1 /*****************************************************************************
2 * ratecontrol.c: ratecontrol
3 *****************************************************************************
4 * Copyright (C) 2005-2012 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 */
91 int qpa_aq; /* average of macroblocks' qp after aq */
93 float qp_novbv; /* QP for the current frame if 1-pass VBV was disabled. */
97 int64_t buffer_fill_final;
98 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
99 double buffer_rate; /* # of bits added to buffer_fill after each frame */
100 double vbv_max_rate; /* # of bits added to buffer_fill per second */
101 predictor_t *pred; /* predict frame size from satd */
102 int single_frame_vbv;
103 double rate_factor_max_increment; /* Don't allow RF above (CRF + this value). */
108 double cplxr_sum; /* sum of bits*qscale/rceq */
109 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow, only includes finished frames */
110 int64_t filler_bits_sum; /* sum in bits of finished frames' filler data */
111 double wanted_bits_window; /* target bitrate * window */
113 double short_term_cplxsum;
114 double short_term_cplxcount;
115 double rate_factor_constant;
120 FILE *p_stat_file_out;
121 char *psz_stat_file_tmpname;
122 FILE *p_mbtree_stat_file_out;
123 char *psz_mbtree_stat_file_tmpname;
124 char *psz_mbtree_stat_file_name;
125 FILE *p_mbtree_stat_file_in;
127 int num_entries; /* number of ratecontrol_entry_ts */
128 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
130 double last_qscale_for[3]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
131 int last_non_b_pict_type;
132 double accum_p_qp; /* for determining I-frame quant */
134 double last_accum_p_norm;
135 double lmin[3]; /* min qscale by frame type */
137 double lstep; /* max change (multiply) in qscale per frame */
138 uint16_t *qp_buffer[2]; /* Global buffers for converting MB-tree quantizer data. */
139 int qpbuf_pos; /* In order to handle pyramid reordering, QP buffer acts as a stack.
140 * This value is the current position (0 or 1). */
143 float frame_size_estimated; /* Access to this variable must be atomic: double is
144 * not atomic on all arches we care about */
145 double frame_size_maximum; /* Maximum frame size due to MinCR */
146 double frame_size_planned;
147 double slice_size_planned;
148 predictor_t (*row_pred)[2];
149 predictor_t row_preds[3][2];
150 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
151 int bframes; /* # consecutive B-frames before this P-frame */
152 int bframe_bits; /* total cost of those frames */
156 x264_zone_t *prev_zone;
159 int initial_cpb_removal_delay;
160 int initial_cpb_removal_delay_offset;
161 double nrt_first_access_unit; /* nominal removal time */
162 double previous_cpb_final_arrival_time;
163 uint64_t hrd_multiply_denom;
167 static int parse_zones( x264_t *h );
168 static int init_pass2(x264_t *);
169 static float rate_estimate_qscale( x264_t *h );
170 static int update_vbv( x264_t *h, int bits );
171 static void update_vbv_plan( x264_t *h, int overhead );
172 static float predict_size( predictor_t *p, float q, float var );
173 static void update_predictor( predictor_t *p, float q, float var, float bits );
175 #define CMP_OPT_FIRST_PASS( opt, param_val )\
177 if( ( p = strstr( opts, opt "=" ) ) && sscanf( p, opt "=%d" , &i ) && param_val != i )\
179 x264_log( h, X264_LOG_ERROR, "different " opt " setting than first pass (%d vs %d)\n", param_val, i );\
185 * qp = h.264's quantizer
186 * qscale = linearized quantizer = Lagrange multiplier
188 static inline float qp2qscale( float qp )
190 return 0.85f * powf( 2.0f, ( qp - 12.0f ) / 6.0f );
192 static inline float qscale2qp( float qscale )
194 return 12.0f + 6.0f * log2f( qscale/0.85f );
197 /* Texture bitrate is not quite inversely proportional to qscale,
198 * probably due the the changing number of SKIP blocks.
199 * MV bits level off at about qp<=12, because the lambda used
200 * for motion estimation is constant there. */
201 static inline double qscale2bits( ratecontrol_entry_t *rce, double qscale )
205 return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
206 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
210 static ALWAYS_INLINE uint32_t ac_energy_var( uint64_t sum_ssd, int shift, x264_frame_t *frame, int i, int b_store )
212 uint32_t sum = sum_ssd;
213 uint32_t ssd = sum_ssd >> 32;
216 frame->i_pixel_sum[i] += sum;
217 frame->i_pixel_ssd[i] += ssd;
219 return ssd - ((uint64_t)sum * sum >> shift);
222 static ALWAYS_INLINE uint32_t ac_energy_plane( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame, int i, int b_chroma, int b_field, int b_store )
224 int height = b_chroma ? 16>>CHROMA_V_SHIFT : 16;
225 int stride = frame->i_stride[i];
227 ? 16 * mb_x + height * (mb_y&~1) * stride + (mb_y&1) * stride
228 : 16 * mb_x + height * mb_y * stride;
232 ALIGNED_ARRAY_16( pixel, pix,[FENC_STRIDE*16] );
233 int chromapix = h->luma2chroma_pixel[PIXEL_16x16];
234 int shift = 7 - CHROMA_V_SHIFT;
236 h->mc.load_deinterleave_chroma_fenc( pix, frame->plane[1] + offset, stride, height );
237 return ac_energy_var( h->pixf.var[chromapix]( pix, FENC_STRIDE ), shift, frame, 1, b_store )
238 + ac_energy_var( h->pixf.var[chromapix]( pix+FENC_STRIDE/2, FENC_STRIDE ), shift, frame, 2, b_store );
241 return ac_energy_var( h->pixf.var[PIXEL_16x16]( frame->plane[i] + offset, stride ), 8, frame, i, b_store );
244 // Find the total AC energy of the block in all planes.
245 static NOINLINE uint32_t x264_ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
247 /* This function contains annoying hacks because GCC has a habit of reordering emms
248 * and putting it after floating point ops. As a result, we put the emms at the end of the
249 * function and make sure that its always called before the float math. Noinline makes
250 * sure no reordering goes on. */
252 x264_prefetch_fenc( h, frame, mb_x, mb_y );
253 if( h->mb.b_adaptive_mbaff )
255 /* We don't know the super-MB mode we're going to pick yet, so
256 * simply try both and pick the lower of the two. */
257 uint32_t var_interlaced, var_progressive;
258 var_interlaced = ac_energy_plane( h, mb_x, mb_y, frame, 0, 0, 1, 1 );
259 var_progressive = ac_energy_plane( h, mb_x, mb_y, frame, 0, 0, 0, 0 );
262 var_interlaced += ac_energy_plane( h, mb_x, mb_y, frame, 1, 0, 1, 1 );
263 var_progressive += ac_energy_plane( h, mb_x, mb_y, frame, 1, 0, 0, 0 );
264 var_interlaced += ac_energy_plane( h, mb_x, mb_y, frame, 2, 0, 1, 1 );
265 var_progressive += ac_energy_plane( h, mb_x, mb_y, frame, 2, 0, 0, 0 );
269 var_interlaced += ac_energy_plane( h, mb_x, mb_y, frame, 1, 1, 1, 1 );
270 var_progressive += ac_energy_plane( h, mb_x, mb_y, frame, 1, 1, 0, 0 );
272 var = X264_MIN( var_interlaced, var_progressive );
276 var = ac_energy_plane( h, mb_x, mb_y, frame, 0, 0, PARAM_INTERLACED, 1 );
279 var += ac_energy_plane( h, mb_x, mb_y, frame, 1, 0, PARAM_INTERLACED, 1 );
280 var += ac_energy_plane( h, mb_x, mb_y, frame, 2, 0, PARAM_INTERLACED, 1 );
283 var += ac_energy_plane( h, mb_x, mb_y, frame, 1, 1, PARAM_INTERLACED, 1 );
289 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame, float *quant_offsets )
291 /* constants chosen to result in approximately the same overall bitrate as without AQ.
292 * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
295 /* Initialize frame stats */
296 for( int i = 0; i < 3; i++ )
298 frame->i_pixel_sum[i] = 0;
299 frame->i_pixel_ssd[i] = 0;
302 /* Degenerate cases */
303 if( h->param.rc.i_aq_mode == X264_AQ_NONE || h->param.rc.f_aq_strength == 0 )
305 /* Need to init it anyways for MB tree */
306 if( h->param.rc.i_aq_mode && h->param.rc.f_aq_strength == 0 )
310 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
311 frame->f_qp_offset[mb_xy] = frame->f_qp_offset_aq[mb_xy] = quant_offsets[mb_xy];
312 if( h->frames.b_have_lowres )
313 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
314 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8( frame->f_qp_offset[mb_xy] );
318 memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
319 memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
320 if( h->frames.b_have_lowres )
321 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
322 frame->i_inv_qscale_factor[mb_xy] = 256;
325 /* Need variance data for weighted prediction */
326 if( h->param.analyse.i_weighted_pred )
328 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
329 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
330 x264_ac_energy_mb( h, mb_x, mb_y, frame );
335 /* Actual adaptive quantization */
338 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
340 float bit_depth_correction = powf(1 << (BIT_DEPTH-8), 0.5f);
341 float avg_adj_pow2 = 0.f;
342 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
343 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
345 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
346 float qp_adj = powf( energy + 1, 0.125f );
347 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
349 avg_adj_pow2 += qp_adj * qp_adj;
351 avg_adj /= h->mb.i_mb_count;
352 avg_adj_pow2 /= h->mb.i_mb_count;
353 strength = h->param.rc.f_aq_strength * avg_adj / bit_depth_correction;
354 avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - (14.f * bit_depth_correction)) / avg_adj;
357 strength = h->param.rc.f_aq_strength * 1.0397f;
359 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
360 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
363 int mb_xy = mb_x + mb_y*h->mb.i_mb_stride;
364 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
366 qp_adj = frame->f_qp_offset[mb_xy];
367 qp_adj = strength * (qp_adj - avg_adj);
371 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
372 qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - (14.427f + 2*(BIT_DEPTH-8)));
375 qp_adj += quant_offsets[mb_xy];
376 frame->f_qp_offset[mb_xy] =
377 frame->f_qp_offset_aq[mb_xy] = qp_adj;
378 if( h->frames.b_have_lowres )
379 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8(qp_adj);
383 /* Remove mean from SSD calculation */
384 for( int i = 0; i < 3; i++ )
386 uint64_t ssd = frame->i_pixel_ssd[i];
387 uint64_t sum = frame->i_pixel_sum[i];
388 int width = 16*h->mb.i_mb_width >> (i && CHROMA_H_SHIFT);
389 int height = 16*h->mb.i_mb_height >> (i && CHROMA_V_SHIFT);
390 frame->i_pixel_ssd[i] = ssd - (sum * sum + width * height / 2) / (width * height);
394 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame, float *quant_offsets )
396 x264_ratecontrol_t *rc = h->rc;
397 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
399 if( rc->entry[frame->i_frame].kept_as_ref )
402 if( rc->qpbuf_pos < 0 )
408 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
410 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 )
413 if( i_type != i_type_actual && rc->qpbuf_pos == 1 )
415 x264_log( h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type, i_type_actual );
418 } while( i_type != i_type_actual );
421 for( int i = 0; i < h->mb.i_mb_count; i++ )
423 frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[rc->qpbuf_pos][i] )) * (1/256.0);
424 if( h->frames.b_have_lowres )
425 frame->i_inv_qscale_factor[i] = x264_exp2fix8(frame->f_qp_offset[i]);
430 x264_stack_align( x264_adaptive_quant_frame, h, frame, quant_offsets );
433 x264_log( h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n" );
437 int x264_reference_build_list_optimal( x264_t *h )
439 ratecontrol_entry_t *rce = h->rc->rce;
440 x264_frame_t *frames[16];
441 x264_weight_t weights[16][3];
444 if( rce->refs != h->i_ref[0] )
447 memcpy( frames, h->fref[0], sizeof(frames) );
448 memcpy( refcount, rce->refcount, sizeof(refcount) );
449 memcpy( weights, h->fenc->weight, sizeof(weights) );
450 memset( &h->fenc->weight[1][0], 0, sizeof(x264_weight_t[15][3]) );
452 /* For now don't reorder ref 0; it seems to lower quality
453 in most cases due to skips. */
454 for( int ref = 1; ref < h->i_ref[0]; ref++ )
459 for( int i = 1; i < h->i_ref[0]; i++ )
460 /* Favor lower POC as a tiebreaker. */
461 COPY2_IF_GT( max, refcount[i], bestref, i );
463 /* FIXME: If there are duplicates from frames other than ref0 then it is possible
464 * that the optimal ordering doesnt place every duplicate. */
466 refcount[bestref] = -1;
467 h->fref[0][ref] = frames[bestref];
468 memcpy( h->fenc->weight[ref], weights[bestref], sizeof(weights[bestref]) );
474 static char *x264_strcat_filename( char *input, char *suffix )
476 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
479 strcpy( output, input );
480 strcat( output, suffix );
484 void x264_ratecontrol_init_reconfigurable( x264_t *h, int b_init )
486 x264_ratecontrol_t *rc = h->rc;
487 if( !b_init && rc->b_2pass )
490 if( h->param.rc.i_rc_method == X264_RC_CRF )
492 /* Arbitrary rescaling to make CRF somewhat similar to QP.
