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
31 #include "common/common.h"
32 #include "common/cpu.h"
33 #include "ratecontrol.h"
45 uint64_t expected_bits; /*total expected bits up to the current frame (current one excluded)*/
52 float blurred_complexity;
55 int16_t i_weight_denom;
58 } ratecontrol_entry_t;
68 struct x264_ratecontrol_t
77 double rate_tolerance;
79 int nmb; /* number of macroblocks in a frame */
83 ratecontrol_entry_t *rce;
84 int qp; /* qp for current frame */
85 int qpm; /* qp for current macroblock */
86 float f_qpm; /* qp for current macroblock: precise float for AQ */
87 float qpa_rc; /* average of macroblocks' qp before aq */
88 float qpa_aq; /* average of macroblocks' qp after aq */
89 float qp_novbv; /* QP for the current frame if 1-pass VBV was disabled. */
94 double buffer_fill_final; /* real buffer as of the last finished frame */
95 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
96 double buffer_rate; /* # of bits added to buffer_fill after each frame */
97 predictor_t *pred; /* predict frame size from satd */
103 double cplxr_sum; /* sum of bits*qscale/rceq */
104 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow, only includes finished frames */
105 double wanted_bits_window; /* target bitrate * window */
107 double short_term_cplxsum;
108 double short_term_cplxcount;
109 double rate_factor_constant;
114 FILE *p_stat_file_out;
115 char *psz_stat_file_tmpname;
116 FILE *p_mbtree_stat_file_out;
117 char *psz_mbtree_stat_file_tmpname;
118 char *psz_mbtree_stat_file_name;
119 FILE *p_mbtree_stat_file_in;
121 int num_entries; /* number of ratecontrol_entry_ts */
122 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
124 double last_qscale_for[5]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
125 int last_non_b_pict_type;
126 double accum_p_qp; /* for determining I-frame quant */
128 double last_accum_p_norm;
129 double lmin[5]; /* min qscale by frame type */
131 double lstep; /* max change (multiply) in qscale per frame */
132 uint16_t *qp_buffer[2]; /* Global buffers for converting MB-tree quantizer data. */
133 int qpbuf_pos; /* In order to handle pyramid reordering, QP buffer acts as a stack.
134 * This value is the current position (0 or 1). */
137 float frame_size_estimated; /* Access to this variable must be atomic: double is
138 * not atomic on all arches we care about */
139 double frame_size_maximum; /* Maximum frame size due to MinCR */
140 double frame_size_planned;
141 double slice_size_planned;
142 double max_frame_error;
143 predictor_t (*row_pred)[2];
144 predictor_t row_preds[5][2];
145 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
146 int bframes; /* # consecutive B-frames before this P-frame */
147 int bframe_bits; /* total cost of those frames */
151 x264_zone_t *prev_zone;
155 static int parse_zones( x264_t *h );
156 static int init_pass2(x264_t *);
157 static float rate_estimate_qscale( x264_t *h );
158 static void update_vbv( x264_t *h, int bits );
159 static void update_vbv_plan( x264_t *h, int overhead );
160 static double predict_size( predictor_t *p, double q, double var );
161 static void update_predictor( predictor_t *p, double q, double var, double bits );
163 #define CMP_OPT_FIRST_PASS( opt, param_val )\
165 if( ( p = strstr( opts, opt "=" ) ) && sscanf( p, opt "=%d" , &i ) && param_val != i )\
167 x264_log( h, X264_LOG_ERROR, "different " opt " setting than first pass (%d vs %d)\n", param_val, i );\
173 * qp = h.264's quantizer
174 * qscale = linearized quantizer = Lagrange multiplier
176 static inline double qp2qscale(double qp)
178 return 0.85 * pow(2.0, ( qp - 12.0 ) / 6.0);
180 static inline double qscale2qp(double qscale)
182 return 12.0 + 6.0 * log(qscale/0.85) / log(2.0);
185 /* Texture bitrate is not quite inversely proportional to qscale,
186 * probably due the the changing number of SKIP blocks.
187 * MV bits level off at about qp<=12, because the lambda used
188 * for motion estimation is constant there. */
189 static inline double qscale2bits(ratecontrol_entry_t *rce, double qscale)
193 return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
194 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
198 static ALWAYS_INLINE uint32_t ac_energy_plane( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame, int i )
201 int shift = i ? 6 : 8;
202 int stride = frame->i_stride[i];
203 int offset = h->mb.b_interlaced
204 ? w * (mb_x + (mb_y&~1) * stride) + (mb_y&1) * stride
205 : w * (mb_x + mb_y * stride);
206 int pix = i ? PIXEL_8x8 : PIXEL_16x16;
207 stride <<= h->mb.b_interlaced;
208 uint64_t res = h->pixf.var[pix]( frame->plane[i] + offset, stride );
209 uint32_t sum = (uint32_t)res;
210 uint32_t sqr = res >> 32;
211 return sqr - (sum * sum >> shift);
214 // Find the total AC energy of the block in all planes.
215 static NOINLINE uint32_t ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
217 /* This function contains annoying hacks because GCC has a habit of reordering emms
218 * and putting it after floating point ops. As a result, we put the emms at the end of the
219 * function and make sure that its always called before the float math. Noinline makes
220 * sure no reordering goes on. */
221 uint32_t var = ac_energy_plane( h, mb_x, mb_y, frame, 0 );
222 var += ac_energy_plane( h, mb_x, mb_y, frame, 1 );
223 var += ac_energy_plane( h, mb_x, mb_y, frame, 2 );
228 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame )
230 /* constants chosen to result in approximately the same overall bitrate as without AQ.
231 * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
235 /* Need to init it anyways for MB tree. */
236 if( h->param.rc.f_aq_strength == 0 )
239 memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
240 memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
241 if( h->frames.b_have_lowres )
242 for( mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
243 frame->i_inv_qscale_factor[mb_xy] = 256;
247 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
249 float avg_adj_pow2 = 0.f;
250 for( mb_y = 0; mb_y < h->sps->i_mb_height; mb_y++ )
251 for( mb_x = 0; mb_x < h->sps->i_mb_width; mb_x++ )
253 uint32_t energy = ac_energy_mb( h, mb_x, mb_y, frame );
254 float qp_adj = powf( energy + 1, 0.125f );
255 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
257 avg_adj_pow2 += qp_adj * qp_adj;
259 avg_adj /= h->mb.i_mb_count;
260 avg_adj_pow2 /= h->mb.i_mb_count;
261 strength = h->param.rc.f_aq_strength * avg_adj;
262 avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - 14.f) / avg_adj;
265 strength = h->param.rc.f_aq_strength * 1.0397f;
267 for( mb_y = 0; mb_y < h->sps->i_mb_height; mb_y++ )
268 for( mb_x = 0; mb_x < h->sps->i_mb_width; mb_x++ )
271 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
273 qp_adj = frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride];
274 qp_adj = strength * (qp_adj - avg_adj);
278 uint32_t energy = ac_energy_mb( h, mb_x, mb_y, frame );
279 qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - 14.427f);
281 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] =
282 frame->f_qp_offset_aq[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
283 if( h->frames.b_have_lowres )
284 frame->i_inv_qscale_factor[mb_x + mb_y*h->mb.i_mb_stride] = x264_exp2fix8(qp_adj);
289 /*****************************************************************************
290 * x264_adaptive_quant:
291 * adjust macroblock QP based on variance (AC energy) of the MB.
292 * high variance = higher QP
293 * low variance = lower QP
294 * This generally increases SSIM and lowers PSNR.
