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 "ratecontrol.h"
44 uint64_t expected_bits; /*total expected bits up to the current frame (current one excluded)*/
51 float blurred_complexity;
54 int16_t i_weight_denom;
59 } ratecontrol_entry_t;
69 struct x264_ratecontrol_t
78 double rate_tolerance;
80 int nmb; /* number of macroblocks in a frame */
84 ratecontrol_entry_t *rce;
85 int qp; /* qp for current frame */
86 float 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 double vbv_max_rate; /* # of bits added to buffer_fill per second */
98 predictor_t *pred; /* predict frame size from satd */
100 double rate_factor_max_increment; /* Don't allow RF above (CRF + this value). */
105 double cplxr_sum; /* sum of bits*qscale/rceq */
106 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow, only includes finished frames */
107 double wanted_bits_window; /* target bitrate * window */
109 double short_term_cplxsum;
110 double short_term_cplxcount;
111 double rate_factor_constant;
116 FILE *p_stat_file_out;
117 char *psz_stat_file_tmpname;
118 FILE *p_mbtree_stat_file_out;
119 char *psz_mbtree_stat_file_tmpname;
120 char *psz_mbtree_stat_file_name;
121 FILE *p_mbtree_stat_file_in;
123 int num_entries; /* number of ratecontrol_entry_ts */
124 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
126 double last_qscale_for[5]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
127 int last_non_b_pict_type;
128 double accum_p_qp; /* for determining I-frame quant */
130 double last_accum_p_norm;
131 double lmin[5]; /* min qscale by frame type */
133 double lstep; /* max change (multiply) in qscale per frame */
134 uint16_t *qp_buffer[2]; /* Global buffers for converting MB-tree quantizer data. */
135 int qpbuf_pos; /* In order to handle pyramid reordering, QP buffer acts as a stack.
136 * This value is the current position (0 or 1). */
139 float frame_size_estimated; /* Access to this variable must be atomic: double is
140 * not atomic on all arches we care about */
141 double frame_size_maximum; /* Maximum frame size due to MinCR */
142 double frame_size_planned;
143 double slice_size_planned;
144 double max_frame_error;
145 predictor_t (*row_pred)[2];
146 predictor_t row_preds[5][2];
147 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
148 int bframes; /* # consecutive B-frames before this P-frame */
149 int bframe_bits; /* total cost of those frames */
153 x264_zone_t *prev_zone;
156 int initial_cpb_removal_delay;
157 int initial_cpb_removal_delay_offset;
158 double nrt_first_access_unit; /* nominal removal time */
159 double previous_cpb_final_arrival_time;
163 static int parse_zones( x264_t *h );
164 static int init_pass2(x264_t *);
165 static float rate_estimate_qscale( x264_t *h );
166 static int update_vbv( x264_t *h, int bits );
167 static void update_vbv_plan( x264_t *h, int overhead );
168 static double predict_size( predictor_t *p, double q, double var );
169 static void update_predictor( predictor_t *p, double q, double var, double bits );
171 #define CMP_OPT_FIRST_PASS( opt, param_val )\
173 if( ( p = strstr( opts, opt "=" ) ) && sscanf( p, opt "=%d" , &i ) && param_val != i )\
175 x264_log( h, X264_LOG_ERROR, "different " opt " setting than first pass (%d vs %d)\n", param_val, i );\
181 * qp = h.264's quantizer
182 * qscale = linearized quantizer = Lagrange multiplier
184 static inline double qp2qscale( double qp )
186 return 0.85 * pow( 2.0, ( qp - 12.0 ) / 6.0 );
188 static inline double qscale2qp( double qscale )
190 return 12.0 + 6.0 * log2( qscale/0.85 );
193 /* Texture bitrate is not quite inversely proportional to qscale,
194 * probably due the the changing number of SKIP blocks.
195 * MV bits level off at about qp<=12, because the lambda used
196 * for motion estimation is constant there. */
197 static inline double qscale2bits( ratecontrol_entry_t *rce, double qscale )
201 return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
202 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
206 static ALWAYS_INLINE uint32_t ac_energy_plane( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame, int i )
209 int shift = i ? 6 : 8;
210 int stride = frame->i_stride[i];
211 int offset = h->mb.b_interlaced
212 ? w * (mb_x + (mb_y&~1) * stride) + (mb_y&1) * stride
213 : w * (mb_x + mb_y * stride);
214 int pix = i ? PIXEL_8x8 : PIXEL_16x16;
215 stride <<= h->mb.b_interlaced;
216 uint64_t res = h->pixf.var[pix]( frame->plane[i] + offset, stride );
217 uint32_t sum = (uint32_t)res;
218 uint32_t sqr = res >> 32;
219 return sqr - (sum * sum >> shift);
222 // Find the total AC energy of the block in all planes.
223 static NOINLINE uint32_t ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
225 /* This function contains annoying hacks because GCC has a habit of reordering emms
226 * and putting it after floating point ops. As a result, we put the emms at the end of the
227 * function and make sure that its always called before the float math. Noinline makes
228 * sure no reordering goes on. */
229 uint32_t var = ac_energy_plane( h, mb_x, mb_y, frame, 0 );
230 var += ac_energy_plane( h, mb_x, mb_y, frame, 1 );
231 var += ac_energy_plane( h, mb_x, mb_y, frame, 2 );
236 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame )
238 /* constants chosen to result in approximately the same overall bitrate as without AQ.
239 * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
242 /* Need to init it anyways for MB tree. */
243 if( h->param.rc.f_aq_strength == 0 )
245 memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
246 memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
247 if( h->frames.b_have_lowres )
248 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
249 frame->i_inv_qscale_factor[mb_xy] = 256;
253 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
255 float avg_adj_pow2 = 0.f;
256 for( int mb_y = 0; mb_y < h->sps->i_mb_height; mb_y++ )
257 for( int mb_x = 0; mb_x < h->sps->i_mb_width; mb_x++ )
259 uint32_t energy = ac_energy_mb( h, mb_x, mb_y, frame );
260 float qp_adj = powf( energy + 1, 0.125f );
261 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
263 avg_adj_pow2 += qp_adj * qp_adj;
265 avg_adj /= h->mb.i_mb_count;
266 avg_adj_pow2 /= h->mb.i_mb_count;
267 strength = h->param.rc.f_aq_strength * avg_adj;
268 avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - 14.f) / avg_adj;
271 strength = h->param.rc.f_aq_strength * 1.0397f;
273 for( int mb_y = 0; mb_y < h->sps->i_mb_height; mb_y++ )
274 for( int mb_x = 0; mb_x < h->sps->i_mb_width; mb_x++ )
277 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
279 qp_adj = frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride];
280 qp_adj = strength * (qp_adj - avg_adj);
284 uint32_t energy = ac_energy_mb( h, mb_x, mb_y, frame );
285 qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - 14.427f);
287 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] =
288 frame->f_qp_offset_aq[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
289 if( h->frames.b_have_lowres )
290 frame->i_inv_qscale_factor[mb_x + mb_y*h->mb.i_mb_stride] = x264_exp2fix8(qp_adj);
294 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame )
296 x264_ratecontrol_t *rc = h->rc;
297 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
299 if( rc->entry[frame->i_frame].kept_as_ref )
302 if( rc->qpbuf_pos < 0 )
308 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
310 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 )
313 if( i_type != i_type_actual && rc->qpbuf_pos == 1 )
315 x264_log(h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type, i_type_actual);
318 } while( i_type != i_type_actual );
321 for( int i = 0; i < h->mb.i_mb_count; i++ )
323 frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[rc->qpbuf_pos][i] )) * (1/256.0);
324 if( h->frames.b_have_lowres )
325 frame->i_inv_qscale_factor[i] = x264_exp2fix8(frame->f_qp_offset[i]);
330 x264_adaptive_quant_frame( h, frame );
333 x264_log(h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n");
337 int x264_reference_build_list_optimal( x264_t *h )
339 ratecontrol_entry_t *rce = h->rc->rce;
340 x264_frame_t *frames[16];
341 x264_weight_t weights[16][3];
344 if( rce->refs != h->i_ref0 )
347 memcpy( frames, h->fref0, sizeof(frames) );
348 memcpy( refcount, rce->refcount, sizeof(refcount) );
349 memcpy( weights, h->fenc->weight, sizeof(weights) );
350 memset( &h->fenc->weight[1][0], 0, sizeof(x264_weight_t[15][3]) );
352 /* For now don't reorder ref 0; it seems to lower quality
353 in most cases due to skips. */
354 for( int ref = 1; ref < h->i_ref0; ref++ )
359 for( int i = 1; i < h->i_ref0; i++ )
360 /* Favor lower POC as a tiebreaker. */
361 COPY2_IF_GT( max, refcount[i], bestref, i );
363 /* FIXME: If there are duplicates from frames other than ref0 then it is possible
364 * that the optimal ordering doesnt place every duplicate. */
366 refcount[bestref] = -1;
367 h->fref0[ref] = frames[bestref];
368 memcpy( h->fenc->weight[ref], weights[bestref], sizeof(weights[bestref]) );
374 static char *x264_strcat_filename( char *input, char *suffix )
376 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
379 strcpy( output, input );
380 strcat( output, suffix );
384 void x264_ratecontrol_init_reconfigurable( x264_t *h, int b_init )
386 x264_ratecontrol_t *rc = h->rc;
387 if( !b_init && rc->b_2pass )
390 if( h->param.rc.i_rc_method == X264_RC_CRF )
392 /* Arbitrary rescaling to make CRF somewhat similar to QP.