493 * Try to compensate for MB-tree's effects as well. */
494 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
495 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
496 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
497 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset + QP_BD_OFFSET );
500 if( h->param.rc.i_vbv_max_bitrate > 0 && h->param.rc.i_vbv_buffer_size > 0 )
502 /* We don't support changing the ABR bitrate right now,
503 so if the stream starts as CBR, keep it CBR. */
504 if( rc->b_vbv_min_rate )
505 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
507 if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
509 h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
510 x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
511 h->param.rc.i_vbv_buffer_size );
514 int vbv_buffer_size = h->param.rc.i_vbv_buffer_size * 1000;
515 int vbv_max_bitrate = h->param.rc.i_vbv_max_bitrate * 1000;
518 if( h->param.i_nal_hrd && b_init )
520 h->sps->vui.hrd.i_cpb_cnt = 1;
521 h->sps->vui.hrd.b_cbr_hrd = h->param.i_nal_hrd == X264_NAL_HRD_CBR;
522 h->sps->vui.hrd.i_time_offset_length = 0;
527 int bitrate = 1000*h->param.rc.i_vbv_max_bitrate;
528 int bufsize = 1000*h->param.rc.i_vbv_buffer_size;
530 // normalize HRD size and rate to the value / scale notation
531 h->sps->vui.hrd.i_bit_rate_scale = x264_clip3( x264_ctz( bitrate ) - BR_SHIFT, 0, 15 );
532 h->sps->vui.hrd.i_bit_rate_value = bitrate >> ( h->sps->vui.hrd.i_bit_rate_scale + BR_SHIFT );
533 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 );
534 h->sps->vui.hrd.i_cpb_size_scale = x264_clip3( x264_ctz( bufsize ) - CPB_SHIFT, 0, 15 );
535 h->sps->vui.hrd.i_cpb_size_value = bufsize >> ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
536 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 );
542 #define MAX_DURATION 0.5
544 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 );
545 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;
546 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);
548 h->sps->vui.hrd.i_initial_cpb_removal_delay_length = 2 + x264_clip3( 32 - x264_clz( max_delay ), 4, 22 );
549 h->sps->vui.hrd.i_cpb_removal_delay_length = x264_clip3( 32 - x264_clz( max_cpb_output_delay ), 4, 31 );
550 h->sps->vui.hrd.i_dpb_output_delay_length = x264_clip3( 32 - x264_clz( max_dpb_output_delay ), 4, 31 );
554 vbv_buffer_size = h->sps->vui.hrd.i_cpb_size_unscaled;
555 vbv_max_bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
557 else if( h->param.i_nal_hrd && !b_init )
559 x264_log( h, X264_LOG_WARNING, "VBV parameters cannot be changed when NAL HRD is in use\n" );
562 h->sps->vui.hrd.i_bit_rate_unscaled = vbv_max_bitrate;
563 h->sps->vui.hrd.i_cpb_size_unscaled = vbv_buffer_size;
565 if( rc->b_vbv_min_rate )
566 rc->bitrate = h->param.rc.i_bitrate * 1000.;
567 rc->buffer_rate = vbv_max_bitrate / rc->fps;
568 rc->vbv_max_rate = vbv_max_bitrate;
569 rc->buffer_size = vbv_buffer_size;
570 rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
571 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
572 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
573 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.f_rf_constant_max )
575 rc->rate_factor_max_increment = h->param.rc.f_rf_constant_max - h->param.rc.f_rf_constant;
576 if( rc->rate_factor_max_increment <= 0 )
578 x264_log( h, X264_LOG_WARNING, "CRF max must be greater than CRF\n" );
579 rc->rate_factor_max_increment = 0;
584 if( h->param.rc.f_vbv_buffer_init > 1. )
585 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 );
586 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);
587 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init * h->sps->vui.i_time_scale;
589 rc->b_vbv_min_rate = !rc->b_2pass
590 && h->param.rc.i_rc_method == X264_RC_ABR
591 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
596 int x264_ratecontrol_new( x264_t *h )
598 x264_ratecontrol_t *rc;
602 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
605 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
606 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
608 /* FIXME: use integers */
609 if( h->param.i_fps_num > 0 && h->param.i_fps_den > 0 )
610 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
614 if( h->param.rc.b_mb_tree )
616 h->param.rc.f_pb_factor = 1;
620 rc->qcompress = h->param.rc.f_qcompress;
622 rc->bitrate = h->param.rc.i_bitrate * 1000.;
623 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
624 rc->nmb = h->mb.i_mb_count;
625 rc->last_non_b_pict_type = -1;
628 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
630 x264_log( h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n" );
634 x264_ratecontrol_init_reconfigurable( h, 1 );
636 if( h->param.i_nal_hrd )
638 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale;
639 uint64_t num = 180000;
640 x264_reduce_fraction64( &num, &denom );
641 rc->hrd_multiply_denom = 180000 / num;
643 double bits_required = log2( 180000 / rc->hrd_multiply_denom )
644 + log2( h->sps->vui.i_time_scale )
645 + log2( h->sps->vui.hrd.i_cpb_size_unscaled );
646 if( bits_required >= 63 )
648 x264_log( h, X264_LOG_ERROR, "HRD with very large timescale and bufsize not supported\n" );
653 if( rc->rate_tolerance < 0.01 )
655 x264_log( h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n" );
656 rc->rate_tolerance = 0.01;
659 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
663 /* FIXME ABR_INIT_QP is actually used only in CRF */
664 #define ABR_INIT_QP (( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 ) + QP_BD_OFFSET)
665 rc->accum_p_norm = .01;
666 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
667 /* estimated ratio that produces a reasonable QP for the first I-frame */
668 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
669 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
670 rc->last_non_b_pict_type = SLICE_TYPE_I;
673 rc->ip_offset = 6.0 * log2f( h->param.rc.f_ip_factor );
674 rc->pb_offset = 6.0 * log2f( h->param.rc.f_pb_factor );
675 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
676 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, QP_MAX );
677 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, QP_MAX );
678 h->mb.ip_offset = rc->ip_offset + 0.5;
680 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
681 rc->last_qscale = qp2qscale( 26 );
682 int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
683 CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
684 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
685 for( int i = 0; i < 3; i++ )
687 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
688 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
689 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
690 for( int j = 0; j < num_preds; j++ )
692 rc->pred[i+j*5].coeff= 2.0;
693 rc->pred[i+j*5].count= 1.0;
694 rc->pred[i+j*5].decay= 0.5;
695 rc->pred[i+j*5].offset= 0.0;
697 for( int j = 0; j < 2; j++ )
699 rc->row_preds[i][j].coeff= .25;
700 rc->row_preds[i][j].count= 1.0;
701 rc->row_preds[i][j].decay= 0.5;
702 rc->row_preds[i][j].offset= 0.0;
705 *rc->pred_b_from_p = rc->pred[0];
707 if( parse_zones( h ) < 0 )
709 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
713 /* Load stat file and init 2pass algo */
714 if( h->param.rc.b_stat_read )
716 char *p, *stats_in, *stats_buf;
718 /* read 1st pass stats */
719 assert( h->param.rc.psz_stat_in );
720 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
723 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n" );
726 if( h->param.rc.b_mb_tree )
728 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
729 if( !mbtree_stats_in )
731 rc->p_mbtree_stat_file_in = fopen( mbtree_stats_in, "rb" );
732 x264_free( mbtree_stats_in );
733 if( !rc->p_mbtree_stat_file_in )
735 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n" );
740 /* check whether 1st pass options were compatible with current options */
741 if( strncmp( stats_buf, "#options:", 9 ) )
743 x264_log( h, X264_LOG_ERROR, "options list in stats file not valid\n" );
747 float res_factor, res_factor_bits;
751 char *opts = stats_buf;
752 stats_in = strchr( stats_buf, '\n' );
757 if( sscanf( opts, "#options: %dx%d", &i, &j ) != 2 )
759 x264_log( h, X264_LOG_ERROR, "resolution specified in stats file not valid\n" );
762 else if( h->param.rc.b_mb_tree && (i != h->param.i_width || j != h->param.i_height) )
764 x264_log( h, X264_LOG_ERROR, "MB-tree doesn't support different resolution than 1st pass (%dx%d vs %dx%d)\n",
765 h->param.i_width, h->param.i_height, i, j );
768 res_factor = (float)h->param.i_width * h->param.i_height / (i*j);
769 /* Change in bits relative to resolution isn't quite linear on typical sources,
770 * so we'll at least try to roughly approximate this effect. */
771 res_factor_bits = powf( res_factor, 0.7 );
773 if( ( p = strstr( opts, "timebase=" ) ) && sscanf( p, "timebase=%u/%u", &k, &l ) != 2 )
775 x264_log( h, X264_LOG_ERROR, "timebase specified in stats file not valid\n" );
778 if( k != h->param.i_timebase_num || l != h->param.i_timebase_den )
780 x264_log( h, X264_LOG_ERROR, "timebase mismatch with 1st pass (%u/%u vs %u/%u)\n",
781 h->param.i_timebase_num, h->param.i_timebase_den, k, l );
785 CMP_OPT_FIRST_PASS( "bitdepth", BIT_DEPTH );
786 CMP_OPT_FIRST_PASS( "weightp", X264_MAX( 0, h->param.analyse.i_weighted_pred ) );
787 CMP_OPT_FIRST_PASS( "bframes", h->param.i_bframe );
788 CMP_OPT_FIRST_PASS( "b_pyramid", h->param.i_bframe_pyramid );
789 CMP_OPT_FIRST_PASS( "intra_refresh", h->param.b_intra_refresh );
790 CMP_OPT_FIRST_PASS( "open_gop", h->param.b_open_gop );
791 CMP_OPT_FIRST_PASS( "bluray_compat", h->param.b_bluray_compat );
793 if( (p = strstr( opts, "interlaced=" )) )
795 char *current = h->param.b_interlaced ? h->param.b_tff ? "tff" : "bff" : h->param.b_fake_interlaced ? "fake" : "0";
797 sscanf( p, "interlaced=%4s", buf );
798 if( strcmp( current, buf ) )
800 x264_log( h, X264_LOG_ERROR, "different interlaced setting than first pass (%s vs %s)\n", current, buf );
805 if( (p = strstr( opts, "keyint=" )) )
808 char buf[13] = "infinite ";
809 if( h->param.i_keyint_max != X264_KEYINT_MAX_INFINITE )
810 sprintf( buf, "%d ", h->param.i_keyint_max );
811 if( strncmp( p, buf, strlen(buf) ) )
813 x264_log( h, X264_LOG_ERROR, "different keyint setting than first pass (%.*s vs %.*s)\n",
