1 /*****************************************************************************
2 * ratecontrol.c: ratecontrol
3 *****************************************************************************
4 * Copyright (C) 2005-2016 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 #undef NDEBUG // always check asserts, the speed effect is far too small to disable them
32 #include "common/common.h"
33 #include "ratecontrol.h"
45 double expected_bits; /* total expected bits up to the current frame (current one excluded) */
52 float blurred_complexity;
55 int16_t i_weight_denom[2];
59 int64_t i_cpb_duration;
61 } ratecontrol_entry_t;
72 struct x264_ratecontrol_t
81 double rate_tolerance;
83 int nmb; /* number of macroblocks in a frame */
87 ratecontrol_entry_t *rce;
88 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 int64_t buffer_fill_final_min;
99 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
100 double buffer_rate; /* # of bits added to buffer_fill after each frame */
101 double vbv_max_rate; /* # of bits added to buffer_fill per second */
102 predictor_t *pred; /* predict frame size from satd */
103 int single_frame_vbv;
104 float rate_factor_max_increment; /* Don't allow RF above (CRF + this value). */
109 double cplxr_sum; /* sum of bits*qscale/rceq */
110 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow, only includes finished frames */
111 int64_t filler_bits_sum; /* sum in bits of finished frames' filler data */
112 double wanted_bits_window; /* target bitrate * window */
114 double short_term_cplxsum;
115 double short_term_cplxcount;
116 double rate_factor_constant;
121 FILE *p_stat_file_out;
122 char *psz_stat_file_tmpname;
123 FILE *p_mbtree_stat_file_out;
124 char *psz_mbtree_stat_file_tmpname;
125 char *psz_mbtree_stat_file_name;
126 FILE *p_mbtree_stat_file_in;
128 int num_entries; /* number of ratecontrol_entry_ts */
129 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
130 ratecontrol_entry_t **entry_out;
132 double last_qscale_for[3]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
133 int last_non_b_pict_type;
134 double accum_p_qp; /* for determining I-frame quant */
136 double last_accum_p_norm;
137 double lmin[3]; /* min qscale by frame type */
139 double lstep; /* max change (multiply) in qscale per frame */
142 uint16_t *qp_buffer[2]; /* Global buffers for converting MB-tree quantizer data. */
143 int qpbuf_pos; /* In order to handle pyramid reordering, QP buffer acts as a stack.
144 * This value is the current position (0 or 1). */
149 float *scale_buffer[2]; /* Intermediate buffers */
150 int filtersize[2]; /* filter size (H/V) */
153 int srcdim[2]; /* Source dimensions (W/H) */
157 float frame_size_estimated; /* Access to this variable must be atomic: double is
158 * not atomic on all arches we care about */
159 double frame_size_maximum; /* Maximum frame size due to MinCR */
160 double frame_size_planned;
161 double slice_size_planned;
162 predictor_t *row_pred;
163 predictor_t row_preds[3][2];
164 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
165 int bframes; /* # consecutive B-frames before this P-frame */
166 int bframe_bits; /* total cost of those frames */
170 x264_zone_t *prev_zone;
173 int initial_cpb_removal_delay;
174 int initial_cpb_removal_delay_offset;
175 double nrt_first_access_unit; /* nominal removal time */
176 double previous_cpb_final_arrival_time;
177 uint64_t hrd_multiply_denom;
181 static int parse_zones( x264_t *h );
182 static int init_pass2(x264_t *);
183 static float rate_estimate_qscale( x264_t *h );
184 static int update_vbv( x264_t *h, int bits );
185 static void update_vbv_plan( x264_t *h, int overhead );
186 static float predict_size( predictor_t *p, float q, float var );
187 static void update_predictor( predictor_t *p, float q, float var, float bits );
189 #define CMP_OPT_FIRST_PASS( opt, param_val )\
191 if( ( p = strstr( opts, opt "=" ) ) && sscanf( p, opt "=%d" , &i ) && param_val != i )\
193 x264_log( h, X264_LOG_ERROR, "different " opt " setting than first pass (%d vs %d)\n", param_val, i );\
199 * qp = h.264's quantizer
200 * qscale = linearized quantizer = Lagrange multiplier
202 static inline float qp2qscale( float qp )
204 return 0.85f * powf( 2.0f, ( qp - (12.0f + QP_BD_OFFSET) ) / 6.0f );
206 static inline float qscale2qp( float qscale )
208 return (12.0f + QP_BD_OFFSET) + 6.0f * log2f( qscale/0.85f );
211 /* Texture bitrate is not quite inversely proportional to qscale,
212 * probably due the the changing number of SKIP blocks.
213 * MV bits level off at about qp<=12, because the lambda used
214 * for motion estimation is constant there. */
215 static inline double qscale2bits( ratecontrol_entry_t *rce, double qscale )
219 return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
220 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
224 static ALWAYS_INLINE uint32_t ac_energy_var( uint64_t sum_ssd, int shift, x264_frame_t *frame, int i, int b_store )
226 uint32_t sum = sum_ssd;
227 uint32_t ssd = sum_ssd >> 32;
230 frame->i_pixel_sum[i] += sum;
231 frame->i_pixel_ssd[i] += ssd;
233 return ssd - ((uint64_t)sum * sum >> shift);
236 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 )
238 int height = b_chroma ? 16>>CHROMA_V_SHIFT : 16;
239 int stride = frame->i_stride[i];
241 ? 16 * mb_x + height * (mb_y&~1) * stride + (mb_y&1) * stride
242 : 16 * mb_x + height * mb_y * stride;
246 ALIGNED_ARRAY_16( pixel, pix,[FENC_STRIDE*16] );
247 int chromapix = h->luma2chroma_pixel[PIXEL_16x16];
248 int shift = 7 - CHROMA_V_SHIFT;
250 h->mc.load_deinterleave_chroma_fenc( pix, frame->plane[1] + offset, stride, height );
251 return ac_energy_var( h->pixf.var[chromapix]( pix, FENC_STRIDE ), shift, frame, 1, b_store )
252 + ac_energy_var( h->pixf.var[chromapix]( pix+FENC_STRIDE/2, FENC_STRIDE ), shift, frame, 2, b_store );
255 return ac_energy_var( h->pixf.var[PIXEL_16x16]( frame->plane[i] + offset, stride ), 8, frame, i, b_store );
258 // Find the total AC energy of the block in all planes.
259 static NOINLINE uint32_t x264_ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
261 /* This function contains annoying hacks because GCC has a habit of reordering emms
262 * and putting it after floating point ops. As a result, we put the emms at the end of the
263 * function and make sure that its always called before the float math. Noinline makes
264 * sure no reordering goes on. */
266 x264_prefetch_fenc( h, frame, mb_x, mb_y );
267 if( h->mb.b_adaptive_mbaff )
269 /* We don't know the super-MB mode we're going to pick yet, so
270 * simply try both and pick the lower of the two. */
271 uint32_t var_interlaced, var_progressive;
272 var_interlaced = ac_energy_plane( h, mb_x, mb_y, frame, 0, 0, 1, 1 );
273 var_progressive = ac_energy_plane( h, mb_x, mb_y, frame, 0, 0, 0, 0 );
276 var_interlaced += ac_energy_plane( h, mb_x, mb_y, frame, 1, 0, 1, 1 );
277 var_progressive += ac_energy_plane( h, mb_x, mb_y, frame, 1, 0, 0, 0 );
278 var_interlaced += ac_energy_plane( h, mb_x, mb_y, frame, 2, 0, 1, 1 );
279 var_progressive += ac_energy_plane( h, mb_x, mb_y, frame, 2, 0, 0, 0 );
283 var_interlaced += ac_energy_plane( h, mb_x, mb_y, frame, 1, 1, 1, 1 );
284 var_progressive += ac_energy_plane( h, mb_x, mb_y, frame, 1, 1, 0, 0 );
286 var = X264_MIN( var_interlaced, var_progressive );
290 var = ac_energy_plane( h, mb_x, mb_y, frame, 0, 0, PARAM_INTERLACED, 1 );
293 var += ac_energy_plane( h, mb_x, mb_y, frame, 1, 0, PARAM_INTERLACED, 1 );
294 var += ac_energy_plane( h, mb_x, mb_y, frame, 2, 0, PARAM_INTERLACED, 1 );
297 var += ac_energy_plane( h, mb_x, mb_y, frame, 1, 1, PARAM_INTERLACED, 1 );
303 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame, float *quant_offsets )
305 /* Initialize frame stats */
306 for( int i = 0; i < 3; i++ )
308 frame->i_pixel_sum[i] = 0;
309 frame->i_pixel_ssd[i] = 0;
312 /* Degenerate cases */
313 if( h->param.rc.i_aq_mode == X264_AQ_NONE || h->param.rc.f_aq_strength == 0 )
315 /* Need to init it anyways for MB tree */
316 if( h->param.rc.i_aq_mode && h->param.rc.f_aq_strength == 0 )
320 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
321 frame->f_qp_offset[mb_xy] = frame->f_qp_offset_aq[mb_xy] = quant_offsets[mb_xy];
322 if( h->frames.b_have_lowres )
323 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
324 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8( frame->f_qp_offset[mb_xy] );
328 memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
329 memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
330 if( h->frames.b_have_lowres )
331 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
332 frame->i_inv_qscale_factor[mb_xy] = 256;
335 /* Need variance data for weighted prediction */
336 if( h->param.analyse.i_weighted_pred )
338 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
339 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
340 x264_ac_energy_mb( h, mb_x, mb_y, frame );
345 /* Actual adaptive quantization */
348 /* constants chosen to result in approximately the same overall bitrate as without AQ.
