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
33 #include "common/common.h"
34 #include "ratecontrol.h"
46 uint64_t expected_bits; /*total expected bits up to the current frame (current one excluded)*/
53 float blurred_complexity;
56 int16_t i_weight_denom[2];
60 int64_t i_cpb_duration;
61 } ratecontrol_entry_t;
71 struct x264_ratecontrol_t
80 double rate_tolerance;
82 int nmb; /* number of macroblocks in a frame */
86 ratecontrol_entry_t *rce;
87 int qp; /* qp for current frame */
88 float qpm; /* qp for current macroblock: precise float for AQ */
89 float qpa_rc; /* average of macroblocks' qp before aq */
90 float qpa_aq; /* average of macroblocks' qp after aq */
91 float qp_novbv; /* QP for the current frame if 1-pass VBV was disabled. */
95 int64_t buffer_fill_final;
96 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
97 double buffer_rate; /* # of bits added to buffer_fill after each frame */
98 double vbv_max_rate; /* # of bits added to buffer_fill per second */
99 predictor_t *pred; /* predict frame size from satd */
100 int single_frame_vbv;
101 double rate_factor_max_increment; /* Don't allow RF above (CRF + this value). */
106 double cplxr_sum; /* sum of bits*qscale/rceq */
107 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow, only includes finished frames */
108 int64_t filler_bits_sum; /* sum in bits of finished frames' filler data */
109 double wanted_bits_window; /* target bitrate * window */
111 double short_term_cplxsum;
112 double short_term_cplxcount;
113 double rate_factor_constant;
118 FILE *p_stat_file_out;
119 char *psz_stat_file_tmpname;
120 FILE *p_mbtree_stat_file_out;
121 char *psz_mbtree_stat_file_tmpname;
122 char *psz_mbtree_stat_file_name;
123 FILE *p_mbtree_stat_file_in;
125 int num_entries; /* number of ratecontrol_entry_ts */
126 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
128 double last_qscale_for[3]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
129 int last_non_b_pict_type;
130 double accum_p_qp; /* for determining I-frame quant */
132 double last_accum_p_norm;
133 double lmin[3]; /* min qscale by frame type */
135 double lstep; /* max change (multiply) in qscale per frame */
136 uint16_t *qp_buffer[2]; /* Global buffers for converting MB-tree quantizer data. */
137 int qpbuf_pos; /* In order to handle pyramid reordering, QP buffer acts as a stack.
138 * This value is the current position (0 or 1). */
141 float frame_size_estimated; /* Access to this variable must be atomic: double is
142 * not atomic on all arches we care about */
143 double frame_size_maximum; /* Maximum frame size due to MinCR */
144 double frame_size_planned;
145 double slice_size_planned;
146 predictor_t (*row_pred)[2];
147 predictor_t row_preds[3][2];
148 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
149 int bframes; /* # consecutive B-frames before this P-frame */
150 int bframe_bits; /* total cost of those frames */
154 x264_zone_t *prev_zone;
157 int initial_cpb_removal_delay;
158 int initial_cpb_removal_delay_offset;
159 double nrt_first_access_unit; /* nominal removal time */
160 double previous_cpb_final_arrival_time;
161 uint64_t hrd_multiply_denom;
165 static int parse_zones( x264_t *h );
166 static int init_pass2(x264_t *);
167 static float rate_estimate_qscale( x264_t *h );
168 static int update_vbv( x264_t *h, int bits );
169 static void update_vbv_plan( x264_t *h, int overhead );
170 static double predict_size( predictor_t *p, double q, double var );
171 static void update_predictor( predictor_t *p, double q, double var, double bits );
173 #define CMP_OPT_FIRST_PASS( opt, param_val )\
175 if( ( p = strstr( opts, opt "=" ) ) && sscanf( p, opt "=%d" , &i ) && param_val != i )\
177 x264_log( h, X264_LOG_ERROR, "different " opt " setting than first pass (%d vs %d)\n", param_val, i );\
183 * qp = h.264's quantizer
184 * qscale = linearized quantizer = Lagrange multiplier
186 static inline double qp2qscale( double qp )
188 return 0.85 * pow( 2.0, ( qp - 12.0 ) / 6.0 );
190 static inline double qscale2qp( double qscale )
192 return 12.0 + 6.0 * log2( qscale/0.85 );
195 /* Texture bitrate is not quite inversely proportional to qscale,
196 * probably due the the changing number of SKIP blocks.
197 * MV bits level off at about qp<=12, because the lambda used
198 * for motion estimation is constant there. */
199 static inline double qscale2bits( ratecontrol_entry_t *rce, double qscale )
203 return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
204 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
208 static ALWAYS_INLINE uint32_t ac_energy_var( uint64_t sum_ssd, int shift, x264_frame_t *frame, int i )
210 uint32_t sum = sum_ssd;
211 uint32_t ssd = sum_ssd >> 32;
212 frame->i_pixel_sum[i] += sum;
213 frame->i_pixel_ssd[i] += ssd;
214 return ssd - ((uint64_t)sum * sum >> shift);
217 static ALWAYS_INLINE uint32_t ac_energy_plane( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame, int i )
220 int stride = frame->i_stride[i];
221 int offset = h->mb.b_interlaced
222 ? 16 * mb_x + w * (mb_y&~1) * stride + (mb_y&1) * stride
223 : 16 * mb_x + w * mb_y * stride;
224 stride <<= h->mb.b_interlaced;
227 ALIGNED_ARRAY_16( pixel, pix,[FENC_STRIDE*8] );
228 h->mc.load_deinterleave_8x8x2_fenc( pix, frame->plane[1] + offset, stride );
229 return ac_energy_var( h->pixf.var[PIXEL_8x8]( pix, FENC_STRIDE ), 6, frame, 1 )
230 + ac_energy_var( h->pixf.var[PIXEL_8x8]( pix+FENC_STRIDE/2, FENC_STRIDE ), 6, frame, 2 );
233 return ac_energy_var( h->pixf.var[PIXEL_16x16]( frame->plane[0] + offset, stride ), 8, frame, 0 );
236 // Find the total AC energy of the block in all planes.
237 static NOINLINE uint32_t x264_ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
239 /* This function contains annoying hacks because GCC has a habit of reordering emms
240 * and putting it after floating point ops. As a result, we put the emms at the end of the
241 * function and make sure that its always called before the float math. Noinline makes
242 * sure no reordering goes on. */
243 uint32_t var = ac_energy_plane( h, mb_x, mb_y, frame, 0 );
244 var += ac_energy_plane( h, mb_x, mb_y, frame, 1 );
249 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame, float *quant_offsets )
251 /* constants chosen to result in approximately the same overall bitrate as without AQ.
252 * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
255 /* Initialize frame stats */
256 for( int i = 0; i < 3; i++ )
258 frame->i_pixel_sum[i] = 0;
259 frame->i_pixel_ssd[i] = 0;
262 /* Degenerate cases */
263 if( h->param.rc.i_aq_mode == X264_AQ_NONE || h->param.rc.f_aq_strength == 0 )
265 /* Need to init it anyways for MB tree */
266 if( h->param.rc.i_aq_mode && h->param.rc.f_aq_strength == 0 )
270 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
271 frame->f_qp_offset[mb_xy] = frame->f_qp_offset_aq[mb_xy] = quant_offsets[mb_xy];
272 if( h->frames.b_have_lowres )
273 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
274 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8( frame->f_qp_offset[mb_xy] );
278 memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
279 memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
280 if( h->frames.b_have_lowres )
281 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
282 frame->i_inv_qscale_factor[mb_xy] = 256;
285 /* Need variance data for weighted prediction */
286 if( h->param.analyse.i_weighted_pred )
288 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
289 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
290 x264_ac_energy_mb( h, mb_x, mb_y, frame );
295 /* Actual adaptive quantization */
298 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
300 float bit_depth_correction = powf(1 << (BIT_DEPTH-8), 0.5f);
301 float avg_adj_pow2 = 0.f;
302 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
303 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
305 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
306 float qp_adj = powf( energy + 1, 0.125f );
307 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
309 avg_adj_pow2 += qp_adj * qp_adj;
311 avg_adj /= h->mb.i_mb_count;
312 avg_adj_pow2 /= h->mb.i_mb_count;
313 strength = h->param.rc.f_aq_strength * avg_adj / bit_depth_correction;
314 avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - (14.f * bit_depth_correction)) / avg_adj;
317 strength = h->param.rc.f_aq_strength * 1.0397f;
319 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
320 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
323 int mb_xy = mb_x + mb_y*h->mb.i_mb_stride;
324 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
326 qp_adj = frame->f_qp_offset[mb_xy];
327 qp_adj = strength * (qp_adj - avg_adj);
331 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
332 qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - (14.427f + 2*(BIT_DEPTH-8)));
335 qp_adj += quant_offsets[mb_xy];
336 frame->f_qp_offset[mb_xy] =
337 frame->f_qp_offset_aq[mb_xy] = qp_adj;
338 if( h->frames.b_have_lowres )
339 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8(qp_adj);
343 /* Remove mean from SSD calculation */
344 for( int i = 0; i < 3; i++ )
346 uint64_t ssd = frame->i_pixel_ssd[i];
347 uint64_t sum = frame->i_pixel_sum[i];
348 int width = h->mb.i_mb_width*16>>!!i;
349 int height = h->mb.i_mb_height*16>>!!i;
350 frame->i_pixel_ssd[i] = ssd - (sum * sum + width * height / 2) / (width * height);
354 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame, float *quant_offsets )
356 x264_ratecontrol_t *rc = h->rc;
357 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
359 if( rc->entry[frame->i_frame].kept_as_ref )
362 if( rc->qpbuf_pos < 0 )
368 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
370 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 )
373 if( i_type != i_type_actual && rc->qpbuf_pos == 1 )
375 x264_log(h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type, i_type_actual);
378 } while( i_type != i_type_actual );
381 for( int i = 0; i < h->mb.i_mb_count; i++ )
383 frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[rc->qpbuf_pos][i] )) * (1/256.0);
384 if( h->frames.b_have_lowres )
385 frame->i_inv_qscale_factor[i] = x264_exp2fix8(frame->f_qp_offset[i]);
390 x264_stack_align( x264_adaptive_quant_frame, h, frame, quant_offsets );
393 x264_log(h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n");
397 int x264_reference_build_list_optimal( x264_t *h )
399 ratecontrol_entry_t *rce = h->rc->rce;
400 x264_frame_t *frames[16];
401 x264_weight_t weights[16][3];
404 if( rce->refs != h->i_ref[0] )
407 memcpy( frames, h->fref[0], sizeof(frames) );
408 memcpy( refcount, rce->refcount, sizeof(refcount) );
409 memcpy( weights, h->fenc->weight, sizeof(weights) );
410 memset( &h->fenc->weight[1][0], 0, sizeof(x264_weight_t[15][3]) );
412 /* For now don't reorder ref 0; it seems to lower quality
413 in most cases due to skips. */
414 for( int ref = 1; ref < h->i_ref[0]; ref++ )
419 for( int i = 1; i < h->i_ref[0]; i++ )
420 /* Favor lower POC as a tiebreaker. */
421 COPY2_IF_GT( max, refcount[i], bestref, i );
423 /* FIXME: If there are duplicates from frames other than ref0 then it is possible
424 * that the optimal ordering doesnt place every duplicate. */
426 refcount[bestref] = -1;
427 h->fref[0][ref] = frames[bestref];
428 memcpy( h->fenc->weight[ref], weights[bestref], sizeof(weights[bestref]) );
434 static char *x264_strcat_filename( char *input, char *suffix )
436 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
439 strcpy( output, input );
440 strcat( output, suffix );
444 void x264_ratecontrol_init_reconfigurable( x264_t *h, int b_init )
446 x264_ratecontrol_t *rc = h->rc;
447 if( !b_init && rc->b_2pass )
450 if( h->param.rc.i_rc_method == X264_RC_CRF )
452 /* Arbitrary rescaling to make CRF somewhat similar to QP.