814 strlen(buf)-1, buf, strcspn(p, " "), p );
819 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
820 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
822 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
824 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
825 h->mb.b_direct_auto_write = 1;
828 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
829 h->param.i_bframe_adaptive = i;
830 else if( h->param.i_bframe )
832 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
836 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 ) )
837 h->param.rc.i_lookahead = i;
840 /* find number of pics */
843 for( num_entries = -1; p; num_entries++ )
844 p = strchr( p + 1, ';' );
847 x264_log( h, X264_LOG_ERROR, "empty stats file\n" );
850 rc->num_entries = num_entries;
852 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
854 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
855 h->param.i_frame_total, rc->num_entries );
857 if( h->param.i_frame_total > rc->num_entries )
859 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
860 h->param.i_frame_total, rc->num_entries );
864 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
866 /* init all to skipped p frames */
867 for( int i = 0; i < rc->num_entries; i++ )
869 ratecontrol_entry_t *rce = &rc->entry[i];
870 rce->pict_type = SLICE_TYPE_P;
871 rce->qscale = rce->new_qscale = qp2qscale( 20 );
872 rce->misc_bits = rc->nmb + 10;
878 for( int i = 0; i < rc->num_entries; i++ )
880 ratecontrol_entry_t *rce;
888 next= strchr(p, ';');
890 *next++ = 0; //sscanf is unbelievably slow on long strings
891 e = sscanf( p, " in:%d ", &frame_number );
893 if( frame_number < 0 || frame_number >= rc->num_entries )
895 x264_log( h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i );
898 rce = &rc->entry[frame_number];
899 rce->direct_mode = 0;
901 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",
902 &pict_type, &rce->i_duration, &rce->i_cpb_duration, &qp, &rce->tex_bits,
903 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
904 &rce->s_count, &rce->direct_mode );
905 rce->tex_bits *= res_factor_bits;
906 rce->mv_bits *= res_factor_bits;
907 rce->misc_bits *= res_factor_bits;
908 rce->i_count *= res_factor;
909 rce->p_count *= res_factor;
910 rce->s_count *= res_factor;
912 p = strstr( p, "ref:" );
916 for( ref = 0; ref < 16; ref++ )
918 if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
920 p = strchr( p+1, ' ' );
927 rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
928 char *w = strchr( p, 'w' );
931 int count = sscanf( w, "w:%hd,%hd,%hd,%hd,%hd,%hd,%hd,%hd",
932 &rce->i_weight_denom[0], &rce->weight[0][0], &rce->weight[0][1],
933 &rce->i_weight_denom[1], &rce->weight[1][0], &rce->weight[1][1],
934 &rce->weight[2][0], &rce->weight[2][1] );
936 rce->i_weight_denom[1] = -1;
937 else if ( count != 8 )
938 rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
941 if( pict_type != 'b' )
942 rce->kept_as_ref = 1;
946 rce->frame_type = X264_TYPE_IDR;
947 rce->pict_type = SLICE_TYPE_I;
950 rce->frame_type = X264_TYPE_I;
951 rce->pict_type = SLICE_TYPE_I;
954 rce->frame_type = X264_TYPE_P;
955 rce->pict_type = SLICE_TYPE_P;
958 rce->frame_type = X264_TYPE_BREF;
959 rce->pict_type = SLICE_TYPE_B;
962 rce->frame_type = X264_TYPE_B;
963 rce->pict_type = SLICE_TYPE_B;
965 default: e = -1; break;
970 x264_log( h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e );
973 rce->qscale = qp2qscale( qp );
977 x264_free( stats_buf );
979 if( h->param.rc.i_rc_method == X264_RC_ABR )
981 if( init_pass2( h ) < 0 )
983 } /* else we're using constant quant, so no need to run the bitrate allocation */
986 /* Open output file */
987 /* If input and output files are the same, output to a temp file
988 * and move it to the real name only when it's complete */
989 if( h->param.rc.b_stat_write )
992 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
993 if( !rc->psz_stat_file_tmpname )
996 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
997 if( rc->p_stat_file_out == NULL )
999 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n" );
1003 p = x264_param2string( &h->param, 1 );
1005 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
1007 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
1009 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
1010 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
1011 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
1014 rc->p_mbtree_stat_file_out = fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
1015 if( rc->p_mbtree_stat_file_out == NULL )
1017 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n" );
1023 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
1025 CHECKED_MALLOC( rc->qp_buffer[0], h->mb.i_mb_count * sizeof(uint16_t) );
1026 if( h->param.i_bframe_pyramid && h->param.rc.b_stat_read )
1027 CHECKED_MALLOC( rc->qp_buffer[1], h->mb.i_mb_count * sizeof(uint16_t) );
1031 for( int i = 0; i<h->param.i_threads; i++ )
1033 h->thread[i]->rc = rc+i;
1037 h->thread[i]->param = h->param;
1038 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
1039 h->thread[i]->mb.ip_offset = h->mb.ip_offset;
1048 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
1051 char *tok, UNUSED *saveptr=NULL;
1053 z->f_bitrate_factor = 1;
1054 if( 3 <= sscanf(p, "%d,%d,q=%d%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
1056 else if( 3 <= sscanf(p, "%d,%d,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
1058 else if( 2 <= sscanf(p, "%d,%d%n", &z->i_start, &z->i_end, &len) )
1062 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
1068 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
1069 memcpy( z->param, &h->param, sizeof(x264_param_t) );
1070 z->param->param_free = x264_free;
1071 while( (tok = strtok_r( p, ",", &saveptr )) )
1073 char *val = strchr( tok, '=' );
1079 if( x264_param_parse( z->param, tok, val ) )
1081 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
1091 static int parse_zones( x264_t *h )
1093 x264_ratecontrol_t *rc = h->rc;
1094 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
1096 char *psz_zones, *p;
1097 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
1098 strcpy( psz_zones, h->param.rc.psz_zones );
1099 h->param.rc.i_zones = 1;
1100 for( p = psz_zones; *p; p++ )
1101 h->param.rc.i_zones += (*p == '/');
1102 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
1104 for( int i = 0; i < h->param.rc.i_zones; i++ )
1106 int i_tok = strcspn( p, "/" );
1108 if( parse_zone( h, &h->param.rc.zones[i], p ) )
1112 x264_free( psz_zones );
1115 if( h->param.rc.i_zones > 0 )
1117 for( int i = 0; i < h->param.rc.i_zones; i++ )
1119 x264_zone_t z = h->param.rc.zones[i];
1120 if( z.i_start < 0 || z.i_start > z.i_end )
1122 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
1123 z.i_start, z.i_end );
1126 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
1128 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
1129 z.f_bitrate_factor );
1134 rc->i_zones = h->param.rc.i_zones + 1;
1135 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
1136 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
1138 // default zone to fall back to if none of the others match
1139 rc->zones[0].i_start = 0;
1140 rc->zones[0].i_end = INT_MAX;
1141 rc->zones[0].b_force_qp = 0;
1142 rc->zones[0].f_bitrate_factor = 1;
1143 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
1144 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
1145 for( int i = 1; i < rc->i_zones; i++ )
1147 if( !rc->zones[i].param )
1148 rc->zones[i].param = rc->zones[0].param;
1157 static x264_zone_t *get_zone( x264_t *h, int frame_num )
1159 for( int i = h->rc->i_zones - 1; i >= 0; i-- )
1161 x264_zone_t *z = &h->rc->zones[i];
1162 if( frame_num >= z->i_start && frame_num <= z->i_end )
1168 void x264_ratecontrol_summary( x264_t *h )
1170 x264_ratecontrol_t *rc = h->rc;
1171 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
1173 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
1174 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
1175 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
1176 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
1177 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset - QP_BD_OFFSET );
1181 void x264_ratecontrol_delete( x264_t *h )
1183 x264_ratecontrol_t *rc = h->rc;
1186 if( rc->p_stat_file_out )
1188 b_regular_file = x264_is_regular_file( rc->p_stat_file_out );
1189 fclose( rc->p_stat_file_out );
1190 if( h->i_frame >= rc->num_entries && b_regular_file )
1191 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
1193 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1194 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
1196 x264_free( rc->psz_stat_file_tmpname );
1198 if( rc->p_mbtree_stat_file_out )
1200 b_regular_file = x264_is_regular_file( rc->p_mbtree_stat_file_out );
1201 fclose( rc->p_mbtree_stat_file_out );
1202 if( h->i_frame >= rc->num_entries && b_regular_file )
1203 if( rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
1205 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1206 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
1208 x264_free( rc->psz_mbtree_stat_file_tmpname );
1209 x264_free( rc->psz_mbtree_stat_file_name );
1211 if( rc->p_mbtree_stat_file_in )
1212 fclose( rc->p_mbtree_stat_file_in );
1213 x264_free( rc->pred );
1214 x264_free( rc->pred_b_from_p );
1215 x264_free( rc->entry );
1216 x264_free( rc->qp_buffer[0] );
1217 x264_free( rc->qp_buffer[1] );
1220 x264_free( rc->zones[0].param );
1221 for( int i = 1; i < rc->i_zones; i++ )
1222 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
1223 rc->zones[i].param->param_free( rc->zones[i].param );
1224 x264_free( rc->zones );
1229 static void accum_p_qp_update( x264_t *h, float qp )
1231 x264_ratecontrol_t *rc = h->rc;
1232 rc->accum_p_qp *= .95;
1233 rc->accum_p_norm *= .95;
1234 rc->accum_p_norm += 1;
1235 if( h->sh.i_type == SLICE_TYPE_I )
1236 rc->accum_p_qp += qp + rc->ip_offset;
1238 rc->accum_p_qp += qp;
1241 /* Before encoding a frame, choose a QP for it */
1242 void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
1244 x264_ratecontrol_t *rc = h->rc;
1245 ratecontrol_entry_t *rce = NULL;
1246 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
1251 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