349 * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
352 float bias_strength = 0.f;
354 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE || h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE_BIASED )
356 float bit_depth_correction = 1.f / (1 << (2*(BIT_DEPTH-8)));
357 float avg_adj_pow2 = 0.f;
358 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
359 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
361 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
362 float qp_adj = powf( energy * bit_depth_correction + 1, 0.125f );
363 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
365 avg_adj_pow2 += qp_adj * qp_adj;
367 avg_adj /= h->mb.i_mb_count;
368 avg_adj_pow2 /= h->mb.i_mb_count;
369 strength = h->param.rc.f_aq_strength * avg_adj;
370 avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - 14.f) / avg_adj;
371 bias_strength = h->param.rc.f_aq_strength;
374 strength = h->param.rc.f_aq_strength * 1.0397f;
376 for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
377 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
380 int mb_xy = mb_x + mb_y*h->mb.i_mb_stride;
381 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE_BIASED )
383 qp_adj = frame->f_qp_offset[mb_xy];
384 qp_adj = strength * (qp_adj - avg_adj) + bias_strength * (1.f - 14.f / (qp_adj * qp_adj));
386 else if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
388 qp_adj = frame->f_qp_offset[mb_xy];
389 qp_adj = strength * (qp_adj - avg_adj);
393 uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
394 qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - (14.427f + 2*(BIT_DEPTH-8)));
397 qp_adj += quant_offsets[mb_xy];
398 frame->f_qp_offset[mb_xy] =
399 frame->f_qp_offset_aq[mb_xy] = qp_adj;
400 if( h->frames.b_have_lowres )
401 frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8(qp_adj);
405 /* Remove mean from SSD calculation */
406 for( int i = 0; i < 3; i++ )
408 uint64_t ssd = frame->i_pixel_ssd[i];
409 uint64_t sum = frame->i_pixel_sum[i];
410 int width = 16*h->mb.i_mb_width >> (i && CHROMA_H_SHIFT);
411 int height = 16*h->mb.i_mb_height >> (i && CHROMA_V_SHIFT);
412 frame->i_pixel_ssd[i] = ssd - (sum * sum + width * height / 2) / (width * height);
416 static int x264_macroblock_tree_rescale_init( x264_t *h, x264_ratecontrol_t *rc )
418 /* Use fractional QP array dimensions to compensate for edge padding */
419 float srcdim[2] = {rc->mbtree.srcdim[0] / 16.f, rc->mbtree.srcdim[1] / 16.f};
420 float dstdim[2] = { h->param.i_width / 16.f, h->param.i_height / 16.f};
421 int srcdimi[2] = {ceil(srcdim[0]), ceil(srcdim[1])};
422 int dstdimi[2] = {ceil(dstdim[0]), ceil(dstdim[1])};
423 if( PARAM_INTERLACED )
425 srcdimi[1] = (srcdimi[1]+1)&~1;
426 dstdimi[1] = (dstdimi[1]+1)&~1;
429 rc->mbtree.src_mb_count = srcdimi[0] * srcdimi[1];
431 CHECKED_MALLOC( rc->mbtree.qp_buffer[0], rc->mbtree.src_mb_count * sizeof(uint16_t) );
432 if( h->param.i_bframe_pyramid && h->param.rc.b_stat_read )
433 CHECKED_MALLOC( rc->mbtree.qp_buffer[1], rc->mbtree.src_mb_count * sizeof(uint16_t) );
434 rc->mbtree.qpbuf_pos = -1;
436 /* No rescaling to do */
437 if( srcdimi[0] == dstdimi[0] && srcdimi[1] == dstdimi[1] )
440 rc->mbtree.rescale_enabled = 1;
442 /* Allocate intermediate scaling buffers */
443 CHECKED_MALLOC( rc->mbtree.scale_buffer[0], srcdimi[0] * srcdimi[1] * sizeof(float) );
444 CHECKED_MALLOC( rc->mbtree.scale_buffer[1], dstdimi[0] * srcdimi[1] * sizeof(float) );
446 /* Allocate and calculate resize filter parameters and coefficients */
447 for( int i = 0; i < 2; i++ )
449 if( srcdim[i] > dstdim[i] ) // downscale
450 rc->mbtree.filtersize[i] = 1 + (2 * srcdimi[i] + dstdimi[i] - 1) / dstdimi[i];
452 rc->mbtree.filtersize[i] = 3;
454 CHECKED_MALLOC( rc->mbtree.coeffs[i], rc->mbtree.filtersize[i] * dstdimi[i] * sizeof(float) );
455 CHECKED_MALLOC( rc->mbtree.pos[i], dstdimi[i] * sizeof(int) );
457 /* Initialize filter coefficients */
458 float inc = srcdim[i] / dstdim[i];
459 float dmul = inc > 1.f ? dstdim[i] / srcdim[i] : 1.f;
460 float dstinsrc = 0.5f * inc - 0.5f;
461 int filtersize = rc->mbtree.filtersize[i];
462 for( int j = 0; j < dstdimi[i]; j++ )
464 int pos = dstinsrc - (filtersize - 2.f) * 0.5f;
466 rc->mbtree.pos[i][j] = pos;
467 for( int k = 0; k < filtersize; k++ )
469 float d = fabs( pos + k - dstinsrc ) * dmul;
470 float coeff = X264_MAX( 1.f - d, 0 );
471 rc->mbtree.coeffs[i][j * filtersize + k] = coeff;
475 for( int k = 0; k < filtersize; k++ )
476 rc->mbtree.coeffs[i][j * filtersize + k] *= sum;
481 /* Write back actual qp array dimensions */
482 rc->mbtree.srcdim[0] = srcdimi[0];
483 rc->mbtree.srcdim[1] = srcdimi[1];
489 static void x264_macroblock_tree_rescale_destroy( x264_ratecontrol_t *rc )
491 for( int i = 0; i < 2; i++ )
493 x264_free( rc->mbtree.qp_buffer[i] );
494 x264_free( rc->mbtree.scale_buffer[i] );
495 x264_free( rc->mbtree.coeffs[i] );
496 x264_free( rc->mbtree.pos[i] );
500 static ALWAYS_INLINE float tapfilter( float *src, int pos, int max, int stride, float *coeff, int filtersize )
503 for( int i = 0; i < filtersize; i++, pos++ )
504 sum += src[x264_clip3( pos, 0, max-1 )*stride] * coeff[i];
508 static void x264_macroblock_tree_rescale( x264_t *h, x264_ratecontrol_t *rc, float *dst )
510 float *input, *output;
511 int filtersize, stride, height;
514 input = rc->mbtree.scale_buffer[0];
515 output = rc->mbtree.scale_buffer[1];
516 filtersize = rc->mbtree.filtersize[0];
517 stride = rc->mbtree.srcdim[0];
518 height = rc->mbtree.srcdim[1];
519 for( int y = 0; y < height; y++, input += stride, output += h->mb.i_mb_width )
521 float *coeff = rc->mbtree.coeffs[0];
522 for( int x = 0; x < h->mb.i_mb_width; x++, coeff+=filtersize )
523 output[x] = tapfilter( input, rc->mbtree.pos[0][x], stride, 1, coeff, filtersize );
527 input = rc->mbtree.scale_buffer[1];
529 filtersize = rc->mbtree.filtersize[1];
530 stride = h->mb.i_mb_width;
531 height = rc->mbtree.srcdim[1];
532 for( int x = 0; x < h->mb.i_mb_width; x++, input++, output++ )
534 float *coeff = rc->mbtree.coeffs[1];
535 for( int y = 0; y < h->mb.i_mb_height; y++, coeff+=filtersize )
536 output[y*stride] = tapfilter( input, rc->mbtree.pos[1][y], height, stride, coeff, filtersize );
540 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame, float *quant_offsets )
542 x264_ratecontrol_t *rc = h->rc;
543 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
545 if( rc->entry[frame->i_frame].kept_as_ref )
548 if( rc->mbtree.qpbuf_pos < 0 )
552 rc->mbtree.qpbuf_pos++;
554 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
556 if( fread( rc->mbtree.qp_buffer[rc->mbtree.qpbuf_pos], sizeof(uint16_t), rc->mbtree.src_mb_count, rc->p_mbtree_stat_file_in ) != rc->mbtree.src_mb_count )
559 if( i_type != i_type_actual && rc->mbtree.qpbuf_pos == 1 )
561 x264_log( h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type, i_type_actual );
564 } while( i_type != i_type_actual );
567 float *dst = rc->mbtree.rescale_enabled ? rc->mbtree.scale_buffer[0] : frame->f_qp_offset;
568 h->mc.mbtree_fix8_unpack( dst, rc->mbtree.qp_buffer[rc->mbtree.qpbuf_pos], rc->mbtree.src_mb_count );
569 if( rc->mbtree.rescale_enabled )
570 x264_macroblock_tree_rescale( h, rc, frame->f_qp_offset );
571 if( h->frames.b_have_lowres )
572 for( int i = 0; i < h->mb.i_mb_count; i++ )
573 frame->i_inv_qscale_factor[i] = x264_exp2fix8( frame->f_qp_offset[i] );
574 rc->mbtree.qpbuf_pos--;
577 x264_stack_align( x264_adaptive_quant_frame, h, frame, quant_offsets );
580 x264_log( h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n" );
584 int x264_reference_build_list_optimal( x264_t *h )
586 ratecontrol_entry_t *rce = h->rc->rce;
587 x264_frame_t *frames[16];
588 x264_weight_t weights[16][3];
591 if( rce->refs != h->i_ref[0] )
594 memcpy( frames, h->fref[0], sizeof(frames) );
595 memcpy( refcount, rce->refcount, sizeof(refcount) );
596 memcpy( weights, h->fenc->weight, sizeof(weights) );
597 memset( &h->fenc->weight[1][0], 0, sizeof(x264_weight_t[15][3]) );
599 /* For now don't reorder ref 0; it seems to lower quality
600 in most cases due to skips. */
601 for( int ref = 1; ref < h->i_ref[0]; ref++ )
606 for( int i = 1; i < h->i_ref[0]; i++ )
607 /* Favor lower POC as a tiebreaker. */
608 COPY2_IF_GT( max, refcount[i], bestref, i );
610 /* FIXME: If there are duplicates from frames other than ref0 then it is possible
611 * that the optimal ordering doesnt place every duplicate. */
613 refcount[bestref] = -1;
614 h->fref[0][ref] = frames[bestref];
615 memcpy( h->fenc->weight[ref], weights[bestref], sizeof(weights[bestref]) );
621 static char *x264_strcat_filename( char *input, char *suffix )
623 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
626 strcpy( output, input );
627 strcat( output, suffix );
631 void x264_ratecontrol_init_reconfigurable( x264_t *h, int b_init )
633 x264_ratecontrol_t *rc = h->rc;
634 if( !b_init && rc->b_2pass )
637 if( h->param.rc.i_rc_method == X264_RC_CRF )
639 /* Arbitrary rescaling to make CRF somewhat similar to QP.
640 * Try to compensate for MB-tree's effects as well. */
641 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
642 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
643 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
644 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset + QP_BD_OFFSET );
647 if( h->param.rc.i_vbv_max_bitrate > 0 && h->param.rc.i_vbv_buffer_size > 0 )
649 /* We don't support changing the ABR bitrate right now,
650 so if the stream starts as CBR, keep it CBR. */
651 if( rc->b_vbv_min_rate )
652 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
654 if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
656 h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
657 x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
658 h->param.rc.i_vbv_buffer_size );
661 int kilobit_size = h->param.i_avcintra_class ? 1024 : 1000;
662 int vbv_buffer_size = h->param.rc.i_vbv_buffer_size * kilobit_size;
663 int vbv_max_bitrate = h->param.rc.i_vbv_max_bitrate * kilobit_size;
666 if( h->param.i_nal_hrd && b_init )
668 h->sps->vui.hrd.i_cpb_cnt = 1;
669 h->sps->vui.hrd.b_cbr_hrd = h->param.i_nal_hrd == X264_NAL_HRD_CBR;
670 h->sps->vui.hrd.i_time_offset_length = 0;
675 // normalize HRD size and rate to the value / scale notation
676 h->sps->vui.hrd.i_bit_rate_scale = x264_clip3( x264_ctz( vbv_max_bitrate ) - BR_SHIFT, 0, 15 );
677 h->sps->vui.hrd.i_bit_rate_value = vbv_max_bitrate >> ( h->sps->vui.hrd.i_bit_rate_scale + BR_SHIFT );
678 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 );
679 h->sps->vui.hrd.i_cpb_size_scale = x264_clip3( x264_ctz( vbv_buffer_size ) - CPB_SHIFT, 0, 15 );
680 h->sps->vui.hrd.i_cpb_size_value = vbv_buffer_size >> ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
681 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 );
687 #define MAX_DURATION 0.5
689 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 );
690 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;
691 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);
693 h->sps->vui.hrd.i_initial_cpb_removal_delay_length = 2 + x264_clip3( 32 - x264_clz( max_delay ), 4, 22 );
694 h->sps->vui.hrd.i_cpb_removal_delay_length = x264_clip3( 32 - x264_clz( max_cpb_output_delay ), 4, 31 );
695 h->sps->vui.hrd.i_dpb_output_delay_length = x264_clip3( 32 - x264_clz( max_dpb_output_delay ), 4, 31 );
699 vbv_buffer_size = h->sps->vui.hrd.i_cpb_size_unscaled;
700 vbv_max_bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
702 else if( h->param.i_nal_hrd && !b_init )
704 x264_log( h, X264_LOG_WARNING, "VBV parameters cannot be changed when NAL HRD is in use\n" );