453 * Try to compensate for MB-tree's effects as well. */
454 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
455 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
456 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
457 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset + QP_BD_OFFSET );
460 if( h->param.rc.i_vbv_max_bitrate > 0 && h->param.rc.i_vbv_buffer_size > 0 )
462 if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
464 h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
465 x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
466 h->param.rc.i_vbv_buffer_size );
469 /* We don't support changing the ABR bitrate right now,
470 so if the stream starts as CBR, keep it CBR. */
471 if( rc->b_vbv_min_rate )
472 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
474 int vbv_buffer_size = h->param.rc.i_vbv_buffer_size * 1000;
475 int vbv_max_bitrate = h->param.rc.i_vbv_max_bitrate * 1000;
478 h->sps->vui.hrd.i_bit_rate_unscaled = vbv_max_bitrate;
479 h->sps->vui.hrd.i_cpb_size_unscaled = vbv_buffer_size;
480 if( h->param.i_nal_hrd && b_init )
482 h->sps->vui.hrd.i_cpb_cnt = 1;
483 h->sps->vui.hrd.b_cbr_hrd = h->param.i_nal_hrd == X264_NAL_HRD_CBR;
484 h->sps->vui.hrd.i_time_offset_length = 0;
489 int bitrate = 1000*h->param.rc.i_vbv_max_bitrate;
490 int bufsize = 1000*h->param.rc.i_vbv_buffer_size;
492 // normalize HRD size and rate to the value / scale notation
493 h->sps->vui.hrd.i_bit_rate_scale = x264_clip3( x264_ctz( bitrate ) - BR_SHIFT, 0, 15 );
494 h->sps->vui.hrd.i_bit_rate_value = bitrate >> ( h->sps->vui.hrd.i_bit_rate_scale + BR_SHIFT );
495 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 );
496 h->sps->vui.hrd.i_cpb_size_scale = x264_clip3( x264_ctz( bufsize ) - CPB_SHIFT, 0, 15 );
497 h->sps->vui.hrd.i_cpb_size_value = bufsize >> ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
498 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 );
504 #define MAX_DURATION 0.5
506 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 );
507 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;
508 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);
510 h->sps->vui.hrd.i_initial_cpb_removal_delay_length = 2 + x264_clip3( 32 - x264_clz( max_delay ), 4, 22 );
511 h->sps->vui.hrd.i_cpb_removal_delay_length = x264_clip3( 32 - x264_clz( max_cpb_output_delay ), 4, 31 );
512 h->sps->vui.hrd.i_dpb_output_delay_length = x264_clip3( 32 - x264_clz( max_dpb_output_delay ), 4, 31 );
516 vbv_buffer_size = h->sps->vui.hrd.i_cpb_size_unscaled;
517 vbv_max_bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
519 else if( h->param.i_nal_hrd && !b_init )
521 x264_log( h, X264_LOG_WARNING, "VBV parameters cannot be changed when NAL HRD is in use\n" );
525 rc->buffer_rate = vbv_max_bitrate / rc->fps;
526 rc->vbv_max_rate = vbv_max_bitrate;
527 rc->buffer_size = vbv_buffer_size;
528 rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
529 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
530 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
531 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.f_rf_constant_max )
533 rc->rate_factor_max_increment = h->param.rc.f_rf_constant_max - h->param.rc.f_rf_constant;
534 if( rc->rate_factor_max_increment <= 0 )
536 x264_log( h, X264_LOG_WARNING, "CRF max must be greater than CRF\n" );
537 rc->rate_factor_max_increment = 0;
542 if( h->param.rc.f_vbv_buffer_init > 1. )
543 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 );
544 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);
545 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init * h->sps->vui.i_time_scale;
547 rc->b_vbv_min_rate = !rc->b_2pass
548 && h->param.rc.i_rc_method == X264_RC_ABR
549 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
554 int x264_ratecontrol_new( x264_t *h )
556 x264_ratecontrol_t *rc;
560 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
563 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
564 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
566 /* FIXME: use integers */
567 if( h->param.i_fps_num > 0 && h->param.i_fps_den > 0 )
568 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
572 if( h->param.rc.b_mb_tree )
574 h->param.rc.f_pb_factor = 1;
578 rc->qcompress = h->param.rc.f_qcompress;
580 rc->bitrate = h->param.rc.i_bitrate * 1000.;
581 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
582 rc->nmb = h->mb.i_mb_count;
583 rc->last_non_b_pict_type = -1;
586 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
588 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
592 x264_ratecontrol_init_reconfigurable( h, 1 );
594 if( h->param.i_nal_hrd )
596 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale;
597 uint64_t num = 180000;
598 x264_reduce_fraction64( &num, &denom );
599 rc->hrd_multiply_denom = 180000 / num;
601 double bits_required = log2( 180000 / rc->hrd_multiply_denom )
602 + log2( h->sps->vui.i_time_scale )
603 + log2( h->sps->vui.hrd.i_cpb_size_unscaled );
604 if( bits_required >= 63 )
606 x264_log( h, X264_LOG_ERROR, "HRD with very large timescale and bufsize not supported\n" );
611 if( rc->rate_tolerance < 0.01 )
613 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
614 rc->rate_tolerance = 0.01;
617 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
621 /* FIXME ABR_INIT_QP is actually used only in CRF */
622 #define ABR_INIT_QP (( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 ) + QP_BD_OFFSET)
623 rc->accum_p_norm = .01;
624 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
625 /* estimated ratio that produces a reasonable QP for the first I-frame */
626 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
627 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
628 rc->last_non_b_pict_type = SLICE_TYPE_I;
631 rc->ip_offset = 6.0 * log2f( h->param.rc.f_ip_factor );
632 rc->pb_offset = 6.0 * log2f( h->param.rc.f_pb_factor );
633 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
634 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, QP_MAX );
635 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, QP_MAX );
636 h->mb.ip_offset = rc->ip_offset + 0.5;
638 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
639 rc->last_qscale = qp2qscale( 26 );
640 int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
641 CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
642 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
643 for( int i = 0; i < 3; i++ )
645 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
646 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
647 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
648 for( int j = 0; j < num_preds; j++ )
650 rc->pred[i+j*5].coeff= 2.0;
651 rc->pred[i+j*5].count= 1.0;
652 rc->pred[i+j*5].decay= 0.5;
653 rc->pred[i+j*5].offset= 0.0;
655 for( int j = 0; j < 2; j++ )
657 rc->row_preds[i][j].coeff= .25;
658 rc->row_preds[i][j].count= 1.0;
659 rc->row_preds[i][j].decay= 0.5;
660 rc->row_preds[i][j].offset= 0.0;
663 *rc->pred_b_from_p = rc->pred[0];
665 if( parse_zones( h ) < 0 )
667 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
671 /* Load stat file and init 2pass algo */
672 if( h->param.rc.b_stat_read )
674 char *p, *stats_in, *stats_buf;
676 /* read 1st pass stats */
677 assert( h->param.rc.psz_stat_in );
678 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
681 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
684 if( h->param.rc.b_mb_tree )
686 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
687 if( !mbtree_stats_in )
689 rc->p_mbtree_stat_file_in = fopen( mbtree_stats_in, "rb" );
690 x264_free( mbtree_stats_in );
691 if( !rc->p_mbtree_stat_file_in )
693 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
698 /* check whether 1st pass options were compatible with current options */
699 if( !strncmp( stats_buf, "#options:", 9 ) )
703 char *opts = stats_buf;
704 stats_in = strchr( stats_buf, '\n' );
709 if( sscanf( opts, "#options: %dx%d", &i, &j ) != 2 )
711 x264_log( h, X264_LOG_ERROR, "resolution specified in stats file not valid\n" );
714 else if( h->param.rc.b_mb_tree && (i != h->param.i_width || j != h->param.i_height) )
716 x264_log( h, X264_LOG_ERROR, "MB-tree doesn't support different resolution than 1st pass (%dx%d vs %dx%d)\n",
717 h->param.i_width, h->param.i_height, i, j );
721 if( ( p = strstr( opts, "timebase=" ) ) && sscanf( p, "timebase=%u/%u", &k, &l ) != 2 )
723 x264_log( h, X264_LOG_ERROR, "timebase specified in stats file not valid\n" );
726 if( k != h->param.i_timebase_num || l != h->param.i_timebase_den )
728 x264_log( h, X264_LOG_ERROR, "timebase mismatch with 1st pass (%u/%u vs %u/%u)\n",
729 h->param.i_timebase_num, h->param.i_timebase_den, k, l );
733 CMP_OPT_FIRST_PASS( "bitdepth", BIT_DEPTH );
734 CMP_OPT_FIRST_PASS( "weightp", X264_MAX( 0, h->param.analyse.i_weighted_pred ) );
735 CMP_OPT_FIRST_PASS( "bframes", h->param.i_bframe );
736 CMP_OPT_FIRST_PASS( "b_pyramid", h->param.i_bframe_pyramid );
737 CMP_OPT_FIRST_PASS( "intra_refresh", h->param.b_intra_refresh );
738 CMP_OPT_FIRST_PASS( "open_gop", h->param.i_open_gop );
740 if( (p = strstr( opts, "keyint=" )) )
743 char buf[13] = "infinite ";
744 if( h->param.i_keyint_max != X264_KEYINT_MAX_INFINITE )
745 sprintf( buf, "%d ", h->param.i_keyint_max );
746 if( strncmp( p, buf, strlen(buf) ) )
