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 int64_t buffer_fill_final;
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[3]; /* 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[3]; /* 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 predictor_t (*row_pred)[2];
145 predictor_t row_preds[3][2];
146 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
147 int bframes; /* # consecutive B-frames before this P-frame */
148 int bframe_bits; /* total cost of those frames */
152 x264_zone_t *prev_zone;
155 int initial_cpb_removal_delay;
156 int initial_cpb_removal_delay_offset;
157 double nrt_first_access_unit; /* nominal removal time */
158 double previous_cpb_final_arrival_time;
159 uint64_t hrd_multiply_denom;
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_var( uint64_t sum_ssd, int shift, x264_frame_t *frame, int i )
208 uint32_t sum = sum_ssd;
209 uint32_t ssd = sum_ssd >> 32;
210 frame->i_pixel_sum[i] += sum;
211 frame->i_pixel_ssd[i] += ssd;
212 return ssd - ((uint64_t)sum * sum >> shift);
215 static ALWAYS_INLINE uint32_t ac_energy_plane( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame, int i )
218 int stride = frame->i_stride[i];
219 int offset = h->mb.b_interlaced
220 ? 16 * mb_x + w * (mb_y&~1) * stride + (mb_y&1) * stride
221 : 16 * mb_x + w * mb_y * stride;
222 stride <<= h->mb.b_interlaced;
225 ALIGNED_ARRAY_16( pixel, pix,[FENC_STRIDE*8] );
226 h->mc.load_deinterleave_8x8x2_fenc( pix, frame->plane[1] + offset, stride );
227 return ac_energy_var( h->pixf.var[PIXEL_8x8]( pix, FENC_STRIDE ), 6, frame, i )
228 + ac_energy_var( h->pixf.var[PIXEL_8x8]( pix+FENC_STRIDE/2, FENC_STRIDE ), 6, frame, i );
231 return ac_energy_var( h->pixf.var[PIXEL_16x16]( frame->plane[0] + offset, stride ), 8, frame, i );
234 // Find the total AC energy of the block in all planes.
235 static NOINLINE uint32_t x264_ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
237 /* This function contains annoying hacks because GCC has a habit of reordering emms
238 * and putting it after floating point ops. As a result, we put the emms at the end of the
239 * function and make sure that its always called before the float math. Noinline makes
240 * sure no reordering goes on. */
241 uint32_t var = ac_energy_plane( h, mb_x, mb_y, frame, 0 );
242 var += ac_energy_plane( h, mb_x, mb_y, frame, 1 );
247 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame, float *quant_offsets )
249 /* constants chosen to result in approximately the same overall bitrate as without AQ.
250 * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
253 /* Initialize frame stats */
254 for( int i = 0; i < 3; i++ )
256 frame->i_pixel_sum[i] = 0;
257 frame->i_pixel_ssd[i] = 0;
260 /* Degenerate cases */
261 if( h->param.rc.i_aq_mode == X264_AQ_NONE || h->param.rc.f_aq_strength == 0 )
263 /* Need to init it anyways for MB tree */
264 if( h->param.rc.i_aq_mode && h->param.rc.f_aq_strength == 0 )
268 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
269 frame->f_qp_offset[mb_xy] = frame->f_qp_offset_aq[mb_xy] = quant_offsets[mb_xy];
270 if( h->frames.b_have_lowres )
271 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
272 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8( frame->f_qp_offset[mb_xy] );
276 memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
277 memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
278 if( h->frames.b_have_lowres )
279 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
280 frame->i_inv_qscale_factor[mb_xy] = 256;
283 /* Need variance data for weighted prediction */
284 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE || h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
286 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
287 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
288 x264_ac_energy_mb( h, mb_x, mb_y, frame );
293 /* Actual adaptive quantization */
296 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
298 float bit_depth_correction = powf(1 << (BIT_DEPTH-8), 0.5f);
299 float avg_adj_pow2 = 0.f;
300 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
301 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
303 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
304 float qp_adj = powf( energy + 1, 0.125f );
305 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
307 avg_adj_pow2 += qp_adj * qp_adj;
309 avg_adj /= h->mb.i_mb_count;
310 avg_adj_pow2 /= h->mb.i_mb_count;
311 strength = h->param.rc.f_aq_strength * avg_adj / bit_depth_correction;
312 avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - (14.f * bit_depth_correction)) / avg_adj;
315 strength = h->param.rc.f_aq_strength * 1.0397f;
317 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
318 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
321 int mb_xy = mb_x + mb_y*h->mb.i_mb_stride;
322 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
324 qp_adj = frame->f_qp_offset[mb_xy];
325 qp_adj = strength * (qp_adj - avg_adj);
329 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
330 qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - (14.427f + 2*(BIT_DEPTH-8)));
333 qp_adj += quant_offsets[mb_xy];
334 frame->f_qp_offset[mb_xy] =
335 frame->f_qp_offset_aq[mb_xy] = qp_adj;
336 if( h->frames.b_have_lowres )
337 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8(qp_adj);
341 /* Remove mean from SSD calculation */
342 for( int i = 0; i < 3; i++ )
344 uint64_t ssd = frame->i_pixel_ssd[i];
345 uint64_t sum = frame->i_pixel_sum[i];
346 int width = h->mb.i_mb_width*16>>!!i;
347 int height = h->mb.i_mb_height*16>>!!i;
348 frame->i_pixel_ssd[i] = ssd - (sum * sum + width * height / 2) / (width * height);
352 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame, float *quant_offsets )
354 x264_ratecontrol_t *rc = h->rc;
355 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
357 if( rc->entry[frame->i_frame].kept_as_ref )
360 if( rc->qpbuf_pos < 0 )
366 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
368 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 )
371 if( i_type != i_type_actual && rc->qpbuf_pos == 1 )
373 x264_log(h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type, i_type_actual);
376 } while( i_type != i_type_actual );
379 for( int i = 0; i < h->mb.i_mb_count; i++ )
381 frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[rc->qpbuf_pos][i] )) * (1/256.0);
382 if( h->frames.b_have_lowres )
383 frame->i_inv_qscale_factor[i] = x264_exp2fix8(frame->f_qp_offset[i]);
388 x264_stack_align( x264_adaptive_quant_frame, h, frame, quant_offsets );
391 x264_log(h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n");
395 int x264_reference_build_list_optimal( x264_t *h )
397 ratecontrol_entry_t *rce = h->rc->rce;
398 x264_frame_t *frames[16];
399 x264_weight_t weights[16][3];
402 if( rce->refs != h->i_ref0 )
405 memcpy( frames, h->fref0, sizeof(frames) );
406 memcpy( refcount, rce->refcount, sizeof(refcount) );
407 memcpy( weights, h->fenc->weight, sizeof(weights) );
408 memset( &h->fenc->weight[1][0], 0, sizeof(x264_weight_t[15][3]) );
410 /* For now don't reorder ref 0; it seems to lower quality
411 in most cases due to skips. */
412 for( int ref = 1; ref < h->i_ref0; ref++ )
417 for( int i = 1; i < h->i_ref0; i++ )
418 /* Favor lower POC as a tiebreaker. */
419 COPY2_IF_GT( max, refcount[i], bestref, i );
421 /* FIXME: If there are duplicates from frames other than ref0 then it is possible
422 * that the optimal ordering doesnt place every duplicate. */
424 refcount[bestref] = -1;
425 h->fref0[ref] = frames[bestref];
426 memcpy( h->fenc->weight[ref], weights[bestref], sizeof(weights[bestref]) );
432 static char *x264_strcat_filename( char *input, char *suffix )
434 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
437 strcpy( output, input );
438 strcat( output, suffix );
442 void x264_ratecontrol_init_reconfigurable( x264_t *h, int b_init )
444 x264_ratecontrol_t *rc = h->rc;
445 if( !b_init && rc->b_2pass )
448 if( h->param.rc.i_rc_method == X264_RC_CRF )
450 /* Arbitrary rescaling to make CRF somewhat similar to QP.