393 * Try to compensate for MB-tree's effects as well. */
394 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
395 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
396 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
397 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset );
400 if( h->param.rc.i_vbv_max_bitrate > 0 && h->param.rc.i_vbv_buffer_size > 0 )
402 if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
404 h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
405 x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
406 h->param.rc.i_vbv_buffer_size );
409 /* We don't support changing the ABR bitrate right now,
410 so if the stream starts as CBR, keep it CBR. */
411 if( rc->b_vbv_min_rate )
412 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
414 int vbv_buffer_size = h->param.rc.i_vbv_buffer_size * 1000;
415 int vbv_max_bitrate = h->param.rc.i_vbv_max_bitrate * 1000;
418 if( h->param.i_nal_hrd && b_init )
420 h->sps->vui.hrd.i_cpb_cnt = 1;
421 h->sps->vui.hrd.b_cbr_hrd = h->param.i_nal_hrd == X264_NAL_HRD_CBR;
422 h->sps->vui.hrd.i_time_offset_length = 0;
427 int bitrate = 1000*h->param.rc.i_vbv_max_bitrate;
428 int bufsize = 1000*h->param.rc.i_vbv_buffer_size;
430 // normalize HRD size and rate to the value / scale notation
431 h->sps->vui.hrd.i_bit_rate_scale = x264_clip3( x264_ctz( bitrate ) - BR_SHIFT, 0, 15 );
432 h->sps->vui.hrd.i_bit_rate_value = bitrate >> ( h->sps->vui.hrd.i_bit_rate_scale + BR_SHIFT );
433 h->sps->vui.hrd.i_bit_rate_unscaled = h->sps->vui.hrd.i_bit_rate_value << ( h->sps->vui.hrd.i_bit_rate_scale + BR_SHIFT );
434 h->sps->vui.hrd.i_cpb_size_scale = x264_clip3( x264_ctz( bufsize ) - CPB_SHIFT, 0, 15 );
435 h->sps->vui.hrd.i_cpb_size_value = bufsize >> ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
436 h->sps->vui.hrd.i_cpb_size_unscaled = h->sps->vui.hrd.i_cpb_size_value << ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
442 #define MAX_DURATION 0.5
444 int max_cpb_output_delay = h->param.i_keyint_max * MAX_DURATION * h->sps->vui.i_time_scale / h->sps->vui.i_num_units_in_tick;
445 int max_dpb_output_delay = h->sps->vui.i_max_dec_frame_buffering * MAX_DURATION * h->sps->vui.i_time_scale / h->sps->vui.i_num_units_in_tick;
446 int max_delay = (int)(90000.0 * (double)h->sps->vui.hrd.i_cpb_size_unscaled / h->sps->vui.hrd.i_bit_rate_unscaled + 0.5);
448 h->sps->vui.hrd.i_initial_cpb_removal_delay_length = 2 + x264_clip3( 32 - x264_clz( max_delay ), 4, 22 );
449 h->sps->vui.hrd.i_cpb_removal_delay_length = x264_clip3( 32 - x264_clz( max_cpb_output_delay ), 4, 32 );
450 h->sps->vui.hrd.i_dpb_output_delay_length = x264_clip3( 32 - x264_clz( max_dpb_output_delay ), 4, 32 );
454 vbv_buffer_size = X264_MIN( vbv_buffer_size, h->sps->vui.hrd.i_cpb_size_unscaled );
455 vbv_max_bitrate = X264_MIN( vbv_max_bitrate, h->sps->vui.hrd.i_bit_rate_unscaled );
457 else if( h->param.i_nal_hrd && !b_init )
459 x264_log( h, X264_LOG_WARNING, "VBV parameters cannot be changed when NAL HRD is in use\n" );
463 rc->buffer_rate = vbv_max_bitrate / rc->fps;
464 rc->vbv_max_rate = vbv_max_bitrate;
465 rc->buffer_size = vbv_buffer_size;
466 rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
467 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
468 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
469 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.f_rf_constant_max )
471 rc->rate_factor_max_increment = h->param.rc.f_rf_constant_max - h->param.rc.f_rf_constant;
472 if( rc->rate_factor_max_increment <= 0 )
474 x264_log( h, X264_LOG_WARNING, "CRF max must be greater than CRF\n" );
475 rc->rate_factor_max_increment = 0;
480 if( h->param.rc.f_vbv_buffer_init > 1. )
481 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 );
482 h->param.rc.f_vbv_buffer_init = x264_clip3f( X264_MAX( h->param.rc.f_vbv_buffer_init, rc->buffer_rate / rc->buffer_size ), 0, 1);
483 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
485 rc->b_vbv_min_rate = !rc->b_2pass
486 && h->param.rc.i_rc_method == X264_RC_ABR
487 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
492 int x264_ratecontrol_new( x264_t *h )
494 x264_ratecontrol_t *rc;
498 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
501 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
502 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
504 /* FIXME: use integers */
505 if( h->param.i_fps_num > 0 && h->param.i_fps_den > 0 )
506 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
510 if( h->param.rc.b_mb_tree )
512 h->param.rc.f_pb_factor = 1;
516 rc->qcompress = h->param.rc.f_qcompress;
518 rc->bitrate = h->param.rc.i_bitrate * 1000.;
519 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
520 rc->nmb = h->mb.i_mb_count;
521 rc->last_non_b_pict_type = -1;
524 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
526 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
530 x264_ratecontrol_init_reconfigurable( h, 1 );
532 if( rc->rate_tolerance < 0.01 )
534 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
535 rc->rate_tolerance = 0.01;
538 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
542 /* FIXME ABR_INIT_QP is actually used only in CRF */
543 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
544 rc->accum_p_norm = .01;
545 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
546 /* estimated ratio that produces a reasonable QP for the first I-frame */
547 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
548 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
549 rc->last_non_b_pict_type = SLICE_TYPE_I;
552 rc->ip_offset = 6.0 * log2f( h->param.rc.f_ip_factor );
553 rc->pb_offset = 6.0 * log2f( h->param.rc.f_pb_factor );
554 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
555 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
556 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
557 h->mb.ip_offset = rc->ip_offset + 0.5;
559 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
560 rc->last_qscale = qp2qscale( 26 );
561 int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
562 CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
563 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
564 for( int i = 0; i < 5; i++ )
566 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
567 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
568 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
569 for( int j = 0; j < num_preds; j++ )
571 rc->pred[i+j*5].coeff= 2.0;
572 rc->pred[i+j*5].count= 1.0;
573 rc->pred[i+j*5].decay= 0.5;
574 rc->pred[i+j*5].offset= 0.0;
576 for( int j = 0; j < 2; j++ )
578 rc->row_preds[i][j].coeff= .25;
579 rc->row_preds[i][j].count= 1.0;
580 rc->row_preds[i][j].decay= 0.5;
581 rc->row_preds[i][j].offset= 0.0;
584 *rc->pred_b_from_p = rc->pred[0];
586 if( parse_zones( h ) < 0 )
588 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
592 /* Load stat file and init 2pass algo */
593 if( h->param.rc.b_stat_read )
595 char *p, *stats_in, *stats_buf;
597 /* read 1st pass stats */
598 assert( h->param.rc.psz_stat_in );
599 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
602 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
605 if( h->param.rc.b_mb_tree )
607 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
608 if( !mbtree_stats_in )
610 rc->p_mbtree_stat_file_in = fopen( mbtree_stats_in, "rb" );
611 x264_free( mbtree_stats_in );
612 if( !rc->p_mbtree_stat_file_in )
614 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
619 /* check whether 1st pass options were compatible with current options */
620 if( !strncmp( stats_buf, "#options:", 9 ) )
624 char *opts = stats_buf;
625 stats_in = strchr( stats_buf, '\n' );
630 if( sscanf( opts, "#options: %dx%d", &i, &j ) != 2 )
632 x264_log( h, X264_LOG_ERROR, "resolution specified in stats file not valid\n" );
635 else if( h->param.rc.b_mb_tree && (i != h->param.i_width || j != h->param.i_height) )
637 x264_log( h, X264_LOG_ERROR, "MB-tree doesn't support different resolution than 1st pass (%dx%d vs %dx%d)\n",
638 h->param.i_width, h->param.i_height, i, j );
642 if( ( p = strstr( opts, "timebase=" ) ) && sscanf( p, "timebase=%u/%u", &k, &l ) != 2 )
644 x264_log( h, X264_LOG_ERROR, "timebase specified in stats file not valid\n" );
647 if( k != h->param.i_timebase_num || l != h->param.i_timebase_den )
649 x264_log( h, X264_LOG_ERROR, "timebase mismatch with 1st pass (%u/%u vs %u/%u)\n",
650 h->param.i_timebase_num, h->param.i_timebase_den, k, l );
654 CMP_OPT_FIRST_PASS( "weightp", X264_MAX( 0, h->param.analyse.i_weighted_pred ) );
655 CMP_OPT_FIRST_PASS( "bframes", h->param.i_bframe );
656 CMP_OPT_FIRST_PASS( "b_pyramid", h->param.i_bframe_pyramid );
657 CMP_OPT_FIRST_PASS( "intra_refresh", h->param.b_intra_refresh );
658 CMP_OPT_FIRST_PASS( "keyint", h->param.i_keyint_max );