295 *****************************************************************************/
296 void x264_adaptive_quant( x264_t *h )
299 /* MB-tree currently doesn't adjust quantizers in unreferenced frames. */
300 float qp_offset = h->fdec->b_kept_as_ref ? h->fenc->f_qp_offset[h->mb.i_mb_xy] : h->fenc->f_qp_offset_aq[h->mb.i_mb_xy];
301 h->mb.i_qp = x264_clip3( h->rc->f_qpm + qp_offset + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
304 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame )
306 x264_ratecontrol_t *rc = h->rc;
307 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
310 if( rc->entry[frame->i_frame].kept_as_ref )
313 if( rc->qpbuf_pos < 0 )
319 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
321 if( fread( rc->qp_buffer[rc->qpbuf_pos], sizeof(uint16_t), h->mb.i_mb_count, rc->p_mbtree_stat_file_in ) != h->mb.i_mb_count )
324 if( i_type != i_type_actual && rc->qpbuf_pos == 1 )
326 x264_log(h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type, i_type_actual);
329 } while( i_type != i_type_actual );
332 for( i = 0; i < h->mb.i_mb_count; i++ )
334 frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[rc->qpbuf_pos][i] )) * (1/256.0);
335 if( h->frames.b_have_lowres )
336 frame->i_inv_qscale_factor[i] = x264_exp2fix8(frame->f_qp_offset[i]);
341 x264_adaptive_quant_frame( h, frame );
344 x264_log(h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n");
348 int x264_reference_build_list_optimal( x264_t *h )
350 ratecontrol_entry_t *rce = h->rc->rce;
351 x264_frame_t *frames[16];
352 x264_weight_t weights[16][3];
356 if( rce->refs != h->i_ref0 )
359 memcpy( frames, h->fref0, sizeof(frames) );
360 memcpy( refcount, rce->refcount, sizeof(refcount) );
361 memcpy( weights, h->fenc->weight, sizeof(weights) );
362 memset( &h->fenc->weight[1][0], 0, sizeof(x264_weight_t[15][3]) );
364 /* For now don't reorder ref 0; it seems to lower quality
365 in most cases due to skips. */
366 for( ref = 1; ref < h->i_ref0; ref++ )
371 for( i = 1; i < h->i_ref0; i++ )
372 if( !frames[i]->b_duplicate || frames[i]->i_frame != h->fref0[ref-1]->i_frame )
373 /* Favor lower POC as a tiebreaker. */
374 COPY2_IF_GT( max, refcount[i], bestref, i );
376 /* FIXME: If there are duplicates from frames other than ref0 then it is possible
377 * that the optimal ordering doesnt place every duplicate. */
379 refcount[bestref] = -1;
380 h->fref0[ref] = frames[bestref];
381 memcpy( h->fenc->weight[ref], weights[bestref], sizeof(weights[bestref]) );
387 static char *x264_strcat_filename( char *input, char *suffix )
389 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
392 strcpy( output, input );
393 strcat( output, suffix );
397 void x264_ratecontrol_init_reconfigurable( x264_t *h, int b_init )
399 x264_ratecontrol_t *rc = h->rc;
400 if( !b_init && rc->b_2pass )
403 if( h->param.rc.i_vbv_max_bitrate > 0 && h->param.rc.i_vbv_buffer_size > 0 )
405 if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
407 h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
408 x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
409 h->param.rc.i_vbv_buffer_size );
412 /* We don't support changing the ABR bitrate right now,
413 so if the stream starts as CBR, keep it CBR. */
414 if( rc->b_vbv_min_rate )
415 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
416 rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000. / rc->fps;
417 rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000.;
418 rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
419 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
420 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
423 if( h->param.rc.f_vbv_buffer_init > 1. )
424 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 );
425 h->param.rc.f_vbv_buffer_init = X264_MAX( h->param.rc.f_vbv_buffer_init, rc->buffer_rate / rc->buffer_size );
426 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
428 rc->b_vbv_min_rate = !rc->b_2pass
429 && h->param.rc.i_rc_method == X264_RC_ABR
430 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
433 if( h->param.rc.i_rc_method == X264_RC_CRF )
435 /* Arbitrary rescaling to make CRF somewhat similar to QP.
436 * Try to compensate for MB-tree's effects as well. */
437 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
438 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
439 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
440 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset );
444 int x264_ratecontrol_new( x264_t *h )
446 x264_ratecontrol_t *rc;
451 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
454 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
455 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
457 /* FIXME: use integers */
458 if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0)
459 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
463 if( h->param.rc.b_mb_tree )
465 h->param.rc.f_pb_factor = 1;
469 rc->qcompress = h->param.rc.f_qcompress;
471 rc->bitrate = h->param.rc.i_bitrate * 1000.;
472 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
473 rc->nmb = h->mb.i_mb_count;
474 rc->last_non_b_pict_type = -1;
477 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
479 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
483 x264_ratecontrol_init_reconfigurable( h, 1 );
485 if( rc->rate_tolerance < 0.01 )
487 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
488 rc->rate_tolerance = 0.01;
491 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
495 /* FIXME ABR_INIT_QP is actually used only in CRF */
496 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
497 rc->accum_p_norm = .01;
498 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
499 /* estimated ratio that produces a reasonable QP for the first I-frame */
500 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
501 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
502 rc->last_non_b_pict_type = SLICE_TYPE_I;
505 rc->ip_offset = 6.0 * log(h->param.rc.f_ip_factor) / log(2.0);
506 rc->pb_offset = 6.0 * log(h->param.rc.f_pb_factor) / log(2.0);
507 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
508 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
509 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
510 h->mb.ip_offset = rc->ip_offset + 0.5;
512 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
513 rc->last_qscale = qp2qscale(26);
514 int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
515 CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
516 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
517 for( i = 0; i < 5; i++ )
519 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
520 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
521 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
522 for( j = 0; j < num_preds; j++ )
524 rc->pred[i+j*5].coeff= 2.0;
525 rc->pred[i+j*5].count= 1.0;
526 rc->pred[i+j*5].decay= 0.5;
527 rc->pred[i+j*5].offset= 0.0;
529 for( j = 0; j < 2; j++ )
531 rc->row_preds[i][j].coeff= .25;
532 rc->row_preds[i][j].count= 1.0;
533 rc->row_preds[i][j].decay= 0.5;
534 rc->row_preds[i][j].offset= 0.0;
537 *rc->pred_b_from_p = rc->pred[0];
539 if( parse_zones( h ) < 0 )
541 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
545 /* Load stat file and init 2pass algo */
546 if( h->param.rc.b_stat_read )
548 char *p, *stats_in, *stats_buf;
550 /* read 1st pass stats */
551 assert( h->param.rc.psz_stat_in );
552 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
555 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
558 if( h->param.rc.b_mb_tree )
560 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
561 if( !mbtree_stats_in )
563 rc->p_mbtree_stat_file_in = fopen( mbtree_stats_in, "rb" );
564 x264_free( mbtree_stats_in );
565 if( !rc->p_mbtree_stat_file_in )
567 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
572 /* check whether 1st pass options were compatible with current options */
573 if( !strncmp( stats_buf, "#options:", 9 ) )
576 char *opts = stats_buf;
577 stats_in = strchr( stats_buf, '\n' );
582 if( sscanf( opts, "#options: %dx%d", &i, &j ) != 2 )
584 x264_log( h, X264_LOG_ERROR, "resolution specified in stats file not valid\n" );
587 else if( h->param.rc.b_mb_tree && (i != h->param.i_width || j != h->param.i_height) )
589 x264_log( h, X264_LOG_ERROR, "MB-tree doesn't support different resolution than 1st pass (%dx%d vs %dx%d)\n",