451 * Try to compensate for MB-tree's effects as well. */
452 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
453 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
454 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
455 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset );
458 if( h->param.rc.i_vbv_max_bitrate > 0 && h->param.rc.i_vbv_buffer_size > 0 )
460 if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
462 h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
463 x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
464 h->param.rc.i_vbv_buffer_size );
467 /* We don't support changing the ABR bitrate right now,
468 so if the stream starts as CBR, keep it CBR. */
469 if( rc->b_vbv_min_rate )
470 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
472 int vbv_buffer_size = h->param.rc.i_vbv_buffer_size * 1000;
473 int vbv_max_bitrate = h->param.rc.i_vbv_max_bitrate * 1000;
476 h->sps->vui.hrd.i_bit_rate_unscaled = vbv_max_bitrate;
477 h->sps->vui.hrd.i_cpb_size_unscaled = vbv_buffer_size;
478 if( h->param.i_nal_hrd && b_init )
480 h->sps->vui.hrd.i_cpb_cnt = 1;
481 h->sps->vui.hrd.b_cbr_hrd = h->param.i_nal_hrd == X264_NAL_HRD_CBR;
482 h->sps->vui.hrd.i_time_offset_length = 0;
487 int bitrate = 1000*h->param.rc.i_vbv_max_bitrate;
488 int bufsize = 1000*h->param.rc.i_vbv_buffer_size;
490 // normalize HRD size and rate to the value / scale notation
491 h->sps->vui.hrd.i_bit_rate_scale = x264_clip3( x264_ctz( bitrate ) - BR_SHIFT, 0, 15 );
492 h->sps->vui.hrd.i_bit_rate_value = bitrate >> ( h->sps->vui.hrd.i_bit_rate_scale + BR_SHIFT );
493 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 );
494 h->sps->vui.hrd.i_cpb_size_scale = x264_clip3( x264_ctz( bufsize ) - CPB_SHIFT, 0, 15 );
495 h->sps->vui.hrd.i_cpb_size_value = bufsize >> ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
496 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 );
502 #define MAX_DURATION 0.5
504 int max_cpb_output_delay = X264_MIN( h->param.i_keyint_max * MAX_DURATION * h->sps->vui.i_time_scale / h->sps->vui.i_num_units_in_tick, INT_MAX );
505 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;
506 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);
508 h->sps->vui.hrd.i_initial_cpb_removal_delay_length = 2 + x264_clip3( 32 - x264_clz( max_delay ), 4, 22 );
509 h->sps->vui.hrd.i_cpb_removal_delay_length = x264_clip3( 32 - x264_clz( max_cpb_output_delay ), 4, 31 );
510 h->sps->vui.hrd.i_dpb_output_delay_length = x264_clip3( 32 - x264_clz( max_dpb_output_delay ), 4, 31 );
514 vbv_buffer_size = h->sps->vui.hrd.i_cpb_size_unscaled;
515 vbv_max_bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
517 else if( h->param.i_nal_hrd && !b_init )
519 x264_log( h, X264_LOG_WARNING, "VBV parameters cannot be changed when NAL HRD is in use\n" );
523 rc->buffer_rate = vbv_max_bitrate / rc->fps;
524 rc->vbv_max_rate = vbv_max_bitrate;
525 rc->buffer_size = vbv_buffer_size;
526 rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
527 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
528 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
529 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.f_rf_constant_max )
531 rc->rate_factor_max_increment = h->param.rc.f_rf_constant_max - h->param.rc.f_rf_constant;
532 if( rc->rate_factor_max_increment <= 0 )
534 x264_log( h, X264_LOG_WARNING, "CRF max must be greater than CRF\n" );
535 rc->rate_factor_max_increment = 0;
540 if( h->param.rc.f_vbv_buffer_init > 1. )
541 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 );
542 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);
543 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init * h->sps->vui.i_time_scale;
545 rc->b_vbv_min_rate = !rc->b_2pass
546 && h->param.rc.i_rc_method == X264_RC_ABR
547 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
552 int x264_ratecontrol_new( x264_t *h )
554 x264_ratecontrol_t *rc;
558 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
561 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
562 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
564 /* FIXME: use integers */
565 if( h->param.i_fps_num > 0 && h->param.i_fps_den > 0 )
566 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
570 if( h->param.rc.b_mb_tree )
572 h->param.rc.f_pb_factor = 1;
576 rc->qcompress = h->param.rc.f_qcompress;
578 rc->bitrate = h->param.rc.i_bitrate * 1000.;
579 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
580 rc->nmb = h->mb.i_mb_count;
581 rc->last_non_b_pict_type = -1;
584 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
586 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
590 x264_ratecontrol_init_reconfigurable( h, 1 );
592 if( h->param.i_nal_hrd )
594 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale;
595 uint64_t num = 180000;
596 x264_reduce_fraction64( &num, &denom );
597 rc->hrd_multiply_denom = 180000 / num;
599 double bits_required = log2( 180000 / rc->hrd_multiply_denom )
600 + log2( h->sps->vui.i_time_scale )
601 + log2( h->sps->vui.hrd.i_cpb_size_unscaled );
602 if( bits_required >= 63 )
604 x264_log( h, X264_LOG_ERROR, "HRD with very large timescale and bufsize not supported\n" );
609 if( rc->rate_tolerance < 0.01 )
611 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
612 rc->rate_tolerance = 0.01;
615 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
619 /* FIXME ABR_INIT_QP is actually used only in CRF */
620 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
621 rc->accum_p_norm = .01;
622 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
623 /* estimated ratio that produces a reasonable QP for the first I-frame */
624 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
625 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
626 rc->last_non_b_pict_type = SLICE_TYPE_I;
629 rc->ip_offset = 6.0 * log2f( h->param.rc.f_ip_factor );
630 rc->pb_offset = 6.0 * log2f( h->param.rc.f_pb_factor );
631 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
632 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, QP_MAX );
633 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, QP_MAX );
634 h->mb.ip_offset = rc->ip_offset + 0.5;
636 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
637 rc->last_qscale = qp2qscale( 26 );
638 int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
639 CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
640 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
641 for( int i = 0; i < 3; i++ )
643 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
644 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
645 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
646 for( int j = 0; j < num_preds; j++ )
648 rc->pred[i+j*5].coeff= 2.0;
649 rc->pred[i+j*5].count= 1.0;
650 rc->pred[i+j*5].decay= 0.5;
651 rc->pred[i+j*5].offset= 0.0;
653 for( int j = 0; j < 2; j++ )
655 rc->row_preds[i][j].coeff= .25;
656 rc->row_preds[i][j].count= 1.0;
657 rc->row_preds[i][j].decay= 0.5;
658 rc->row_preds[i][j].offset= 0.0;
661 *rc->pred_b_from_p = rc->pred[0];
663 if( parse_zones( h ) < 0 )
665 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
669 /* Load stat file and init 2pass algo */
670 if( h->param.rc.b_stat_read )
672 char *p, *stats_in, *stats_buf;
674 /* read 1st pass stats */
675 assert( h->param.rc.psz_stat_in );
676 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
679 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
682 if( h->param.rc.b_mb_tree )
684 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
685 if( !mbtree_stats_in )
687 rc->p_mbtree_stat_file_in = fopen( mbtree_stats_in, "rb" );
688 x264_free( mbtree_stats_in );
689 if( !rc->p_mbtree_stat_file_in )
691 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
696 /* check whether 1st pass options were compatible with current options */
697 if( !strncmp( stats_buf, "#options:", 9 ) )
701 char *opts = stats_buf;
702 stats_in = strchr( stats_buf, '\n' );
707 if( sscanf( opts, "#options: %dx%d", &i, &j ) != 2 )
709 x264_log( h, X264_LOG_ERROR, "resolution specified in stats file not valid\n" );
712 else if( h->param.rc.b_mb_tree && (i != h->param.i_width || j != h->param.i_height) )
714 x264_log( h, X264_LOG_ERROR, "MB-tree doesn't support different resolution than 1st pass (%dx%d vs %dx%d)\n",
715 h->param.i_width, h->param.i_height, i, j );
719 if( ( p = strstr( opts, "timebase=" ) ) && sscanf( p, "timebase=%u/%u", &k, &l ) != 2 )
721 x264_log( h, X264_LOG_ERROR, "timebase specified in stats file not valid\n" );
724 if( k != h->param.i_timebase_num || l != h->param.i_timebase_den )
726 x264_log( h, X264_LOG_ERROR, "timebase mismatch with 1st pass (%u/%u vs %u/%u)\n",
727 h->param.i_timebase_num, h->param.i_timebase_den, k, l );
731 CMP_OPT_FIRST_PASS( "weightp", X264_MAX( 0, h->param.analyse.i_weighted_pred ) );
732 CMP_OPT_FIRST_PASS( "bframes", h->param.i_bframe );
733 CMP_OPT_FIRST_PASS( "b_pyramid", h->param.i_bframe_pyramid );
734 CMP_OPT_FIRST_PASS( "intra_refresh", h->param.b_intra_refresh );
735 CMP_OPT_FIRST_PASS( "open_gop", h->param.i_open_gop );
737 if( (p = strstr( opts, "keyint=" )) )
740 char buf[13] = "infinite ";
741 if( h->param.i_keyint_max != X264_KEYINT_MAX_INFINITE )
742 sprintf( buf, "%d ", h->param.i_keyint_max );
743 if( strncmp( p, buf, strlen(buf) ) )
745 x264_log( h, X264_LOG_ERROR, "different keyint setting than first pass (%.*s vs %.*s)\n",
746 strlen(buf)-1, buf, strcspn(p, " "), p );
751 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
752 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
754 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
756 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
757 h->mb.b_direct_auto_write = 1;
760 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
761 h->param.i_bframe_adaptive = i;
762 else if( h->param.i_bframe )
764 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
768 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 ) )