748 x264_log( h, X264_LOG_ERROR, "different keyint setting than first pass (%.*s vs %.*s)\n",
749 strlen(buf)-1, buf, strcspn(p, " "), p );
754 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
755 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
757 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
759 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
760 h->mb.b_direct_auto_write = 1;
763 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
764 h->param.i_bframe_adaptive = i;
765 else if( h->param.i_bframe )
767 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
771 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 ) )
772 h->param.rc.i_lookahead = i;
775 /* find number of pics */
778 for( num_entries = -1; p; num_entries++ )
779 p = strchr( p + 1, ';' );
782 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
785 rc->num_entries = num_entries;
787 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
789 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
790 h->param.i_frame_total, rc->num_entries );
792 if( h->param.i_frame_total > rc->num_entries )
794 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
795 h->param.i_frame_total, rc->num_entries );
799 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
801 /* init all to skipped p frames */
802 for( int i = 0; i < rc->num_entries; i++ )
804 ratecontrol_entry_t *rce = &rc->entry[i];
805 rce->pict_type = SLICE_TYPE_P;
806 rce->qscale = rce->new_qscale = qp2qscale( 20 );
807 rce->misc_bits = rc->nmb + 10;
813 for( int i = 0; i < rc->num_entries; i++ )
815 ratecontrol_entry_t *rce;
823 next= strchr(p, ';');
825 *next++ = 0; //sscanf is unbelievably slow on long strings
826 e = sscanf( p, " in:%d ", &frame_number );
828 if( frame_number < 0 || frame_number >= rc->num_entries )
830 x264_log( h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i );
833 rce = &rc->entry[frame_number];
834 rce->direct_mode = 0;
836 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",
837 &pict_type, &rce->i_duration, &rce->i_cpb_duration, &qp, &rce->tex_bits,
838 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
839 &rce->s_count, &rce->direct_mode );
841 p = strstr( p, "ref:" );
845 for( ref = 0; ref < 16; ref++ )
847 if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
849 p = strchr( p+1, ' ' );
856 rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
857 char *w = strchr( p, 'w' );
860 int count = sscanf( w, "w:%hd,%hd,%hd,%hd,%hd,%hd,%hd,%hd",
861 &rce->i_weight_denom[0], &rce->weight[0][0], &rce->weight[0][1],
862 &rce->i_weight_denom[1], &rce->weight[1][0], &rce->weight[1][1],
863 &rce->weight[2][0], &rce->weight[2][1] );
865 rce->i_weight_denom[1] = -1;
866 else if ( count != 8 )
867 rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
870 if( pict_type != 'b' )
871 rce->kept_as_ref = 1;
875 rce->frame_type = X264_TYPE_IDR;
876 rce->pict_type = SLICE_TYPE_I;
879 rce->frame_type = X264_TYPE_I;
880 rce->pict_type = SLICE_TYPE_I;
883 rce->frame_type = X264_TYPE_P;
884 rce->pict_type = SLICE_TYPE_P;
887 rce->frame_type = X264_TYPE_BREF;
888 rce->pict_type = SLICE_TYPE_B;
891 rce->frame_type = X264_TYPE_B;
892 rce->pict_type = SLICE_TYPE_B;
894 default: e = -1; break;
899 x264_log( h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e );
902 rce->qscale = qp2qscale( qp );
906 x264_free( stats_buf );
908 if( h->param.rc.i_rc_method == X264_RC_ABR )
910 if( init_pass2( h ) < 0 )
912 } /* else we're using constant quant, so no need to run the bitrate allocation */
915 /* Open output file */
916 /* If input and output files are the same, output to a temp file
917 * and move it to the real name only when it's complete */
918 if( h->param.rc.b_stat_write )
921 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
922 if( !rc->psz_stat_file_tmpname )
925 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
926 if( rc->p_stat_file_out == NULL )
928 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
932 p = x264_param2string( &h->param, 1 );
934 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
936 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
938 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
939 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
940 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
943 rc->p_mbtree_stat_file_out = fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
944 if( rc->p_mbtree_stat_file_out == NULL )
946 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
952 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
954 CHECKED_MALLOC( rc->qp_buffer[0], h->mb.i_mb_count * sizeof(uint16_t) );
955 if( h->param.i_bframe_pyramid && h->param.rc.b_stat_read )
956 CHECKED_MALLOC( rc->qp_buffer[1], h->mb.i_mb_count * sizeof(uint16_t) );
960 for( int i = 0; i<h->param.i_threads; i++ )
962 h->thread[i]->rc = rc+i;
966 h->thread[i]->param = h->param;
967 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
968 h->thread[i]->mb.ip_offset = h->mb.ip_offset;
977 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
980 char *tok, UNUSED *saveptr=NULL;
982 z->f_bitrate_factor = 1;
983 if( 3 <= sscanf(p, "%d,%d,q=%d%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
985 else if( 3 <= sscanf(p, "%d,%d,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
987 else if( 2 <= sscanf(p, "%d,%d%n", &z->i_start, &z->i_end, &len) )
991 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
997 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
998 memcpy( z->param, &h->param, sizeof(x264_param_t) );
999 z->param->param_free = x264_free;
1000 while( (tok = strtok_r( p, ",", &saveptr )) )
1002 char *val = strchr( tok, '=' );
1008 if( x264_param_parse( z->param, tok, val ) )
1010 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
1020 static int parse_zones( x264_t *h )
1022 x264_ratecontrol_t *rc = h->rc;
1023 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
1025 char *psz_zones, *p;
1026 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
1027 strcpy( psz_zones, h->param.rc.psz_zones );
1028 h->param.rc.i_zones = 1;
1029 for( p = psz_zones; *p; p++ )
1030 h->param.rc.i_zones += (*p == '/');
1031 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
1033 for( int i = 0; i < h->param.rc.i_zones; i++ )
1035 int i_tok = strcspn( p, "/" );
1037 if( parse_zone( h, &h->param.rc.zones[i], p ) )
1041 x264_free( psz_zones );
1044 if( h->param.rc.i_zones > 0 )
1046 for( int i = 0; i < h->param.rc.i_zones; i++ )
1048 x264_zone_t z = h->param.rc.zones[i];
1049 if( z.i_start < 0 || z.i_start > z.i_end )
1051 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
1052 z.i_start, z.i_end );
1055 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
1057 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
1058 z.f_bitrate_factor );
1063 rc->i_zones = h->param.rc.i_zones + 1;
1064 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
1065 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
1067 // default zone to fall back to if none of the others match
1068 rc->zones[0].i_start = 0;
1069 rc->zones[0].i_end = INT_MAX;
1070 rc->zones[0].b_force_qp = 0;
1071 rc->zones[0].f_bitrate_factor = 1;
1072 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
1073 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
1074 for( int i = 1; i < rc->i_zones; i++ )
1076 if( !rc->zones[i].param )
1077 rc->zones[i].param = rc->zones[0].param;
1086 static x264_zone_t *get_zone( x264_t *h, int frame_num )
1088 for( int i = h->rc->i_zones - 1; i >= 0; i-- )
1090 x264_zone_t *z = &h->rc->zones[i];
1091 if( frame_num >= z->i_start && frame_num <= z->i_end )
1097 void x264_ratecontrol_summary( x264_t *h )
1099 x264_ratecontrol_t *rc = h->rc;
1100 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
1102 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
1103 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
1104 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
1105 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
1106 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset - QP_BD_OFFSET );
1110 void x264_ratecontrol_delete( x264_t *h )
1112 x264_ratecontrol_t *rc = h->rc;
1115 if( rc->p_stat_file_out )
1117 b_regular_file = x264_is_regular_file( rc->p_stat_file_out );
1118 fclose( rc->p_stat_file_out );
1119 if( h->i_frame >= rc->num_entries && b_regular_file )
1120 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
1122 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1123 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
1125 x264_free( rc->psz_stat_file_tmpname );
1127 if( rc->p_mbtree_stat_file_out )
1129 b_regular_file = x264_is_regular_file( rc->p_mbtree_stat_file_out );
1130 fclose( rc->p_mbtree_stat_file_out );
1131 if( h->i_frame >= rc->num_entries && b_regular_file )
1132 if( rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
1134 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1135 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
1137 x264_free( rc->psz_mbtree_stat_file_tmpname );
1138 x264_free( rc->psz_mbtree_stat_file_name );
1140 if( rc->p_mbtree_stat_file_in )
1141 fclose( rc->p_mbtree_stat_file_in );
1142 x264_free( rc->pred );
1143 x264_free( rc->pred_b_from_p );
1144 x264_free( rc->entry );
1145 x264_free( rc->qp_buffer[0] );
1146 x264_free( rc->qp_buffer[1] );
1149 x264_free( rc->zones[0].param );
