1 /*****************************************************************************
2 * ratecontrol.c: ratecontrol
3 *****************************************************************************
4 * Copyright (C) 2005-2011 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.
26 * This program is also available under a commercial proprietary license.
27 * For more information, contact us at licensing@x264.com.
28 *****************************************************************************/
30 #define _ISOC99_SOURCE
31 #undef NDEBUG // always check asserts, the speed effect is far too small to disable them
34 #include "common/common.h"
35 #include "ratecontrol.h"
47 uint64_t expected_bits; /*total expected bits up to the current frame (current one excluded)*/
54 float blurred_complexity;
57 int16_t i_weight_denom[2];
61 int64_t i_cpb_duration;
62 } ratecontrol_entry_t;
72 struct x264_ratecontrol_t
81 double rate_tolerance;
83 int nmb; /* number of macroblocks in a frame */
87 ratecontrol_entry_t *rce;
88 int qp; /* qp for current frame */
89 float qpm; /* qp for current macroblock: precise float for AQ */
90 float qpa_rc; /* average of macroblocks' qp before aq */
91 float qpa_aq; /* average of macroblocks' qp after aq */
92 float qp_novbv; /* QP for the current frame if 1-pass VBV was disabled. */
96 int64_t buffer_fill_final;
97 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
98 double buffer_rate; /* # of bits added to buffer_fill after each frame */
99 double vbv_max_rate; /* # of bits added to buffer_fill per second */
100 predictor_t *pred; /* predict frame size from satd */
101 int single_frame_vbv;
102 double rate_factor_max_increment; /* Don't allow RF above (CRF + this value). */
107 double cplxr_sum; /* sum of bits*qscale/rceq */
108 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow, only includes finished frames */
109 int64_t filler_bits_sum; /* sum in bits of finished frames' filler data */
110 double wanted_bits_window; /* target bitrate * window */
112 double short_term_cplxsum;
113 double short_term_cplxcount;
114 double rate_factor_constant;
119 FILE *p_stat_file_out;
120 char *psz_stat_file_tmpname;
121 FILE *p_mbtree_stat_file_out;
122 char *psz_mbtree_stat_file_tmpname;
123 char *psz_mbtree_stat_file_name;
124 FILE *p_mbtree_stat_file_in;
126 int num_entries; /* number of ratecontrol_entry_ts */
127 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
129 double last_qscale_for[3]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
130 int last_non_b_pict_type;
131 double accum_p_qp; /* for determining I-frame quant */
133 double last_accum_p_norm;
134 double lmin[3]; /* min qscale by frame type */
136 double lstep; /* max change (multiply) in qscale per frame */
137 uint16_t *qp_buffer[2]; /* Global buffers for converting MB-tree quantizer data. */
138 int qpbuf_pos; /* In order to handle pyramid reordering, QP buffer acts as a stack.
139 * This value is the current position (0 or 1). */
142 float frame_size_estimated; /* Access to this variable must be atomic: double is
143 * not atomic on all arches we care about */
144 double frame_size_maximum; /* Maximum frame size due to MinCR */
145 double frame_size_planned;
146 double slice_size_planned;
147 predictor_t (*row_pred)[2];
148 predictor_t row_preds[3][2];
149 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
150 int bframes; /* # consecutive B-frames before this P-frame */
151 int bframe_bits; /* total cost of those frames */
155 x264_zone_t *prev_zone;
158 int initial_cpb_removal_delay;
159 int initial_cpb_removal_delay_offset;
160 double nrt_first_access_unit; /* nominal removal time */
161 double previous_cpb_final_arrival_time;
162 uint64_t hrd_multiply_denom;
166 static int parse_zones( x264_t *h );
167 static int init_pass2(x264_t *);
168 static float rate_estimate_qscale( x264_t *h );
169 static int update_vbv( x264_t *h, int bits );
170 static void update_vbv_plan( x264_t *h, int overhead );
171 static double predict_size( predictor_t *p, double q, double var );
172 static void update_predictor( predictor_t *p, double q, double var, double bits );
174 #define CMP_OPT_FIRST_PASS( opt, param_val )\
176 if( ( p = strstr( opts, opt "=" ) ) && sscanf( p, opt "=%d" , &i ) && param_val != i )\
178 x264_log( h, X264_LOG_ERROR, "different " opt " setting than first pass (%d vs %d)\n", param_val, i );\
184 * qp = h.264's quantizer
185 * qscale = linearized quantizer = Lagrange multiplier
187 static inline double qp2qscale( double qp )
189 return 0.85 * pow( 2.0, ( qp - 12.0 ) / 6.0 );
191 static inline double qscale2qp( double qscale )
193 return 12.0 + 6.0 * log2( qscale/0.85 );
196 /* Texture bitrate is not quite inversely proportional to qscale,
197 * probably due the the changing number of SKIP blocks.
198 * MV bits level off at about qp<=12, because the lambda used
199 * for motion estimation is constant there. */
200 static inline double qscale2bits( ratecontrol_entry_t *rce, double qscale )
204 return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
205 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
209 static ALWAYS_INLINE uint32_t ac_energy_var( uint64_t sum_ssd, int shift, x264_frame_t *frame, int i )
211 uint32_t sum = sum_ssd;
212 uint32_t ssd = sum_ssd >> 32;
213 frame->i_pixel_sum[i] += sum;
214 frame->i_pixel_ssd[i] += ssd;
215 return ssd - ((uint64_t)sum * sum >> shift);
218 static ALWAYS_INLINE uint32_t ac_energy_plane( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame, int i )
221 int stride = frame->i_stride[i];
222 int offset = h->mb.b_interlaced
223 ? 16 * mb_x + w * (mb_y&~1) * stride + (mb_y&1) * stride
224 : 16 * mb_x + w * mb_y * stride;
225 stride <<= h->mb.b_interlaced;
228 ALIGNED_ARRAY_16( pixel, pix,[FENC_STRIDE*8] );
229 h->mc.load_deinterleave_8x8x2_fenc( pix, frame->plane[1] + offset, stride );
230 return ac_energy_var( h->pixf.var[PIXEL_8x8]( pix, FENC_STRIDE ), 6, frame, 1 )
231 + ac_energy_var( h->pixf.var[PIXEL_8x8]( pix+FENC_STRIDE/2, FENC_STRIDE ), 6, frame, 2 );
234 return ac_energy_var( h->pixf.var[PIXEL_16x16]( frame->plane[0] + offset, stride ), 8, frame, 0 );
237 // Find the total AC energy of the block in all planes.
238 static NOINLINE uint32_t x264_ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
240 /* This function contains annoying hacks because GCC has a habit of reordering emms
241 * and putting it after floating point ops. As a result, we put the emms at the end of the
242 * function and make sure that its always called before the float math. Noinline makes
243 * sure no reordering goes on. */
244 uint32_t var = ac_energy_plane( h, mb_x, mb_y, frame, 0 );
245 var += ac_energy_plane( h, mb_x, mb_y, frame, 1 );
250 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame, float *quant_offsets )
252 /* constants chosen to result in approximately the same overall bitrate as without AQ.
253 * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
256 /* Initialize frame stats */
257 for( int i = 0; i < 3; i++ )
259 frame->i_pixel_sum[i] = 0;
260 frame->i_pixel_ssd[i] = 0;
263 /* Degenerate cases */
264 if( h->param.rc.i_aq_mode == X264_AQ_NONE || h->param.rc.f_aq_strength == 0 )
266 /* Need to init it anyways for MB tree */
267 if( h->param.rc.i_aq_mode && h->param.rc.f_aq_strength == 0 )
271 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
272 frame->f_qp_offset[mb_xy] = frame->f_qp_offset_aq[mb_xy] = quant_offsets[mb_xy];
273 if( h->frames.b_have_lowres )
274 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
275 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8( frame->f_qp_offset[mb_xy] );
279 memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
280 memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
281 if( h->frames.b_have_lowres )
282 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
283 frame->i_inv_qscale_factor[mb_xy] = 256;
286 /* Need variance data for weighted prediction */
287 if( h->param.analyse.i_weighted_pred )
289 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
290 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
291 x264_ac_energy_mb( h, mb_x, mb_y, frame );
296 /* Actual adaptive quantization */
299 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
301 float bit_depth_correction = powf(1 << (BIT_DEPTH-8), 0.5f);
302 float avg_adj_pow2 = 0.f;
303 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
304 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
306 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
307 float qp_adj = powf( energy + 1, 0.125f );
308 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
310 avg_adj_pow2 += qp_adj * qp_adj;
312 avg_adj /= h->mb.i_mb_count;
313 avg_adj_pow2 /= h->mb.i_mb_count;
314 strength = h->param.rc.f_aq_strength * avg_adj / bit_depth_correction;
315 avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - (14.f * bit_depth_correction)) / avg_adj;
318 strength = h->param.rc.f_aq_strength * 1.0397f;
320 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
321 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
324 int mb_xy = mb_x + mb_y*h->mb.i_mb_stride;
325 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
327 qp_adj = frame->f_qp_offset[mb_xy];
328 qp_adj = strength * (qp_adj - avg_adj);
332 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
333 qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - (14.427f + 2*(BIT_DEPTH-8)));
336 qp_adj += quant_offsets[mb_xy];
337 frame->f_qp_offset[mb_xy] =
338 frame->f_qp_offset_aq[mb_xy] = qp_adj;
339 if( h->frames.b_have_lowres )
340 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8(qp_adj);
344 /* Remove mean from SSD calculation */
345 for( int i = 0; i < 3; i++ )
347 uint64_t ssd = frame->i_pixel_ssd[i];
348 uint64_t sum = frame->i_pixel_sum[i];
349 int width = h->mb.i_mb_width*16>>!!i;
350 int height = h->mb.i_mb_height*16>>!!i;
351 frame->i_pixel_ssd[i] = ssd - (sum * sum + width * height / 2) / (width * height);
355 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame, float *quant_offsets )
357 x264_ratecontrol_t *rc = h->rc;
358 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
360 if( rc->entry[frame->i_frame].kept_as_ref )
363 if( rc->qpbuf_pos < 0 )
369 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
371 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 )
374 if( i_type != i_type_actual && rc->qpbuf_pos == 1 )
376 x264_log(h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type, i_type_actual);
379 } while( i_type != i_type_actual );
382 for( int i = 0; i < h->mb.i_mb_count; i++ )
384 frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[rc->qpbuf_pos][i] )) * (1/256.0);
385 if( h->frames.b_have_lowres )
386 frame->i_inv_qscale_factor[i] = x264_exp2fix8(frame->f_qp_offset[i]);
391 x264_stack_align( x264_adaptive_quant_frame, h, frame, quant_offsets );
394 x264_log(h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n");
398 int x264_reference_build_list_optimal( x264_t *h )
400 ratecontrol_entry_t *rce = h->rc->rce;
401 x264_frame_t *frames[16];
402 x264_weight_t weights[16][3];
405 if( rce->refs != h->i_ref[0] )
408 memcpy( frames, h->fref[0], sizeof(frames) );
409 memcpy( refcount, rce->refcount, sizeof(refcount) );
410 memcpy( weights, h->fenc->weight, sizeof(weights) );
411 memset( &h->fenc->weight[1][0], 0, sizeof(x264_weight_t[15][3]) );
413 /* For now don't reorder ref 0; it seems to lower quality
414 in most cases due to skips. */
415 for( int ref = 1; ref < h->i_ref[0]; ref++ )
420 for( int i = 1; i < h->i_ref[0]; i++ )
421 /* Favor lower POC as a tiebreaker. */
422 COPY2_IF_GT( max, refcount[i], bestref, i );
424 /* FIXME: If there are duplicates from frames other than ref0 then it is possible
425 * that the optimal ordering doesnt place every duplicate. */
427 refcount[bestref] = -1;
428 h->fref[0][ref] = frames[bestref];
429 memcpy( h->fenc->weight[ref], weights[bestref], sizeof(weights[bestref]) );
435 static char *x264_strcat_filename( char *input, char *suffix )
437 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
440 strcpy( output, input );
441 strcat( output, suffix );
445 void x264_ratecontrol_init_reconfigurable( x264_t *h, int b_init )
447 x264_ratecontrol_t *rc = h->rc;
448 if( !b_init && rc->b_2pass )
451 if( h->param.rc.i_rc_method == X264_RC_CRF )
453 /* Arbitrary rescaling to make CRF somewhat similar to QP.
