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 int 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 float predict_size( predictor_t *p, float q, float var );
171 static void update_predictor( predictor_t *p, float q, float var, float 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 float qp2qscale( float qp )
188 return 0.85f * powf( 2.0f, ( qp - 12.0f ) / 6.0f );
190 static inline float qscale2qp( float qscale )
192 return 12.0f + 6.0f * log2f( qscale/0.85f );
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, int b_store )
210 uint32_t sum = sum_ssd;
211 uint32_t ssd = sum_ssd >> 32;
214 frame->i_pixel_sum[i] += sum;
215 frame->i_pixel_ssd[i] += ssd;
217 return ssd - ((uint64_t)sum * sum >> shift);
220 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 )
222 int height = b_chroma ? 16>>CHROMA_V_SHIFT : 16;
223 int stride = frame->i_stride[i];
225 ? 16 * mb_x + height * (mb_y&~1) * stride + (mb_y&1) * stride
226 : 16 * mb_x + height * mb_y * stride;
230 ALIGNED_ARRAY_16( pixel, pix,[FENC_STRIDE*16] );
231 int chromapix = h->luma2chroma_pixel[PIXEL_16x16];
232 int shift = 7 - CHROMA_V_SHIFT;
234 h->mc.load_deinterleave_chroma_fenc( pix, frame->plane[1] + offset, stride, height );
235 return ac_energy_var( h->pixf.var[chromapix]( pix, FENC_STRIDE ), shift, frame, 1, b_store )
236 + ac_energy_var( h->pixf.var[chromapix]( pix+FENC_STRIDE/2, FENC_STRIDE ), shift, frame, 2, b_store );
239 return ac_energy_var( h->pixf.var[PIXEL_16x16]( frame->plane[i] + offset, stride ), 8, frame, i, b_store );
242 // Find the total AC energy of the block in all planes.
243 static NOINLINE uint32_t x264_ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
245 /* This function contains annoying hacks because GCC has a habit of reordering emms
246 * and putting it after floating point ops. As a result, we put the emms at the end of the
247 * function and make sure that its always called before the float math. Noinline makes
248 * sure no reordering goes on. */
250 x264_prefetch_fenc( h, frame, mb_x, mb_y );
251 if( h->mb.b_adaptive_mbaff )
253 /* We don't know the super-MB mode we're going to pick yet, so
254 * simply try both and pick the lower of the two. */
255 uint32_t var_interlaced, var_progressive;
256 var_interlaced = ac_energy_plane( h, mb_x, mb_y, frame, 0, 0, 1, 1 );
257 var_progressive = ac_energy_plane( h, mb_x, mb_y, frame, 0, 0, 0, 0 );
260 var_interlaced += ac_energy_plane( h, mb_x, mb_y, frame, 1, 0, 1, 1 );
261 var_progressive += ac_energy_plane( h, mb_x, mb_y, frame, 1, 0, 0, 0 );
262 var_interlaced += ac_energy_plane( h, mb_x, mb_y, frame, 2, 0, 1, 1 );
263 var_progressive += ac_energy_plane( h, mb_x, mb_y, frame, 2, 0, 0, 0 );
267 var_interlaced += ac_energy_plane( h, mb_x, mb_y, frame, 1, 1, 1, 1 );
268 var_progressive += ac_energy_plane( h, mb_x, mb_y, frame, 1, 1, 0, 0 );
270 var = X264_MIN( var_interlaced, var_progressive );
274 var = ac_energy_plane( h, mb_x, mb_y, frame, 0, 0, PARAM_INTERLACED, 1 );
277 var += ac_energy_plane( h, mb_x, mb_y, frame, 1, 0, PARAM_INTERLACED, 1 );
278 var += ac_energy_plane( h, mb_x, mb_y, frame, 2, 0, PARAM_INTERLACED, 1 );
281 var += ac_energy_plane( h, mb_x, mb_y, frame, 1, 1, PARAM_INTERLACED, 1 );
287 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame, float *quant_offsets )
289 /* constants chosen to result in approximately the same overall bitrate as without AQ.
290 * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
293 /* Initialize frame stats */
294 for( int i = 0; i < 3; i++ )
296 frame->i_pixel_sum[i] = 0;
297 frame->i_pixel_ssd[i] = 0;
300 /* Degenerate cases */
301 if( h->param.rc.i_aq_mode == X264_AQ_NONE || h->param.rc.f_aq_strength == 0 )
303 /* Need to init it anyways for MB tree */
304 if( h->param.rc.i_aq_mode && h->param.rc.f_aq_strength == 0 )
308 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
309 frame->f_qp_offset[mb_xy] = frame->f_qp_offset_aq[mb_xy] = quant_offsets[mb_xy];
310 if( h->frames.b_have_lowres )
311 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
312 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8( frame->f_qp_offset[mb_xy] );
316 memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
317 memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
318 if( h->frames.b_have_lowres )
319 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
320 frame->i_inv_qscale_factor[mb_xy] = 256;
323 /* Need variance data for weighted prediction */
324 if( h->param.analyse.i_weighted_pred )
326 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
327 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
328 x264_ac_energy_mb( h, mb_x, mb_y, frame );
333 /* Actual adaptive quantization */
336 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
338 float bit_depth_correction = powf(1 << (BIT_DEPTH-8), 0.5f);
339 float avg_adj_pow2 = 0.f;
340 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
341 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
343 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
344 float qp_adj = powf( energy + 1, 0.125f );
345 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
347 avg_adj_pow2 += qp_adj * qp_adj;
349 avg_adj /= h->mb.i_mb_count;
350 avg_adj_pow2 /= h->mb.i_mb_count;
351 strength = h->param.rc.f_aq_strength * avg_adj / bit_depth_correction;
352 avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - (14.f * bit_depth_correction)) / avg_adj;
355 strength = h->param.rc.f_aq_strength * 1.0397f;
357 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
358 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
361 int mb_xy = mb_x + mb_y*h->mb.i_mb_stride;
362 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
364 qp_adj = frame->f_qp_offset[mb_xy];
365 qp_adj = strength * (qp_adj - avg_adj);
369 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
370 qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - (14.427f + 2*(BIT_DEPTH-8)));
373 qp_adj += quant_offsets[mb_xy];
374 frame->f_qp_offset[mb_xy] =
375 frame->f_qp_offset_aq[mb_xy] = qp_adj;
376 if( h->frames.b_have_lowres )
377 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8(qp_adj);
381 /* Remove mean from SSD calculation */
382 for( int i = 0; i < 3; i++ )
384 uint64_t ssd = frame->i_pixel_ssd[i];
385 uint64_t sum = frame->i_pixel_sum[i];
386 int width = 16*h->mb.i_mb_width >> (i && CHROMA_H_SHIFT);
387 int height = 16*h->mb.i_mb_height >> (i && CHROMA_V_SHIFT);
388 frame->i_pixel_ssd[i] = ssd - (sum * sum + width * height / 2) / (width * height);
392 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame, float *quant_offsets )
394 x264_ratecontrol_t *rc = h->rc;
395 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
397 if( rc->entry[frame->i_frame].kept_as_ref )
400 if( rc->qpbuf_pos < 0 )
406 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
408 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 )
411 if( i_type != i_type_actual && rc->qpbuf_pos == 1 )
413 x264_log( h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type, i_type_actual );
416 } while( i_type != i_type_actual );
419 for( int i = 0; i < h->mb.i_mb_count; i++ )
421 frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[rc->qpbuf_pos][i] )) * (1/256.0);
422 if( h->frames.b_have_lowres )
423 frame->i_inv_qscale_factor[i] = x264_exp2fix8(frame->f_qp_offset[i]);
428 x264_stack_align( x264_adaptive_quant_frame, h, frame, quant_offsets );
431 x264_log( h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n" );
435 int x264_reference_build_list_optimal( x264_t *h )
437 ratecontrol_entry_t *rce = h->rc->rce;
438 x264_frame_t *frames[16];
439 x264_weight_t weights[16][3];
442 if( rce->refs != h->i_ref[0] )
445 memcpy( frames, h->fref[0], sizeof(frames) );
446 memcpy( refcount, rce->refcount, sizeof(refcount) );
447 memcpy( weights, h->fenc->weight, sizeof(weights) );
448 memset( &h->fenc->weight[1][0], 0, sizeof(x264_weight_t[15][3]) );
450 /* For now don't reorder ref 0; it seems to lower quality
451 in most cases due to skips. */
452 for( int ref = 1; ref < h->i_ref[0]; ref++ )
457 for( int i = 1; i < h->i_ref[0]; i++ )
458 /* Favor lower POC as a tiebreaker. */
459 COPY2_IF_GT( max, refcount[i], bestref, i );
461 /* FIXME: If there are duplicates from frames other than ref0 then it is possible
462 * that the optimal ordering doesnt place every duplicate. */
464 refcount[bestref] = -1;
465 h->fref[0][ref] = frames[bestref];
466 memcpy( h->fenc->weight[ref], weights[bestref], sizeof(weights[bestref]) );
472 static char *x264_strcat_filename( char *input, char *suffix )
474 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
477 strcpy( output, input );
478 strcat( output, suffix );
482 void x264_ratecontrol_init_reconfigurable( x264_t *h, int b_init )
484 x264_ratecontrol_t *rc = h->rc;
485 if( !b_init && rc->b_2pass )
488 if( h->param.rc.i_rc_method == X264_RC_CRF )
490 /* Arbitrary rescaling to make CRF somewhat similar to QP.