1150 for( int i = 1; i < rc->i_zones; i++ )
1151 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
1152 rc->zones[i].param->param_free( rc->zones[i].param );
1153 x264_free( rc->zones );
1158 static void accum_p_qp_update( x264_t *h, float qp )
1160 x264_ratecontrol_t *rc = h->rc;
1161 rc->accum_p_qp *= .95;
1162 rc->accum_p_norm *= .95;
1163 rc->accum_p_norm += 1;
1164 if( h->sh.i_type == SLICE_TYPE_I )
1165 rc->accum_p_qp += qp + rc->ip_offset;
1167 rc->accum_p_qp += qp;
1170 /* Before encoding a frame, choose a QP for it */
1171 void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
1173 x264_ratecontrol_t *rc = h->rc;
1174 ratecontrol_entry_t *rce = NULL;
1175 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
1180 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
1181 x264_encoder_reconfig( h, zone->param );
1182 rc->prev_zone = zone;
1184 if( h->param.rc.b_stat_read )
1186 int frame = h->fenc->i_frame;
1187 assert( frame >= 0 && frame < rc->num_entries );
1188 rce = h->rc->rce = &h->rc->entry[frame];
1190 if( h->sh.i_type == SLICE_TYPE_B
1191 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
1193 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
1194 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
1200 memset( h->fdec->i_row_bits, 0, h->mb.i_mb_height * sizeof(int) );
1201 rc->row_pred = &rc->row_preds[h->sh.i_type];
1202 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;
1203 update_vbv_plan( h, overhead );
1205 const x264_level_t *l = x264_levels;
1206 while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
1209 int mincr = l->mincr;
1211 /* Blu-ray requires this */
1212 if( l->level_idc == 41 && h->param.i_nal_hrd )
1215 /* High 10 doesn't require minCR, so just set the maximum to a large value. */
1216 if( h->sps->i_profile_idc == PROFILE_HIGH10 )
1217 rc->frame_size_maximum = 1e9;
1220 /* The spec has a bizarre special case for the first frame. */
1221 if( h->i_frame == 0 )
1223 //384 * ( Max( PicSizeInMbs, fR * MaxMBPS ) + MaxMBPS * ( tr( 0 ) - tr,n( 0 ) ) ) / MinCR
1224 double fr = 1. / 172;
1225 int pic_size_in_mbs = h->mb.i_mb_width * h->mb.i_mb_height;
1226 rc->frame_size_maximum = 384 * BIT_DEPTH * X264_MAX( pic_size_in_mbs, fr*l->mbps ) / mincr;
1230 //384 * MaxMBPS * ( tr( n ) - tr( n - 1 ) ) / MinCR
1231 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;
1236 if( h->sh.i_type != SLICE_TYPE_B )
1237 rc->bframes = h->fenc->i_bframes;
1239 if( i_force_qp != X264_QP_AUTO )
1243 else if( rc->b_abr )
1245 q = qscale2qp( rate_estimate_qscale( h ) );
1247 else if( rc->b_2pass )
1249 rce->new_qscale = rate_estimate_qscale( h );
1250 q = qscale2qp( rce->new_qscale );
1254 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
1255 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
1257 q = rc->qp_constant[ h->sh.i_type ];
1261 if( zone->b_force_qp )
1262 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1264 q -= 6*log2f( zone->f_bitrate_factor );
1268 q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1272 rc->qp = x264_clip3( (int)(q + 0.5), 0, QP_MAX );
1273 h->fdec->f_qp_avg_rc =
1274 h->fdec->f_qp_avg_aq =
1277 rce->new_qp = rc->qp;
1279 accum_p_qp_update( h, rc->qpm );
1281 if( h->sh.i_type != SLICE_TYPE_B )
1282 rc->last_non_b_pict_type = h->sh.i_type;
1285 static double predict_row_size( x264_t *h, int y, double qp )
1287 /* average between two predictors:
1288 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1289 x264_ratecontrol_t *rc = h->rc;
1290 double pred_s = predict_size( rc->row_pred[0], qp2qscale( qp ), h->fdec->i_row_satd[y] );
1292 if( h->sh.i_type == SLICE_TYPE_I || qp >= h->fref[0][0]->f_row_qp[y] )
1294 if( h->sh.i_type == SLICE_TYPE_P
1295 && h->fref[0][0]->i_type == h->fdec->i_type
1296 && h->fref[0][0]->i_row_satd[y] > 0
1297 && (abs(h->fref[0][0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1299 pred_t = h->fref[0][0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref[0][0]->i_row_satd[y]
1300 * qp2qscale( h->fref[0][0]->f_row_qp[y] ) / qp2qscale( qp );
1304 return (pred_s + pred_t) / 2;
1306 /* Our QP is lower than the reference! */
1309 double pred_intra = predict_size( rc->row_pred[1], qp2qscale( qp ), h->fdec->i_row_satds[0][0][y] );
1310 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1311 return pred_intra + pred_s;
1315 static double row_bits_so_far( x264_t *h, int y )
1318 for( int i = h->i_threadslice_start; i <= y; i++ )
1319 bits += h->fdec->i_row_bits[i];
1323 static double predict_row_size_sum( x264_t *h, int y, double qp )
1325 double bits = row_bits_so_far(h, y);
1326 for( int i = y+1; i < h->i_threadslice_end; i++ )
1327 bits += predict_row_size( h, i, qp );
1332 void x264_ratecontrol_mb( x264_t *h, int bits )
1334 x264_ratecontrol_t *rc = h->rc;
1335 const int y = h->mb.i_mb_y;
1339 h->fdec->i_row_bits[y] += bits;
1340 rc->qpa_rc += rc->qpm;
1341 rc->qpa_aq += h->mb.i_qp;
1343 if( h->mb.i_mb_x != h->mb.i_mb_width - 1 || !rc->b_vbv )
1346 h->fdec->f_row_qp[y] = rc->qpm;
1348 update_predictor( rc->row_pred[0], qp2qscale( rc->qpm ), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1349 if( h->sh.i_type == SLICE_TYPE_P && rc->qpm < h->fref[0][0]->f_row_qp[y] )
1350 update_predictor( rc->row_pred[1], qp2qscale( rc->qpm ), h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1352 /* tweak quality based on difference from predicted size */
1353 if( y < h->i_threadslice_end-1 )
1355 float prev_row_qp = h->fdec->f_row_qp[y];
1356 float qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1357 float qp_absolute_max = h->param.rc.i_qp_max;
1358 if( rc->rate_factor_max_increment )
1359 qp_absolute_max = X264_MIN( qp_absolute_max, rc->qp_novbv + rc->rate_factor_max_increment );
1360 float qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, qp_absolute_max );
1361 float step_size = 0.5;
1363 /* B-frames shouldn't use lower QP than their reference frames. */
1364 if( h->sh.i_type == SLICE_TYPE_B )
1366 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] ) );
1367 rc->qpm = X264_MAX( rc->qpm, qp_min );
1370 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1371 float slice_size_planned = h->param.b_sliced_threads ? rc->slice_size_planned : rc->frame_size_planned;
1372 float max_frame_error = X264_MAX( 0.05, 1.0 / (h->mb.i_mb_height) );
1373 float size_of_other_slices = 0;
1374 if( h->param.b_sliced_threads )
1376 float size_of_other_slices_planned = 0;
1377 for( int i = 0; i < h->param.i_threads; i++ )
1378 if( h != h->thread[i] )
1380 size_of_other_slices += h->thread[i]->rc->frame_size_estimated;
1381 size_of_other_slices_planned += h->thread[i]->rc->slice_size_planned;
1383 float weight = rc->slice_size_planned / rc->frame_size_planned;
1384 size_of_other_slices = (size_of_other_slices - size_of_other_slices_planned) * weight + size_of_other_slices_planned;
1387 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1388 float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
1389 int b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1391 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1392 /* area at the top of the frame was measured inaccurately. */
1393 if( row_bits_so_far( h, y ) < 0.05 * slice_size_planned )
1396 if( h->sh.i_type != SLICE_TYPE_I )
1399 if( !rc->b_vbv_min_rate )
1400 qp_min = X264_MAX( qp_min, rc->qp_novbv );
1402 while( rc->qpm < qp_max
1403 && ((b1 > rc->frame_size_planned + rc_tol) ||
1404 (rc->buffer_fill - b1 < buffer_left_planned * 0.5) ||
1405 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1407 rc->qpm += step_size;
1408 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1411 while( rc->qpm > qp_min
1412 && (rc->qpm > h->fdec->f_row_qp[0] || rc->single_frame_vbv)
1413 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
1414 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
1416 rc->qpm -= step_size;
1417 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1420 /* avoid VBV underflow or MinCR violation */
1421 while( (rc->qpm < qp_absolute_max)
1422 && ((rc->buffer_fill - b1 < rc->buffer_rate * max_frame_error) ||
1423 (rc->frame_size_maximum - b1 < rc->frame_size_maximum * max_frame_error)))
1425 rc->qpm += step_size;
1426 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1429 h->rc->frame_size_estimated = predict_row_size_sum( h, y, rc->qpm );
1433 int x264_ratecontrol_qp( x264_t *h )
1436 return x264_clip3( h->rc->qpm + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1439 int x264_ratecontrol_mb_qp( x264_t *h )
1442 float qp = h->rc->qpm;
1443 if( h->param.rc.i_aq_mode )
1445 /* MB-tree currently doesn't adjust quantizers in unreferenced frames. */
1446 float qp_offset = h->fdec->b_kept_as_ref ? h->fenc->f_qp_offset[h->mb.i_mb_xy] : h->fenc->f_qp_offset_aq[h->mb.i_mb_xy];
1447 /* Scale AQ's effect towards zero in emergency mode. */
1448 if( qp > QP_MAX_SPEC )
1449 qp_offset *= (QP_MAX - qp) / (QP_MAX - QP_MAX_SPEC);
1452 return x264_clip3( qp + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1455 /* In 2pass, force the same frame types as in the 1st pass */
1456 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1458 x264_ratecontrol_t *rc = h->rc;
1459 if( h->param.rc.b_stat_read )
1461 if( frame_num >= rc->num_entries )
1463 /* We could try to initialize everything required for ABR and
1464 * adaptive B-frames, but that would be complicated.