707 h->sps->vui.hrd.i_bit_rate_unscaled = vbv_max_bitrate;
708 h->sps->vui.hrd.i_cpb_size_unscaled = vbv_buffer_size;
710 if( rc->b_vbv_min_rate )
711 rc->bitrate = (double)h->param.rc.i_bitrate * kilobit_size;
712 rc->buffer_rate = vbv_max_bitrate / rc->fps;
713 rc->vbv_max_rate = vbv_max_bitrate;
714 rc->buffer_size = vbv_buffer_size;
715 rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
716 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
717 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
718 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.f_rf_constant_max )
720 rc->rate_factor_max_increment = h->param.rc.f_rf_constant_max - h->param.rc.f_rf_constant;
721 if( rc->rate_factor_max_increment <= 0 )
723 x264_log( h, X264_LOG_WARNING, "CRF max must be greater than CRF\n" );
724 rc->rate_factor_max_increment = 0;
729 if( h->param.rc.f_vbv_buffer_init > 1. )
730 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 );
731 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);
732 rc->buffer_fill_final =
733 rc->buffer_fill_final_min = rc->buffer_size * h->param.rc.f_vbv_buffer_init * h->sps->vui.i_time_scale;
735 rc->b_vbv_min_rate = !rc->b_2pass
736 && h->param.rc.i_rc_method == X264_RC_ABR
737 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
742 int x264_ratecontrol_new( x264_t *h )
744 x264_ratecontrol_t *rc;
748 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
751 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
752 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
754 /* FIXME: use integers */
755 if( h->param.i_fps_num > 0 && h->param.i_fps_den > 0 )
756 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
760 if( h->param.rc.b_mb_tree )
762 h->param.rc.f_pb_factor = 1;
766 rc->qcompress = h->param.rc.f_qcompress;
768 rc->bitrate = h->param.rc.i_bitrate * (h->param.i_avcintra_class ? 1024. : 1000.);
769 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
770 rc->nmb = h->mb.i_mb_count;
771 rc->last_non_b_pict_type = -1;
774 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
776 x264_log( h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n" );
780 x264_ratecontrol_init_reconfigurable( h, 1 );
782 if( h->param.i_nal_hrd )
784 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale;
785 uint64_t num = 90000;
786 x264_reduce_fraction64( &num, &denom );
787 rc->hrd_multiply_denom = 90000 / num;
789 double bits_required = log2( 90000 / rc->hrd_multiply_denom )
790 + log2( h->sps->vui.i_time_scale )
791 + log2( h->sps->vui.hrd.i_cpb_size_unscaled );
792 if( bits_required >= 63 )
794 x264_log( h, X264_LOG_ERROR, "HRD with very large timescale and bufsize not supported\n" );
799 if( rc->rate_tolerance < 0.01 )
801 x264_log( h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n" );
802 rc->rate_tolerance = 0.01;
805 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
809 /* FIXME ABR_INIT_QP is actually used only in CRF */
810 #define ABR_INIT_QP (( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 ) + QP_BD_OFFSET)
811 rc->accum_p_norm = .01;
812 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
813 /* estimated ratio that produces a reasonable QP for the first I-frame */
814 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
815 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
816 rc->last_non_b_pict_type = SLICE_TYPE_I;
819 rc->ip_offset = 6.0 * log2f( h->param.rc.f_ip_factor );
820 rc->pb_offset = 6.0 * log2f( h->param.rc.f_pb_factor );
821 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
822 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, QP_MAX );
823 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, QP_MAX );
824 h->mb.ip_offset = rc->ip_offset + 0.5;
826 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
827 rc->last_qscale = qp2qscale( 26 + QP_BD_OFFSET );
828 int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
829 CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
830 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
831 static const float pred_coeff_table[3] = { 1.0, 1.0, 1.5 };
832 for( int i = 0; i < 3; i++ )
834 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
835 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
836 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
837 for( int j = 0; j < num_preds; j++ )
839 rc->pred[i+j*5].coeff_min = pred_coeff_table[i] / 2;
840 rc->pred[i+j*5].coeff = pred_coeff_table[i];
841 rc->pred[i+j*5].count = 1.0;
842 rc->pred[i+j*5].decay = 0.5;
843 rc->pred[i+j*5].offset = 0.0;
845 for( int j = 0; j < 2; j++ )
847 rc->row_preds[i][j].coeff_min = .25 / 4;
848 rc->row_preds[i][j].coeff = .25;
849 rc->row_preds[i][j].count = 1.0;
850 rc->row_preds[i][j].decay = 0.5;
851 rc->row_preds[i][j].offset = 0.0;
854 rc->pred_b_from_p->coeff_min = 0.5 / 2;
855 rc->pred_b_from_p->coeff = 0.5;
856 rc->pred_b_from_p->count = 1.0;
857 rc->pred_b_from_p->decay = 0.5;
858 rc->pred_b_from_p->offset = 0.0;
860 if( parse_zones( h ) < 0 )
862 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
866 /* Load stat file and init 2pass algo */
867 if( h->param.rc.b_stat_read )
869 char *p, *stats_in, *stats_buf;
871 /* read 1st pass stats */
872 assert( h->param.rc.psz_stat_in );
873 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
876 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n" );
879 if( h->param.rc.b_mb_tree )
881 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
882 if( !mbtree_stats_in )
884 rc->p_mbtree_stat_file_in = x264_fopen( mbtree_stats_in, "rb" );
885 x264_free( mbtree_stats_in );
886 if( !rc->p_mbtree_stat_file_in )
888 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n" );
893 /* check whether 1st pass options were compatible with current options */
894 if( strncmp( stats_buf, "#options:", 9 ) )
896 x264_log( h, X264_LOG_ERROR, "options list in stats file not valid\n" );
900 float res_factor, res_factor_bits;
904 char *opts = stats_buf;
905 stats_in = strchr( stats_buf, '\n' );
910 if( sscanf( opts, "#options: %dx%d", &i, &j ) != 2 )
912 x264_log( h, X264_LOG_ERROR, "resolution specified in stats file not valid\n" );
915 else if( h->param.rc.b_mb_tree )
917 rc->mbtree.srcdim[0] = i;
918 rc->mbtree.srcdim[1] = j;
920 res_factor = (float)h->param.i_width * h->param.i_height / (i*j);
921 /* Change in bits relative to resolution isn't quite linear on typical sources,
922 * so we'll at least try to roughly approximate this effect. */
923 res_factor_bits = powf( res_factor, 0.7 );
925 if( !( p = strstr( opts, "timebase=" ) ) || sscanf( p, "timebase=%u/%u", &k, &l ) != 2 )
927 x264_log( h, X264_LOG_ERROR, "timebase specified in stats file not valid\n" );
930 if( k != h->param.i_timebase_num || l != h->param.i_timebase_den )
932 x264_log( h, X264_LOG_ERROR, "timebase mismatch with 1st pass (%u/%u vs %u/%u)\n",
933 h->param.i_timebase_num, h->param.i_timebase_den, k, l );
937 CMP_OPT_FIRST_PASS( "bitdepth", BIT_DEPTH );
938 CMP_OPT_FIRST_PASS( "weightp", X264_MAX( 0, h->param.analyse.i_weighted_pred ) );
939 CMP_OPT_FIRST_PASS( "bframes", h->param.i_bframe );
940 CMP_OPT_FIRST_PASS( "b_pyramid", h->param.i_bframe_pyramid );
941 CMP_OPT_FIRST_PASS( "intra_refresh", h->param.b_intra_refresh );
942 CMP_OPT_FIRST_PASS( "open_gop", h->param.b_open_gop );
943 CMP_OPT_FIRST_PASS( "bluray_compat", h->param.b_bluray_compat );
945 if( (p = strstr( opts, "interlaced=" )) )
947 char *current = h->param.b_interlaced ? h->param.b_tff ? "tff" : "bff" : h->param.b_fake_interlaced ? "fake" : "0";
949 sscanf( p, "interlaced=%4s", buf );
950 if( strcmp( current, buf ) )
952 x264_log( h, X264_LOG_ERROR, "different interlaced setting than first pass (%s vs %s)\n", current, buf );
957 if( (p = strstr( opts, "keyint=" )) )
960 char buf[13] = "infinite ";
961 if( h->param.i_keyint_max != X264_KEYINT_MAX_INFINITE )
962 sprintf( buf, "%d ", h->param.i_keyint_max );
963 if( strncmp( p, buf, strlen(buf) ) )
965 x264_log( h, X264_LOG_ERROR, "different keyint setting than first pass (%.*s vs %.*s)\n",
966 strlen(buf)-1, buf, strcspn(p, " "), p );
971 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
972 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
974 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
976 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
977 h->mb.b_direct_auto_write = 1;
980 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
981 h->param.i_bframe_adaptive = i;
982 else if( h->param.i_bframe )
984 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
988 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 ) )
989 h->param.rc.i_lookahead = i;
992 /* find number of pics */
995 for( num_entries = -1; p; num_entries++ )
996 p = strchr( p + 1, ';' );
999 x264_log( h, X264_LOG_ERROR, "empty stats file\n" );
1002 rc->num_entries = num_entries;
1004 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
1006 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
1007 h->param.i_frame_total, rc->num_entries );
1009 if( h->param.i_frame_total > rc->num_entries )
1011 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
1012 h->param.i_frame_total, rc->num_entries );
1016 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
1017 CHECKED_MALLOC( rc->entry_out, rc->num_entries * sizeof(ratecontrol_entry_t*) );
1019 /* init all to skipped p frames */
1020 for( int i = 0; i < rc->num_entries; i++ )
1022 ratecontrol_entry_t *rce = &rc->entry[i];
1023 rce->pict_type = SLICE_TYPE_P;
1024 rce->qscale = rce->new_qscale = qp2qscale( 20 + QP_BD_OFFSET );
1025 rce->misc_bits = rc->nmb + 10;
1027 rc->entry_out[i] = rce;
1032 double total_qp_aq = 0;
1033 for( int i = 0; i < rc->num_entries; i++ )
1035 ratecontrol_entry_t *rce;
1036 int frame_number = 0;
1037 int frame_out_number = 0;
1044 next= strchr(p, ';');
1046 *next++ = 0; //sscanf is unbelievably slow on long strings
1047 e = sscanf( p, " in:%d out:%d ", &frame_number, &frame_out_number );
1049 if( frame_number < 0 || frame_number >= rc->num_entries )
1051 x264_log( h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i );
1054 if( frame_out_number < 0 || frame_out_number >= rc->num_entries )
1056 x264_log( h, X264_LOG_ERROR, "bad frame output number (%d) at stats line %d\n", frame_out_number, i );
1059 rce = &rc->entry[frame_number];
1060 rc->entry_out[frame_out_number] = rce;
1061 rce->direct_mode = 0;
1063 e += sscanf( p, " in:%*d out:%*d type:%c dur:%"SCNd64" cpbdur:%"SCNd64" q:%f aq:%f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c",
1064 &pict_type, &rce->i_duration, &rce->i_cpb_duration, &qp_rc, &qp_aq, &rce->tex_bits,
1065 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
1066 &rce->s_count, &rce->direct_mode );
1067 rce->tex_bits *= res_factor_bits;
1068 rce->mv_bits *= res_factor_bits;
1069 rce->misc_bits *= res_factor_bits;
1070 rce->i_count *= res_factor;
1071 rce->p_count *= res_factor;
1072 rce->s_count *= res_factor;
1074 p = strstr( p, "ref:" );
1078 for( ref = 0; ref < 16; ref++ )
1080 if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
1082 p = strchr( p+1, ' ' );
1089 rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
1090 char *w = strchr( p, 'w' );
1093 int count = sscanf( w, "w:%hd,%hd,%hd,%hd,%hd,%hd,%hd,%hd",
1094 &rce->i_weight_denom[0], &rce->weight[0][0], &rce->weight[0][1],
1095 &rce->i_weight_denom[1], &rce->weight[1][0], &rce->weight[1][1],
1096 &rce->weight[2][0], &rce->weight[2][1] );
1098 rce->i_weight_denom[1] = -1;
1099 else if ( count != 8 )
1100 rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
1103 if( pict_type != 'b' )
1104 rce->kept_as_ref = 1;