491 * Try to compensate for MB-tree's effects as well. */
492 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
493 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
494 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
495 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset + QP_BD_OFFSET );
498 if( h->param.rc.i_vbv_max_bitrate > 0 && h->param.rc.i_vbv_buffer_size > 0 )
500 /* We don't support changing the ABR bitrate right now,
501 so if the stream starts as CBR, keep it CBR. */
502 if( rc->b_vbv_min_rate )
503 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
505 if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
507 h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
508 x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
509 h->param.rc.i_vbv_buffer_size );
512 int vbv_buffer_size = h->param.rc.i_vbv_buffer_size * 1000;
513 int vbv_max_bitrate = h->param.rc.i_vbv_max_bitrate * 1000;
516 if( h->param.i_nal_hrd && b_init )
518 h->sps->vui.hrd.i_cpb_cnt = 1;
519 h->sps->vui.hrd.b_cbr_hrd = h->param.i_nal_hrd == X264_NAL_HRD_CBR;
520 h->sps->vui.hrd.i_time_offset_length = 0;
525 int bitrate = 1000*h->param.rc.i_vbv_max_bitrate;
526 int bufsize = 1000*h->param.rc.i_vbv_buffer_size;
528 // normalize HRD size and rate to the value / scale notation
529 h->sps->vui.hrd.i_bit_rate_scale = x264_clip3( x264_ctz( bitrate ) - BR_SHIFT, 0, 15 );
530 h->sps->vui.hrd.i_bit_rate_value = bitrate >> ( h->sps->vui.hrd.i_bit_rate_scale + BR_SHIFT );
531 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 );
532 h->sps->vui.hrd.i_cpb_size_scale = x264_clip3( x264_ctz( bufsize ) - CPB_SHIFT, 0, 15 );
533 h->sps->vui.hrd.i_cpb_size_value = bufsize >> ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
534 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 );
540 #define MAX_DURATION 0.5
542 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 );
543 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;
544 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);
546 h->sps->vui.hrd.i_initial_cpb_removal_delay_length = 2 + x264_clip3( 32 - x264_clz( max_delay ), 4, 22 );
547 h->sps->vui.hrd.i_cpb_removal_delay_length = x264_clip3( 32 - x264_clz( max_cpb_output_delay ), 4, 31 );
548 h->sps->vui.hrd.i_dpb_output_delay_length = x264_clip3( 32 - x264_clz( max_dpb_output_delay ), 4, 31 );
552 vbv_buffer_size = h->sps->vui.hrd.i_cpb_size_unscaled;
553 vbv_max_bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
555 else if( h->param.i_nal_hrd && !b_init )
557 x264_log( h, X264_LOG_WARNING, "VBV parameters cannot be changed when NAL HRD is in use\n" );
560 h->sps->vui.hrd.i_bit_rate_unscaled = vbv_max_bitrate;
561 h->sps->vui.hrd.i_cpb_size_unscaled = vbv_buffer_size;
563 if( rc->b_vbv_min_rate )
564 rc->bitrate = h->param.rc.i_bitrate * 1000.;
565 rc->buffer_rate = vbv_max_bitrate / rc->fps;
566 rc->vbv_max_rate = vbv_max_bitrate;
567 rc->buffer_size = vbv_buffer_size;
568 rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
569 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
570 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
571 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.f_rf_constant_max )
573 rc->rate_factor_max_increment = h->param.rc.f_rf_constant_max - h->param.rc.f_rf_constant;
574 if( rc->rate_factor_max_increment <= 0 )
576 x264_log( h, X264_LOG_WARNING, "CRF max must be greater than CRF\n" );
577 rc->rate_factor_max_increment = 0;
582 if( h->param.rc.f_vbv_buffer_init > 1. )
583 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 );
584 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);
585 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init * h->sps->vui.i_time_scale;
587 rc->b_vbv_min_rate = !rc->b_2pass
588 && h->param.rc.i_rc_method == X264_RC_ABR
589 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
594 int x264_ratecontrol_new( x264_t *h )
596 x264_ratecontrol_t *rc;
600 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
603 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
604 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
606 /* FIXME: use integers */
607 if( h->param.i_fps_num > 0 && h->param.i_fps_den > 0 )
608 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
612 if( h->param.rc.b_mb_tree )
614 h->param.rc.f_pb_factor = 1;
618 rc->qcompress = h->param.rc.f_qcompress;
620 rc->bitrate = h->param.rc.i_bitrate * 1000.;
621 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
622 rc->nmb = h->mb.i_mb_count;
623 rc->last_non_b_pict_type = -1;
626 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
628 x264_log( h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n" );
632 x264_ratecontrol_init_reconfigurable( h, 1 );
634 if( h->param.i_nal_hrd )
636 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale;
637 uint64_t num = 180000;
638 x264_reduce_fraction64( &num, &denom );
639 rc->hrd_multiply_denom = 180000 / num;
641 double bits_required = log2( 180000 / rc->hrd_multiply_denom )
642 + log2( h->sps->vui.i_time_scale )
643 + log2( h->sps->vui.hrd.i_cpb_size_unscaled );
644 if( bits_required >= 63 )
646 x264_log( h, X264_LOG_ERROR, "HRD with very large timescale and bufsize not supported\n" );
651 if( rc->rate_tolerance < 0.01 )
653 x264_log( h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n" );
654 rc->rate_tolerance = 0.01;
657 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
661 /* FIXME ABR_INIT_QP is actually used only in CRF */
662 #define ABR_INIT_QP (( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 ) + QP_BD_OFFSET)
663 rc->accum_p_norm = .01;
664 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
665 /* estimated ratio that produces a reasonable QP for the first I-frame */
666 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
667 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
668 rc->last_non_b_pict_type = SLICE_TYPE_I;
671 rc->ip_offset = 6.0 * log2f( h->param.rc.f_ip_factor );
672 rc->pb_offset = 6.0 * log2f( h->param.rc.f_pb_factor );
673 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
674 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, QP_MAX );
675 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, QP_MAX );
676 h->mb.ip_offset = rc->ip_offset + 0.5;
678 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
679 rc->last_qscale = qp2qscale( 26 );
680 int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
681 CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
682 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
683 for( int i = 0; i < 3; i++ )
685 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
686 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
687 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
688 for( int j = 0; j < num_preds; j++ )
690 rc->pred[i+j*5].coeff= 2.0;
691 rc->pred[i+j*5].count= 1.0;
692 rc->pred[i+j*5].decay= 0.5;
693 rc->pred[i+j*5].offset= 0.0;
695 for( int j = 0; j < 2; j++ )
697 rc->row_preds[i][j].coeff= .25;
698 rc->row_preds[i][j].count= 1.0;
699 rc->row_preds[i][j].decay= 0.5;
700 rc->row_preds[i][j].offset= 0.0;
703 *rc->pred_b_from_p = rc->pred[0];
705 if( parse_zones( h ) < 0 )
707 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
711 /* Load stat file and init 2pass algo */
712 if( h->param.rc.b_stat_read )
714 char *p, *stats_in, *stats_buf;
716 /* read 1st pass stats */
717 assert( h->param.rc.psz_stat_in );
718 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
721 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n" );
724 if( h->param.rc.b_mb_tree )
726 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
727 if( !mbtree_stats_in )
729 rc->p_mbtree_stat_file_in = fopen( mbtree_stats_in, "rb" );
730 x264_free( mbtree_stats_in );
731 if( !rc->p_mbtree_stat_file_in )
733 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n" );
738 /* check whether 1st pass options were compatible with current options */
739 if( strncmp( stats_buf, "#options:", 9 ) )
741 x264_log( h, X264_LOG_ERROR, "options list in stats file not valid\n" );
745 float res_factor, res_factor_bits;
749 char *opts = stats_buf;
750 stats_in = strchr( stats_buf, '\n' );
755 if( sscanf( opts, "#options: %dx%d", &i, &j ) != 2 )
757 x264_log( h, X264_LOG_ERROR, "resolution specified in stats file not valid\n" );
760 else if( h->param.rc.b_mb_tree && (i != h->param.i_width || j != h->param.i_height) )
762 x264_log( h, X264_LOG_ERROR, "MB-tree doesn't support different resolution than 1st pass (%dx%d vs %dx%d)\n",
763 h->param.i_width, h->param.i_height, i, j );
766 res_factor = (float)h->param.i_width * h->param.i_height / (i*j);
767 /* Change in bits relative to resolution isn't quite linear on typical sources,
768 * so we'll at least try to roughly approximate this effect. */
769 res_factor_bits = powf( res_factor, 0.7 );
771 if( ( p = strstr( opts, "timebase=" ) ) && sscanf( p, "timebase=%u/%u", &k, &l ) != 2 )
773 x264_log( h, X264_LOG_ERROR, "timebase specified in stats file not valid\n" );
776 if( k != h->param.i_timebase_num || l != h->param.i_timebase_den )
778 x264_log( h, X264_LOG_ERROR, "timebase mismatch with 1st pass (%u/%u vs %u/%u)\n",
779 h->param.i_timebase_num, h->param.i_timebase_den, k, l );
783 CMP_OPT_FIRST_PASS( "bitdepth", BIT_DEPTH );
784 CMP_OPT_FIRST_PASS( "weightp", X264_MAX( 0, h->param.analyse.i_weighted_pred ) );
785 CMP_OPT_FIRST_PASS( "bframes", h->param.i_bframe );
786 CMP_OPT_FIRST_PASS( "b_pyramid", h->param.i_bframe_pyramid );
787 CMP_OPT_FIRST_PASS( "intra_refresh", h->param.b_intra_refresh );
788 CMP_OPT_FIRST_PASS( "open_gop", h->param.b_open_gop );
789 CMP_OPT_FIRST_PASS( "bluray_compat", h->param.b_bluray_compat );
791 if( (p = strstr( opts, "interlaced=" )) )
793 char *current = h->param.b_interlaced ? h->param.b_tff ? "tff" : "bff" : h->param.b_fake_interlaced ? "fake" : "0";
795 sscanf( p, "interlaced=%4s", buf );