769 h->param.rc.i_lookahead = i;
772 /* find number of pics */
775 for( num_entries = -1; p; num_entries++ )
776 p = strchr( p + 1, ';' );
779 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
782 rc->num_entries = num_entries;
784 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
786 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
787 h->param.i_frame_total, rc->num_entries );
789 if( h->param.i_frame_total > rc->num_entries )
791 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
792 h->param.i_frame_total, rc->num_entries );
796 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
798 /* init all to skipped p frames */
799 for( int i = 0; i < rc->num_entries; i++ )
801 ratecontrol_entry_t *rce = &rc->entry[i];
802 rce->pict_type = SLICE_TYPE_P;
803 rce->qscale = rce->new_qscale = qp2qscale( 20 );
804 rce->misc_bits = rc->nmb + 10;
810 for( int i = 0; i < rc->num_entries; i++ )
812 ratecontrol_entry_t *rce;
820 next= strchr(p, ';');
822 *next++ = 0; //sscanf is unbelievably slow on long strings
823 e = sscanf( p, " in:%d ", &frame_number );
825 if( frame_number < 0 || frame_number >= rc->num_entries )
827 x264_log( h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i );
830 rce = &rc->entry[frame_number];
831 rce->direct_mode = 0;
833 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",
834 &pict_type, &rce->i_duration, &rce->i_cpb_duration, &qp, &rce->tex_bits,
835 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
836 &rce->s_count, &rce->direct_mode );
838 p = strstr( p, "ref:" );
842 for( ref = 0; ref < 16; ref++ )
844 if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
846 p = strchr( p+1, ' ' );
853 rce->i_weight_denom = -1;
854 char *w = strchr( p, 'w' );
856 if( sscanf( w, "w:%hd,%hd,%hd", &rce->i_weight_denom, &rce->weight[0], &rce->weight[1] ) != 3 )
857 rce->i_weight_denom = -1;
859 if( pict_type != 'b' )
860 rce->kept_as_ref = 1;
864 rce->frame_type = X264_TYPE_IDR;
865 rce->pict_type = SLICE_TYPE_I;
868 rce->frame_type = X264_TYPE_I;
869 rce->pict_type = SLICE_TYPE_I;
872 rce->frame_type = X264_TYPE_P;
873 rce->pict_type = SLICE_TYPE_P;
876 rce->frame_type = X264_TYPE_BREF;
877 rce->pict_type = SLICE_TYPE_B;
880 rce->frame_type = X264_TYPE_B;
881 rce->pict_type = SLICE_TYPE_B;
883 default: e = -1; break;
888 x264_log( h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e );
891 rce->qscale = qp2qscale( qp );
895 x264_free( stats_buf );
897 if( h->param.rc.i_rc_method == X264_RC_ABR )
899 if( init_pass2( h ) < 0 )
901 } /* else we're using constant quant, so no need to run the bitrate allocation */
904 /* Open output file */
905 /* If input and output files are the same, output to a temp file
906 * and move it to the real name only when it's complete */
907 if( h->param.rc.b_stat_write )
910 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
911 if( !rc->psz_stat_file_tmpname )
914 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
915 if( rc->p_stat_file_out == NULL )
917 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
921 p = x264_param2string( &h->param, 1 );
923 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
925 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
927 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
928 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
929 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
932 rc->p_mbtree_stat_file_out = fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
933 if( rc->p_mbtree_stat_file_out == NULL )
935 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
941 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
943 CHECKED_MALLOC( rc->qp_buffer[0], h->mb.i_mb_count * sizeof(uint16_t) );
944 if( h->param.i_bframe_pyramid && h->param.rc.b_stat_read )
945 CHECKED_MALLOC( rc->qp_buffer[1], h->mb.i_mb_count * sizeof(uint16_t) );
949 for( int i = 0; i<h->param.i_threads; i++ )
951 h->thread[i]->rc = rc+i;
955 h->thread[i]->param = h->param;
956 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
965 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
968 char *tok, UNUSED *saveptr=NULL;
970 z->f_bitrate_factor = 1;
971 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
973 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
975 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
979 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
985 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
986 memcpy( z->param, &h->param, sizeof(x264_param_t) );
987 z->param->param_free = x264_free;
988 while( (tok = strtok_r( p, ",", &saveptr )) )
990 char *val = strchr( tok, '=' );
996 if( x264_param_parse( z->param, tok, val ) )
998 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
1008 static int parse_zones( x264_t *h )
1010 x264_ratecontrol_t *rc = h->rc;
1011 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
1013 char *psz_zones, *p;
1014 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
1015 strcpy( psz_zones, h->param.rc.psz_zones );
1016 h->param.rc.i_zones = 1;
1017 for( p = psz_zones; *p; p++ )
1018 h->param.rc.i_zones += (*p == '/');
1019 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
1021 for( int i = 0; i < h->param.rc.i_zones; i++ )
1023 int i_tok = strcspn( p, "/" );
1025 if( parse_zone( h, &h->param.rc.zones[i], p ) )
1029 x264_free( psz_zones );
1032 if( h->param.rc.i_zones > 0 )
1034 for( int i = 0; i < h->param.rc.i_zones; i++ )
1036 x264_zone_t z = h->param.rc.zones[i];
1037 if( z.i_start < 0 || z.i_start > z.i_end )
1039 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
1040 z.i_start, z.i_end );
1043 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
1045 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
1046 z.f_bitrate_factor );
1051 rc->i_zones = h->param.rc.i_zones + 1;
1052 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
1053 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
1055 // default zone to fall back to if none of the others match
1056 rc->zones[0].i_start = 0;
1057 rc->zones[0].i_end = INT_MAX;
1058 rc->zones[0].b_force_qp = 0;
1059 rc->zones[0].f_bitrate_factor = 1;
1060 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
1061 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
1062 for( int i = 1; i < rc->i_zones; i++ )
1064 if( !rc->zones[i].param )
1065 rc->zones[i].param = rc->zones[0].param;
1074 static x264_zone_t *get_zone( x264_t *h, int frame_num )
1076 for( int i = h->rc->i_zones - 1; i >= 0; i-- )
1078 x264_zone_t *z = &h->rc->zones[i];
1079 if( frame_num >= z->i_start && frame_num <= z->i_end )
1085 void x264_ratecontrol_summary( x264_t *h )
1087 x264_ratecontrol_t *rc = h->rc;
1088 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
1090 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
1091 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
1092 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
1093 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
1094 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset );
1098 void x264_ratecontrol_delete( x264_t *h )
1100 x264_ratecontrol_t *rc = h->rc;
1103 if( rc->p_stat_file_out )
1105 b_regular_file = x264_is_regular_file( rc->p_stat_file_out );
1106 fclose( rc->p_stat_file_out );
1107 if( h->i_frame >= rc->num_entries && b_regular_file )
1108 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
1110 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1111 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
1113 x264_free( rc->psz_stat_file_tmpname );
1115 if( rc->p_mbtree_stat_file_out )
1117 b_regular_file = x264_is_regular_file( rc->p_mbtree_stat_file_out );
1118 fclose( rc->p_mbtree_stat_file_out );
1119 if( h->i_frame >= rc->num_entries && b_regular_file )
1120 if( rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
1122 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1123 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
1125 x264_free( rc->psz_mbtree_stat_file_tmpname );
1126 x264_free( rc->psz_mbtree_stat_file_name );
1128 if( rc->p_mbtree_stat_file_in )
1129 fclose( rc->p_mbtree_stat_file_in );
1130 x264_free( rc->pred );
1131 x264_free( rc->pred_b_from_p );
1132 x264_free( rc->entry );
1133 x264_free( rc->qp_buffer[0] );
1134 x264_free( rc->qp_buffer[1] );
1137 x264_free( rc->zones[0].param );
1138 for( int i = 1; i < rc->i_zones; i++ )
1139 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
1140 rc->zones[i].param->param_free( rc->zones[i].param );
1141 x264_free( rc->zones );
1146 static void accum_p_qp_update( x264_t *h, float qp )
1148 x264_ratecontrol_t *rc = h->rc;
1149 rc->accum_p_qp *= .95;
1150 rc->accum_p_norm *= .95;
1151 rc->accum_p_norm += 1;
1152 if( h->sh.i_type == SLICE_TYPE_I )
1153 rc->accum_p_qp += qp + rc->ip_offset;
1155 rc->accum_p_qp += qp;
1158 /* Before encoding a frame, choose a QP for it */
1159 void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
1161 x264_ratecontrol_t *rc = h->rc;
1162 ratecontrol_entry_t *rce = NULL;
1163 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
1168 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
1169 x264_encoder_reconfig( h, zone->param );
1170 rc->prev_zone = zone;
1172 rc->qp_force = i_force_qp;
1174 if( h->param.rc.b_stat_read )
1176 int frame = h->fenc->i_frame;
1177 assert( frame >= 0 && frame < rc->num_entries );
1178 rce = h->rc->rce = &h->rc->entry[frame];
1180 if( h->sh.i_type == SLICE_TYPE_B
1181 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
1183 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
1184 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
1190 memset( h->fdec->i_row_bits, 0, h->mb.i_mb_height * sizeof(int) );
1191 rc->row_pred = &rc->row_preds[h->sh.i_type];
1192 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;
1193 update_vbv_plan( h, overhead );
1195 const x264_level_t *l = x264_levels;
1196 while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
1199 int mincr = l->mincr;
1201 /* Blu-ray requires this */