1108 rce->frame_type = X264_TYPE_IDR;
1109 rce->pict_type = SLICE_TYPE_I;
1112 rce->frame_type = X264_TYPE_I;
1113 rce->pict_type = SLICE_TYPE_I;
1116 rce->frame_type = X264_TYPE_P;
1117 rce->pict_type = SLICE_TYPE_P;
1120 rce->frame_type = X264_TYPE_BREF;
1121 rce->pict_type = SLICE_TYPE_B;
1124 rce->frame_type = X264_TYPE_B;
1125 rce->pict_type = SLICE_TYPE_B;
1127 default: e = -1; break;
1132 x264_log( h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e );
1135 rce->qscale = qp2qscale( qp_rc );
1136 total_qp_aq += qp_aq;
1139 if( !h->param.b_stitchable )
1140 h->pps->i_pic_init_qp = SPEC_QP( (int)(total_qp_aq / rc->num_entries + 0.5) );
1142 x264_free( stats_buf );
1144 if( h->param.rc.i_rc_method == X264_RC_ABR )
1146 if( init_pass2( h ) < 0 )
1148 } /* else we're using constant quant, so no need to run the bitrate allocation */
1151 /* Open output file */
1152 /* If input and output files are the same, output to a temp file
1153 * and move it to the real name only when it's complete */
1154 if( h->param.rc.b_stat_write )
1157 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
1158 if( !rc->psz_stat_file_tmpname )
1161 rc->p_stat_file_out = x264_fopen( rc->psz_stat_file_tmpname, "wb" );
1162 if( rc->p_stat_file_out == NULL )
1164 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n" );
1168 p = x264_param2string( &h->param, 1 );
1170 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
1172 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
1174 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
1175 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
1176 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
1179 rc->p_mbtree_stat_file_out = x264_fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
1180 if( rc->p_mbtree_stat_file_out == NULL )
1182 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n" );
1188 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
1190 if( !h->param.rc.b_stat_read )
1192 rc->mbtree.srcdim[0] = h->param.i_width;
1193 rc->mbtree.srcdim[1] = h->param.i_height;
1195 if( x264_macroblock_tree_rescale_init( h, rc ) < 0 )
1199 for( int i = 0; i<h->param.i_threads; i++ )
1201 h->thread[i]->rc = rc+i;
1205 h->thread[i]->param = h->param;
1206 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
1207 h->thread[i]->mb.ip_offset = h->mb.ip_offset;
1216 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
1219 char *tok, UNUSED *saveptr=NULL;
1221 z->f_bitrate_factor = 1;
1222 if( 3 <= sscanf(p, "%d,%d,q=%d%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
1224 else if( 3 <= sscanf(p, "%d,%d,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
1226 else if( 2 <= sscanf(p, "%d,%d%n", &z->i_start, &z->i_end, &len) )
1230 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
1236 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
1237 memcpy( z->param, &h->param, sizeof(x264_param_t) );
1238 z->param->param_free = x264_free;
1239 while( (tok = strtok_r( p, ",", &saveptr )) )
1241 char *val = strchr( tok, '=' );
1247 if( x264_param_parse( z->param, tok, val ) )
1249 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
1259 static int parse_zones( x264_t *h )
1261 x264_ratecontrol_t *rc = h->rc;
1262 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
1264 char *psz_zones, *p;
1265 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
1266 strcpy( psz_zones, h->param.rc.psz_zones );
1267 h->param.rc.i_zones = 1;
1268 for( p = psz_zones; *p; p++ )
1269 h->param.rc.i_zones += (*p == '/');
1270 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
1272 for( int i = 0; i < h->param.rc.i_zones; i++ )
1274 int i_tok = strcspn( p, "/" );
1276 if( parse_zone( h, &h->param.rc.zones[i], p ) )
1278 x264_free( psz_zones );
1283 x264_free( psz_zones );
1286 if( h->param.rc.i_zones > 0 )
1288 for( int i = 0; i < h->param.rc.i_zones; i++ )
1290 x264_zone_t z = h->param.rc.zones[i];
1291 if( z.i_start < 0 || z.i_start > z.i_end )
1293 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
1294 z.i_start, z.i_end );
1297 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
1299 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
1300 z.f_bitrate_factor );
1305 rc->i_zones = h->param.rc.i_zones + 1;
1306 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
1307 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
1309 // default zone to fall back to if none of the others match
1310 rc->zones[0].i_start = 0;
1311 rc->zones[0].i_end = INT_MAX;
1312 rc->zones[0].b_force_qp = 0;
1313 rc->zones[0].f_bitrate_factor = 1;
1314 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
1315 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
1316 for( int i = 1; i < rc->i_zones; i++ )
1318 if( !rc->zones[i].param )
1319 rc->zones[i].param = rc->zones[0].param;
1328 static x264_zone_t *get_zone( x264_t *h, int frame_num )
1330 for( int i = h->rc->i_zones - 1; i >= 0; i-- )
1332 x264_zone_t *z = &h->rc->zones[i];
1333 if( frame_num >= z->i_start && frame_num <= z->i_end )
1339 void x264_ratecontrol_summary( x264_t *h )
1341 x264_ratecontrol_t *rc = h->rc;
1342 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
1344 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
1345 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
1346 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
1347 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
1348 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset - QP_BD_OFFSET );
1352 void x264_ratecontrol_delete( x264_t *h )
1354 x264_ratecontrol_t *rc = h->rc;
1357 if( rc->p_stat_file_out )
1359 b_regular_file = x264_is_regular_file( rc->p_stat_file_out );
1360 fclose( rc->p_stat_file_out );
1361 if( h->i_frame >= rc->num_entries && b_regular_file )
1362 if( x264_rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
1364 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1365 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
1367 x264_free( rc->psz_stat_file_tmpname );
1369 if( rc->p_mbtree_stat_file_out )
1371 b_regular_file = x264_is_regular_file( rc->p_mbtree_stat_file_out );
1372 fclose( rc->p_mbtree_stat_file_out );
1373 if( h->i_frame >= rc->num_entries && b_regular_file )
1374 if( x264_rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
1376 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1377 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
1379 x264_free( rc->psz_mbtree_stat_file_tmpname );
1380 x264_free( rc->psz_mbtree_stat_file_name );
1382 if( rc->p_mbtree_stat_file_in )
1383 fclose( rc->p_mbtree_stat_file_in );
1384 x264_free( rc->pred );
1385 x264_free( rc->pred_b_from_p );
1386 x264_free( rc->entry );
1387 x264_free( rc->entry_out );
1388 x264_macroblock_tree_rescale_destroy( rc );
1391 x264_free( rc->zones[0].param );
1392 for( int i = 1; i < rc->i_zones; i++ )
1393 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
1394 rc->zones[i].param->param_free( rc->zones[i].param );
1395 x264_free( rc->zones );
1400 static void accum_p_qp_update( x264_t *h, float qp )
1402 x264_ratecontrol_t *rc = h->rc;
1403 rc->accum_p_qp *= .95;
1404 rc->accum_p_norm *= .95;
1405 rc->accum_p_norm += 1;
1406 if( h->sh.i_type == SLICE_TYPE_I )
1407 rc->accum_p_qp += qp + rc->ip_offset;
1409 rc->accum_p_qp += qp;
1412 /* Before encoding a frame, choose a QP for it */
1413 void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
1415 x264_ratecontrol_t *rc = h->rc;
1416 ratecontrol_entry_t *rce = NULL;
1417 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
1422 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
1423 x264_encoder_reconfig_apply( h, zone->param );
1424 rc->prev_zone = zone;
1426 if( h->param.rc.b_stat_read )
1428 int frame = h->fenc->i_frame;
1429 assert( frame >= 0 && frame < rc->num_entries );
1430 rce = h->rc->rce = &h->rc->entry[frame];
1432 if( h->sh.i_type == SLICE_TYPE_B
1433 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
1435 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
1436 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
1442 memset( h->fdec->i_row_bits, 0, h->mb.i_mb_height * sizeof(int) );
1443 memset( h->fdec->f_row_qp, 0, h->mb.i_mb_height * sizeof(float) );
1444 memset( h->fdec->f_row_qscale, 0, h->mb.i_mb_height * sizeof(float) );
1445 rc->row_pred = rc->row_preds[h->sh.i_type];
1446 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;
1447 update_vbv_plan( h, overhead );
1449 const x264_level_t *l = x264_levels;
1450 while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
1453 int mincr = l->mincr;
1455 if( h->param.b_bluray_compat )
1458 /* Profiles above High don't require minCR, so just set the maximum to a large value. */
1459 if( h->sps->i_profile_idc > PROFILE_HIGH )
1460 rc->frame_size_maximum = 1e9;
1463 /* The spec has a bizarre special case for the first frame. */
1464 if( h->i_frame == 0 )
1466 //384 * ( Max( PicSizeInMbs, fR * MaxMBPS ) + MaxMBPS * ( tr( 0 ) - tr,n( 0 ) ) ) / MinCR
1467 double fr = 1. / 172;
1468 int pic_size_in_mbs = h->mb.i_mb_width * h->mb.i_mb_height;
1469 rc->frame_size_maximum = 384 * BIT_DEPTH * X264_MAX( pic_size_in_mbs, fr*l->mbps ) / mincr;
1473 //384 * MaxMBPS * ( tr( n ) - tr( n - 1 ) ) / MinCR
1474 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;
1479 if( h->sh.i_type != SLICE_TYPE_B )
1480 rc->bframes = h->fenc->i_bframes;
1484 q = qscale2qp( rate_estimate_qscale( h ) );
1486 else if( rc->b_2pass )
1488 rce->new_qscale = rate_estimate_qscale( h );
1489 q = qscale2qp( rce->new_qscale );
1493 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
1494 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
1496 q = rc->qp_constant[ h->sh.i_type ];
1500 if( zone->b_force_qp )
1501 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1503 q -= 6*log2f( zone->f_bitrate_factor );
1506 if( i_force_qp != X264_QP_AUTO )
1509 q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1511 rc->qpa_rc = rc->qpa_rc_prev =
1512 rc->qpa_aq = rc->qpa_aq_prev = 0;
1513 h->fdec->f_qp_avg_rc =
1514 h->fdec->f_qp_avg_aq =
1519 accum_p_qp_update( h, rc->qpm );
1521 if( h->sh.i_type != SLICE_TYPE_B )
1522 rc->last_non_b_pict_type = h->sh.i_type;
1525 static float predict_row_size( x264_t *h, int y, float qscale )
1527 /* average between two predictors:
1528 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1529 x264_ratecontrol_t *rc = h->rc;
1530 float pred_s = predict_size( &rc->row_pred[0], qscale, h->fdec->i_row_satd[y] );
1531 if( h->sh.i_type == SLICE_TYPE_I || qscale >= h->fref[0][0]->f_row_qscale[y] )
1533 if( h->sh.i_type == SLICE_TYPE_P
1534 && h->fref[0][0]->i_type == h->fdec->i_type
1535 && h->fref[0][0]->f_row_qscale[y] > 0
1536 && h->fref[0][0]->i_row_satd[y] > 0
1537 && (abs(h->fref[0][0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1539 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]
1540 * h->fref[0][0]->f_row_qscale[y] / qscale;
1541 return (pred_s + pred_t) * 0.5f;
1545 /* Our QP is lower than the reference! */
1548 float pred_intra = predict_size( &rc->row_pred[1], qscale, h->fdec->i_row_satds[0][0][y] );
1549 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1550 return pred_intra + pred_s;
1554 static int row_bits_so_far( x264_t *h, int y )
1557 for( int i = h->i_threadslice_start; i <= y; i++ )
1558 bits += h->fdec->i_row_bits[i];
1562 static float predict_row_size_to_end( x264_t *h, int y, float qp )
1564 float qscale = qp2qscale( qp );
1566 for( int i = y+1; i < h->i_threadslice_end; i++ )
1567 bits += predict_row_size( h, i, qscale );
1572 * eliminate all use of qp in row ratecontrol: make it entirely qscale-based.