796 if( strcmp( current, buf ) )
798 x264_log( h, X264_LOG_ERROR, "different interlaced setting than first pass (%s vs %s)\n", current, buf );
803 if( (p = strstr( opts, "keyint=" )) )
806 char buf[13] = "infinite ";
807 if( h->param.i_keyint_max != X264_KEYINT_MAX_INFINITE )
808 sprintf( buf, "%d ", h->param.i_keyint_max );
809 if( strncmp( p, buf, strlen(buf) ) )
811 x264_log( h, X264_LOG_ERROR, "different keyint setting than first pass (%.*s vs %.*s)\n",
812 strlen(buf)-1, buf, strcspn(p, " "), p );
817 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
818 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
820 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
822 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
823 h->mb.b_direct_auto_write = 1;
826 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
827 h->param.i_bframe_adaptive = i;
828 else if( h->param.i_bframe )
830 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
834 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 ) )
835 h->param.rc.i_lookahead = i;
838 /* find number of pics */
841 for( num_entries = -1; p; num_entries++ )
842 p = strchr( p + 1, ';' );
845 x264_log( h, X264_LOG_ERROR, "empty stats file\n" );
848 rc->num_entries = num_entries;
850 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
852 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
853 h->param.i_frame_total, rc->num_entries );
855 if( h->param.i_frame_total > rc->num_entries )
857 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
858 h->param.i_frame_total, rc->num_entries );
862 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
864 /* init all to skipped p frames */
865 for( int i = 0; i < rc->num_entries; i++ )
867 ratecontrol_entry_t *rce = &rc->entry[i];
868 rce->pict_type = SLICE_TYPE_P;
869 rce->qscale = rce->new_qscale = qp2qscale( 20 );
870 rce->misc_bits = rc->nmb + 10;
876 for( int i = 0; i < rc->num_entries; i++ )
878 ratecontrol_entry_t *rce;
886 next= strchr(p, ';');
888 *next++ = 0; //sscanf is unbelievably slow on long strings
889 e = sscanf( p, " in:%d ", &frame_number );
891 if( frame_number < 0 || frame_number >= rc->num_entries )
893 x264_log( h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i );
896 rce = &rc->entry[frame_number];
897 rce->direct_mode = 0;
899 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",
900 &pict_type, &rce->i_duration, &rce->i_cpb_duration, &qp, &rce->tex_bits,
901 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
902 &rce->s_count, &rce->direct_mode );
903 rce->tex_bits *= res_factor_bits;
904 rce->mv_bits *= res_factor_bits;
905 rce->misc_bits *= res_factor_bits;
906 rce->i_count *= res_factor;
907 rce->p_count *= res_factor;
908 rce->s_count *= res_factor;
910 p = strstr( p, "ref:" );
914 for( ref = 0; ref < 16; ref++ )
916 if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
918 p = strchr( p+1, ' ' );
925 rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
926 char *w = strchr( p, 'w' );
929 int count = sscanf( w, "w:%hd,%hd,%hd,%hd,%hd,%hd,%hd,%hd",
930 &rce->i_weight_denom[0], &rce->weight[0][0], &rce->weight[0][1],
931 &rce->i_weight_denom[1], &rce->weight[1][0], &rce->weight[1][1],
932 &rce->weight[2][0], &rce->weight[2][1] );
934 rce->i_weight_denom[1] = -1;
935 else if ( count != 8 )
936 rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
939 if( pict_type != 'b' )
940 rce->kept_as_ref = 1;
944 rce->frame_type = X264_TYPE_IDR;
945 rce->pict_type = SLICE_TYPE_I;
948 rce->frame_type = X264_TYPE_I;
949 rce->pict_type = SLICE_TYPE_I;
952 rce->frame_type = X264_TYPE_P;
953 rce->pict_type = SLICE_TYPE_P;
956 rce->frame_type = X264_TYPE_BREF;
957 rce->pict_type = SLICE_TYPE_B;
960 rce->frame_type = X264_TYPE_B;
961 rce->pict_type = SLICE_TYPE_B;
963 default: e = -1; break;
968 x264_log( h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e );
971 rce->qscale = qp2qscale( qp );
975 x264_free( stats_buf );
977 if( h->param.rc.i_rc_method == X264_RC_ABR )
979 if( init_pass2( h ) < 0 )
981 } /* else we're using constant quant, so no need to run the bitrate allocation */
984 /* Open output file */
985 /* If input and output files are the same, output to a temp file
986 * and move it to the real name only when it's complete */
987 if( h->param.rc.b_stat_write )
990 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
991 if( !rc->psz_stat_file_tmpname )
994 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
995 if( rc->p_stat_file_out == NULL )
997 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n" );
1001 p = x264_param2string( &h->param, 1 );
1003 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
1005 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
1007 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
1008 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
1009 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
1012 rc->p_mbtree_stat_file_out = fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
1013 if( rc->p_mbtree_stat_file_out == NULL )
1015 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n" );
1021 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
1023 CHECKED_MALLOC( rc->qp_buffer[0], h->mb.i_mb_count * sizeof(uint16_t) );
1024 if( h->param.i_bframe_pyramid && h->param.rc.b_stat_read )
1025 CHECKED_MALLOC( rc->qp_buffer[1], h->mb.i_mb_count * sizeof(uint16_t) );
1029 for( int i = 0; i<h->param.i_threads; i++ )
1031 h->thread[i]->rc = rc+i;
1035 h->thread[i]->param = h->param;
1036 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
1037 h->thread[i]->mb.ip_offset = h->mb.ip_offset;
1046 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
1049 char *tok, UNUSED *saveptr=NULL;
1051 z->f_bitrate_factor = 1;
1052 if( 3 <= sscanf(p, "%d,%d,q=%d%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
1054 else if( 3 <= sscanf(p, "%d,%d,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
1056 else if( 2 <= sscanf(p, "%d,%d%n", &z->i_start, &z->i_end, &len) )
1060 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
1066 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
1067 memcpy( z->param, &h->param, sizeof(x264_param_t) );
1068 z->param->param_free = x264_free;
1069 while( (tok = strtok_r( p, ",", &saveptr )) )
1071 char *val = strchr( tok, '=' );
1077 if( x264_param_parse( z->param, tok, val ) )
1079 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
1089 static int parse_zones( x264_t *h )
1091 x264_ratecontrol_t *rc = h->rc;
1092 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
1094 char *psz_zones, *p;
1095 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
1096 strcpy( psz_zones, h->param.rc.psz_zones );
1097 h->param.rc.i_zones = 1;
1098 for( p = psz_zones; *p; p++ )
1099 h->param.rc.i_zones += (*p == '/');
1100 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
1102 for( int i = 0; i < h->param.rc.i_zones; i++ )
1104 int i_tok = strcspn( p, "/" );
1106 if( parse_zone( h, &h->param.rc.zones[i], p ) )
1110 x264_free( psz_zones );
1113 if( h->param.rc.i_zones > 0 )
1115 for( int i = 0; i < h->param.rc.i_zones; i++ )
1117 x264_zone_t z = h->param.rc.zones[i];
1118 if( z.i_start < 0 || z.i_start > z.i_end )
1120 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
1121 z.i_start, z.i_end );
1124 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
1126 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
1127 z.f_bitrate_factor );
1132 rc->i_zones = h->param.rc.i_zones + 1;
1133 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
1134 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
1136 // default zone to fall back to if none of the others match
1137 rc->zones[0].i_start = 0;
1138 rc->zones[0].i_end = INT_MAX;
1139 rc->zones[0].b_force_qp = 0;
1140 rc->zones[0].f_bitrate_factor = 1;
1141 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
1142 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
1143 for( int i = 1; i < rc->i_zones; i++ )
1145 if( !rc->zones[i].param )
1146 rc->zones[i].param = rc->zones[0].param;
1155 static x264_zone_t *get_zone( x264_t *h, int frame_num )
1157 for( int i = h->rc->i_zones - 1; i >= 0; i-- )
1159 x264_zone_t *z = &h->rc->zones[i];
1160 if( frame_num >= z->i_start && frame_num <= z->i_end )
1166 void x264_ratecontrol_summary( x264_t *h )
1168 x264_ratecontrol_t *rc = h->rc;
1169 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
1171 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
1172 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
1173 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
1174 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
1175 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset - QP_BD_OFFSET );
1179 void x264_ratecontrol_delete( x264_t *h )
1181 x264_ratecontrol_t *rc = h->rc;
1184 if( rc->p_stat_file_out )
1186 b_regular_file = x264_is_regular_file( rc->p_stat_file_out );
1187 fclose( rc->p_stat_file_out );
1188 if( h->i_frame >= rc->num_entries && b_regular_file )
1189 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
1191 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1192 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
1194 x264_free( rc->psz_stat_file_tmpname );
1196 if( rc->p_mbtree_stat_file_out )
1198 b_regular_file = x264_is_regular_file( rc->p_mbtree_stat_file_out );
1199 fclose( rc->p_mbtree_stat_file_out );
1200 if( h->i_frame >= rc->num_entries && b_regular_file )
1201 if( rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
1203 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1204 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
1206 x264_free( rc->psz_mbtree_stat_file_tmpname );
1207 x264_free( rc->psz_mbtree_stat_file_name );
1209 if( rc->p_mbtree_stat_file_in )
1210 fclose( rc->p_mbtree_stat_file_in );
1211 x264_free( rc->pred );
1212 x264_free( rc->pred_b_from_p );
1213 x264_free( rc->entry );
1214 x264_free( rc->qp_buffer[0] );