1465 * So just calculate the average QP used so far. */
1466 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24 + QP_BD_OFFSET
1467 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1468 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, QP_MAX );
1469 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 );
1470 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 );
1472 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1473 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1474 if( h->param.i_bframe_adaptive )
1475 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1477 for( int i = 0; i < h->param.i_threads; i++ )
1479 h->thread[i]->rc->b_abr = 0;
1480 h->thread[i]->rc->b_2pass = 0;
1481 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1482 h->thread[i]->param.rc.b_stat_read = 0;
1483 h->thread[i]->param.i_bframe_adaptive = 0;
1484 h->thread[i]->param.i_scenecut_threshold = 0;
1485 h->thread[i]->param.rc.b_mb_tree = 0;
1486 if( h->thread[i]->param.i_bframe > 1 )
1487 h->thread[i]->param.i_bframe = 1;
1489 return X264_TYPE_AUTO;
1491 return rc->entry[frame_num].frame_type;
1494 return X264_TYPE_AUTO;
1497 void x264_ratecontrol_set_weights( x264_t *h, x264_frame_t *frm )
1499 ratecontrol_entry_t *rce = &h->rc->entry[frm->i_frame];
1500 if( h->param.analyse.i_weighted_pred <= 0 )
1503 if( rce->i_weight_denom[0] >= 0 )
1504 SET_WEIGHT( frm->weight[0][0], 1, rce->weight[0][0], rce->i_weight_denom[0], rce->weight[0][1] );
1506 if( rce->i_weight_denom[1] >= 0 )
1508 SET_WEIGHT( frm->weight[0][1], 1, rce->weight[1][0], rce->i_weight_denom[1], rce->weight[1][1] );
1509 SET_WEIGHT( frm->weight[0][2], 1, rce->weight[2][0], rce->i_weight_denom[1], rce->weight[2][1] );
1513 /* After encoding one frame, save stats and update ratecontrol state */
1514 int x264_ratecontrol_end( x264_t *h, int bits, int *filler )
1516 x264_ratecontrol_t *rc = h->rc;
1517 const int *mbs = h->stat.frame.i_mb_count;
1521 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1522 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1523 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1524 for( int i = B_DIRECT; i < B_8x8; i++ )
1525 h->stat.frame.i_mb_count_p += mbs[i];
1527 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1528 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1530 if( h->param.rc.b_stat_write )
1532 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1533 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1534 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1535 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1536 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1537 char c_direct = h->mb.b_direct_auto_write ?
1538 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1539 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1541 if( fprintf( rc->p_stat_file_out,
1542 "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:",
1543 h->fenc->i_frame, h->i_frame,
1544 c_type, h->fenc->i_duration,
1545 h->fenc->i_cpb_duration, rc->qpa_rc,
1546 h->stat.frame.i_tex_bits,
1547 h->stat.frame.i_mv_bits,
1548 h->stat.frame.i_misc_bits,
1549 h->stat.frame.i_mb_count_i,
1550 h->stat.frame.i_mb_count_p,
1551 h->stat.frame.i_mb_count_skip,
1555 /* Only write information for reference reordering once. */
1556 int use_old_stats = h->param.rc.b_stat_read && rc->rce->refs > 1;
1557 for( int i = 0; i < (use_old_stats ? rc->rce->refs : h->i_ref[0]); i++ )
1559 int refcount = use_old_stats ? rc->rce->refcount[i]
1560 : h->param.b_interlaced ? h->stat.frame.i_mb_count_ref[0][i*2]
1561 + h->stat.frame.i_mb_count_ref[0][i*2+1]
1562 : h->stat.frame.i_mb_count_ref[0][i];
1563 if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1567 if( h->param.analyse.i_weighted_pred >= X264_WEIGHTP_SIMPLE && h->sh.weight[0][0].weightfn )
1569 if( fprintf( rc->p_stat_file_out, "w:%d,%d,%d",
1570 h->sh.weight[0][0].i_denom, h->sh.weight[0][0].i_scale, h->sh.weight[0][0].i_offset ) < 0 )
1572 if( h->sh.weight[0][1].weightfn || h->sh.weight[0][2].weightfn )
1574 if( fprintf( rc->p_stat_file_out, ",%d,%d,%d,%d,%d ",
1575 h->sh.weight[0][1].i_denom, h->sh.weight[0][1].i_scale, h->sh.weight[0][1].i_offset,
1576 h->sh.weight[0][2].i_scale, h->sh.weight[0][2].i_offset ) < 0 )
1579 else if( fprintf( rc->p_stat_file_out, " " ) < 0 )
1583 if( fprintf( rc->p_stat_file_out, ";\n") < 0 )
1586 /* Don't re-write the data in multi-pass mode. */
1587 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1589 uint8_t i_type = h->sh.i_type;
1590 /* Values are stored as big-endian FIX8.8 */
1591 for( int i = 0; i < h->mb.i_mb_count; i++ )
1592 rc->qp_buffer[0][i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1593 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1595 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 )
1602 if( h->sh.i_type != SLICE_TYPE_B )
1603 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / rc->last_rceq;
1606 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1607 * Not perfectly accurate with B-refs, but good enough. */
1608 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / (rc->last_rceq * fabs( h->param.rc.f_pb_factor ));
1610 rc->cplxr_sum *= rc->cbr_decay;
1611 rc->wanted_bits_window += h->fenc->f_duration * rc->bitrate;
1612 rc->wanted_bits_window *= rc->cbr_decay;
1616 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale( rc->rce->new_qp ) );
1618 if( h->mb.b_variable_qp )
1620 if( h->sh.i_type == SLICE_TYPE_B )
1622 rc->bframe_bits += bits;
1623 if( h->fenc->b_last_minigop_bframe )
1625 update_predictor( rc->pred_b_from_p, qp2qscale( rc->qpa_rc ),
1626 h->fref[1][h->i_ref[1]-1]->i_satd, rc->bframe_bits / rc->bframes );
1627 rc->bframe_bits = 0;
1632 *filler = update_vbv( h, bits );
1633 rc->filler_bits_sum += *filler * 8;
1635 if( h->sps->vui.b_nal_hrd_parameters_present )
1637 if( h->fenc->i_frame == 0 )
1639 // access unit initialises the HRD
1640 h->fenc->hrd_timing.cpb_initial_arrival_time = 0;
1641 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1642 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1643 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit = (double)rc->initial_cpb_removal_delay / 90000;
1647 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit + (double)(h->fenc->i_cpb_delay - h->i_cpb_delay_pir_offset) *
1648 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1650 double cpb_earliest_arrival_time = h->fenc->hrd_timing.cpb_removal_time - (double)rc->initial_cpb_removal_delay / 90000;
1651 if( h->fenc->b_keyframe )
1653 rc->nrt_first_access_unit = h->fenc->hrd_timing.cpb_removal_time;
1654 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1655 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1658 cpb_earliest_arrival_time -= (double)rc->initial_cpb_removal_delay_offset / 90000;
1660 if( h->sps->vui.hrd.b_cbr_hrd )
1661 h->fenc->hrd_timing.cpb_initial_arrival_time = rc->previous_cpb_final_arrival_time;
1663 h->fenc->hrd_timing.cpb_initial_arrival_time = X264_MAX( rc->previous_cpb_final_arrival_time, cpb_earliest_arrival_time );
1665 int filler_bits = *filler ? X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), *filler )*8 : 0;
1667 h->fenc->hrd_timing.cpb_final_arrival_time = rc->previous_cpb_final_arrival_time = h->fenc->hrd_timing.cpb_initial_arrival_time +
1668 (double)(bits + filler_bits) / h->sps->vui.hrd.i_bit_rate_unscaled;
1670 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 +
1671 h->fenc->hrd_timing.cpb_removal_time;
1676 x264_log(h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n");
1680 /****************************************************************************
1682 ***************************************************************************/
1685 * modify the bitrate curve from pass1 for one frame
1687 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1689 x264_ratecontrol_t *rcc= h->rc;
1690 x264_zone_t *zone = get_zone( h, frame_num );
1691 double q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1693 // avoid NaN's in the rc_eq
1694 if( !isfinite(q) || rce->tex_bits + rce->mv_bits == 0 )
1695 q = rcc->last_qscale_for[rce->pict_type];
1700 rcc->last_qscale = q;
1705 if( zone->b_force_qp )
1706 q = qp2qscale( zone->i_qp );
1708 q /= zone->f_bitrate_factor;
1714 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1716 x264_ratecontrol_t *rcc = h->rc;
1717 const int pict_type = rce->pict_type;
1719 // force I/B quants as a function of P quants
1720 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1721 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1722 if( pict_type == SLICE_TYPE_I )
1725 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1726 double ip_factor = fabs( h->param.rc.f_ip_factor );
1727 /* don't apply ip_factor if the following frame is also I */
1728 if( rcc->accum_p_norm <= 0 )
1730 else if( h->param.rc.f_ip_factor < 0 )
1732 else if( rcc->accum_p_norm >= 1 )
1735 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1737 else if( pict_type == SLICE_TYPE_B )
1739 if( h->param.rc.f_pb_factor > 0 )
1741 if( !rce->kept_as_ref )
1742 q *= fabs( h->param.rc.f_pb_factor );
1744 else if( pict_type == SLICE_TYPE_P
1745 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1746 && rce->tex_bits == 0 )
1751 /* last qscale / qdiff stuff */
1752 if( rcc->last_non_b_pict_type == pict_type &&
1753 (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1) )
1755 double last_q = rcc->last_qscale_for[pict_type];
1756 double max_qscale = last_q * rcc->lstep;
1757 double min_qscale = last_q / rcc->lstep;
1759 if ( q > max_qscale ) q = max_qscale;
1760 else if( q < min_qscale ) q = min_qscale;
1763 rcc->last_qscale_for[pict_type] = q;
1764 if( pict_type != SLICE_TYPE_B )
1765 rcc->last_non_b_pict_type = pict_type;
1766 if( pict_type == SLICE_TYPE_I )
1768 rcc->last_accum_p_norm = rcc->accum_p_norm;
1769 rcc->accum_p_norm = 0;
1770 rcc->accum_p_qp = 0;
1772 if( pict_type == SLICE_TYPE_P )
1774 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1775 rcc->accum_p_qp = mask * (qscale2qp( q ) + rcc->accum_p_qp);
1776 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1781 static double predict_size( predictor_t *p, double q, double var )
1783 return (p->coeff*var + p->offset) / (q*p->count);
1786 static void update_predictor( predictor_t *p, double q, double var, double bits )
1788 const double range = 1.5;
1791 double old_coeff = p->coeff / p->count;
1792 double new_coeff = bits*q / var;
1793 double new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
1794 double new_offset = bits*q - new_coeff_clipped * var;
1795 if( new_offset >= 0 )
1796 new_coeff = new_coeff_clipped;
1799 p->count *= p->decay;
1800 p->coeff *= p->decay;
1801 p->offset *= p->decay;
1803 p->coeff += new_coeff;
1804 p->offset += new_offset;
1807 // update VBV after encoding a frame
1808 static int update_vbv( x264_t *h, int bits )
1811 int bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
1812 x264_ratecontrol_t *rcc = h->rc;
1813 x264_ratecontrol_t *rct = h->thread[0]->rc;
1814 uint64_t buffer_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
1816 if( rcc->last_satd >= h->mb.i_mb_count )
1817 update_predictor( &rct->pred[h->sh.i_type], qp2qscale( rcc->qpa_rc ), rcc->last_satd, bits );
1822 rct->buffer_fill_final -= (uint64_t)bits * h->sps->vui.i_time_scale;