660 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
661 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
663 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
665 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
666 h->mb.b_direct_auto_write = 1;
669 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
670 h->param.i_bframe_adaptive = i;
671 else if( h->param.i_bframe )
673 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
677 if( (h->param.rc.b_mb_tree || h->param.rc.i_vbv_buffer_size) && ( p = strstr( opts, "rc_lookahead=" ) ) && sscanf( p, "rc_lookahead=%d", &i ) )
678 h->param.rc.i_lookahead = i;
681 /* find number of pics */
684 for( num_entries = -1; p; num_entries++ )
685 p = strchr( p + 1, ';' );
688 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
691 rc->num_entries = num_entries;
693 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
695 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
696 h->param.i_frame_total, rc->num_entries );
698 if( h->param.i_frame_total > rc->num_entries )
700 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
701 h->param.i_frame_total, rc->num_entries );
705 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
707 /* init all to skipped p frames */
708 for( int i = 0; i < rc->num_entries; i++ )
710 ratecontrol_entry_t *rce = &rc->entry[i];
711 rce->pict_type = SLICE_TYPE_P;
712 rce->qscale = rce->new_qscale = qp2qscale( 20 );
713 rce->misc_bits = rc->nmb + 10;
719 for( int i = 0; i < rc->num_entries; i++ )
721 ratecontrol_entry_t *rce;
729 next= strchr(p, ';');
731 *next++ = 0; //sscanf is unbelievably slow on long strings
732 e = sscanf( p, " in:%d ", &frame_number );
734 if( frame_number < 0 || frame_number >= rc->num_entries )
736 x264_log( h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i );
739 rce = &rc->entry[frame_number];
740 rce->direct_mode = 0;
742 e += sscanf( p, " in:%*d out:%*d type:%c dur:%d cpbdur:%d q:%f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c",
743 &pict_type, &rce->i_duration, &rce->i_cpb_duration, &qp, &rce->tex_bits,
744 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
745 &rce->s_count, &rce->direct_mode );
747 p = strstr( p, "ref:" );
751 for( ref = 0; ref < 16; ref++ )
753 if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
755 p = strchr( p+1, ' ' );
762 rce->i_weight_denom = -1;
763 char *w = strchr( p, 'w' );
765 if( sscanf( w, "w:%hd,%hd,%hd", &rce->i_weight_denom, &rce->weight[0], &rce->weight[1] ) != 3 )
766 rce->i_weight_denom = -1;
768 if( pict_type != 'b' )
769 rce->kept_as_ref = 1;
773 rce->frame_type = X264_TYPE_IDR;
774 rce->pict_type = SLICE_TYPE_I;
777 rce->frame_type = X264_TYPE_I;
778 rce->pict_type = SLICE_TYPE_I;
781 rce->frame_type = X264_TYPE_P;
782 rce->pict_type = SLICE_TYPE_P;
785 rce->frame_type = X264_TYPE_BREF;
786 rce->pict_type = SLICE_TYPE_B;
789 rce->frame_type = X264_TYPE_B;
790 rce->pict_type = SLICE_TYPE_B;
792 default: e = -1; break;
797 x264_log( h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e );
800 rce->qscale = qp2qscale( qp );
804 x264_free( stats_buf );
806 if( h->param.rc.i_rc_method == X264_RC_ABR )
808 if( init_pass2( h ) < 0 )
810 } /* else we're using constant quant, so no need to run the bitrate allocation */
813 /* Open output file */
814 /* If input and output files are the same, output to a temp file
815 * and move it to the real name only when it's complete */
816 if( h->param.rc.b_stat_write )
819 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
820 if( !rc->psz_stat_file_tmpname )
823 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
824 if( rc->p_stat_file_out == NULL )
826 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
830 p = x264_param2string( &h->param, 1 );
832 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
834 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
836 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
837 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
838 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
841 rc->p_mbtree_stat_file_out = fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
842 if( rc->p_mbtree_stat_file_out == NULL )
844 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
850 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
852 CHECKED_MALLOC( rc->qp_buffer[0], h->mb.i_mb_count * sizeof(uint16_t) );
853 if( h->param.i_bframe_pyramid && h->param.rc.b_stat_read )
854 CHECKED_MALLOC( rc->qp_buffer[1], h->mb.i_mb_count * sizeof(uint16_t) );
858 for( int i = 0; i<h->param.i_threads; i++ )
860 h->thread[i]->rc = rc+i;
864 h->thread[i]->param = h->param;
865 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
874 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
877 char *tok, UNUSED *saveptr=NULL;
879 z->f_bitrate_factor = 1;
880 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
882 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
884 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
888 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
894 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
895 memcpy( z->param, &h->param, sizeof(x264_param_t) );
896 z->param->param_free = x264_free;
897 while( (tok = strtok_r( p, ",", &saveptr )) )
899 char *val = strchr( tok, '=' );
905 if( x264_param_parse( z->param, tok, val ) )
907 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
917 static int parse_zones( x264_t *h )
919 x264_ratecontrol_t *rc = h->rc;
920 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
923 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
924 strcpy( psz_zones, h->param.rc.psz_zones );
925 h->param.rc.i_zones = 1;
926 for( p = psz_zones; *p; p++ )
927 h->param.rc.i_zones += (*p == '/');
928 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
930 for( int i = 0; i < h->param.rc.i_zones; i++ )
932 int i_tok = strcspn( p, "/" );
934 if( parse_zone( h, &h->param.rc.zones[i], p ) )
938 x264_free( psz_zones );
941 if( h->param.rc.i_zones > 0 )
943 for( int i = 0; i < h->param.rc.i_zones; i++ )
945 x264_zone_t z = h->param.rc.zones[i];
946 if( z.i_start < 0 || z.i_start > z.i_end )
948 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
949 z.i_start, z.i_end );
952 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
954 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
955 z.f_bitrate_factor );
960 rc->i_zones = h->param.rc.i_zones + 1;
961 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
962 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
964 // default zone to fall back to if none of the others match
965 rc->zones[0].i_start = 0;
966 rc->zones[0].i_end = INT_MAX;
967 rc->zones[0].b_force_qp = 0;
968 rc->zones[0].f_bitrate_factor = 1;
969 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
970 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
971 for( int i = 1; i < rc->i_zones; i++ )
973 if( !rc->zones[i].param )
974 rc->zones[i].param = rc->zones[0].param;
983 static x264_zone_t *get_zone( x264_t *h, int frame_num )
985 for( int i = h->rc->i_zones - 1; i >= 0; i-- )
987 x264_zone_t *z = &h->rc->zones[i];
988 if( frame_num >= z->i_start && frame_num <= z->i_end )
994 void x264_ratecontrol_summary( x264_t *h )
996 x264_ratecontrol_t *rc = h->rc;
997 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
999 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
1000 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
1001 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
1002 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
1003 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset );
1007 void x264_ratecontrol_delete( x264_t *h )
1009 x264_ratecontrol_t *rc = h->rc;
1012 if( rc->p_stat_file_out )
1014 b_regular_file = x264_is_regular_file( rc->p_stat_file_out );
1015 fclose( rc->p_stat_file_out );
1016 if( h->i_frame >= rc->num_entries && b_regular_file )
1017 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
1019 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1020 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
1022 x264_free( rc->psz_stat_file_tmpname );
1024 if( rc->p_mbtree_stat_file_out )
1026 b_regular_file = x264_is_regular_file( rc->p_mbtree_stat_file_out );
1027 fclose( rc->p_mbtree_stat_file_out );
1028 if( h->i_frame >= rc->num_entries && b_regular_file )
1029 if( rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
1031 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1032 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
1034 x264_free( rc->psz_mbtree_stat_file_tmpname );
1035 x264_free( rc->psz_mbtree_stat_file_name );
1037 if( rc->p_mbtree_stat_file_in )
1038 fclose( rc->p_mbtree_stat_file_in );
1039 x264_free( rc->pred );
1040 x264_free( rc->pred_b_from_p );
1041 x264_free( rc->entry );
1042 x264_free( rc->qp_buffer[0] );
1043 x264_free( rc->qp_buffer[1] );