1215 x264_free( rc->qp_buffer[1] );
1218 x264_free( rc->zones[0].param );
1219 for( int i = 1; i < rc->i_zones; i++ )
1220 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
1221 rc->zones[i].param->param_free( rc->zones[i].param );
1222 x264_free( rc->zones );
1227 static void accum_p_qp_update( x264_t *h, float qp )
1229 x264_ratecontrol_t *rc = h->rc;
1230 rc->accum_p_qp *= .95;
1231 rc->accum_p_norm *= .95;
1232 rc->accum_p_norm += 1;
1233 if( h->sh.i_type == SLICE_TYPE_I )
1234 rc->accum_p_qp += qp + rc->ip_offset;
1236 rc->accum_p_qp += qp;
1239 /* Before encoding a frame, choose a QP for it */
1240 void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
1242 x264_ratecontrol_t *rc = h->rc;
1243 ratecontrol_entry_t *rce = NULL;
1244 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
1249 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
1250 x264_encoder_reconfig( h, zone->param );
1251 rc->prev_zone = zone;
1253 if( h->param.rc.b_stat_read )
1255 int frame = h->fenc->i_frame;
1256 assert( frame >= 0 && frame < rc->num_entries );
1257 rce = h->rc->rce = &h->rc->entry[frame];
1259 if( h->sh.i_type == SLICE_TYPE_B
1260 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
1262 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
1263 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
1269 memset( h->fdec->i_row_bits, 0, h->mb.i_mb_height * sizeof(int) );
1270 memset( h->fdec->f_row_qp, 0, h->mb.i_mb_height * sizeof(float) );
1271 memset( h->fdec->f_row_qscale, 0, h->mb.i_mb_height * sizeof(float) );
1272 rc->row_pred = &rc->row_preds[h->sh.i_type];
1273 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;
1274 update_vbv_plan( h, overhead );
1276 const x264_level_t *l = x264_levels;
1277 while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
1280 int mincr = l->mincr;
1282 if( h->param.b_bluray_compat )
1285 /* Profiles above High don't require minCR, so just set the maximum to a large value. */
1286 if( h->sps->i_profile_idc > PROFILE_HIGH )
1287 rc->frame_size_maximum = 1e9;
1290 /* The spec has a bizarre special case for the first frame. */
1291 if( h->i_frame == 0 )
1293 //384 * ( Max( PicSizeInMbs, fR * MaxMBPS ) + MaxMBPS * ( tr( 0 ) - tr,n( 0 ) ) ) / MinCR
1294 double fr = 1. / 172;
1295 int pic_size_in_mbs = h->mb.i_mb_width * h->mb.i_mb_height;
1296 rc->frame_size_maximum = 384 * BIT_DEPTH * X264_MAX( pic_size_in_mbs, fr*l->mbps ) / mincr;
1300 //384 * MaxMBPS * ( tr( n ) - tr( n - 1 ) ) / MinCR
1301 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;
1306 if( h->sh.i_type != SLICE_TYPE_B )
1307 rc->bframes = h->fenc->i_bframes;
1311 q = qscale2qp( rate_estimate_qscale( h ) );
1313 else if( rc->b_2pass )
1315 rce->new_qscale = rate_estimate_qscale( h );
1316 q = qscale2qp( rce->new_qscale );
1320 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
1321 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
1323 q = rc->qp_constant[ h->sh.i_type ];
1327 if( zone->b_force_qp )
1328 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1330 q -= 6*log2f( zone->f_bitrate_factor );
1333 if( i_force_qp != X264_QP_AUTO )
1336 q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1340 rc->qp = x264_clip3( q + 0.5f, 0, QP_MAX );
1341 h->fdec->f_qp_avg_rc =
1342 h->fdec->f_qp_avg_aq =
1345 rce->new_qp = rc->qp;
1347 accum_p_qp_update( h, rc->qpm );
1349 if( h->sh.i_type != SLICE_TYPE_B )
1350 rc->last_non_b_pict_type = h->sh.i_type;
1353 static float predict_row_size( x264_t *h, int y, float qscale )
1355 /* average between two predictors:
1356 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1357 x264_ratecontrol_t *rc = h->rc;
1358 float pred_s = predict_size( rc->row_pred[0], qscale, h->fdec->i_row_satd[y] );
1359 if( h->sh.i_type == SLICE_TYPE_I || qscale >= h->fref[0][0]->f_row_qscale[y] )
1361 if( h->sh.i_type == SLICE_TYPE_P
1362 && h->fref[0][0]->i_type == h->fdec->i_type
1363 && h->fref[0][0]->f_row_qscale[y] > 0
1364 && h->fref[0][0]->i_row_satd[y] > 0
1365 && (abs(h->fref[0][0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1367 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]
1368 * h->fref[0][0]->f_row_qscale[y] / qscale;
1369 return (pred_s + pred_t) * 0.5f;
1373 /* Our QP is lower than the reference! */
1376 float pred_intra = predict_size( rc->row_pred[1], qscale, h->fdec->i_row_satds[0][0][y] );
1377 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1378 return pred_intra + pred_s;
1382 static int row_bits_so_far( x264_t *h, int y )
1385 for( int i = h->i_threadslice_start; i <= y; i++ )
1386 bits += h->fdec->i_row_bits[i];
1390 static float predict_row_size_sum( x264_t *h, int y, float qp )
1392 float qscale = qp2qscale( qp );
1393 float bits = row_bits_so_far( h, y );
1394 for( int i = y+1; i < h->i_threadslice_end; i++ )
1395 bits += predict_row_size( h, i, qscale );
1400 * eliminate all use of qp in row ratecontrol: make it entirely qscale-based.
1401 * make this function stop being needlessly O(N^2)
1402 * update more often than once per row? */
1403 void x264_ratecontrol_mb( x264_t *h, int bits )
1405 x264_ratecontrol_t *rc = h->rc;
1406 const int y = h->mb.i_mb_y;
1408 h->fdec->i_row_bits[y] += bits;
1409 rc->qpa_aq += h->mb.i_qp;
1411 if( h->mb.i_mb_x != h->mb.i_mb_width - 1 )
1415 rc->qpa_rc += rc->qpm * h->mb.i_mb_width;
1420 float qscale = qp2qscale( rc->qpm );
1421 h->fdec->f_row_qp[y] = rc->qpm;
1422 h->fdec->f_row_qscale[y] = qscale;
1424 update_predictor( rc->row_pred[0], qscale, h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1425 if( h->sh.i_type == SLICE_TYPE_P && rc->qpm < h->fref[0][0]->f_row_qp[y] )
1426 update_predictor( rc->row_pred[1], qscale, h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1428 /* update ratecontrol per-mbpair in MBAFF */
1429 if( SLICE_MBAFF && !(y&1) )
1432 /* tweak quality based on difference from predicted size */
1433 if( y < h->i_threadslice_end-1 )
1435 float prev_row_qp = h->fdec->f_row_qp[y];
1436 float qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1437 float qp_absolute_max = h->param.rc.i_qp_max;
1438 if( rc->rate_factor_max_increment )
1439 qp_absolute_max = X264_MIN( qp_absolute_max, rc->qp_novbv + rc->rate_factor_max_increment );
1440 float qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, qp_absolute_max );
1441 float step_size = 0.5f;
1443 /* B-frames shouldn't use lower QP than their reference frames. */
1444 if( h->sh.i_type == SLICE_TYPE_B )
1446 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] ) );
1447 rc->qpm = X264_MAX( rc->qpm, qp_min );
1450 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1451 float slice_size_planned = h->param.b_sliced_threads ? rc->slice_size_planned : rc->frame_size_planned;
1452 float max_frame_error = X264_MAX( 0.05f, 1.0f / h->mb.i_mb_height );
1453 float size_of_other_slices = 0;
1454 if( h->param.b_sliced_threads )
1456 float size_of_other_slices_planned = 0;
1457 for( int i = 0; i < h->param.i_threads; i++ )
1458 if( h != h->thread[i] )
1460 size_of_other_slices += h->thread[i]->rc->frame_size_estimated;
1461 size_of_other_slices_planned += h->thread[i]->rc->slice_size_planned;
1463 float weight = rc->slice_size_planned / rc->frame_size_planned;
1464 size_of_other_slices = (size_of_other_slices - size_of_other_slices_planned) * weight + size_of_other_slices_planned;
1467 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1468 float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
1469 float b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1471 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1472 /* area at the top of the frame was measured inaccurately. */
1473 if( row_bits_so_far( h, y ) < 0.05f * slice_size_planned )
1476 if( h->sh.i_type != SLICE_TYPE_I )
1479 if( !rc->b_vbv_min_rate )
1480 qp_min = X264_MAX( qp_min, rc->qp_novbv );
1482 while( rc->qpm < qp_max
1483 && ((b1 > rc->frame_size_planned + rc_tol) ||
1484 (rc->buffer_fill - b1 < buffer_left_planned * 0.5f) ||
1485 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1487 rc->qpm += step_size;
1488 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1491 while( rc->qpm > qp_min
1492 && (rc->qpm > h->fdec->f_row_qp[0] || rc->single_frame_vbv)
1493 && ((b1 < rc->frame_size_planned * 0.8f && rc->qpm <= prev_row_qp)
1494 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1f) )
1496 rc->qpm -= step_size;
1497 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1500 /* avoid VBV underflow or MinCR violation */
1501 while( (rc->qpm < qp_absolute_max)
1502 && ((rc->buffer_fill - b1 < rc->buffer_rate * max_frame_error) ||
1503 (rc->frame_size_maximum - b1 < rc->frame_size_maximum * max_frame_error)))
1505 rc->qpm += step_size;
1506 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1509 h->rc->frame_size_estimated = b1 - size_of_other_slices;
1512 h->rc->frame_size_estimated = predict_row_size_sum( h, y, rc->qpm );
1515 int x264_ratecontrol_qp( x264_t *h )
1518 return x264_clip3( h->rc->qpm + 0.5f, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1521 int x264_ratecontrol_mb_qp( x264_t *h )
1524 float qp = h->rc->qpm;
1525 if( h->param.rc.i_aq_mode )
1527 /* MB-tree currently doesn't adjust quantizers in unreferenced frames. */
1528 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];
1529 /* Scale AQ's effect towards zero in emergency mode. */
1530 if( qp > QP_MAX_SPEC )
1531 qp_offset *= (QP_MAX - qp) / (QP_MAX - QP_MAX_SPEC);
1534 return x264_clip3( qp + 0.5f, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1537 /* In 2pass, force the same frame types as in the 1st pass */
1538 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1540 x264_ratecontrol_t *rc = h->rc;
1541 if( h->param.rc.b_stat_read )
1543 if( frame_num >= rc->num_entries )
1545 /* We could try to initialize everything required for ABR and
1546 * adaptive B-frames, but that would be complicated.