1202 if( l->level_idc == 41 && h->param.i_nal_hrd )
1205 /* High 10 doesn't require minCR, so just set the maximum to a large value. */
1206 if( h->sps->i_profile_idc == PROFILE_HIGH10 )
1207 rc->frame_size_maximum = 1e9;
1210 /* The spec has a bizarre special case for the first frame. */
1211 if( h->i_frame == 0 )
1213 //384 * ( Max( PicSizeInMbs, fR * MaxMBPS ) + MaxMBPS * ( tr( 0 ) - tr,n( 0 ) ) ) / MinCR
1214 double fr = 1. / 172;
1215 int pic_size_in_mbs = h->mb.i_mb_width * h->mb.i_mb_height;
1216 rc->frame_size_maximum = 384 * BIT_DEPTH * X264_MAX( pic_size_in_mbs, fr*l->mbps ) / mincr;
1220 //384 * MaxMBPS * ( tr( n ) - tr( n - 1 ) ) / MinCR
1221 rc->frame_size_maximum = 384 * BIT_DEPTH * ((double)h->fenc->i_cpb_duration * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale) * l->mbps / mincr;
1226 if( h->sh.i_type != SLICE_TYPE_B )
1227 rc->bframes = h->fenc->i_bframes;
1233 else if( rc->b_abr )
1235 q = qscale2qp( rate_estimate_qscale( h ) );
1237 else if( rc->b_2pass )
1239 rce->new_qscale = rate_estimate_qscale( h );
1240 q = qscale2qp( rce->new_qscale );
1244 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
1245 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
1247 q = rc->qp_constant[ h->sh.i_type ];
1251 if( zone->b_force_qp )
1252 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1254 q -= 6*log2f( zone->f_bitrate_factor );
1258 q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1262 rc->qp = x264_clip3( (int)(q + 0.5), 0, QP_MAX );
1263 h->fdec->f_qp_avg_rc =
1264 h->fdec->f_qp_avg_aq =
1267 rce->new_qp = rc->qp;
1269 accum_p_qp_update( h, rc->qpm );
1271 if( h->sh.i_type != SLICE_TYPE_B )
1272 rc->last_non_b_pict_type = h->sh.i_type;
1275 static double predict_row_size( x264_t *h, int y, double qp )
1277 /* average between two predictors:
1278 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1279 x264_ratecontrol_t *rc = h->rc;
1280 double pred_s = predict_size( rc->row_pred[0], qp2qscale( qp ), h->fdec->i_row_satd[y] );
1282 if( h->sh.i_type == SLICE_TYPE_I || qp >= h->fref0[0]->f_row_qp[y] )
1284 if( h->sh.i_type == SLICE_TYPE_P
1285 && h->fref0[0]->i_type == h->fdec->i_type
1286 && h->fref0[0]->i_row_satd[y] > 0
1287 && (abs(h->fref0[0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1289 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
1290 * qp2qscale( h->fref0[0]->f_row_qp[y] ) / qp2qscale( qp );
1294 return (pred_s + pred_t) / 2;
1296 /* Our QP is lower than the reference! */
1299 double pred_intra = predict_size( rc->row_pred[1], qp2qscale( qp ), h->fdec->i_row_satds[0][0][y] );
1300 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1301 return pred_intra + pred_s;
1305 static double row_bits_so_far( x264_t *h, int y )
1308 for( int i = h->i_threadslice_start; i <= y; i++ )
1309 bits += h->fdec->i_row_bits[i];
1313 static double predict_row_size_sum( x264_t *h, int y, double qp )
1315 double bits = row_bits_so_far(h, y);
1316 for( int i = y+1; i < h->i_threadslice_end; i++ )
1317 bits += predict_row_size( h, i, qp );
1322 void x264_ratecontrol_mb( x264_t *h, int bits )
1324 x264_ratecontrol_t *rc = h->rc;
1325 const int y = h->mb.i_mb_y;
1329 h->fdec->i_row_bits[y] += bits;
1330 rc->qpa_rc += rc->qpm;
1331 rc->qpa_aq += h->mb.i_qp;
1333 if( h->mb.i_mb_x != h->mb.i_mb_width - 1 || !rc->b_vbv )
1336 h->fdec->f_row_qp[y] = rc->qpm;
1338 update_predictor( rc->row_pred[0], qp2qscale( rc->qpm ), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1339 if( h->sh.i_type == SLICE_TYPE_P && rc->qpm < h->fref0[0]->f_row_qp[y] )
1340 update_predictor( rc->row_pred[1], qp2qscale( rc->qpm ), h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1342 /* tweak quality based on difference from predicted size */
1343 if( y < h->i_threadslice_end-1 )
1345 float prev_row_qp = h->fdec->f_row_qp[y];
1346 float qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1347 float qp_absolute_max = h->param.rc.i_qp_max;
1348 if( rc->rate_factor_max_increment )
1349 qp_absolute_max = X264_MIN( qp_absolute_max, rc->qp_novbv + rc->rate_factor_max_increment );
1350 float qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, qp_absolute_max );
1351 float step_size = 0.5;
1353 /* B-frames shouldn't use lower QP than their reference frames. */
1354 if( h->sh.i_type == SLICE_TYPE_B )
1356 qp_min = X264_MAX( qp_min, X264_MAX( h->fref0[0]->f_row_qp[y+1], h->fref1[0]->f_row_qp[y+1] ) );
1357 rc->qpm = X264_MAX( rc->qpm, qp_min );
1360 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1361 float slice_size_planned = h->param.b_sliced_threads ? rc->slice_size_planned : rc->frame_size_planned;
1362 float max_frame_error = X264_MAX( 0.05, 1.0 / (h->mb.i_mb_height) );
1363 float size_of_other_slices = 0;
1364 if( h->param.b_sliced_threads )
1366 float size_of_other_slices_planned = 0;
1367 for( int i = 0; i < h->param.i_threads; i++ )
1368 if( h != h->thread[i] )
1370 size_of_other_slices += h->thread[i]->rc->frame_size_estimated;
1371 size_of_other_slices_planned += h->thread[i]->rc->slice_size_planned;
1373 float weight = rc->slice_size_planned / rc->frame_size_planned;
1374 size_of_other_slices = (size_of_other_slices - size_of_other_slices_planned) * weight + size_of_other_slices_planned;
1377 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1378 float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
1379 int b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1381 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1382 /* area at the top of the frame was measured inaccurately. */
1383 if( row_bits_so_far( h, y ) < 0.05 * slice_size_planned )
1386 if( h->sh.i_type != SLICE_TYPE_I )
1389 if( !rc->b_vbv_min_rate )
1390 qp_min = X264_MAX( qp_min, rc->qp_novbv );
1392 while( rc->qpm < qp_max
1393 && ((b1 > rc->frame_size_planned + rc_tol) ||
1394 (rc->buffer_fill - b1 < buffer_left_planned * 0.5) ||
1395 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1397 rc->qpm += step_size;
1398 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1401 while( rc->qpm > qp_min
1402 && (rc->qpm > h->fdec->f_row_qp[0] || rc->single_frame_vbv)
1403 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
1404 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
1406 rc->qpm -= step_size;
1407 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1410 /* avoid VBV underflow or MinCR violation */
1411 while( (rc->qpm < qp_absolute_max)
1412 && ((rc->buffer_fill - b1 < rc->buffer_rate * max_frame_error) ||
1413 (rc->frame_size_maximum - b1 < rc->frame_size_maximum * max_frame_error)))
1415 rc->qpm += step_size;
1416 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1419 h->rc->frame_size_estimated = predict_row_size_sum( h, y, rc->qpm );
1423 int x264_ratecontrol_qp( x264_t *h )
1426 return x264_clip3( h->rc->qpm + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1429 int x264_ratecontrol_mb_qp( x264_t *h )
1432 float qp = h->rc->qpm;
1433 if( h->param.rc.i_aq_mode )
1434 /* MB-tree currently doesn't adjust quantizers in unreferenced frames. */
1435 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];
1436 return x264_clip3( qp + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1439 /* In 2pass, force the same frame types as in the 1st pass */
1440 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1442 x264_ratecontrol_t *rc = h->rc;
1443 if( h->param.rc.b_stat_read )
1445 if( frame_num >= rc->num_entries )
1447 /* We could try to initialize everything required for ABR and
1448 * adaptive B-frames, but that would be complicated.
1449 * So just calculate the average QP used so far. */
1450 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24
1451 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1452 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, QP_MAX );
1453 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, QP_MAX );
1454 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, QP_MAX );
1456 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1457 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1458 if( h->param.i_bframe_adaptive )
1459 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1461 for( int i = 0; i < h->param.i_threads; i++ )
1463 h->thread[i]->rc->b_abr = 0;
1464 h->thread[i]->rc->b_2pass = 0;
1465 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1466 h->thread[i]->param.rc.b_stat_read = 0;
1467 h->thread[i]->param.i_bframe_adaptive = 0;
1468 h->thread[i]->param.i_scenecut_threshold = 0;
1469 h->thread[i]->param.rc.b_mb_tree = 0;
1470 if( h->thread[i]->param.i_bframe > 1 )
1471 h->thread[i]->param.i_bframe = 1;
1473 return X264_TYPE_AUTO;
1475 return rc->entry[frame_num].frame_type;
1478 return X264_TYPE_AUTO;
1481 void x264_ratecontrol_set_weights( x264_t *h, x264_frame_t *frm )
1483 ratecontrol_entry_t *rce = &h->rc->entry[frm->i_frame];
1484 if( h->param.analyse.i_weighted_pred <= 0 )
1486 if( rce->i_weight_denom >= 0 )
1487 SET_WEIGHT( frm->weight[0][0], 1, rce->weight[0], rce->i_weight_denom, rce->weight[1] );
1490 /* After encoding one frame, save stats and update ratecontrol state */
1491 int x264_ratecontrol_end( x264_t *h, int bits, int *filler )
1493 x264_ratecontrol_t *rc = h->rc;
1494 const int *mbs = h->stat.frame.i_mb_count;
1498 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1499 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1500 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1501 for( int i = B_DIRECT; i < B_8x8; i++ )
1502 h->stat.frame.i_mb_count_p += mbs[i];
1504 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1505 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1507 if( h->param.rc.b_stat_write )
1509 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1510 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1511 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1512 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1513 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1514 char c_direct = h->mb.b_direct_auto_write ?
