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 double max_frame_error;
145 predictor_t (*row_pred)[2];
146 predictor_t row_preds[3][2];
147 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
148 int bframes; /* # consecutive B-frames before this P-frame */
149 int bframe_bits; /* total cost of those frames */
153 x264_zone_t *prev_zone;
156 int initial_cpb_removal_delay;
157 int initial_cpb_removal_delay_offset;
158 double nrt_first_access_unit; /* nominal removal time */
159 double previous_cpb_final_arrival_time;
160 uint64_t hrd_multiply_denom;
164 static int parse_zones( x264_t *h );
165 static int init_pass2(x264_t *);
166 static float rate_estimate_qscale( x264_t *h );
167 static int update_vbv( x264_t *h, int bits );
168 static void update_vbv_plan( x264_t *h, int overhead );
169 static double predict_size( predictor_t *p, double q, double var );
170 static void update_predictor( predictor_t *p, double q, double var, double bits );
172 #define CMP_OPT_FIRST_PASS( opt, param_val )\
174 if( ( p = strstr( opts, opt "=" ) ) && sscanf( p, opt "=%d" , &i ) && param_val != i )\
176 x264_log( h, X264_LOG_ERROR, "different " opt " setting than first pass (%d vs %d)\n", param_val, i );\
182 * qp = h.264's quantizer
183 * qscale = linearized quantizer = Lagrange multiplier
185 static inline double qp2qscale( double qp )
187 return 0.85 * pow( 2.0, ( qp - 12.0 ) / 6.0 );
189 static inline double qscale2qp( double qscale )
191 return 12.0 + 6.0 * log2( qscale/0.85 );
194 /* Texture bitrate is not quite inversely proportional to qscale,
195 * probably due the the changing number of SKIP blocks.
196 * MV bits level off at about qp<=12, because the lambda used
197 * for motion estimation is constant there. */
198 static inline double qscale2bits( ratecontrol_entry_t *rce, double qscale )
202 return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
203 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
207 static ALWAYS_INLINE uint32_t ac_energy_var( uint64_t sum_ssd, int shift, x264_frame_t *frame, int i )
209 uint32_t sum = sum_ssd;
210 uint32_t ssd = sum_ssd >> 32;
211 frame->i_pixel_sum[i] += sum;
212 frame->i_pixel_ssd[i] += ssd;
213 return ssd - ((uint64_t)sum * sum >> shift);
216 static ALWAYS_INLINE uint32_t ac_energy_plane( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame, int i )
219 int stride = frame->i_stride[i];
220 int offset = h->mb.b_interlaced
221 ? 16 * mb_x + w * (mb_y&~1) * stride + (mb_y&1) * stride
222 : 16 * mb_x + w * mb_y * stride;
223 stride <<= h->mb.b_interlaced;
226 ALIGNED_ARRAY_16( pixel, pix,[FENC_STRIDE*8] );
227 h->mc.load_deinterleave_8x8x2_fenc( pix, frame->plane[1] + offset, stride );
228 return ac_energy_var( h->pixf.var[PIXEL_8x8]( pix, FENC_STRIDE ), 6, frame, i )
229 + ac_energy_var( h->pixf.var[PIXEL_8x8]( pix+FENC_STRIDE/2, FENC_STRIDE ), 6, frame, i );
232 return ac_energy_var( h->pixf.var[PIXEL_16x16]( frame->plane[0] + offset, stride ), 8, frame, i );
235 // Find the total AC energy of the block in all planes.
236 static NOINLINE uint32_t x264_ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
238 /* This function contains annoying hacks because GCC has a habit of reordering emms
239 * and putting it after floating point ops. As a result, we put the emms at the end of the
240 * function and make sure that its always called before the float math. Noinline makes
241 * sure no reordering goes on. */
242 uint32_t var = ac_energy_plane( h, mb_x, mb_y, frame, 0 );
243 var += ac_energy_plane( h, mb_x, mb_y, frame, 1 );
248 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame, float *quant_offsets )
250 /* constants chosen to result in approximately the same overall bitrate as without AQ.
251 * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
254 /* Initialize frame stats */
255 for( int i = 0; i < 3; i++ )
257 frame->i_pixel_sum[i] = 0;
258 frame->i_pixel_ssd[i] = 0;
261 /* Degenerate cases */
262 if( h->param.rc.i_aq_mode == X264_AQ_NONE || h->param.rc.f_aq_strength == 0 )
264 /* Need to init it anyways for MB tree */
265 if( h->param.rc.i_aq_mode && h->param.rc.f_aq_strength == 0 )
269 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
270 frame->f_qp_offset[mb_xy] = frame->f_qp_offset_aq[mb_xy] = quant_offsets[mb_xy];
271 if( h->frames.b_have_lowres )
272 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
273 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8( frame->f_qp_offset[mb_xy] );
277 memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
278 memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
279 if( h->frames.b_have_lowres )
280 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
281 frame->i_inv_qscale_factor[mb_xy] = 256;
284 /* Need variance data for weighted prediction */
285 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE || h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
287 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
288 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
289 x264_ac_energy_mb( h, mb_x, mb_y, frame );
294 /* Actual adaptive quantization */
297 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
299 float bit_depth_correction = powf(1 << (BIT_DEPTH-8), 0.5f);
300 float avg_adj_pow2 = 0.f;
301 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
302 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
304 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
305 float qp_adj = powf( energy + 1, 0.125f );
306 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
308 avg_adj_pow2 += qp_adj * qp_adj;
310 avg_adj /= h->mb.i_mb_count;
311 avg_adj_pow2 /= h->mb.i_mb_count;
312 strength = h->param.rc.f_aq_strength * avg_adj / bit_depth_correction;
313 avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - (14.f * bit_depth_correction)) / avg_adj;
316 strength = h->param.rc.f_aq_strength * 1.0397f;
318 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
319 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
322 int mb_xy = mb_x + mb_y*h->mb.i_mb_stride;
323 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
325 qp_adj = frame->f_qp_offset[mb_xy];
326 qp_adj = strength * (qp_adj - avg_adj);
330 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
331 qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - (14.427f + 2*(BIT_DEPTH-8)));
334 qp_adj += quant_offsets[mb_xy];
335 frame->f_qp_offset[mb_xy] =
336 frame->f_qp_offset_aq[mb_xy] = qp_adj;
337 if( h->frames.b_have_lowres )
338 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8(qp_adj);
342 /* Remove mean from SSD calculation */
343 for( int i = 0; i < 3; i++ )
345 uint64_t ssd = frame->i_pixel_ssd[i];
346 uint64_t sum = frame->i_pixel_sum[i];
347 int width = h->mb.i_mb_width*16>>!!i;
348 int height = h->mb.i_mb_height*16>>!!i;
349 frame->i_pixel_ssd[i] = ssd - (sum * sum + width * height / 2) / (width * height);
353 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame, float *quant_offsets )
355 x264_ratecontrol_t *rc = h->rc;
356 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
358 if( rc->entry[frame->i_frame].kept_as_ref )
361 if( rc->qpbuf_pos < 0 )
367 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
369 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 )
372 if( i_type != i_type_actual && rc->qpbuf_pos == 1 )
374 x264_log(h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type, i_type_actual);
377 } while( i_type != i_type_actual );
380 for( int i = 0; i < h->mb.i_mb_count; i++ )
382 frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[rc->qpbuf_pos][i] )) * (1/256.0);
383 if( h->frames.b_have_lowres )
384 frame->i_inv_qscale_factor[i] = x264_exp2fix8(frame->f_qp_offset[i]);
389 x264_stack_align( x264_adaptive_quant_frame, h, frame, quant_offsets );
392 x264_log(h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n");
396 int x264_reference_build_list_optimal( x264_t *h )
398 ratecontrol_entry_t *rce = h->rc->rce;
399 x264_frame_t *frames[16];
400 x264_weight_t weights[16][3];
403 if( rce->refs != h->i_ref0 )
406 memcpy( frames, h->fref0, sizeof(frames) );
407 memcpy( refcount, rce->refcount, sizeof(refcount) );
408 memcpy( weights, h->fenc->weight, sizeof(weights) );
409 memset( &h->fenc->weight[1][0], 0, sizeof(x264_weight_t[15][3]) );
411 /* For now don't reorder ref 0; it seems to lower quality
412 in most cases due to skips. */
413 for( int ref = 1; ref < h->i_ref0; ref++ )
418 for( int i = 1; i < h->i_ref0; i++ )
419 /* Favor lower POC as a tiebreaker. */
420 COPY2_IF_GT( max, refcount[i], bestref, i );
422 /* FIXME: If there are duplicates from frames other than ref0 then it is possible
423 * that the optimal ordering doesnt place every duplicate. */
425 refcount[bestref] = -1;
426 h->fref0[ref] = frames[bestref];
427 memcpy( h->fenc->weight[ref], weights[bestref], sizeof(weights[bestref]) );
433 static char *x264_strcat_filename( char *input, char *suffix )
435 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
438 strcpy( output, input );
439 strcat( output, suffix );
443 void x264_ratecontrol_init_reconfigurable( x264_t *h, int b_init )
445 x264_ratecontrol_t *rc = h->rc;
446 if( !b_init && rc->b_2pass )
449 if( h->param.rc.i_rc_method == X264_RC_CRF )
451 /* Arbitrary rescaling to make CRF somewhat similar to QP.