454 * Try to compensate for MB-tree's effects as well. */
455 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
456 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
457 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
458 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset + QP_BD_OFFSET );
461 if( h->param.rc.i_vbv_max_bitrate > 0 && h->param.rc.i_vbv_buffer_size > 0 )
463 if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
465 h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
466 x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
467 h->param.rc.i_vbv_buffer_size );
470 /* We don't support changing the ABR bitrate right now,
471 so if the stream starts as CBR, keep it CBR. */
472 if( rc->b_vbv_min_rate )
473 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
475 int vbv_buffer_size = h->param.rc.i_vbv_buffer_size * 1000;
476 int vbv_max_bitrate = h->param.rc.i_vbv_max_bitrate * 1000;
479 h->sps->vui.hrd.i_bit_rate_unscaled = vbv_max_bitrate;
480 h->sps->vui.hrd.i_cpb_size_unscaled = vbv_buffer_size;
481 if( h->param.i_nal_hrd && b_init )
483 h->sps->vui.hrd.i_cpb_cnt = 1;
484 h->sps->vui.hrd.b_cbr_hrd = h->param.i_nal_hrd == X264_NAL_HRD_CBR;
485 h->sps->vui.hrd.i_time_offset_length = 0;
490 int bitrate = 1000*h->param.rc.i_vbv_max_bitrate;
491 int bufsize = 1000*h->param.rc.i_vbv_buffer_size;
493 // normalize HRD size and rate to the value / scale notation
494 h->sps->vui.hrd.i_bit_rate_scale = x264_clip3( x264_ctz( bitrate ) - BR_SHIFT, 0, 15 );
495 h->sps->vui.hrd.i_bit_rate_value = bitrate >> ( h->sps->vui.hrd.i_bit_rate_scale + BR_SHIFT );
496 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 );
497 h->sps->vui.hrd.i_cpb_size_scale = x264_clip3( x264_ctz( bufsize ) - CPB_SHIFT, 0, 15 );
498 h->sps->vui.hrd.i_cpb_size_value = bufsize >> ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
499 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 );
505 #define MAX_DURATION 0.5
507 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 );
508 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;
509 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);
511 h->sps->vui.hrd.i_initial_cpb_removal_delay_length = 2 + x264_clip3( 32 - x264_clz( max_delay ), 4, 22 );
512 h->sps->vui.hrd.i_cpb_removal_delay_length = x264_clip3( 32 - x264_clz( max_cpb_output_delay ), 4, 31 );
513 h->sps->vui.hrd.i_dpb_output_delay_length = x264_clip3( 32 - x264_clz( max_dpb_output_delay ), 4, 31 );
517 vbv_buffer_size = h->sps->vui.hrd.i_cpb_size_unscaled;
518 vbv_max_bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
520 else if( h->param.i_nal_hrd && !b_init )
522 x264_log( h, X264_LOG_WARNING, "VBV parameters cannot be changed when NAL HRD is in use\n" );
526 rc->buffer_rate = vbv_max_bitrate / rc->fps;
527 rc->vbv_max_rate = vbv_max_bitrate;
528 rc->buffer_size = vbv_buffer_size;
529 rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
530 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
531 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
532 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.f_rf_constant_max )
534 rc->rate_factor_max_increment = h->param.rc.f_rf_constant_max - h->param.rc.f_rf_constant;
535 if( rc->rate_factor_max_increment <= 0 )
537 x264_log( h, X264_LOG_WARNING, "CRF max must be greater than CRF\n" );
538 rc->rate_factor_max_increment = 0;
543 if( h->param.rc.f_vbv_buffer_init > 1. )
544 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 );
545 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);
546 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init * h->sps->vui.i_time_scale;
548 rc->b_vbv_min_rate = !rc->b_2pass
549 && h->param.rc.i_rc_method == X264_RC_ABR
550 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
555 int x264_ratecontrol_new( x264_t *h )
557 x264_ratecontrol_t *rc;
561 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
564 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
565 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
567 /* FIXME: use integers */
568 if( h->param.i_fps_num > 0 && h->param.i_fps_den > 0 )
569 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
573 if( h->param.rc.b_mb_tree )
575 h->param.rc.f_pb_factor = 1;
579 rc->qcompress = h->param.rc.f_qcompress;
581 rc->bitrate = h->param.rc.i_bitrate * 1000.;
582 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
583 rc->nmb = h->mb.i_mb_count;
584 rc->last_non_b_pict_type = -1;
587 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
589 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
593 x264_ratecontrol_init_reconfigurable( h, 1 );
595 if( h->param.i_nal_hrd )
597 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale;
598 uint64_t num = 180000;
599 x264_reduce_fraction64( &num, &denom );
600 rc->hrd_multiply_denom = 180000 / num;
602 double bits_required = log2( 180000 / rc->hrd_multiply_denom )
603 + log2( h->sps->vui.i_time_scale )
604 + log2( h->sps->vui.hrd.i_cpb_size_unscaled );
605 if( bits_required >= 63 )
607 x264_log( h, X264_LOG_ERROR, "HRD with very large timescale and bufsize not supported\n" );
612 if( rc->rate_tolerance < 0.01 )
614 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
615 rc->rate_tolerance = 0.01;
618 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
622 /* FIXME ABR_INIT_QP is actually used only in CRF */
623 #define ABR_INIT_QP (( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 ) + QP_BD_OFFSET)
624 rc->accum_p_norm = .01;
625 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
626 /* estimated ratio that produces a reasonable QP for the first I-frame */
627 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
628 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
629 rc->last_non_b_pict_type = SLICE_TYPE_I;
632 rc->ip_offset = 6.0 * log2f( h->param.rc.f_ip_factor );
633 rc->pb_offset = 6.0 * log2f( h->param.rc.f_pb_factor );
634 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
635 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, QP_MAX );
636 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, QP_MAX );
637 h->mb.ip_offset = rc->ip_offset + 0.5;
639 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
640 rc->last_qscale = qp2qscale( 26 );
641 int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
642 CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
643 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
644 for( int i = 0; i < 3; i++ )
646 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
647 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
648 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
649 for( int j = 0; j < num_preds; j++ )
651 rc->pred[i+j*5].coeff= 2.0;
652 rc->pred[i+j*5].count= 1.0;
653 rc->pred[i+j*5].decay= 0.5;
654 rc->pred[i+j*5].offset= 0.0;
656 for( int j = 0; j < 2; j++ )
658 rc->row_preds[i][j].coeff= .25;
659 rc->row_preds[i][j].count= 1.0;
660 rc->row_preds[i][j].decay= 0.5;
661 rc->row_preds[i][j].offset= 0.0;
664 *rc->pred_b_from_p = rc->pred[0];
666 if( parse_zones( h ) < 0 )
668 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
672 /* Load stat file and init 2pass algo */
673 if( h->param.rc.b_stat_read )
675 char *p, *stats_in, *stats_buf;
677 /* read 1st pass stats */
678 assert( h->param.rc.psz_stat_in );
679 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
682 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
685 if( h->param.rc.b_mb_tree )
687 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
688 if( !mbtree_stats_in )
690 rc->p_mbtree_stat_file_in = fopen( mbtree_stats_in, "rb" );
691 x264_free( mbtree_stats_in );
692 if( !rc->p_mbtree_stat_file_in )
694 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
699 /* check whether 1st pass options were compatible with current options */
700 if( !strncmp( stats_buf, "#options:", 9 ) )
704 char *opts = stats_buf;
705 stats_in = strchr( stats_buf, '\n' );
710 if( sscanf( opts, "#options: %dx%d", &i, &j ) != 2 )
712 x264_log( h, X264_LOG_ERROR, "resolution specified in stats file not valid\n" );
715 else if( h->param.rc.b_mb_tree && (i != h->param.i_width || j != h->param.i_height) )
717 x264_log( h, X264_LOG_ERROR, "MB-tree doesn't support different resolution than 1st pass (%dx%d vs %dx%d)\n",
718 h->param.i_width, h->param.i_height, i, j );
722 if( ( p = strstr( opts, "timebase=" ) ) && sscanf( p, "timebase=%u/%u", &k, &l ) != 2 )
724 x264_log( h, X264_LOG_ERROR, "timebase specified in stats file not valid\n" );
727 if( k != h->param.i_timebase_num || l != h->param.i_timebase_den )
729 x264_log( h, X264_LOG_ERROR, "timebase mismatch with 1st pass (%u/%u vs %u/%u)\n",
730 h->param.i_timebase_num, h->param.i_timebase_den, k, l );
734 CMP_OPT_FIRST_PASS( "bitdepth", BIT_DEPTH );
735 CMP_OPT_FIRST_PASS( "weightp", X264_MAX( 0, h->param.analyse.i_weighted_pred ) );
736 CMP_OPT_FIRST_PASS( "bframes", h->param.i_bframe );
737 CMP_OPT_FIRST_PASS( "b_pyramid", h->param.i_bframe_pyramid );
738 CMP_OPT_FIRST_PASS( "intra_refresh", h->param.b_intra_refresh );
739 CMP_OPT_FIRST_PASS( "open_gop", h->param.i_open_gop );
741 if( (p = strstr( opts, "keyint=" )) )
744 char buf[13] = "infinite ";
745 if( h->param.i_keyint_max != X264_KEYINT_MAX_INFINITE )
746 sprintf( buf, "%d ", h->param.i_keyint_max );