1046 x264_free( rc->zones[0].param );
1047 for( int i = 1; i < rc->i_zones; i++ )
1048 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
1049 rc->zones[i].param->param_free( rc->zones[i].param );
1050 x264_free( rc->zones );
1055 static void accum_p_qp_update( x264_t *h, float qp )
1057 x264_ratecontrol_t *rc = h->rc;
1058 rc->accum_p_qp *= .95;
1059 rc->accum_p_norm *= .95;
1060 rc->accum_p_norm += 1;
1061 if( h->sh.i_type == SLICE_TYPE_I )
1062 rc->accum_p_qp += qp + rc->ip_offset;
1064 rc->accum_p_qp += qp;
1067 /* Before encoding a frame, choose a QP for it */
1068 void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
1070 x264_ratecontrol_t *rc = h->rc;
1071 ratecontrol_entry_t *rce = NULL;
1072 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
1077 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
1078 x264_encoder_reconfig( h, zone->param );
1079 rc->prev_zone = zone;
1081 rc->qp_force = i_force_qp;
1083 if( h->param.rc.b_stat_read )
1085 int frame = h->fenc->i_frame;
1086 assert( frame >= 0 && frame < rc->num_entries );
1087 rce = h->rc->rce = &h->rc->entry[frame];
1089 if( h->sh.i_type == SLICE_TYPE_B
1090 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
1092 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
1093 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
1099 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
1100 rc->row_pred = &rc->row_preds[h->sh.i_type];
1101 rc->buffer_rate = h->fenc->i_cpb_duration * rc->vbv_max_rate * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1102 update_vbv_plan( h, overhead );
1104 const x264_level_t *l = x264_levels;
1105 while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
1108 int mincr = l->mincr;
1110 /* Blu-ray requires this */
1111 if( l->level_idc == 41 && h->param.i_nal_hrd )
1114 /* The spec has a bizarre special case for the first frame. */
1115 if( h->i_frame == 0 )
1117 //384 * ( Max( PicSizeInMbs, fR * MaxMBPS ) + MaxMBPS * ( tr( 0 ) - tr,n( 0 ) ) ) / MinCR
1118 double fr = 1. / 172;
1119 int pic_size_in_mbs = h->sps->i_mb_width * h->sps->i_mb_height;
1120 rc->frame_size_maximum = 384 * 8 * X264_MAX( pic_size_in_mbs, fr*l->mbps ) / mincr;
1124 //384 * MaxMBPS * ( tr( n ) - tr( n - 1 ) ) / MinCR
1125 rc->frame_size_maximum = 384 * 8 * ((double)h->fenc->i_cpb_duration * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale) * l->mbps / mincr;
1129 if( h->sh.i_type != SLICE_TYPE_B )
1130 rc->bframes = h->fenc->i_bframes;
1136 else if( rc->b_abr )
1138 q = qscale2qp( rate_estimate_qscale( h ) );
1140 else if( rc->b_2pass )
1142 rce->new_qscale = rate_estimate_qscale( h );
1143 q = qscale2qp( rce->new_qscale );
1147 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
1148 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
1150 q = rc->qp_constant[ h->sh.i_type ];
1154 if( zone->b_force_qp )
1155 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1157 q -= 6*log2f( zone->f_bitrate_factor );
1161 q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1165 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
1166 h->fdec->f_qp_avg_rc =
1167 h->fdec->f_qp_avg_aq =
1170 rce->new_qp = rc->qp;
1172 accum_p_qp_update( h, rc->qpm );
1174 if( h->sh.i_type != SLICE_TYPE_B )
1175 rc->last_non_b_pict_type = h->sh.i_type;
1178 static double predict_row_size( x264_t *h, int y, double qp )
1180 /* average between two predictors:
1181 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1182 x264_ratecontrol_t *rc = h->rc;
1183 double pred_s = predict_size( rc->row_pred[0], qp2qscale( qp ), h->fdec->i_row_satd[y] );
1185 if( h->sh.i_type == SLICE_TYPE_I || qp >= h->fref0[0]->f_row_qp[y] )
1187 if( h->sh.i_type == SLICE_TYPE_P
1188 && h->fref0[0]->i_type == h->fdec->i_type
1189 && h->fref0[0]->i_row_satd[y] > 0
1190 && (abs(h->fref0[0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1192 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
1193 * qp2qscale( h->fref0[0]->f_row_qp[y] ) / qp2qscale( qp );
1197 return (pred_s + pred_t) / 2;
1199 /* Our QP is lower than the reference! */
1202 double pred_intra = predict_size( rc->row_pred[1], qp2qscale( qp ), h->fdec->i_row_satds[0][0][y] );
1203 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1204 return pred_intra + pred_s;
1208 static double row_bits_so_far( x264_t *h, int y )
1211 for( int i = h->i_threadslice_start; i <= y; i++ )
1212 bits += h->fdec->i_row_bits[i];
1216 static double predict_row_size_sum( x264_t *h, int y, double qp )
1218 double bits = row_bits_so_far(h, y);
1219 for( int i = y+1; i < h->i_threadslice_end; i++ )
1220 bits += predict_row_size( h, i, qp );
1225 void x264_ratecontrol_mb( x264_t *h, int bits )
1227 x264_ratecontrol_t *rc = h->rc;
1228 const int y = h->mb.i_mb_y;
1232 h->fdec->i_row_bits[y] += bits;
1233 rc->qpa_rc += rc->qpm;
1234 rc->qpa_aq += h->mb.i_qp;
1236 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !rc->b_vbv )
1239 h->fdec->f_row_qp[y] = rc->qpm;
1241 update_predictor( rc->row_pred[0], qp2qscale( rc->qpm ), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1242 if( h->sh.i_type == SLICE_TYPE_P && rc->qpm < h->fref0[0]->f_row_qp[y] )
1243 update_predictor( rc->row_pred[1], qp2qscale( rc->qpm ), h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1245 /* tweak quality based on difference from predicted size */
1246 if( y < h->i_threadslice_end-1 )
1248 float prev_row_qp = h->fdec->f_row_qp[y];
1249 float qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1250 float qp_absolute_max = h->param.rc.i_qp_max;
1251 if( rc->rate_factor_max_increment )
1252 qp_absolute_max = X264_MIN( qp_absolute_max, rc->qp_novbv + rc->rate_factor_max_increment );
1253 float qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, qp_absolute_max );
1254 float step_size = 0.5;
1256 /* B-frames shouldn't use lower QP than their reference frames. */
1257 if( h->sh.i_type == SLICE_TYPE_B )
1259 qp_min = X264_MAX( qp_min, X264_MAX( h->fref0[0]->f_row_qp[y+1], h->fref1[0]->f_row_qp[y+1] ) );
1260 rc->qpm = X264_MAX( rc->qpm, qp_min );
1263 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1264 float slice_size_planned = h->param.b_sliced_threads ? rc->slice_size_planned : rc->frame_size_planned;
1265 float size_of_other_slices = 0;
1266 if( h->param.b_sliced_threads )
1268 for( int i = 0; i < h->param.i_threads; i++ )
1269 if( h != h->thread[i] )
1270 size_of_other_slices += h->thread[i]->rc->frame_size_estimated;
1273 rc->max_frame_error = X264_MAX( 0.05, 1.0 / (h->sps->i_mb_width) );
1275 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1276 float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
1277 int b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1279 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1280 /* area at the top of the frame was measured inaccurately. */
1281 if( row_bits_so_far( h, y ) < 0.05 * slice_size_planned )
1284 if( h->sh.i_type != SLICE_TYPE_I )
1287 if( !rc->b_vbv_min_rate )
1288 qp_min = X264_MAX( qp_min, rc->qp_novbv );
1290 while( rc->qpm < qp_max
1291 && ((b1 > rc->frame_size_planned + rc_tol) ||
1292 (rc->buffer_fill - b1 < buffer_left_planned * 0.5) ||
1293 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1295 rc->qpm += step_size;
1296 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1299 while( rc->qpm > qp_min
1300 && (rc->qpm > h->fdec->f_row_qp[0] || rc->single_frame_vbv)
1301 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
1302 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
1304 rc->qpm -= step_size;
1305 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1308 /* avoid VBV underflow or MinCR violation */
1309 while( (rc->qpm < qp_absolute_max)
1310 && ((rc->buffer_fill - b1 < rc->buffer_rate * rc->max_frame_error) ||
1311 (rc->frame_size_maximum - b1 < rc->frame_size_maximum * rc->max_frame_error)))
1313 rc->qpm += step_size;
1314 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1317 h->rc->frame_size_estimated = predict_row_size_sum( h, y, rc->qpm );
1321 int x264_ratecontrol_qp( x264_t *h )
1324 return x264_clip3( h->rc->qpm + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1327 int x264_ratecontrol_mb_qp( x264_t *h )
1330 float qp = h->rc->qpm;
1331 if( h->param.rc.i_aq_mode )
1332 /* MB-tree currently doesn't adjust quantizers in unreferenced frames. */
1333 qp += 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];
1334 return x264_clip3( qp + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1337 /* In 2pass, force the same frame types as in the 1st pass */
1338 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1340 x264_ratecontrol_t *rc = h->rc;
1341 if( h->param.rc.b_stat_read )
1343 if( frame_num >= rc->num_entries )
1345 /* We could try to initialize everything required for ABR and
1346 * adaptive B-frames, but that would be complicated.