1824 if( rct->buffer_fill_final < 0 )
1825 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 );
1826 rct->buffer_fill_final = X264_MAX( rct->buffer_fill_final, 0 );
1827 rct->buffer_fill_final += (uint64_t)bitrate * h->sps->vui.i_num_units_in_tick * h->fenc->i_cpb_duration;
1829 if( h->sps->vui.hrd.b_cbr_hrd && rct->buffer_fill_final > buffer_size )
1831 int64_t scale = (int64_t)h->sps->vui.i_time_scale * 8;
1832 filler = (rct->buffer_fill_final - buffer_size + scale - 1) / scale;
1833 bits = X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), filler ) * 8;
1834 rct->buffer_fill_final -= (uint64_t)bits * h->sps->vui.i_time_scale;
1837 rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, buffer_size );
1842 void x264_hrd_fullness( x264_t *h )
1844 x264_ratecontrol_t *rct = h->thread[0]->rc;
1845 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale / rct->hrd_multiply_denom;
1846 uint64_t cpb_state = rct->buffer_fill_final;
1847 uint64_t cpb_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
1848 uint64_t multiply_factor = 180000 / rct->hrd_multiply_denom;
1850 if( rct->buffer_fill_final < 0 || rct->buffer_fill_final > cpb_size )
1852 x264_log( h, X264_LOG_WARNING, "CPB %s: %.0lf bits in a %.0lf-bit buffer\n",
1853 rct->buffer_fill_final < 0 ? "underflow" : "overflow", (float)rct->buffer_fill_final/denom, (float)cpb_size/denom );
1856 h->initial_cpb_removal_delay = (multiply_factor * cpb_state + denom) / (2*denom);
1857 h->initial_cpb_removal_delay_offset = (multiply_factor * cpb_size + denom) / (2*denom) - h->initial_cpb_removal_delay;
1860 // provisionally update VBV according to the planned size of all frames currently in progress
1861 static void update_vbv_plan( x264_t *h, int overhead )
1863 x264_ratecontrol_t *rcc = h->rc;
1864 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final / h->sps->vui.i_time_scale;
1865 if( h->i_thread_frames > 1 )
1867 int j = h->rc - h->thread[0]->rc;
1868 for( int i = 1; i < h->i_thread_frames; i++ )
1870 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
1871 double bits = t->rc->frame_size_planned;
1872 if( !t->b_thread_active )
1874 bits = X264_MAX(bits, t->rc->frame_size_estimated);
1875 rcc->buffer_fill -= bits;
1876 rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
1877 rcc->buffer_fill += t->rc->buffer_rate;
1878 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1881 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1882 rcc->buffer_fill -= overhead;
1885 // apply VBV constraints and clip qscale to between lmin and lmax
1886 static double clip_qscale( x264_t *h, int pict_type, double q )
1888 x264_ratecontrol_t *rcc = h->rc;
1889 double lmin = rcc->lmin[pict_type];
1890 double lmax = rcc->lmax[pict_type];
1891 if( rcc->rate_factor_max_increment )
1892 lmax = X264_MIN( lmax, qp2qscale( rcc->qp_novbv + rcc->rate_factor_max_increment ) );
1895 /* B-frames are not directly subject to VBV,
1896 * since they are controlled by the P-frames' QPs. */
1898 if( rcc->b_vbv && rcc->last_satd > 0 )
1900 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
1901 * the lookahead overflow and such that the buffer is in a reasonable state
1902 * by the end of the lookahead. */
1903 if( h->param.rc.i_lookahead )
1907 /* Avoid an infinite loop. */
1908 for( int iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
1911 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1912 double buffer_fill_cur = rcc->buffer_fill - cur_bits;
1914 double total_duration = 0;
1915 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
1916 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
1917 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
1919 /* Loop over the planned future frames. */
1920 for( int j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
1922 total_duration += h->fenc->f_planned_cpb_duration[j];
1923 buffer_fill_cur += rcc->vbv_max_rate * h->fenc->f_planned_cpb_duration[j];
1924 int i_type = h->fenc->i_planned_type[j];
1925 int i_satd = h->fenc->i_planned_satd[j];
1926 if( i_type == X264_TYPE_AUTO )
1928 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
1929 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
1930 buffer_fill_cur -= cur_bits;
1932 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
1933 target_fill = X264_MIN( rcc->buffer_fill + total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.5 );
1934 if( buffer_fill_cur < target_fill )
1940 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
1941 target_fill = x264_clip3f( rcc->buffer_fill - total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
1942 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
1951 /* Fallback to old purely-reactive algorithm: no lookahead. */
1954 if( ( pict_type == SLICE_TYPE_P ||
1955 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
1956 rcc->buffer_fill/rcc->buffer_size < 0.5 )
1958 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1961 /* Now a hard threshold to make sure the frame fits in VBV.
1962 * This one is mostly for I-frames. */
1963 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1965 /* For small VBVs, allow the frame to use up the entire VBV. */
1966 double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
1967 /* For single-frame VBVs, request that the frame use up the entire VBV. */
1968 double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
1970 if( bits > rcc->buffer_fill/max_fill_factor )
1971 qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
1974 if( bits < rcc->buffer_rate/min_fill_factor )
1975 q *= bits*min_fill_factor/rcc->buffer_rate;
1976 q = X264_MAX( q0, q );
1979 /* Apply MinCR restrictions */
1980 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1981 if( bits > rcc->frame_size_maximum )
1982 q *= bits / rcc->frame_size_maximum;
1983 bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1985 /* Check B-frame complexity, and use up any bits that would
1986 * overflow before the next P-frame. */
1987 if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
1989 int nb = rcc->bframes;
1990 double pbbits = bits;
1991 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1993 double bframe_cpb_duration = 0;
1994 double minigop_cpb_duration;
1995 for( int i = 0; i < nb; i++ )
1996 bframe_cpb_duration += h->fenc->f_planned_cpb_duration[1+i];
1998 if( bbits * nb > bframe_cpb_duration * rcc->vbv_max_rate )
2000 pbbits += nb * bbits;
2002 minigop_cpb_duration = bframe_cpb_duration + h->fenc->f_planned_cpb_duration[0];
2003 space = rcc->buffer_fill + minigop_cpb_duration*rcc->vbv_max_rate - rcc->buffer_size;
2004 if( pbbits < space )
2006 q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
2008 q = X264_MAX( q0/2, q );
2011 if( !rcc->b_vbv_min_rate )
2012 q = X264_MAX( q0, q );
2017 else if( rcc->b_2pass )
2019 double min2 = log( lmin );
2020 double max2 = log( lmax );
2021 q = (log(q) - min2)/(max2-min2) - 0.5;
2022 q = 1.0/(1.0 + exp( -4*q ));
2023 q = q*(max2-min2) + min2;
2027 return x264_clip3f( q, lmin, lmax );
2030 // update qscale for 1 frame based on actual bits used so far
2031 static float rate_estimate_qscale( x264_t *h )
2034 x264_ratecontrol_t *rcc = h->rc;
2035 ratecontrol_entry_t UNINIT(rce);
2036 int pict_type = h->sh.i_type;
2037 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
2038 + h->stat.i_frame_size[SLICE_TYPE_P]
2039 + h->stat.i_frame_size[SLICE_TYPE_B])
2040 - rcc->filler_bits_sum;
2045 if( pict_type != rce.pict_type )
2047 x264_log( h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
2048 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type] );
2052 if( pict_type == SLICE_TYPE_B )
2054 /* B-frames don't have independent ratecontrol, but rather get the
2055 * average QP of the two adjacent P-frames + an offset */
2057 int i0 = IS_X264_TYPE_I(h->fref_nearest[0]->i_type);
2058 int i1 = IS_X264_TYPE_I(h->fref_nearest[1]->i_type);
2059 int dt0 = abs(h->fenc->i_poc - h->fref_nearest[0]->i_poc);
2060 int dt1 = abs(h->fenc->i_poc - h->fref_nearest[1]->i_poc);
2061 float q0 = h->fref_nearest[0]->f_qp_avg_rc;
2062 float q1 = h->fref_nearest[1]->f_qp_avg_rc;
2064 if( h->fref_nearest[0]->i_type == X264_TYPE_BREF )
2065 q0 -= rcc->pb_offset/2;
2066 if( h->fref_nearest[1]->i_type == X264_TYPE_BREF )
2067 q1 -= rcc->pb_offset/2;
2070 q = (q0 + q1) / 2 + rcc->ip_offset;
2076 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
2078 if( h->fenc->b_kept_as_ref )
2079 q += rcc->pb_offset/2;
2081 q += rcc->pb_offset;
2083 if( rcc->b_2pass && rcc->b_vbv )
2084 rcc->frame_size_planned = qscale2bits( &rce, qp2qscale( q ) );
2086 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, qp2qscale( q ), h->fref[1][h->i_ref[1]-1]->i_satd );
2087 h->rc->frame_size_estimated = rcc->frame_size_planned;
2091 rcc->last_satd = x264_rc_analyse_slice( h );
2093 return qp2qscale( q );
2097 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
2101 double lmin = rcc->lmin[pict_type];
2102 double lmax = rcc->lmax[pict_type];
2104 int64_t predicted_bits = total_bits;
2108 if( h->i_thread_frames > 1 )
2110 int j = h->rc - h->thread[0]->rc;
2111 for( int i = 1; i < h->i_thread_frames; i++ )
2113 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
2114 double bits = t->rc->frame_size_planned;
2115 if( !t->b_thread_active )
2117 bits = X264_MAX(bits, t->rc->frame_size_estimated);
2118 predicted_bits += (int64_t)bits;
2124 if( h->i_frame < h->i_thread_frames )
2125 predicted_bits += (int64_t)h->i_frame * rcc->bitrate / rcc->fps;
2127 predicted_bits += (int64_t)(h->i_thread_frames - 1) * rcc->bitrate / rcc->fps;
2130 /* Adjust ABR buffer based on distance to the end of the video. */
2131 if( rcc->num_entries > h->i_frame )
2133 double final_bits = rcc->entry[rcc->num_entries-1].expected_bits;
2134 double video_pos = rce.expected_bits / final_bits;
2135 double scale_factor = sqrt( (1 - video_pos) * rcc->num_entries );
2136 abr_buffer *= 0.5 * X264_MAX( scale_factor, 0.5 );
2139 diff = predicted_bits - (int64_t)rce.expected_bits;
2141 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
2142 if( ((h->i_frame + 1 - h->i_thread_frames) >= rcc->fps) &&
2143 (rcc->expected_bits_sum > 0))
2145 /* Adjust quant based on the difference between
2146 * achieved and expected bitrate so far */
2147 double cur_time = (double)h->i_frame / rcc->num_entries;
2148 double w = x264_clip3f( cur_time*100, 0.0, 1.0 );
2149 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
2153 /* Do not overflow vbv */
2154 double expected_size = qscale2bits( &rce, q );
2155 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2156 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
2157 double qmax = q*(2 - expected_fullness);
2158 double size_constraint = 1 + expected_fullness;
2159 qmax = X264_MAX( qmax, rce.new_qscale );
2160 if( expected_fullness < .05 )
2162 qmax = X264_MIN(qmax, lmax);
2163 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
2164 ((expected_vbv < 0) && (q < lmax)))
2167 expected_size = qscale2bits(&rce, q);
2168 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2170 rcc->last_satd = x264_rc_analyse_slice( h );
2172 q = x264_clip3f( q, lmin, lmax );
2174 else /* 1pass ABR */
2176 /* Calculate the quantizer which would have produced the desired
2177 * average bitrate if it had been applied to all frames so far.