747 if( strncmp( p, buf, strlen(buf) ) )
749 x264_log( h, X264_LOG_ERROR, "different keyint setting than first pass (%.*s vs %.*s)\n",
750 strlen(buf)-1, buf, strcspn(p, " "), p );
755 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
756 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
758 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
760 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
761 h->mb.b_direct_auto_write = 1;
764 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
765 h->param.i_bframe_adaptive = i;
766 else if( h->param.i_bframe )
768 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
772 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 ) )
773 h->param.rc.i_lookahead = i;
776 /* find number of pics */
779 for( num_entries = -1; p; num_entries++ )
780 p = strchr( p + 1, ';' );
783 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
786 rc->num_entries = num_entries;
788 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
790 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
791 h->param.i_frame_total, rc->num_entries );
793 if( h->param.i_frame_total > rc->num_entries )
795 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
796 h->param.i_frame_total, rc->num_entries );
800 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
802 /* init all to skipped p frames */
803 for( int i = 0; i < rc->num_entries; i++ )
805 ratecontrol_entry_t *rce = &rc->entry[i];
806 rce->pict_type = SLICE_TYPE_P;
807 rce->qscale = rce->new_qscale = qp2qscale( 20 );
808 rce->misc_bits = rc->nmb + 10;
814 for( int i = 0; i < rc->num_entries; i++ )
816 ratecontrol_entry_t *rce;
824 next= strchr(p, ';');
826 *next++ = 0; //sscanf is unbelievably slow on long strings
827 e = sscanf( p, " in:%d ", &frame_number );
829 if( frame_number < 0 || frame_number >= rc->num_entries )
831 x264_log( h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i );
834 rce = &rc->entry[frame_number];
835 rce->direct_mode = 0;
837 e += sscanf( p, " in:%*d out:%*d type:%c dur:%"SCNd64" cpbdur:%"SCNd64" q:%f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c",
838 &pict_type, &rce->i_duration, &rce->i_cpb_duration, &qp, &rce->tex_bits,
839 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
840 &rce->s_count, &rce->direct_mode );
842 p = strstr( p, "ref:" );
846 for( ref = 0; ref < 16; ref++ )
848 if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
850 p = strchr( p+1, ' ' );
857 rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
858 char *w = strchr( p, 'w' );
861 int count = sscanf( w, "w:%hd,%hd,%hd,%hd,%hd,%hd,%hd,%hd",
862 &rce->i_weight_denom[0], &rce->weight[0][0], &rce->weight[0][1],
863 &rce->i_weight_denom[1], &rce->weight[1][0], &rce->weight[1][1],
864 &rce->weight[2][0], &rce->weight[2][1] );
866 rce->i_weight_denom[1] = -1;
867 else if ( count != 8 )
868 rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
871 if( pict_type != 'b' )
872 rce->kept_as_ref = 1;
876 rce->frame_type = X264_TYPE_IDR;
877 rce->pict_type = SLICE_TYPE_I;
880 rce->frame_type = X264_TYPE_I;
881 rce->pict_type = SLICE_TYPE_I;
884 rce->frame_type = X264_TYPE_P;
885 rce->pict_type = SLICE_TYPE_P;
888 rce->frame_type = X264_TYPE_BREF;
889 rce->pict_type = SLICE_TYPE_B;
892 rce->frame_type = X264_TYPE_B;
893 rce->pict_type = SLICE_TYPE_B;
895 default: e = -1; break;
900 x264_log( h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e );
903 rce->qscale = qp2qscale( qp );
907 x264_free( stats_buf );
909 if( h->param.rc.i_rc_method == X264_RC_ABR )
911 if( init_pass2( h ) < 0 )
913 } /* else we're using constant quant, so no need to run the bitrate allocation */
916 /* Open output file */
917 /* If input and output files are the same, output to a temp file
918 * and move it to the real name only when it's complete */
919 if( h->param.rc.b_stat_write )
922 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
923 if( !rc->psz_stat_file_tmpname )
926 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
927 if( rc->p_stat_file_out == NULL )
929 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
933 p = x264_param2string( &h->param, 1 );
935 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
937 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
939 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
940 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
941 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
944 rc->p_mbtree_stat_file_out = fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
945 if( rc->p_mbtree_stat_file_out == NULL )
947 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
953 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
955 CHECKED_MALLOC( rc->qp_buffer[0], h->mb.i_mb_count * sizeof(uint16_t) );
956 if( h->param.i_bframe_pyramid && h->param.rc.b_stat_read )
957 CHECKED_MALLOC( rc->qp_buffer[1], h->mb.i_mb_count * sizeof(uint16_t) );
961 for( int i = 0; i<h->param.i_threads; i++ )
963 h->thread[i]->rc = rc+i;
967 h->thread[i]->param = h->param;
968 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
969 h->thread[i]->mb.ip_offset = h->mb.ip_offset;
978 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
981 char *tok, UNUSED *saveptr=NULL;
983 z->f_bitrate_factor = 1;
984 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
986 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
988 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
992 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
998 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
999 memcpy( z->param, &h->param, sizeof(x264_param_t) );
1000 z->param->param_free = x264_free;
1001 while( (tok = strtok_r( p, ",", &saveptr )) )
1003 char *val = strchr( tok, '=' );
1009 if( x264_param_parse( z->param, tok, val ) )
1011 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
1021 static int parse_zones( x264_t *h )
1023 x264_ratecontrol_t *rc = h->rc;
1024 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
1026 char *psz_zones, *p;
1027 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
1028 strcpy( psz_zones, h->param.rc.psz_zones );
1029 h->param.rc.i_zones = 1;
1030 for( p = psz_zones; *p; p++ )
1031 h->param.rc.i_zones += (*p == '/');
1032 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
1034 for( int i = 0; i < h->param.rc.i_zones; i++ )
1036 int i_tok = strcspn( p, "/" );
1038 if( parse_zone( h, &h->param.rc.zones[i], p ) )
1042 x264_free( psz_zones );
1045 if( h->param.rc.i_zones > 0 )
1047 for( int i = 0; i < h->param.rc.i_zones; i++ )
1049 x264_zone_t z = h->param.rc.zones[i];
1050 if( z.i_start < 0 || z.i_start > z.i_end )
1052 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
1053 z.i_start, z.i_end );
1056 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
1058 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
1059 z.f_bitrate_factor );
1064 rc->i_zones = h->param.rc.i_zones + 1;
1065 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
1066 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
1068 // default zone to fall back to if none of the others match
1069 rc->zones[0].i_start = 0;
1070 rc->zones[0].i_end = INT_MAX;
1071 rc->zones[0].b_force_qp = 0;
1072 rc->zones[0].f_bitrate_factor = 1;
1073 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
1074 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
1075 for( int i = 1; i < rc->i_zones; i++ )
1077 if( !rc->zones[i].param )
1078 rc->zones[i].param = rc->zones[0].param;
1087 static x264_zone_t *get_zone( x264_t *h, int frame_num )
1089 for( int i = h->rc->i_zones - 1; i >= 0; i-- )
1091 x264_zone_t *z = &h->rc->zones[i];
1092 if( frame_num >= z->i_start && frame_num <= z->i_end )
1098 void x264_ratecontrol_summary( x264_t *h )
1100 x264_ratecontrol_t *rc = h->rc;
1101 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
1103 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
1104 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
1105 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
1106 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
1107 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset - QP_BD_OFFSET );
1111 void x264_ratecontrol_delete( x264_t *h )
1113 x264_ratecontrol_t *rc = h->rc;
1116 if( rc->p_stat_file_out )
1118 b_regular_file = x264_is_regular_file( rc->p_stat_file_out );
1119 fclose( rc->p_stat_file_out );
1120 if( h->i_frame >= rc->num_entries && b_regular_file )
1121 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
1123 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1124 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
1126 x264_free( rc->psz_stat_file_tmpname );
1128 if( rc->p_mbtree_stat_file_out )
1130 b_regular_file = x264_is_regular_file( rc->p_mbtree_stat_file_out );
1131 fclose( rc->p_mbtree_stat_file_out );
1132 if( h->i_frame >= rc->num_entries && b_regular_file )
1133 if( rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
1135 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1136 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
1138 x264_free( rc->psz_mbtree_stat_file_tmpname );
1139 x264_free( rc->psz_mbtree_stat_file_name );
1141 if( rc->p_mbtree_stat_file_in )
1142 fclose( rc->p_mbtree_stat_file_in );
1143 x264_free( rc->pred );
1144 x264_free( rc->pred_b_from_p );
1145 x264_free( rc->entry );