1515 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1516 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1518 if( fprintf( rc->p_stat_file_out,
1519 "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:",
1520 h->fenc->i_frame, h->i_frame,
1521 c_type, h->fenc->i_duration,
1522 h->fenc->i_cpb_duration, rc->qpa_rc,
1523 h->stat.frame.i_tex_bits,
1524 h->stat.frame.i_mv_bits,
1525 h->stat.frame.i_misc_bits,
1526 h->stat.frame.i_mb_count_i,
1527 h->stat.frame.i_mb_count_p,
1528 h->stat.frame.i_mb_count_skip,
1532 /* Only write information for reference reordering once. */
1533 int use_old_stats = h->param.rc.b_stat_read && rc->rce->refs > 1;
1534 for( int i = 0; i < (use_old_stats ? rc->rce->refs : h->i_ref0); i++ )
1536 int refcount = use_old_stats ? rc->rce->refcount[i]
1537 : h->param.b_interlaced ? h->stat.frame.i_mb_count_ref[0][i*2]
1538 + h->stat.frame.i_mb_count_ref[0][i*2+1]
1539 : h->stat.frame.i_mb_count_ref[0][i];
1540 if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1544 if( h->sh.weight[0][0].weightfn )
1546 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 )
1550 if( fprintf( rc->p_stat_file_out, ";\n") < 0 )
1553 /* Don't re-write the data in multi-pass mode. */
1554 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1556 uint8_t i_type = h->sh.i_type;
1557 /* Values are stored as big-endian FIX8.8 */
1558 for( int i = 0; i < h->mb.i_mb_count; i++ )
1559 rc->qp_buffer[0][i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1560 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1562 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 )
1569 if( h->sh.i_type != SLICE_TYPE_B )
1570 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / rc->last_rceq;
1573 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1574 * Not perfectly accurate with B-refs, but good enough. */
1575 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / (rc->last_rceq * fabs( h->param.rc.f_pb_factor ));
1577 rc->cplxr_sum *= rc->cbr_decay;
1578 double frame_duration = (double)h->fenc->i_duration * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1580 rc->wanted_bits_window += frame_duration * rc->bitrate;
1581 rc->wanted_bits_window *= rc->cbr_decay;
1585 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale( rc->rce->new_qp ) );
1587 if( h->mb.b_variable_qp )
1589 if( h->sh.i_type == SLICE_TYPE_B )
1591 rc->bframe_bits += bits;
1592 if( h->fenc->b_last_minigop_bframe )
1594 update_predictor( rc->pred_b_from_p, qp2qscale( rc->qpa_rc ),
1595 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1596 rc->bframe_bits = 0;
1601 *filler = update_vbv( h, bits );
1603 if( h->sps->vui.b_nal_hrd_parameters_present )
1605 if( h->fenc->i_frame == 0 )
1607 // access unit initialises the HRD
1608 h->fenc->hrd_timing.cpb_initial_arrival_time = 0;
1609 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1610 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1611 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit = (double)rc->initial_cpb_removal_delay / 90000;
1615 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit + (double)h->fenc->i_cpb_delay *
1616 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1618 double cpb_earliest_arrival_time = h->fenc->hrd_timing.cpb_removal_time - (double)rc->initial_cpb_removal_delay / 90000;
1619 if( h->fenc->b_keyframe )
1621 rc->nrt_first_access_unit = h->fenc->hrd_timing.cpb_removal_time;
1622 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1623 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1626 cpb_earliest_arrival_time -= (double)rc->initial_cpb_removal_delay_offset / 90000;
1628 if( h->sps->vui.hrd.b_cbr_hrd )
1629 h->fenc->hrd_timing.cpb_initial_arrival_time = rc->previous_cpb_final_arrival_time;
1631 h->fenc->hrd_timing.cpb_initial_arrival_time = X264_MAX( rc->previous_cpb_final_arrival_time, cpb_earliest_arrival_time );
1633 int filler_bits = *filler ? X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), *filler )*8 : 0;
1635 h->fenc->hrd_timing.cpb_final_arrival_time = rc->previous_cpb_final_arrival_time = h->fenc->hrd_timing.cpb_initial_arrival_time +
1636 (double)(bits + filler_bits) / h->sps->vui.hrd.i_bit_rate_unscaled;
1638 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 +
1639 h->fenc->hrd_timing.cpb_removal_time;
1644 x264_log(h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n");
1648 /****************************************************************************
1650 ***************************************************************************/
1653 * modify the bitrate curve from pass1 for one frame
1655 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1657 x264_ratecontrol_t *rcc= h->rc;
1658 x264_zone_t *zone = get_zone( h, frame_num );
1659 double q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1661 // avoid NaN's in the rc_eq
1662 if( !isfinite(q) || rce->tex_bits + rce->mv_bits == 0 )
1663 q = rcc->last_qscale_for[rce->pict_type];
1668 rcc->last_qscale = q;
1673 if( zone->b_force_qp )
1674 q = qp2qscale( zone->i_qp );
1676 q /= zone->f_bitrate_factor;
1682 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1684 x264_ratecontrol_t *rcc = h->rc;
1685 const int pict_type = rce->pict_type;
1687 // force I/B quants as a function of P quants
1688 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1689 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1690 if( pict_type == SLICE_TYPE_I )
1693 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1694 double ip_factor = fabs( h->param.rc.f_ip_factor );
1695 /* don't apply ip_factor if the following frame is also I */
1696 if( rcc->accum_p_norm <= 0 )
1698 else if( h->param.rc.f_ip_factor < 0 )
1700 else if( rcc->accum_p_norm >= 1 )
1703 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1705 else if( pict_type == SLICE_TYPE_B )
1707 if( h->param.rc.f_pb_factor > 0 )
1709 if( !rce->kept_as_ref )
1710 q *= fabs( h->param.rc.f_pb_factor );
1712 else if( pict_type == SLICE_TYPE_P
1713 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1714 && rce->tex_bits == 0 )
1719 /* last qscale / qdiff stuff */
1720 if( rcc->last_non_b_pict_type == pict_type &&
1721 (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1) )
1723 double last_q = rcc->last_qscale_for[pict_type];
1724 double max_qscale = last_q * rcc->lstep;
1725 double min_qscale = last_q / rcc->lstep;
1727 if ( q > max_qscale ) q = max_qscale;
1728 else if( q < min_qscale ) q = min_qscale;
1731 rcc->last_qscale_for[pict_type] = q;
1732 if( pict_type != SLICE_TYPE_B )
1733 rcc->last_non_b_pict_type = pict_type;
1734 if( pict_type == SLICE_TYPE_I )
1736 rcc->last_accum_p_norm = rcc->accum_p_norm;
1737 rcc->accum_p_norm = 0;
1738 rcc->accum_p_qp = 0;
1740 if( pict_type == SLICE_TYPE_P )
1742 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1743 rcc->accum_p_qp = mask * (qscale2qp( q ) + rcc->accum_p_qp);
1744 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1749 static double predict_size( predictor_t *p, double q, double var )
1751 return (p->coeff*var + p->offset) / (q*p->count);
1754 static void update_predictor( predictor_t *p, double q, double var, double bits )
1756 const double range = 1.5;
1759 double old_coeff = p->coeff / p->count;
1760 double new_coeff = bits*q / var;
1761 double new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
1762 double new_offset = bits*q - new_coeff_clipped * var;
1763 if( new_offset >= 0 )
1764 new_coeff = new_coeff_clipped;
1767 p->count *= p->decay;
1768 p->coeff *= p->decay;
1769 p->offset *= p->decay;
1771 p->coeff += new_coeff;
1772 p->offset += new_offset;
1775 // update VBV after encoding a frame
1776 static int update_vbv( x264_t *h, int bits )
1779 int bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
1780 x264_ratecontrol_t *rcc = h->rc;
1781 x264_ratecontrol_t *rct = h->thread[0]->rc;
1782 uint64_t buffer_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
1784 if( rcc->last_satd >= h->mb.i_mb_count )
1785 update_predictor( &rct->pred[h->sh.i_type], qp2qscale( rcc->qpa_rc ), rcc->last_satd, bits );
1790 rct->buffer_fill_final -= (uint64_t)bits * h->sps->vui.i_time_scale;
1792 if( rct->buffer_fill_final < 0 )
1793 x264_log( h, X264_LOG_WARNING, "VBV underflow (frame %d, %.0f bits)\n", h->i_frame, (double)rct->buffer_fill_final / h->sps->vui.i_time_scale );
1794 rct->buffer_fill_final = X264_MAX( rct->buffer_fill_final, 0 );
1795 rct->buffer_fill_final += (uint64_t)bitrate * h->sps->vui.i_num_units_in_tick * h->fenc->i_cpb_duration;
1797 if( h->sps->vui.hrd.b_cbr_hrd && rct->buffer_fill_final > buffer_size )
1799 filler = ceil( (rct->buffer_fill_final - buffer_size) / (8. * h->sps->vui.i_time_scale) );
1800 bits = X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), filler ) * 8;
1801 rct->buffer_fill_final -= (uint64_t)bits * h->sps->vui.i_time_scale;
1804 rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, buffer_size );
1809 void x264_hrd_fullness( x264_t *h )
1811 x264_ratecontrol_t *rct = h->thread[0]->rc;
1812 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale / rct->hrd_multiply_denom;
1813 uint64_t cpb_state = rct->buffer_fill_final;
1814 uint64_t cpb_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
1815 uint64_t multiply_factor = 180000 / rct->hrd_multiply_denom;
1817 if( rct->buffer_fill_final < 0 || rct->buffer_fill_final > cpb_size )
1819 x264_log( h, X264_LOG_WARNING, "CPB %s: %.0lf bits in a %.0lf-bit buffer\n",
1820 rct->buffer_fill_final < 0 ? "underflow" : "overflow", (float)rct->buffer_fill_final/denom, (float)cpb_size/denom );
1823 h->initial_cpb_removal_delay = (multiply_factor * cpb_state + denom) / (2*denom);