590 h->param.i_width, h->param.i_height, i, j );
594 CMP_OPT_FIRST_PASS( "wpredp", X264_MAX( 0, h->param.analyse.i_weighted_pred ) );
595 CMP_OPT_FIRST_PASS( "bframes", h->param.i_bframe );
596 CMP_OPT_FIRST_PASS( "b_pyramid", h->param.i_bframe_pyramid );
597 CMP_OPT_FIRST_PASS( "intra_refresh", h->param.b_intra_refresh );
598 CMP_OPT_FIRST_PASS( "keyint", h->param.i_keyint_max );
600 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
601 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
603 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
605 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
606 h->mb.b_direct_auto_write = 1;
609 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
610 h->param.i_bframe_adaptive = i;
611 else if( h->param.i_bframe )
613 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
617 if( h->param.rc.b_mb_tree && ( p = strstr( opts, "rc_lookahead=" ) ) && sscanf( p, "rc_lookahead=%d", &i ) )
618 h->param.rc.i_lookahead = i;
621 /* find number of pics */
624 p = strchr(p+1, ';');
627 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
632 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
634 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
635 h->param.i_frame_total, rc->num_entries );
637 if( h->param.i_frame_total > rc->num_entries )
639 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
640 h->param.i_frame_total, rc->num_entries );
644 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
646 /* init all to skipped p frames */
647 for(i=0; i<rc->num_entries; i++)
649 ratecontrol_entry_t *rce = &rc->entry[i];
650 rce->pict_type = SLICE_TYPE_P;
651 rce->qscale = rce->new_qscale = qp2qscale(20);
652 rce->misc_bits = rc->nmb + 10;
658 for(i=0; i < rc->num_entries; i++)
660 ratecontrol_entry_t *rce;
668 next= strchr(p, ';');
671 (*next)=0; //sscanf is unbelievably slow on long strings
674 e = sscanf(p, " in:%d ", &frame_number);
676 if(frame_number < 0 || frame_number >= rc->num_entries)
678 x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i);
681 rce = &rc->entry[frame_number];
682 rce->direct_mode = 0;
684 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",
685 &pict_type, &qp, &rce->tex_bits,
686 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
687 &rce->s_count, &rce->direct_mode);
689 p = strstr( p, "ref:" );
693 for( ref = 0; ref < 16; ref++ )
695 if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
697 p = strchr( p+1, ' ' );
704 rce->i_weight_denom = -1;
705 char *w = strchr( p, 'w' );
707 if( sscanf( w, "w:%hd,%hd,%hd", &rce->i_weight_denom, &rce->weight[0], &rce->weight[1] ) != 3 )
708 rce->i_weight_denom = -1;
710 if( pict_type != 'b' )
711 rce->kept_as_ref = 1;
715 rce->frame_type = X264_TYPE_IDR;
716 rce->pict_type = SLICE_TYPE_I;
719 rce->frame_type = X264_TYPE_I;
720 rce->pict_type = SLICE_TYPE_I;
723 rce->frame_type = X264_TYPE_P;
724 rce->pict_type = SLICE_TYPE_P;
727 rce->frame_type = X264_TYPE_BREF;
728 rce->pict_type = SLICE_TYPE_B;
731 rce->frame_type = X264_TYPE_B;
732 rce->pict_type = SLICE_TYPE_B;
734 default: e = -1; break;
739 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
742 rce->qscale = qp2qscale(qp);
746 x264_free(stats_buf);
748 if(h->param.rc.i_rc_method == X264_RC_ABR)
750 if(init_pass2(h) < 0) return -1;
751 } /* else we're using constant quant, so no need to run the bitrate allocation */
754 /* Open output file */
755 /* If input and output files are the same, output to a temp file
756 * and move it to the real name only when it's complete */
757 if( h->param.rc.b_stat_write )
760 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
761 if( !rc->psz_stat_file_tmpname )
764 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
765 if( rc->p_stat_file_out == NULL )
767 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
771 p = x264_param2string( &h->param, 1 );
773 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
775 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
777 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
778 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
779 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
782 rc->p_mbtree_stat_file_out = fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
783 if( rc->p_mbtree_stat_file_out == NULL )
785 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
791 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
793 CHECKED_MALLOC( rc->qp_buffer[0], h->mb.i_mb_count * sizeof(uint16_t) );
794 if( h->param.i_bframe_pyramid && h->param.rc.b_stat_read )
795 CHECKED_MALLOC( rc->qp_buffer[1], h->mb.i_mb_count * sizeof(uint16_t) );
799 for( i=0; i<h->param.i_threads; i++ )
801 h->thread[i]->rc = rc+i;
805 h->thread[i]->param = h->param;
806 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
815 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
818 char *tok, UNUSED *saveptr=NULL;
820 z->f_bitrate_factor = 1;
821 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
823 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
825 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
829 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
835 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
836 memcpy( z->param, &h->param, sizeof(x264_param_t) );
837 z->param->param_free = x264_free;
838 while( (tok = strtok_r( p, ",", &saveptr )) )
840 char *val = strchr( tok, '=' );
846 if( x264_param_parse( z->param, tok, val ) )
848 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
858 static int parse_zones( x264_t *h )
860 x264_ratecontrol_t *rc = h->rc;
862 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
865 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
866 strcpy( psz_zones, h->param.rc.psz_zones );
867 h->param.rc.i_zones = 1;
868 for( p = psz_zones; *p; p++ )
869 h->param.rc.i_zones += (*p == '/');
870 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
872 for( i = 0; i < h->param.rc.i_zones; i++ )
874 int i_tok = strcspn( p, "/" );
876 if( parse_zone( h, &h->param.rc.zones[i], p ) )
880 x264_free( psz_zones );
883 if( h->param.rc.i_zones > 0 )
885 for( i = 0; i < h->param.rc.i_zones; i++ )
887 x264_zone_t z = h->param.rc.zones[i];
888 if( z.i_start < 0 || z.i_start > z.i_end )
890 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
891 z.i_start, z.i_end );
894 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
896 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
897 z.f_bitrate_factor );
902 rc->i_zones = h->param.rc.i_zones + 1;
903 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
904 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
906 // default zone to fall back to if none of the others match
907 rc->zones[0].i_start = 0;
908 rc->zones[0].i_end = INT_MAX;
909 rc->zones[0].b_force_qp = 0;
910 rc->zones[0].f_bitrate_factor = 1;
911 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
912 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
913 for( i = 1; i < rc->i_zones; i++ )
915 if( !rc->zones[i].param )
916 rc->zones[i].param = rc->zones[0].param;
925 static x264_zone_t *get_zone( x264_t *h, int frame_num )
928 for( i = h->rc->i_zones-1; i >= 0; i-- )
930 x264_zone_t *z = &h->rc->zones[i];
931 if( frame_num >= z->i_start && frame_num <= z->i_end )
937 void x264_ratecontrol_summary( x264_t *h )
939 x264_ratecontrol_t *rc = h->rc;
940 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
942 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
943 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
944 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
945 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
946 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset );
950 void x264_ratecontrol_delete( x264_t *h )
952 x264_ratecontrol_t *rc = h->rc;
956 if( rc->p_stat_file_out )
958 b_regular_file = x264_is_regular_file( rc->p_stat_file_out );
959 fclose( rc->p_stat_file_out );
960 if( h->i_frame >= rc->num_entries && b_regular_file )
961 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
963 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
964 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
966 x264_free( rc->psz_stat_file_tmpname );
968 if( rc->p_mbtree_stat_file_out )
970 b_regular_file = x264_is_regular_file( rc->p_mbtree_stat_file_out );
971 fclose( rc->p_mbtree_stat_file_out );