1146 x264_free( rc->qp_buffer[0] );
1147 x264_free( rc->qp_buffer[1] );
1150 x264_free( rc->zones[0].param );
1151 for( int i = 1; i < rc->i_zones; i++ )
1152 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
1153 rc->zones[i].param->param_free( rc->zones[i].param );
1154 x264_free( rc->zones );
1159 static void accum_p_qp_update( x264_t *h, float qp )
1161 x264_ratecontrol_t *rc = h->rc;
1162 rc->accum_p_qp *= .95;
1163 rc->accum_p_norm *= .95;
1164 rc->accum_p_norm += 1;
1165 if( h->sh.i_type == SLICE_TYPE_I )
1166 rc->accum_p_qp += qp + rc->ip_offset;
1168 rc->accum_p_qp += qp;
1171 /* Before encoding a frame, choose a QP for it */
1172 void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
1174 x264_ratecontrol_t *rc = h->rc;
1175 ratecontrol_entry_t *rce = NULL;
1176 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
1181 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
1182 x264_encoder_reconfig( h, zone->param );
1183 rc->prev_zone = zone;
1185 if( h->param.rc.b_stat_read )
1187 int frame = h->fenc->i_frame;
1188 assert( frame >= 0 && frame < rc->num_entries );
1189 rce = h->rc->rce = &h->rc->entry[frame];
1191 if( h->sh.i_type == SLICE_TYPE_B
1192 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
1194 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
1195 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
1201 memset( h->fdec->i_row_bits, 0, h->mb.i_mb_height * sizeof(int) );
1202 rc->row_pred = &rc->row_preds[h->sh.i_type];
1203 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;
1204 update_vbv_plan( h, overhead );
1206 const x264_level_t *l = x264_levels;
1207 while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
1210 int mincr = l->mincr;
1212 /* Blu-ray requires this */
1213 if( l->level_idc == 41 && h->param.i_nal_hrd )
1216 /* High 10 doesn't require minCR, so just set the maximum to a large value. */
1217 if( h->sps->i_profile_idc == PROFILE_HIGH10 )
1218 rc->frame_size_maximum = 1e9;
1221 /* The spec has a bizarre special case for the first frame. */
1222 if( h->i_frame == 0 )
1224 //384 * ( Max( PicSizeInMbs, fR * MaxMBPS ) + MaxMBPS * ( tr( 0 ) - tr,n( 0 ) ) ) / MinCR
1225 double fr = 1. / 172;
1226 int pic_size_in_mbs = h->mb.i_mb_width * h->mb.i_mb_height;
1227 rc->frame_size_maximum = 384 * BIT_DEPTH * X264_MAX( pic_size_in_mbs, fr*l->mbps ) / mincr;
1231 //384 * MaxMBPS * ( tr( n ) - tr( n - 1 ) ) / MinCR
1232 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;
1237 if( h->sh.i_type != SLICE_TYPE_B )
1238 rc->bframes = h->fenc->i_bframes;
1240 if( i_force_qp != X264_QP_AUTO )
1244 else if( rc->b_abr )
1246 q = qscale2qp( rate_estimate_qscale( h ) );
1248 else if( rc->b_2pass )
1250 rce->new_qscale = rate_estimate_qscale( h );
1251 q = qscale2qp( rce->new_qscale );
1255 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
1256 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
1258 q = rc->qp_constant[ h->sh.i_type ];
1262 if( zone->b_force_qp )
1263 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1265 q -= 6*log2f( zone->f_bitrate_factor );
1269 q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1273 rc->qp = x264_clip3( (int)(q + 0.5), 0, QP_MAX );
1274 h->fdec->f_qp_avg_rc =
1275 h->fdec->f_qp_avg_aq =
1278 rce->new_qp = rc->qp;
1280 accum_p_qp_update( h, rc->qpm );
1282 if( h->sh.i_type != SLICE_TYPE_B )
1283 rc->last_non_b_pict_type = h->sh.i_type;
1286 static double predict_row_size( x264_t *h, int y, double qp )
1288 /* average between two predictors:
1289 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1290 x264_ratecontrol_t *rc = h->rc;
1291 double pred_s = predict_size( rc->row_pred[0], qp2qscale( qp ), h->fdec->i_row_satd[y] );
1293 if( h->sh.i_type == SLICE_TYPE_I || qp >= h->fref[0][0]->f_row_qp[y] )
1295 if( h->sh.i_type == SLICE_TYPE_P
1296 && h->fref[0][0]->i_type == h->fdec->i_type
1297 && h->fref[0][0]->i_row_satd[y] > 0
1298 && (abs(h->fref[0][0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1300 pred_t = h->fref[0][0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref[0][0]->i_row_satd[y]
1301 * qp2qscale( h->fref[0][0]->f_row_qp[y] ) / qp2qscale( qp );
1305 return (pred_s + pred_t) / 2;
1307 /* Our QP is lower than the reference! */
1310 double pred_intra = predict_size( rc->row_pred[1], qp2qscale( qp ), h->fdec->i_row_satds[0][0][y] );
1311 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1312 return pred_intra + pred_s;
1316 static double row_bits_so_far( x264_t *h, int y )
1319 for( int i = h->i_threadslice_start; i <= y; i++ )
1320 bits += h->fdec->i_row_bits[i];
1324 static double predict_row_size_sum( x264_t *h, int y, double qp )
1326 double bits = row_bits_so_far(h, y);
1327 for( int i = y+1; i < h->i_threadslice_end; i++ )
1328 bits += predict_row_size( h, i, qp );
1333 void x264_ratecontrol_mb( x264_t *h, int bits )
1335 x264_ratecontrol_t *rc = h->rc;
1336 const int y = h->mb.i_mb_y;
1340 h->fdec->i_row_bits[y] += bits;
1341 rc->qpa_rc += rc->qpm;
1342 rc->qpa_aq += h->mb.i_qp;
1344 if( h->mb.i_mb_x != h->mb.i_mb_width - 1 || !rc->b_vbv )
1347 h->fdec->f_row_qp[y] = rc->qpm;
1349 update_predictor( rc->row_pred[0], qp2qscale( rc->qpm ), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1350 if( h->sh.i_type == SLICE_TYPE_P && rc->qpm < h->fref[0][0]->f_row_qp[y] )
1351 update_predictor( rc->row_pred[1], qp2qscale( rc->qpm ), h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1353 /* tweak quality based on difference from predicted size */
1354 if( y < h->i_threadslice_end-1 )
1356 float prev_row_qp = h->fdec->f_row_qp[y];
1357 float qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1358 float qp_absolute_max = h->param.rc.i_qp_max;
1359 if( rc->rate_factor_max_increment )
1360 qp_absolute_max = X264_MIN( qp_absolute_max, rc->qp_novbv + rc->rate_factor_max_increment );
1361 float qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, qp_absolute_max );
1362 float step_size = 0.5;
1364 /* B-frames shouldn't use lower QP than their reference frames. */
1365 if( h->sh.i_type == SLICE_TYPE_B )
1367 qp_min = X264_MAX( qp_min, X264_MAX( h->fref[0][0]->f_row_qp[y+1], h->fref[1][0]->f_row_qp[y+1] ) );
1368 rc->qpm = X264_MAX( rc->qpm, qp_min );
1371 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1372 float slice_size_planned = h->param.b_sliced_threads ? rc->slice_size_planned : rc->frame_size_planned;
1373 float max_frame_error = X264_MAX( 0.05, 1.0 / (h->mb.i_mb_height) );
1374 float size_of_other_slices = 0;
1375 if( h->param.b_sliced_threads )
1377 float size_of_other_slices_planned = 0;
1378 for( int i = 0; i < h->param.i_threads; i++ )
1379 if( h != h->thread[i] )
1381 size_of_other_slices += h->thread[i]->rc->frame_size_estimated;
1382 size_of_other_slices_planned += h->thread[i]->rc->slice_size_planned;
1384 float weight = rc->slice_size_planned / rc->frame_size_planned;
1385 size_of_other_slices = (size_of_other_slices - size_of_other_slices_planned) * weight + size_of_other_slices_planned;
1388 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1389 float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
1390 int b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1392 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1393 /* area at the top of the frame was measured inaccurately. */
1394 if( row_bits_so_far( h, y ) < 0.05 * slice_size_planned )
1397 if( h->sh.i_type != SLICE_TYPE_I )
1400 if( !rc->b_vbv_min_rate )
1401 qp_min = X264_MAX( qp_min, rc->qp_novbv );
1403 while( rc->qpm < qp_max
1404 && ((b1 > rc->frame_size_planned + rc_tol) ||
1405 (rc->buffer_fill - b1 < buffer_left_planned * 0.5) ||
1406 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1408 rc->qpm += step_size;
1409 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1412 while( rc->qpm > qp_min
1413 && (rc->qpm > h->fdec->f_row_qp[0] || rc->single_frame_vbv)
1414 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
1415 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
1417 rc->qpm -= step_size;
1418 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1421 /* avoid VBV underflow or MinCR violation */
1422 while( (rc->qpm < qp_absolute_max)
1423 && ((rc->buffer_fill - b1 < rc->buffer_rate * max_frame_error) ||
1424 (rc->frame_size_maximum - b1 < rc->frame_size_maximum * max_frame_error)))
1426 rc->qpm += step_size;
1427 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1430 h->rc->frame_size_estimated = predict_row_size_sum( h, y, rc->qpm );
1434 int x264_ratecontrol_qp( x264_t *h )
1437 return x264_clip3( h->rc->qpm + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1440 int x264_ratecontrol_mb_qp( x264_t *h )
1443 float qp = h->rc->qpm;
1444 if( h->param.rc.i_aq_mode )
1445 /* MB-tree currently doesn't adjust quantizers in unreferenced frames. */
1446 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];
1447 return x264_clip3( qp + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1450 /* In 2pass, force the same frame types as in the 1st pass */
1451 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1453 x264_ratecontrol_t *rc = h->rc;
1454 if( h->param.rc.b_stat_read )
1456 if( frame_num >= rc->num_entries )
1458 /* We could try to initialize everything required for ABR and
1459 * adaptive B-frames, but that would be complicated.