1824 h->initial_cpb_removal_delay_offset = (multiply_factor * cpb_size + denom) / (2*denom) - h->initial_cpb_removal_delay;
1827 // provisionally update VBV according to the planned size of all frames currently in progress
1828 static void update_vbv_plan( x264_t *h, int overhead )
1830 x264_ratecontrol_t *rcc = h->rc;
1831 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final / h->sps->vui.i_time_scale;
1832 if( h->i_thread_frames > 1 )
1834 int j = h->rc - h->thread[0]->rc;
1835 for( int i = 1; i < h->i_thread_frames; i++ )
1837 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
1838 double bits = t->rc->frame_size_planned;
1839 if( !t->b_thread_active )
1841 bits = X264_MAX(bits, t->rc->frame_size_estimated);
1842 rcc->buffer_fill -= bits;
1843 rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
1844 rcc->buffer_fill += t->rc->buffer_rate;
1845 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1848 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1849 rcc->buffer_fill -= overhead;
1852 // apply VBV constraints and clip qscale to between lmin and lmax
1853 static double clip_qscale( x264_t *h, int pict_type, double q )
1855 x264_ratecontrol_t *rcc = h->rc;
1856 double lmin = rcc->lmin[pict_type];
1857 double lmax = rcc->lmax[pict_type];
1858 if( rcc->rate_factor_max_increment )
1859 lmax = X264_MIN( lmax, qp2qscale( rcc->qp_novbv + rcc->rate_factor_max_increment ) );
1862 /* B-frames are not directly subject to VBV,
1863 * since they are controlled by the P-frames' QPs. */
1865 if( rcc->b_vbv && rcc->last_satd > 0 )
1867 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
1868 * the lookahead overflow and such that the buffer is in a reasonable state
1869 * by the end of the lookahead. */
1870 if( h->param.rc.i_lookahead )
1874 /* Avoid an infinite loop. */
1875 for( int iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
1878 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1879 double buffer_fill_cur = rcc->buffer_fill - cur_bits;
1881 double total_duration = 0;
1882 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
1883 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
1884 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
1886 /* Loop over the planned future frames. */
1887 for( int j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
1889 total_duration += h->fenc->f_planned_cpb_duration[j];
1890 buffer_fill_cur += rcc->vbv_max_rate * h->fenc->f_planned_cpb_duration[j];
1891 int i_type = h->fenc->i_planned_type[j];
1892 int i_satd = h->fenc->i_planned_satd[j];
1893 if( i_type == X264_TYPE_AUTO )
1895 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
1896 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
1897 buffer_fill_cur -= cur_bits;
1899 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
1900 target_fill = X264_MIN( rcc->buffer_fill + total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.5 );
1901 if( buffer_fill_cur < target_fill )
1907 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
1908 target_fill = x264_clip3f( rcc->buffer_fill - total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
1909 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
1918 /* Fallback to old purely-reactive algorithm: no lookahead. */
1921 if( ( pict_type == SLICE_TYPE_P ||
1922 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
1923 rcc->buffer_fill/rcc->buffer_size < 0.5 )
1925 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1928 /* Now a hard threshold to make sure the frame fits in VBV.
1929 * This one is mostly for I-frames. */
1930 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1932 /* For small VBVs, allow the frame to use up the entire VBV. */
1933 double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
1934 /* For single-frame VBVs, request that the frame use up the entire VBV. */
1935 double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
1937 if( bits > rcc->buffer_fill/max_fill_factor )
1938 qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
1941 if( bits < rcc->buffer_rate/min_fill_factor )
1942 q *= bits*min_fill_factor/rcc->buffer_rate;
1943 q = X264_MAX( q0, q );
1946 /* Apply MinCR restrictions */
1947 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1948 if( bits > rcc->frame_size_maximum )
1949 q *= bits / rcc->frame_size_maximum;
1950 bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1952 /* Check B-frame complexity, and use up any bits that would
1953 * overflow before the next P-frame. */
1954 if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
1956 int nb = rcc->bframes;
1957 double pbbits = bits;
1958 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1960 double bframe_cpb_duration = 0;
1961 double minigop_cpb_duration;
1962 for( int i = 0; i < nb; i++ )
1963 bframe_cpb_duration += h->fenc->f_planned_cpb_duration[1+i];
1965 if( bbits * nb > bframe_cpb_duration * rcc->vbv_max_rate )
1967 pbbits += nb * bbits;
1969 minigop_cpb_duration = bframe_cpb_duration + h->fenc->f_planned_cpb_duration[0];
1970 space = rcc->buffer_fill + minigop_cpb_duration*rcc->vbv_max_rate - rcc->buffer_size;
1971 if( pbbits < space )
1973 q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
1975 q = X264_MAX( q0-5, q );
1978 if( !rcc->b_vbv_min_rate )
1979 q = X264_MAX( q0, q );
1984 else if( rcc->b_2pass )
1986 double min2 = log( lmin );
1987 double max2 = log( lmax );
1988 q = (log(q) - min2)/(max2-min2) - 0.5;
1989 q = 1.0/(1.0 + exp( -4*q ));
1990 q = q*(max2-min2) + min2;
1994 return x264_clip3f( q, lmin, lmax );
1997 // update qscale for 1 frame based on actual bits used so far
1998 static float rate_estimate_qscale( x264_t *h )
2001 x264_ratecontrol_t *rcc = h->rc;
2002 ratecontrol_entry_t rce;
2003 int pict_type = h->sh.i_type;
2004 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
2005 + h->stat.i_frame_size[SLICE_TYPE_P]
2006 + h->stat.i_frame_size[SLICE_TYPE_B]);
2011 if( pict_type != rce.pict_type )
2013 x264_log( h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
2014 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type] );
2018 if( pict_type == SLICE_TYPE_B )
2020 /* B-frames don't have independent ratecontrol, but rather get the
2021 * average QP of the two adjacent P-frames + an offset */
2023 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
2024 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
2025 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
2026 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
2027 float q0 = h->fref0[0]->f_qp_avg_rc;
2028 float q1 = h->fref1[0]->f_qp_avg_rc;
2030 if( h->fref0[0]->i_type == X264_TYPE_BREF )
2031 q0 -= rcc->pb_offset/2;
2032 if( h->fref1[0]->i_type == X264_TYPE_BREF )
2033 q1 -= rcc->pb_offset/2;
2036 q = (q0 + q1) / 2 + rcc->ip_offset;
2042 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
2044 if( h->fenc->b_kept_as_ref )
2045 q += rcc->pb_offset/2;
2047 q += rcc->pb_offset;
2049 if( rcc->b_2pass && rcc->b_vbv )
2050 rcc->frame_size_planned = qscale2bits( &rce, q );
2052 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
2053 h->rc->frame_size_estimated = rcc->frame_size_planned;
2057 rcc->last_satd = x264_rc_analyse_slice( h );
2059 return qp2qscale( q );
2063 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
2067 double lmin = rcc->lmin[pict_type];
2068 double lmax = rcc->lmax[pict_type];
2070 int64_t predicted_bits = total_bits;
2074 if( h->i_thread_frames > 1 )
2076 int j = h->rc - h->thread[0]->rc;
2077 for( int i = 1; i < h->i_thread_frames; i++ )
2079 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
2080 double bits = t->rc->frame_size_planned;
2081 if( !t->b_thread_active )
2083 bits = X264_MAX(bits, t->rc->frame_size_estimated);
2084 predicted_bits += (int64_t)bits;
2090 if( h->i_frame < h->i_thread_frames )
2091 predicted_bits += (int64_t)h->i_frame * rcc->bitrate / rcc->fps;
2093 predicted_bits += (int64_t)(h->i_thread_frames - 1) * rcc->bitrate / rcc->fps;
2096 /* Adjust ABR buffer based on distance to the end of the video. */
2097 if( rcc->num_entries > h->i_frame )
2099 double final_bits = rcc->entry[rcc->num_entries-1].expected_bits;
2100 double video_pos = rce.expected_bits / final_bits;
2101 double scale_factor = sqrt( (1 - video_pos) * rcc->num_entries );
2102 abr_buffer *= 0.5 * X264_MAX( scale_factor, 0.5 );
2105 diff = predicted_bits - (int64_t)rce.expected_bits;
2107 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
2108 if( ((h->i_frame + 1 - h->i_thread_frames) >= rcc->fps) &&
2109 (rcc->expected_bits_sum > 0))
2111 /* Adjust quant based on the difference between
2112 * achieved and expected bitrate so far */
2113 double cur_time = (double)h->i_frame / rcc->num_entries;
2114 double w = x264_clip3f( cur_time*100, 0.0, 1.0 );
2115 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
2119 /* Do not overflow vbv */
2120 double expected_size = qscale2bits( &rce, q );
2121 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2122 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
2123 double qmax = q*(2 - expected_fullness);
2124 double size_constraint = 1 + expected_fullness;
2125 qmax = X264_MAX( qmax, rce.new_qscale );
2126 if( expected_fullness < .05 )
2128 qmax = X264_MIN(qmax, lmax);
2129 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
2130 ((expected_vbv < 0) && (q < lmax)))
2133 expected_size = qscale2bits(&rce, q);
2134 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2136 rcc->last_satd = x264_rc_analyse_slice( h );
2138 q = x264_clip3f( q, lmin, lmax );
2140 else /* 1pass ABR */
2142 /* Calculate the quantizer which would have produced the desired
2143 * average bitrate if it had been applied to all frames so far.