972 if( h->i_frame >= rc->num_entries && b_regular_file )
973 if( rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
975 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
976 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
978 x264_free( rc->psz_mbtree_stat_file_tmpname );
979 x264_free( rc->psz_mbtree_stat_file_name );
981 if( rc->p_mbtree_stat_file_in )
982 fclose( rc->p_mbtree_stat_file_in );
983 x264_free( rc->pred );
984 x264_free( rc->pred_b_from_p );
985 x264_free( rc->entry );
986 x264_free( rc->qp_buffer[0] );
987 x264_free( rc->qp_buffer[1] );
990 x264_free( rc->zones[0].param );
991 for( i=1; i<rc->i_zones; i++ )
992 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
993 rc->zones[i].param->param_free( rc->zones[i].param );
994 x264_free( rc->zones );
999 static void accum_p_qp_update( x264_t *h, float qp )
1001 x264_ratecontrol_t *rc = h->rc;
1002 rc->accum_p_qp *= .95;
1003 rc->accum_p_norm *= .95;
1004 rc->accum_p_norm += 1;
1005 if( h->sh.i_type == SLICE_TYPE_I )
1006 rc->accum_p_qp += qp + rc->ip_offset;
1008 rc->accum_p_qp += qp;
1011 /* Before encoding a frame, choose a QP for it */
1012 void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
1014 x264_ratecontrol_t *rc = h->rc;
1015 ratecontrol_entry_t *rce = NULL;
1016 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
1021 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
1022 x264_encoder_reconfig( h, zone->param );
1023 rc->prev_zone = zone;
1025 rc->qp_force = i_force_qp;
1027 if( h->param.rc.b_stat_read )
1029 int frame = h->fenc->i_frame;
1030 assert( frame >= 0 && frame < rc->num_entries );
1031 rce = h->rc->rce = &h->rc->entry[frame];
1033 if( h->sh.i_type == SLICE_TYPE_B
1034 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
1036 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
1037 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
1043 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
1044 rc->row_pred = &rc->row_preds[h->sh.i_type];
1045 update_vbv_plan( h, overhead );
1047 const x264_level_t *l = x264_levels;
1048 while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
1051 /* The spec has a bizarre special case for the first frame. */
1052 if( h->i_frame == 0 )
1054 //384 * ( Max( PicSizeInMbs, fR * MaxMBPS ) + MaxMBPS * ( tr( 0 ) - tr,n( 0 ) ) ) / MinCR
1055 double fr = 1. / 172;
1056 int pic_size_in_mbs = h->sps->i_mb_width * h->sps->i_mb_height;
1057 rc->frame_size_maximum = 384 * 8 * X264_MAX( pic_size_in_mbs, fr*l->mbps ) / l->mincr;
1061 //384 * MaxMBPS * ( tr( n ) - tr( n - 1 ) ) / MinCR
1062 rc->frame_size_maximum = 384 * 8 * (1 / rc->fps) * l->mbps / l->mincr;
1066 if( h->sh.i_type != SLICE_TYPE_B )
1067 rc->bframes = h->fenc->i_bframes;
1073 else if( rc->b_abr )
1075 q = qscale2qp( rate_estimate_qscale( h ) );
1077 else if( rc->b_2pass )
1079 rce->new_qscale = rate_estimate_qscale( h );
1080 q = qscale2qp( rce->new_qscale );
1084 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
1085 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
1087 q = rc->qp_constant[ h->sh.i_type ];
1091 if( zone->b_force_qp )
1092 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1094 q -= 6*log(zone->f_bitrate_factor)/log(2);
1098 q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1103 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
1104 h->fdec->f_qp_avg_rc =
1105 h->fdec->f_qp_avg_aq =
1108 rce->new_qp = rc->qp;
1110 accum_p_qp_update( h, rc->f_qpm );
1112 if( h->sh.i_type != SLICE_TYPE_B )
1113 rc->last_non_b_pict_type = h->sh.i_type;
1116 static double predict_row_size( x264_t *h, int y, int qp )
1118 /* average between two predictors:
1119 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1120 x264_ratecontrol_t *rc = h->rc;
1121 double pred_s = predict_size( rc->row_pred[0], qp2qscale(qp), h->fdec->i_row_satd[y] );
1123 if( h->sh.i_type == SLICE_TYPE_I || qp >= h->fref0[0]->i_row_qp[y] )
1125 if( h->sh.i_type == SLICE_TYPE_P
1126 && h->fref0[0]->i_type == h->fdec->i_type
1127 && h->fref0[0]->i_row_satd[y] > 0
1128 && (abs(h->fref0[0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1130 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
1131 * qp2qscale(h->fref0[0]->i_row_qp[y]) / qp2qscale(qp);
1135 return (pred_s + pred_t) / 2;
1137 /* Our QP is lower than the reference! */
1140 double pred_intra = predict_size( rc->row_pred[1], qp2qscale(qp), h->fdec->i_row_satds[0][0][y] );
1141 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1142 return pred_intra + pred_s;
1146 static double row_bits_so_far( x264_t *h, int y )
1150 for( i = h->i_threadslice_start; i <= y; i++ )
1151 bits += h->fdec->i_row_bits[i];
1155 static double predict_row_size_sum( x264_t *h, int y, int qp )
1158 double bits = row_bits_so_far(h, y);
1159 for( i = y+1; i < h->i_threadslice_end; i++ )
1160 bits += predict_row_size( h, i, qp );
1165 void x264_ratecontrol_mb( x264_t *h, int bits )
1167 x264_ratecontrol_t *rc = h->rc;
1168 const int y = h->mb.i_mb_y;
1172 h->fdec->i_row_bits[y] += bits;
1173 rc->qpa_rc += rc->f_qpm;
1174 rc->qpa_aq += h->mb.i_qp;
1176 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !rc->b_vbv )
1179 h->fdec->i_row_qp[y] = rc->qpm;
1181 update_predictor( rc->row_pred[0], qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1182 if( h->sh.i_type == SLICE_TYPE_P && rc->qpm < h->fref0[0]->i_row_qp[y] )
1183 update_predictor( rc->row_pred[1], qp2qscale(rc->qpm), h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1185 /* tweak quality based on difference from predicted size */
1186 if( y < h->i_threadslice_end-1 )
1189 int prev_row_qp = h->fdec->i_row_qp[y];
1190 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
1191 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1193 /* B-frames shouldn't use lower QP than their reference frames. */
1194 if( h->sh.i_type == SLICE_TYPE_B )
1196 i_qp_min = X264_MAX( i_qp_min, X264_MAX( h->fref0[0]->i_row_qp[y+1], h->fref1[0]->i_row_qp[y+1] ) );
1197 rc->qpm = X264_MAX( rc->qpm, i_qp_min );
1200 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1201 float slice_size_planned = h->param.b_sliced_threads ? rc->slice_size_planned : rc->frame_size_planned;
1202 float size_of_other_slices = 0;
1203 if( h->param.b_sliced_threads )
1205 for( i = 0; i < h->param.i_threads; i++ )
1206 if( h != h->thread[i] )
1207 size_of_other_slices += h->thread[i]->rc->frame_size_estimated;
1210 rc->max_frame_error = X264_MAX( 0.05, 1.0 / (h->sps->i_mb_width) );
1212 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1213 float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
1214 int b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1216 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1217 /* area at the top of the frame was measured inaccurately. */
1218 if( row_bits_so_far( h, y ) < 0.05 * slice_size_planned )
1221 if( h->sh.i_type != SLICE_TYPE_I )
1224 if( !rc->b_vbv_min_rate )
1225 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
1227 while( rc->qpm < i_qp_max
1228 && ((b1 > rc->frame_size_planned + rc_tol) ||
1229 (rc->buffer_fill - b1 < buffer_left_planned * 0.5) ||
1230 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1233 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1236 while( rc->qpm > i_qp_min
1237 && (rc->qpm > h->fdec->i_row_qp[0] || rc->single_frame_vbv)
1238 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
1239 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
1242 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1245 /* avoid VBV underflow or MinCR violation */
1246 while( (rc->qpm < h->param.rc.i_qp_max)
1247 && ((rc->buffer_fill - b1 < rc->buffer_rate * rc->max_frame_error) ||
1248 (rc->frame_size_maximum - b1 < rc->frame_size_maximum * rc->max_frame_error)))
1251 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1254 h->rc->frame_size_estimated = predict_row_size_sum( h, y, rc->qpm );
1257 /* loses the fractional part of the frame-wise qp */
1258 rc->f_qpm = rc->qpm;
1261 int x264_ratecontrol_qp( x264_t *h )
1266 /* In 2pass, force the same frame types as in the 1st pass */
1267 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1269 x264_ratecontrol_t *rc = h->rc;
1270 if( h->param.rc.b_stat_read )
1272 if( frame_num >= rc->num_entries )
1274 /* We could try to initialize everything required for ABR and
1275 * adaptive B-frames, but that would be complicated.