1547 * So just calculate the average QP used so far. */
1548 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24 + QP_BD_OFFSET
1549 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1550 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, QP_MAX );
1551 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 );
1552 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 );
1554 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries );
1555 x264_log( h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant );
1556 if( h->param.i_bframe_adaptive )
1557 x264_log( h, X264_LOG_ERROR, "disabling adaptive B-frames\n" );
1559 for( int i = 0; i < h->param.i_threads; i++ )
1561 h->thread[i]->rc->b_abr = 0;
1562 h->thread[i]->rc->b_2pass = 0;
1563 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1564 h->thread[i]->param.rc.b_stat_read = 0;
1565 h->thread[i]->param.i_bframe_adaptive = 0;
1566 h->thread[i]->param.i_scenecut_threshold = 0;
1567 h->thread[i]->param.rc.b_mb_tree = 0;
1568 if( h->thread[i]->param.i_bframe > 1 )
1569 h->thread[i]->param.i_bframe = 1;
1571 return X264_TYPE_AUTO;
1573 return rc->entry[frame_num].frame_type;
1576 return X264_TYPE_AUTO;
1579 void x264_ratecontrol_set_weights( x264_t *h, x264_frame_t *frm )
1581 ratecontrol_entry_t *rce = &h->rc->entry[frm->i_frame];
1582 if( h->param.analyse.i_weighted_pred <= 0 )
1585 if( rce->i_weight_denom[0] >= 0 )
1586 SET_WEIGHT( frm->weight[0][0], 1, rce->weight[0][0], rce->i_weight_denom[0], rce->weight[0][1] );
1588 if( rce->i_weight_denom[1] >= 0 )
1590 SET_WEIGHT( frm->weight[0][1], 1, rce->weight[1][0], rce->i_weight_denom[1], rce->weight[1][1] );
1591 SET_WEIGHT( frm->weight[0][2], 1, rce->weight[2][0], rce->i_weight_denom[1], rce->weight[2][1] );
1595 /* After encoding one frame, save stats and update ratecontrol state */
1596 int x264_ratecontrol_end( x264_t *h, int bits, int *filler )
1598 x264_ratecontrol_t *rc = h->rc;
1599 const int *mbs = h->stat.frame.i_mb_count;
1603 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1604 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1605 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1606 for( int i = B_DIRECT; i < B_8x8; i++ )
1607 h->stat.frame.i_mb_count_p += mbs[i];
1609 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1610 h->fdec->f_qp_avg_aq = (float)rc->qpa_aq / h->mb.i_mb_count;
1612 if( h->param.rc.b_stat_write )
1614 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1615 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1616 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1617 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1618 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1619 char c_direct = h->mb.b_direct_auto_write ?
1620 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1621 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1623 if( fprintf( rc->p_stat_file_out,
1624 "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:",
1625 h->fenc->i_frame, h->i_frame,
1626 c_type, h->fenc->i_duration,
1627 h->fenc->i_cpb_duration, rc->qpa_rc,
1628 h->stat.frame.i_tex_bits,
1629 h->stat.frame.i_mv_bits,
1630 h->stat.frame.i_misc_bits,
1631 h->stat.frame.i_mb_count_i,
1632 h->stat.frame.i_mb_count_p,
1633 h->stat.frame.i_mb_count_skip,
1637 /* Only write information for reference reordering once. */
1638 int use_old_stats = h->param.rc.b_stat_read && rc->rce->refs > 1;
1639 for( int i = 0; i < (use_old_stats ? rc->rce->refs : h->i_ref[0]); i++ )
1641 int refcount = use_old_stats ? rc->rce->refcount[i]
1642 : PARAM_INTERLACED ? h->stat.frame.i_mb_count_ref[0][i*2]
1643 + h->stat.frame.i_mb_count_ref[0][i*2+1]
1644 : h->stat.frame.i_mb_count_ref[0][i];
1645 if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1649 if( h->param.analyse.i_weighted_pred >= X264_WEIGHTP_SIMPLE && h->sh.weight[0][0].weightfn )
1651 if( fprintf( rc->p_stat_file_out, "w:%d,%d,%d",
1652 h->sh.weight[0][0].i_denom, h->sh.weight[0][0].i_scale, h->sh.weight[0][0].i_offset ) < 0 )
1654 if( h->sh.weight[0][1].weightfn || h->sh.weight[0][2].weightfn )
1656 if( fprintf( rc->p_stat_file_out, ",%d,%d,%d,%d,%d ",
1657 h->sh.weight[0][1].i_denom, h->sh.weight[0][1].i_scale, h->sh.weight[0][1].i_offset,
1658 h->sh.weight[0][2].i_scale, h->sh.weight[0][2].i_offset ) < 0 )
1661 else if( fprintf( rc->p_stat_file_out, " " ) < 0 )
1665 if( fprintf( rc->p_stat_file_out, ";\n") < 0 )
1668 /* Don't re-write the data in multi-pass mode. */
1669 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1671 uint8_t i_type = h->sh.i_type;
1672 /* Values are stored as big-endian FIX8.8 */
1673 for( int i = 0; i < h->mb.i_mb_count; i++ )
1674 rc->qp_buffer[0][i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1675 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1677 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 )
1684 if( h->sh.i_type != SLICE_TYPE_B )
1685 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / rc->last_rceq;
1688 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1689 * Not perfectly accurate with B-refs, but good enough. */
1690 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / (rc->last_rceq * fabs( h->param.rc.f_pb_factor ));
1692 rc->cplxr_sum *= rc->cbr_decay;
1693 rc->wanted_bits_window += h->fenc->f_duration * rc->bitrate;
1694 rc->wanted_bits_window *= rc->cbr_decay;
1698 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale( rc->rce->new_qp ) );
1700 if( h->mb.b_variable_qp )
1702 if( h->sh.i_type == SLICE_TYPE_B )
1704 rc->bframe_bits += bits;
1705 if( h->fenc->b_last_minigop_bframe )
1707 update_predictor( rc->pred_b_from_p, qp2qscale( rc->qpa_rc ),
1708 h->fref[1][h->i_ref[1]-1]->i_satd, rc->bframe_bits / rc->bframes );
1709 rc->bframe_bits = 0;
1714 *filler = update_vbv( h, bits );
1715 rc->filler_bits_sum += *filler * 8;
1717 if( h->sps->vui.b_nal_hrd_parameters_present )
1719 if( h->fenc->i_frame == 0 )
1721 // access unit initialises the HRD
1722 h->fenc->hrd_timing.cpb_initial_arrival_time = 0;
1723 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1724 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1725 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit = (double)rc->initial_cpb_removal_delay / 90000;
1729 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit + (double)(h->fenc->i_cpb_delay - h->i_cpb_delay_pir_offset) *
1730 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1732 double cpb_earliest_arrival_time = h->fenc->hrd_timing.cpb_removal_time - (double)rc->initial_cpb_removal_delay / 90000;
1733 if( h->fenc->b_keyframe )
1735 rc->nrt_first_access_unit = h->fenc->hrd_timing.cpb_removal_time;
1736 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1737 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1740 cpb_earliest_arrival_time -= (double)rc->initial_cpb_removal_delay_offset / 90000;
1742 if( h->sps->vui.hrd.b_cbr_hrd )
1743 h->fenc->hrd_timing.cpb_initial_arrival_time = rc->previous_cpb_final_arrival_time;
1745 h->fenc->hrd_timing.cpb_initial_arrival_time = X264_MAX( rc->previous_cpb_final_arrival_time, cpb_earliest_arrival_time );
1747 int filler_bits = *filler ? X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), *filler )*8 : 0;
1749 h->fenc->hrd_timing.cpb_final_arrival_time = rc->previous_cpb_final_arrival_time = h->fenc->hrd_timing.cpb_initial_arrival_time +
1750 (double)(bits + filler_bits) / h->sps->vui.hrd.i_bit_rate_unscaled;
1752 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 +
1753 h->fenc->hrd_timing.cpb_removal_time;
1758 x264_log( h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n" );
1762 /****************************************************************************
1764 ***************************************************************************/
1767 * modify the bitrate curve from pass1 for one frame
1769 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1771 x264_ratecontrol_t *rcc= h->rc;
1772 x264_zone_t *zone = get_zone( h, frame_num );
1774 if( h->param.rc.b_mb_tree )
1776 double timescale = (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1777 q = pow( BASE_FRAME_DURATION / CLIP_DURATION(rce->i_duration * timescale), 1 - h->param.rc.f_qcompress );
1780 q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1782 // avoid NaN's in the rc_eq
1783 if( !isfinite(q) || rce->tex_bits + rce->mv_bits == 0 )
1784 q = rcc->last_qscale_for[rce->pict_type];
1789 rcc->last_qscale = q;
1794 if( zone->b_force_qp )
1795 q = qp2qscale( zone->i_qp );
1797 q /= zone->f_bitrate_factor;
1803 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q, int frame_num)
1805 x264_ratecontrol_t *rcc = h->rc;
1806 const int pict_type = rce->pict_type;
1807 x264_zone_t *zone = get_zone( h, frame_num );
1809 // force I/B quants as a function of P quants
1810 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1811 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1812 if( pict_type == SLICE_TYPE_I )
1815 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1816 double ip_factor = fabs( h->param.rc.f_ip_factor );
1817 /* don't apply ip_factor if the following frame is also I */
1818 if( rcc->accum_p_norm <= 0 )
1820 else if( h->param.rc.f_ip_factor < 0 )
1822 else if( rcc->accum_p_norm >= 1 )
1825 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1827 else if( pict_type == SLICE_TYPE_B )
1829 if( h->param.rc.f_pb_factor > 0 )
1831 if( !rce->kept_as_ref )
1832 q *= fabs( h->param.rc.f_pb_factor );
1834 else if( pict_type == SLICE_TYPE_P
1835 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1836 && rce->tex_bits == 0 )
1841 /* last qscale / qdiff stuff */
1842 if( rcc->last_non_b_pict_type == pict_type &&
1843 (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1) )
1845 double last_q = rcc->last_qscale_for[pict_type];
1846 double max_qscale = last_q * rcc->lstep;
1847 double min_qscale = last_q / rcc->lstep;
1849 if ( q > max_qscale ) q = max_qscale;
1850 else if( q < min_qscale ) q = min_qscale;
1853 rcc->last_qscale_for[pict_type] = q;
1854 if( pict_type != SLICE_TYPE_B )
1855 rcc->last_non_b_pict_type = pict_type;
1856 if( pict_type == SLICE_TYPE_I )
1858 rcc->last_accum_p_norm = rcc->accum_p_norm;
1859 rcc->accum_p_norm = 0;
1860 rcc->accum_p_qp = 0;
1862 if( pict_type == SLICE_TYPE_P )
1864 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1865 rcc->accum_p_qp = mask * (qscale2qp( q ) + rcc->accum_p_qp);
1866 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1871 if( zone->b_force_qp )
1872 q = qp2qscale( zone->i_qp );
1874 q /= zone->f_bitrate_factor;
1880 static float predict_size( predictor_t *p, float q, float var )
1882 return (p->coeff*var + p->offset) / (q*p->count);
1885 static void update_predictor( predictor_t *p, float q, float var, float bits )
1890 float old_coeff = p->coeff / p->count;
1891 float new_coeff = bits*q / var;
1892 float new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
1893 float new_offset = bits*q - new_coeff_clipped * var;
1894 if( new_offset >= 0 )
1895 new_coeff = new_coeff_clipped;
1898 p->count *= p->decay;
1899 p->coeff *= p->decay;
1900 p->offset *= p->decay;
1902 p->coeff += new_coeff;
1903 p->offset += new_offset;
1906 // update VBV after encoding a frame
1907 static int update_vbv( x264_t *h, int bits )
1910 int bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
1911 x264_ratecontrol_t *rcc = h->rc;
1912 x264_ratecontrol_t *rct = h->thread[0]->rc;
1913 uint64_t buffer_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
1915 if( rcc->last_satd >= h->mb.i_mb_count )