1573 * make this function stop being needlessly O(N^2)
1574 * update more often than once per row? */
1575 int x264_ratecontrol_mb( x264_t *h, int bits )
1577 x264_ratecontrol_t *rc = h->rc;
1578 const int y = h->mb.i_mb_y;
1580 h->fdec->i_row_bits[y] += bits;
1581 rc->qpa_aq += h->mb.i_qp;
1583 if( h->mb.i_mb_x != h->mb.i_mb_width - 1 )
1587 rc->qpa_rc += rc->qpm * h->mb.i_mb_width;
1592 float qscale = qp2qscale( rc->qpm );
1593 h->fdec->f_row_qp[y] = rc->qpm;
1594 h->fdec->f_row_qscale[y] = qscale;
1596 update_predictor( &rc->row_pred[0], qscale, h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1597 if( h->sh.i_type != SLICE_TYPE_I && rc->qpm < h->fref[0][0]->f_row_qp[y] )
1598 update_predictor( &rc->row_pred[1], qscale, h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1600 /* update ratecontrol per-mbpair in MBAFF */
1601 if( SLICE_MBAFF && !(y&1) )
1604 /* FIXME: We don't currently support the case where there's a slice
1605 * boundary in between. */
1606 int can_reencode_row = h->sh.i_first_mb <= ((h->mb.i_mb_y - SLICE_MBAFF) * h->mb.i_mb_stride);
1608 /* tweak quality based on difference from predicted size */
1609 float prev_row_qp = h->fdec->f_row_qp[y];
1610 float qp_absolute_max = h->param.rc.i_qp_max;
1611 if( rc->rate_factor_max_increment )
1612 qp_absolute_max = X264_MIN( qp_absolute_max, rc->qp_novbv + rc->rate_factor_max_increment );
1613 float qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, qp_absolute_max );
1614 float qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1615 float step_size = 0.5f;
1616 float slice_size_planned = h->param.b_sliced_threads ? rc->slice_size_planned : rc->frame_size_planned;
1617 float bits_so_far = row_bits_so_far( h, y );
1618 float max_frame_error = x264_clip3f( 1.0 / h->mb.i_mb_height, 0.05, 0.25 );
1619 float max_frame_size = rc->frame_size_maximum - rc->frame_size_maximum * max_frame_error;
1620 max_frame_size = X264_MIN( max_frame_size, rc->buffer_fill - rc->buffer_rate * max_frame_error );
1621 float size_of_other_slices = 0;
1622 if( h->param.b_sliced_threads )
1624 float size_of_other_slices_planned = 0;
1625 for( int i = 0; i < h->param.i_threads; i++ )
1626 if( h != h->thread[i] )
1628 size_of_other_slices += h->thread[i]->rc->frame_size_estimated;
1629 size_of_other_slices_planned += h->thread[i]->rc->slice_size_planned;
1631 float weight = rc->slice_size_planned / rc->frame_size_planned;
1632 size_of_other_slices = (size_of_other_slices - size_of_other_slices_planned) * weight + size_of_other_slices_planned;
1634 if( y < h->i_threadslice_end-1 )
1636 /* B-frames shouldn't use lower QP than their reference frames. */
1637 if( h->sh.i_type == SLICE_TYPE_B )
1639 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] ) );
1640 rc->qpm = X264_MAX( rc->qpm, qp_min );
1643 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1644 buffer_left_planned = X264_MAX( buffer_left_planned, 0.f );
1645 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1646 float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
1647 float b1 = bits_so_far + predict_row_size_to_end( h, y, rc->qpm ) + size_of_other_slices;
1648 float trust_coeff = x264_clip3f( bits_so_far / slice_size_planned, 0.0, 1.0 );
1650 /* Don't increase the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1651 /* area at the top of the frame was measured inaccurately. */
1652 if( trust_coeff < 0.05f )
1653 qp_max = qp_absolute_max = prev_row_qp;
1655 if( h->sh.i_type != SLICE_TYPE_I )
1658 if( !rc->b_vbv_min_rate )
1659 qp_min = X264_MAX( qp_min, rc->qp_novbv );
1661 while( rc->qpm < qp_max
1662 && ((b1 > rc->frame_size_planned + rc_tol) ||
1663 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv) ||
1664 (b1 > rc->buffer_fill - buffer_left_planned * 0.5f)) )
1666 rc->qpm += step_size;
1667 b1 = bits_so_far + predict_row_size_to_end( h, y, rc->qpm ) + size_of_other_slices;
1670 float b_max = b1 + ((rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 0.90f - b1) * trust_coeff;
1671 rc->qpm -= step_size;
1672 float b2 = bits_so_far + predict_row_size_to_end( h, y, rc->qpm ) + size_of_other_slices;
1673 while( rc->qpm > qp_min && rc->qpm < prev_row_qp
1674 && (rc->qpm > h->fdec->f_row_qp[0] || rc->single_frame_vbv)
1675 && (b2 < max_frame_size)
1676 && ((b2 < rc->frame_size_planned * 0.8f) || (b2 < b_max)) )
1679 rc->qpm -= step_size;
1680 b2 = bits_so_far + predict_row_size_to_end( h, y, rc->qpm ) + size_of_other_slices;
1682 rc->qpm += step_size;
1684 /* avoid VBV underflow or MinCR violation */
1685 while( rc->qpm < qp_absolute_max && (b1 > max_frame_size) )
1687 rc->qpm += step_size;
1688 b1 = bits_so_far + predict_row_size_to_end( h, y, rc->qpm ) + size_of_other_slices;
1691 h->rc->frame_size_estimated = b1 - size_of_other_slices;
1693 /* If the current row was large enough to cause a large QP jump, try re-encoding it. */
1694 if( rc->qpm > qp_max && prev_row_qp < qp_max && can_reencode_row )
1696 /* Bump QP to halfway in between... close enough. */
1697 rc->qpm = x264_clip3f( (prev_row_qp + rc->qpm)*0.5f, prev_row_qp + 1.0f, qp_max );
1698 rc->qpa_rc = rc->qpa_rc_prev;
1699 rc->qpa_aq = rc->qpa_aq_prev;
1700 h->fdec->i_row_bits[y] = 0;
1701 h->fdec->i_row_bits[y-SLICE_MBAFF] = 0;
1707 h->rc->frame_size_estimated = bits_so_far;
1709 /* Last-ditch attempt: if the last row of the frame underflowed the VBV,
1711 if( rc->qpm < qp_max && can_reencode_row
1712 && (h->rc->frame_size_estimated + size_of_other_slices > X264_MIN( rc->frame_size_maximum, rc->buffer_fill )) )
1715 rc->qpa_rc = rc->qpa_rc_prev;
1716 rc->qpa_aq = rc->qpa_aq_prev;
1717 h->fdec->i_row_bits[y] = 0;
1718 h->fdec->i_row_bits[y-SLICE_MBAFF] = 0;
1723 rc->qpa_rc_prev = rc->qpa_rc;
1724 rc->qpa_aq_prev = rc->qpa_aq;
1729 int x264_ratecontrol_qp( x264_t *h )
1732 return x264_clip3( h->rc->qpm + 0.5f, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1735 int x264_ratecontrol_mb_qp( x264_t *h )
1738 float qp = h->rc->qpm;
1739 if( h->param.rc.i_aq_mode )
1741 /* MB-tree currently doesn't adjust quantizers in unreferenced frames. */
1742 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];
1743 /* Scale AQ's effect towards zero in emergency mode. */
1744 if( qp > QP_MAX_SPEC )
1745 qp_offset *= (QP_MAX - qp) / (QP_MAX - QP_MAX_SPEC);
1748 return x264_clip3( qp + 0.5f, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1751 /* In 2pass, force the same frame types as in the 1st pass */
1752 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1754 x264_ratecontrol_t *rc = h->rc;
1755 if( h->param.rc.b_stat_read )
1757 if( frame_num >= rc->num_entries )
1759 /* We could try to initialize everything required for ABR and
1760 * adaptive B-frames, but that would be complicated.
1761 * So just calculate the average QP used so far. */
1762 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24 + QP_BD_OFFSET
1763 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1764 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, QP_MAX );
1765 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 );
1766 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 );
1768 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries );
1769 x264_log( h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant );
1770 if( h->param.i_bframe_adaptive )
1771 x264_log( h, X264_LOG_ERROR, "disabling adaptive B-frames\n" );
1773 for( int i = 0; i < h->param.i_threads; i++ )
1775 h->thread[i]->rc->b_abr = 0;
1776 h->thread[i]->rc->b_2pass = 0;
1777 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1778 h->thread[i]->param.rc.b_stat_read = 0;
1779 h->thread[i]->param.i_bframe_adaptive = 0;
1780 h->thread[i]->param.i_scenecut_threshold = 0;
1781 h->thread[i]->param.rc.b_mb_tree = 0;
1782 if( h->thread[i]->param.i_bframe > 1 )
1783 h->thread[i]->param.i_bframe = 1;
1785 return X264_TYPE_AUTO;
1787 return rc->entry[frame_num].frame_type;
1790 return X264_TYPE_AUTO;
1793 void x264_ratecontrol_set_weights( x264_t *h, x264_frame_t *frm )
1795 ratecontrol_entry_t *rce = &h->rc->entry[frm->i_frame];
1796 if( h->param.analyse.i_weighted_pred <= 0 )
1799 if( rce->i_weight_denom[0] >= 0 )
1800 SET_WEIGHT( frm->weight[0][0], 1, rce->weight[0][0], rce->i_weight_denom[0], rce->weight[0][1] );
1802 if( rce->i_weight_denom[1] >= 0 )
1804 SET_WEIGHT( frm->weight[0][1], 1, rce->weight[1][0], rce->i_weight_denom[1], rce->weight[1][1] );
1805 SET_WEIGHT( frm->weight[0][2], 1, rce->weight[2][0], rce->i_weight_denom[1], rce->weight[2][1] );
1809 /* After encoding one frame, save stats and update ratecontrol state */
1810 int x264_ratecontrol_end( x264_t *h, int bits, int *filler )
1812 x264_ratecontrol_t *rc = h->rc;
1813 const int *mbs = h->stat.frame.i_mb_count;
1817 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1818 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1819 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1820 for( int i = B_DIRECT; i < B_8x8; i++ )
1821 h->stat.frame.i_mb_count_p += mbs[i];
1823 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1824 h->fdec->f_qp_avg_aq = (float)rc->qpa_aq / h->mb.i_mb_count;
1825 h->fdec->f_crf_avg = h->param.rc.f_rf_constant + h->fdec->f_qp_avg_rc - rc->qp_novbv;
1827 if( h->param.rc.b_stat_write )
1829 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1830 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1831 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1832 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1833 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1834 char c_direct = h->mb.b_direct_auto_write ?
1835 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1836 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1838 if( fprintf( rc->p_stat_file_out,
1839 "in:%d out:%d type:%c dur:%"PRId64" cpbdur:%"PRId64" q:%.2f aq:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c ref:",
1840 h->fenc->i_frame, h->i_frame,
1841 c_type, h->fenc->i_duration,
1842 h->fenc->i_cpb_duration,
1843 rc->qpa_rc, h->fdec->f_qp_avg_aq,
1844 h->stat.frame.i_tex_bits,
1845 h->stat.frame.i_mv_bits,
1846 h->stat.frame.i_misc_bits,
1847 h->stat.frame.i_mb_count_i,
1848 h->stat.frame.i_mb_count_p,
1849 h->stat.frame.i_mb_count_skip,
1853 /* Only write information for reference reordering once. */
1854 int use_old_stats = h->param.rc.b_stat_read && rc->rce->refs > 1;
1855 for( int i = 0; i < (use_old_stats ? rc->rce->refs : h->i_ref[0]); i++ )
1857 int refcount = use_old_stats ? rc->rce->refcount[i]
1858 : PARAM_INTERLACED ? h->stat.frame.i_mb_count_ref[0][i*2]
1859 + h->stat.frame.i_mb_count_ref[0][i*2+1]
1860 : h->stat.frame.i_mb_count_ref[0][i];
1861 if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1865 if( h->param.analyse.i_weighted_pred >= X264_WEIGHTP_SIMPLE && h->sh.weight[0][0].weightfn )
1867 if( fprintf( rc->p_stat_file_out, "w:%d,%d,%d",
1868 h->sh.weight[0][0].i_denom, h->sh.weight[0][0].i_scale, h->sh.weight[0][0].i_offset ) < 0 )
1870 if( h->sh.weight[0][1].weightfn || h->sh.weight[0][2].weightfn )
1872 if( fprintf( rc->p_stat_file_out, ",%d,%d,%d,%d,%d ",
1873 h->sh.weight[0][1].i_denom, h->sh.weight[0][1].i_scale, h->sh.weight[0][1].i_offset,
1874 h->sh.weight[0][2].i_scale, h->sh.weight[0][2].i_offset ) < 0 )
1877 else if( fprintf( rc->p_stat_file_out, " " ) < 0 )
1881 if( fprintf( rc->p_stat_file_out, ";\n") < 0 )
1884 /* Don't re-write the data in multi-pass mode. */
1885 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1887 uint8_t i_type = h->sh.i_type;
1888 h->mc.mbtree_fix8_pack( rc->mbtree.qp_buffer[0], h->fenc->f_qp_offset, h->mb.i_mb_count );
1889 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1891 if( fwrite( rc->mbtree.qp_buffer[0], sizeof(uint16_t), h->mb.i_mb_count, rc->p_mbtree_stat_file_out ) < h->mb.i_mb_count )
1898 if( h->sh.i_type != SLICE_TYPE_B )
1899 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / rc->last_rceq;
1902 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1903 * Not perfectly accurate with B-refs, but good enough. */
1904 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / (rc->last_rceq * fabs( h->param.rc.f_pb_factor ));
1906 rc->cplxr_sum *= rc->cbr_decay;
1907 rc->wanted_bits_window += h->fenc->f_duration * rc->bitrate;
1908 rc->wanted_bits_window *= rc->cbr_decay;
1912 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale( rc->rce->new_qp ) );
1914 if( h->mb.b_variable_qp )
1916 if( h->sh.i_type == SLICE_TYPE_B )
1918 rc->bframe_bits += bits;
1919 if( h->fenc->b_last_minigop_bframe )
