits = i_frame_done * rcc->bitrate / rcc->fps;
1509 if( wanted_bits > 0 )
1511 abr_buffer *= X264_MAX( 1, sqrt(i_frame_done/25) );
1512 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1517 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1518 /* should test _next_ pict type, but that isn't decided yet */
1519 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1521 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1522 q /= fabs( h->param.rc.f_ip_factor );
1524 else if( h->i_frame > 0 )
1526 /* Asymmetric clipping, because symmetric would prevent
1527 * overflow control in areas of rapidly oscillating complexity */
1528 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1529 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1530 if( overflow > 1.1 && h->i_frame > 3 )
1532 else if( overflow < 0.9 )
1535 q = x264_clip3f(q, lmin, lmax);
1537 else if( h->param.rc.i_rc_method == X264_RC_CRF )
1539 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1542 //FIXME use get_diff_limited_q() ?
1543 q = clip_qscale( h, pict_type, q );
1546 rcc->last_qscale_for[pict_type] =
1547 rcc->last_qscale = q;
1549 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
1550 rcc->last_qscale_for[SLICE_TYPE_P] = q;
1552 if( rcc->b_2pass && rcc->b_vbv)
1553 rcc->frame_size_planned = qscale2bits(&rce, q);
1555 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1556 x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
1561 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
1565 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
1566 /* these vars are updated in x264_ratecontrol_start()
1567 * so copy them from the context that most recently started (prev)
1568 * to the context that's about to start (cur).
1574 COPY(last_qscale_for);
1575 COPY(last_non_b_pict_type);
1576 COPY(short_term_cplxsum);
1577 COPY(short_term_cplxcount);
1584 #define COPY(var) next->rc->var = cur->rc->var
1585 /* these vars are updated in x264_ratecontrol_end()
1586 * so copy them from the context that most recently ended (cur)
1587 * to the context that's about to end (next)
1590 COPY(expected_bits_sum);
1591 COPY(wanted_bits_window);
1595 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
1596 /* the rest of the variables are either constant or thread-local */
1599 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
1601 /* find an interval ending on an overflow or underflow (depending on whether
1602 * we're adding or removing bits), and starting on the earliest frame that
1603 * can influence the buffer fill of that end frame. */
1604 x264_ratecontrol_t *rcc = h->rc;
1605 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
1606 const double buffer_max = .9 * rcc->buffer_size;
1607 double fill = fills[*t0-1];
1608 double parity = over ? 1. : -1.;
1609 int i, start=-1, end=-1;
1610 for(i = *t0; i < rcc->num_entries; i++)
1612 fill += (rcc->buffer_rate - qscale2bits(&rcc->entry[i], rcc->entry[i].new_qscale)) * parity;
1613 fill = x264_clip3f(fill, 0, rcc->buffer_size);
1615 if(fill <= buffer_min || i == 0)
1621 else if(fill >= buffer_max && start >= 0)
1626 return start>=0 && end>=0;
1629 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
1631 x264_ratecontrol_t *rcc = h->rc;
1632 double qscale_orig, qscale_new;
1637 for(i = t0; i <= t1; i++)
1639 qscale_orig = rcc->entry[i].new_qscale;
1640 qscale_orig = x264_clip3f(qscale_orig, qscale_min, qscale_max);
1641 qscale_new = qscale_orig * adjustment;
1642 qscale_new = x264_clip3f(qscale_new, qscale_min, qscale_max);
1643 rcc->entry[i].new_qscale = qscale_new;
1644 adjusted = adjusted || (qscale_new != qscale_orig);
1649 static double count_expected_bits( x264_t *h )
1651 x264_ratecontrol_t *rcc = h->rc;
1652 double expected_bits = 0;
1654 for(i = 0; i < rcc->num_entries; i++)
1656 ratecontrol_entry_t *rce = &rcc->entry[i];
1657 rce->expected_bits = expected_bits;
1658 expected_bits += qscale2bits(rce, rce->new_qscale);
1660 return expected_bits;
1663 static void vbv_pass2( x264_t *h )
1665 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
1666 * frames in the interval until either buffer is full at some intermediate frame or the
1667 * last frame in the interval no longer underflows. Recompute intervals and repeat.