1916 update_predictor( &rct->pred[h->sh.i_type], qp2qscale( rcc->qpa_rc ), rcc->last_satd, bits );
1921 rct->buffer_fill_final -= (uint64_t)bits * h->sps->vui.i_time_scale;
1923 if( rct->buffer_fill_final < 0 )
1924 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 );
1925 rct->buffer_fill_final = X264_MAX( rct->buffer_fill_final, 0 );
1926 rct->buffer_fill_final += (uint64_t)bitrate * h->sps->vui.i_num_units_in_tick * h->fenc->i_cpb_duration;
1928 if( h->sps->vui.hrd.b_cbr_hrd && rct->buffer_fill_final > buffer_size )
1930 int64_t scale = (int64_t)h->sps->vui.i_time_scale * 8;
1931 filler = (rct->buffer_fill_final - buffer_size + scale - 1) / scale;
1932 bits = X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), filler ) * 8;
1933 rct->buffer_fill_final -= (uint64_t)bits * h->sps->vui.i_time_scale;
1936 rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, buffer_size );
1941 void x264_hrd_fullness( x264_t *h )
1943 x264_ratecontrol_t *rct = h->thread[0]->rc;
1944 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale / rct->hrd_multiply_denom;
1945 uint64_t cpb_state = rct->buffer_fill_final;
1946 uint64_t cpb_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
1947 uint64_t multiply_factor = 180000 / rct->hrd_multiply_denom;
1949 if( rct->buffer_fill_final < 0 || rct->buffer_fill_final > cpb_size )
1951 x264_log( h, X264_LOG_WARNING, "CPB %s: %.0lf bits in a %.0lf-bit buffer\n",
1952 rct->buffer_fill_final < 0 ? "underflow" : "overflow", (float)rct->buffer_fill_final/denom, (float)cpb_size/denom );
1955 h->initial_cpb_removal_delay = (multiply_factor * cpb_state + denom) / (2*denom);
1956 h->initial_cpb_removal_delay_offset = (multiply_factor * cpb_size + denom) / (2*denom) - h->initial_cpb_removal_delay;
1959 // provisionally update VBV according to the planned size of all frames currently in progress
1960 static void update_vbv_plan( x264_t *h, int overhead )
1962 x264_ratecontrol_t *rcc = h->rc;
1963 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final / h->sps->vui.i_time_scale;
1964 if( h->i_thread_frames > 1 )
1966 int j = h->rc - h->thread[0]->rc;
1967 for( int i = 1; i < h->i_thread_frames; i++ )
1969 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
1970 double bits = t->rc->frame_size_planned;
1971 if( !t->b_thread_active )
1973 bits = X264_MAX(bits, t->rc->frame_size_estimated);
1974 rcc->buffer_fill -= bits;
1975 rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
1976 rcc->buffer_fill += t->rc->buffer_rate;
1977 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1980 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1981 rcc->buffer_fill -= overhead;
1984 // apply VBV constraints and clip qscale to between lmin and lmax
1985 static double clip_qscale( x264_t *h, int pict_type, double q )
1987 x264_ratecontrol_t *rcc = h->rc;
1988 double lmin = rcc->lmin[pict_type];
1989 double lmax = rcc->lmax[pict_type];
1990 if( rcc->rate_factor_max_increment )
1991 lmax = X264_MIN( lmax, qp2qscale( rcc->qp_novbv + rcc->rate_factor_max_increment ) );
1994 /* B-frames are not directly subject to VBV,
1995 * since they are controlled by the P-frames' QPs. */
1997 if( rcc->b_vbv && rcc->last_satd > 0 )
1999 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
2000 * the lookahead overflow and such that the buffer is in a reasonable state
2001 * by the end of the lookahead. */
2002 if( h->param.rc.i_lookahead )
2006 /* Avoid an infinite loop. */
2007 for( int iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
2010 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2011 double buffer_fill_cur = rcc->buffer_fill - cur_bits;
2013 double total_duration = 0;
2014 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
2015 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
2016 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
2018 /* Loop over the planned future frames. */
2019 for( int j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
2021 total_duration += h->fenc->f_planned_cpb_duration[j];
2022 buffer_fill_cur += rcc->vbv_max_rate * h->fenc->f_planned_cpb_duration[j];
2023 int i_type = h->fenc->i_planned_type[j];
2024 int i_satd = h->fenc->i_planned_satd[j];
2025 if( i_type == X264_TYPE_AUTO )
2027 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
2028 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
2029 buffer_fill_cur -= cur_bits;
2031 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
2032 target_fill = X264_MIN( rcc->buffer_fill + total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.5 );
2033 if( buffer_fill_cur < target_fill )
2039 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
2040 target_fill = x264_clip3f( rcc->buffer_fill - total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
2041 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
2050 /* Fallback to old purely-reactive algorithm: no lookahead. */
2053 if( ( pict_type == SLICE_TYPE_P ||
2054 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
2055 rcc->buffer_fill/rcc->buffer_size < 0.5 )
2057 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
2060 /* Now a hard threshold to make sure the frame fits in VBV.
2061 * This one is mostly for I-frames. */
2062 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2064 /* For small VBVs, allow the frame to use up the entire VBV. */
2065 double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
2066 /* For single-frame VBVs, request that the frame use up the entire VBV. */
2067 double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
2069 if( bits > rcc->buffer_fill/max_fill_factor )
2070 qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
2073 if( bits < rcc->buffer_rate/min_fill_factor )
2074 q *= bits*min_fill_factor/rcc->buffer_rate;
2075 q = X264_MAX( q0, q );
2078 /* Apply MinCR restrictions */
2079 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2080 if( bits > rcc->frame_size_maximum )
2081 q *= bits / rcc->frame_size_maximum;
2082 bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2084 /* Check B-frame complexity, and use up any bits that would
2085 * overflow before the next P-frame. */
2086 if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
2088 int nb = rcc->bframes;
2089 double pbbits = bits;
2090 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
2092 double bframe_cpb_duration = 0;
2093 double minigop_cpb_duration;
2094 for( int i = 0; i < nb; i++ )
2095 bframe_cpb_duration += h->fenc->f_planned_cpb_duration[1+i];
2097 if( bbits * nb > bframe_cpb_duration * rcc->vbv_max_rate )
2099 pbbits += nb * bbits;
2101 minigop_cpb_duration = bframe_cpb_duration + h->fenc->f_planned_cpb_duration[0];
2102 space = rcc->buffer_fill + minigop_cpb_duration*rcc->vbv_max_rate - rcc->buffer_size;
2103 if( pbbits < space )
2105 q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
2107 q = X264_MAX( q0/2, q );
2110 if( !rcc->b_vbv_min_rate )
2111 q = X264_MAX( q0, q );
2116 else if( rcc->b_2pass )
2118 double min2 = log( lmin );
2119 double max2 = log( lmax );
2120 q = (log(q) - min2)/(max2-min2) - 0.5;
2121 q = 1.0/(1.0 + exp( -4*q ));
2122 q = q*(max2-min2) + min2;
2126 return x264_clip3f( q, lmin, lmax );
2129 // update qscale for 1 frame based on actual bits used so far
2130 static float rate_estimate_qscale( x264_t *h )
2133 x264_ratecontrol_t *rcc = h->rc;
2134 ratecontrol_entry_t UNINIT(rce);
2135 int pict_type = h->sh.i_type;
2136 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
2137 + h->stat.i_frame_size[SLICE_TYPE_P]
2138 + h->stat.i_frame_size[SLICE_TYPE_B])
2139 - rcc->filler_bits_sum;
2144 if( pict_type != rce.pict_type )
2146 x264_log( h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
2147 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type] );
2151 if( pict_type == SLICE_TYPE_B )
2153 /* B-frames don't have independent ratecontrol, but rather get the
2154 * average QP of the two adjacent P-frames + an offset */
2156 int i0 = IS_X264_TYPE_I(h->fref_nearest[0]->i_type);
2157 int i1 = IS_X264_TYPE_I(h->fref_nearest[1]->i_type);
2158 int dt0 = abs(h->fenc->i_poc - h->fref_nearest[0]->i_poc);
2159 int dt1 = abs(h->fenc->i_poc - h->fref_nearest[1]->i_poc);
2160 float q0 = h->fref_nearest[0]->f_qp_avg_rc;
2161 float q1 = h->fref_nearest[1]->f_qp_avg_rc;
2163 if( h->fref_nearest[0]->i_type == X264_TYPE_BREF )
2164 q0 -= rcc->pb_offset/2;
2165 if( h->fref_nearest[1]->i_type == X264_TYPE_BREF )
2166 q1 -= rcc->pb_offset/2;
2169 q = (q0 + q1) / 2 + rcc->ip_offset;
2175 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
2177 if( h->fenc->b_kept_as_ref )
2178 q += rcc->pb_offset/2;
2180 q += rcc->pb_offset;
2182 if( rcc->b_2pass && rcc->b_vbv )
2183 rcc->frame_size_planned = qscale2bits( &rce, qp2qscale( q ) );
2185 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, qp2qscale( q ), h->fref[1][h->i_ref[1]-1]->i_satd );
2186 /* Limit planned size by MinCR */
2188 rcc->frame_size_planned = X264_MIN( rcc->frame_size_planned, rcc->frame_size_maximum );
2189 h->rc->frame_size_estimated = rcc->frame_size_planned;
2193 rcc->last_satd = x264_rc_analyse_slice( h );
2195 return qp2qscale( q );
2199 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
2203 double lmin = rcc->lmin[pict_type];
2204 double lmax = rcc->lmax[pict_type];
2206 int64_t predicted_bits = total_bits;
2210 if( h->i_thread_frames > 1 )
2212 int j = h->rc - h->thread[0]->rc;
2213 for( int i = 1; i < h->i_thread_frames; i++ )
2215 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
2216 double bits = t->rc->frame_size_planned;
2217 if( !t->b_thread_active )
2219 bits = X264_MAX(bits, t->rc->frame_size_estimated);
2220 predicted_bits += (int64_t)bits;
2226 if( h->i_frame < h->i_thread_frames )
2227 predicted_bits += (int64_t)h->i_frame * rcc->bitrate / rcc->fps;
2229 predicted_bits += (int64_t)(h->i_thread_frames - 1) * rcc->bitrate / rcc->fps;
2232 /* Adjust ABR buffer based on distance to the end of the video. */
2233 if( rcc->num_entries > h->i_frame )
2235 double final_bits = rcc->entry[rcc->num_entries-1].expected_bits;
2236 double video_pos = rce.expected_bits / final_bits;
2237 double scale_factor = sqrt( (1 - video_pos) * rcc->num_entries );
2238 abr_buffer *= 0.5 * X264_MAX( scale_factor, 0.5 );
2241 diff = predicted_bits - (int64_t)rce.expected_bits;
2243 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
2244 if( ((h->i_frame + 1 - h->i_thread_frames) >= rcc->fps) &&
2245 (rcc->expected_bits_sum > 0))
2247 /* Adjust quant based on the difference between
2248 * achieved and expected bitrate so far */
2249 double cur_time = (double)h->i_frame / rcc->num_entries;
2250 double w = x264_clip3f( cur_time*100, 0.0, 1.0 );
2251 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
2255 /* Do not overflow vbv */
2256 double expected_size = qscale2bits( &rce, q );
2257 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2258 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
2259 double qmax = q*(2 - expected_fullness);
2260 double size_constraint = 1 + expected_fullness;
2261 qmax = X264_MAX( qmax, rce.new_qscale );
2262 if( expected_fullness < .05 )
2264 qmax = X264_MIN(qmax, lmax);
2265 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
2266 ((expected_vbv < 0) && (q < lmax)))
2269 expected_size = qscale2bits(&rce, q);
2270 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2272 rcc->last_satd = x264_rc_analyse_slice( h );
2274 q = x264_clip3f( q, lmin, lmax );
2276 else /* 1pass ABR */
2278 /* Calculate the quantizer which would have produced the desired
2279 * average bitrate if it had been applied to all frames so far.