1460 * So just calculate the average QP used so far. */
1461 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24 + QP_BD_OFFSET
1462 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1463 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, QP_MAX );
1464 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 );
1465 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 );
1467 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1468 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1469 if( h->param.i_bframe_adaptive )
1470 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1472 for( int i = 0; i < h->param.i_threads; i++ )
1474 h->thread[i]->rc->b_abr = 0;
1475 h->thread[i]->rc->b_2pass = 0;
1476 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1477 h->thread[i]->param.rc.b_stat_read = 0;
1478 h->thread[i]->param.i_bframe_adaptive = 0;
1479 h->thread[i]->param.i_scenecut_threshold = 0;
1480 h->thread[i]->param.rc.b_mb_tree = 0;
1481 if( h->thread[i]->param.i_bframe > 1 )
1482 h->thread[i]->param.i_bframe = 1;
1484 return X264_TYPE_AUTO;
1486 return rc->entry[frame_num].frame_type;
1489 return X264_TYPE_AUTO;
1492 void x264_ratecontrol_set_weights( x264_t *h, x264_frame_t *frm )
1494 ratecontrol_entry_t *rce = &h->rc->entry[frm->i_frame];
1495 if( h->param.analyse.i_weighted_pred <= 0 )
1498 if( rce->i_weight_denom[0] >= 0 )
1499 SET_WEIGHT( frm->weight[0][0], 1, rce->weight[0][0], rce->i_weight_denom[0], rce->weight[0][1] );
1501 if( rce->i_weight_denom[1] >= 0 )
1503 SET_WEIGHT( frm->weight[0][1], 1, rce->weight[1][0], rce->i_weight_denom[1], rce->weight[1][1] );
1504 SET_WEIGHT( frm->weight[0][2], 1, rce->weight[2][0], rce->i_weight_denom[1], rce->weight[2][1] );
1508 /* After encoding one frame, save stats and update ratecontrol state */
1509 int x264_ratecontrol_end( x264_t *h, int bits, int *filler )
1511 x264_ratecontrol_t *rc = h->rc;
1512 const int *mbs = h->stat.frame.i_mb_count;
1516 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1517 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1518 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1519 for( int i = B_DIRECT; i < B_8x8; i++ )
1520 h->stat.frame.i_mb_count_p += mbs[i];
1522 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1523 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1525 if( h->param.rc.b_stat_write )
1527 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1528 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1529 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1530 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1531 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1532 char c_direct = h->mb.b_direct_auto_write ?
1533 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1534 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1536 if( fprintf( rc->p_stat_file_out,
1537 "in:%d out:%d type:%c dur:%"PRId64" cpbdur:%"PRId64" q:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c ref:",
1538 h->fenc->i_frame, h->i_frame,
1539 c_type, h->fenc->i_duration,
1540 h->fenc->i_cpb_duration, rc->qpa_rc,
1541 h->stat.frame.i_tex_bits,
1542 h->stat.frame.i_mv_bits,
1543 h->stat.frame.i_misc_bits,
1544 h->stat.frame.i_mb_count_i,
1545 h->stat.frame.i_mb_count_p,
1546 h->stat.frame.i_mb_count_skip,
1550 /* Only write information for reference reordering once. */
1551 int use_old_stats = h->param.rc.b_stat_read && rc->rce->refs > 1;
1552 for( int i = 0; i < (use_old_stats ? rc->rce->refs : h->i_ref[0]); i++ )
1554 int refcount = use_old_stats ? rc->rce->refcount[i]
1555 : h->param.b_interlaced ? h->stat.frame.i_mb_count_ref[0][i*2]
1556 + h->stat.frame.i_mb_count_ref[0][i*2+1]
1557 : h->stat.frame.i_mb_count_ref[0][i];
1558 if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1562 if( h->param.analyse.i_weighted_pred >= X264_WEIGHTP_SIMPLE && h->sh.weight[0][0].weightfn )
1564 if( fprintf( rc->p_stat_file_out, "w:%d,%d,%d",
1565 h->sh.weight[0][0].i_denom, h->sh.weight[0][0].i_scale, h->sh.weight[0][0].i_offset ) < 0 )
1567 if( h->sh.weight[0][1].weightfn || h->sh.weight[0][2].weightfn )
1569 if( fprintf( rc->p_stat_file_out, ",%d,%d,%d,%d,%d ",
1570 h->sh.weight[0][1].i_denom, h->sh.weight[0][1].i_scale, h->sh.weight[0][1].i_offset,
1571 h->sh.weight[0][2].i_scale, h->sh.weight[0][2].i_offset ) < 0 )
1574 else if( fprintf( rc->p_stat_file_out, " " ) < 0 )
1578 if( fprintf( rc->p_stat_file_out, ";\n") < 0 )
1581 /* Don't re-write the data in multi-pass mode. */
1582 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1584 uint8_t i_type = h->sh.i_type;
1585 /* Values are stored as big-endian FIX8.8 */
1586 for( int i = 0; i < h->mb.i_mb_count; i++ )
1587 rc->qp_buffer[0][i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1588 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1590 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 )
1597 if( h->sh.i_type != SLICE_TYPE_B )
1598 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / rc->last_rceq;
1601 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1602 * Not perfectly accurate with B-refs, but good enough. */
1603 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / (rc->last_rceq * fabs( h->param.rc.f_pb_factor ));
1605 rc->cplxr_sum *= rc->cbr_decay;
1606 rc->wanted_bits_window += h->fenc->f_duration * rc->bitrate;
1607 rc->wanted_bits_window *= rc->cbr_decay;
1611 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale( rc->rce->new_qp ) );
1613 if( h->mb.b_variable_qp )
1615 if( h->sh.i_type == SLICE_TYPE_B )
1617 rc->bframe_bits += bits;
1618 if( h->fenc->b_last_minigop_bframe )
1620 update_predictor( rc->pred_b_from_p, qp2qscale( rc->qpa_rc ),
1621 h->fref[1][h->i_ref[1]-1]->i_satd, rc->bframe_bits / rc->bframes );
1622 rc->bframe_bits = 0;
1627 *filler = update_vbv( h, bits );
1628 rc->filler_bits_sum += *filler * 8;
1630 if( h->sps->vui.b_nal_hrd_parameters_present )
1632 if( h->fenc->i_frame == 0 )
1634 // access unit initialises the HRD
1635 h->fenc->hrd_timing.cpb_initial_arrival_time = 0;
1636 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1637 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1638 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit = (double)rc->initial_cpb_removal_delay / 90000;
1642 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit + (double)(h->fenc->i_cpb_delay - h->i_cpb_delay_pir_offset) *
1643 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1645 double cpb_earliest_arrival_time = h->fenc->hrd_timing.cpb_removal_time - (double)rc->initial_cpb_removal_delay / 90000;
1646 if( h->fenc->b_keyframe )
1648 rc->nrt_first_access_unit = h->fenc->hrd_timing.cpb_removal_time;
1649 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1650 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1653 cpb_earliest_arrival_time -= (double)rc->initial_cpb_removal_delay_offset / 90000;
1655 if( h->sps->vui.hrd.b_cbr_hrd )
1656 h->fenc->hrd_timing.cpb_initial_arrival_time = rc->previous_cpb_final_arrival_time;
1658 h->fenc->hrd_timing.cpb_initial_arrival_time = X264_MAX( rc->previous_cpb_final_arrival_time, cpb_earliest_arrival_time );
1660 int filler_bits = *filler ? X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), *filler )*8 : 0;
1662 h->fenc->hrd_timing.cpb_final_arrival_time = rc->previous_cpb_final_arrival_time = h->fenc->hrd_timing.cpb_initial_arrival_time +
1663 (double)(bits + filler_bits) / h->sps->vui.hrd.i_bit_rate_unscaled;
1665 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 +
1666 h->fenc->hrd_timing.cpb_removal_time;
1671 x264_log(h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n");
1675 /****************************************************************************
1677 ***************************************************************************/
1680 * modify the bitrate curve from pass1 for one frame
1682 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1684 x264_ratecontrol_t *rcc= h->rc;
1685 x264_zone_t *zone = get_zone( h, frame_num );
1686 double q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1688 // avoid NaN's in the rc_eq
1689 if( !isfinite(q) || rce->tex_bits + rce->mv_bits == 0 )
1690 q = rcc->last_qscale_for[rce->pict_type];
1695 rcc->last_qscale = q;
1700 if( zone->b_force_qp )
1701 q = qp2qscale( zone->i_qp );
1703 q /= zone->f_bitrate_factor;
1709 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1711 x264_ratecontrol_t *rcc = h->rc;
1712 const int pict_type = rce->pict_type;
1714 // force I/B quants as a function of P quants
1715 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1716 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1717 if( pict_type == SLICE_TYPE_I )
1720 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1721 double ip_factor = fabs( h->param.rc.f_ip_factor );
1722 /* don't apply ip_factor if the following frame is also I */
1723 if( rcc->accum_p_norm <= 0 )
1725 else if( h->param.rc.f_ip_factor < 0 )
1727 else if( rcc->accum_p_norm >= 1 )
1730 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1732 else if( pict_type == SLICE_TYPE_B )
1734 if( h->param.rc.f_pb_factor > 0 )
1736 if( !rce->kept_as_ref )
1737 q *= fabs( h->param.rc.f_pb_factor );
1739 else if( pict_type == SLICE_TYPE_P
1740 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1741 && rce->tex_bits == 0 )
1746 /* last qscale / qdiff stuff */
1747 if( rcc->last_non_b_pict_type == pict_type &&
1748 (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1) )
1750 double last_q = rcc->last_qscale_for[pict_type];
1751 double max_qscale = last_q * rcc->lstep;
1752 double min_qscale = last_q / rcc->lstep;
1754 if ( q > max_qscale ) q = max_qscale;
1755 else if( q < min_qscale ) q = min_qscale;
1758 rcc->last_qscale_for[pict_type] = q;
1759 if( pict_type != SLICE_TYPE_B )
1760 rcc->last_non_b_pict_type = pict_type;
1761 if( pict_type == SLICE_TYPE_I )
1763 rcc->last_accum_p_norm = rcc->accum_p_norm;
1764 rcc->accum_p_norm = 0;
1765 rcc->accum_p_qp = 0;
1767 if( pict_type == SLICE_TYPE_P )
1769 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1770 rcc->accum_p_qp = mask * (qscale2qp( q ) + rcc->accum_p_qp);
1771 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1776 static double predict_size( predictor_t *p, double q, double var )
1778 return (p->coeff*var + p->offset) / (q*p->count);
1781 static void update_predictor( predictor_t *p, double q, double var, double bits )
1783 const double range = 1.5;
1786 double old_coeff = p->coeff / p->count;
1787 double new_coeff = bits*q / var;
1788 double new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
1789 double new_offset = bits*q - new_coeff_clipped * var;
1790 if( new_offset >= 0 )
1791 new_coeff = new_coeff_clipped;
1794 p->count *= p->decay;
1795 p->coeff *= p->decay;
1796 p->offset *= p->decay;
1798 p->coeff += new_coeff;
1799 p->offset += new_offset;
1802 // update VBV after encoding a frame
1803 static int update_vbv( x264_t *h, int bits )
1806 int bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
1807 x264_ratecontrol_t *rcc = h->rc;
1808 x264_ratecontrol_t *rct = h->thread[0]->rc;
1809 uint64_t buffer_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
1811 if( rcc->last_satd >= h->mb.i_mb_count )
1812 update_predictor( &rct->pred[h->sh.i_type], qp2qscale( rcc->qpa_rc ), rcc->last_satd, bits );