2144 * Then modulate that quant based on the current frame's complexity
2145 * relative to the average complexity so far (using the 2pass RCEQ).
2146 * Then bias the quant up or down if total size so far was far from
2148 * Result: Depending on the value of rate_tolerance, there is a
2149 * tradeoff between quality and bitrate precision. But at large
2150 * tolerances, the bit distribution approaches that of 2pass. */
2152 double wanted_bits, overflow = 1;
2154 rcc->last_satd = x264_rc_analyse_slice( h );
2155 rcc->short_term_cplxsum *= 0.5;
2156 rcc->short_term_cplxcount *= 0.5;
2157 rcc->short_term_cplxsum += rcc->last_satd;
2158 rcc->short_term_cplxcount ++;
2160 rce.tex_bits = rcc->last_satd;
2161 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
2163 rce.p_count = rcc->nmb;
2167 rce.pict_type = pict_type;
2169 if( h->param.rc.i_rc_method == X264_RC_CRF )
2171 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
2175 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
2177 /* ABR code can potentially be counterproductive in CBR, so just don't bother.
2178 * Don't run it if the frame complexity is zero either. */
2179 if( !rcc->b_vbv_min_rate && rcc->last_satd )
2181 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
2182 int i_frame_done = h->i_frame + 1 - h->i_thread_frames;
2183 double time_done = i_frame_done / rcc->fps;
2184 if( h->param.b_vfr_input && i_frame_done > 0 )
2185 time_done = ((double)(h->fenc->i_reordered_pts - h->i_reordered_pts_delay)) * h->param.i_timebase_num / h->param.i_timebase_den;
2186 wanted_bits = time_done * rcc->bitrate;
2187 if( wanted_bits > 0 )
2189 abr_buffer *= X264_MAX( 1, sqrt( time_done ) );
2190 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
2196 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
2197 /* should test _next_ pict type, but that isn't decided yet */
2198 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
2200 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2201 q /= fabs( h->param.rc.f_ip_factor );
2203 else if( h->i_frame > 0 )
2205 /* Asymmetric clipping, because symmetric would prevent
2206 * overflow control in areas of rapidly oscillating complexity */
2207 double lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
2208 double lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
2209 if( overflow > 1.1 && h->i_frame > 3 )
2211 else if( overflow < 0.9 )
2214 q = x264_clip3f(q, lmin, lmax);
2216 else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
2218 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
2220 rcc->qp_novbv = qscale2qp( q );
2222 //FIXME use get_diff_limited_q() ?
2223 q = clip_qscale( h, pict_type, q );
2226 rcc->last_qscale_for[pict_type] =
2227 rcc->last_qscale = q;
2229 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
2230 rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );
2232 if( rcc->b_2pass && rcc->b_vbv )
2233 rcc->frame_size_planned = qscale2bits(&rce, q);
2235 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2237 /* Always use up the whole VBV in this case. */
2238 if( rcc->single_frame_vbv )
2239 rcc->frame_size_planned = rcc->buffer_rate;
2240 h->rc->frame_size_estimated = rcc->frame_size_planned;
2245 void x264_threads_normalize_predictors( x264_t *h )
2247 double totalsize = 0;
2248 for( int i = 0; i < h->param.i_threads; i++ )
2249 totalsize += h->thread[i]->rc->slice_size_planned;
2250 double factor = h->rc->frame_size_planned / totalsize;
2251 for( int i = 0; i < h->param.i_threads; i++ )
2252 h->thread[i]->rc->slice_size_planned *= factor;
2255 void x264_threads_distribute_ratecontrol( x264_t *h )
2258 x264_ratecontrol_t *rc = h->rc;
2260 /* Initialize row predictors */
2261 if( h->i_frame == 0 )
2262 for( int i = 0; i < h->param.i_threads; i++ )
2264 x264_ratecontrol_t *t = h->thread[i]->rc;
2265 memcpy( t->row_preds, rc->row_preds, sizeof(rc->row_preds) );
2268 for( int i = 0; i < h->param.i_threads; i++ )
2270 x264_t *t = h->thread[i];
2271 memcpy( t->rc, rc, offsetof(x264_ratecontrol_t, row_pred) );
2272 t->rc->row_pred = &t->rc->row_preds[h->sh.i_type];
2273 /* Calculate the planned slice size. */
2274 if( rc->b_vbv && rc->frame_size_planned )
2277 for( row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2278 size += h->fdec->i_row_satd[row];
2279 t->rc->slice_size_planned = predict_size( &rc->pred[h->sh.i_type + (i+1)*5], rc->qpm, size );
2282 t->rc->slice_size_planned = 0;
2284 if( rc->b_vbv && rc->frame_size_planned )
2286 x264_threads_normalize_predictors( h );
2288 if( rc->single_frame_vbv )
2290 /* Compensate for our max frame error threshold: give more bits (proportionally) to smaller slices. */
2291 for( int i = 0; i < h->param.i_threads; i++ )
2293 x264_t *t = h->thread[i];
2294 float max_frame_error = X264_MAX( 0.05, 1.0 / (t->i_threadslice_end - t->i_threadslice_start) );
2295 t->rc->slice_size_planned += 2 * max_frame_error * rc->frame_size_planned;
2297 x264_threads_normalize_predictors( h );
2300 for( int i = 0; i < h->param.i_threads; i++ )
2301 h->thread[i]->rc->frame_size_estimated = h->thread[i]->rc->slice_size_planned;
2305 void x264_threads_merge_ratecontrol( x264_t *h )
2307 x264_ratecontrol_t *rc = h->rc;
2310 for( int i = 0; i < h->param.i_threads; i++ )
2312 x264_t *t = h->thread[i];
2313 x264_ratecontrol_t *rct = h->thread[i]->rc;
2314 if( h->param.rc.i_vbv_buffer_size )
2317 for( int row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2318 size += h->fdec->i_row_satd[row];
2319 int bits = t->stat.frame.i_mv_bits + t->stat.frame.i_tex_bits + t->stat.frame.i_misc_bits;
2320 int mb_count = (t->i_threadslice_end - t->i_threadslice_start) * h->mb.i_mb_width;
2321 update_predictor( &rc->pred[h->sh.i_type+(i+1)*5], qp2qscale( rct->qpa_rc/mb_count ), size, bits );
2325 rc->qpa_rc += rct->qpa_rc;
2326 rc->qpa_aq += rct->qpa_aq;
2330 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
2334 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
2335 /* these vars are updated in x264_ratecontrol_start()
2336 * so copy them from the context that most recently started (prev)
2337 * to the context that's about to start (cur). */
2342 COPY(last_qscale_for);
2343 COPY(last_non_b_pict_type);
2344 COPY(short_term_cplxsum);
2345 COPY(short_term_cplxcount);
2349 /* these vars can be updated by x264_ratecontrol_init_reconfigurable */
2352 COPY(single_frame_vbv);
2354 COPY(b_vbv_min_rate);
2355 COPY(rate_factor_constant);
2361 #define COPY(var) next->rc->var = cur->rc->var
2362 /* these vars are updated in x264_ratecontrol_end()
2363 * so copy them from the context that most recently ended (cur)
2364 * to the context that's about to end (next) */
2366 COPY(expected_bits_sum);
2367 COPY(wanted_bits_window);
2369 COPY(initial_cpb_removal_delay);
2370 COPY(initial_cpb_removal_delay_offset);
2371 COPY(nrt_first_access_unit);
2372 COPY(previous_cpb_final_arrival_time);
2375 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
2376 /* the rest of the variables are either constant or thread-local */
2379 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
2381 /* find an interval ending on an overflow or underflow (depending on whether
2382 * we're adding or removing bits), and starting on the earliest frame that
2383 * can influence the buffer fill of that end frame. */
2384 x264_ratecontrol_t *rcc = h->rc;
2385 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
2386 const double buffer_max = .9 * rcc->buffer_size;
2387 double fill = fills[*t0-1];
2388 double parity = over ? 1. : -1.;
2389 int start = -1, end = -1;
2390 for( int i = *t0; i < rcc->num_entries; i++ )
2392 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 -
2393 qscale2bits( &rcc->entry[i], rcc->entry[i].new_qscale )) * parity;
2394 fill = x264_clip3f(fill, 0, rcc->buffer_size);
2396 if( fill <= buffer_min || i == 0 )
2402 else if( fill >= buffer_max && start >= 0 )
2407 return start >= 0 && end >= 0;
2410 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
2412 x264_ratecontrol_t *rcc = h->rc;
2413 double qscale_orig, qscale_new;
2417 for( int i = t0; i <= t1; i++ )
2419 qscale_orig = rcc->entry[i].new_qscale;
2420 qscale_orig = x264_clip3f( qscale_orig, qscale_min, qscale_max );
2421 qscale_new = qscale_orig * adjustment;
2422 qscale_new = x264_clip3f( qscale_new, qscale_min, qscale_max );
2423 rcc->entry[i].new_qscale = qscale_new;
2424 adjusted = adjusted || (qscale_new != qscale_orig);
2429 static double count_expected_bits( x264_t *h )
2431 x264_ratecontrol_t *rcc = h->rc;
2432 double expected_bits = 0;
2433 for( int i = 0; i < rcc->num_entries; i++ )
2435 ratecontrol_entry_t *rce = &rcc->entry[i];
2436 rce->expected_bits = expected_bits;
2437 expected_bits += qscale2bits( rce, rce->new_qscale );
2439 return expected_bits;
2442 static int vbv_pass2( x264_t *h, double all_available_bits )
2444 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
2445 * frames in the interval until either buffer is full at some intermediate frame or the
2446 * last frame in the interval no longer underflows. Recompute intervals and repeat.