2178 * Then modulate that quant based on the current frame's complexity
2179 * relative to the average complexity so far (using the 2pass RCEQ).
2180 * Then bias the quant up or down if total size so far was far from
2182 * Result: Depending on the value of rate_tolerance, there is a
2183 * tradeoff between quality and bitrate precision. But at large
2184 * tolerances, the bit distribution approaches that of 2pass. */
2186 double wanted_bits, overflow = 1;
2188 rcc->last_satd = x264_rc_analyse_slice( h );
2189 rcc->short_term_cplxsum *= 0.5;
2190 rcc->short_term_cplxcount *= 0.5;
2191 rcc->short_term_cplxsum += rcc->last_satd / (CLIP_DURATION(h->fenc->f_duration) / BASE_FRAME_DURATION);
2192 rcc->short_term_cplxcount ++;
2194 rce.tex_bits = rcc->last_satd;
2195 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
2197 rce.p_count = rcc->nmb;
2201 rce.pict_type = pict_type;
2203 if( h->param.rc.i_rc_method == X264_RC_CRF )
2205 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
2209 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
2211 /* ABR code can potentially be counterproductive in CBR, so just don't bother.
2212 * Don't run it if the frame complexity is zero either. */
2213 if( !rcc->b_vbv_min_rate && rcc->last_satd )
2215 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
2216 int i_frame_done = h->i_frame + 1 - h->i_thread_frames;
2217 double time_done = i_frame_done / rcc->fps;
2218 if( h->param.b_vfr_input && i_frame_done > 0 )
2219 time_done = ((double)(h->fenc->i_reordered_pts - h->i_reordered_pts_delay)) * h->param.i_timebase_num / h->param.i_timebase_den;
2220 wanted_bits = time_done * rcc->bitrate;
2221 if( wanted_bits > 0 )
2223 abr_buffer *= X264_MAX( 1, sqrt( time_done ) );
2224 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
2230 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
2231 /* should test _next_ pict type, but that isn't decided yet */
2232 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
2234 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2235 q /= fabs( h->param.rc.f_ip_factor );
2237 else if( h->i_frame > 0 )
2239 if( h->param.rc.i_rc_method != X264_RC_CRF )
2241 /* Asymmetric clipping, because symmetric would prevent
2242 * overflow control in areas of rapidly oscillating complexity */
2243 double lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
2244 double lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
2245 if( overflow > 1.1 && h->i_frame > 3 )
2247 else if( overflow < 0.9 )
2250 q = x264_clip3f(q, lmin, lmax);
2253 else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
2255 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
2257 rcc->qp_novbv = qscale2qp( q );
2259 //FIXME use get_diff_limited_q() ?
2260 q = clip_qscale( h, pict_type, q );
2263 rcc->last_qscale_for[pict_type] =
2264 rcc->last_qscale = q;
2266 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
2267 rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );
2269 if( rcc->b_2pass && rcc->b_vbv )
2270 rcc->frame_size_planned = qscale2bits(&rce, q);
2272 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2274 /* Always use up the whole VBV in this case. */
2275 if( rcc->single_frame_vbv )
2276 rcc->frame_size_planned = rcc->buffer_rate;
2277 h->rc->frame_size_estimated = rcc->frame_size_planned;
2282 void x264_threads_normalize_predictors( x264_t *h )
2284 double totalsize = 0;
2285 for( int i = 0; i < h->param.i_threads; i++ )
2286 totalsize += h->thread[i]->rc->slice_size_planned;
2287 double factor = h->rc->frame_size_planned / totalsize;
2288 for( int i = 0; i < h->param.i_threads; i++ )
2289 h->thread[i]->rc->slice_size_planned *= factor;
2292 void x264_threads_distribute_ratecontrol( x264_t *h )
2295 x264_ratecontrol_t *rc = h->rc;
2297 /* Initialize row predictors */
2298 if( h->i_frame == 0 )
2299 for( int i = 0; i < h->param.i_threads; i++ )
2301 x264_ratecontrol_t *t = h->thread[i]->rc;
2302 memcpy( t->row_preds, rc->row_preds, sizeof(rc->row_preds) );
2305 for( int i = 0; i < h->param.i_threads; i++ )
2307 x264_t *t = h->thread[i];
2308 memcpy( t->rc, rc, offsetof(x264_ratecontrol_t, row_pred) );
2309 t->rc->row_pred = &t->rc->row_preds[h->sh.i_type];
2310 /* Calculate the planned slice size. */
2311 if( rc->b_vbv && rc->frame_size_planned )
2314 for( row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2315 size += h->fdec->i_row_satd[row];
2316 t->rc->slice_size_planned = predict_size( &rc->pred[h->sh.i_type + (i+1)*5], rc->qpm, size );
2319 t->rc->slice_size_planned = 0;
2321 if( rc->b_vbv && rc->frame_size_planned )
2323 x264_threads_normalize_predictors( h );
2325 if( rc->single_frame_vbv )
2327 /* Compensate for our max frame error threshold: give more bits (proportionally) to smaller slices. */
2328 for( int i = 0; i < h->param.i_threads; i++ )
2330 x264_t *t = h->thread[i];
2331 float max_frame_error = X264_MAX( 0.05, 1.0 / (t->i_threadslice_end - t->i_threadslice_start) );
2332 t->rc->slice_size_planned += 2 * max_frame_error * rc->frame_size_planned;
2334 x264_threads_normalize_predictors( h );
2337 for( int i = 0; i < h->param.i_threads; i++ )
2338 h->thread[i]->rc->frame_size_estimated = h->thread[i]->rc->slice_size_planned;
2342 void x264_threads_merge_ratecontrol( x264_t *h )
2344 x264_ratecontrol_t *rc = h->rc;
2347 for( int i = 0; i < h->param.i_threads; i++ )
2349 x264_t *t = h->thread[i];
2350 x264_ratecontrol_t *rct = h->thread[i]->rc;
2351 if( h->param.rc.i_vbv_buffer_size )
2354 for( int row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2355 size += h->fdec->i_row_satd[row];
2356 int bits = t->stat.frame.i_mv_bits + t->stat.frame.i_tex_bits + t->stat.frame.i_misc_bits;
2357 int mb_count = (t->i_threadslice_end - t->i_threadslice_start) * h->mb.i_mb_width;
2358 update_predictor( &rc->pred[h->sh.i_type+(i+1)*5], qp2qscale( rct->qpa_rc/mb_count ), size, bits );
2362 rc->qpa_rc += rct->qpa_rc;
2363 rc->qpa_aq += rct->qpa_aq;
2367 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
2371 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
2372 /* these vars are updated in x264_ratecontrol_start()
2373 * so copy them from the context that most recently started (prev)
2374 * to the context that's about to start (cur). */
2379 COPY(last_qscale_for);
2380 COPY(last_non_b_pict_type);
2381 COPY(short_term_cplxsum);
2382 COPY(short_term_cplxcount);
2386 /* these vars can be updated by x264_ratecontrol_init_reconfigurable */
2389 COPY(single_frame_vbv);
2391 COPY(b_vbv_min_rate);
2392 COPY(rate_factor_constant);
2398 #define COPY(var) next->rc->var = cur->rc->var
2399 /* these vars are updated in x264_ratecontrol_end()
2400 * so copy them from the context that most recently ended (cur)
2401 * to the context that's about to end (next) */
2403 COPY(expected_bits_sum);
2404 COPY(filler_bits_sum);
2405 COPY(wanted_bits_window);
2407 COPY(initial_cpb_removal_delay);
2408 COPY(initial_cpb_removal_delay_offset);
2409 COPY(nrt_first_access_unit);
2410 COPY(previous_cpb_final_arrival_time);
2413 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
2414 /* the rest of the variables are either constant or thread-local */
2417 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
2419 /* find an interval ending on an overflow or underflow (depending on whether
2420 * we're adding or removing bits), and starting on the earliest frame that
2421 * can influence the buffer fill of that end frame. */
2422 x264_ratecontrol_t *rcc = h->rc;
2423 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
2424 const double buffer_max = .9 * rcc->buffer_size;
2425 double fill = fills[*t0-1];
2426 double parity = over ? 1. : -1.;
2427 int start = -1, end = -1;
2428 for( int i = *t0; i < rcc->num_entries; i++ )
2430 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 -
2431 qscale2bits( &rcc->entry[i], rcc->entry[i].new_qscale )) * parity;
2432 fill = x264_clip3f(fill, 0, rcc->buffer_size);
2434 if( fill <= buffer_min || i == 0 )
2440 else if( fill >= buffer_max && start >= 0 )
2445 return start >= 0 && end >= 0;
2448 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
2450 x264_ratecontrol_t *rcc = h->rc;
2451 double qscale_orig, qscale_new;
2455 for( int i = t0; i <= t1; i++ )
2457 qscale_orig = rcc->entry[i].new_qscale;
2458 qscale_orig = x264_clip3f( qscale_orig, qscale_min, qscale_max );
2459 qscale_new = qscale_orig * adjustment;
2460 qscale_new = x264_clip3f( qscale_new, qscale_min, qscale_max );
2461 rcc->entry[i].new_qscale = qscale_new;
2462 adjusted = adjusted || (qscale_new != qscale_orig);
2467 static double count_expected_bits( x264_t *h )
2469 x264_ratecontrol_t *rcc = h->rc;
2470 double expected_bits = 0;
2471 for( int i = 0; i < rcc->num_entries; i++ )
2473 ratecontrol_entry_t *rce = &rcc->entry[i];
2474 rce->expected_bits = expected_bits;
2475 expected_bits += qscale2bits( rce, rce->new_qscale );
2477 return expected_bits;
2480 static int vbv_pass2( x264_t *h, double all_available_bits )
2482 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
2483 * frames in the interval until either buffer is full at some intermediate frame or the
2484 * last frame in the interval no longer underflows. Recompute intervals and repeat.