1668 * Then do the converse to put bits back into overflow areas until target size is met */
1670 x264_ratecontrol_t *rcc = h->rc;
1671 double *fills = x264_malloc((rcc->num_entries+1)*sizeof(double));
1672 double all_available_bits = h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps;
1673 double expected_bits = 0;
1675 double prev_bits = 0;
1677 double qscale_min = qp2qscale(h->param.rc.i_qp_min);
1678 double qscale_max = qp2qscale(h->param.rc.i_qp_max);
1680 int adj_min, adj_max;
1684 /* adjust overall stream size */
1688 prev_bits = expected_bits;
1690 if(expected_bits != 0)
1691 { /* not first iteration */
1692 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
1693 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
1697 while(adj_min && find_underflow(h, fills, &t0, &t1, 1))
1699 adj_min = fix_underflow(h, t0, t1, adjustment, qscale_min, qscale_max);
1704 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
1706 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
1708 while(adj_max && find_underflow(h, fills, &t0, &t1, 0))
1709 adj_max = fix_underflow(h, t0, t1, 1.001, qscale_min, qscale_max);
1711 expected_bits = count_expected_bits(h);
1712 } while(expected_bits < .995 * all_available_bits && expected_bits > prev_bits);
1715 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
1717 /* store expected vbv filling values for tracking when encoding */
1718 for(i = 0; i < rcc->num_entries; i++)
1719 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
1724 static int init_pass2( x264_t *h )
1726 x264_ratecontrol_t *rcc = h->rc;
1727 uint64_t all_const_bits = 0;
1728 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps);
1729 double rate_factor, step, step_mult;
1730 double qblur = h->param.rc.f_qblur;
1731 double cplxblur = h->param.rc.f_complexity_blur;
1732 const int filter_size = (int)(qblur*4) | 1;
1733 double expected_bits;
1734 double *qscale, *blurred_qscale;
1737 /* find total/average complexity & const_bits */
1738 for(i=0; i<rcc->num_entries; i++)
1740 ratecontrol_entry_t *rce = &rcc->entry[i];
1741 all_const_bits += rce->misc_bits;
1744 if( all_available_bits < all_const_bits)
1746 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
1747 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)));
1751 /* Blur complexities, to reduce local fluctuation of QP.
1752 * We don't blur the QPs directly, because then one very simple frame
1753 * could drag down the QP of a nearby complex frame and give it more
1754 * bits than intended. */
1755 for(i=0; i<rcc->num_entries; i++)
1757 ratecontrol_entry_t *rce = &rcc->entry[i];
1758 double weight_sum = 0;
1759 double cplx_sum = 0;
1760 double weight = 1.0;
1761 double gaussian_weight;
1763 /* weighted average of cplx of future frames */
1764 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++)
1766 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
1767 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1770 gaussian_weight = weight * exp(-j*j/200.0);
1771 weight_sum += gaussian_weight;
1772 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
1774 /* weighted average of cplx of past frames */
1776 for(j=0; j<=cplxblur*2 && j<=i; j++)
1778 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
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);
1782 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
1786 rce->blurred_complexity = cplx_sum / weight_sum;
1789 qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1791 blurred_qscale = x264_malloc(sizeof(double)*rcc->num_entries);
1793 blurred_qscale = qscale;
1795 /* Search for a factor which, when multiplied by the RCEQ values from
1796 * each frame, adds up to the desired total size.
1797 * There is no exact closed-form solution because of VBV constraints and
1798 * because qscale2bits is not invertible, but we can start with the simple
1799 * approximation of scaling the 1st pass by the ratio of bitrates.
1800 * The search range is probably overkill, but speed doesn't matter here. */
1803 for(i=0; i<rcc->num_entries; i++)
1804 expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i));
1805 step_mult = all_available_bits / expected_bits;
1808 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
1811 rate_factor += step;
1813 rcc->last_non_b_pict_type = -1;
1814 rcc->last_accum_p_norm = 1;
1815 rcc->accum_p_norm = 0;
1818 for(i=0; i<rcc->num_entries; i++)
1820 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
1823 /* fixed I/B qscale relative to P */
1824 for(i=rcc->num_entries-1; i>=0; i--)
1826 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
1827 assert(qscale[i] >= 0);
1833 assert(filter_size%2==1);
1834 for(i=0; i<rcc->num_entries; i++)
1836 ratecontrol_entry_t *rce = &rcc->entry[i];
1838 double q=0.0, sum=0.0;
1840 for(j=0; j<filter_size; j++)
1842 int index = i+j-filter_size/2;
1844 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
1845 if(index < 0 || index >= rcc->num_entries)
1847 if(rce->pict_type != rcc->entry[index].pict_type)
1849 q += qscale[index] * coeff;
1852 blurred_qscale[i] = q/sum;
1856 /* find expected bits */
1857 for(i=0; i<rcc->num_entries; i++)
1859 ratecontrol_entry_t *rce = &rcc->entry[i];
1860 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
1861 assert(rce->new_qscale >= 0);
1862 expected_bits += qscale2bits(rce, rce->new_qscale);
1865 if(expected_bits > all_available_bits) rate_factor -= step;
1870 x264_free(blurred_qscale);
1874 expected_bits = count_expected_bits(h);
1876 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
1879 for(i=0; i<rcc->num_entries; i++)
1880 avgq += rcc->entry[i].new_qscale;
1881 avgq = qscale2qp(avgq / rcc->num_entries);
1883 if ((expected_bits > all_available_bits) || (!rcc->b_vbv))
1884 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
1885 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
1886 (float)h->param.rc.i_bitrate,
1887 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
1889 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
1891 if(h->param.rc.i_qp_min > 0)
1892 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
1894 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
1896 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
1898 if(h->param.rc.i_qp_max < 51)
1899 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
1901 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
1903 else if(!(rcc->b_2pass && rcc->b_vbv))
1904 x264_log(h, X264_LOG_WARNING, "internal error\n");