1921 update_predictor( rc->pred_b_from_p, qp2qscale( rc->qpa_rc ),
1922 h->fref[1][h->i_ref[1]-1]->i_satd, rc->bframe_bits / rc->bframes );
1923 rc->bframe_bits = 0;
1928 *filler = update_vbv( h, bits );
1929 rc->filler_bits_sum += *filler * 8;
1931 if( h->sps->vui.b_nal_hrd_parameters_present )
1933 if( h->fenc->i_frame == 0 )
1935 // access unit initialises the HRD
1936 h->fenc->hrd_timing.cpb_initial_arrival_time = 0;
1937 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1938 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1939 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit = (double)rc->initial_cpb_removal_delay / 90000;
1943 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit + (double)(h->fenc->i_cpb_delay - h->i_cpb_delay_pir_offset) *
1944 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1946 if( h->fenc->b_keyframe )
1948 rc->nrt_first_access_unit = h->fenc->hrd_timing.cpb_removal_time;
1949 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1950 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1953 double cpb_earliest_arrival_time = h->fenc->hrd_timing.cpb_removal_time - (double)rc->initial_cpb_removal_delay / 90000;
1954 if( !h->fenc->b_keyframe )
1955 cpb_earliest_arrival_time -= (double)rc->initial_cpb_removal_delay_offset / 90000;
1957 if( h->sps->vui.hrd.b_cbr_hrd )
1958 h->fenc->hrd_timing.cpb_initial_arrival_time = rc->previous_cpb_final_arrival_time;
1960 h->fenc->hrd_timing.cpb_initial_arrival_time = X264_MAX( rc->previous_cpb_final_arrival_time, cpb_earliest_arrival_time );
1962 int filler_bits = *filler ? X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), *filler )*8 : 0;
1964 h->fenc->hrd_timing.cpb_final_arrival_time = rc->previous_cpb_final_arrival_time = h->fenc->hrd_timing.cpb_initial_arrival_time +
1965 (double)(bits + filler_bits) / h->sps->vui.hrd.i_bit_rate_unscaled;
1967 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 +
1968 h->fenc->hrd_timing.cpb_removal_time;
1973 x264_log( h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n" );
1977 /****************************************************************************
1979 ***************************************************************************/
1982 * modify the bitrate curve from pass1 for one frame
1984 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1986 x264_ratecontrol_t *rcc= h->rc;
1987 x264_zone_t *zone = get_zone( h, frame_num );
1989 if( h->param.rc.b_mb_tree )
1991 double timescale = (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1992 q = pow( BASE_FRAME_DURATION / CLIP_DURATION(rce->i_duration * timescale), 1 - h->param.rc.f_qcompress );
1995 q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1997 // avoid NaN's in the rc_eq
1998 if( !isfinite(q) || rce->tex_bits + rce->mv_bits == 0 )
1999 q = rcc->last_qscale_for[rce->pict_type];
2004 rcc->last_qscale = q;
2009 if( zone->b_force_qp )
2010 q = qp2qscale( zone->i_qp );
2012 q /= zone->f_bitrate_factor;
2018 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q, int frame_num)
2020 x264_ratecontrol_t *rcc = h->rc;
2021 const int pict_type = rce->pict_type;
2022 x264_zone_t *zone = get_zone( h, frame_num );
2024 // force I/B quants as a function of P quants
2025 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
2026 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
2027 if( pict_type == SLICE_TYPE_I )
2030 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2031 double ip_factor = fabs( h->param.rc.f_ip_factor );
2032 /* don't apply ip_factor if the following frame is also I */
2033 if( rcc->accum_p_norm <= 0 )
2035 else if( h->param.rc.f_ip_factor < 0 )
2037 else if( rcc->accum_p_norm >= 1 )
2040 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
2042 else if( pict_type == SLICE_TYPE_B )
2044 if( h->param.rc.f_pb_factor > 0 )
2046 if( !rce->kept_as_ref )
2047 q *= fabs( h->param.rc.f_pb_factor );
2049 else if( pict_type == SLICE_TYPE_P
2050 && rcc->last_non_b_pict_type == SLICE_TYPE_P
2051 && rce->tex_bits == 0 )
2056 /* last qscale / qdiff stuff */
2057 if( rcc->last_non_b_pict_type == pict_type &&
2058 (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1) )
2060 double last_q = rcc->last_qscale_for[pict_type];
2061 double max_qscale = last_q * rcc->lstep;
2062 double min_qscale = last_q / rcc->lstep;
2064 if ( q > max_qscale ) q = max_qscale;
2065 else if( q < min_qscale ) q = min_qscale;
2068 rcc->last_qscale_for[pict_type] = q;
2069 if( pict_type != SLICE_TYPE_B )
2070 rcc->last_non_b_pict_type = pict_type;
2071 if( pict_type == SLICE_TYPE_I )
2073 rcc->last_accum_p_norm = rcc->accum_p_norm;
2074 rcc->accum_p_norm = 0;
2075 rcc->accum_p_qp = 0;
2077 if( pict_type == SLICE_TYPE_P )
2079 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
2080 rcc->accum_p_qp = mask * (qscale2qp( q ) + rcc->accum_p_qp);
2081 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
2086 if( zone->b_force_qp )
2087 q = qp2qscale( zone->i_qp );
2089 q /= zone->f_bitrate_factor;
2095 static float predict_size( predictor_t *p, float q, float var )
2097 return (p->coeff*var + p->offset) / (q*p->count);
2100 static void update_predictor( predictor_t *p, float q, float var, float bits )
2105 float old_coeff = p->coeff / p->count;
2106 float old_offset = p->offset / p->count;
2107 float new_coeff = X264_MAX( (bits*q - old_offset) / var, p->coeff_min );
2108 float new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
2109 float new_offset = bits*q - new_coeff_clipped * var;
2110 if( new_offset >= 0 )
2111 new_coeff = new_coeff_clipped;
2114 p->count *= p->decay;
2115 p->coeff *= p->decay;
2116 p->offset *= p->decay;
2118 p->coeff += new_coeff;
2119 p->offset += new_offset;
2122 // update VBV after encoding a frame
2123 static int update_vbv( x264_t *h, int bits )
2126 int bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
2127 x264_ratecontrol_t *rcc = h->rc;
2128 x264_ratecontrol_t *rct = h->thread[0]->rc;
2129 int64_t buffer_size = (int64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
2131 if( rcc->last_satd >= h->mb.i_mb_count )
2132 update_predictor( &rct->pred[h->sh.i_type], qp2qscale( rcc->qpa_rc ), rcc->last_satd, bits );
2137 uint64_t buffer_diff = (uint64_t)bits * h->sps->vui.i_time_scale;
2138 rct->buffer_fill_final -= buffer_diff;
2139 rct->buffer_fill_final_min -= buffer_diff;
2141 if( rct->buffer_fill_final_min < 0 )
2143 double underflow = (double)rct->buffer_fill_final_min / h->sps->vui.i_time_scale;
2144 if( rcc->rate_factor_max_increment && rcc->qpm >= rcc->qp_novbv + rcc->rate_factor_max_increment )
2145 x264_log( h, X264_LOG_DEBUG, "VBV underflow due to CRF-max (frame %d, %.0f bits)\n", h->i_frame, underflow );
2147 x264_log( h, X264_LOG_WARNING, "VBV underflow (frame %d, %.0f bits)\n", h->i_frame, underflow );
2148 rct->buffer_fill_final =
2149 rct->buffer_fill_final_min = 0;
2152 if( h->param.i_avcintra_class )
2153 buffer_diff = buffer_size;
2155 buffer_diff = (uint64_t)bitrate * h->sps->vui.i_num_units_in_tick * h->fenc->i_cpb_duration;
2156 rct->buffer_fill_final += buffer_diff;
2157 rct->buffer_fill_final_min += buffer_diff;
2159 if( rct->buffer_fill_final > buffer_size )
2161 if( h->param.rc.b_filler )
2163 int64_t scale = (int64_t)h->sps->vui.i_time_scale * 8;
2164 filler = (rct->buffer_fill_final - buffer_size + scale - 1) / scale;
2165 bits = h->param.i_avcintra_class ? filler * 8 : X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), filler ) * 8;
2166 buffer_diff = (uint64_t)bits * h->sps->vui.i_time_scale;
2167 rct->buffer_fill_final -= buffer_diff;
2168 rct->buffer_fill_final_min -= buffer_diff;
2172 rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, buffer_size );
2173 rct->buffer_fill_final_min = X264_MIN( rct->buffer_fill_final_min, buffer_size );
2180 void x264_hrd_fullness( x264_t *h )
2182 x264_ratecontrol_t *rct = h->thread[0]->rc;
2183 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale / rct->hrd_multiply_denom;
2184 uint64_t cpb_state = rct->buffer_fill_final;
2185 uint64_t cpb_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
2186 uint64_t multiply_factor = 90000 / rct->hrd_multiply_denom;
2188 if( rct->buffer_fill_final < 0 || rct->buffer_fill_final > (int64_t)cpb_size )
2190 x264_log( h, X264_LOG_WARNING, "CPB %s: %.0f bits in a %.0f-bit buffer\n",
2191 rct->buffer_fill_final < 0 ? "underflow" : "overflow",
2192 (double)rct->buffer_fill_final / h->sps->vui.i_time_scale, (double)cpb_size / h->sps->vui.i_time_scale );
2195 h->initial_cpb_removal_delay = (multiply_factor * cpb_state) / denom;
2196 h->initial_cpb_removal_delay_offset = (multiply_factor * cpb_size) / denom - h->initial_cpb_removal_delay;
2198 int64_t decoder_buffer_fill = h->initial_cpb_removal_delay * denom / multiply_factor;
2199 rct->buffer_fill_final_min = X264_MIN( rct->buffer_fill_final_min, decoder_buffer_fill );
2202 // provisionally update VBV according to the planned size of all frames currently in progress
2203 static void update_vbv_plan( x264_t *h, int overhead )
2205 x264_ratecontrol_t *rcc = h->rc;
2206 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final_min / h->sps->vui.i_time_scale;
2207 if( h->i_thread_frames > 1 )
2209 int j = h->rc - h->thread[0]->rc;
2210 for( int i = 1; i < h->i_thread_frames; i++ )
2212 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
2213 double bits = t->rc->frame_size_planned;
2214 if( !t->b_thread_active )
2216 bits = X264_MAX(bits, t->rc->frame_size_estimated);
2217 rcc->buffer_fill -= bits;
2218 rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
2219 rcc->buffer_fill += t->rc->buffer_rate;
2220 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
2223 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
2224 rcc->buffer_fill -= overhead;
2227 // apply VBV constraints and clip qscale to between lmin and lmax
2228 static double clip_qscale( x264_t *h, int pict_type, double q )
2230 x264_ratecontrol_t *rcc = h->rc;
2231 double lmin = rcc->lmin[pict_type];
2232 double lmax = rcc->lmax[pict_type];
2233 if( rcc->rate_factor_max_increment )
2234 lmax = X264_MIN( lmax, qp2qscale( rcc->qp_novbv + rcc->rate_factor_max_increment ) );
2237 /* B-frames are not directly subject to VBV,
2238 * since they are controlled by the P-frames' QPs. */
2240 if( rcc->b_vbv && rcc->last_satd > 0 )
2242 double fenc_cpb_duration = (double)h->fenc->i_cpb_duration *
2243 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2244 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
2245 * the lookahead overflow and such that the buffer is in a reasonable state
2246 * by the end of the lookahead. */
2247 if( h->param.rc.i_lookahead )
2251 /* Avoid an infinite loop. */
2252 for( int iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
2255 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2256 double buffer_fill_cur = rcc->buffer_fill - cur_bits;
2258 double total_duration = 0;
2259 double last_duration = fenc_cpb_duration;
2260 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
2261 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
2262 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
2264 /* Loop over the planned future frames. */
2265 for( int j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
2267 total_duration += last_duration;
2268 buffer_fill_cur += rcc->vbv_max_rate * last_duration;
2269 int i_type = h->fenc->i_planned_type[j];
2270 int i_satd = h->fenc->i_planned_satd[j];
2271 if( i_type == X264_TYPE_AUTO )
2273 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
2274 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
2275 buffer_fill_cur -= cur_bits;
2276 last_duration = h->fenc->f_planned_cpb_duration[j];
2278 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
2279 target_fill = X264_MIN( rcc->buffer_fill + total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.5 );
2280 if( buffer_fill_cur < target_fill )
2286 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
2287 target_fill = x264_clip3f( rcc->buffer_fill - total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
2288 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
2297 /* Fallback to old purely-reactive algorithm: no lookahead. */
2300 if( ( pict_type == SLICE_TYPE_P ||
2301 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
2302 rcc->buffer_fill/rcc->buffer_size < 0.5 )
2304 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
2307 /* Now a hard threshold to make sure the frame fits in VBV.