1252 x264_encoder_reconfig( h, zone->param );
1253 rc->prev_zone = zone;
1255 if( h->param.rc.b_stat_read )
1257 int frame = h->fenc->i_frame;
1258 assert( frame >= 0 && frame < rc->num_entries );
1259 rce = h->rc->rce = &h->rc->entry[frame];
1261 if( h->sh.i_type == SLICE_TYPE_B
1262 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
1264 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
1265 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
1271 memset( h->fdec->i_row_bits, 0, h->mb.i_mb_height * sizeof(int) );
1272 memset( h->fdec->f_row_qp, 0, h->mb.i_mb_height * sizeof(float) );
1273 memset( h->fdec->f_row_qscale, 0, h->mb.i_mb_height * sizeof(float) );
1274 rc->row_pred = &rc->row_preds[h->sh.i_type];
1275 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;
1276 update_vbv_plan( h, overhead );
1278 const x264_level_t *l = x264_levels;
1279 while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
1282 int mincr = l->mincr;
1284 if( h->param.b_bluray_compat )
1287 /* Profiles above High don't require minCR, so just set the maximum to a large value. */
1288 if( h->sps->i_profile_idc > PROFILE_HIGH )
1289 rc->frame_size_maximum = 1e9;
1292 /* The spec has a bizarre special case for the first frame. */
1293 if( h->i_frame == 0 )
1295 //384 * ( Max( PicSizeInMbs, fR * MaxMBPS ) + MaxMBPS * ( tr( 0 ) - tr,n( 0 ) ) ) / MinCR
1296 double fr = 1. / 172;
1297 int pic_size_in_mbs = h->mb.i_mb_width * h->mb.i_mb_height;
1298 rc->frame_size_maximum = 384 * BIT_DEPTH * X264_MAX( pic_size_in_mbs, fr*l->mbps ) / mincr;
1302 //384 * MaxMBPS * ( tr( n ) - tr( n - 1 ) ) / MinCR
1303 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;
1308 if( h->sh.i_type != SLICE_TYPE_B )
1309 rc->bframes = h->fenc->i_bframes;
1313 q = qscale2qp( rate_estimate_qscale( h ) );
1315 else if( rc->b_2pass )
1317 rce->new_qscale = rate_estimate_qscale( h );
1318 q = qscale2qp( rce->new_qscale );
1322 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
1323 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
1325 q = rc->qp_constant[ h->sh.i_type ];
1329 if( zone->b_force_qp )
1330 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1332 q -= 6*log2f( zone->f_bitrate_factor );
1335 if( i_force_qp != X264_QP_AUTO )
1338 q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1340 rc->qpa_rc = rc->qpa_rc_prev =
1341 rc->qpa_aq = rc->qpa_aq_prev = 0;
1342 rc->qp = x264_clip3( q + 0.5f, 0, QP_MAX );
1343 h->fdec->f_qp_avg_rc =
1344 h->fdec->f_qp_avg_aq =
1347 rce->new_qp = rc->qp;
1349 accum_p_qp_update( h, rc->qpm );
1351 if( h->sh.i_type != SLICE_TYPE_B )
1352 rc->last_non_b_pict_type = h->sh.i_type;
1355 static float predict_row_size( x264_t *h, int y, float qscale )
1357 /* average between two predictors:
1358 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1359 x264_ratecontrol_t *rc = h->rc;
1360 float pred_s = predict_size( rc->row_pred[0], qscale, h->fdec->i_row_satd[y] );
1361 if( h->sh.i_type == SLICE_TYPE_I || qscale >= h->fref[0][0]->f_row_qscale[y] )
1363 if( h->sh.i_type == SLICE_TYPE_P
1364 && h->fref[0][0]->i_type == h->fdec->i_type
1365 && h->fref[0][0]->f_row_qscale[y] > 0
1366 && h->fref[0][0]->i_row_satd[y] > 0
1367 && (abs(h->fref[0][0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1369 float pred_t = h->fref[0][0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref[0][0]->i_row_satd[y]
1370 * h->fref[0][0]->f_row_qscale[y] / qscale;
1371 return (pred_s + pred_t) * 0.5f;
1375 /* Our QP is lower than the reference! */
1378 float pred_intra = predict_size( rc->row_pred[1], qscale, h->fdec->i_row_satds[0][0][y] );
1379 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1380 return pred_intra + pred_s;
1384 static int row_bits_so_far( x264_t *h, int y )
1387 for( int i = h->i_threadslice_start; i <= y; i++ )
1388 bits += h->fdec->i_row_bits[i];
1392 static float predict_row_size_sum( x264_t *h, int y, float qp )
1394 float qscale = qp2qscale( qp );
1395 float bits = row_bits_so_far( h, y );
1396 for( int i = y+1; i < h->i_threadslice_end; i++ )
1397 bits += predict_row_size( h, i, qscale );
1402 * eliminate all use of qp in row ratecontrol: make it entirely qscale-based.
1403 * make this function stop being needlessly O(N^2)
1404 * update more often than once per row? */
1405 int x264_ratecontrol_mb( x264_t *h, int bits )
1407 x264_ratecontrol_t *rc = h->rc;
1408 const int y = h->mb.i_mb_y;
1410 h->fdec->i_row_bits[y] += bits;
1411 rc->qpa_aq += h->mb.i_qp;
1413 if( h->mb.i_mb_x != h->mb.i_mb_width - 1 )
1417 rc->qpa_rc += rc->qpm * h->mb.i_mb_width;
1422 float qscale = qp2qscale( rc->qpm );
1423 h->fdec->f_row_qp[y] = rc->qpm;
1424 h->fdec->f_row_qscale[y] = qscale;
1426 update_predictor( rc->row_pred[0], qscale, h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1427 if( h->sh.i_type == SLICE_TYPE_P && rc->qpm < h->fref[0][0]->f_row_qp[y] )
1428 update_predictor( rc->row_pred[1], qscale, h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1430 /* update ratecontrol per-mbpair in MBAFF */
1431 if( SLICE_MBAFF && !(y&1) )
1434 /* FIXME: We don't currently support the case where there's a slice
1435 * boundary in between. */
1436 int can_reencode_row = h->sh.i_first_mb <= ((h->mb.i_mb_y - SLICE_MBAFF) * h->mb.i_mb_stride);
1438 /* tweak quality based on difference from predicted size */
1439 float prev_row_qp = h->fdec->f_row_qp[y];
1440 float qp_absolute_max = h->param.rc.i_qp_max;
1441 if( rc->rate_factor_max_increment )
1442 qp_absolute_max = X264_MIN( qp_absolute_max, rc->qp_novbv + rc->rate_factor_max_increment );
1443 float qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, qp_absolute_max );
1444 float qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1445 float step_size = 0.5f;
1446 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1447 float slice_size_planned = h->param.b_sliced_threads ? rc->slice_size_planned : rc->frame_size_planned;
1448 float max_frame_error = X264_MAX( 0.05f, 1.0f / h->mb.i_mb_height );
1449 float size_of_other_slices = 0;
1450 if( h->param.b_sliced_threads )
1452 float size_of_other_slices_planned = 0;
1453 for( int i = 0; i < h->param.i_threads; i++ )
1454 if( h != h->thread[i] )
1456 size_of_other_slices += h->thread[i]->rc->frame_size_estimated;
1457 size_of_other_slices_planned += h->thread[i]->rc->slice_size_planned;
1459 float weight = rc->slice_size_planned / rc->frame_size_planned;
1460 size_of_other_slices = (size_of_other_slices - size_of_other_slices_planned) * weight + size_of_other_slices_planned;
1462 if( y < h->i_threadslice_end-1 )
1464 /* B-frames shouldn't use lower QP than their reference frames. */
1465 if( h->sh.i_type == SLICE_TYPE_B )
1467 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] ) );
1468 rc->qpm = X264_MAX( rc->qpm, qp_min );
1471 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1472 float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
1473 float b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1475 /* Don't increase the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1476 /* area at the top of the frame was measured inaccurately. */
1477 if( row_bits_so_far( h, y ) < 0.05f * slice_size_planned )
1478 qp_max = qp_absolute_max = prev_row_qp;
1480 if( h->sh.i_type != SLICE_TYPE_I )
1483 if( !rc->b_vbv_min_rate )
1484 qp_min = X264_MAX( qp_min, rc->qp_novbv );
1486 while( rc->qpm < qp_max
1487 && ((b1 > rc->frame_size_planned + rc_tol) ||
1488 (rc->buffer_fill - b1 < buffer_left_planned * 0.5f) ||
1489 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1491 rc->qpm += step_size;
1492 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1495 while( rc->qpm > qp_min
1496 && (rc->qpm > h->fdec->f_row_qp[0] || rc->single_frame_vbv)
1497 && ((b1 < rc->frame_size_planned * 0.8f && rc->qpm <= prev_row_qp)
1498 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1f) )
1500 rc->qpm -= step_size;
1501 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1504 /* avoid VBV underflow or MinCR violation */
1505 while( (rc->qpm < qp_absolute_max)
1506 && ((rc->buffer_fill - b1 < rc->buffer_rate * max_frame_error) ||
1507 (rc->frame_size_maximum - b1 < rc->frame_size_maximum * max_frame_error)))
1509 rc->qpm += step_size;
1510 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1513 h->rc->frame_size_estimated = b1 - size_of_other_slices;
1515 /* If the current row was large enough to cause a large QP jump, try re-encoding it. */
1516 if( rc->qpm > qp_max && prev_row_qp < qp_max && can_reencode_row )
1518 /* Bump QP to halfway in between... close enough. */
1519 rc->qpm = x264_clip3f( (prev_row_qp + rc->qpm)*0.5f, prev_row_qp + 1.0f, qp_max );
1520 rc->qpa_rc = rc->qpa_rc_prev;
1521 rc->qpa_aq = rc->qpa_aq_prev;
1522 h->fdec->i_row_bits[y] = h->fdec->i_row_bits[y-SLICE_MBAFF] = 0;
1528 h->rc->frame_size_estimated = predict_row_size_sum( h, y, rc->qpm );
1530 /* Last-ditch attempt: if the last row of the frame underflowed the VBV,
1532 if( (h->rc->frame_size_estimated + size_of_other_slices) > (rc->buffer_fill - rc->buffer_rate * max_frame_error) &&
1533 rc->qpm < qp_max && can_reencode_row )
1536 rc->qpa_rc = rc->qpa_rc_prev;
1537 rc->qpa_aq = rc->qpa_aq_prev;
1538 h->fdec->i_row_bits[y] = h->fdec->i_row_bits[y-SLICE_MBAFF] = 0;
1543 rc->qpa_rc_prev = rc->qpa_rc;
1544 rc->qpa_aq_prev = rc->qpa_aq;
1549 int x264_ratecontrol_qp( x264_t *h )
1552 return x264_clip3( h->rc->qpm + 0.5f, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1555 int x264_ratecontrol_mb_qp( x264_t *h )
1558 float qp = h->rc->qpm;
1559 if( h->param.rc.i_aq_mode )
1561 /* MB-tree currently doesn't adjust quantizers in unreferenced frames. */
1562 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];
1563 /* Scale AQ's effect towards zero in emergency mode. */
1564 if( qp > QP_MAX_SPEC )
1565 qp_offset *= (QP_MAX - qp) / (QP_MAX - QP_MAX_SPEC);
1568 return x264_clip3( qp + 0.5f, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1571 /* In 2pass, force the same frame types as in the 1st pass */
1572 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1574 x264_ratecontrol_t *rc = h->rc;
1575 if( h->param.rc.b_stat_read )
1577 if( frame_num >= rc->num_entries )
1579 /* We could try to initialize everything required for ABR and
1580 * adaptive B-frames, but that would be complicated.