2280 * Then modulate that quant based on the current frame's complexity
2281 * relative to the average complexity so far (using the 2pass RCEQ).
2282 * Then bias the quant up or down if total size so far was far from
2284 * Result: Depending on the value of rate_tolerance, there is a
2285 * tradeoff between quality and bitrate precision. But at large
2286 * tolerances, the bit distribution approaches that of 2pass. */
2288 double wanted_bits, overflow = 1;
2290 rcc->last_satd = x264_rc_analyse_slice( h );
2291 rcc->short_term_cplxsum *= 0.5;
2292 rcc->short_term_cplxcount *= 0.5;
2293 rcc->short_term_cplxsum += rcc->last_satd / (CLIP_DURATION(h->fenc->f_duration) / BASE_FRAME_DURATION);
2294 rcc->short_term_cplxcount ++;
2296 rce.tex_bits = rcc->last_satd;
2297 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
2299 rce.p_count = rcc->nmb;
2303 rce.pict_type = pict_type;
2304 rce.i_duration = h->fenc->i_duration;
2306 if( h->param.rc.i_rc_method == X264_RC_CRF )
2308 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
2312 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
2314 /* ABR code can potentially be counterproductive in CBR, so just don't bother.
2315 * Don't run it if the frame complexity is zero either. */
2316 if( !rcc->b_vbv_min_rate && rcc->last_satd )
2318 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
2319 int i_frame_done = h->i_frame + 1 - h->i_thread_frames;
2320 double time_done = i_frame_done / rcc->fps;
2321 if( h->param.b_vfr_input && i_frame_done > 0 )
2322 time_done = ((double)(h->fenc->i_reordered_pts - h->i_reordered_pts_delay)) * h->param.i_timebase_num / h->param.i_timebase_den;
2323 wanted_bits = time_done * rcc->bitrate;
2324 if( wanted_bits > 0 )
2326 abr_buffer *= X264_MAX( 1, sqrt( time_done ) );
2327 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
2333 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
2334 /* should test _next_ pict type, but that isn't decided yet */
2335 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
2337 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2338 q /= fabs( h->param.rc.f_ip_factor );
2340 else if( h->i_frame > 0 )
2342 if( h->param.rc.i_rc_method != X264_RC_CRF )
2344 /* Asymmetric clipping, because symmetric would prevent
2345 * overflow control in areas of rapidly oscillating complexity */
2346 double lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
2347 double lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
2348 if( overflow > 1.1 && h->i_frame > 3 )
2350 else if( overflow < 0.9 )
2353 q = x264_clip3f(q, lmin, lmax);
2356 else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
2358 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
2360 rcc->qp_novbv = qscale2qp( q );
2362 //FIXME use get_diff_limited_q() ?
2363 q = clip_qscale( h, pict_type, q );
2366 rcc->last_qscale_for[pict_type] =
2367 rcc->last_qscale = q;
2369 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
2370 rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );
2372 if( rcc->b_2pass && rcc->b_vbv )
2373 rcc->frame_size_planned = qscale2bits(&rce, q);
2375 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2377 /* Always use up the whole VBV in this case. */
2378 if( rcc->single_frame_vbv )
2379 rcc->frame_size_planned = rcc->buffer_rate;
2380 /* Limit planned size by MinCR */
2382 rcc->frame_size_planned = X264_MIN( rcc->frame_size_planned, rcc->frame_size_maximum );
2383 h->rc->frame_size_estimated = rcc->frame_size_planned;
2388 static void x264_threads_normalize_predictors( x264_t *h )
2390 double totalsize = 0;
2391 for( int i = 0; i < h->param.i_threads; i++ )
2392 totalsize += h->thread[i]->rc->slice_size_planned;
2393 double factor = h->rc->frame_size_planned / totalsize;
2394 for( int i = 0; i < h->param.i_threads; i++ )
2395 h->thread[i]->rc->slice_size_planned *= factor;
2398 void x264_threads_distribute_ratecontrol( x264_t *h )
2401 x264_ratecontrol_t *rc = h->rc;
2403 /* Initialize row predictors */
2404 if( h->i_frame == 0 )
2405 for( int i = 0; i < h->param.i_threads; i++ )
2407 x264_ratecontrol_t *t = h->thread[i]->rc;
2408 memcpy( t->row_preds, rc->row_preds, sizeof(rc->row_preds) );
2411 for( int i = 0; i < h->param.i_threads; i++ )
2413 x264_t *t = h->thread[i];
2414 memcpy( t->rc, rc, offsetof(x264_ratecontrol_t, row_pred) );
2415 t->rc->row_pred = &t->rc->row_preds[h->sh.i_type];
2416 /* Calculate the planned slice size. */
2417 if( rc->b_vbv && rc->frame_size_planned )
2420 for( row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2421 size += h->fdec->i_row_satd[row];
2422 t->rc->slice_size_planned = predict_size( &rc->pred[h->sh.i_type + (i+1)*5], rc->qpm, size );
2425 t->rc->slice_size_planned = 0;
2427 if( rc->b_vbv && rc->frame_size_planned )
2429 x264_threads_normalize_predictors( h );
2431 if( rc->single_frame_vbv )
2433 /* Compensate for our max frame error threshold: give more bits (proportionally) to smaller slices. */
2434 for( int i = 0; i < h->param.i_threads; i++ )
2436 x264_t *t = h->thread[i];
2437 float max_frame_error = X264_MAX( 0.05, 1.0 / (t->i_threadslice_end - t->i_threadslice_start) );
2438 t->rc->slice_size_planned += 2 * max_frame_error * rc->frame_size_planned;
2440 x264_threads_normalize_predictors( h );
2443 for( int i = 0; i < h->param.i_threads; i++ )
2444 h->thread[i]->rc->frame_size_estimated = h->thread[i]->rc->slice_size_planned;
2448 void x264_threads_merge_ratecontrol( x264_t *h )
2450 x264_ratecontrol_t *rc = h->rc;
2453 for( int i = 0; i < h->param.i_threads; i++ )
2455 x264_t *t = h->thread[i];
2456 x264_ratecontrol_t *rct = h->thread[i]->rc;
2457 if( h->param.rc.i_vbv_buffer_size )
2460 for( int row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2461 size += h->fdec->i_row_satd[row];
2462 int bits = t->stat.frame.i_mv_bits + t->stat.frame.i_tex_bits + t->stat.frame.i_misc_bits;
2463 int mb_count = (t->i_threadslice_end - t->i_threadslice_start) * h->mb.i_mb_width;
2464 update_predictor( &rc->pred[h->sh.i_type+(i+1)*5], qp2qscale( rct->qpa_rc/mb_count ), size, bits );
2468 rc->qpa_rc += rct->qpa_rc;
2469 rc->qpa_aq += rct->qpa_aq;
2473 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
2477 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
2478 /* these vars are updated in x264_ratecontrol_start()
2479 * so copy them from the context that most recently started (prev)
2480 * to the context that's about to start (cur). */
2485 COPY(last_qscale_for);
2486 COPY(last_non_b_pict_type);
2487 COPY(short_term_cplxsum);
2488 COPY(short_term_cplxcount);
2492 /* these vars can be updated by x264_ratecontrol_init_reconfigurable */
2497 COPY(single_frame_vbv);
2499 COPY(rate_factor_constant);
2500 COPY(rate_factor_max_increment);
2505 #define COPY(var) next->rc->var = cur->rc->var
2506 /* these vars are updated in x264_ratecontrol_end()
2507 * so copy them from the context that most recently ended (cur)
2508 * to the context that's about to end (next) */
2510 COPY(expected_bits_sum);
2511 COPY(filler_bits_sum);
2512 COPY(wanted_bits_window);
2514 COPY(initial_cpb_removal_delay);
2515 COPY(initial_cpb_removal_delay_offset);
2516 COPY(nrt_first_access_unit);
2517 COPY(previous_cpb_final_arrival_time);
2520 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
2521 /* the rest of the variables are either constant or thread-local */
2524 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
2526 /* find an interval ending on an overflow or underflow (depending on whether
2527 * we're adding or removing bits), and starting on the earliest frame that
2528 * can influence the buffer fill of that end frame. */
2529 x264_ratecontrol_t *rcc = h->rc;
2530 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
2531 const double buffer_max = .9 * rcc->buffer_size;
2532 double fill = fills[*t0-1];
2533 double parity = over ? 1. : -1.;
2534 int start = -1, end = -1;
2535 for( int i = *t0; i < rcc->num_entries; i++ )
2537 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 -
2538 qscale2bits( &rcc->entry[i], rcc->entry[i].new_qscale )) * parity;
2539 fill = x264_clip3f(fill, 0, rcc->buffer_size);
2541 if( fill <= buffer_min || i == 0 )
2547 else if( fill >= buffer_max && start >= 0 )
2552 return start >= 0 && end >= 0;
2555 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
2557 x264_ratecontrol_t *rcc = h->rc;
2558 double qscale_orig, qscale_new;
2562 for( int i = t0; i <= t1; i++ )
2564 qscale_orig = rcc->entry[i].new_qscale;
2565 qscale_orig = x264_clip3f( qscale_orig, qscale_min, qscale_max );
2566 qscale_new = qscale_orig * adjustment;
2567 qscale_new = x264_clip3f( qscale_new, qscale_min, qscale_max );
2568 rcc->entry[i].new_qscale = qscale_new;
2569 adjusted = adjusted || (qscale_new != qscale_orig);
2574 static double count_expected_bits( x264_t *h )
2576 x264_ratecontrol_t *rcc = h->rc;
2577 double expected_bits = 0;
2578 for( int i = 0; i < rcc->num_entries; i++ )
2580 ratecontrol_entry_t *rce = &rcc->entry[i];
2581 rce->expected_bits = expected_bits;
2582 expected_bits += qscale2bits( rce, rce->new_qscale );
2584 return expected_bits;
2587 static int vbv_pass2( x264_t *h, double all_available_bits )
2589 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
2590 * frames in the interval until either buffer is full at some intermediate frame or the
2591 * last frame in the interval no longer underflows. Recompute intervals and repeat.