452 * Try to compensate for MB-tree's effects as well. */
453 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
454 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
455 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
456 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset );
459 if( h->param.rc.i_vbv_max_bitrate > 0 && h->param.rc.i_vbv_buffer_size > 0 )
461 if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
463 h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
464 x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
465 h->param.rc.i_vbv_buffer_size );
468 /* We don't support changing the ABR bitrate right now,
469 so if the stream starts as CBR, keep it CBR. */
470 if( rc->b_vbv_min_rate )
471 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
473 int vbv_buffer_size = h->param.rc.i_vbv_buffer_size * 1000;
474 int vbv_max_bitrate = h->param.rc.i_vbv_max_bitrate * 1000;
477 h->sps->vui.hrd.i_bit_rate_unscaled = vbv_max_bitrate;
478 h->sps->vui.hrd.i_cpb_size_unscaled = vbv_buffer_size;
479 if( h->param.i_nal_hrd && b_init )
481 h->sps->vui.hrd.i_cpb_cnt = 1;
482 h->sps->vui.hrd.b_cbr_hrd = h->param.i_nal_hrd == X264_NAL_HRD_CBR;
483 h->sps->vui.hrd.i_time_offset_length = 0;
488 int bitrate = 1000*h->param.rc.i_vbv_max_bitrate;
489 int bufsize = 1000*h->param.rc.i_vbv_buffer_size;
491 // normalize HRD size and rate to the value / scale notation
492 h->sps->vui.hrd.i_bit_rate_scale = x264_clip3( x264_ctz( bitrate ) - BR_SHIFT, 0, 15 );
493 h->sps->vui.hrd.i_bit_rate_value = bitrate >> ( h->sps->vui.hrd.i_bit_rate_scale + BR_SHIFT );
494 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 );
495 h->sps->vui.hrd.i_cpb_size_scale = x264_clip3( x264_ctz( bufsize ) - CPB_SHIFT, 0, 15 );
496 h->sps->vui.hrd.i_cpb_size_value = bufsize >> ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
497 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 );
503 #define MAX_DURATION 0.5
505 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 );
506 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;
507 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);
509 h->sps->vui.hrd.i_initial_cpb_removal_delay_length = 2 + x264_clip3( 32 - x264_clz( max_delay ), 4, 22 );
510 h->sps->vui.hrd.i_cpb_removal_delay_length = x264_clip3( 32 - x264_clz( max_cpb_output_delay ), 4, 31 );
511 h->sps->vui.hrd.i_dpb_output_delay_length = x264_clip3( 32 - x264_clz( max_dpb_output_delay ), 4, 31 );
515 vbv_buffer_size = h->sps->vui.hrd.i_cpb_size_unscaled;
516 vbv_max_bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
518 else if( h->param.i_nal_hrd && !b_init )
520 x264_log( h, X264_LOG_WARNING, "VBV parameters cannot be changed when NAL HRD is in use\n" );
524 rc->buffer_rate = vbv_max_bitrate / rc->fps;
525 rc->vbv_max_rate = vbv_max_bitrate;
526 rc->buffer_size = vbv_buffer_size;
527 rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
528 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
529 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
530 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.f_rf_constant_max )
532 rc->rate_factor_max_increment = h->param.rc.f_rf_constant_max - h->param.rc.f_rf_constant;
533 if( rc->rate_factor_max_increment <= 0 )
535 x264_log( h, X264_LOG_WARNING, "CRF max must be greater than CRF\n" );
536 rc->rate_factor_max_increment = 0;
541 if( h->param.rc.f_vbv_buffer_init > 1. )
542 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 );
543 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);
544 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init * h->sps->vui.i_time_scale;
546 rc->b_vbv_min_rate = !rc->b_2pass
547 && h->param.rc.i_rc_method == X264_RC_ABR
548 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
553 int x264_ratecontrol_new( x264_t *h )
555 x264_ratecontrol_t *rc;
559 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
562 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
563 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
565 /* FIXME: use integers */
566 if( h->param.i_fps_num > 0 && h->param.i_fps_den > 0 )
567 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
571 if( h->param.rc.b_mb_tree )
573 h->param.rc.f_pb_factor = 1;
577 rc->qcompress = h->param.rc.f_qcompress;
579 rc->bitrate = h->param.rc.i_bitrate * 1000.;
580 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
581 rc->nmb = h->mb.i_mb_count;
582 rc->last_non_b_pict_type = -1;
585 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
587 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
591 x264_ratecontrol_init_reconfigurable( h, 1 );
593 if( h->param.i_nal_hrd )
595 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale;
596 uint64_t num = 180000;
597 x264_reduce_fraction64( &num, &denom );
598 rc->hrd_multiply_denom = 180000 / num;
600 double bits_required = log2( 180000 / rc->hrd_multiply_denom )
601 + log2( h->sps->vui.i_time_scale )
602 + log2( h->sps->vui.hrd.i_cpb_size_unscaled );
603 if( bits_required >= 63 )
605 x264_log( h, X264_LOG_ERROR, "HRD with very large timescale and bufsize not supported\n" );
610 if( rc->rate_tolerance < 0.01 )
612 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
613 rc->rate_tolerance = 0.01;
616 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
620 /* FIXME ABR_INIT_QP is actually used only in CRF */
621 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
622 rc->accum_p_norm = .01;
623 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
624 /* estimated ratio that produces a reasonable QP for the first I-frame */
625 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
626 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
627 rc->last_non_b_pict_type = SLICE_TYPE_I;
630 rc->ip_offset = 6.0 * log2f( h->param.rc.f_ip_factor );
631 rc->pb_offset = 6.0 * log2f( h->param.rc.f_pb_factor );
632 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
633 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, QP_MAX );
634 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, QP_MAX );
635 h->mb.ip_offset = rc->ip_offset + 0.5;
637 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
638 rc->last_qscale = qp2qscale( 26 );
639 int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
640 CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
641 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
642 for( int i = 0; i < 3; i++ )
644 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
645 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
646 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
647 for( int j = 0; j < num_preds; j++ )
649 rc->pred[i+j*5].coeff= 2.0;
650 rc->pred[i+j*5].count= 1.0;
651 rc->pred[i+j*5].decay= 0.5;
652 rc->pred[i+j*5].offset= 0.0;
654 for( int j = 0; j < 2; j++ )
656 rc->row_preds[i][j].coeff= .25;
657 rc->row_preds[i][j].count= 1.0;
658 rc->row_preds[i][j].decay= 0.5;
659 rc->row_preds[i][j].offset= 0.0;
662 *rc->pred_b_from_p = rc->pred[0];
664 if( parse_zones( h ) < 0 )
666 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
670 /* Load stat file and init 2pass algo */
671 if( h->param.rc.b_stat_read )
673 char *p, *stats_in, *stats_buf;
675 /* read 1st pass stats */
676 assert( h->param.rc.psz_stat_in );
677 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
680 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
683 if( h->param.rc.b_mb_tree )
685 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
686 if( !mbtree_stats_in )
688 rc->p_mbtree_stat_file_in = fopen( mbtree_stats_in, "rb" );
689 x264_free( mbtree_stats_in );
690 if( !rc->p_mbtree_stat_file_in )
692 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
697 /* check whether 1st pass options were compatible with current options */
698 if( !strncmp( stats_buf, "#options:", 9 ) )
702 char *opts = stats_buf;
703 stats_in = strchr( stats_buf, '\n' );
708 if( sscanf( opts, "#options: %dx%d", &i, &j ) != 2 )
710 x264_log( h, X264_LOG_ERROR, "resolution specified in stats file not valid\n" );
713 else if( h->param.rc.b_mb_tree && (i != h->param.i_width || j != h->param.i_height) )
715 x264_log( h, X264_LOG_ERROR, "MB-tree doesn't support different resolution than 1st pass (%dx%d vs %dx%d)\n",
716 h->param.i_width, h->param.i_height, i, j );
720 if( ( p = strstr( opts, "timebase=" ) ) && sscanf( p, "timebase=%u/%u", &k, &l ) != 2 )
722 x264_log( h, X264_LOG_ERROR, "timebase specified in stats file not valid\n" );
725 if( k != h->param.i_timebase_num || l != h->param.i_timebase_den )
727 x264_log( h, X264_LOG_ERROR, "timebase mismatch with 1st pass (%u/%u vs %u/%u)\n",
728 h->param.i_timebase_num, h->param.i_timebase_den, k, l );
732 CMP_OPT_FIRST_PASS( "weightp", X264_MAX( 0, h->param.analyse.i_weighted_pred ) );
733 CMP_OPT_FIRST_PASS( "bframes", h->param.i_bframe );
734 CMP_OPT_FIRST_PASS( "b_pyramid", h->param.i_bframe_pyramid );
735 CMP_OPT_FIRST_PASS( "intra_refresh", h->param.b_intra_refresh );
736 CMP_OPT_FIRST_PASS( "open_gop", h->param.i_open_gop );