1581 * So just calculate the average QP used so far. */
1582 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24 + QP_BD_OFFSET
1583 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1584 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, QP_MAX );
1585 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 );
1586 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 );
1588 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries );
1589 x264_log( h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant );
1590 if( h->param.i_bframe_adaptive )
1591 x264_log( h, X264_LOG_ERROR, "disabling adaptive B-frames\n" );
1593 for( int i = 0; i < h->param.i_threads; i++ )
1595 h->thread[i]->rc->b_abr = 0;
1596 h->thread[i]->rc->b_2pass = 0;
1597 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1598 h->thread[i]->param.rc.b_stat_read = 0;
1599 h->thread[i]->param.i_bframe_adaptive = 0;
1600 h->thread[i]->param.i_scenecut_threshold = 0;
1601 h->thread[i]->param.rc.b_mb_tree = 0;
1602 if( h->thread[i]->param.i_bframe > 1 )
1603 h->thread[i]->param.i_bframe = 1;
1605 return X264_TYPE_AUTO;
1607 return rc->entry[frame_num].frame_type;
1610 return X264_TYPE_AUTO;
1613 void x264_ratecontrol_set_weights( x264_t *h, x264_frame_t *frm )
1615 ratecontrol_entry_t *rce = &h->rc->entry[frm->i_frame];
1616 if( h->param.analyse.i_weighted_pred <= 0 )
1619 if( rce->i_weight_denom[0] >= 0 )
1620 SET_WEIGHT( frm->weight[0][0], 1, rce->weight[0][0], rce->i_weight_denom[0], rce->weight[0][1] );
1622 if( rce->i_weight_denom[1] >= 0 )
1624 SET_WEIGHT( frm->weight[0][1], 1, rce->weight[1][0], rce->i_weight_denom[1], rce->weight[1][1] );
1625 SET_WEIGHT( frm->weight[0][2], 1, rce->weight[2][0], rce->i_weight_denom[1], rce->weight[2][1] );
1629 /* After encoding one frame, save stats and update ratecontrol state */
1630 int x264_ratecontrol_end( x264_t *h, int bits, int *filler )
1632 x264_ratecontrol_t *rc = h->rc;
1633 const int *mbs = h->stat.frame.i_mb_count;
1637 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1638 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1639 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1640 for( int i = B_DIRECT; i < B_8x8; i++ )
1641 h->stat.frame.i_mb_count_p += mbs[i];
1643 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1644 h->fdec->f_qp_avg_aq = (float)rc->qpa_aq / h->mb.i_mb_count;
1646 if( h->param.rc.b_stat_write )
1648 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1649 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1650 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1651 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1652 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1653 char c_direct = h->mb.b_direct_auto_write ?
1654 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1655 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1657 if( fprintf( rc->p_stat_file_out,
1658 "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:",
1659 h->fenc->i_frame, h->i_frame,
1660 c_type, h->fenc->i_duration,
1661 h->fenc->i_cpb_duration, rc->qpa_rc,
1662 h->stat.frame.i_tex_bits,
1663 h->stat.frame.i_mv_bits,
1664 h->stat.frame.i_misc_bits,
1665 h->stat.frame.i_mb_count_i,
1666 h->stat.frame.i_mb_count_p,
1667 h->stat.frame.i_mb_count_skip,
1671 /* Only write information for reference reordering once. */
1672 int use_old_stats = h->param.rc.b_stat_read && rc->rce->refs > 1;
1673 for( int i = 0; i < (use_old_stats ? rc->rce->refs : h->i_ref[0]); i++ )
1675 int refcount = use_old_stats ? rc->rce->refcount[i]
1676 : PARAM_INTERLACED ? h->stat.frame.i_mb_count_ref[0][i*2]
1677 + h->stat.frame.i_mb_count_ref[0][i*2+1]
1678 : h->stat.frame.i_mb_count_ref[0][i];
1679 if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1683 if( h->param.analyse.i_weighted_pred >= X264_WEIGHTP_SIMPLE && h->sh.weight[0][0].weightfn )
1685 if( fprintf( rc->p_stat_file_out, "w:%d,%d,%d",
1686 h->sh.weight[0][0].i_denom, h->sh.weight[0][0].i_scale, h->sh.weight[0][0].i_offset ) < 0 )
1688 if( h->sh.weight[0][1].weightfn || h->sh.weight[0][2].weightfn )
1690 if( fprintf( rc->p_stat_file_out, ",%d,%d,%d,%d,%d ",
1691 h->sh.weight[0][1].i_denom, h->sh.weight[0][1].i_scale, h->sh.weight[0][1].i_offset,
1692 h->sh.weight[0][2].i_scale, h->sh.weight[0][2].i_offset ) < 0 )
1695 else if( fprintf( rc->p_stat_file_out, " " ) < 0 )
1699 if( fprintf( rc->p_stat_file_out, ";\n") < 0 )
1702 /* Don't re-write the data in multi-pass mode. */
1703 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1705 uint8_t i_type = h->sh.i_type;
1706 /* Values are stored as big-endian FIX8.8 */
1707 for( int i = 0; i < h->mb.i_mb_count; i++ )
1708 rc->qp_buffer[0][i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1709 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1711 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 )
1718 if( h->sh.i_type != SLICE_TYPE_B )
1719 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / rc->last_rceq;
1722 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1723 * Not perfectly accurate with B-refs, but good enough. */
1724 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / (rc->last_rceq * fabs( h->param.rc.f_pb_factor ));
1726 rc->cplxr_sum *= rc->cbr_decay;
1727 rc->wanted_bits_window += h->fenc->f_duration * rc->bitrate;
1728 rc->wanted_bits_window *= rc->cbr_decay;
1732 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale( rc->rce->new_qp ) );
1734 if( h->mb.b_variable_qp )
1736 if( h->sh.i_type == SLICE_TYPE_B )
1738 rc->bframe_bits += bits;
1739 if( h->fenc->b_last_minigop_bframe )
1741 update_predictor( rc->pred_b_from_p, qp2qscale( rc->qpa_rc ),
1742 h->fref[1][h->i_ref[1]-1]->i_satd, rc->bframe_bits / rc->bframes );
1743 rc->bframe_bits = 0;
1748 *filler = update_vbv( h, bits );
1749 rc->filler_bits_sum += *filler * 8;
1751 if( h->sps->vui.b_nal_hrd_parameters_present )
1753 if( h->fenc->i_frame == 0 )
1755 // access unit initialises the HRD
1756 h->fenc->hrd_timing.cpb_initial_arrival_time = 0;
1757 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1758 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1759 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit = (double)rc->initial_cpb_removal_delay / 90000;
1763 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit + (double)(h->fenc->i_cpb_delay - h->i_cpb_delay_pir_offset) *
1764 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1766 double cpb_earliest_arrival_time = h->fenc->hrd_timing.cpb_removal_time - (double)rc->initial_cpb_removal_delay / 90000;
1767 if( h->fenc->b_keyframe )
1769 rc->nrt_first_access_unit = h->fenc->hrd_timing.cpb_removal_time;
1770 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1771 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1774 cpb_earliest_arrival_time -= (double)rc->initial_cpb_removal_delay_offset / 90000;
1776 if( h->sps->vui.hrd.b_cbr_hrd )
1777 h->fenc->hrd_timing.cpb_initial_arrival_time = rc->previous_cpb_final_arrival_time;
1779 h->fenc->hrd_timing.cpb_initial_arrival_time = X264_MAX( rc->previous_cpb_final_arrival_time, cpb_earliest_arrival_time );
1781 int filler_bits = *filler ? X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), *filler )*8 : 0;
1783 h->fenc->hrd_timing.cpb_final_arrival_time = rc->previous_cpb_final_arrival_time = h->fenc->hrd_timing.cpb_initial_arrival_time +
1784 (double)(bits + filler_bits) / h->sps->vui.hrd.i_bit_rate_unscaled;
1786 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 +
1787 h->fenc->hrd_timing.cpb_removal_time;
1792 x264_log( h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n" );
1796 /****************************************************************************
1798 ***************************************************************************/
1801 * modify the bitrate curve from pass1 for one frame
1803 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1805 x264_ratecontrol_t *rcc= h->rc;
1806 x264_zone_t *zone = get_zone( h, frame_num );
1808 if( h->param.rc.b_mb_tree )
1810 double timescale = (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1811 q = pow( BASE_FRAME_DURATION / CLIP_DURATION(rce->i_duration * timescale), 1 - h->param.rc.f_qcompress );
1814 q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1816 // avoid NaN's in the rc_eq
1817 if( !isfinite(q) || rce->tex_bits + rce->mv_bits == 0 )