1347 * So just calculate the average QP used so far. */
1348 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24
1349 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1350 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
1351 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 );
1352 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 );
1354 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1355 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1356 if( h->param.i_bframe_adaptive )
1357 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1359 for( int i = 0; i < h->param.i_threads; i++ )
1361 h->thread[i]->rc->b_abr = 0;
1362 h->thread[i]->rc->b_2pass = 0;
1363 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1364 h->thread[i]->param.rc.b_stat_read = 0;
1365 h->thread[i]->param.i_bframe_adaptive = 0;
1366 h->thread[i]->param.i_scenecut_threshold = 0;
1367 h->thread[i]->param.rc.b_mb_tree = 0;
1368 if( h->thread[i]->param.i_bframe > 1 )
1369 h->thread[i]->param.i_bframe = 1;
1371 return X264_TYPE_AUTO;
1373 return rc->entry[frame_num].frame_type;
1376 return X264_TYPE_AUTO;
1379 void x264_ratecontrol_set_weights( x264_t *h, x264_frame_t *frm )
1381 ratecontrol_entry_t *rce = &h->rc->entry[frm->i_frame];
1382 if( h->param.analyse.i_weighted_pred <= 0 )
1384 if( rce->i_weight_denom >= 0 )
1385 SET_WEIGHT( frm->weight[0][0], 1, rce->weight[0], rce->i_weight_denom, rce->weight[1] );
1388 /* After encoding one frame, save stats and update ratecontrol state */
1389 int x264_ratecontrol_end( x264_t *h, int bits, int *filler )
1391 x264_ratecontrol_t *rc = h->rc;
1392 const int *mbs = h->stat.frame.i_mb_count;
1396 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1397 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1398 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1399 for( int i = B_DIRECT; i < B_8x8; i++ )
1400 h->stat.frame.i_mb_count_p += mbs[i];
1402 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1403 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1405 if( h->param.rc.b_stat_write )
1407 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1408 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1409 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1410 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1411 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1412 char c_direct = h->mb.b_direct_auto_write ?
1413 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1414 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1416 if( fprintf( rc->p_stat_file_out,
1417 "in:%d out:%d type:%c dur:%d cpbdur:%d q:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c ref:",
1418 h->fenc->i_frame, h->i_frame,
1419 c_type, h->fenc->i_duration,
1420 h->fenc->i_cpb_duration, rc->qpa_rc,
1421 h->stat.frame.i_tex_bits,
1422 h->stat.frame.i_mv_bits,
1423 h->stat.frame.i_misc_bits,
1424 h->stat.frame.i_mb_count_i,
1425 h->stat.frame.i_mb_count_p,
1426 h->stat.frame.i_mb_count_skip,
1430 /* Only write information for reference reordering once. */
1431 int use_old_stats = h->param.rc.b_stat_read && rc->rce->refs > 1;
1432 for( int i = 0; i < (use_old_stats ? rc->rce->refs : h->i_ref0); i++ )
1434 int refcount = use_old_stats ? rc->rce->refcount[i]
1435 : h->param.b_interlaced ? h->stat.frame.i_mb_count_ref[0][i*2]
1436 + h->stat.frame.i_mb_count_ref[0][i*2+1]
1437 : h->stat.frame.i_mb_count_ref[0][i];
1438 if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1442 if( h->sh.weight[0][0].weightfn )
1444 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 )
1448 if( fprintf( rc->p_stat_file_out, ";\n") < 0 )
1451 /* Don't re-write the data in multi-pass mode. */
1452 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1454 uint8_t i_type = h->sh.i_type;
1455 /* Values are stored as big-endian FIX8.8 */
1456 for( int i = 0; i < h->mb.i_mb_count; i++ )
1457 rc->qp_buffer[0][i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1458 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1460 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 )
1467 if( h->sh.i_type != SLICE_TYPE_B )
1468 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / rc->last_rceq;
1471 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1472 * Not perfectly accurate with B-refs, but good enough. */
1473 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / (rc->last_rceq * fabs( h->param.rc.f_pb_factor ));
1475 rc->cplxr_sum *= rc->cbr_decay;
1476 double frame_duration = (double)h->fenc->i_duration * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1478 rc->wanted_bits_window += frame_duration * rc->bitrate;
1479 rc->wanted_bits_window *= rc->cbr_decay;
1483 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale( rc->rce->new_qp ) );
1485 if( h->mb.b_variable_qp )
1487 if( h->sh.i_type == SLICE_TYPE_B )
1489 rc->bframe_bits += bits;
1490 if( h->fenc->b_last_minigop_bframe )
1492 update_predictor( rc->pred_b_from_p, qp2qscale( rc->qpa_rc ),
1493 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1494 rc->bframe_bits = 0;
1499 *filler = update_vbv( h, bits );
1501 if( h->sps->vui.b_nal_hrd_parameters_present )
1503 if( h->fenc->i_frame == 0 )
1505 // access unit initialises the HRD
1506 h->fenc->hrd_timing.cpb_initial_arrival_time = 0;
1507 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1508 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1509 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit = (double)rc->initial_cpb_removal_delay / 90000;
1513 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit + (double)h->fenc->i_cpb_delay *
1514 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1516 double cpb_earliest_arrival_time = h->fenc->hrd_timing.cpb_removal_time - (double)rc->initial_cpb_removal_delay / 90000;
1517 if( h->fenc->b_keyframe )
1519 rc->nrt_first_access_unit = h->fenc->hrd_timing.cpb_removal_time;
1520 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1521 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1524 cpb_earliest_arrival_time -= (double)rc->initial_cpb_removal_delay_offset / 90000;
1526 if( h->sps->vui.hrd.b_cbr_hrd )
1527 h->fenc->hrd_timing.cpb_initial_arrival_time = rc->previous_cpb_final_arrival_time;
1529 h->fenc->hrd_timing.cpb_initial_arrival_time = X264_MAX( rc->previous_cpb_final_arrival_time, cpb_earliest_arrival_time );
1531 int filler_bits = *filler ? X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), *filler )*8 : 0;
1533 h->fenc->hrd_timing.cpb_final_arrival_time = rc->previous_cpb_final_arrival_time = h->fenc->hrd_timing.cpb_initial_arrival_time +
1534 (double)(bits + filler_bits) / h->sps->vui.hrd.i_bit_rate_unscaled;
1536 h->fenc->hrd_timing.dpb_output_time = (double)h->fenc->i_dpb_output_delay * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale +
1537 h->fenc->hrd_timing.cpb_removal_time;
1542 x264_log(h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n");
1546 /****************************************************************************
1548 ***************************************************************************/
1551 * modify the bitrate curve from pass1 for one frame
1553 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1555 x264_ratecontrol_t *rcc= h->rc;
1556 x264_zone_t *zone = get_zone( h, frame_num );
1557 double q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1559 // avoid NaN's in the rc_eq
1560 if( !isfinite(q) || rce->tex_bits + rce->mv_bits == 0 )
1561 q = rcc->last_qscale_for[rce->pict_type];
1566 rcc->last_qscale = q;
1571 if( zone->b_force_qp )
1572 q = qp2qscale( zone->i_qp );
1574 q /= zone->f_bitrate_factor;
1580 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1582 x264_ratecontrol_t *rcc = h->rc;
1583 const int pict_type = rce->pict_type;
1585 // force I/B quants as a function of P quants
1586 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1587 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1588 if( pict_type == SLICE_TYPE_I )
1591 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1592 double ip_factor = fabs( h->param.rc.f_ip_factor );
1593 /* don't apply ip_factor if the following frame is also I */
1594 if( rcc->accum_p_norm <= 0 )
1596 else if( h->param.rc.f_ip_factor < 0 )
1598 else if( rcc->accum_p_norm >= 1 )
1601 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1603 else if( pict_type == SLICE_TYPE_B )
1605 if( h->param.rc.f_pb_factor > 0 )
1607 if( !rce->kept_as_ref )
1608 q *= fabs( h->param.rc.f_pb_factor );
1610 else if( pict_type == SLICE_TYPE_P
1611 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1612 && rce->tex_bits == 0 )
1617 /* last qscale / qdiff stuff */
1618 if( rcc->last_non_b_pict_type == pict_type &&
1619 (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1) )
1621 double last_q = rcc->last_qscale_for[pict_type];
1622 double max_qscale = last_q * rcc->lstep;
1623 double min_qscale = last_q / rcc->lstep;
1625 if ( q > max_qscale ) q = max_qscale;
1626 else if( q < min_qscale ) q = min_qscale;
1629 rcc->last_qscale_for[pict_type] = q;
1630 if( pict_type != SLICE_TYPE_B )
1631 rcc->last_non_b_pict_type = pict_type;
1632 if( pict_type == SLICE_TYPE_I )
1634 rcc->last_accum_p_norm = rcc->accum_p_norm;
1635 rcc->accum_p_norm = 0;
1636 rcc->accum_p_qp = 0;
1638 if( pict_type == SLICE_TYPE_P )
1640 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1641 rcc->accum_p_qp = mask * (qscale2qp( q ) + rcc->accum_p_qp);
1642 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1647 static double predict_size( predictor_t *p, double q, double var )
1649 return (p->coeff*var + p->offset) / (q*p->count);
1652 static void update_predictor( predictor_t *p, double q, double var, double bits )
1654 const double range = 1.5;
1657 double old_coeff = p->coeff / p->count;
1658 double new_coeff = bits*q / var;
1659 double new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
1660 double new_offset = bits*q - new_coeff_clipped * var;
1661 if( new_offset >= 0 )
1662 new_coeff = new_coeff_clipped;
1665 p->count *= p->decay;
1666 p->coeff *= p->decay;
1667 p->offset *= p->decay;
1669 p->coeff += new_coeff;
1670 p->offset += new_offset;
1673 // update VBV after encoding a frame
1674 static int update_vbv( x264_t *h, int bits )
1678 x264_ratecontrol_t *rcc = h->rc;
1679 x264_ratecontrol_t *rct = h->thread[0]->rc;
1681 if( rcc->last_satd >= h->mb.i_mb_count )
1682 update_predictor( &rct->pred[h->sh.i_type], qp2qscale( rcc->qpa_rc ), rcc->last_satd, bits );
1687 rct->buffer_fill_final -= bits;
1689 if( rct->buffer_fill_final < 0 )