1817 rct->buffer_fill_final -= (uint64_t)bits * h->sps->vui.i_time_scale;
1819 if( rct->buffer_fill_final < 0 )
1820 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 );
1821 rct->buffer_fill_final = X264_MAX( rct->buffer_fill_final, 0 );
1822 rct->buffer_fill_final += (uint64_t)bitrate * h->sps->vui.i_num_units_in_tick * h->fenc->i_cpb_duration;
1824 if( h->sps->vui.hrd.b_cbr_hrd && rct->buffer_fill_final > buffer_size )
1826 filler = ceil( (rct->buffer_fill_final - buffer_size) / (8. * h->sps->vui.i_time_scale) );
1827 bits = X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), filler ) * 8;
1828 rct->buffer_fill_final -= (uint64_t)bits * h->sps->vui.i_time_scale;
1831 rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, buffer_size );
1836 void x264_hrd_fullness( x264_t *h )
1838 x264_ratecontrol_t *rct = h->thread[0]->rc;
1839 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale / rct->hrd_multiply_denom;
1840 uint64_t cpb_state = rct->buffer_fill_final;
1841 uint64_t cpb_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
1842 uint64_t multiply_factor = 180000 / rct->hrd_multiply_denom;
1844 if( rct->buffer_fill_final < 0 || rct->buffer_fill_final > cpb_size )
1846 x264_log( h, X264_LOG_WARNING, "CPB %s: %.0lf bits in a %.0lf-bit buffer\n",
1847 rct->buffer_fill_final < 0 ? "underflow" : "overflow", (float)rct->buffer_fill_final/denom, (float)cpb_size/denom );
1850 h->initial_cpb_removal_delay = (multiply_factor * cpb_state + denom) / (2*denom);
1851 h->initial_cpb_removal_delay_offset = (multiply_factor * cpb_size + denom) / (2*denom) - h->initial_cpb_removal_delay;
1854 // provisionally update VBV according to the planned size of all frames currently in progress
1855 static void update_vbv_plan( x264_t *h, int overhead )
1857 x264_ratecontrol_t *rcc = h->rc;
1858 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final / h->sps->vui.i_time_scale;
1859 if( h->i_thread_frames > 1 )
1861 int j = h->rc - h->thread[0]->rc;
1862 for( int i = 1; i < h->i_thread_frames; i++ )
1864 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
1865 double bits = t->rc->frame_size_planned;
1866 if( !t->b_thread_active )
1868 bits = X264_MAX(bits, t->rc->frame_size_estimated);
1869 rcc->buffer_fill -= bits;
1870 rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
1871 rcc->buffer_fill += t->rc->buffer_rate;
1872 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1875 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1876 rcc->buffer_fill -= overhead;
1879 // apply VBV constraints and clip qscale to between lmin and lmax
1880 static double clip_qscale( x264_t *h, int pict_type, double q )
1882 x264_ratecontrol_t *rcc = h->rc;
1883 double lmin = rcc->lmin[pict_type];
1884 double lmax = rcc->lmax[pict_type];
1885 if( rcc->rate_factor_max_increment )
1886 lmax = X264_MIN( lmax, qp2qscale( rcc->qp_novbv + rcc->rate_factor_max_increment ) );
1889 /* B-frames are not directly subject to VBV,
1890 * since they are controlled by the P-frames' QPs. */
1892 if( rcc->b_vbv && rcc->last_satd > 0 )
1894 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
1895 * the lookahead overflow and such that the buffer is in a reasonable state
1896 * by the end of the lookahead. */
1897 if( h->param.rc.i_lookahead )
1901 /* Avoid an infinite loop. */
1902 for( int iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
1905 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1906 double buffer_fill_cur = rcc->buffer_fill - cur_bits;
1908 double total_duration = 0;
1909 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
1910 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
1911 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
1913 /* Loop over the planned future frames. */
1914 for( int j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
1916 total_duration += h->fenc->f_planned_cpb_duration[j];
1917 buffer_fill_cur += rcc->vbv_max_rate * h->fenc->f_planned_cpb_duration[j];
1918 int i_type = h->fenc->i_planned_type[j];
1919 int i_satd = h->fenc->i_planned_satd[j];
1920 if( i_type == X264_TYPE_AUTO )
1922 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
1923 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
1924 buffer_fill_cur -= cur_bits;
1926 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
1927 target_fill = X264_MIN( rcc->buffer_fill + total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.5 );
1928 if( buffer_fill_cur < target_fill )
1934 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
1935 target_fill = x264_clip3f( rcc->buffer_fill - total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
1936 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
1945 /* Fallback to old purely-reactive algorithm: no lookahead. */
1948 if( ( pict_type == SLICE_TYPE_P ||
1949 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
1950 rcc->buffer_fill/rcc->buffer_size < 0.5 )
1952 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1955 /* Now a hard threshold to make sure the frame fits in VBV.
1956 * This one is mostly for I-frames. */
1957 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1959 /* For small VBVs, allow the frame to use up the entire VBV. */
1960 double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
1961 /* For single-frame VBVs, request that the frame use up the entire VBV. */
1962 double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
1964 if( bits > rcc->buffer_fill/max_fill_factor )
1965 qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
1968 if( bits < rcc->buffer_rate/min_fill_factor )
1969 q *= bits*min_fill_factor/rcc->buffer_rate;
1970 q = X264_MAX( q0, q );
1973 /* Apply MinCR restrictions */
1974 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1975 if( bits > rcc->frame_size_maximum )
1976 q *= bits / rcc->frame_size_maximum;
1977 bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1979 /* Check B-frame complexity, and use up any bits that would
1980 * overflow before the next P-frame. */
1981 if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
1983 int nb = rcc->bframes;
1984 double pbbits = bits;
1985 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1987 double bframe_cpb_duration = 0;
1988 double minigop_cpb_duration;
1989 for( int i = 0; i < nb; i++ )
1990 bframe_cpb_duration += h->fenc->f_planned_cpb_duration[1+i];
1992 if( bbits * nb > bframe_cpb_duration * rcc->vbv_max_rate )
1994 pbbits += nb * bbits;
1996 minigop_cpb_duration = bframe_cpb_duration + h->fenc->f_planned_cpb_duration[0];
1997 space = rcc->buffer_fill + minigop_cpb_duration*rcc->vbv_max_rate - rcc->buffer_size;
1998 if( pbbits < space )
2000 q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
2002 q = X264_MAX( q0/2, q );
2005 if( !rcc->b_vbv_min_rate )
2006 q = X264_MAX( q0, q );
2011 else if( rcc->b_2pass )
2013 double min2 = log( lmin );
2014 double max2 = log( lmax );
2015 q = (log(q) - min2)/(max2-min2) - 0.5;
2016 q = 1.0/(1.0 + exp( -4*q ));
2017 q = q*(max2-min2) + min2;
2021 return x264_clip3f( q, lmin, lmax );
2024 // update qscale for 1 frame based on actual bits used so far
2025 static float rate_estimate_qscale( x264_t *h )
2028 x264_ratecontrol_t *rcc = h->rc;
2029 ratecontrol_entry_t UNINIT(rce);
2030 int pict_type = h->sh.i_type;
2031 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
2032 + h->stat.i_frame_size[SLICE_TYPE_P]
2033 + h->stat.i_frame_size[SLICE_TYPE_B])
2034 - rcc->filler_bits_sum;
2039 if( pict_type != rce.pict_type )
2041 x264_log( h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
2042 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type] );
2046 if( pict_type == SLICE_TYPE_B )
2048 /* B-frames don't have independent ratecontrol, but rather get the
2049 * average QP of the two adjacent P-frames + an offset */
2051 int i0 = IS_X264_TYPE_I(h->fref_nearest[0]->i_type);
2052 int i1 = IS_X264_TYPE_I(h->fref_nearest[1]->i_type);
2053 int dt0 = abs(h->fenc->i_poc - h->fref_nearest[0]->i_poc);
2054 int dt1 = abs(h->fenc->i_poc - h->fref_nearest[1]->i_poc);
2055 float q0 = h->fref_nearest[0]->f_qp_avg_rc;
2056 float q1 = h->fref_nearest[1]->f_qp_avg_rc;
2058 if( h->fref_nearest[0]->i_type == X264_TYPE_BREF )
2059 q0 -= rcc->pb_offset/2;
2060 if( h->fref_nearest[1]->i_type == X264_TYPE_BREF )
2061 q1 -= rcc->pb_offset/2;
2064 q = (q0 + q1) / 2 + rcc->ip_offset;
2070 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
2072 if( h->fenc->b_kept_as_ref )
2073 q += rcc->pb_offset/2;
2075 q += rcc->pb_offset;
2077 if( rcc->b_2pass && rcc->b_vbv )
2078 rcc->frame_size_planned = qscale2bits( &rce, qp2qscale( q ) );
2080 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, qp2qscale( q ), h->fref[1][h->i_ref[1]-1]->i_satd );
2081 h->rc->frame_size_estimated = rcc->frame_size_planned;
2085 rcc->last_satd = x264_rc_analyse_slice( h );
2087 return qp2qscale( q );
2091 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
2095 double lmin = rcc->lmin[pict_type];
2096 double lmax = rcc->lmax[pict_type];
2098 int64_t predicted_bits = total_bits;
2102 if( h->i_thread_frames > 1 )
2104 int j = h->rc - h->thread[0]->rc;
2105 for( int i = 1; i < h->i_thread_frames; i++ )
2107 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
2108 double bits = t->rc->frame_size_planned;
2109 if( !t->b_thread_active )
2111 bits = X264_MAX(bits, t->rc->frame_size_estimated);
2112 predicted_bits += (int64_t)bits;
2118 if( h->i_frame < h->i_thread_frames )
2119 predicted_bits += (int64_t)h->i_frame * rcc->bitrate / rcc->fps;
2121 predicted_bits += (int64_t)(h->i_thread_frames - 1) * rcc->bitrate / rcc->fps;
2124 /* Adjust ABR buffer based on distance to the end of the video. */
2125 if( rcc->num_entries > h->i_frame )
2127 double final_bits = rcc->entry[rcc->num_entries-1].expected_bits;
2128 double video_pos = rce.expected_bits / final_bits;
2129 double scale_factor = sqrt( (1 - video_pos) * rcc->num_entries );
2130 abr_buffer *= 0.5 * X264_MAX( scale_factor, 0.5 );
2133 diff = predicted_bits - (int64_t)rce.expected_bits;
2135 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
2136 if( ((h->i_frame + 1 - h->i_thread_frames) >= rcc->fps) &&
2137 (rcc->expected_bits_sum > 0))
2139 /* Adjust quant based on the difference between
2140 * achieved and expected bitrate so far */
2141 double cur_time = (double)h->i_frame / rcc->num_entries;
2142 double w = x264_clip3f( cur_time*100, 0.0, 1.0 );
2143 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
2147 /* Do not overflow vbv */
2148 double expected_size = qscale2bits( &rce, q );
2149 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2150 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
2151 double qmax = q*(2 - expected_fullness);
2152 double size_constraint = 1 + expected_fullness;
2153 qmax = X264_MAX( qmax, rce.new_qscale );
2154 if( expected_fullness < .05 )
2156 qmax = X264_MIN(qmax, lmax);
2157 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
2158 ((expected_vbv < 0) && (q < lmax)))
2161 expected_size = qscale2bits(&rce, q);
2162 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2164 rcc->last_satd = x264_rc_analyse_slice( h );
2166 q = x264_clip3f( q, lmin, lmax );
2168 else /* 1pass ABR */
2170 /* Calculate the quantizer which would have produced the desired
2171 * average bitrate if it had been applied to all frames so far.