738 if( (p = strstr( opts, "keyint=" )) )
741 char buf[13] = "infinite ";
742 if( h->param.i_keyint_max != X264_KEYINT_MAX_INFINITE )
743 sprintf( buf, "%d ", h->param.i_keyint_max );
744 if( strncmp( p, buf, strlen(buf) ) )
746 x264_log( h, X264_LOG_ERROR, "different keyint setting than first pass (%.*s vs %.*s)\n",
747 strlen(buf)-1, buf, strcspn(p, " "), p );
752 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
753 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
755 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
757 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
758 h->mb.b_direct_auto_write = 1;
761 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
762 h->param.i_bframe_adaptive = i;
763 else if( h->param.i_bframe )
765 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
769 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 ) )
770 h->param.rc.i_lookahead = i;
773 /* find number of pics */
776 for( num_entries = -1; p; num_entries++ )
777 p = strchr( p + 1, ';' );
780 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
783 rc->num_entries = num_entries;
785 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
787 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
788 h->param.i_frame_total, rc->num_entries );
790 if( h->param.i_frame_total > rc->num_entries )
792 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
793 h->param.i_frame_total, rc->num_entries );
797 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
799 /* init all to skipped p frames */
800 for( int i = 0; i < rc->num_entries; i++ )
802 ratecontrol_entry_t *rce = &rc->entry[i];
803 rce->pict_type = SLICE_TYPE_P;
804 rce->qscale = rce->new_qscale = qp2qscale( 20 );
805 rce->misc_bits = rc->nmb + 10;
811 for( int i = 0; i < rc->num_entries; i++ )
813 ratecontrol_entry_t *rce;
821 next= strchr(p, ';');
823 *next++ = 0; //sscanf is unbelievably slow on long strings
824 e = sscanf( p, " in:%d ", &frame_number );
826 if( frame_number < 0 || frame_number >= rc->num_entries )
828 x264_log( h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i );
831 rce = &rc->entry[frame_number];
832 rce->direct_mode = 0;
834 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",
835 &pict_type, &rce->i_duration, &rce->i_cpb_duration, &qp, &rce->tex_bits,
836 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
837 &rce->s_count, &rce->direct_mode );
839 p = strstr( p, "ref:" );
843 for( ref = 0; ref < 16; ref++ )
845 if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
847 p = strchr( p+1, ' ' );
854 rce->i_weight_denom = -1;
855 char *w = strchr( p, 'w' );
857 if( sscanf( w, "w:%hd,%hd,%hd", &rce->i_weight_denom, &rce->weight[0], &rce->weight[1] ) != 3 )
858 rce->i_weight_denom = -1;
860 if( pict_type != 'b' )
861 rce->kept_as_ref = 1;
865 rce->frame_type = X264_TYPE_IDR;
866 rce->pict_type = SLICE_TYPE_I;
869 rce->frame_type = X264_TYPE_I;
870 rce->pict_type = SLICE_TYPE_I;
873 rce->frame_type = X264_TYPE_P;
874 rce->pict_type = SLICE_TYPE_P;
877 rce->frame_type = X264_TYPE_BREF;
878 rce->pict_type = SLICE_TYPE_B;
881 rce->frame_type = X264_TYPE_B;
882 rce->pict_type = SLICE_TYPE_B;
884 default: e = -1; break;
889 x264_log( h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e );
892 rce->qscale = qp2qscale( qp );
896 x264_free( stats_buf );
898 if( h->param.rc.i_rc_method == X264_RC_ABR )
900 if( init_pass2( h ) < 0 )
902 } /* else we're using constant quant, so no need to run the bitrate allocation */
905 /* Open output file */
906 /* If input and output files are the same, output to a temp file
907 * and move it to the real name only when it's complete */
908 if( h->param.rc.b_stat_write )
911 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
912 if( !rc->psz_stat_file_tmpname )
915 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
916 if( rc->p_stat_file_out == NULL )
918 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
922 p = x264_param2string( &h->param, 1 );
924 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
926 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
928 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
929 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
930 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
933 rc->p_mbtree_stat_file_out = fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
934 if( rc->p_mbtree_stat_file_out == NULL )
936 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
942 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
944 CHECKED_MALLOC( rc->qp_buffer[0], h->mb.i_mb_count * sizeof(uint16_t) );
945 if( h->param.i_bframe_pyramid && h->param.rc.b_stat_read )
946 CHECKED_MALLOC( rc->qp_buffer[1], h->mb.i_mb_count * sizeof(uint16_t) );
950 for( int i = 0; i<h->param.i_threads; i++ )
952 h->thread[i]->rc = rc+i;
956 h->thread[i]->param = h->param;
957 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
966 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
969 char *tok, UNUSED *saveptr=NULL;
971 z->f_bitrate_factor = 1;
972 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
974 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
976 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
980 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
986 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
987 memcpy( z->param, &h->param, sizeof(x264_param_t) );
988 z->param->param_free = x264_free;
989 while( (tok = strtok_r( p, ",", &saveptr )) )
991 char *val = strchr( tok, '=' );
997 if( x264_param_parse( z->param, tok, val ) )
999 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
1009 static int parse_zones( x264_t *h )
1011 x264_ratecontrol_t *rc = h->rc;
1012 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
1014 char *psz_zones, *p;
1015 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
1016 strcpy( psz_zones, h->param.rc.psz_zones );
1017 h->param.rc.i_zones = 1;
1018 for( p = psz_zones; *p; p++ )
1019 h->param.rc.i_zones += (*p == '/');
1020 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
1022 for( int i = 0; i < h->param.rc.i_zones; i++ )
1024 int i_tok = strcspn( p, "/" );
1026 if( parse_zone( h, &h->param.rc.zones[i], p ) )
1030 x264_free( psz_zones );
1033 if( h->param.rc.i_zones > 0 )
1035 for( int i = 0; i < h->param.rc.i_zones; i++ )
1037 x264_zone_t z = h->param.rc.zones[i];
1038 if( z.i_start < 0 || z.i_start > z.i_end )
1040 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
1041 z.i_start, z.i_end );
1044 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
1046 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
1047 z.f_bitrate_factor );
1052 rc->i_zones = h->param.rc.i_zones + 1;
1053 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
1054 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
1056 // default zone to fall back to if none of the others match
1057 rc->zones[0].i_start = 0;
1058 rc->zones[0].i_end = INT_MAX;
1059 rc->zones[0].b_force_qp = 0;
1060 rc->zones[0].f_bitrate_factor = 1;
1061 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
1062 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
1063 for( int i = 1; i < rc->i_zones; i++ )
1065 if( !rc->zones[i].param )
1066 rc->zones[i].param = rc->zones[0].param;
1075 static x264_zone_t *get_zone( x264_t *h, int frame_num )
1077 for( int i = h->rc->i_zones - 1; i >= 0; i-- )
1079 x264_zone_t *z = &h->rc->zones[i];
1080 if( frame_num >= z->i_start && frame_num <= z->i_end )
1086 void x264_ratecontrol_summary( x264_t *h )
1088 x264_ratecontrol_t *rc = h->rc;
1089 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
1091 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
1092 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
1093 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
1094 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
1095 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset );
1099 void x264_ratecontrol_delete( x264_t *h )
1101 x264_ratecontrol_t *rc = h->rc;
1104 if( rc->p_stat_file_out )
1106 b_regular_file = x264_is_regular_file( rc->p_stat_file_out );
1107 fclose( rc->p_stat_file_out );
1108 if( h->i_frame >= rc->num_entries && b_regular_file )
1109 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
1111 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1112 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
1114 x264_free( rc->psz_stat_file_tmpname );
1116 if( rc->p_mbtree_stat_file_out )
1118 b_regular_file = x264_is_regular_file( rc->p_mbtree_stat_file_out );
1119 fclose( rc->p_mbtree_stat_file_out );
1120 if( h->i_frame >= rc->num_entries && b_regular_file )
1121 if( rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
1123 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1124 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
1126 x264_free( rc->psz_mbtree_stat_file_tmpname );
1127 x264_free( rc->psz_mbtree_stat_file_name );
1129 if( rc->p_mbtree_stat_file_in )
1130 fclose( rc->p_mbtree_stat_file_in );
1131 x264_free( rc->pred );