1818 q = rcc->last_qscale_for[rce->pict_type];
1823 rcc->last_qscale = q;
1828 if( zone->b_force_qp )
1829 q = qp2qscale( zone->i_qp );
1831 q /= zone->f_bitrate_factor;
1837 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q, int frame_num)
1839 x264_ratecontrol_t *rcc = h->rc;
1840 const int pict_type = rce->pict_type;
1841 x264_zone_t *zone = get_zone( h, frame_num );
1843 // force I/B quants as a function of P quants
1844 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1845 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1846 if( pict_type == SLICE_TYPE_I )
1849 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1850 double ip_factor = fabs( h->param.rc.f_ip_factor );
1851 /* don't apply ip_factor if the following frame is also I */
1852 if( rcc->accum_p_norm <= 0 )
1854 else if( h->param.rc.f_ip_factor < 0 )
1856 else if( rcc->accum_p_norm >= 1 )
1859 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1861 else if( pict_type == SLICE_TYPE_B )
1863 if( h->param.rc.f_pb_factor > 0 )
1865 if( !rce->kept_as_ref )
1866 q *= fabs( h->param.rc.f_pb_factor );
1868 else if( pict_type == SLICE_TYPE_P
1869 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1870 && rce->tex_bits == 0 )
1875 /* last qscale / qdiff stuff */
1876 if( rcc->last_non_b_pict_type == pict_type &&
1877 (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1) )
1879 double last_q = rcc->last_qscale_for[pict_type];
1880 double max_qscale = last_q * rcc->lstep;
1881 double min_qscale = last_q / rcc->lstep;
1883 if ( q > max_qscale ) q = max_qscale;
1884 else if( q < min_qscale ) q = min_qscale;
1887 rcc->last_qscale_for[pict_type] = q;
1888 if( pict_type != SLICE_TYPE_B )
1889 rcc->last_non_b_pict_type = pict_type;
1890 if( pict_type == SLICE_TYPE_I )
1892 rcc->last_accum_p_norm = rcc->accum_p_norm;
1893 rcc->accum_p_norm = 0;
1894 rcc->accum_p_qp = 0;
1896 if( pict_type == SLICE_TYPE_P )
1898 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1899 rcc->accum_p_qp = mask * (qscale2qp( q ) + rcc->accum_p_qp);
1900 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1905 if( zone->b_force_qp )
1906 q = qp2qscale( zone->i_qp );
1908 q /= zone->f_bitrate_factor;
1914 static float predict_size( predictor_t *p, float q, float var )
1916 return (p->coeff*var + p->offset) / (q*p->count);
1919 static void update_predictor( predictor_t *p, float q, float var, float bits )
1924 float old_coeff = p->coeff / p->count;
1925 float new_coeff = bits*q / var;
1926 float new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
1927 float new_offset = bits*q - new_coeff_clipped * var;
1928 if( new_offset >= 0 )
1929 new_coeff = new_coeff_clipped;
1932 p->count *= p->decay;
1933 p->coeff *= p->decay;
1934 p->offset *= p->decay;
1936 p->coeff += new_coeff;
1937 p->offset += new_offset;
1940 // update VBV after encoding a frame
1941 static int update_vbv( x264_t *h, int bits )
1944 int bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
1945 x264_ratecontrol_t *rcc = h->rc;
1946 x264_ratecontrol_t *rct = h->thread[0]->rc;
1947 uint64_t buffer_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
1949 if( rcc->last_satd >= h->mb.i_mb_count )
1950 update_predictor( &rct->pred[h->sh.i_type], qp2qscale( rcc->qpa_rc ), rcc->last_satd, bits );
1955 rct->buffer_fill_final -= (uint64_t)bits * h->sps->vui.i_time_scale;
1957 if( rct->buffer_fill_final < 0 )
1958 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 );
1959 rct->buffer_fill_final = X264_MAX( rct->buffer_fill_final, 0 );
1960 rct->buffer_fill_final += (uint64_t)bitrate * h->sps->vui.i_num_units_in_tick * h->fenc->i_cpb_duration;
1962 if( h->sps->vui.hrd.b_cbr_hrd && rct->buffer_fill_final > buffer_size )
1964 int64_t scale = (int64_t)h->sps->vui.i_time_scale * 8;
1965 filler = (rct->buffer_fill_final - buffer_size + scale - 1) / scale;
1966 bits = X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), filler ) * 8;
1967 rct->buffer_fill_final -= (uint64_t)bits * h->sps->vui.i_time_scale;
1970 rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, buffer_size );
1975 void x264_hrd_fullness( x264_t *h )
1977 x264_ratecontrol_t *rct = h->thread[0]->rc;
1978 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale / rct->hrd_multiply_denom;
1979 uint64_t cpb_state = rct->buffer_fill_final;
1980 uint64_t cpb_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
1981 uint64_t multiply_factor = 180000 / rct->hrd_multiply_denom;
1983 if( rct->buffer_fill_final < 0 || rct->buffer_fill_final > cpb_size )
1985 x264_log( h, X264_LOG_WARNING, "CPB %s: %.0lf bits in a %.0lf-bit buffer\n",
1986 rct->buffer_fill_final < 0 ? "underflow" : "overflow", (float)rct->buffer_fill_final/denom, (float)cpb_size/denom );
1989 h->initial_cpb_removal_delay = (multiply_factor * cpb_state + denom) / (2*denom);
1990 h->initial_cpb_removal_delay_offset = (multiply_factor * cpb_size + denom) / (2*denom) - h->initial_cpb_removal_delay;
1993 // provisionally update VBV according to the planned size of all frames currently in progress
1994 static void update_vbv_plan( x264_t *h, int overhead )
1996 x264_ratecontrol_t *rcc = h->rc;
1997 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final / h->sps->vui.i_time_scale;
1998 if( h->i_thread_frames > 1 )
2000 int j = h->rc - h->thread[0]->rc;
2001 for( int i = 1; i < h->i_thread_frames; i++ )
2003 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
2004 double bits = t->rc->frame_size_planned;
2005 if( !t->b_thread_active )
2007 bits = X264_MAX(bits, t->rc->frame_size_estimated);
2008 rcc->buffer_fill -= bits;
2009 rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
2010 rcc->buffer_fill += t->rc->buffer_rate;
2011 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
2014 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
2015 rcc->buffer_fill -= overhead;
2018 // apply VBV constraints and clip qscale to between lmin and lmax
2019 static double clip_qscale( x264_t *h, int pict_type, double q )
2021 x264_ratecontrol_t *rcc = h->rc;
2022 double lmin = rcc->lmin[pict_type];
2023 double lmax = rcc->lmax[pict_type];
2024 if( rcc->rate_factor_max_increment )
2025 lmax = X264_MIN( lmax, qp2qscale( rcc->qp_novbv + rcc->rate_factor_max_increment ) );
2028 /* B-frames are not directly subject to VBV,
2029 * since they are controlled by the P-frames' QPs. */
2031 if( rcc->b_vbv && rcc->last_satd > 0 )
2033 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
2034 * the lookahead overflow and such that the buffer is in a reasonable state
2035 * by the end of the lookahead. */
2036 if( h->param.rc.i_lookahead )
2040 /* Avoid an infinite loop. */
2041 for( int iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
2044 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2045 double buffer_fill_cur = rcc->buffer_fill - cur_bits;
2047 double total_duration = 0;
2048 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
2049 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
2050 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
2052 /* Loop over the planned future frames. */
2053 for( int j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
2055 total_duration += h->fenc->f_planned_cpb_duration[j];
2056 buffer_fill_cur += rcc->vbv_max_rate * h->fenc->f_planned_cpb_duration[j];
2057 int i_type = h->fenc->i_planned_type[j];
2058 int i_satd = h->fenc->i_planned_satd[j];
2059 if( i_type == X264_TYPE_AUTO )
2061 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
2062 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
2063 buffer_fill_cur -= cur_bits;
2065 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
2066 target_fill = X264_MIN( rcc->buffer_fill + total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.5 );
2067 if( buffer_fill_cur < target_fill )
2073 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
2074 target_fill = x264_clip3f( rcc->buffer_fill - total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
2075 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
2084 /* Fallback to old purely-reactive algorithm: no lookahead. */
2087 if( ( pict_type == SLICE_TYPE_P ||
2088 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
2089 rcc->buffer_fill/rcc->buffer_size < 0.5 )
2091 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
2094 /* Now a hard threshold to make sure the frame fits in VBV.