2172 * Then modulate that quant based on the current frame's complexity
2173 * relative to the average complexity so far (using the 2pass RCEQ).
2174 * Then bias the quant up or down if total size so far was far from
2176 * Result: Depending on the value of rate_tolerance, there is a
2177 * tradeoff between quality and bitrate precision. But at large
2178 * tolerances, the bit distribution approaches that of 2pass. */
2180 double wanted_bits, overflow = 1;
2182 rcc->last_satd = x264_rc_analyse_slice( h );
2183 rcc->short_term_cplxsum *= 0.5;
2184 rcc->short_term_cplxcount *= 0.5;
2185 rcc->short_term_cplxsum += rcc->last_satd / (CLIP_DURATION(h->fenc->f_duration) / BASE_FRAME_DURATION);
2186 rcc->short_term_cplxcount ++;
2188 rce.tex_bits = rcc->last_satd;
2189 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
2191 rce.p_count = rcc->nmb;
2195 rce.pict_type = pict_type;
2197 if( h->param.rc.i_rc_method == X264_RC_CRF )
2199 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
2203 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
2205 /* ABR code can potentially be counterproductive in CBR, so just don't bother.
2206 * Don't run it if the frame complexity is zero either. */
2207 if( !rcc->b_vbv_min_rate && rcc->last_satd )
2209 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
2210 int i_frame_done = h->i_frame + 1 - h->i_thread_frames;
2211 double time_done = i_frame_done / rcc->fps;
2212 if( h->param.b_vfr_input && i_frame_done > 0 )
2213 time_done = ((double)(h->fenc->i_reordered_pts - h->i_reordered_pts_delay)) * h->param.i_timebase_num / h->param.i_timebase_den;
2214 wanted_bits = time_done * rcc->bitrate;
2215 if( wanted_bits > 0 )
2217 abr_buffer *= X264_MAX( 1, sqrt( time_done ) );
2218 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
2224 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
2225 /* should test _next_ pict type, but that isn't decided yet */
2226 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
2228 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2229 q /= fabs( h->param.rc.f_ip_factor );
2231 else if( h->i_frame > 0 )
2233 if( h->param.rc.i_rc_method != X264_RC_CRF )
2235 /* Asymmetric clipping, because symmetric would prevent
2236 * overflow control in areas of rapidly oscillating complexity */
2237 double lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
2238 double lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
2239 if( overflow > 1.1 && h->i_frame > 3 )
2241 else if( overflow < 0.9 )
2244 q = x264_clip3f(q, lmin, lmax);
2247 else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
2249 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
2251 rcc->qp_novbv = qscale2qp( q );
2253 //FIXME use get_diff_limited_q() ?
2254 q = clip_qscale( h, pict_type, q );
2257 rcc->last_qscale_for[pict_type] =
2258 rcc->last_qscale = q;
2260 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
2261 rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );
2263 if( rcc->b_2pass && rcc->b_vbv )
2264 rcc->frame_size_planned = qscale2bits(&rce, q);
2266 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2268 /* Always use up the whole VBV in this case. */
2269 if( rcc->single_frame_vbv )
2270 rcc->frame_size_planned = rcc->buffer_rate;
2271 h->rc->frame_size_estimated = rcc->frame_size_planned;
2276 void x264_threads_normalize_predictors( x264_t *h )
2278 double totalsize = 0;
2279 for( int i = 0; i < h->param.i_threads; i++ )
2280 totalsize += h->thread[i]->rc->slice_size_planned;
2281 double factor = h->rc->frame_size_planned / totalsize;
2282 for( int i = 0; i < h->param.i_threads; i++ )
2283 h->thread[i]->rc->slice_size_planned *= factor;
2286 void x264_threads_distribute_ratecontrol( x264_t *h )
2289 x264_ratecontrol_t *rc = h->rc;
2291 /* Initialize row predictors */
2292 if( h->i_frame == 0 )
2293 for( int i = 0; i < h->param.i_threads; i++ )
2295 x264_ratecontrol_t *t = h->thread[i]->rc;
2296 memcpy( t->row_preds, rc->row_preds, sizeof(rc->row_preds) );
2299 for( int i = 0; i < h->param.i_threads; i++ )
2301 x264_t *t = h->thread[i];
2302 memcpy( t->rc, rc, offsetof(x264_ratecontrol_t, row_pred) );
2303 t->rc->row_pred = &t->rc->row_preds[h->sh.i_type];
2304 /* Calculate the planned slice size. */
2305 if( rc->b_vbv && rc->frame_size_planned )
2308 for( row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2309 size += h->fdec->i_row_satd[row];
2310 t->rc->slice_size_planned = predict_size( &rc->pred[h->sh.i_type + (i+1)*5], rc->qpm, size );
2313 t->rc->slice_size_planned = 0;
2315 if( rc->b_vbv && rc->frame_size_planned )
2317 x264_threads_normalize_predictors( h );
2319 if( rc->single_frame_vbv )
2321 /* Compensate for our max frame error threshold: give more bits (proportionally) to smaller slices. */
2322 for( int i = 0; i < h->param.i_threads; i++ )
2324 x264_t *t = h->thread[i];
2325 float max_frame_error = X264_MAX( 0.05, 1.0 / (t->i_threadslice_end - t->i_threadslice_start) );
2326 t->rc->slice_size_planned += 2 * max_frame_error * rc->frame_size_planned;
2328 x264_threads_normalize_predictors( h );
2331 for( int i = 0; i < h->param.i_threads; i++ )
2332 h->thread[i]->rc->frame_size_estimated = h->thread[i]->rc->slice_size_planned;
2336 void x264_threads_merge_ratecontrol( x264_t *h )
2338 x264_ratecontrol_t *rc = h->rc;
2341 for( int i = 0; i < h->param.i_threads; i++ )
2343 x264_t *t = h->thread[i];
2344 x264_ratecontrol_t *rct = h->thread[i]->rc;
2345 if( h->param.rc.i_vbv_buffer_size )
2348 for( int row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2349 size += h->fdec->i_row_satd[row];
2350 int bits = t->stat.frame.i_mv_bits + t->stat.frame.i_tex_bits + t->stat.frame.i_misc_bits;
2351 int mb_count = (t->i_threadslice_end - t->i_threadslice_start) * h->mb.i_mb_width;
2352 update_predictor( &rc->pred[h->sh.i_type+(i+1)*5], qp2qscale( rct->qpa_rc/mb_count ), size, bits );
2356 rc->qpa_rc += rct->qpa_rc;
2357 rc->qpa_aq += rct->qpa_aq;
2361 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
2365 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
2366 /* these vars are updated in x264_ratecontrol_start()
2367 * so copy them from the context that most recently started (prev)
2368 * to the context that's about to start (cur). */
2373 COPY(last_qscale_for);
2374 COPY(last_non_b_pict_type);
2375 COPY(short_term_cplxsum);
2376 COPY(short_term_cplxcount);
2380 /* these vars can be updated by x264_ratecontrol_init_reconfigurable */
2383 COPY(single_frame_vbv);
2385 COPY(b_vbv_min_rate);
2386 COPY(rate_factor_constant);
2392 #define COPY(var) next->rc->var = cur->rc->var
2393 /* these vars are updated in x264_ratecontrol_end()
2394 * so copy them from the context that most recently ended (cur)
2395 * to the context that's about to end (next) */
2397 COPY(expected_bits_sum);
2398 COPY(filler_bits_sum);
2399 COPY(wanted_bits_window);
2401 COPY(initial_cpb_removal_delay);
2402 COPY(initial_cpb_removal_delay_offset);
2403 COPY(nrt_first_access_unit);
2404 COPY(previous_cpb_final_arrival_time);
2407 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
2408 /* the rest of the variables are either constant or thread-local */
2411 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
2413 /* find an interval ending on an overflow or underflow (depending on whether
2414 * we're adding or removing bits), and starting on the earliest frame that
2415 * can influence the buffer fill of that end frame. */
2416 x264_ratecontrol_t *rcc = h->rc;
2417 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
2418 const double buffer_max = .9 * rcc->buffer_size;
2419 double fill = fills[*t0-1];
2420 double parity = over ? 1. : -1.;
2421 int start = -1, end = -1;
2422 for( int i = *t0; i < rcc->num_entries; i++ )
2424 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 -
2425 qscale2bits( &rcc->entry[i], rcc->entry[i].new_qscale )) * parity;
2426 fill = x264_clip3f(fill, 0, rcc->buffer_size);
2428 if( fill <= buffer_min || i == 0 )
2434 else if( fill >= buffer_max && start >= 0 )
2439 return start >= 0 && end >= 0;
2442 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
2444 x264_ratecontrol_t *rcc = h->rc;
2445 double qscale_orig, qscale_new;
2449 for( int i = t0; i <= t1; i++ )
2451 qscale_orig = rcc->entry[i].new_qscale;
2452 qscale_orig = x264_clip3f( qscale_orig, qscale_min, qscale_max );
2453 qscale_new = qscale_orig * adjustment;
2454 qscale_new = x264_clip3f( qscale_new, qscale_min, qscale_max );
2455 rcc->entry[i].new_qscale = qscale_new;
2456 adjusted = adjusted || (qscale_new != qscale_orig);
2461 static double count_expected_bits( x264_t *h )
2463 x264_ratecontrol_t *rcc = h->rc;
2464 double expected_bits = 0;
2465 for( int i = 0; i < rcc->num_entries; i++ )
2467 ratecontrol_entry_t *rce = &rcc->entry[i];
2468 rce->expected_bits = expected_bits;
2469 expected_bits += qscale2bits( rce, rce->new_qscale );
2471 return expected_bits;
2474 static int vbv_pass2( x264_t *h, double all_available_bits )
2476 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
2477 * frames in the interval until either buffer is full at some intermediate frame or the
2478 * last frame in the interval no longer underflows. Recompute intervals and repeat.