1132 x264_free( rc->pred_b_from_p );
1133 x264_free( rc->entry );
1134 x264_free( rc->qp_buffer[0] );
1135 x264_free( rc->qp_buffer[1] );
1138 x264_free( rc->zones[0].param );
1139 for( int i = 1; i < rc->i_zones; i++ )
1140 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
1141 rc->zones[i].param->param_free( rc->zones[i].param );
1142 x264_free( rc->zones );
1147 static void accum_p_qp_update( x264_t *h, float qp )
1149 x264_ratecontrol_t *rc = h->rc;
1150 rc->accum_p_qp *= .95;
1151 rc->accum_p_norm *= .95;
1152 rc->accum_p_norm += 1;
1153 if( h->sh.i_type == SLICE_TYPE_I )
1154 rc->accum_p_qp += qp + rc->ip_offset;
1156 rc->accum_p_qp += qp;
1159 /* Before encoding a frame, choose a QP for it */
1160 void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
1162 x264_ratecontrol_t *rc = h->rc;
1163 ratecontrol_entry_t *rce = NULL;
1164 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
1169 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
1170 x264_encoder_reconfig( h, zone->param );
1171 rc->prev_zone = zone;
1173 rc->qp_force = i_force_qp;
1175 if( h->param.rc.b_stat_read )
1177 int frame = h->fenc->i_frame;
1178 assert( frame >= 0 && frame < rc->num_entries );
1179 rce = h->rc->rce = &h->rc->entry[frame];
1181 if( h->sh.i_type == SLICE_TYPE_B
1182 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
1184 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
1185 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
1191 memset( h->fdec->i_row_bits, 0, h->mb.i_mb_height * sizeof(int) );
1192 rc->row_pred = &rc->row_preds[h->sh.i_type];
1193 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;
1194 update_vbv_plan( h, overhead );
1196 const x264_level_t *l = x264_levels;
1197 while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
1200 int mincr = l->mincr;
1202 /* Blu-ray requires this */
1203 if( l->level_idc == 41 && h->param.i_nal_hrd )
1206 /* High 10 doesn't require minCR, so just set the maximum to a large value. */
1207 if( h->sps->i_profile_idc == PROFILE_HIGH10 )
1208 rc->frame_size_maximum = 1e9;
1211 /* The spec has a bizarre special case for the first frame. */
1212 if( h->i_frame == 0 )
1214 //384 * ( Max( PicSizeInMbs, fR * MaxMBPS ) + MaxMBPS * ( tr( 0 ) - tr,n( 0 ) ) ) / MinCR
1215 double fr = 1. / 172;
1216 int pic_size_in_mbs = h->mb.i_mb_width * h->mb.i_mb_height;
1217 rc->frame_size_maximum = 384 * BIT_DEPTH * X264_MAX( pic_size_in_mbs, fr*l->mbps ) / mincr;
1221 //384 * MaxMBPS * ( tr( n ) - tr( n - 1 ) ) / MinCR
1222 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;
1227 if( h->sh.i_type != SLICE_TYPE_B )
1228 rc->bframes = h->fenc->i_bframes;
1234 else if( rc->b_abr )
1236 q = qscale2qp( rate_estimate_qscale( h ) );
1238 else if( rc->b_2pass )
1240 rce->new_qscale = rate_estimate_qscale( h );
1241 q = qscale2qp( rce->new_qscale );
1245 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
1246 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
1248 q = rc->qp_constant[ h->sh.i_type ];
1252 if( zone->b_force_qp )
1253 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1255 q -= 6*log2f( zone->f_bitrate_factor );
1259 q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1263 rc->qp = x264_clip3( (int)(q + 0.5), 0, QP_MAX );
1264 h->fdec->f_qp_avg_rc =
1265 h->fdec->f_qp_avg_aq =
1268 rce->new_qp = rc->qp;
1270 accum_p_qp_update( h, rc->qpm );
1272 if( h->sh.i_type != SLICE_TYPE_B )
1273 rc->last_non_b_pict_type = h->sh.i_type;
1276 static double predict_row_size( x264_t *h, int y, double qp )
1278 /* average between two predictors:
1279 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1280 x264_ratecontrol_t *rc = h->rc;
1281 double pred_s = predict_size( rc->row_pred[0], qp2qscale( qp ), h->fdec->i_row_satd[y] );
1283 if( h->sh.i_type == SLICE_TYPE_I || qp >= h->fref0[0]->f_row_qp[y] )
1285 if( h->sh.i_type == SLICE_TYPE_P
1286 && h->fref0[0]->i_type == h->fdec->i_type
1287 && h->fref0[0]->i_row_satd[y] > 0
1288 && (abs(h->fref0[0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1290 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
1291 * qp2qscale( h->fref0[0]->f_row_qp[y] ) / qp2qscale( qp );
1295 return (pred_s + pred_t) / 2;
1297 /* Our QP is lower than the reference! */
1300 double pred_intra = predict_size( rc->row_pred[1], qp2qscale( qp ), h->fdec->i_row_satds[0][0][y] );
1301 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1302 return pred_intra + pred_s;
1306 static double row_bits_so_far( x264_t *h, int y )
1309 for( int i = h->i_threadslice_start; i <= y; i++ )
1310 bits += h->fdec->i_row_bits[i];
1314 static double predict_row_size_sum( x264_t *h, int y, double qp )
1316 double bits = row_bits_so_far(h, y);
1317 for( int i = y+1; i < h->i_threadslice_end; i++ )
1318 bits += predict_row_size( h, i, qp );
1323 void x264_ratecontrol_mb( x264_t *h, int bits )
1325 x264_ratecontrol_t *rc = h->rc;
1326 const int y = h->mb.i_mb_y;
1330 h->fdec->i_row_bits[y] += bits;
1331 rc->qpa_rc += rc->qpm;
1332 rc->qpa_aq += h->mb.i_qp;
1334 if( h->mb.i_mb_x != h->mb.i_mb_width - 1 || !rc->b_vbv )
1337 h->fdec->f_row_qp[y] = rc->qpm;
1339 update_predictor( rc->row_pred[0], qp2qscale( rc->qpm ), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1340 if( h->sh.i_type == SLICE_TYPE_P && rc->qpm < h->fref0[0]->f_row_qp[y] )
1341 update_predictor( rc->row_pred[1], qp2qscale( rc->qpm ), h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1343 /* tweak quality based on difference from predicted size */
1344 if( y < h->i_threadslice_end-1 )
1346 float prev_row_qp = h->fdec->f_row_qp[y];
1347 float qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1348 float qp_absolute_max = h->param.rc.i_qp_max;
1349 if( rc->rate_factor_max_increment )
1350 qp_absolute_max = X264_MIN( qp_absolute_max, rc->qp_novbv + rc->rate_factor_max_increment );
1351 float qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, qp_absolute_max );
1352 float step_size = 0.5;
1354 /* B-frames shouldn't use lower QP than their reference frames. */
1355 if( h->sh.i_type == SLICE_TYPE_B )
1357 qp_min = X264_MAX( qp_min, X264_MAX( h->fref0[0]->f_row_qp[y+1], h->fref1[0]->f_row_qp[y+1] ) );
1358 rc->qpm = X264_MAX( rc->qpm, qp_min );
1361 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1362 float slice_size_planned = h->param.b_sliced_threads ? rc->slice_size_planned : rc->frame_size_planned;
1363 float size_of_other_slices = 0;
1364 if( h->param.b_sliced_threads )
1366 for( int i = 0; i < h->param.i_threads; i++ )
1367 if( h != h->thread[i] )
1368 size_of_other_slices += h->thread[i]->rc->frame_size_estimated;
1371 rc->max_frame_error = X264_MAX( 0.05, 1.0 / (h->mb.i_mb_width) );
1373 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1374 float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
1375 int b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1377 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1378 /* area at the top of the frame was measured inaccurately. */
1379 if( row_bits_so_far( h, y ) < 0.05 * slice_size_planned )
1382 if( h->sh.i_type != SLICE_TYPE_I )
1385 if( !rc->b_vbv_min_rate )
1386 qp_min = X264_MAX( qp_min, rc->qp_novbv );
1388 while( rc->qpm < qp_max
1389 && ((b1 > rc->frame_size_planned + rc_tol) ||
1390 (rc->buffer_fill - b1 < buffer_left_planned * 0.5) ||
1391 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1393 rc->qpm += step_size;
1394 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1397 while( rc->qpm > qp_min
1398 && (rc->qpm > h->fdec->f_row_qp[0] || rc->single_frame_vbv)
1399 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
1400 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
1402 rc->qpm -= step_size;
1403 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1406 /* avoid VBV underflow or MinCR violation */
1407 while( (rc->qpm < qp_absolute_max)
1408 && ((rc->buffer_fill - b1 < rc->buffer_rate * rc->max_frame_error) ||
1409 (rc->frame_size_maximum - b1 < rc->frame_size_maximum * rc->max_frame_error)))
1411 rc->qpm += step_size;
1412 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1415 h->rc->frame_size_estimated = predict_row_size_sum( h, y, rc->qpm );
1419 int x264_ratecontrol_qp( x264_t *h )
1422 return x264_clip3( h->rc->qpm + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1425 int x264_ratecontrol_mb_qp( x264_t *h )
1428 float qp = h->rc->qpm;
1429 if( h->param.rc.i_aq_mode )
1430 /* MB-tree currently doesn't adjust quantizers in unreferenced frames. */
1431 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];
1432 return x264_clip3( qp + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1435 /* In 2pass, force the same frame types as in the 1st pass */
1436 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1438 x264_ratecontrol_t *rc = h->rc;
1439 if( h->param.rc.b_stat_read )
1441 if( frame_num >= rc->num_entries )
1443 /* We could try to initialize everything required for ABR and
1444 * adaptive B-frames, but that would be complicated.