2308 * This one is mostly for I-frames. */
2309 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2310 /* For small VBVs, allow the frame to use up the entire VBV. */
2311 double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
2312 /* For single-frame VBVs, request that the frame use up the entire VBV. */
2313 double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
2315 if( bits > rcc->buffer_fill/max_fill_factor )
2317 double qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
2321 if( bits < rcc->buffer_rate/min_fill_factor )
2323 double qf = x264_clip3f( bits*min_fill_factor/rcc->buffer_rate, 0.001, 1.0 );
2326 q = X264_MAX( q0, q );
2329 /* Check B-frame complexity, and use up any bits that would
2330 * overflow before the next P-frame. */
2331 if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
2333 int nb = rcc->bframes;
2334 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2335 double pbbits = bits;
2336 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
2338 double bframe_cpb_duration = 0;
2339 double minigop_cpb_duration;
2340 for( int i = 0; i < nb; i++ )
2341 bframe_cpb_duration += h->fenc->f_planned_cpb_duration[i];
2343 if( bbits * nb > bframe_cpb_duration * rcc->vbv_max_rate )
2345 pbbits += nb * bbits;
2347 minigop_cpb_duration = bframe_cpb_duration + fenc_cpb_duration;
2348 space = rcc->buffer_fill + minigop_cpb_duration*rcc->vbv_max_rate - rcc->buffer_size;
2349 if( pbbits < space )
2351 q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
2353 q = X264_MAX( q0/2, q );
2356 /* Apply MinCR and buffer fill restrictions */
2357 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2358 double frame_size_maximum = X264_MIN( rcc->frame_size_maximum, X264_MAX( rcc->buffer_fill, 0.001 ) );
2359 if( bits > frame_size_maximum )
2360 q *= bits / frame_size_maximum;
2362 if( !rcc->b_vbv_min_rate )
2363 q = X264_MAX( q0, q );
2368 else if( rcc->b_2pass )
2370 double min2 = log( lmin );
2371 double max2 = log( lmax );
2372 q = (log(q) - min2)/(max2-min2) - 0.5;
2373 q = 1.0/(1.0 + exp( -4*q ));
2374 q = q*(max2-min2) + min2;
2378 return x264_clip3f( q, lmin, lmax );
2381 // update qscale for 1 frame based on actual bits used so far
2382 static float rate_estimate_qscale( x264_t *h )
2385 x264_ratecontrol_t *rcc = h->rc;
2386 ratecontrol_entry_t rce = {0};
2387 int pict_type = h->sh.i_type;
2388 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
2389 + h->stat.i_frame_size[SLICE_TYPE_P]
2390 + h->stat.i_frame_size[SLICE_TYPE_B])
2391 - rcc->filler_bits_sum;
2396 if( pict_type != rce.pict_type )
2398 x264_log( h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
2399 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type] );
2403 if( pict_type == SLICE_TYPE_B )
2405 /* B-frames don't have independent ratecontrol, but rather get the
2406 * average QP of the two adjacent P-frames + an offset */
2408 int i0 = IS_X264_TYPE_I(h->fref_nearest[0]->i_type);
2409 int i1 = IS_X264_TYPE_I(h->fref_nearest[1]->i_type);
2410 int dt0 = abs(h->fenc->i_poc - h->fref_nearest[0]->i_poc);
2411 int dt1 = abs(h->fenc->i_poc - h->fref_nearest[1]->i_poc);
2412 float q0 = h->fref_nearest[0]->f_qp_avg_rc;
2413 float q1 = h->fref_nearest[1]->f_qp_avg_rc;
2415 if( h->fref_nearest[0]->i_type == X264_TYPE_BREF )
2416 q0 -= rcc->pb_offset/2;
2417 if( h->fref_nearest[1]->i_type == X264_TYPE_BREF )
2418 q1 -= rcc->pb_offset/2;
2421 q = (q0 + q1) / 2 + rcc->ip_offset;
2427 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
2429 if( h->fenc->b_kept_as_ref )
2430 q += rcc->pb_offset/2;
2432 q += rcc->pb_offset;
2437 rcc->frame_size_planned = qscale2bits( &rce, q );
2439 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref[1][h->i_ref[1]-1]->i_satd );
2440 /* Limit planned size by MinCR */
2442 rcc->frame_size_planned = X264_MIN( rcc->frame_size_planned, rcc->frame_size_maximum );
2443 h->rc->frame_size_estimated = rcc->frame_size_planned;
2447 rcc->last_satd = x264_rc_analyse_slice( h );
2452 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
2453 double predicted_bits = total_bits;
2454 if( h->i_thread_frames > 1 )
2456 int j = h->rc - h->thread[0]->rc;
2457 for( int i = 1; i < h->i_thread_frames; i++ )
2459 x264_t *t = h->thread[(j+i) % h->i_thread_frames];
2460 double bits = t->rc->frame_size_planned;
2461 if( !t->b_thread_active )
2463 bits = X264_MAX(bits, t->rc->frame_size_estimated);
2464 predicted_bits += bits;
2470 double lmin = rcc->lmin[pict_type];
2471 double lmax = rcc->lmax[pict_type];
2474 /* Adjust ABR buffer based on distance to the end of the video. */
2475 if( rcc->num_entries > h->i_frame )
2477 double final_bits = rcc->entry_out[rcc->num_entries-1]->expected_bits;
2478 double video_pos = rce.expected_bits / final_bits;
2479 double scale_factor = sqrt( (1 - video_pos) * rcc->num_entries );
2480 abr_buffer *= 0.5 * X264_MAX( scale_factor, 0.5 );
2483 diff = predicted_bits - rce.expected_bits;
2485 q /= x264_clip3f((abr_buffer - diff) / abr_buffer, .5, 2);
2486 if( h->i_frame >= rcc->fps && rcc->expected_bits_sum >= 1 )
2488 /* Adjust quant based on the difference between
2489 * achieved and expected bitrate so far */
2490 double cur_time = (double)h->i_frame / rcc->num_entries;
2491 double w = x264_clip3f( cur_time*100, 0.0, 1.0 );
2492 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
2494 rcc->qp_novbv = qscale2qp( q );
2497 /* Do not overflow vbv */
2498 double expected_size = qscale2bits( &rce, q );
2499 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2500 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
2501 double qmax = q*(2 - expected_fullness);
2502 double size_constraint = 1 + expected_fullness;
2503 qmax = X264_MAX( qmax, rce.new_qscale );
2504 if( expected_fullness < .05 )
2506 qmax = X264_MIN(qmax, lmax);
2507 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
2508 ((expected_vbv < 0) && (q < lmax)))
2511 expected_size = qscale2bits(&rce, q);
2512 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2514 rcc->last_satd = x264_rc_analyse_slice( h );
2516 q = x264_clip3f( q, lmin, lmax );
2518 else /* 1pass ABR */
2520 /* Calculate the quantizer which would have produced the desired
2521 * average bitrate if it had been applied to all frames so far.
2522 * Then modulate that quant based on the current frame's complexity
2523 * relative to the average complexity so far (using the 2pass RCEQ).
2524 * Then bias the quant up or down if total size so far was far from
2526 * Result: Depending on the value of rate_tolerance, there is a
2527 * tradeoff between quality and bitrate precision. But at large
2528 * tolerances, the bit distribution approaches that of 2pass. */
2530 double wanted_bits, overflow = 1;
2532 rcc->last_satd = x264_rc_analyse_slice( h );
2533 rcc->short_term_cplxsum *= 0.5;
2534 rcc->short_term_cplxcount *= 0.5;
2535 rcc->short_term_cplxsum += rcc->last_satd / (CLIP_DURATION(h->fenc->f_duration) / BASE_FRAME_DURATION);
2536 rcc->short_term_cplxcount ++;
2538 rce.tex_bits = rcc->last_satd;
2539 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
2541 rce.p_count = rcc->nmb;
2545 rce.pict_type = pict_type;
2546 rce.i_duration = h->fenc->i_duration;
2548 if( h->param.rc.i_rc_method == X264_RC_CRF )
2550 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
2554 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
2556 /* ABR code can potentially be counterproductive in CBR, so just don't bother.
2557 * Don't run it if the frame complexity is zero either. */
2558 if( !rcc->b_vbv_min_rate && rcc->last_satd )
2560 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
2561 int i_frame_done = h->i_frame;
2562 double time_done = i_frame_done / rcc->fps;
2563 if( h->param.b_vfr_input && i_frame_done > 0 )
2564 time_done = ((double)(h->fenc->i_reordered_pts - h->i_reordered_pts_delay)) * h->param.i_timebase_num / h->param.i_timebase_den;
2565 wanted_bits = time_done * rcc->bitrate;
2566 if( wanted_bits > 0 )
2568 abr_buffer *= X264_MAX( 1, sqrt( time_done ) );
2569 overflow = x264_clip3f( 1.0 + (predicted_bits - wanted_bits) / abr_buffer, .5, 2 );
2575 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
2576 /* should test _next_ pict type, but that isn't decided yet */
2577 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
2579 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2580 q /= fabs( h->param.rc.f_ip_factor );
2582 else if( h->i_frame > 0 )
2584 if( h->param.rc.i_rc_method != X264_RC_CRF )
2586 /* Asymmetric clipping, because symmetric would prevent
2587 * overflow control in areas of rapidly oscillating complexity */
2588 double lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
2589 double lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
2590 if( overflow > 1.1 && h->i_frame > 3 )
2592 else if( overflow < 0.9 )
2595 q = x264_clip3f(q, lmin, lmax);
2598 else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
2600 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
2602 rcc->qp_novbv = qscale2qp( q );
2604 //FIXME use get_diff_limited_q() ?