2485 * Then do the converse to put bits back into overflow areas until target size is met */
2487 x264_ratecontrol_t *rcc = h->rc;
2489 double expected_bits = 0;
2491 double prev_bits = 0;
2493 double qscale_min = qp2qscale( h->param.rc.i_qp_min );
2494 double qscale_max = qp2qscale( h->param.rc.i_qp_max );
2496 int adj_min, adj_max;
2497 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2501 /* adjust overall stream size */
2505 prev_bits = expected_bits;
2508 { /* not first iteration */
2509 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2510 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2514 while(adj_min && find_underflow( h, fills, &t0, &t1, 1 ))
2516 adj_min = fix_underflow( h, t0, t1, adjustment, qscale_min, qscale_max );
2521 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2523 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2525 while( adj_max && find_underflow( h, fills, &t0, &t1, 0 ) )
2526 adj_max = fix_underflow( h, t0, t1, 1.001, qscale_min, qscale_max );
2528 expected_bits = count_expected_bits( h );
2529 } while( (expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2532 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2534 /* store expected vbv filling values for tracking when encoding */
2535 for( int i = 0; i < rcc->num_entries; i++ )
2536 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
2538 x264_free( fills-1 );
2544 static int init_pass2( x264_t *h )
2546 x264_ratecontrol_t *rcc = h->rc;
2547 uint64_t all_const_bits = 0;
2548 double timescale = (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2549 double duration = 0;
2550 for( int i = 0; i < rcc->num_entries; i++ )
2551 duration += rcc->entry[i].i_duration;
2552 duration *= timescale;
2553 uint64_t all_available_bits = h->param.rc.i_bitrate * 1000. * duration;
2554 double rate_factor, step_mult;
2555 double qblur = h->param.rc.f_qblur;
2556 double cplxblur = h->param.rc.f_complexity_blur;
2557 const int filter_size = (int)(qblur*4) | 1;
2558 double expected_bits;
2559 double *qscale, *blurred_qscale;
2560 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
2562 /* find total/average complexity & const_bits */
2563 for( int i = 0; i < rcc->num_entries; i++ )
2565 ratecontrol_entry_t *rce = &rcc->entry[i];
2566 all_const_bits += rce->misc_bits;
2569 if( all_available_bits < all_const_bits)
2571 x264_log( h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2572 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)) );
2576 /* Blur complexities, to reduce local fluctuation of QP.
2577 * We don't blur the QPs directly, because then one very simple frame
2578 * could drag down the QP of a nearby complex frame and give it more
2579 * bits than intended. */
2580 for( int i = 0; i < rcc->num_entries; i++ )
2582 ratecontrol_entry_t *rce = &rcc->entry[i];
2583 double weight_sum = 0;
2584 double cplx_sum = 0;
2585 double weight = 1.0;
2586 double gaussian_weight;
2587 /* weighted average of cplx of future frames */
2588 for( int j = 1; j < cplxblur*2 && j < rcc->num_entries-i; j++ )
2590 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2591 double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2592 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2593 if( weight < .0001 )
2595 gaussian_weight = weight * exp( -j*j/200.0 );
2596 weight_sum += gaussian_weight;
2597 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2599 /* weighted average of cplx of past frames */
2601 for( int j = 0; j <= cplxblur*2 && j <= i; j++ )
2603 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2604 double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2605 gaussian_weight = weight * exp( -j*j/200.0 );
2606 weight_sum += gaussian_weight;
2607 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2608 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2609 if( weight < .0001 )
2612 rce->blurred_complexity = cplx_sum / weight_sum;
2615 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2616 if( filter_size > 1 )
2617 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2619 blurred_qscale = qscale;
2621 /* Search for a factor which, when multiplied by the RCEQ values from
2622 * each frame, adds up to the desired total size.
2623 * There is no exact closed-form solution because of VBV constraints and
2624 * because qscale2bits is not invertible, but we can start with the simple
2625 * approximation of scaling the 1st pass by the ratio of bitrates.
2626 * The search range is probably overkill, but speed doesn't matter here. */
2629 for( int i = 0; i < rcc->num_entries; i++ )
2631 double q = get_qscale(h, &rcc->entry[i], 1.0, i);
2632 expected_bits += qscale2bits(&rcc->entry[i], q);
2633 rcc->last_qscale_for[rcc->entry[i].pict_type] = q;
2635 step_mult = all_available_bits / expected_bits;
2638 for( double step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2641 rate_factor += step;
2643 rcc->last_non_b_pict_type = -1;
2644 rcc->last_accum_p_norm = 1;
2645 rcc->accum_p_norm = 0;
2647 rcc->last_qscale_for[0] =
2648 rcc->last_qscale_for[1] =
2649 rcc->last_qscale_for[2] = pow( base_cplx, 1 - rcc->qcompress ) / rate_factor;
2652 for( int i = 0; i < rcc->num_entries; i++ )
2654 qscale[i] = get_qscale( h, &rcc->entry[i], rate_factor, i );
2655 rcc->last_qscale_for[rcc->entry[i].pict_type] = qscale[i];
2658 /* fixed I/B qscale relative to P */
2659 for( int i = rcc->num_entries-1; i >= 0; i-- )
2661 qscale[i] = get_diff_limited_q( h, &rcc->entry[i], qscale[i] );
2662 assert(qscale[i] >= 0);
2666 if( filter_size > 1 )
2668 assert( filter_size%2 == 1 );
2669 for( int i = 0; i < rcc->num_entries; i++ )
2671 ratecontrol_entry_t *rce = &rcc->entry[i];
2672 double q = 0.0, sum = 0.0;
2674 for( int j = 0; j < filter_size; j++ )
2676 int idx = i+j-filter_size/2;
2678 double coeff = qblur==0 ? 1.0 : exp( -d*d/(qblur*qblur) );
2679 if( idx < 0 || idx >= rcc->num_entries )
2681 if( rce->pict_type != rcc->entry[idx].pict_type )
2683 q += qscale[idx] * coeff;
2686 blurred_qscale[i] = q/sum;
2690 /* find expected bits */
2691 for( int i = 0; i < rcc->num_entries; i++ )
2693 ratecontrol_entry_t *rce = &rcc->entry[i];
2694 rce->new_qscale = clip_qscale( h, rce->pict_type, blurred_qscale[i] );
2695 assert(rce->new_qscale >= 0);
2696 expected_bits += qscale2bits( rce, rce->new_qscale );
2699 if( expected_bits > all_available_bits )
2700 rate_factor -= step;
2703 x264_free( qscale );
2704 if( filter_size > 1 )
2705 x264_free( blurred_qscale );
2708 if( vbv_pass2( h, all_available_bits ) )
2710 expected_bits = count_expected_bits( h );
2712 if( fabs( expected_bits/all_available_bits - 1.0 ) > 0.01 )
2715 for( int i = 0; i < rcc->num_entries; i++ )
2716 avgq += rcc->entry[i].new_qscale;
2717 avgq = qscale2qp( avgq / rcc->num_entries );
2719 if( expected_bits > all_available_bits || !rcc->b_vbv )
2720 x264_log( h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n" );
2721 x264_log( h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
2722 (float)h->param.rc.i_bitrate,
2723 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
2725 if( expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2 )
2727 if( h->param.rc.i_qp_min > 0 )
2728 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min );
2730 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate\n" );
2732 else if( expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2 )
2734 if( h->param.rc.i_qp_max < QP_MAX )
2735 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max );
2737 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate\n");
2739 else if( !(rcc->b_2pass && rcc->b_vbv) )
2740 x264_log( h, X264_LOG_WARNING, "internal error\n" );