1445 * So just calculate the average QP used so far. */
1446 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24
1447 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1448 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, QP_MAX );
1449 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 );
1450 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 );
1452 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1453 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1454 if( h->param.i_bframe_adaptive )
1455 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1457 for( int i = 0; i < h->param.i_threads; i++ )
1459 h->thread[i]->rc->b_abr = 0;
1460 h->thread[i]->rc->b_2pass = 0;
1461 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1462 h->thread[i]->param.rc.b_stat_read = 0;
1463 h->thread[i]->param.i_bframe_adaptive = 0;
1464 h->thread[i]->param.i_scenecut_threshold = 0;
1465 h->thread[i]->param.rc.b_mb_tree = 0;
1466 if( h->thread[i]->param.i_bframe > 1 )
1467 h->thread[i]->param.i_bframe = 1;
1469 return X264_TYPE_AUTO;
1471 return rc->entry[frame_num].frame_type;
1474 return X264_TYPE_AUTO;
1477 void x264_ratecontrol_set_weights( x264_t *h, x264_frame_t *frm )
1479 ratecontrol_entry_t *rce = &h->rc->entry[frm->i_frame];
1480 if( h->param.analyse.i_weighted_pred <= 0 )
1482 if( rce->i_weight_denom >= 0 )
1483 SET_WEIGHT( frm->weight[0][0], 1, rce->weight[0], rce->i_weight_denom, rce->weight[1] );
1486 /* After encoding one frame, save stats and update ratecontrol state */
1487 int x264_ratecontrol_end( x264_t *h, int bits, int *filler )
1489 x264_ratecontrol_t *rc = h->rc;
1490 const int *mbs = h->stat.frame.i_mb_count;
1494 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1495 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1496 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1497 for( int i = B_DIRECT; i < B_8x8; i++ )
1498 h->stat.frame.i_mb_count_p += mbs[i];
1500 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1501 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1503 if( h->param.rc.b_stat_write )
1505 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1506 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1507 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1508 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1509 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1510 char c_direct = h->mb.b_direct_auto_write ?
1511 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1512 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1514 if( fprintf( rc->p_stat_file_out,
1515 "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:",
1516 h->fenc->i_frame, h->i_frame,
1517 c_type, h->fenc->i_duration,
1518 h->fenc->i_cpb_duration, rc->qpa_rc,
1519 h->stat.frame.i_tex_bits,
1520 h->stat.frame.i_mv_bits,
1521 h->stat.frame.i_misc_bits,
1522 h->stat.frame.i_mb_count_i,
1523 h->stat.frame.i_mb_count_p,
1524 h->stat.frame.i_mb_count_skip,
1528 /* Only write information for reference reordering once. */
1529 int use_old_stats = h->param.rc.b_stat_read && rc->rce->refs > 1;
1530 for( int i = 0; i < (use_old_stats ? rc->rce->refs : h->i_ref0); i++ )
1532 int refcount = use_old_stats ? rc->rce->refcount[i]
1533 : h->param.b_interlaced ? h->stat.frame.i_mb_count_ref[0][i*2]
1534 + h->stat.frame.i_mb_count_ref[0][i*2+1]
1535 : h->stat.frame.i_mb_count_ref[0][i];
1536 if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1540 if( h->sh.weight[0][0].weightfn )
1542 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 )
1546 if( fprintf( rc->p_stat_file_out, ";\n") < 0 )
1549 /* Don't re-write the data in multi-pass mode. */
1550 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1552 uint8_t i_type = h->sh.i_type;
1553 /* Values are stored as big-endian FIX8.8 */
1554 for( int i = 0; i < h->mb.i_mb_count; i++ )
1555 rc->qp_buffer[0][i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1556 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1558 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 )
1565 if( h->sh.i_type != SLICE_TYPE_B )
1566 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / rc->last_rceq;
1569 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1570 * Not perfectly accurate with B-refs, but good enough. */
1571 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / (rc->last_rceq * fabs( h->param.rc.f_pb_factor ));
1573 rc->cplxr_sum *= rc->cbr_decay;
1574 double frame_duration = (double)h->fenc->i_duration * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1576 rc->wanted_bits_window += frame_duration * rc->bitrate;
1577 rc->wanted_bits_window *= rc->cbr_decay;
1581 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale( rc->rce->new_qp ) );
1583 if( h->mb.b_variable_qp )
1585 if( h->sh.i_type == SLICE_TYPE_B )
1587 rc->bframe_bits += bits;
1588 if( h->fenc->b_last_minigop_bframe )
1590 update_predictor( rc->pred_b_from_p, qp2qscale( rc->qpa_rc ),
1591 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1592 rc->bframe_bits = 0;
1597 *filler = update_vbv( h, bits );
1599 if( h->sps->vui.b_nal_hrd_parameters_present )
1601 if( h->fenc->i_frame == 0 )
1603 // access unit initialises the HRD
1604 h->fenc->hrd_timing.cpb_initial_arrival_time = 0;
1605 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1606 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1607 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit = (double)rc->initial_cpb_removal_delay / 90000;
1611 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit + (double)h->fenc->i_cpb_delay *
1612 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1614 double cpb_earliest_arrival_time = h->fenc->hrd_timing.cpb_removal_time - (double)rc->initial_cpb_removal_delay / 90000;
1615 if( h->fenc->b_keyframe )
1617 rc->nrt_first_access_unit = h->fenc->hrd_timing.cpb_removal_time;
1618 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1619 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1622 cpb_earliest_arrival_time -= (double)rc->initial_cpb_removal_delay_offset / 90000;
1624 if( h->sps->vui.hrd.b_cbr_hrd )
1625 h->fenc->hrd_timing.cpb_initial_arrival_time = rc->previous_cpb_final_arrival_time;
1627 h->fenc->hrd_timing.cpb_initial_arrival_time = X264_MAX( rc->previous_cpb_final_arrival_time, cpb_earliest_arrival_time );
1629 int filler_bits = *filler ? X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), *filler )*8 : 0;
1631 h->fenc->hrd_timing.cpb_final_arrival_time = rc->previous_cpb_final_arrival_time = h->fenc->hrd_timing.cpb_initial_arrival_time +
1632 (double)(bits + filler_bits) / h->sps->vui.hrd.i_bit_rate_unscaled;
1634 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 +
1635 h->fenc->hrd_timing.cpb_removal_time;
1640 x264_log(h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n");
1644 /****************************************************************************
1646 ***************************************************************************/
1649 * modify the bitrate curve from pass1 for one frame
1651 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1653 x264_ratecontrol_t *rcc= h->rc;
1654 x264_zone_t *zone = get_zone( h, frame_num );
1655 double q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1657 // avoid NaN's in the rc_eq
1658 if( !isfinite(q) || rce->tex_bits + rce->mv_bits == 0 )
1659 q = rcc->last_qscale_for[rce->pict_type];
1664 rcc->last_qscale = q;
1669 if( zone->b_force_qp )
1670 q = qp2qscale( zone->i_qp );
1672 q /= zone->f_bitrate_factor;
1678 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1680 x264_ratecontrol_t *rcc = h->rc;
1681 const int pict_type = rce->pict_type;
1683 // force I/B quants as a function of P quants
1684 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1685 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1686 if( pict_type == SLICE_TYPE_I )
1689 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1690 double ip_factor = fabs( h->param.rc.f_ip_factor );
1691 /* don't apply ip_factor if the following frame is also I */
1692 if( rcc->accum_p_norm <= 0 )
1694 else if( h->param.rc.f_ip_factor < 0 )
1696 else if( rcc->accum_p_norm >= 1 )
1699 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1701 else if( pict_type == SLICE_TYPE_B )
1703 if( h->param.rc.f_pb_factor > 0 )
1705 if( !rce->kept_as_ref )
1706 q *= fabs( h->param.rc.f_pb_factor );
1708 else if( pict_type == SLICE_TYPE_P
1709 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1710 && rce->tex_bits == 0 )
1715 /* last qscale / qdiff stuff */
1716 if( rcc->last_non_b_pict_type == pict_type &&
1717 (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1) )
1719 double last_q = rcc->last_qscale_for[pict_type];
1720 double max_qscale = last_q * rcc->lstep;
1721 double min_qscale = last_q / rcc->lstep;
1723 if ( q > max_qscale ) q = max_qscale;
1724 else if( q < min_qscale ) q = min_qscale;
1727 rcc->last_qscale_for[pict_type] = q;
1728 if( pict_type != SLICE_TYPE_B )
1729 rcc->last_non_b_pict_type = pict_type;
1730 if( pict_type == SLICE_TYPE_I )
1732 rcc->last_accum_p_norm = rcc->accum_p_norm;
1733 rcc->accum_p_norm = 0;
1734 rcc->accum_p_qp = 0;
1736 if( pict_type == SLICE_TYPE_P )
1738 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1739 rcc->accum_p_qp = mask * (qscale2qp( q ) + rcc->accum_p_qp);
1740 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1745 static double predict_size( predictor_t *p, double q, double var )
1747 return (p->coeff*var + p->offset) / (q*p->count);
1750 static void update_predictor( predictor_t *p, double q, double var, double bits )
1752 const double range = 1.5;
1755 double old_coeff = p->coeff / p->count;
1756 double new_coeff = bits*q / var;
1757 double new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
1758 double new_offset = bits*q - new_coeff_clipped * var;
1759 if( new_offset >= 0 )
1760 new_coeff = new_coeff_clipped;
1763 p->count *= p->decay;
1764 p->coeff *= p->decay;
1765 p->offset *= p->decay;
1767 p->coeff += new_coeff;
1768 p->offset += new_offset;
1771 // update VBV after encoding a frame
1772 static int update_vbv( x264_t *h, int bits )
1775 int bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
1776 x264_ratecontrol_t *rcc = h->rc;
1777 x264_ratecontrol_t *rct = h->thread[0]->rc;
1778 uint64_t buffer_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
1780 if( rcc->last_satd >= h->mb.i_mb_count )
1781 update_predictor( &rct->pred[h->sh.i_type], qp2qscale( rcc->qpa_rc ), rcc->last_satd, bits );
1786 rct->buffer_fill_final -= (uint64_t)bits * h->sps->vui.i_time_scale;
1788 if( rct->buffer_fill_final < 0 )
1789 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 );
1790 rct->buffer_fill_final = X264_MAX( rct->buffer_fill_final, 0 );
1791 rct->buffer_fill_final += (uint64_t)bitrate * h->sps->vui.i_num_units_in_tick * h->fenc->i_cpb_duration;