1276 * So just calculate the average QP used so far. */
1279 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24
1280 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1281 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
1282 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 );
1283 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 );
1285 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1286 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1287 if( h->param.i_bframe_adaptive )
1288 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1290 for( i = 0; i < h->param.i_threads; i++ )
1292 h->thread[i]->rc->b_abr = 0;
1293 h->thread[i]->rc->b_2pass = 0;
1294 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1295 h->thread[i]->param.rc.b_stat_read = 0;
1296 h->thread[i]->param.i_bframe_adaptive = 0;
1297 h->thread[i]->param.i_scenecut_threshold = 0;
1298 h->thread[i]->param.rc.b_mb_tree = 0;
1299 if( h->thread[i]->param.i_bframe > 1 )
1300 h->thread[i]->param.i_bframe = 1;
1302 return X264_TYPE_AUTO;
1304 return rc->entry[frame_num].frame_type;
1307 return X264_TYPE_AUTO;
1310 void x264_ratecontrol_set_weights( x264_t *h, x264_frame_t *frm )
1312 ratecontrol_entry_t *rce = &h->rc->entry[frm->i_frame];
1313 if( h->param.analyse.i_weighted_pred <= 0 )
1315 if( rce->i_weight_denom >= 0 )
1316 SET_WEIGHT( frm->weight[0][0], 1, rce->weight[0], rce->i_weight_denom, rce->weight[1] );
1319 /* After encoding one frame, save stats and update ratecontrol state */
1320 int x264_ratecontrol_end( x264_t *h, int bits )
1322 x264_ratecontrol_t *rc = h->rc;
1323 const int *mbs = h->stat.frame.i_mb_count;
1328 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1329 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1330 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1331 for( i = B_DIRECT; i < B_8x8; i++ )
1332 h->stat.frame.i_mb_count_p += mbs[i];
1334 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1335 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1337 if( h->param.rc.b_stat_write )
1339 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1340 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1341 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1342 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1343 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1344 char c_direct = h->mb.b_direct_auto_write ?
1345 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1346 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1348 if( fprintf( rc->p_stat_file_out,
1349 "in:%d out:%d type:%c q:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c ref:",
1350 h->fenc->i_frame, h->i_frame,
1352 h->stat.frame.i_tex_bits,
1353 h->stat.frame.i_mv_bits,
1354 h->stat.frame.i_misc_bits,
1355 h->stat.frame.i_mb_count_i,
1356 h->stat.frame.i_mb_count_p,
1357 h->stat.frame.i_mb_count_skip,
1361 /* Only write information for reference reordering once. */
1362 int use_old_stats = h->param.rc.b_stat_read && rc->rce->refs > 1;
1363 for( i = 0; i < (use_old_stats ? rc->rce->refs : h->i_ref0); i++ )
1365 int refcount = use_old_stats ? rc->rce->refcount[i]
1366 : h->param.b_interlaced ? h->stat.frame.i_mb_count_ref[0][i*2]
1367 + h->stat.frame.i_mb_count_ref[0][i*2+1]
1368 : h->stat.frame.i_mb_count_ref[0][i];
1369 if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1373 if( h->sh.weight[0][0].weightfn )
1375 if( fprintf( rc->p_stat_file_out, "w:%"PRId32",%"PRId32",%"PRId32, h->sh.weight[0][0].i_denom, h->sh.weight[0][0].i_scale, h->sh.weight[0][0].i_offset ) < 0 )
1379 if( fprintf( rc->p_stat_file_out, ";\n") < 0 )
1382 /* Don't re-write the data in multi-pass mode. */
1383 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1385 uint8_t i_type = h->sh.i_type;
1387 /* Values are stored as big-endian FIX8.8 */
1388 for( i = 0; i < h->mb.i_mb_count; i++ )
1389 rc->qp_buffer[0][i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1390 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1392 if( fwrite( rc->qp_buffer[0], sizeof(uint16_t), h->mb.i_mb_count, rc->p_mbtree_stat_file_out ) < h->mb.i_mb_count )
1399 if( h->sh.i_type != SLICE_TYPE_B )
1400 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / rc->last_rceq;
1403 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1404 * Not perfectly accurate with B-refs, but good enough. */
1405 rc->cplxr_sum += bits * qp2qscale(rc->qpa_rc) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor));
1407 rc->cplxr_sum *= rc->cbr_decay;
1408 rc->wanted_bits_window += rc->bitrate / rc->fps;
1409 rc->wanted_bits_window *= rc->cbr_decay;
1414 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) );
1417 if( h->mb.b_variable_qp )
1419 if( h->sh.i_type == SLICE_TYPE_B )
1421 rc->bframe_bits += bits;
1422 if( h->fenc->b_last_minigop_bframe )
1424 update_predictor( rc->pred_b_from_p, qp2qscale(rc->qpa_rc),
1425 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1426 rc->bframe_bits = 0;
1431 update_vbv( h, bits );
1434 x264_log(h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n");
1438 /****************************************************************************
1440 ***************************************************************************/
1443 * modify the bitrate curve from pass1 for one frame
1445 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1447 x264_ratecontrol_t *rcc= h->rc;
1449 x264_zone_t *zone = get_zone( h, frame_num );
1451 q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1453 // avoid NaN's in the rc_eq
1454 if(!isfinite(q) || rce->tex_bits + rce->mv_bits == 0)
1455 q = rcc->last_qscale_for[rce->pict_type];
1460 rcc->last_qscale = q;
1465 if( zone->b_force_qp )
1466 q = qp2qscale(zone->i_qp);
1468 q /= zone->f_bitrate_factor;
1474 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1476 x264_ratecontrol_t *rcc = h->rc;
1477 const int pict_type = rce->pict_type;
1479 // force I/B quants as a function of P quants
1480 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1481 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1482 if( pict_type == SLICE_TYPE_I )
1485 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1486 double ip_factor = fabs( h->param.rc.f_ip_factor );
1487 /* don't apply ip_factor if the following frame is also I */
1488 if( rcc->accum_p_norm <= 0 )
1490 else if( h->param.rc.f_ip_factor < 0 )
1492 else if( rcc->accum_p_norm >= 1 )
1495 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1497 else if( pict_type == SLICE_TYPE_B )
1499 if( h->param.rc.f_pb_factor > 0 )
1501 if( !rce->kept_as_ref )
1502 q *= fabs( h->param.rc.f_pb_factor );
1504 else if( pict_type == SLICE_TYPE_P
1505 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1506 && rce->tex_bits == 0 )
1511 /* last qscale / qdiff stuff */
1512 if(rcc->last_non_b_pict_type==pict_type
1513 && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1))
1515 double last_q = rcc->last_qscale_for[pict_type];
1516 double max_qscale = last_q * rcc->lstep;
1517 double min_qscale = last_q / rcc->lstep;
1519 if (q > max_qscale) q = max_qscale;
1520 else if(q < min_qscale) q = min_qscale;
1523 rcc->last_qscale_for[pict_type] = q;
1524 if(pict_type!=SLICE_TYPE_B)
1525 rcc->last_non_b_pict_type = pict_type;
1526 if(pict_type==SLICE_TYPE_I)
1528 rcc->last_accum_p_norm = rcc->accum_p_norm;
1529 rcc->accum_p_norm = 0;
1530 rcc->accum_p_qp = 0;
1532 if(pict_type==SLICE_TYPE_P)
1534 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1535 rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp);
1536 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1541 static double predict_size( predictor_t *p, double q, double var )
1543 return (p->coeff*var + p->offset) / (q*p->count);
1546 static void update_predictor( predictor_t *p, double q, double var, double bits )
1548 const double range = 1.5;
1551 double old_coeff = p->coeff / p->count;
1552 double new_coeff = bits*q / var;
1553 double new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
1554 double new_offset = bits*q - new_coeff_clipped * var;
1555 if( new_offset >= 0 )
1556 new_coeff = new_coeff_clipped;
1559 p->count *= p->decay;
1560 p->coeff *= p->decay;
1561 p->offset *= p->decay;
1563 p->coeff += new_coeff;
1564 p->offset += new_offset;
1567 // update VBV after encoding a frame
1568 static void update_vbv( x264_t *h, int bits )
1570 x264_ratecontrol_t *rcc = h->rc;
1571 x264_ratecontrol_t *rct = h->thread[0]->rc;
1573 if( rcc->last_satd >= h->mb.i_mb_count )
1574 update_predictor( &rct->pred[h->sh.i_type], qp2qscale(rcc->qpa_rc), rcc->last_satd, bits );
1579 rct->buffer_fill_final -= bits;
1580 if( rct->buffer_fill_final < 0 )
1581 x264_log( h, X264_LOG_WARNING, "VBV underflow (frame %d, %.0f bits)\n", h->i_frame, rct->buffer_fill_final );
1582 rct->buffer_fill_final = X264_MAX( rct->buffer_fill_final, 0 );
1583 rct->buffer_fill_final += rcc->buffer_rate;