2447 * Then do the converse to put bits back into overflow areas until target size is met */
2449 x264_ratecontrol_t *rcc = h->rc;
2451 double expected_bits = 0;
2453 double prev_bits = 0;
2455 double qscale_min = qp2qscale( h->param.rc.i_qp_min );
2456 double qscale_max = qp2qscale( h->param.rc.i_qp_max );
2458 int adj_min, adj_max;
2459 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2463 /* adjust overall stream size */
2467 prev_bits = expected_bits;
2470 { /* not first iteration */
2471 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2472 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2476 while(adj_min && find_underflow( h, fills, &t0, &t1, 1 ))
2478 adj_min = fix_underflow( h, t0, t1, adjustment, qscale_min, qscale_max );
2483 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2485 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2487 while( adj_max && find_underflow( h, fills, &t0, &t1, 0 ) )
2488 adj_max = fix_underflow( h, t0, t1, 1.001, qscale_min, qscale_max );
2490 expected_bits = count_expected_bits( h );
2491 } while( (expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2494 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2496 /* store expected vbv filling values for tracking when encoding */
2497 for( int i = 0; i < rcc->num_entries; i++ )
2498 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
2500 x264_free( fills-1 );
2506 static int init_pass2( x264_t *h )
2508 x264_ratecontrol_t *rcc = h->rc;
2509 uint64_t all_const_bits = 0;
2510 double duration = 0;
2511 for( int i = 0; i < rcc->num_entries; i++ )
2512 duration += rcc->entry[i].i_duration;
2513 duration *= (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2514 uint64_t all_available_bits = h->param.rc.i_bitrate * 1000. * duration;
2515 double rate_factor, step_mult;
2516 double qblur = h->param.rc.f_qblur;
2517 double cplxblur = h->param.rc.f_complexity_blur;
2518 const int filter_size = (int)(qblur*4) | 1;
2519 double expected_bits;
2520 double *qscale, *blurred_qscale;
2521 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
2523 /* find total/average complexity & const_bits */
2524 for( int i = 0; i < rcc->num_entries; i++ )
2526 ratecontrol_entry_t *rce = &rcc->entry[i];
2527 all_const_bits += rce->misc_bits;
2530 if( all_available_bits < all_const_bits)
2532 x264_log( h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2533 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)) );
2537 /* Blur complexities, to reduce local fluctuation of QP.
2538 * We don't blur the QPs directly, because then one very simple frame
2539 * could drag down the QP of a nearby complex frame and give it more
2540 * bits than intended. */
2541 for( int i = 0; i < rcc->num_entries; i++ )
2543 ratecontrol_entry_t *rce = &rcc->entry[i];
2544 double weight_sum = 0;
2545 double cplx_sum = 0;
2546 double weight = 1.0;
2547 double gaussian_weight;
2548 /* weighted average of cplx of future frames */
2549 for( int j = 1; j < cplxblur*2 && j < rcc->num_entries-i; j++ )
2551 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2552 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2553 if( weight < .0001 )
2555 gaussian_weight = weight * exp( -j*j/200.0 );
2556 weight_sum += gaussian_weight;
2557 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2559 /* weighted average of cplx of past frames */
2561 for( int j = 0; j <= cplxblur*2 && j <= i; j++ )
2563 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2564 gaussian_weight = weight * exp( -j*j/200.0 );
2565 weight_sum += gaussian_weight;
2566 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits);
2567 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2568 if( weight < .0001 )
2571 rce->blurred_complexity = cplx_sum / weight_sum;
2574 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2575 if( filter_size > 1 )
2576 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2578 blurred_qscale = qscale;
2580 /* Search for a factor which, when multiplied by the RCEQ values from
2581 * each frame, adds up to the desired total size.
2582 * There is no exact closed-form solution because of VBV constraints and
2583 * because qscale2bits is not invertible, but we can start with the simple
2584 * approximation of scaling the 1st pass by the ratio of bitrates.
2585 * The search range is probably overkill, but speed doesn't matter here. */
2588 for( int i = 0; i < rcc->num_entries; i++ )
2590 double q = get_qscale(h, &rcc->entry[i], 1.0, i);
2591 expected_bits += qscale2bits(&rcc->entry[i], q);
2592 rcc->last_qscale_for[rcc->entry[i].pict_type] = q;
2594 step_mult = all_available_bits / expected_bits;
2597 for( double step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2600 rate_factor += step;
2602 rcc->last_non_b_pict_type = -1;
2603 rcc->last_accum_p_norm = 1;
2604 rcc->accum_p_norm = 0;
2606 rcc->last_qscale_for[0] =
2607 rcc->last_qscale_for[1] =
2608 rcc->last_qscale_for[2] = pow( base_cplx, 1 - rcc->qcompress ) / rate_factor;
2611 for( int i = 0; i < rcc->num_entries; i++ )
2613 qscale[i] = get_qscale( h, &rcc->entry[i], rate_factor, i );
2614 rcc->last_qscale_for[rcc->entry[i].pict_type] = qscale[i];
2617 /* fixed I/B qscale relative to P */
2618 for( int i = rcc->num_entries-1; i >= 0; i-- )
2620 qscale[i] = get_diff_limited_q( h, &rcc->entry[i], qscale[i] );
2621 assert(qscale[i] >= 0);
2625 if( filter_size > 1 )
2627 assert( filter_size%2 == 1 );
2628 for( int i = 0; i < rcc->num_entries; i++ )
2630 ratecontrol_entry_t *rce = &rcc->entry[i];
2631 double q = 0.0, sum = 0.0;
2633 for( int j = 0; j < filter_size; j++ )
2635 int idx = i+j-filter_size/2;
2637 double coeff = qblur==0 ? 1.0 : exp( -d*d/(qblur*qblur) );
2638 if( idx < 0 || idx >= rcc->num_entries )
2640 if( rce->pict_type != rcc->entry[idx].pict_type )
2642 q += qscale[idx] * coeff;
2645 blurred_qscale[i] = q/sum;
2649 /* find expected bits */
2650 for( int i = 0; i < rcc->num_entries; i++ )
2652 ratecontrol_entry_t *rce = &rcc->entry[i];
2653 rce->new_qscale = clip_qscale( h, rce->pict_type, blurred_qscale[i] );
2654 assert(rce->new_qscale >= 0);
2655 expected_bits += qscale2bits( rce, rce->new_qscale );
2658 if( expected_bits > all_available_bits )
2659 rate_factor -= step;
2662 x264_free( qscale );
2663 if( filter_size > 1 )
2664 x264_free( blurred_qscale );
2667 if( vbv_pass2( h, all_available_bits ) )
2669 expected_bits = count_expected_bits( h );
2671 if( fabs( expected_bits/all_available_bits - 1.0 ) > 0.01 )
2674 for( int i = 0; i < rcc->num_entries; i++ )
2675 avgq += rcc->entry[i].new_qscale;
2676 avgq = qscale2qp( avgq / rcc->num_entries );
2678 if( expected_bits > all_available_bits || !rcc->b_vbv )
2679 x264_log( h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n" );
2680 x264_log( h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
2681 (float)h->param.rc.i_bitrate,
2682 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
2684 if( expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2 )
2686 if( h->param.rc.i_qp_min > 0 )
2687 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min );
2689 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate\n" );
2691 else if( expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2 )
2693 if( h->param.rc.i_qp_max < QP_MAX )
2694 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max );
2696 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate\n");
2698 else if( !(rcc->b_2pass && rcc->b_vbv) )
2699 x264_log( h, X264_LOG_WARNING, "internal error\n" );