1793 if( h->sps->vui.hrd.b_cbr_hrd && rct->buffer_fill_final > buffer_size )
1795 filler = ceil( (rct->buffer_fill_final - buffer_size) / (8. * h->sps->vui.i_time_scale) );
1796 bits = X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), filler ) * 8;
1797 rct->buffer_fill_final -= (uint64_t)bits * h->sps->vui.i_time_scale;
1800 rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, buffer_size );
1805 void x264_hrd_fullness( x264_t *h )
1807 x264_ratecontrol_t *rct = h->thread[0]->rc;
1808 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale / rct->hrd_multiply_denom;
1809 uint64_t cpb_state = rct->buffer_fill_final;
1810 uint64_t cpb_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
1811 uint64_t multiply_factor = 180000 / rct->hrd_multiply_denom;
1813 if( rct->buffer_fill_final < 0 || rct->buffer_fill_final > cpb_size )
1815 x264_log( h, X264_LOG_WARNING, "CPB %s: %.0lf bits in a %.0lf-bit buffer\n",
1816 rct->buffer_fill_final < 0 ? "underflow" : "overflow", (float)rct->buffer_fill_final/denom, (float)cpb_size/denom );
1819 h->initial_cpb_removal_delay = (multiply_factor * cpb_state + denom) / (2*denom);
1820 h->initial_cpb_removal_delay_offset = (multiply_factor * cpb_size + denom) / (2*denom) - h->initial_cpb_removal_delay;
1823 // provisionally update VBV according to the planned size of all frames currently in progress
1824 static void update_vbv_plan( x264_t *h, int overhead )
1826 x264_ratecontrol_t *rcc = h->rc;
1827 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final / h->sps->vui.i_time_scale;
1828 if( h->i_thread_frames > 1 )
1830 int j = h->rc - h->thread[0]->rc;
1831 for( int i = 1; i < h->i_thread_frames; i++ )
1833 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
1834 double bits = t->rc->frame_size_planned;
1835 if( !t->b_thread_active )
1837 bits = X264_MAX(bits, t->rc->frame_size_estimated);
1838 rcc->buffer_fill -= bits;
1839 rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
1840 rcc->buffer_fill += t->rc->buffer_rate;
1841 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1844 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1845 rcc->buffer_fill -= overhead;
1848 // apply VBV constraints and clip qscale to between lmin and lmax
1849 static double clip_qscale( x264_t *h, int pict_type, double q )
1851 x264_ratecontrol_t *rcc = h->rc;
1852 double lmin = rcc->lmin[pict_type];
1853 double lmax = rcc->lmax[pict_type];
1854 if( rcc->rate_factor_max_increment )
1855 lmax = X264_MIN( lmax, qp2qscale( rcc->qp_novbv + rcc->rate_factor_max_increment ) );
1858 /* B-frames are not directly subject to VBV,
1859 * since they are controlled by the P-frames' QPs. */
1861 if( rcc->b_vbv && rcc->last_satd > 0 )
1863 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
1864 * the lookahead overflow and such that the buffer is in a reasonable state
1865 * by the end of the lookahead. */
1866 if( h->param.rc.i_lookahead )
1870 /* Avoid an infinite loop. */
1871 for( int iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
1874 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1875 double buffer_fill_cur = rcc->buffer_fill - cur_bits;
1877 double total_duration = 0;
1878 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
1879 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
1880 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
1882 /* Loop over the planned future frames. */
1883 for( int j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
1885 total_duration += h->fenc->f_planned_cpb_duration[j];
1886 buffer_fill_cur += rcc->vbv_max_rate * h->fenc->f_planned_cpb_duration[j];
1887 int i_type = h->fenc->i_planned_type[j];
1888 int i_satd = h->fenc->i_planned_satd[j];
1889 if( i_type == X264_TYPE_AUTO )
1891 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
1892 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
1893 buffer_fill_cur -= cur_bits;
1895 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
1896 target_fill = X264_MIN( rcc->buffer_fill + total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.5 );
1897 if( buffer_fill_cur < target_fill )
1903 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
1904 target_fill = x264_clip3f( rcc->buffer_fill - total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
1905 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
1914 /* Fallback to old purely-reactive algorithm: no lookahead. */
1917 if( ( pict_type == SLICE_TYPE_P ||
1918 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
1919 rcc->buffer_fill/rcc->buffer_size < 0.5 )
1921 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1924 /* Now a hard threshold to make sure the frame fits in VBV.
1925 * This one is mostly for I-frames. */
1926 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1928 /* For small VBVs, allow the frame to use up the entire VBV. */
1929 double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
1930 /* For single-frame VBVs, request that the frame use up the entire VBV. */
1931 double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
1933 if( bits > rcc->buffer_fill/max_fill_factor )
1934 qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
1937 if( bits < rcc->buffer_rate/min_fill_factor )
1938 q *= bits*min_fill_factor/rcc->buffer_rate;
1939 q = X264_MAX( q0, q );
1942 /* Apply MinCR restrictions */
1943 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1944 if( bits > rcc->frame_size_maximum )
1945 q *= bits / rcc->frame_size_maximum;
1946 bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1948 /* Check B-frame complexity, and use up any bits that would
1949 * overflow before the next P-frame. */
1950 if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
1952 int nb = rcc->bframes;
1953 double pbbits = bits;
1954 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1956 double bframe_cpb_duration = 0;
1957 double minigop_cpb_duration;
1958 for( int i = 0; i < nb; i++ )
1959 bframe_cpb_duration += h->fenc->f_planned_cpb_duration[1+i];
1961 if( bbits * nb > bframe_cpb_duration * rcc->vbv_max_rate )
1963 pbbits += nb * bbits;
1965 minigop_cpb_duration = bframe_cpb_duration + h->fenc->f_planned_cpb_duration[0];
1966 space = rcc->buffer_fill + minigop_cpb_duration*rcc->vbv_max_rate - rcc->buffer_size;
1967 if( pbbits < space )
1969 q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
1971 q = X264_MAX( q0-5, q );
1974 if( !rcc->b_vbv_min_rate )
1975 q = X264_MAX( q0, q );
1980 else if( rcc->b_2pass )
1982 double min2 = log( lmin );
1983 double max2 = log( lmax );
1984 q = (log(q) - min2)/(max2-min2) - 0.5;
1985 q = 1.0/(1.0 + exp( -4*q ));
1986 q = q*(max2-min2) + min2;
1990 return x264_clip3f( q, lmin, lmax );
1993 // update qscale for 1 frame based on actual bits used so far
1994 static float rate_estimate_qscale( x264_t *h )
1997 x264_ratecontrol_t *rcc = h->rc;
1998 ratecontrol_entry_t rce;
1999 int pict_type = h->sh.i_type;
2000 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
2001 + h->stat.i_frame_size[SLICE_TYPE_P]
2002 + h->stat.i_frame_size[SLICE_TYPE_B]);
2007 if( pict_type != rce.pict_type )
2009 x264_log( h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
2010 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type] );
2014 if( pict_type == SLICE_TYPE_B )
2016 /* B-frames don't have independent ratecontrol, but rather get the
2017 * average QP of the two adjacent P-frames + an offset */
2019 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
2020 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
2021 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
2022 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
2023 float q0 = h->fref0[0]->f_qp_avg_rc;
2024 float q1 = h->fref1[0]->f_qp_avg_rc;
2026 if( h->fref0[0]->i_type == X264_TYPE_BREF )
2027 q0 -= rcc->pb_offset/2;
2028 if( h->fref1[0]->i_type == X264_TYPE_BREF )
2029 q1 -= rcc->pb_offset/2;
2032 q = (q0 + q1) / 2 + rcc->ip_offset;
2038 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
2040 if( h->fenc->b_kept_as_ref )
2041 q += rcc->pb_offset/2;
2043 q += rcc->pb_offset;
2045 if( rcc->b_2pass && rcc->b_vbv )
2046 rcc->frame_size_planned = qscale2bits( &rce, q );
2048 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
2049 h->rc->frame_size_estimated = rcc->frame_size_planned;
2053 rcc->last_satd = x264_rc_analyse_slice( h );
2055 return qp2qscale( q );
2059 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
2063 double lmin = rcc->lmin[pict_type];
2064 double lmax = rcc->lmax[pict_type];
2066 int64_t predicted_bits = total_bits;
2070 if( h->i_thread_frames > 1 )
2072 int j = h->rc - h->thread[0]->rc;
2073 for( int i = 1; i < h->i_thread_frames; i++ )
2075 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
2076 double bits = t->rc->frame_size_planned;
2077 if( !t->b_thread_active )
2079 bits = X264_MAX(bits, t->rc->frame_size_estimated);
2080 predicted_bits += (int64_t)bits;
2086 if( h->i_frame < h->i_thread_frames )
2087 predicted_bits += (int64_t)h->i_frame * rcc->bitrate / rcc->fps;
2089 predicted_bits += (int64_t)(h->i_thread_frames - 1) * rcc->bitrate / rcc->fps;
2092 /* Adjust ABR buffer based on distance to the end of the video. */
2093 if( rcc->num_entries > h->i_frame )
2095 double final_bits = rcc->entry[rcc->num_entries-1].expected_bits;
2096 double video_pos = rce.expected_bits / final_bits;
2097 double scale_factor = sqrt( (1 - video_pos) * rcc->num_entries );
2098 abr_buffer *= 0.5 * X264_MAX( scale_factor, 0.5 );
2101 diff = predicted_bits - (int64_t)rce.expected_bits;
2103 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
2104 if( ((h->i_frame + 1 - h->i_thread_frames) >= rcc->fps) &&
2105 (rcc->expected_bits_sum > 0))
2107 /* Adjust quant based on the difference between
2108 * achieved and expected bitrate so far */
2109 double cur_time = (double)h->i_frame / rcc->num_entries;
2110 double w = x264_clip3f( cur_time*100, 0.0, 1.0 );
2111 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
2115 /* Do not overflow vbv */
2116 double expected_size = qscale2bits( &rce, q );
2117 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2118 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
2119 double qmax = q*(2 - expected_fullness);
2120 double size_constraint = 1 + expected_fullness;
2121 qmax = X264_MAX( qmax, rce.new_qscale );
2122 if( expected_fullness < .05 )
2124 qmax = X264_MIN(qmax, lmax);
2125 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
2126 ((expected_vbv < 0) && (q < lmax)))
2129 expected_size = qscale2bits(&rce, q);
2130 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2132 rcc->last_satd = x264_rc_analyse_slice( h );
2134 q = x264_clip3f( q, lmin, lmax );
2136 else /* 1pass ABR */
2138 /* Calculate the quantizer which would have produced the desired
2139 * average bitrate if it had been applied to all frames so far.