1690 x264_log( h, X264_LOG_WARNING, "VBV underflow (frame %d, %.0f bits)\n", h->i_frame, rct->buffer_fill_final );
1691 rct->buffer_fill_final = X264_MAX( rct->buffer_fill_final, 0 );
1692 rct->buffer_fill_final += rcc->buffer_rate;
1694 if( h->sps->vui.hrd.b_cbr_hrd && rct->buffer_fill_final > rcc->buffer_size )
1696 filler = ceil( (rct->buffer_fill_final - rcc->buffer_size) / 8 );
1697 rct->buffer_fill_final -= X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), filler ) * 8;
1700 rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, rcc->buffer_size );
1705 int x264_hrd_fullness( x264_t *h )
1707 x264_ratecontrol_t *rct = h->thread[0]->rc;
1708 double cpb_bits = rct->buffer_fill_final;
1709 double bps = h->sps->vui.hrd.i_bit_rate_unscaled;
1710 double cpb_size = h->sps->vui.hrd.i_cpb_size_unscaled;
1711 double cpb_fullness = 90000.0*cpb_bits/bps;
1713 if( cpb_bits < 0 || cpb_bits > cpb_size )
1715 x264_log( h, X264_LOG_WARNING, "CPB %s: %.0lf bits in a %.0lf-bit buffer\n",
1716 cpb_bits < 0 ? "underflow" : "overflow", cpb_bits, cpb_size );
1719 h->initial_cpb_removal_delay_offset = 90000.0*(cpb_size - cpb_bits)/bps;
1721 return x264_clip3f( cpb_fullness + 0.5, 0, 90000.0*cpb_size/bps ); // just lie if we are in a weird state
1724 // provisionally update VBV according to the planned size of all frames currently in progress
1725 static void update_vbv_plan( x264_t *h, int overhead )
1727 x264_ratecontrol_t *rcc = h->rc;
1728 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final;
1729 if( h->i_thread_frames > 1 )
1731 int j = h->rc - h->thread[0]->rc;
1732 for( int i = 1; i < h->i_thread_frames; i++ )
1734 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
1735 double bits = t->rc->frame_size_planned;
1736 if( !t->b_thread_active )
1738 bits = X264_MAX(bits, t->rc->frame_size_estimated);
1739 rcc->buffer_fill -= bits;
1740 rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
1741 rcc->buffer_fill += t->rc->buffer_rate;
1742 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1745 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1746 rcc->buffer_fill -= overhead;
1749 // apply VBV constraints and clip qscale to between lmin and lmax
1750 static double clip_qscale( x264_t *h, int pict_type, double q )
1752 x264_ratecontrol_t *rcc = h->rc;
1753 double lmin = rcc->lmin[pict_type];
1754 double lmax = rcc->lmax[pict_type];
1755 if( rcc->rate_factor_max_increment )
1756 lmax = X264_MIN( lmax, qp2qscale( rcc->qp_novbv + rcc->rate_factor_max_increment ) );
1759 /* B-frames are not directly subject to VBV,
1760 * since they are controlled by the P-frames' QPs. */
1762 if( rcc->b_vbv && rcc->last_satd > 0 )
1764 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
1765 * the lookahead overflow and such that the buffer is in a reasonable state
1766 * by the end of the lookahead. */
1767 if( h->param.rc.i_lookahead )
1771 /* Avoid an infinite loop. */
1772 for( int iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
1775 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1776 double buffer_fill_cur = rcc->buffer_fill - cur_bits;
1778 double total_duration = 0;
1779 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
1780 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
1781 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
1783 /* Loop over the planned future frames. */
1784 for( int j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
1786 total_duration += h->fenc->f_planned_cpb_duration[j];
1787 buffer_fill_cur += rcc->vbv_max_rate * h->fenc->f_planned_cpb_duration[j];
1788 int i_type = h->fenc->i_planned_type[j];
1789 int i_satd = h->fenc->i_planned_satd[j];
1790 if( i_type == X264_TYPE_AUTO )
1792 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
1793 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
1794 buffer_fill_cur -= cur_bits;
1796 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
1797 target_fill = X264_MIN( rcc->buffer_fill + total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.5 );
1798 if( buffer_fill_cur < target_fill )
1804 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
1805 target_fill = x264_clip3f( rcc->buffer_fill - total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
1806 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
1815 /* Fallback to old purely-reactive algorithm: no lookahead. */
1818 if( ( pict_type == SLICE_TYPE_P ||
1819 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
1820 rcc->buffer_fill/rcc->buffer_size < 0.5 )
1822 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1825 /* Now a hard threshold to make sure the frame fits in VBV.
1826 * This one is mostly for I-frames. */
1827 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1829 /* For small VBVs, allow the frame to use up the entire VBV. */
1830 double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
1831 /* For single-frame VBVs, request that the frame use up the entire VBV. */
1832 double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
1834 if( bits > rcc->buffer_fill/max_fill_factor )
1835 qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
1838 if( bits < rcc->buffer_rate/min_fill_factor )
1839 q *= bits*min_fill_factor/rcc->buffer_rate;
1840 q = X264_MAX( q0, q );
1843 /* Apply MinCR restrictions */
1844 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1845 if( bits > rcc->frame_size_maximum )
1846 q *= bits / rcc->frame_size_maximum;
1847 bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1849 /* Check B-frame complexity, and use up any bits that would
1850 * overflow before the next P-frame. */
1851 if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
1853 int nb = rcc->bframes;
1854 double pbbits = bits;
1855 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1857 double bframe_cpb_duration = 0;
1858 double minigop_cpb_duration;
1859 for( int i = 0; i < nb; i++ )
1860 bframe_cpb_duration += h->fenc->f_planned_cpb_duration[1+i];
1862 if( bbits * nb > bframe_cpb_duration * rcc->vbv_max_rate )
1864 pbbits += nb * bbits;
1866 minigop_cpb_duration = bframe_cpb_duration + h->fenc->f_planned_cpb_duration[0];
1867 space = rcc->buffer_fill + minigop_cpb_duration*rcc->vbv_max_rate - rcc->buffer_size;
1868 if( pbbits < space )
1870 q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
1872 q = X264_MAX( q0-5, q );
1875 if( !rcc->b_vbv_min_rate )
1876 q = X264_MAX( q0, q );
1881 else if( rcc->b_2pass )
1883 double min2 = log( lmin );
1884 double max2 = log( lmax );
1885 q = (log(q) - min2)/(max2-min2) - 0.5;
1886 q = 1.0/(1.0 + exp( -4*q ));
1887 q = q*(max2-min2) + min2;
1891 return x264_clip3f( q, lmin, lmax );
1894 // update qscale for 1 frame based on actual bits used so far
1895 static float rate_estimate_qscale( x264_t *h )
1898 x264_ratecontrol_t *rcc = h->rc;
1899 ratecontrol_entry_t rce;
1900 int pict_type = h->sh.i_type;
1901 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
1902 + h->stat.i_frame_size[SLICE_TYPE_P]
1903 + h->stat.i_frame_size[SLICE_TYPE_B]);
1908 if( pict_type != rce.pict_type )
1910 x264_log( h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1911 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type] );
1915 if( pict_type == SLICE_TYPE_B )
1917 /* B-frames don't have independent ratecontrol, but rather get the
1918 * average QP of the two adjacent P-frames + an offset */
1920 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1921 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1922 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1923 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1924 float q0 = h->fref0[0]->f_qp_avg_rc;
1925 float q1 = h->fref1[0]->f_qp_avg_rc;
1927 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1928 q0 -= rcc->pb_offset/2;
1929 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1930 q1 -= rcc->pb_offset/2;
1933 q = (q0 + q1) / 2 + rcc->ip_offset;
1939 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1941 if( h->fenc->b_kept_as_ref )
1942 q += rcc->pb_offset/2;
1944 q += rcc->pb_offset;
1946 if( rcc->b_2pass && rcc->b_vbv )
1947 rcc->frame_size_planned = qscale2bits( &rce, q );
1949 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1950 h->rc->frame_size_estimated = rcc->frame_size_planned;
1954 rcc->last_satd = x264_rc_analyse_slice( h );
1956 return qp2qscale( q );
1960 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1964 double lmin = rcc->lmin[pict_type];
1965 double lmax = rcc->lmax[pict_type];
1967 int64_t predicted_bits = total_bits;
1968 /* Adjust ABR buffer based on distance to the end of the video. */
1969 if( rcc->num_entries > h->fenc->i_frame )
1970 abr_buffer *= 0.5 * sqrt( rcc->num_entries - h->fenc->i_frame );
1974 if( h->i_thread_frames > 1 )
1976 int j = h->rc - h->thread[0]->rc;
1977 for( int i = 1; i < h->i_thread_frames; i++ )
1979 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
1980 double bits = t->rc->frame_size_planned;
1981 if( !t->b_thread_active )
1983 bits = X264_MAX(bits, t->rc->frame_size_estimated);
1984 predicted_bits += (int64_t)bits;
1990 if( h->fenc->i_frame < h->i_thread_frames )
1991 predicted_bits += (int64_t)h->fenc->i_frame * rcc->bitrate / rcc->fps;
1993 predicted_bits += (int64_t)(h->i_thread_frames - 1) * rcc->bitrate / rcc->fps;
1996 diff = predicted_bits - (int64_t)rce.expected_bits;
1998 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
1999 if( ((h->fenc->i_frame + 1 - h->i_thread_frames) >= rcc->fps) &&
2000 (rcc->expected_bits_sum > 0))
2002 /* Adjust quant based on the difference between
2003 * achieved and expected bitrate so far */
2004 double cur_time = (double)h->fenc->i_frame / rcc->num_entries;
2005 double w = x264_clip3f( cur_time*100, 0.0, 1.0 );
2006 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
2010 /* Do not overflow vbv */
2011 double expected_size = qscale2bits( &rce, q );
2012 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2013 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
2014 double qmax = q*(2 - expected_fullness);
2015 double size_constraint = 1 + expected_fullness;
2016 qmax = X264_MAX( qmax, rce.new_qscale );
2017 if( expected_fullness < .05 )
2019 qmax = X264_MIN(qmax, lmax);
2020 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
2021 ((expected_vbv < 0) && (q < lmax)))
2024 expected_size = qscale2bits(&rce, q);
2025 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2027 rcc->last_satd = x264_rc_analyse_slice( h );
2029 q = x264_clip3f( q, lmin, lmax );
2031 else /* 1pass ABR */
2033 /* Calculate the quantizer which would have produced the desired
2034 * average bitrate if it had been applied to all frames so far.