1584 rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, rcc->buffer_size );
1587 // provisionally update VBV according to the planned size of all frames currently in progress
1588 static void update_vbv_plan( x264_t *h, int overhead )
1590 x264_ratecontrol_t *rcc = h->rc;
1591 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final;
1592 if( h->i_thread_frames > 1 )
1594 int j = h->rc - h->thread[0]->rc;
1596 for( i=1; i<h->i_thread_frames; i++ )
1598 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
1599 double bits = t->rc->frame_size_planned;
1600 if( !t->b_thread_active )
1602 bits = X264_MAX(bits, t->rc->frame_size_estimated);
1603 rcc->buffer_fill -= bits;
1604 rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
1605 rcc->buffer_fill += rcc->buffer_rate;
1606 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1609 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1610 rcc->buffer_fill -= overhead;
1613 // apply VBV constraints and clip qscale to between lmin and lmax
1614 static double clip_qscale( x264_t *h, int pict_type, double q )
1616 x264_ratecontrol_t *rcc = h->rc;
1617 double lmin = rcc->lmin[pict_type];
1618 double lmax = rcc->lmax[pict_type];
1621 /* B-frames are not directly subject to VBV,
1622 * since they are controlled by the P-frames' QPs. */
1624 if( rcc->b_vbv && rcc->last_satd > 0 )
1626 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
1627 * the lookahead overflow and such that the buffer is in a reasonable state
1628 * by the end of the lookahead. */
1629 if( h->param.rc.i_lookahead )
1631 int j, iterations, terminate = 0;
1633 /* Avoid an infinite loop. */
1634 for( iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
1637 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1638 double buffer_fill_cur = rcc->buffer_fill - cur_bits;
1640 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
1641 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
1642 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
1644 /* Loop over the planned future frames. */
1645 for( j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
1647 buffer_fill_cur += rcc->buffer_rate;
1648 int i_type = h->fenc->i_planned_type[j];
1649 int i_satd = h->fenc->i_planned_satd[j];
1650 if( i_type == X264_TYPE_AUTO )
1652 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
1653 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
1654 buffer_fill_cur -= cur_bits;
1656 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
1657 target_fill = X264_MIN( rcc->buffer_fill + j * rcc->buffer_rate * 0.5, rcc->buffer_size * 0.5 );
1658 if( buffer_fill_cur < target_fill )
1664 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
1665 target_fill = x264_clip3f( rcc->buffer_fill - j * rcc->buffer_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
1666 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
1675 /* Fallback to old purely-reactive algorithm: no lookahead. */
1678 if( ( pict_type == SLICE_TYPE_P ||
1679 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
1680 rcc->buffer_fill/rcc->buffer_size < 0.5 )
1682 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1685 /* Now a hard threshold to make sure the frame fits in VBV.
1686 * This one is mostly for I-frames. */
1687 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1689 /* For small VBVs, allow the frame to use up the entire VBV. */
1690 double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
1691 /* For single-frame VBVs, request that the frame use up the entire VBV. */
1692 double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
1694 if( bits > rcc->buffer_fill/max_fill_factor )
1695 qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
1698 if( bits < rcc->buffer_rate/min_fill_factor )
1699 q *= bits*min_fill_factor/rcc->buffer_rate;
1700 q = X264_MAX( q0, q );
1703 /* Apply MinCR restrictions */
1704 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1705 if( bits > rcc->frame_size_maximum )
1706 q *= bits / rcc->frame_size_maximum;
1708 /* Check B-frame complexity, and use up any bits that would
1709 * overflow before the next P-frame. */
1710 if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
1712 int nb = rcc->bframes;
1713 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1714 double pbbits = bits;
1715 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1717 if( bbits > rcc->buffer_rate )
1719 pbbits += nb * bbits;
1721 space = rcc->buffer_fill + (1+nb)*rcc->buffer_rate - rcc->buffer_size;
1722 if( pbbits < space )
1724 q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
1726 q = X264_MAX( q0-5, q );
1729 if( !rcc->b_vbv_min_rate )
1730 q = X264_MAX( q0, q );
1735 else if(rcc->b_2pass)
1737 double min2 = log(lmin);
1738 double max2 = log(lmax);
1739 q = (log(q) - min2)/(max2-min2) - 0.5;
1740 q = 1.0/(1.0 + exp(-4*q));
1741 q = q*(max2-min2) + min2;
1745 return x264_clip3f(q, lmin, lmax);
1748 // update qscale for 1 frame based on actual bits used so far
1749 static float rate_estimate_qscale( x264_t *h )
1752 x264_ratecontrol_t *rcc = h->rc;
1753 ratecontrol_entry_t rce;
1754 int pict_type = h->sh.i_type;
1755 double lmin = rcc->lmin[pict_type];
1756 double lmax = rcc->lmax[pict_type];
1757 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
1758 + h->stat.i_frame_size[SLICE_TYPE_P]
1759 + h->stat.i_frame_size[SLICE_TYPE_B]);
1764 if(pict_type != rce.pict_type)
1766 x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1767 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]);
1771 if( pict_type == SLICE_TYPE_B )
1773 /* B-frames don't have independent ratecontrol, but rather get the
1774 * average QP of the two adjacent P-frames + an offset */
1776 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1777 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1778 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1779 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1780 float q0 = h->fref0[0]->f_qp_avg_rc;
1781 float q1 = h->fref1[0]->f_qp_avg_rc;
1783 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1784 q0 -= rcc->pb_offset/2;
1785 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1786 q1 -= rcc->pb_offset/2;
1789 q = (q0 + q1) / 2 + rcc->ip_offset;
1795 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1797 if(h->fenc->b_kept_as_ref)
1798 q += rcc->pb_offset/2;
1800 q += rcc->pb_offset;
1802 if( rcc->b_2pass && rcc->b_vbv )
1803 rcc->frame_size_planned = qscale2bits( &rce, q );
1805 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1806 h->rc->frame_size_estimated = rcc->frame_size_planned;
1810 rcc->last_satd = x264_rc_analyse_slice( h );
1811 return qp2qscale(q);
1815 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1820 int64_t predicted_bits = total_bits;
1821 /* Adjust ABR buffer based on distance to the end of the video. */
1822 if( rcc->num_entries > h->fenc->i_frame )
1823 abr_buffer *= 0.5 * sqrt( rcc->num_entries - h->fenc->i_frame );
1827 if( h->i_thread_frames > 1 )
1829 int j = h->rc - h->thread[0]->rc;
1831 for( i=1; i<h->i_thread_frames; i++ )
1833 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
1834 double bits = t->rc->frame_size_planned;
1835 if( !t->b_thread_active )
1837 bits = X264_MAX(bits, t->rc->frame_size_estimated);
1838 predicted_bits += (int64_t)bits;
1844 if( h->fenc->i_frame < h->i_thread_frames )
1845 predicted_bits += (int64_t)h->fenc->i_frame * rcc->bitrate / rcc->fps;
1847 predicted_bits += (int64_t)(h->i_thread_frames - 1) * rcc->bitrate / rcc->fps;
1850 diff = predicted_bits - (int64_t)rce.expected_bits;
1852 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1853 if( ((h->fenc->i_frame + 1 - h->i_thread_frames) >= rcc->fps) &&
1854 (rcc->expected_bits_sum > 0))
1856 /* Adjust quant based on the difference between
1857 * achieved and expected bitrate so far */
1858 double time = (double)h->fenc->i_frame / rcc->num_entries;
1859 double w = x264_clip3f( time*100, 0.0, 1.0 );
1860 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
1864 /* Do not overflow vbv */
1865 double expected_size = qscale2bits(&rce, q);
1866 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1867 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
1868 double qmax = q*(2 - expected_fullness);
1869 double size_constraint = 1 + expected_fullness;
1870 qmax = X264_MAX(qmax, rce.new_qscale);
1871 if (expected_fullness < .05)
1873 qmax = X264_MIN(qmax, lmax);
1874 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
1875 ((expected_vbv < 0) && (q < lmax)))
1878 expected_size = qscale2bits(&rce, q);
1879 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
1881 rcc->last_satd = x264_rc_analyse_slice( h );
1883 q = x264_clip3f( q, lmin, lmax );
1885 else /* 1pass ABR */
1887 /* Calculate the quantizer which would have produced the desired
1888 * average bitrate if it had been applied to all frames so far.