2140 * Then modulate that quant based on the current frame's complexity
2141 * relative to the average complexity so far (using the 2pass RCEQ).
2142 * Then bias the quant up or down if total size so far was far from
2144 * Result: Depending on the value of rate_tolerance, there is a
2145 * tradeoff between quality and bitrate precision. But at large
2146 * tolerances, the bit distribution approaches that of 2pass. */
2148 double wanted_bits, overflow = 1;
2150 rcc->last_satd = x264_rc_analyse_slice( h );
2151 rcc->short_term_cplxsum *= 0.5;
2152 rcc->short_term_cplxcount *= 0.5;
2153 rcc->short_term_cplxsum += rcc->last_satd;
2154 rcc->short_term_cplxcount ++;
2156 rce.tex_bits = rcc->last_satd;
2157 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
2159 rce.p_count = rcc->nmb;
2163 rce.pict_type = pict_type;
2165 if( h->param.rc.i_rc_method == X264_RC_CRF )
2167 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
2171 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
2173 /* ABR code can potentially be counterproductive in CBR, so just don't bother.
2174 * Don't run it if the frame complexity is zero either. */
2175 if( !rcc->b_vbv_min_rate && rcc->last_satd )
2177 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
2178 int i_frame_done = h->i_frame + 1 - h->i_thread_frames;
2179 double time_done = i_frame_done / rcc->fps;
2180 if( h->param.b_vfr_input && i_frame_done > 0 )
2181 time_done = ((double)(h->fenc->i_reordered_pts - h->i_reordered_pts_delay)) * h->param.i_timebase_num / h->param.i_timebase_den;
2182 wanted_bits = time_done * rcc->bitrate;
2183 if( wanted_bits > 0 )
2185 abr_buffer *= X264_MAX( 1, sqrt( time_done ) );
2186 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
2192 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
2193 /* should test _next_ pict type, but that isn't decided yet */
2194 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
2196 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2197 q /= fabs( h->param.rc.f_ip_factor );
2199 else if( h->i_frame > 0 )
2201 /* Asymmetric clipping, because symmetric would prevent
2202 * overflow control in areas of rapidly oscillating complexity */
2203 double lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
2204 double lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
2205 if( overflow > 1.1 && h->i_frame > 3 )
2207 else if( overflow < 0.9 )
2210 q = x264_clip3f(q, lmin, lmax);
2212 else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
2214 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
2216 rcc->qp_novbv = qscale2qp( q );
2218 //FIXME use get_diff_limited_q() ?
2219 q = clip_qscale( h, pict_type, q );
2222 rcc->last_qscale_for[pict_type] =
2223 rcc->last_qscale = q;
2225 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
2226 rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );
2228 if( rcc->b_2pass && rcc->b_vbv )
2229 rcc->frame_size_planned = qscale2bits(&rce, q);
2231 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2233 /* Always use up the whole VBV in this case. */
2234 if( rcc->single_frame_vbv )
2235 rcc->frame_size_planned = rcc->buffer_rate;
2236 h->rc->frame_size_estimated = rcc->frame_size_planned;
2241 void x264_threads_normalize_predictors( x264_t *h )
2243 double totalsize = 0;
2244 for( int i = 0; i < h->param.i_threads; i++ )
2245 totalsize += h->thread[i]->rc->slice_size_planned;
2246 double factor = h->rc->frame_size_planned / totalsize;
2247 for( int i = 0; i < h->param.i_threads; i++ )
2248 h->thread[i]->rc->slice_size_planned *= factor;
2251 void x264_threads_distribute_ratecontrol( x264_t *h )
2254 x264_ratecontrol_t *rc = h->rc;
2256 /* Initialize row predictors */
2257 if( h->i_frame == 0 )
2258 for( int i = 0; i < h->param.i_threads; i++ )
2260 x264_ratecontrol_t *t = h->thread[i]->rc;
2261 memcpy( t->row_preds, rc->row_preds, sizeof(rc->row_preds) );
2264 for( int i = 0; i < h->param.i_threads; i++ )
2266 x264_t *t = h->thread[i];
2267 memcpy( t->rc, rc, offsetof(x264_ratecontrol_t, row_pred) );
2268 t->rc->row_pred = &t->rc->row_preds[h->sh.i_type];
2269 /* Calculate the planned slice size. */
2270 if( rc->b_vbv && rc->frame_size_planned )
2273 for( row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2274 size += h->fdec->i_row_satd[row];
2275 t->rc->slice_size_planned = predict_size( &rc->pred[h->sh.i_type + (i+1)*5], rc->qpm, size );
2278 t->rc->slice_size_planned = 0;
2280 if( rc->b_vbv && rc->frame_size_planned )
2282 x264_threads_normalize_predictors( h );
2284 if( rc->single_frame_vbv )
2286 /* Compensate for our max frame error threshold: give more bits (proportionally) to smaller slices. */
2287 for( int i = 0; i < h->param.i_threads; i++ )
2289 x264_t *t = h->thread[i];
2290 t->rc->max_frame_error = X264_MAX( 0.05, 1.0 / (t->i_threadslice_end - t->i_threadslice_start) );
2291 t->rc->slice_size_planned += 2 * t->rc->max_frame_error * rc->frame_size_planned;
2293 x264_threads_normalize_predictors( h );
2296 for( int i = 0; i < h->param.i_threads; i++ )
2297 h->thread[i]->rc->frame_size_estimated = h->thread[i]->rc->slice_size_planned;
2301 void x264_threads_merge_ratecontrol( x264_t *h )
2303 x264_ratecontrol_t *rc = h->rc;
2306 for( int i = 0; i < h->param.i_threads; i++ )
2308 x264_t *t = h->thread[i];
2309 x264_ratecontrol_t *rct = h->thread[i]->rc;
2310 if( h->param.rc.i_vbv_buffer_size )
2313 for( int row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2314 size += h->fdec->i_row_satd[row];
2315 int bits = t->stat.frame.i_mv_bits + t->stat.frame.i_tex_bits + t->stat.frame.i_misc_bits;
2316 int mb_count = (t->i_threadslice_end - t->i_threadslice_start) * h->mb.i_mb_width;
2317 update_predictor( &rc->pred[h->sh.i_type+(i+1)*5], qp2qscale( rct->qpa_rc/mb_count ), size, bits );
2321 rc->qpa_rc += rct->qpa_rc;
2322 rc->qpa_aq += rct->qpa_aq;
2326 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
2330 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
2331 /* these vars are updated in x264_ratecontrol_start()
2332 * so copy them from the context that most recently started (prev)
2333 * to the context that's about to start (cur). */
2338 COPY(last_qscale_for);
2339 COPY(last_non_b_pict_type);
2340 COPY(short_term_cplxsum);
2341 COPY(short_term_cplxcount);
2345 /* these vars can be updated by x264_ratecontrol_init_reconfigurable */
2348 COPY(single_frame_vbv);
2350 COPY(b_vbv_min_rate);
2351 COPY(rate_factor_constant);
2357 #define COPY(var) next->rc->var = cur->rc->var
2358 /* these vars are updated in x264_ratecontrol_end()
2359 * so copy them from the context that most recently ended (cur)
2360 * to the context that's about to end (next) */
2362 COPY(expected_bits_sum);
2363 COPY(wanted_bits_window);
2365 COPY(initial_cpb_removal_delay);
2366 COPY(initial_cpb_removal_delay_offset);
2367 COPY(nrt_first_access_unit);
2368 COPY(previous_cpb_final_arrival_time);
2371 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
2372 /* the rest of the variables are either constant or thread-local */
2375 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
2377 /* find an interval ending on an overflow or underflow (depending on whether
2378 * we're adding or removing bits), and starting on the earliest frame that
2379 * can influence the buffer fill of that end frame. */
2380 x264_ratecontrol_t *rcc = h->rc;
2381 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
2382 const double buffer_max = .9 * rcc->buffer_size;
2383 double fill = fills[*t0-1];
2384 double parity = over ? 1. : -1.;
2385 int start = -1, end = -1;
2386 for( int i = *t0; i < rcc->num_entries; i++ )
2388 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 -
2389 qscale2bits( &rcc->entry[i], rcc->entry[i].new_qscale )) * parity;
2390 fill = x264_clip3f(fill, 0, rcc->buffer_size);
2392 if( fill <= buffer_min || i == 0 )
2398 else if( fill >= buffer_max && start >= 0 )
2403 return start >= 0 && end >= 0;
2406 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
2408 x264_ratecontrol_t *rcc = h->rc;
2409 double qscale_orig, qscale_new;
2413 for( int i = t0; i <= t1; i++ )
2415 qscale_orig = rcc->entry[i].new_qscale;
2416 qscale_orig = x264_clip3f( qscale_orig, qscale_min, qscale_max );
2417 qscale_new = qscale_orig * adjustment;
2418 qscale_new = x264_clip3f( qscale_new, qscale_min, qscale_max );
2419 rcc->entry[i].new_qscale = qscale_new;
2420 adjusted = adjusted || (qscale_new != qscale_orig);
2425 static double count_expected_bits( x264_t *h )
2427 x264_ratecontrol_t *rcc = h->rc;
2428 double expected_bits = 0;
2429 for( int i = 0; i < rcc->num_entries; i++ )
2431 ratecontrol_entry_t *rce = &rcc->entry[i];
2432 rce->expected_bits = expected_bits;
2433 expected_bits += qscale2bits( rce, rce->new_qscale );
2435 return expected_bits;
2438 static int vbv_pass2( x264_t *h, double all_available_bits )
2440 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
2441 * frames in the interval until either buffer is full at some intermediate frame or the
2442 * last frame in the interval no longer underflows. Recompute intervals and repeat.