2605 q = clip_qscale( h, pict_type, q );
2608 rcc->last_qscale_for[pict_type] =
2609 rcc->last_qscale = q;
2611 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
2612 rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );
2615 rcc->frame_size_planned = qscale2bits( &rce, q );
2617 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2619 /* Always use up the whole VBV in this case. */
2620 if( rcc->single_frame_vbv )
2621 rcc->frame_size_planned = rcc->buffer_rate;
2622 /* Limit planned size by MinCR */
2624 rcc->frame_size_planned = X264_MIN( rcc->frame_size_planned, rcc->frame_size_maximum );
2625 h->rc->frame_size_estimated = rcc->frame_size_planned;
2630 static void x264_threads_normalize_predictors( x264_t *h )
2632 double totalsize = 0;
2633 for( int i = 0; i < h->param.i_threads; i++ )
2634 totalsize += h->thread[i]->rc->slice_size_planned;
2635 double factor = h->rc->frame_size_planned / totalsize;
2636 for( int i = 0; i < h->param.i_threads; i++ )
2637 h->thread[i]->rc->slice_size_planned *= factor;
2640 void x264_threads_distribute_ratecontrol( x264_t *h )
2643 x264_ratecontrol_t *rc = h->rc;
2645 float qscale = qp2qscale( rc->qpm );
2647 /* Initialize row predictors */
2648 if( h->i_frame == 0 )
2649 for( int i = 0; i < h->param.i_threads; i++ )
2651 x264_t *t = h->thread[i];
2653 memcpy( t->rc->row_preds, rc->row_preds, sizeof(rc->row_preds) );
2656 for( int i = 0; i < h->param.i_threads; i++ )
2658 x264_t *t = h->thread[i];
2660 memcpy( t->rc, rc, offsetof(x264_ratecontrol_t, row_pred) );
2661 t->rc->row_pred = t->rc->row_preds[h->sh.i_type];
2662 /* Calculate the planned slice size. */
2663 if( rc->b_vbv && rc->frame_size_planned )
2666 for( row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2667 size += h->fdec->i_row_satd[row];
2668 t->rc->slice_size_planned = predict_size( &rc->pred[h->sh.i_type + (i+1)*5], qscale, size );
2671 t->rc->slice_size_planned = 0;
2673 if( rc->b_vbv && rc->frame_size_planned )
2675 x264_threads_normalize_predictors( h );
2677 if( rc->single_frame_vbv )
2679 /* Compensate for our max frame error threshold: give more bits (proportionally) to smaller slices. */
2680 for( int i = 0; i < h->param.i_threads; i++ )
2682 x264_t *t = h->thread[i];
2683 float max_frame_error = x264_clip3f( 1.0 / (t->i_threadslice_end - t->i_threadslice_start), 0.05, 0.25 );
2684 t->rc->slice_size_planned += 2 * max_frame_error * rc->frame_size_planned;
2686 x264_threads_normalize_predictors( h );
2689 for( int i = 0; i < h->param.i_threads; i++ )
2690 h->thread[i]->rc->frame_size_estimated = h->thread[i]->rc->slice_size_planned;
2694 void x264_threads_merge_ratecontrol( x264_t *h )
2696 x264_ratecontrol_t *rc = h->rc;
2699 for( int i = 0; i < h->param.i_threads; i++ )
2701 x264_t *t = h->thread[i];
2702 x264_ratecontrol_t *rct = h->thread[i]->rc;
2703 if( h->param.rc.i_vbv_buffer_size )
2706 for( int row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2707 size += h->fdec->i_row_satd[row];
2708 int bits = t->stat.frame.i_mv_bits + t->stat.frame.i_tex_bits + t->stat.frame.i_misc_bits;
2709 int mb_count = (t->i_threadslice_end - t->i_threadslice_start) * h->mb.i_mb_width;
2710 update_predictor( &rc->pred[h->sh.i_type+(i+1)*5], qp2qscale( rct->qpa_rc/mb_count ), size, bits );
2714 rc->qpa_rc += rct->qpa_rc;
2715 rc->qpa_aq += rct->qpa_aq;
2719 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
2723 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
2724 /* these vars are updated in x264_ratecontrol_start()
2725 * so copy them from the context that most recently started (prev)
2726 * to the context that's about to start (cur). */
2731 COPY(last_qscale_for);
2732 COPY(last_non_b_pict_type);
2733 COPY(short_term_cplxsum);
2734 COPY(short_term_cplxcount);
2737 COPY(mbtree.qpbuf_pos);
2738 /* these vars can be updated by x264_ratecontrol_init_reconfigurable */
2743 COPY(single_frame_vbv);
2745 COPY(rate_factor_constant);
2746 COPY(rate_factor_max_increment);
2751 #define COPY(var) next->rc->var = cur->rc->var
2752 /* these vars are updated in x264_ratecontrol_end()
2753 * so copy them from the context that most recently ended (cur)
2754 * to the context that's about to end (next) */
2756 COPY(expected_bits_sum);
2757 COPY(filler_bits_sum);
2758 COPY(wanted_bits_window);
2760 COPY(initial_cpb_removal_delay);
2761 COPY(initial_cpb_removal_delay_offset);
2762 COPY(nrt_first_access_unit);
2763 COPY(previous_cpb_final_arrival_time);
2766 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
2767 /* the rest of the variables are either constant or thread-local */
2770 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
2772 /* find an interval ending on an overflow or underflow (depending on whether
2773 * we're adding or removing bits), and starting on the earliest frame that
2774 * can influence the buffer fill of that end frame. */
2775 x264_ratecontrol_t *rcc = h->rc;
2776 const double buffer_min = .1 * rcc->buffer_size;
2777 const double buffer_max = .9 * rcc->buffer_size;
2778 double fill = fills[*t0-1];
2779 double parity = over ? 1. : -1.;
2780 int start = -1, end = -1;
2781 for( int i = *t0; i < rcc->num_entries; i++ )
2783 fill += (rcc->entry_out[i]->i_cpb_duration * rcc->vbv_max_rate * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale -
2784 qscale2bits( rcc->entry_out[i], rcc->entry_out[i]->new_qscale )) * parity;
2785 fill = x264_clip3f(fill, 0, rcc->buffer_size);
2787 if( fill <= buffer_min || i == 0 )
2793 else if( fill >= buffer_max && start >= 0 )
2798 return start >= 0 && end >= 0;
2801 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max )
2803 x264_ratecontrol_t *rcc = h->rc;
2804 double qscale_orig, qscale_new;
2808 for( int i = t0; i <= t1; i++ )
2810 qscale_orig = rcc->entry_out[i]->new_qscale;
2811 qscale_orig = x264_clip3f( qscale_orig, qscale_min, qscale_max );
2812 qscale_new = qscale_orig * adjustment;
2813 qscale_new = x264_clip3f( qscale_new, qscale_min, qscale_max );
2814 rcc->entry_out[i]->new_qscale = qscale_new;
2815 adjusted = adjusted || (qscale_new != qscale_orig);
2820 static double count_expected_bits( x264_t *h )
2822 x264_ratecontrol_t *rcc = h->rc;
2823 double expected_bits = 0;
2824 for( int i = 0; i < rcc->num_entries; i++ )
2826 ratecontrol_entry_t *rce = rcc->entry_out[i];
2827 rce->expected_bits = expected_bits;
2828 expected_bits += qscale2bits( rce, rce->new_qscale );
2830 return expected_bits;
2833 static int vbv_pass2( x264_t *h, double all_available_bits )
2835 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
2836 * frames in the interval until either buffer is full at some intermediate frame or the
2837 * last frame in the interval no longer underflows. Recompute intervals and repeat.
2838 * Then do the converse to put bits back into overflow areas until target size is met */
2840 x264_ratecontrol_t *rcc = h->rc;
2842 double expected_bits = 0;
2844 double prev_bits = 0;
2846 double qscale_min = qp2qscale( h->param.rc.i_qp_min );
2847 double qscale_max = qp2qscale( h->param.rc.i_qp_max );
2849 int adj_min, adj_max;
2850 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2854 /* adjust overall stream size */
2858 prev_bits = expected_bits;
2861 { /* not first iteration */
2862 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2863 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2867 while(adj_min && find_underflow( h, fills, &t0, &t1, 1 ))
2869 adj_min = fix_underflow( h, t0, t1, adjustment, qscale_min, qscale_max );
2874 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2876 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2878 while( adj_max && find_underflow( h, fills, &t0, &t1, 0 ) )
2879 adj_max = fix_underflow( h, t0, t1, 1.001, qscale_min, qscale_max );
2881 expected_bits = count_expected_bits( h );
2882 } while( (expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2885 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2887 /* store expected vbv filling values for tracking when encoding */
2888 for( int i = 0; i < rcc->num_entries; i++ )
2889 rcc->entry_out[i]->expected_vbv = rcc->buffer_size - fills[i];
2891 x264_free( fills-1 );
2897 static int init_pass2( x264_t *h )
2899 x264_ratecontrol_t *rcc = h->rc;
2900 uint64_t all_const_bits = 0;
2901 double timescale = (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2902 double duration = 0;
2903 for( int i = 0; i < rcc->num_entries; i++ )
2904 duration += rcc->entry[i].i_duration;
2905 duration *= timescale;
2906 uint64_t all_available_bits = h->param.rc.i_bitrate * 1000. * duration;
2907 double rate_factor, step_mult;
2908 double qblur = h->param.rc.f_qblur;
2909 double cplxblur = h->param.rc.f_complexity_blur;
2910 const int filter_size = (int)(qblur*4) | 1;
2911 double expected_bits;
2912 double *qscale, *blurred_qscale;
2913 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
2915 /* find total/average complexity & const_bits */
2916 for( int i = 0; i < rcc->num_entries; i++ )
2918 ratecontrol_entry_t *rce = &rcc->entry[i];
2919 all_const_bits += rce->misc_bits;
2922 if( all_available_bits < all_const_bits)
2924 x264_log( h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2925 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)) );
2929 /* Blur complexities, to reduce local fluctuation of QP.
2930 * We don't blur the QPs directly, because then one very simple frame
2931 * could drag down the QP of a nearby complex frame and give it more
2932 * bits than intended. */
2933 for( int i = 0; i < rcc->num_entries; i++ )
2935 ratecontrol_entry_t *rce = &rcc->entry[i];
2936 double weight_sum = 0;
2937 double cplx_sum = 0;
2938 double weight = 1.0;
2939 double gaussian_weight;
2940 /* weighted average of cplx of future frames */
2941 for( int j = 1; j < cplxblur*2 && j < rcc->num_entries-i; j++ )
2943 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2944 double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2945 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2946 if( weight < .0001 )
2948 gaussian_weight = weight * exp( -j*j/200.0 );
2949 weight_sum += gaussian_weight;
2950 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2952 /* weighted average of cplx of past frames */
2954 for( int j = 0; j <= cplxblur*2 && j <= i; j++ )
2956 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2957 double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2958 gaussian_weight = weight * exp( -j*j/200.0 );
2959 weight_sum += gaussian_weight;
2960 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2961 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2962 if( weight < .0001 )
2965 rce->blurred_complexity = cplx_sum / weight_sum;
2968 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2969 if( filter_size > 1 )
2970 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2972 blurred_qscale = qscale;
2974 /* Search for a factor which, when multiplied by the RCEQ values from
2975 * each frame, adds up to the desired total size.
2976 * There is no exact closed-form solution because of VBV constraints and
2977 * because qscale2bits is not invertible, but we can start with the simple
2978 * approximation of scaling the 1st pass by the ratio of bitrates.
2979 * The search range is probably overkill, but speed doesn't matter here. */
2982 for( int i = 0; i < rcc->num_entries; i++ )
2984 double q = get_qscale(h, &rcc->entry[i], 1.0, i);
2985 expected_bits += qscale2bits(&rcc->entry[i], q);
2986 rcc->last_qscale_for[rcc->entry[i].pict_type] = q;
2988 step_mult = all_available_bits / expected_bits;
2991 for( double step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2994 rate_factor += step;
2996 rcc->last_non_b_pict_type = -1;
2997 rcc->last_accum_p_norm = 1;
2998 rcc->accum_p_norm = 0;
3000 rcc->last_qscale_for[0] =
3001 rcc->last_qscale_for[1] =
3002 rcc->last_qscale_for[2] = pow( base_cplx, 1 - rcc->qcompress ) / rate_factor;
3005 for( int i = 0; i < rcc->num_entries; i++ )
3007 qscale[i] = get_qscale( h, &rcc->entry[i], rate_factor, -1 );
3008 rcc->last_qscale_for[rcc->entry[i].pict_type] = qscale[i];
3011 /* fixed I/B qscale relative to P */
3012 for( int i = rcc->num_entries-1; i >= 0; i-- )
3014 qscale[i] = get_diff_limited_q( h, &rcc->entry[i], qscale[i], i );
3015 assert(qscale[i] >= 0);
3019 if( filter_size > 1 )
3021 assert( filter_size%2 == 1 );
3022 for( int i = 0; i < rcc->num_entries; i++ )
3024 ratecontrol_entry_t *rce = &rcc->entry[i];
3025 double q = 0.0, sum = 0.0;
3027 for( int j = 0; j < filter_size; j++ )
3029 int idx = i+j-filter_size/2;
3031 double coeff = qblur==0 ? 1.0 : exp( -d*d/(qblur*qblur) );
3032 if( idx < 0 || idx >= rcc->num_entries )
3034 if( rce->pict_type != rcc->entry[idx].pict_type )
3036 q += qscale[idx] * coeff;
3039 blurred_qscale[i] = q/sum;
3043 /* find expected bits */
3044 for( int i = 0; i < rcc->num_entries; i++ )
3046 ratecontrol_entry_t *rce = &rcc->entry[i];
3047 rce->new_qscale = clip_qscale( h, rce->pict_type, blurred_qscale[i] );
3048 assert(rce->new_qscale >= 0);
3049 expected_bits += qscale2bits( rce, rce->new_qscale );
3052 if( expected_bits > all_available_bits )
3053 rate_factor -= step;
3056 x264_free( qscale );
3057 if( filter_size > 1 )
3058 x264_free( blurred_qscale );
3061 if( vbv_pass2( h, all_available_bits ) )
3063 expected_bits = count_expected_bits( h );
3065 if( fabs( expected_bits/all_available_bits - 1.0 ) > 0.01 )
3068 for( int i = 0; i < rcc->num_entries; i++ )
3069 avgq += rcc->entry[i].new_qscale;
3070 avgq = qscale2qp( avgq / rcc->num_entries );
3072 if( expected_bits > all_available_bits || !rcc->b_vbv )
3073 x264_log( h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n" );
3074 x264_log( h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
3075 (float)h->param.rc.i_bitrate,
3076 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
3078 if( expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2 )
3080 if( h->param.rc.i_qp_min > 0 )
3081 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min );
3083 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate\n" );
3085 else if( expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2 )
3087 if( h->param.rc.i_qp_max < QP_MAX )
3088 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max );
3090 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate\n");
3092 else if( !(rcc->b_2pass && rcc->b_vbv) )
3093 x264_log( h, X264_LOG_WARNING, "internal error\n" );