1889 * Then modulate that quant based on the current frame's complexity
1890 * relative to the average complexity so far (using the 2pass RCEQ).
1891 * Then bias the quant up or down if total size so far was far from
1893 * Result: Depending on the value of rate_tolerance, there is a
1894 * tradeoff between quality and bitrate precision. But at large
1895 * tolerances, the bit distribution approaches that of 2pass. */
1897 double wanted_bits, overflow=1, lmin, lmax;
1899 rcc->last_satd = x264_rc_analyse_slice( h );
1900 rcc->short_term_cplxsum *= 0.5;
1901 rcc->short_term_cplxcount *= 0.5;
1902 rcc->short_term_cplxsum += rcc->last_satd;
1903 rcc->short_term_cplxcount ++;
1905 rce.tex_bits = rcc->last_satd;
1906 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
1908 rce.p_count = rcc->nmb;
1912 rce.pict_type = pict_type;
1914 if( h->param.rc.i_rc_method == X264_RC_CRF )
1916 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
1920 int i_frame_done = h->fenc->i_frame + 1 - h->i_thread_frames;
1922 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
1924 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
1925 wanted_bits = i_frame_done * rcc->bitrate / rcc->fps;
1926 if( wanted_bits > 0 )
1928 abr_buffer *= X264_MAX( 1, sqrt(i_frame_done/25) );
1929 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
1934 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
1935 /* should test _next_ pict type, but that isn't decided yet */
1936 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
1938 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1939 q /= fabs( h->param.rc.f_ip_factor );
1941 else if( h->i_frame > 0 )
1943 /* Asymmetric clipping, because symmetric would prevent
1944 * overflow control in areas of rapidly oscillating complexity */
1945 lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
1946 lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
1947 if( overflow > 1.1 && h->i_frame > 3 )
1949 else if( overflow < 0.9 )
1952 q = x264_clip3f(q, lmin, lmax);
1954 else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
1956 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
1958 rcc->qp_novbv = qscale2qp(q);
1960 //FIXME use get_diff_limited_q() ?
1961 q = clip_qscale( h, pict_type, q );
1964 rcc->last_qscale_for[pict_type] =
1965 rcc->last_qscale = q;
1967 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
1968 rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );
1970 if( rcc->b_2pass && rcc->b_vbv )
1971 rcc->frame_size_planned = qscale2bits(&rce, q);
1973 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1975 /* Always use up the whole VBV in this case. */
1976 if( rcc->single_frame_vbv )
1977 rcc->frame_size_planned = rcc->buffer_rate;
1978 h->rc->frame_size_estimated = rcc->frame_size_planned;
1983 void x264_threads_normalize_predictors( x264_t *h )
1986 double totalsize = 0;
1987 for( i = 0; i < h->param.i_threads; i++ )
1988 totalsize += h->thread[i]->rc->slice_size_planned;
1989 double factor = h->rc->frame_size_planned / totalsize;
1990 for( i = 0; i < h->param.i_threads; i++ )
1991 h->thread[i]->rc->slice_size_planned *= factor;
1994 void x264_threads_distribute_ratecontrol( x264_t *h )
1997 x264_ratecontrol_t *rc = h->rc;
1999 /* Initialize row predictors */
2000 if( h->i_frame == 0 )
2001 for( i = 0; i < h->param.i_threads; i++ )
2003 x264_ratecontrol_t *t = h->thread[i]->rc;
2004 memcpy( t->row_preds, rc->row_preds, sizeof(rc->row_preds) );
2007 for( i = 0; i < h->param.i_threads; i++ )
2009 x264_t *t = h->thread[i];
2010 memcpy( t->rc, rc, offsetof(x264_ratecontrol_t, row_pred) );
2011 t->rc->row_pred = &t->rc->row_preds[h->sh.i_type];
2012 /* Calculate the planned slice size. */
2013 if( rc->b_vbv && rc->frame_size_planned )
2016 for( row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2017 size += h->fdec->i_row_satd[row];
2018 t->rc->slice_size_planned = predict_size( &rc->pred[h->sh.i_type + (i+1)*5], rc->qpm, size );
2021 t->rc->slice_size_planned = 0;
2023 if( rc->b_vbv && rc->frame_size_planned )
2025 x264_threads_normalize_predictors( h );
2027 if( rc->single_frame_vbv )
2029 /* Compensate for our max frame error threshold: give more bits (proportionally) to smaller slices. */
2030 for( i = 0; i < h->param.i_threads; i++ )
2032 x264_t *t = h->thread[i];
2033 t->rc->max_frame_error = X264_MAX( 0.05, 1.0 / (t->i_threadslice_end - t->i_threadslice_start) );
2034 t->rc->slice_size_planned += 2 * t->rc->max_frame_error * rc->frame_size_planned;
2036 x264_threads_normalize_predictors( h );
2039 for( i = 0; i < h->param.i_threads; i++ )
2040 h->thread[i]->rc->frame_size_estimated = h->thread[i]->rc->slice_size_planned;
2044 void x264_threads_merge_ratecontrol( x264_t *h )
2047 x264_ratecontrol_t *rc = h->rc;
2050 for( i = 0; i < h->param.i_threads; i++ )
2052 x264_t *t = h->thread[i];
2053 x264_ratecontrol_t *rct = h->thread[i]->rc;
2054 if( h->param.rc.i_vbv_buffer_size )
2057 for( row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2058 size += h->fdec->i_row_satd[row];
2059 int bits = t->stat.frame.i_mv_bits + t->stat.frame.i_tex_bits + t->stat.frame.i_misc_bits;
2060 int mb_count = (t->i_threadslice_end - t->i_threadslice_start) * h->sps->i_mb_width;
2061 update_predictor( &rc->pred[h->sh.i_type+5*i], qp2qscale(rct->qpa_rc/mb_count), size, bits );
2065 rc->qpa_rc += rct->qpa_rc;
2066 rc->qpa_aq += rct->qpa_aq;
2070 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
2074 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
2075 /* these vars are updated in x264_ratecontrol_start()
2076 * so copy them from the context that most recently started (prev)
2077 * to the context that's about to start (cur). */
2082 COPY(last_qscale_for);
2083 COPY(last_non_b_pict_type);
2084 COPY(short_term_cplxsum);
2085 COPY(short_term_cplxcount);
2089 /* these vars can be updated by x264_ratecontrol_init_reconfigurable */
2092 COPY(single_frame_vbv);
2094 COPY(b_vbv_min_rate);
2095 COPY(rate_factor_constant);
2101 #define COPY(var) next->rc->var = cur->rc->var
2102 /* these vars are updated in x264_ratecontrol_end()
2103 * so copy them from the context that most recently ended (cur)
2104 * to the context that's about to end (next) */
2106 COPY(expected_bits_sum);
2107 COPY(wanted_bits_window);
2111 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
2112 /* the rest of the variables are either constant or thread-local */
2115 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
2117 /* find an interval ending on an overflow or underflow (depending on whether
2118 * we're adding or removing bits), and starting on the earliest frame that
2119 * can influence the buffer fill of that end frame. */
2120 x264_ratecontrol_t *rcc = h->rc;
2121 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
2122 const double buffer_max = .9 * rcc->buffer_size;
2123 double fill = fills[*t0-1];
2124 double parity = over ? 1. : -1.;
2125 int i, start=-1, end=-1;
2126 for(i = *t0; i < rcc->num_entries; i++)
2128 fill += (rcc->buffer_rate - qscale2bits(&rcc->entry[i], rcc->entry[i].new_qscale)) * parity;
2129 fill = x264_clip3f(fill, 0, rcc->buffer_size);
2131 if(fill <= buffer_min || i == 0)
2137 else if(fill >= buffer_max && start >= 0)
2142 return start>=0 && end>=0;
2145 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
2147 x264_ratecontrol_t *rcc = h->rc;
2148 double qscale_orig, qscale_new;
2153 for(i = t0; i <= t1; i++)
2155 qscale_orig = rcc->entry[i].new_qscale;
2156 qscale_orig = x264_clip3f(qscale_orig, qscale_min, qscale_max);
2157 qscale_new = qscale_orig * adjustment;
2158 qscale_new = x264_clip3f(qscale_new, qscale_min, qscale_max);
2159 rcc->entry[i].new_qscale = qscale_new;
2160 adjusted = adjusted || (qscale_new != qscale_orig);
2165 static double count_expected_bits( x264_t *h )
2167 x264_ratecontrol_t *rcc = h->rc;
2168 double expected_bits = 0;
2170 for(i = 0; i < rcc->num_entries; i++)
2172 ratecontrol_entry_t *rce = &rcc->entry[i];
2173 rce->expected_bits = expected_bits;
2174 expected_bits += qscale2bits(rce, rce->new_qscale);
2176 return expected_bits;
2179 static int vbv_pass2( x264_t *h )
2181 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
2182 * frames in the interval until either buffer is full at some intermediate frame or the
2183 * last frame in the interval no longer underflows. Recompute intervals and repeat.