2035 * Then modulate that quant based on the current frame's complexity
2036 * relative to the average complexity so far (using the 2pass RCEQ).
2037 * Then bias the quant up or down if total size so far was far from
2039 * Result: Depending on the value of rate_tolerance, there is a
2040 * tradeoff between quality and bitrate precision. But at large
2041 * tolerances, the bit distribution approaches that of 2pass. */
2043 double wanted_bits, overflow = 1;
2045 rcc->last_satd = x264_rc_analyse_slice( h );
2046 rcc->short_term_cplxsum *= 0.5;
2047 rcc->short_term_cplxcount *= 0.5;
2048 rcc->short_term_cplxsum += rcc->last_satd;
2049 rcc->short_term_cplxcount ++;
2051 rce.tex_bits = rcc->last_satd;
2052 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
2054 rce.p_count = rcc->nmb;
2058 rce.pict_type = pict_type;
2060 if( h->param.rc.i_rc_method == X264_RC_CRF )
2062 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
2066 int i_frame_done = h->fenc->i_frame + 1 - h->i_thread_frames;
2067 double i_time_done = i_frame_done / rcc->fps;
2068 if( h->param.b_vfr_input )
2069 i_time_done = ((double)(h->fenc->i_reordered_pts - h->first_pts)) * h->param.i_timebase_num / h->param.i_timebase_den;
2071 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
2073 /* ABR code can potentially be counterproductive in CBR, so just don't bother.
2074 * Don't run it if the frame complexity is zero either. */
2075 if( !rcc->b_vbv_min_rate && rcc->last_satd )
2077 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
2078 wanted_bits = i_time_done * rcc->bitrate;
2079 if( wanted_bits > 0 )
2081 abr_buffer *= X264_MAX( 1, sqrt(i_time_done) );
2082 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
2088 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
2089 /* should test _next_ pict type, but that isn't decided yet */
2090 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
2092 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2093 q /= fabs( h->param.rc.f_ip_factor );
2095 else if( h->i_frame > 0 )
2097 /* Asymmetric clipping, because symmetric would prevent
2098 * overflow control in areas of rapidly oscillating complexity */
2099 double lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
2100 double lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
2101 if( overflow > 1.1 && h->i_frame > 3 )
2103 else if( overflow < 0.9 )
2106 q = x264_clip3f(q, lmin, lmax);
2108 else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
2110 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
2112 rcc->qp_novbv = qscale2qp( q );
2114 //FIXME use get_diff_limited_q() ?
2115 q = clip_qscale( h, pict_type, q );
2118 rcc->last_qscale_for[pict_type] =
2119 rcc->last_qscale = q;
2121 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
2122 rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );
2124 if( rcc->b_2pass && rcc->b_vbv )
2125 rcc->frame_size_planned = qscale2bits(&rce, q);
2127 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2129 /* Always use up the whole VBV in this case. */
2130 if( rcc->single_frame_vbv )
2131 rcc->frame_size_planned = rcc->buffer_rate;
2132 h->rc->frame_size_estimated = rcc->frame_size_planned;
2137 void x264_threads_normalize_predictors( x264_t *h )
2139 double totalsize = 0;
2140 for( int i = 0; i < h->param.i_threads; i++ )
2141 totalsize += h->thread[i]->rc->slice_size_planned;
2142 double factor = h->rc->frame_size_planned / totalsize;
2143 for( int i = 0; i < h->param.i_threads; i++ )
2144 h->thread[i]->rc->slice_size_planned *= factor;
2147 void x264_threads_distribute_ratecontrol( x264_t *h )
2150 x264_ratecontrol_t *rc = h->rc;
2152 /* Initialize row predictors */
2153 if( h->i_frame == 0 )
2154 for( int i = 0; i < h->param.i_threads; i++ )
2156 x264_ratecontrol_t *t = h->thread[i]->rc;
2157 memcpy( t->row_preds, rc->row_preds, sizeof(rc->row_preds) );
2160 for( int i = 0; i < h->param.i_threads; i++ )
2162 x264_t *t = h->thread[i];
2163 memcpy( t->rc, rc, offsetof(x264_ratecontrol_t, row_pred) );
2164 t->rc->row_pred = &t->rc->row_preds[h->sh.i_type];
2165 /* Calculate the planned slice size. */
2166 if( rc->b_vbv && rc->frame_size_planned )
2169 for( row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2170 size += h->fdec->i_row_satd[row];
2171 t->rc->slice_size_planned = predict_size( &rc->pred[h->sh.i_type + (i+1)*5], rc->qpm, size );
2174 t->rc->slice_size_planned = 0;
2176 if( rc->b_vbv && rc->frame_size_planned )
2178 x264_threads_normalize_predictors( h );
2180 if( rc->single_frame_vbv )
2182 /* Compensate for our max frame error threshold: give more bits (proportionally) to smaller slices. */
2183 for( int i = 0; i < h->param.i_threads; i++ )
2185 x264_t *t = h->thread[i];
2186 t->rc->max_frame_error = X264_MAX( 0.05, 1.0 / (t->i_threadslice_end - t->i_threadslice_start) );
2187 t->rc->slice_size_planned += 2 * t->rc->max_frame_error * rc->frame_size_planned;
2189 x264_threads_normalize_predictors( h );
2192 for( int i = 0; i < h->param.i_threads; i++ )
2193 h->thread[i]->rc->frame_size_estimated = h->thread[i]->rc->slice_size_planned;
2197 void x264_threads_merge_ratecontrol( x264_t *h )
2199 x264_ratecontrol_t *rc = h->rc;
2202 for( int i = 0; i < h->param.i_threads; i++ )
2204 x264_t *t = h->thread[i];
2205 x264_ratecontrol_t *rct = h->thread[i]->rc;
2206 if( h->param.rc.i_vbv_buffer_size )
2209 for( int row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2210 size += h->fdec->i_row_satd[row];
2211 int bits = t->stat.frame.i_mv_bits + t->stat.frame.i_tex_bits + t->stat.frame.i_misc_bits;
2212 int mb_count = (t->i_threadslice_end - t->i_threadslice_start) * h->sps->i_mb_width;
2213 update_predictor( &rc->pred[h->sh.i_type+5*i], qp2qscale( rct->qpa_rc/mb_count ), size, bits );
2217 rc->qpa_rc += rct->qpa_rc;
2218 rc->qpa_aq += rct->qpa_aq;
2222 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
2226 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
2227 /* these vars are updated in x264_ratecontrol_start()
2228 * so copy them from the context that most recently started (prev)
2229 * to the context that's about to start (cur). */
2234 COPY(last_qscale_for);
2235 COPY(last_non_b_pict_type);
2236 COPY(short_term_cplxsum);
2237 COPY(short_term_cplxcount);
2241 /* these vars can be updated by x264_ratecontrol_init_reconfigurable */
2244 COPY(single_frame_vbv);
2246 COPY(b_vbv_min_rate);
2247 COPY(rate_factor_constant);
2253 #define COPY(var) next->rc->var = cur->rc->var
2254 /* these vars are updated in x264_ratecontrol_end()
2255 * so copy them from the context that most recently ended (cur)
2256 * to the context that's about to end (next) */
2258 COPY(expected_bits_sum);
2259 COPY(wanted_bits_window);
2261 COPY(initial_cpb_removal_delay);
2262 COPY(initial_cpb_removal_delay_offset);
2263 COPY(nrt_first_access_unit);
2264 COPY(previous_cpb_final_arrival_time);
2267 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
2268 /* the rest of the variables are either constant or thread-local */
2271 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
2273 /* find an interval ending on an overflow or underflow (depending on whether
2274 * we're adding or removing bits), and starting on the earliest frame that
2275 * can influence the buffer fill of that end frame. */
2276 x264_ratecontrol_t *rcc = h->rc;
2277 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
2278 const double buffer_max = .9 * rcc->buffer_size;
2279 double fill = fills[*t0-1];
2280 double parity = over ? 1. : -1.;
2281 int start = -1, end = -1;
2282 for( int i = *t0; i < rcc->num_entries; i++ )
2284 fill += (rcc->entry[i].i_cpb_duration * rcc->vbv_max_rate * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale -
2285 qscale2bits( &rcc->entry[i], rcc->entry[i].new_qscale )) * parity;
2286 fill = x264_clip3f(fill, 0, rcc->buffer_size);
2288 if( fill <= buffer_min || i == 0 )
2294 else if( fill >= buffer_max && start >= 0 )
2299 return start >= 0 && end >= 0;
2302 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
2304 x264_ratecontrol_t *rcc = h->rc;
2305 double qscale_orig, qscale_new;
2309 for( int i = t0; i <= t1; i++ )
2311 qscale_orig = rcc->entry[i].new_qscale;
2312 qscale_orig = x264_clip3f( qscale_orig, qscale_min, qscale_max );
2313 qscale_new = qscale_orig * adjustment;
2314 qscale_new = x264_clip3f( qscale_new, qscale_min, qscale_max );
2315 rcc->entry[i].new_qscale = qscale_new;
2316 adjusted = adjusted || (qscale_new != qscale_orig);
2321 static double count_expected_bits( x264_t *h )
2323 x264_ratecontrol_t *rcc = h->rc;
2324 double expected_bits = 0;
2325 for( int i = 0; i < rcc->num_entries; i++ )
2327 ratecontrol_entry_t *rce = &rcc->entry[i];
2328 rce->expected_bits = expected_bits;
2329 expected_bits += qscale2bits( rce, rce->new_qscale );
2331 return expected_bits;
2334 static int vbv_pass2( x264_t *h, double all_available_bits )
2336 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
2337 * frames in the interval until either buffer is full at some intermediate frame or the
2338 * last frame in the interval no longer underflows. Recompute intervals and repeat.