2443 * Then do the converse to put bits back into overflow areas until target size is met */
2445 x264_ratecontrol_t *rcc = h->rc;
2447 double expected_bits = 0;
2449 double prev_bits = 0;
2451 double qscale_min = qp2qscale( h->param.rc.i_qp_min );
2452 double qscale_max = qp2qscale( h->param.rc.i_qp_max );
2454 int adj_min, adj_max;
2455 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2459 /* adjust overall stream size */
2463 prev_bits = expected_bits;
2466 { /* not first iteration */
2467 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2468 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2472 while(adj_min && find_underflow( h, fills, &t0, &t1, 1 ))
2474 adj_min = fix_underflow( h, t0, t1, adjustment, qscale_min, qscale_max );
2479 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2481 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2483 while( adj_max && find_underflow( h, fills, &t0, &t1, 0 ) )
2484 adj_max = fix_underflow( h, t0, t1, 1.001, qscale_min, qscale_max );
2486 expected_bits = count_expected_bits( h );
2487 } while( (expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2490 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2492 /* store expected vbv filling values for tracking when encoding */
2493 for( int i = 0; i < rcc->num_entries; i++ )
2494 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
2496 x264_free( fills-1 );
2502 static int init_pass2( x264_t *h )
2504 x264_ratecontrol_t *rcc = h->rc;
2505 uint64_t all_const_bits = 0;
2506 double duration = 0;
2507 for( int i = 0; i < rcc->num_entries; i++ )
2508 duration += rcc->entry[i].i_duration;
2509 duration *= (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2510 uint64_t all_available_bits = h->param.rc.i_bitrate * 1000. * duration;
2511 double rate_factor, step_mult;
2512 double qblur = h->param.rc.f_qblur;
2513 double cplxblur = h->param.rc.f_complexity_blur;
2514 const int filter_size = (int)(qblur*4) | 1;
2515 double expected_bits;
2516 double *qscale, *blurred_qscale;
2518 /* find total/average complexity & const_bits */
2519 for( int i = 0; i < rcc->num_entries; i++ )
2521 ratecontrol_entry_t *rce = &rcc->entry[i];
2522 all_const_bits += rce->misc_bits;
2525 if( all_available_bits < all_const_bits)
2527 x264_log( h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2528 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)) );
2532 /* Blur complexities, to reduce local fluctuation of QP.
2533 * We don't blur the QPs directly, because then one very simple frame
2534 * could drag down the QP of a nearby complex frame and give it more
2535 * bits than intended. */
2536 for( int i = 0; i < rcc->num_entries; i++ )
2538 ratecontrol_entry_t *rce = &rcc->entry[i];
2539 double weight_sum = 0;
2540 double cplx_sum = 0;
2541 double weight = 1.0;
2542 double gaussian_weight;
2543 /* weighted average of cplx of future frames */
2544 for( int j = 1; j < cplxblur*2 && j < rcc->num_entries-i; j++ )
2546 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2547 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2548 if( weight < .0001 )
2550 gaussian_weight = weight * exp( -j*j/200.0 );
2551 weight_sum += gaussian_weight;
2552 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2554 /* weighted average of cplx of past frames */
2556 for( int j = 0; j <= cplxblur*2 && j <= i; j++ )
2558 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2559 gaussian_weight = weight * exp( -j*j/200.0 );
2560 weight_sum += gaussian_weight;
2561 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits);
2562 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2563 if( weight < .0001 )
2566 rce->blurred_complexity = cplx_sum / weight_sum;
2569 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2570 if( filter_size > 1 )
2571 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2573 blurred_qscale = qscale;
2575 /* Search for a factor which, when multiplied by the RCEQ values from
2576 * each frame, adds up to the desired total size.
2577 * There is no exact closed-form solution because of VBV constraints and
2578 * because qscale2bits is not invertible, but we can start with the simple
2579 * approximation of scaling the 1st pass by the ratio of bitrates.
2580 * The search range is probably overkill, but speed doesn't matter here. */
2583 for( int i = 0; i < rcc->num_entries; i++ )
2585 double q = get_qscale(h, &rcc->entry[i], 1.0, i);
2586 expected_bits += qscale2bits(&rcc->entry[i], q);
2587 rcc->last_qscale_for[rcc->entry[i].pict_type] = q;
2589 step_mult = all_available_bits / expected_bits;
2592 for( double step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2595 rate_factor += step;
2597 rcc->last_non_b_pict_type = -1;
2598 rcc->last_accum_p_norm = 1;
2599 rcc->accum_p_norm = 0;
2602 for( int i = 0; i < rcc->num_entries; i++ )
2604 qscale[i] = get_qscale( h, &rcc->entry[i], rate_factor, i );
2605 rcc->last_qscale_for[rcc->entry[i].pict_type] = qscale[i];
2608 /* fixed I/B qscale relative to P */
2609 for( int i = rcc->num_entries-1; i >= 0; i-- )
2611 qscale[i] = get_diff_limited_q( h, &rcc->entry[i], qscale[i] );
2612 assert(qscale[i] >= 0);
2616 if( filter_size > 1 )
2618 assert( filter_size%2 == 1 );
2619 for( int i = 0; i < rcc->num_entries; i++ )
2621 ratecontrol_entry_t *rce = &rcc->entry[i];
2622 double q = 0.0, sum = 0.0;
2624 for( int j = 0; j < filter_size; j++ )
2626 int idx = i+j-filter_size/2;
2628 double coeff = qblur==0 ? 1.0 : exp( -d*d/(qblur*qblur) );
2629 if( idx < 0 || idx >= rcc->num_entries )
2631 if( rce->pict_type != rcc->entry[idx].pict_type )
2633 q += qscale[idx] * coeff;
2636 blurred_qscale[i] = q/sum;
2640 /* find expected bits */
2641 for( int i = 0; i < rcc->num_entries; i++ )
2643 ratecontrol_entry_t *rce = &rcc->entry[i];
2644 rce->new_qscale = clip_qscale( h, rce->pict_type, blurred_qscale[i] );
2645 assert(rce->new_qscale >= 0);
2646 expected_bits += qscale2bits( rce, rce->new_qscale );
2649 if( expected_bits > all_available_bits )
2650 rate_factor -= step;
2653 x264_free( qscale );
2654 if( filter_size > 1 )
2655 x264_free( blurred_qscale );
2658 if( vbv_pass2( h, all_available_bits ) )
2660 expected_bits = count_expected_bits( h );
2662 if( fabs( expected_bits/all_available_bits - 1.0 ) > 0.01 )
2665 for( int i = 0; i < rcc->num_entries; i++ )
2666 avgq += rcc->entry[i].new_qscale;
2667 avgq = qscale2qp( avgq / rcc->num_entries );
2669 if( expected_bits > all_available_bits || !rcc->b_vbv )
2670 x264_log( h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n" );
2671 x264_log( h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
2672 (float)h->param.rc.i_bitrate,
2673 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
2675 if( expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2 )
2677 if( h->param.rc.i_qp_min > 0 )
2678 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min );
2680 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate\n" );
2682 else if( expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2 )
2684 if( h->param.rc.i_qp_max < QP_MAX )
2685 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max );
2687 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate\n");
2689 else if( !(rcc->b_2pass && rcc->b_vbv) )
2690 x264_log( h, X264_LOG_WARNING, "internal error\n" );