2095 * This one is mostly for I-frames. */
2096 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2098 /* For small VBVs, allow the frame to use up the entire VBV. */
2099 double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
2100 /* For single-frame VBVs, request that the frame use up the entire VBV. */
2101 double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
2103 if( bits > rcc->buffer_fill/max_fill_factor )
2104 qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
2107 if( bits < rcc->buffer_rate/min_fill_factor )
2108 q *= bits*min_fill_factor/rcc->buffer_rate;
2109 q = X264_MAX( q0, q );
2112 /* Apply MinCR restrictions */
2113 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2114 if( bits > rcc->frame_size_maximum )
2115 q *= bits / rcc->frame_size_maximum;
2116 bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2118 /* Check B-frame complexity, and use up any bits that would
2119 * overflow before the next P-frame. */
2120 if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
2122 int nb = rcc->bframes;
2123 double pbbits = bits;
2124 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
2126 double bframe_cpb_duration = 0;
2127 double minigop_cpb_duration;
2128 for( int i = 0; i < nb; i++ )
2129 bframe_cpb_duration += h->fenc->f_planned_cpb_duration[1+i];
2131 if( bbits * nb > bframe_cpb_duration * rcc->vbv_max_rate )
2133 pbbits += nb * bbits;
2135 minigop_cpb_duration = bframe_cpb_duration + h->fenc->f_planned_cpb_duration[0];
2136 space = rcc->buffer_fill + minigop_cpb_duration*rcc->vbv_max_rate - rcc->buffer_size;
2137 if( pbbits < space )
2139 q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
2141 q = X264_MAX( q0/2, q );
2144 if( !rcc->b_vbv_min_rate )
2145 q = X264_MAX( q0, q );
2150 else if( rcc->b_2pass )
2152 double min2 = log( lmin );
2153 double max2 = log( lmax );
2154 q = (log(q) - min2)/(max2-min2) - 0.5;
2155 q = 1.0/(1.0 + exp( -4*q ));
2156 q = q*(max2-min2) + min2;
2160 return x264_clip3f( q, lmin, lmax );
2163 // update qscale for 1 frame based on actual bits used so far
2164 static float rate_estimate_qscale( x264_t *h )
2167 x264_ratecontrol_t *rcc = h->rc;
2168 ratecontrol_entry_t UNINIT(rce);
2169 int pict_type = h->sh.i_type;
2170 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
2171 + h->stat.i_frame_size[SLICE_TYPE_P]
2172 + h->stat.i_frame_size[SLICE_TYPE_B])
2173 - rcc->filler_bits_sum;
2178 if( pict_type != rce.pict_type )
2180 x264_log( h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
2181 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type] );
2185 if( pict_type == SLICE_TYPE_B )
2187 /* B-frames don't have independent ratecontrol, but rather get the
2188 * average QP of the two adjacent P-frames + an offset */
2190 int i0 = IS_X264_TYPE_I(h->fref_nearest[0]->i_type);
2191 int i1 = IS_X264_TYPE_I(h->fref_nearest[1]->i_type);
2192 int dt0 = abs(h->fenc->i_poc - h->fref_nearest[0]->i_poc);
2193 int dt1 = abs(h->fenc->i_poc - h->fref_nearest[1]->i_poc);
2194 float q0 = h->fref_nearest[0]->f_qp_avg_rc;
2195 float q1 = h->fref_nearest[1]->f_qp_avg_rc;
2197 if( h->fref_nearest[0]->i_type == X264_TYPE_BREF )
2198 q0 -= rcc->pb_offset/2;
2199 if( h->fref_nearest[1]->i_type == X264_TYPE_BREF )
2200 q1 -= rcc->pb_offset/2;
2203 q = (q0 + q1) / 2 + rcc->ip_offset;
2209 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
2211 if( h->fenc->b_kept_as_ref )
2212 q += rcc->pb_offset/2;
2214 q += rcc->pb_offset;
2216 if( rcc->b_2pass && rcc->b_vbv )
2217 rcc->frame_size_planned = qscale2bits( &rce, qp2qscale( q ) );
2219 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, qp2qscale( q ), h->fref[1][h->i_ref[1]-1]->i_satd );
2220 /* Limit planned size by MinCR */
2222 rcc->frame_size_planned = X264_MIN( rcc->frame_size_planned, rcc->frame_size_maximum );
2223 h->rc->frame_size_estimated = rcc->frame_size_planned;
2227 rcc->last_satd = x264_rc_analyse_slice( h );
2229 return qp2qscale( q );
2233 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
2237 double lmin = rcc->lmin[pict_type];
2238 double lmax = rcc->lmax[pict_type];
2240 int64_t predicted_bits = total_bits;
2244 if( h->i_thread_frames > 1 )
2246 int j = h->rc - h->thread[0]->rc;
2247 for( int i = 1; i < h->i_thread_frames; i++ )
2249 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
2250 double bits = t->rc->frame_size_planned;
2251 if( !t->b_thread_active )
2253 bits = X264_MAX(bits, t->rc->frame_size_estimated);
2254 predicted_bits += (int64_t)bits;
2260 if( h->i_frame < h->i_thread_frames )
2261 predicted_bits += (int64_t)h->i_frame * rcc->bitrate / rcc->fps;
2263 predicted_bits += (int64_t)(h->i_thread_frames - 1) * rcc->bitrate / rcc->fps;
2266 /* Adjust ABR buffer based on distance to the end of the video. */
2267 if( rcc->num_entries > h->i_frame )
2269 double final_bits = rcc->entry[rcc->num_entries-1].expected_bits;
2270 double video_pos = rce.expected_bits / final_bits;
2271 double scale_factor = sqrt( (1 - video_pos) * rcc->num_entries );
2272 abr_buffer *= 0.5 * X264_MAX( scale_factor, 0.5 );
2275 diff = predicted_bits - (int64_t)rce.expected_bits;
2277 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
2278 if( ((h->i_frame + 1 - h->i_thread_frames) >= rcc->fps) &&
2279 (rcc->expected_bits_sum > 0))
2281 /* Adjust quant based on the difference between
2282 * achieved and expected bitrate so far */
2283 double cur_time = (double)h->i_frame / rcc->num_entries;
2284 double w = x264_clip3f( cur_time*100, 0.0, 1.0 );
2285 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
2289 /* Do not overflow vbv */
2290 double expected_size = qscale2bits( &rce, q );
2291 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2292 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
2293 double qmax = q*(2 - expected_fullness);
2294 double size_constraint = 1 + expected_fullness;
2295 qmax = X264_MAX( qmax, rce.new_qscale );
2296 if( expected_fullness < .05 )
2298 qmax = X264_MIN(qmax, lmax);
2299 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
2300 ((expected_vbv < 0) && (q < lmax)))
2303 expected_size = qscale2bits(&rce, q);
2304 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2306 rcc->last_satd = x264_rc_analyse_slice( h );
2308 q = x264_clip3f( q, lmin, lmax );
2310 else /* 1pass ABR */
2312 /* Calculate the quantizer which would have produced the desired
2313 * average bitrate if it had been applied to all frames so far.
2314 * Then modulate that quant based on the current frame's complexity
2315 * relative to the average complexity so far (using the 2pass RCEQ).
2316 * Then bias the quant up or down if total size so far was far from
2318 * Result: Depending on the value of rate_tolerance, there is a
2319 * tradeoff between quality and bitrate precision. But at large
2320 * tolerances, the bit distribution approaches that of 2pass. */
2322 double wanted_bits, overflow = 1;
2324 rcc->last_satd = x264_rc_analyse_slice( h );
2325 rcc->short_term_cplxsum *= 0.5;
2326 rcc->short_term_cplxcount *= 0.5;
2327 rcc->short_term_cplxsum += rcc->last_satd / (CLIP_DURATION(h->fenc->f_duration) / BASE_FRAME_DURATION);
2328 rcc->short_term_cplxcount ++;
2330 rce.tex_bits = rcc->last_satd;
2331 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
2333 rce.p_count = rcc->nmb;
2337 rce.pict_type = pict_type;
2338 rce.i_duration = h->fenc->i_duration;
2340 if( h->param.rc.i_rc_method == X264_RC_CRF )
2342 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
2346 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
2348 /* ABR code can potentially be counterproductive in CBR, so just don't bother.
2349 * Don't run it if the frame complexity is zero either. */
2350 if( !rcc->b_vbv_min_rate && rcc->last_satd )
2352 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
2353 int i_frame_done = h->i_frame + 1 - h->i_thread_frames;
2354 double time_done = i_frame_done / rcc->fps;
2355 if( h->param.b_vfr_input && i_frame_done > 0 )
2356 time_done = ((double)(h->fenc->i_reordered_pts - h->i_reordered_pts_delay)) * h->param.i_timebase_num / h->param.i_timebase_den;
2357 wanted_bits = time_done * rcc->bitrate;
2358 if( wanted_bits > 0 )
2360 abr_buffer *= X264_MAX( 1, sqrt( time_done ) );
2361 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
2367 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
2368 /* should test _next_ pict type, but that isn't decided yet */
2369 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
2371 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2372 q /= fabs( h->param.rc.f_ip_factor );
2374 else if( h->i_frame > 0 )
2376 if( h->param.rc.i_rc_method != X264_RC_CRF )
2378 /* Asymmetric clipping, because symmetric would prevent
2379 * overflow control in areas of rapidly oscillating complexity */
2380 double lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
2381 double lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
2382 if( overflow > 1.1 && h->i_frame > 3 )
2384 else if( overflow < 0.9 )
2387 q = x264_clip3f(q, lmin, lmax);
2390 else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
2392 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
2394 rcc->qp_novbv = qscale2qp( q );
2396 //FIXME use get_diff_limited_q() ?
2397 q = clip_qscale( h, pict_type, q );
2400 rcc->last_qscale_for[pict_type] =
2401 rcc->last_qscale = q;
2403 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
2404 rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );
2406 if( rcc->b_2pass && rcc->b_vbv )
2407 rcc->frame_size_planned = qscale2bits(&rce, q);
2409 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2411 /* Always use up the whole VBV in this case. */
2412 if( rcc->single_frame_vbv )
2413 rcc->frame_size_planned = rcc->buffer_rate;
2414 /* Limit planned size by MinCR */
2416 rcc->frame_size_planned = X264_MIN( rcc->frame_size_planned, rcc->frame_size_maximum );
2417 h->rc->frame_size_estimated = rcc->frame_size_planned;
2422 static void x264_threads_normalize_predictors( x264_t *h )
2424 double totalsize = 0;
2425 for( int i = 0; i < h->param.i_threads; i++ )
2426 totalsize += h->thread[i]->rc->slice_size_planned;
2427 double factor = h->rc->frame_size_planned / totalsize;
2428 for( int i = 0; i < h->param.i_threads; i++ )
2429 h->thread[i]->rc->slice_size_planned *= factor;
2432 void x264_threads_distribute_ratecontrol( x264_t *h )
2435 x264_ratecontrol_t *rc = h->rc;
2437 /* Initialize row predictors */
2438 if( h->i_frame == 0 )
2439 for( int i = 0; i < h->param.i_threads; i++ )
2441 x264_ratecontrol_t *t = h->thread[i]->rc;
2442 memcpy( t->row_preds, rc->row_preds, sizeof(rc->row_preds) );
2445 for( int i = 0; i < h->param.i_threads; i++ )
2447 x264_t *t = h->thread[i];
2448 memcpy( t->rc, rc, offsetof(x264_ratecontrol_t, row_pred) );
2449 t->rc->row_pred = &t->rc->row_preds[h->sh.i_type];
2450 /* Calculate the planned slice size. */
2451 if( rc->b_vbv && rc->frame_size_planned )
2454 for( row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2455 size += h->fdec->i_row_satd[row];
2456 t->rc->slice_size_planned = predict_size( &rc->pred[h->sh.i_type + (i+1)*5], rc->qpm, size );
2459 t->rc->slice_size_planned = 0;
2461 if( rc->b_vbv && rc->frame_size_planned )
2463 x264_threads_normalize_predictors( h );
2465 if( rc->single_frame_vbv )
2467 /* Compensate for our max frame error threshold: give more bits (proportionally) to smaller slices. */
2468 for( int i = 0; i < h->param.i_threads; i++ )
2470 x264_t *t = h->thread[i];
2471 float max_frame_error = X264_MAX( 0.05, 1.0 / (t->i_threadslice_end - t->i_threadslice_start) );
2472 t->rc->slice_size_planned += 2 * max_frame_error * rc->frame_size_planned;
2474 x264_threads_normalize_predictors( h );
2477 for( int i = 0; i < h->param.i_threads; i++ )