2479 * Then do the converse to put bits back into overflow areas until target size is met */
2481 x264_ratecontrol_t *rcc = h->rc;
2483 double expected_bits = 0;
2485 double prev_bits = 0;
2487 double qscale_min = qp2qscale( h->param.rc.i_qp_min );
2488 double qscale_max = qp2qscale( h->param.rc.i_qp_max );
2490 int adj_min, adj_max;
2491 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2495 /* adjust overall stream size */
2499 prev_bits = expected_bits;
2502 { /* not first iteration */
2503 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2504 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2508 while(adj_min && find_underflow( h, fills, &t0, &t1, 1 ))
2510 adj_min = fix_underflow( h, t0, t1, adjustment, qscale_min, qscale_max );
2515 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2517 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2519 while( adj_max && find_underflow( h, fills, &t0, &t1, 0 ) )
2520 adj_max = fix_underflow( h, t0, t1, 1.001, qscale_min, qscale_max );
2522 expected_bits = count_expected_bits( h );
2523 } while( (expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2526 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2528 /* store expected vbv filling values for tracking when encoding */
2529 for( int i = 0; i < rcc->num_entries; i++ )
2530 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
2532 x264_free( fills-1 );
2538 static int init_pass2( x264_t *h )
2540 x264_ratecontrol_t *rcc = h->rc;
2541 uint64_t all_const_bits = 0;
2542 double timescale = (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2543 double duration = 0;
2544 for( int i = 0; i < rcc->num_entries; i++ )
2545 duration += rcc->entry[i].i_duration;
2546 duration *= timescale;
2547 uint64_t all_available_bits = h->param.rc.i_bitrate * 1000. * duration;
2548 double rate_factor, step_mult;
2549 double qblur = h->param.rc.f_qblur;
2550 double cplxblur = h->param.rc.f_complexity_blur;
2551 const int filter_size = (int)(qblur*4) | 1;
2552 double expected_bits;
2553 double *qscale, *blurred_qscale;
2554 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
2556 /* find total/average complexity & const_bits */
2557 for( int i = 0; i < rcc->num_entries; i++ )
2559 ratecontrol_entry_t *rce = &rcc->entry[i];
2560 all_const_bits += rce->misc_bits;
2563 if( all_available_bits < all_const_bits)
2565 x264_log( h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2566 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)) );
2570 /* Blur complexities, to reduce local fluctuation of QP.
2571 * We don't blur the QPs directly, because then one very simple frame
2572 * could drag down the QP of a nearby complex frame and give it more
2573 * bits than intended. */
2574 for( int i = 0; i < rcc->num_entries; i++ )
2576 ratecontrol_entry_t *rce = &rcc->entry[i];
2577 double weight_sum = 0;
2578 double cplx_sum = 0;
2579 double weight = 1.0;
2580 double gaussian_weight;
2581 /* weighted average of cplx of future frames */
2582 for( int j = 1; j < cplxblur*2 && j < rcc->num_entries-i; j++ )
2584 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2585 double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2586 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2587 if( weight < .0001 )
2589 gaussian_weight = weight * exp( -j*j/200.0 );
2590 weight_sum += gaussian_weight;
2591 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2593 /* weighted average of cplx of past frames */
2595 for( int j = 0; j <= cplxblur*2 && j <= i; j++ )
2597 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2598 double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2599 gaussian_weight = weight * exp( -j*j/200.0 );
2600 weight_sum += gaussian_weight;
2601 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2602 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2603 if( weight < .0001 )
2606 rce->blurred_complexity = cplx_sum / weight_sum;
2609 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2610 if( filter_size > 1 )
2611 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2613 blurred_qscale = qscale;
2615 /* Search for a factor which, when multiplied by the RCEQ values from
2616 * each frame, adds up to the desired total size.
2617 * There is no exact closed-form solution because of VBV constraints and
2618 * because qscale2bits is not invertible, but we can start with the simple
2619 * approximation of scaling the 1st pass by the ratio of bitrates.
2620 * The search range is probably overkill, but speed doesn't matter here. */
2623 for( int i = 0; i < rcc->num_entries; i++ )
2625 double q = get_qscale(h, &rcc->entry[i], 1.0, i);
2626 expected_bits += qscale2bits(&rcc->entry[i], q);
2627 rcc->last_qscale_for[rcc->entry[i].pict_type] = q;
2629 step_mult = all_available_bits / expected_bits;
2632 for( double step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2635 rate_factor += step;
2637 rcc->last_non_b_pict_type = -1;
2638 rcc->last_accum_p_norm = 1;
2639 rcc->accum_p_norm = 0;
2641 rcc->last_qscale_for[0] =
2642 rcc->last_qscale_for[1] =
2643 rcc->last_qscale_for[2] = pow( base_cplx, 1 - rcc->qcompress ) / rate_factor;
2646 for( int i = 0; i < rcc->num_entries; i++ )
2648 qscale[i] = get_qscale( h, &rcc->entry[i], rate_factor, i );
2649 rcc->last_qscale_for[rcc->entry[i].pict_type] = qscale[i];
2652 /* fixed I/B qscale relative to P */
2653 for( int i = rcc->num_entries-1; i >= 0; i-- )
2655 qscale[i] = get_diff_limited_q( h, &rcc->entry[i], qscale[i] );
2656 assert(qscale[i] >= 0);
2660 if( filter_size > 1 )
2662 assert( filter_size%2 == 1 );
2663 for( int i = 0; i < rcc->num_entries; i++ )
2665 ratecontrol_entry_t *rce = &rcc->entry[i];
2666 double q = 0.0, sum = 0.0;
2668 for( int j = 0; j < filter_size; j++ )
2670 int idx = i+j-filter_size/2;
2672 double coeff = qblur==0 ? 1.0 : exp( -d*d/(qblur*qblur) );
2673 if( idx < 0 || idx >= rcc->num_entries )
2675 if( rce->pict_type != rcc->entry[idx].pict_type )
2677 q += qscale[idx] * coeff;
2680 blurred_qscale[i] = q/sum;
2684 /* find expected bits */
2685 for( int i = 0; i < rcc->num_entries; i++ )
2687 ratecontrol_entry_t *rce = &rcc->entry[i];
2688 rce->new_qscale = clip_qscale( h, rce->pict_type, blurred_qscale[i] );
2689 assert(rce->new_qscale >= 0);
2690 expected_bits += qscale2bits( rce, rce->new_qscale );
2693 if( expected_bits > all_available_bits )
2694 rate_factor -= step;
2697 x264_free( qscale );
2698 if( filter_size > 1 )
2699 x264_free( blurred_qscale );
2702 if( vbv_pass2( h, all_available_bits ) )
2704 expected_bits = count_expected_bits( h );
2706 if( fabs( expected_bits/all_available_bits - 1.0 ) > 0.01 )
2709 for( int i = 0; i < rcc->num_entries; i++ )
2710 avgq += rcc->entry[i].new_qscale;
2711 avgq = qscale2qp( avgq / rcc->num_entries );
2713 if( expected_bits > all_available_bits || !rcc->b_vbv )
2714 x264_log( h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n" );
2715 x264_log( h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
2716 (float)h->param.rc.i_bitrate,
2717 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
2719 if( expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2 )
2721 if( h->param.rc.i_qp_min > 0 )
2722 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min );
2724 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate\n" );
2726 else if( expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2 )
2728 if( h->param.rc.i_qp_max < QP_MAX )
2729 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max );
2731 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate\n");
2733 else if( !(rcc->b_2pass && rcc->b_vbv) )
2734 x264_log( h, X264_LOG_WARNING, "internal error\n" );