2184 * Then do the converse to put bits back into overflow areas until target size is met */
2186 x264_ratecontrol_t *rcc = h->rc;
2188 double all_available_bits = h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps;
2189 double expected_bits = 0;
2191 double prev_bits = 0;
2193 double qscale_min = qp2qscale(h->param.rc.i_qp_min);
2194 double qscale_max = qp2qscale(h->param.rc.i_qp_max);
2196 int adj_min, adj_max;
2197 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2201 /* adjust overall stream size */
2205 prev_bits = expected_bits;
2207 if(expected_bits != 0)
2208 { /* not first iteration */
2209 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2210 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2214 while(adj_min && find_underflow(h, fills, &t0, &t1, 1))
2216 adj_min = fix_underflow(h, t0, t1, adjustment, qscale_min, qscale_max);
2221 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2223 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2225 while(adj_max && find_underflow(h, fills, &t0, &t1, 0))
2226 adj_max = fix_underflow(h, t0, t1, 1.001, qscale_min, qscale_max);
2228 expected_bits = count_expected_bits(h);
2229 } while((expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2232 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2234 /* store expected vbv filling values for tracking when encoding */
2235 for(i = 0; i < rcc->num_entries; i++)
2236 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
2244 static int init_pass2( x264_t *h )
2246 x264_ratecontrol_t *rcc = h->rc;
2247 uint64_t all_const_bits = 0;
2248 uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000. * rcc->num_entries / rcc->fps);
2249 double rate_factor, step, step_mult;
2250 double qblur = h->param.rc.f_qblur;
2251 double cplxblur = h->param.rc.f_complexity_blur;
2252 const int filter_size = (int)(qblur*4) | 1;
2253 double expected_bits;
2254 double *qscale, *blurred_qscale;
2257 /* find total/average complexity & const_bits */
2258 for(i=0; i<rcc->num_entries; i++)
2260 ratecontrol_entry_t *rce = &rcc->entry[i];
2261 all_const_bits += rce->misc_bits;
2264 if( all_available_bits < all_const_bits)
2266 x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2267 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)));
2271 /* Blur complexities, to reduce local fluctuation of QP.
2272 * We don't blur the QPs directly, because then one very simple frame
2273 * could drag down the QP of a nearby complex frame and give it more
2274 * bits than intended. */
2275 for(i=0; i<rcc->num_entries; i++)
2277 ratecontrol_entry_t *rce = &rcc->entry[i];
2278 double weight_sum = 0;
2279 double cplx_sum = 0;
2280 double weight = 1.0;
2281 double gaussian_weight;
2283 /* weighted average of cplx of future frames */
2284 for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++)
2286 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2287 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2290 gaussian_weight = weight * exp(-j*j/200.0);
2291 weight_sum += gaussian_weight;
2292 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2294 /* weighted average of cplx of past frames */
2296 for(j=0; j<=cplxblur*2 && j<=i; j++)
2298 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2299 gaussian_weight = weight * exp(-j*j/200.0);
2300 weight_sum += gaussian_weight;
2301 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2302 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2306 rce->blurred_complexity = cplx_sum / weight_sum;
2309 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2310 if( filter_size > 1 )
2311 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2313 blurred_qscale = qscale;
2315 /* Search for a factor which, when multiplied by the RCEQ values from
2316 * each frame, adds up to the desired total size.
2317 * There is no exact closed-form solution because of VBV constraints and
2318 * because qscale2bits is not invertible, but we can start with the simple
2319 * approximation of scaling the 1st pass by the ratio of bitrates.
2320 * The search range is probably overkill, but speed doesn't matter here. */
2323 for(i=0; i<rcc->num_entries; i++)
2325 double q = get_qscale(h, &rcc->entry[i], 1.0, i);
2326 expected_bits += qscale2bits(&rcc->entry[i], q);
2327 rcc->last_qscale_for[rcc->entry[i].pict_type] = q;
2329 step_mult = all_available_bits / expected_bits;
2332 for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2335 rate_factor += step;
2337 rcc->last_non_b_pict_type = -1;
2338 rcc->last_accum_p_norm = 1;
2339 rcc->accum_p_norm = 0;
2342 for(i=0; i<rcc->num_entries; i++)
2344 qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i);
2345 rcc->last_qscale_for[rcc->entry[i].pict_type] = qscale[i];
2348 /* fixed I/B qscale relative to P */
2349 for(i=rcc->num_entries-1; i>=0; i--)
2351 qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
2352 assert(qscale[i] >= 0);
2358 assert(filter_size%2==1);
2359 for(i=0; i<rcc->num_entries; i++)
2361 ratecontrol_entry_t *rce = &rcc->entry[i];
2363 double q=0.0, sum=0.0;
2365 for(j=0; j<filter_size; j++)
2367 int index = i+j-filter_size/2;
2369 double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
2370 if(index < 0 || index >= rcc->num_entries)
2372 if(rce->pict_type != rcc->entry[index].pict_type)
2374 q += qscale[index] * coeff;
2377 blurred_qscale[i] = q/sum;
2381 /* find expected bits */
2382 for(i=0; i<rcc->num_entries; i++)
2384 ratecontrol_entry_t *rce = &rcc->entry[i];
2385 rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
2386 assert(rce->new_qscale >= 0);
2387 expected_bits += qscale2bits(rce, rce->new_qscale);
2390 if(expected_bits > all_available_bits) rate_factor -= step;
2395 x264_free(blurred_qscale);
2398 if( vbv_pass2( h ) )
2400 expected_bits = count_expected_bits(h);
2402 if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
2405 for(i=0; i<rcc->num_entries; i++)
2406 avgq += rcc->entry[i].new_qscale;
2407 avgq = qscale2qp(avgq / rcc->num_entries);
2409 if ((expected_bits > all_available_bits) || (!rcc->b_vbv))
2410 x264_log(h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n");
2411 x264_log(h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
2412 (float)h->param.rc.i_bitrate,
2413 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
2415 if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
2417 if(h->param.rc.i_qp_min > 0)
2418 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
2420 x264_log(h, X264_LOG_WARNING, "try reducing target bitrate\n");
2422 else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
2424 if(h->param.rc.i_qp_max < 51)
2425 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
2427 x264_log(h, X264_LOG_WARNING, "try increasing target bitrate\n");
2429 else if(!(rcc->b_2pass && rcc->b_vbv))
2430 x264_log(h, X264_LOG_WARNING, "internal error\n");