2592 * Then do the converse to put bits back into overflow areas until target size is met */
2594 x264_ratecontrol_t *rcc = h->rc;
2596 double expected_bits = 0;
2598 double prev_bits = 0;
2600 double qscale_min = qp2qscale( h->param.rc.i_qp_min );
2601 double qscale_max = qp2qscale( h->param.rc.i_qp_max );
2603 int adj_min, adj_max;
2604 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2608 /* adjust overall stream size */
2612 prev_bits = expected_bits;
2615 { /* not first iteration */
2616 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2617 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2621 while(adj_min && find_underflow( h, fills, &t0, &t1, 1 ))
2623 adj_min = fix_underflow( h, t0, t1, adjustment, qscale_min, qscale_max );
2628 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2630 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2632 while( adj_max && find_underflow( h, fills, &t0, &t1, 0 ) )
2633 adj_max = fix_underflow( h, t0, t1, 1.001, qscale_min, qscale_max );
2635 expected_bits = count_expected_bits( h );
2636 } while( (expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2639 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2641 /* store expected vbv filling values for tracking when encoding */
2642 for( int i = 0; i < rcc->num_entries; i++ )
2643 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
2645 x264_free( fills-1 );
2651 static int init_pass2( x264_t *h )
2653 x264_ratecontrol_t *rcc = h->rc;
2654 uint64_t all_const_bits = 0;
2655 double timescale = (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2656 double duration = 0;
2657 for( int i = 0; i < rcc->num_entries; i++ )
2658 duration += rcc->entry[i].i_duration;
2659 duration *= timescale;
2660 uint64_t all_available_bits = h->param.rc.i_bitrate * 1000. * duration;
2661 double rate_factor, step_mult;
2662 double qblur = h->param.rc.f_qblur;
2663 double cplxblur = h->param.rc.f_complexity_blur;
2664 const int filter_size = (int)(qblur*4) | 1;
2665 double expected_bits;
2666 double *qscale, *blurred_qscale;
2667 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
2669 /* find total/average complexity & const_bits */
2670 for( int i = 0; i < rcc->num_entries; i++ )
2672 ratecontrol_entry_t *rce = &rcc->entry[i];
2673 all_const_bits += rce->misc_bits;
2676 if( all_available_bits < all_const_bits)
2678 x264_log( h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2679 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)) );
2683 /* Blur complexities, to reduce local fluctuation of QP.
2684 * We don't blur the QPs directly, because then one very simple frame
2685 * could drag down the QP of a nearby complex frame and give it more
2686 * bits than intended. */
2687 for( int i = 0; i < rcc->num_entries; i++ )
2689 ratecontrol_entry_t *rce = &rcc->entry[i];
2690 double weight_sum = 0;
2691 double cplx_sum = 0;
2692 double weight = 1.0;
2693 double gaussian_weight;
2694 /* weighted average of cplx of future frames */
2695 for( int j = 1; j < cplxblur*2 && j < rcc->num_entries-i; j++ )
2697 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2698 double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2699 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2700 if( weight < .0001 )
2702 gaussian_weight = weight * exp( -j*j/200.0 );
2703 weight_sum += gaussian_weight;
2704 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2706 /* weighted average of cplx of past frames */
2708 for( int j = 0; j <= cplxblur*2 && j <= i; j++ )
2710 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2711 double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2712 gaussian_weight = weight * exp( -j*j/200.0 );
2713 weight_sum += gaussian_weight;
2714 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2715 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2716 if( weight < .0001 )
2719 rce->blurred_complexity = cplx_sum / weight_sum;
2722 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2723 if( filter_size > 1 )
2724 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2726 blurred_qscale = qscale;
2728 /* Search for a factor which, when multiplied by the RCEQ values from
2729 * each frame, adds up to the desired total size.
2730 * There is no exact closed-form solution because of VBV constraints and
2731 * because qscale2bits is not invertible, but we can start with the simple
2732 * approximation of scaling the 1st pass by the ratio of bitrates.
2733 * The search range is probably overkill, but speed doesn't matter here. */
2736 for( int i = 0; i < rcc->num_entries; i++ )
2738 double q = get_qscale(h, &rcc->entry[i], 1.0, i);
2739 expected_bits += qscale2bits(&rcc->entry[i], q);
2740 rcc->last_qscale_for[rcc->entry[i].pict_type] = q;
2742 step_mult = all_available_bits / expected_bits;
2745 for( double step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2748 rate_factor += step;
2750 rcc->last_non_b_pict_type = -1;
2751 rcc->last_accum_p_norm = 1;
2752 rcc->accum_p_norm = 0;
2754 rcc->last_qscale_for[0] =
2755 rcc->last_qscale_for[1] =
2756 rcc->last_qscale_for[2] = pow( base_cplx, 1 - rcc->qcompress ) / rate_factor;
2759 for( int i = 0; i < rcc->num_entries; i++ )
2761 qscale[i] = get_qscale( h, &rcc->entry[i], rate_factor, -1 );
2762 rcc->last_qscale_for[rcc->entry[i].pict_type] = qscale[i];
2765 /* fixed I/B qscale relative to P */
2766 for( int i = rcc->num_entries-1; i >= 0; i-- )
2768 qscale[i] = get_diff_limited_q( h, &rcc->entry[i], qscale[i], i );
2769 assert(qscale[i] >= 0);
2773 if( filter_size > 1 )
2775 assert( filter_size%2 == 1 );
2776 for( int i = 0; i < rcc->num_entries; i++ )
2778 ratecontrol_entry_t *rce = &rcc->entry[i];
2779 double q = 0.0, sum = 0.0;
2781 for( int j = 0; j < filter_size; j++ )
2783 int idx = i+j-filter_size/2;
2785 double coeff = qblur==0 ? 1.0 : exp( -d*d/(qblur*qblur) );
2786 if( idx < 0 || idx >= rcc->num_entries )
2788 if( rce->pict_type != rcc->entry[idx].pict_type )
2790 q += qscale[idx] * coeff;
2793 blurred_qscale[i] = q/sum;
2797 /* find expected bits */
2798 for( int i = 0; i < rcc->num_entries; i++ )
2800 ratecontrol_entry_t *rce = &rcc->entry[i];
2801 rce->new_qscale = clip_qscale( h, rce->pict_type, blurred_qscale[i] );
2802 assert(rce->new_qscale >= 0);
2803 expected_bits += qscale2bits( rce, rce->new_qscale );
2806 if( expected_bits > all_available_bits )
2807 rate_factor -= step;
2810 x264_free( qscale );
2811 if( filter_size > 1 )
2812 x264_free( blurred_qscale );
2815 if( vbv_pass2( h, all_available_bits ) )
2817 expected_bits = count_expected_bits( h );
2819 if( fabs( expected_bits/all_available_bits - 1.0 ) > 0.01 )
2822 for( int i = 0; i < rcc->num_entries; i++ )
2823 avgq += rcc->entry[i].new_qscale;
2824 avgq = qscale2qp( avgq / rcc->num_entries );
2826 if( expected_bits > all_available_bits || !rcc->b_vbv )
2827 x264_log( h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n" );
2828 x264_log( h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
2829 (float)h->param.rc.i_bitrate,
2830 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
2832 if( expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2 )
2834 if( h->param.rc.i_qp_min > 0 )
2835 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min );
2837 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate\n" );
2839 else if( expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2 )
2841 if( h->param.rc.i_qp_max < QP_MAX )
2842 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max );
2844 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate\n");
2846 else if( !(rcc->b_2pass && rcc->b_vbv) )
2847 x264_log( h, X264_LOG_WARNING, "internal error\n" );