2339 * Then do the converse to put bits back into overflow areas until target size is met */
2341 x264_ratecontrol_t *rcc = h->rc;
2343 double expected_bits = 0;
2345 double prev_bits = 0;
2347 double qscale_min = qp2qscale( h->param.rc.i_qp_min );
2348 double qscale_max = qp2qscale( h->param.rc.i_qp_max );
2350 int adj_min, adj_max;
2351 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2355 /* adjust overall stream size */
2359 prev_bits = expected_bits;
2362 { /* not first iteration */
2363 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2364 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2368 while(adj_min && find_underflow( h, fills, &t0, &t1, 1 ))
2370 adj_min = fix_underflow( h, t0, t1, adjustment, qscale_min, qscale_max );
2375 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2377 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2379 while( adj_max && find_underflow( h, fills, &t0, &t1, 0 ) )
2380 adj_max = fix_underflow( h, t0, t1, 1.001, qscale_min, qscale_max );
2382 expected_bits = count_expected_bits( h );
2383 } while( (expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2386 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2388 /* store expected vbv filling values for tracking when encoding */
2389 for( int i = 0; i < rcc->num_entries; i++ )
2390 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
2392 x264_free( fills-1 );
2398 static int init_pass2( x264_t *h )
2400 x264_ratecontrol_t *rcc = h->rc;
2401 uint64_t all_const_bits = 0;
2402 double duration = 0;
2403 for( int i = 0; i < rcc->num_entries; i++ )
2404 duration += rcc->entry[i].i_duration;
2405 duration *= (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2406 uint64_t all_available_bits = h->param.rc.i_bitrate * 1000. * duration;
2407 double rate_factor, step_mult;
2408 double qblur = h->param.rc.f_qblur;
2409 double cplxblur = h->param.rc.f_complexity_blur;
2410 const int filter_size = (int)(qblur*4) | 1;
2411 double expected_bits;
2412 double *qscale, *blurred_qscale;
2414 /* find total/average complexity & const_bits */
2415 for( int i = 0; i < rcc->num_entries; i++ )
2417 ratecontrol_entry_t *rce = &rcc->entry[i];
2418 all_const_bits += rce->misc_bits;
2421 if( all_available_bits < all_const_bits)
2423 x264_log( h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2424 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)) );
2428 /* Blur complexities, to reduce local fluctuation of QP.
2429 * We don't blur the QPs directly, because then one very simple frame
2430 * could drag down the QP of a nearby complex frame and give it more
2431 * bits than intended. */
2432 for( int i = 0; i < rcc->num_entries; i++ )
2434 ratecontrol_entry_t *rce = &rcc->entry[i];
2435 double weight_sum = 0;
2436 double cplx_sum = 0;
2437 double weight = 1.0;
2438 double gaussian_weight;
2439 /* weighted average of cplx of future frames */
2440 for( int j = 1; j < cplxblur*2 && j < rcc->num_entries-i; j++ )
2442 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2443 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2444 if( weight < .0001 )
2446 gaussian_weight = weight * exp( -j*j/200.0 );
2447 weight_sum += gaussian_weight;
2448 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2450 /* weighted average of cplx of past frames */
2452 for( int j = 0; j <= cplxblur*2 && j <= i; j++ )
2454 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2455 gaussian_weight = weight * exp( -j*j/200.0 );
2456 weight_sum += gaussian_weight;
2457 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits);
2458 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2459 if( weight < .0001 )
2462 rce->blurred_complexity = cplx_sum / weight_sum;
2465 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2466 if( filter_size > 1 )
2467 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2469 blurred_qscale = qscale;
2471 /* Search for a factor which, when multiplied by the RCEQ values from
2472 * each frame, adds up to the desired total size.
2473 * There is no exact closed-form solution because of VBV constraints and
2474 * because qscale2bits is not invertible, but we can start with the simple
2475 * approximation of scaling the 1st pass by the ratio of bitrates.
2476 * The search range is probably overkill, but speed doesn't matter here. */
2479 for( int i = 0; i < rcc->num_entries; i++ )
2481 double q = get_qscale(h, &rcc->entry[i], 1.0, i);
2482 expected_bits += qscale2bits(&rcc->entry[i], q);
2483 rcc->last_qscale_for[rcc->entry[i].pict_type] = q;
2485 step_mult = all_available_bits / expected_bits;
2488 for( double step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2491 rate_factor += step;
2493 rcc->last_non_b_pict_type = -1;
2494 rcc->last_accum_p_norm = 1;
2495 rcc->accum_p_norm = 0;
2498 for( int i = 0; i < rcc->num_entries; i++ )
2500 qscale[i] = get_qscale( h, &rcc->entry[i], rate_factor, i );
2501 rcc->last_qscale_for[rcc->entry[i].pict_type] = qscale[i];
2504 /* fixed I/B qscale relative to P */
2505 for( int i = rcc->num_entries-1; i >= 0; i-- )
2507 qscale[i] = get_diff_limited_q( h, &rcc->entry[i], qscale[i] );
2508 assert(qscale[i] >= 0);
2512 if( filter_size > 1 )
2514 assert( filter_size%2 == 1 );
2515 for( int i = 0; i < rcc->num_entries; i++ )
2517 ratecontrol_entry_t *rce = &rcc->entry[i];
2518 double q = 0.0, sum = 0.0;
2520 for( int j = 0; j < filter_size; j++ )
2522 int index = i+j-filter_size/2;
2524 double coeff = qblur==0 ? 1.0 : exp( -d*d/(qblur*qblur) );
2525 if( index < 0 || index >= rcc->num_entries )
2527 if( rce->pict_type != rcc->entry[index].pict_type )
2529 q += qscale[index] * coeff;
2532 blurred_qscale[i] = q/sum;
2536 /* find expected bits */
2537 for( int i = 0; i < rcc->num_entries; i++ )
2539 ratecontrol_entry_t *rce = &rcc->entry[i];
2540 rce->new_qscale = clip_qscale( h, rce->pict_type, blurred_qscale[i] );
2541 assert(rce->new_qscale >= 0);
2542 expected_bits += qscale2bits( rce, rce->new_qscale );
2545 if( expected_bits > all_available_bits )
2546 rate_factor -= step;
2549 x264_free( qscale );
2550 if( filter_size > 1 )
2551 x264_free( blurred_qscale );
2554 if( vbv_pass2( h, all_available_bits ) )
2556 expected_bits = count_expected_bits( h );
2558 if( fabs( expected_bits/all_available_bits - 1.0 ) > 0.01 )
2561 for( int i = 0; i < rcc->num_entries; i++ )
2562 avgq += rcc->entry[i].new_qscale;
2563 avgq = qscale2qp( avgq / rcc->num_entries );
2565 if( expected_bits > all_available_bits || !rcc->b_vbv )
2566 x264_log( h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n" );
2567 x264_log( h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
2568 (float)h->param.rc.i_bitrate,
2569 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
2571 if( expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2 )
2573 if( h->param.rc.i_qp_min > 0 )
2574 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min );
2576 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate\n" );
2578 else if( expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2 )
2580 if( h->param.rc.i_qp_max < 51 )
2581 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max );
2583 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate\n");
2585 else if( !(rcc->b_2pass && rcc->b_vbv) )
2586 x264_log( h, X264_LOG_WARNING, "internal error\n" );