2478 h->thread[i]->rc->frame_size_estimated = h->thread[i]->rc->slice_size_planned;
2482 void x264_threads_merge_ratecontrol( x264_t *h )
2484 x264_ratecontrol_t *rc = h->rc;
2487 for( int i = 0; i < h->param.i_threads; i++ )
2489 x264_t *t = h->thread[i];
2490 x264_ratecontrol_t *rct = h->thread[i]->rc;
2491 if( h->param.rc.i_vbv_buffer_size )
2494 for( int row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2495 size += h->fdec->i_row_satd[row];
2496 int bits = t->stat.frame.i_mv_bits + t->stat.frame.i_tex_bits + t->stat.frame.i_misc_bits;
2497 int mb_count = (t->i_threadslice_end - t->i_threadslice_start) * h->mb.i_mb_width;
2498 update_predictor( &rc->pred[h->sh.i_type+(i+1)*5], qp2qscale( rct->qpa_rc/mb_count ), size, bits );
2502 rc->qpa_rc += rct->qpa_rc;
2503 rc->qpa_aq += rct->qpa_aq;
2507 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
2511 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
2512 /* these vars are updated in x264_ratecontrol_start()
2513 * so copy them from the context that most recently started (prev)
2514 * to the context that's about to start (cur). */
2519 COPY(last_qscale_for);
2520 COPY(last_non_b_pict_type);
2521 COPY(short_term_cplxsum);
2522 COPY(short_term_cplxcount);
2526 /* these vars can be updated by x264_ratecontrol_init_reconfigurable */
2531 COPY(single_frame_vbv);
2533 COPY(rate_factor_constant);
2534 COPY(rate_factor_max_increment);
2539 #define COPY(var) next->rc->var = cur->rc->var
2540 /* these vars are updated in x264_ratecontrol_end()
2541 * so copy them from the context that most recently ended (cur)
2542 * to the context that's about to end (next) */
2544 COPY(expected_bits_sum);
2545 COPY(filler_bits_sum);
2546 COPY(wanted_bits_window);
2548 COPY(initial_cpb_removal_delay);
2549 COPY(initial_cpb_removal_delay_offset);
2550 COPY(nrt_first_access_unit);
2551 COPY(previous_cpb_final_arrival_time);
2554 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
2555 /* the rest of the variables are either constant or thread-local */
2558 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
2560 /* find an interval ending on an overflow or underflow (depending on whether
2561 * we're adding or removing bits), and starting on the earliest frame that
2562 * can influence the buffer fill of that end frame. */
2563 x264_ratecontrol_t *rcc = h->rc;
2564 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
2565 const double buffer_max = .9 * rcc->buffer_size;
2566 double fill = fills[*t0-1];
2567 double parity = over ? 1. : -1.;
2568 int start = -1, end = -1;
2569 for( int i = *t0; i < rcc->num_entries; i++ )
2571 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 -
2572 qscale2bits( &rcc->entry[i], rcc->entry[i].new_qscale )) * parity;
2573 fill = x264_clip3f(fill, 0, rcc->buffer_size);
2575 if( fill <= buffer_min || i == 0 )
2581 else if( fill >= buffer_max && start >= 0 )
2586 return start >= 0 && end >= 0;
2589 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
2591 x264_ratecontrol_t *rcc = h->rc;
2592 double qscale_orig, qscale_new;
2596 for( int i = t0; i <= t1; i++ )
2598 qscale_orig = rcc->entry[i].new_qscale;
2599 qscale_orig = x264_clip3f( qscale_orig, qscale_min, qscale_max );
2600 qscale_new = qscale_orig * adjustment;
2601 qscale_new = x264_clip3f( qscale_new, qscale_min, qscale_max );
2602 rcc->entry[i].new_qscale = qscale_new;
2603 adjusted = adjusted || (qscale_new != qscale_orig);
2608 static double count_expected_bits( x264_t *h )
2610 x264_ratecontrol_t *rcc = h->rc;
2611 double expected_bits = 0;
2612 for( int i = 0; i < rcc->num_entries; i++ )
2614 ratecontrol_entry_t *rce = &rcc->entry[i];
2615 rce->expected_bits = expected_bits;
2616 expected_bits += qscale2bits( rce, rce->new_qscale );
2618 return expected_bits;
2621 static int vbv_pass2( x264_t *h, double all_available_bits )
2623 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
2624 * frames in the interval until either buffer is full at some intermediate frame or the
2625 * last frame in the interval no longer underflows. Recompute intervals and repeat.
2626 * Then do the converse to put bits back into overflow areas until target size is met */
2628 x264_ratecontrol_t *rcc = h->rc;
2630 double expected_bits = 0;
2632 double prev_bits = 0;
2634 double qscale_min = qp2qscale( h->param.rc.i_qp_min );
2635 double qscale_max = qp2qscale( h->param.rc.i_qp_max );
2637 int adj_min, adj_max;
2638 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2642 /* adjust overall stream size */
2646 prev_bits = expected_bits;
2649 { /* not first iteration */
2650 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2651 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2655 while(adj_min && find_underflow( h, fills, &t0, &t1, 1 ))
2657 adj_min = fix_underflow( h, t0, t1, adjustment, qscale_min, qscale_max );
2662 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2664 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2666 while( adj_max && find_underflow( h, fills, &t0, &t1, 0 ) )
2667 adj_max = fix_underflow( h, t0, t1, 1.001, qscale_min, qscale_max );
2669 expected_bits = count_expected_bits( h );
2670 } while( (expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2673 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2675 /* store expected vbv filling values for tracking when encoding */
2676 for( int i = 0; i < rcc->num_entries; i++ )
2677 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
2679 x264_free( fills-1 );
2685 static int init_pass2( x264_t *h )
2687 x264_ratecontrol_t *rcc = h->rc;
2688 uint64_t all_const_bits = 0;
2689 double timescale = (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2690 double duration = 0;
2691 for( int i = 0; i < rcc->num_entries; i++ )
2692 duration += rcc->entry[i].i_duration;
2693 duration *= timescale;
2694 uint64_t all_available_bits = h->param.rc.i_bitrate * 1000. * duration;
2695 double rate_factor, step_mult;
2696 double qblur = h->param.rc.f_qblur;
2697 double cplxblur = h->param.rc.f_complexity_blur;
2698 const int filter_size = (int)(qblur*4) | 1;
2699 double expected_bits;
2700 double *qscale, *blurred_qscale;
2701 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
2703 /* find total/average complexity & const_bits */
2704 for( int i = 0; i < rcc->num_entries; i++ )
2706 ratecontrol_entry_t *rce = &rcc->entry[i];
2707 all_const_bits += rce->misc_bits;
2710 if( all_available_bits < all_const_bits)
2712 x264_log( h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2713 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)) );
2717 /* Blur complexities, to reduce local fluctuation of QP.
2718 * We don't blur the QPs directly, because then one very simple frame
2719 * could drag down the QP of a nearby complex frame and give it more
2720 * bits than intended. */
2721 for( int i = 0; i < rcc->num_entries; i++ )
2723 ratecontrol_entry_t *rce = &rcc->entry[i];
2724 double weight_sum = 0;
2725 double cplx_sum = 0;
2726 double weight = 1.0;
2727 double gaussian_weight;
2728 /* weighted average of cplx of future frames */
2729 for( int j = 1; j < cplxblur*2 && j < rcc->num_entries-i; j++ )
2731 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2732 double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2733 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2734 if( weight < .0001 )
2736 gaussian_weight = weight * exp( -j*j/200.0 );
2737 weight_sum += gaussian_weight;
2738 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2740 /* weighted average of cplx of past frames */
2742 for( int j = 0; j <= cplxblur*2 && j <= i; j++ )
2744 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2745 double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2746 gaussian_weight = weight * exp( -j*j/200.0 );
2747 weight_sum += gaussian_weight;
2748 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2749 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2750 if( weight < .0001 )
2753 rce->blurred_complexity = cplx_sum / weight_sum;
2756 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2757 if( filter_size > 1 )
2758 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2760 blurred_qscale = qscale;
2762 /* Search for a factor which, when multiplied by the RCEQ values from
2763 * each frame, adds up to the desired total size.
2764 * There is no exact closed-form solution because of VBV constraints and
2765 * because qscale2bits is not invertible, but we can start with the simple
2766 * approximation of scaling the 1st pass by the ratio of bitrates.
2767 * The search range is probably overkill, but speed doesn't matter here. */
2770 for( int i = 0; i < rcc->num_entries; i++ )
2772 double q = get_qscale(h, &rcc->entry[i], 1.0, i);
2773 expected_bits += qscale2bits(&rcc->entry[i], q);
2774 rcc->last_qscale_for[rcc->entry[i].pict_type] = q;
2776 step_mult = all_available_bits / expected_bits;
2779 for( double step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2782 rate_factor += step;
2784 rcc->last_non_b_pict_type = -1;
2785 rcc->last_accum_p_norm = 1;
2786 rcc->accum_p_norm = 0;
2788 rcc->last_qscale_for[0] =
2789 rcc->last_qscale_for[1] =
2790 rcc->last_qscale_for[2] = pow( base_cplx, 1 - rcc->qcompress ) / rate_factor;
2793 for( int i = 0; i < rcc->num_entries; i++ )
2795 qscale[i] = get_qscale( h, &rcc->entry[i], rate_factor, -1 );
2796 rcc->last_qscale_for[rcc->entry[i].pict_type] = qscale[i];
2799 /* fixed I/B qscale relative to P */
2800 for( int i = rcc->num_entries-1; i >= 0; i-- )
2802 qscale[i] = get_diff_limited_q( h, &rcc->entry[i], qscale[i], i );
2803 assert(qscale[i] >= 0);
2807 if( filter_size > 1 )
2809 assert( filter_size%2 == 1 );
2810 for( int i = 0; i < rcc->num_entries; i++ )
2812 ratecontrol_entry_t *rce = &rcc->entry[i];
2813 double q = 0.0, sum = 0.0;
2815 for( int j = 0; j < filter_size; j++ )
2817 int idx = i+j-filter_size/2;
2819 double coeff = qblur==0 ? 1.0 : exp( -d*d/(qblur*qblur) );
2820 if( idx < 0 || idx >= rcc->num_entries )
2822 if( rce->pict_type != rcc->entry[idx].pict_type )
2824 q += qscale[idx] * coeff;
2827 blurred_qscale[i] = q/sum;
2831 /* find expected bits */
2832 for( int i = 0; i < rcc->num_entries; i++ )
2834 ratecontrol_entry_t *rce = &rcc->entry[i];
2835 rce->new_qscale = clip_qscale( h, rce->pict_type, blurred_qscale[i] );
2836 assert(rce->new_qscale >= 0);
2837 expected_bits += qscale2bits( rce, rce->new_qscale );
2840 if( expected_bits > all_available_bits )
2841 rate_factor -= step;
2844 x264_free( qscale );
2845 if( filter_size > 1 )
2846 x264_free( blurred_qscale );
2849 if( vbv_pass2( h, all_available_bits ) )
2851 expected_bits = count_expected_bits( h );
2853 if( fabs( expected_bits/all_available_bits - 1.0 ) > 0.01 )
2856 for( int i = 0; i < rcc->num_entries; i++ )
2857 avgq += rcc->entry[i].new_qscale;
2858 avgq = qscale2qp( avgq / rcc->num_entries );
2860 if( expected_bits > all_available_bits || !rcc->b_vbv )
2861 x264_log( h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n" );
2862 x264_log( h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
2863 (float)h->param.rc.i_bitrate,
2864 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
2866 if( expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2 )
2868 if( h->param.rc.i_qp_min > 0 )
2869 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min );
2871 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate\n" );
2873 else if( expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2 )
2875 if( h->param.rc.i_qp_max < QP_MAX )
2876 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max );
2878 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate\n");
2880 else if( !(rcc->b_2pass && rcc->b_vbv) )
2881 x264_log( h, X264_LOG_WARNING, "internal error\n" );