1 /*****************************************************************************
2 * ratecontrol.c: ratecontrol
3 *****************************************************************************
4 * Copyright (C) 2005-2015 x264 project
6 * Authors: Loren Merritt <lorenm@u.washington.edu>
7 * Michael Niedermayer <michaelni@gmx.at>
8 * Gabriel Bouvigne <gabriel.bouvigne@joost.com>
9 * Fiona Glaser <fiona@x264.com>
10 * Måns Rullgård <mru@mru.ath.cx>
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License as published by
14 * the Free Software Foundation; either version 2 of the License, or
15 * (at your option) any later version.
17 * This program is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 * GNU General Public License for more details.
22 * You should have received a copy of the GNU General Public License
23 * along with this program; if not, write to the Free Software
24 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
26 * This program is also available under a commercial proprietary license.
27 * For more information, contact us at licensing@x264.com.
28 *****************************************************************************/
30 #define _ISOC99_SOURCE
31 #undef NDEBUG // always check asserts, the speed effect is far too small to disable them
33 #include "common/common.h"
34 #include "ratecontrol.h"
46 uint64_t expected_bits; /*total expected bits up to the current frame (current one excluded)*/
53 float blurred_complexity;
56 int16_t i_weight_denom[2];
60 int64_t i_cpb_duration;
61 } ratecontrol_entry_t;
72 struct x264_ratecontrol_t
81 double rate_tolerance;
83 int nmb; /* number of macroblocks in a frame */
87 ratecontrol_entry_t *rce;
88 int qp; /* qp for current frame */
89 float qpm; /* qp for current macroblock: precise float for AQ */
90 float qpa_rc; /* average of macroblocks' qp before aq */
92 int qpa_aq; /* average of macroblocks' qp after aq */
94 float qp_novbv; /* QP for the current frame if 1-pass VBV was disabled. */
98 int64_t buffer_fill_final;
99 int64_t buffer_fill_final_min;
100 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
101 double buffer_rate; /* # of bits added to buffer_fill after each frame */
102 double vbv_max_rate; /* # of bits added to buffer_fill per second */
103 predictor_t *pred; /* predict frame size from satd */
104 int single_frame_vbv;
105 float rate_factor_max_increment; /* Don't allow RF above (CRF + this value). */
110 double cplxr_sum; /* sum of bits*qscale/rceq */
111 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow, only includes finished frames */
112 int64_t filler_bits_sum; /* sum in bits of finished frames' filler data */
113 double wanted_bits_window; /* target bitrate * window */
115 double short_term_cplxsum;
116 double short_term_cplxcount;
117 double rate_factor_constant;
122 FILE *p_stat_file_out;
123 char *psz_stat_file_tmpname;
124 FILE *p_mbtree_stat_file_out;
125 char *psz_mbtree_stat_file_tmpname;
126 char *psz_mbtree_stat_file_name;
127 FILE *p_mbtree_stat_file_in;
129 int num_entries; /* number of ratecontrol_entry_ts */
130 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
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 ) / 6.0f );
206 static inline float qscale2qp( float qscale )
208 return 12.0f + 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 for( int i = 0; i < rc->mbtree.src_mb_count; i++ )
570 int16_t qp_fix8 = endian_fix16( rc->mbtree.qp_buffer[rc->mbtree.qpbuf_pos][i] );
571 dst[i] = qp_fix8 * (1.f/256.f);
573 if( rc->mbtree.rescale_enabled )
574 x264_macroblock_tree_rescale( h, rc, frame->f_qp_offset );
575 if( h->frames.b_have_lowres )
576 for( int i = 0; i < h->mb.i_mb_count; i++ )
577 frame->i_inv_qscale_factor[i] = x264_exp2fix8( frame->f_qp_offset[i] );
578 rc->mbtree.qpbuf_pos--;
581 x264_stack_align( x264_adaptive_quant_frame, h, frame, quant_offsets );
584 x264_log( h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n" );
588 int x264_reference_build_list_optimal( x264_t *h )
590 ratecontrol_entry_t *rce = h->rc->rce;
591 x264_frame_t *frames[16];
592 x264_weight_t weights[16][3];
595 if( rce->refs != h->i_ref[0] )
598 memcpy( frames, h->fref[0], sizeof(frames) );
599 memcpy( refcount, rce->refcount, sizeof(refcount) );
600 memcpy( weights, h->fenc->weight, sizeof(weights) );
601 memset( &h->fenc->weight[1][0], 0, sizeof(x264_weight_t[15][3]) );
603 /* For now don't reorder ref 0; it seems to lower quality
604 in most cases due to skips. */
605 for( int ref = 1; ref < h->i_ref[0]; ref++ )
610 for( int i = 1; i < h->i_ref[0]; i++ )
611 /* Favor lower POC as a tiebreaker. */
612 COPY2_IF_GT( max, refcount[i], bestref, i );
614 /* FIXME: If there are duplicates from frames other than ref0 then it is possible
615 * that the optimal ordering doesnt place every duplicate. */
617 refcount[bestref] = -1;
618 h->fref[0][ref] = frames[bestref];
619 memcpy( h->fenc->weight[ref], weights[bestref], sizeof(weights[bestref]) );
625 static char *x264_strcat_filename( char *input, char *suffix )
627 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
630 strcpy( output, input );
631 strcat( output, suffix );
635 void x264_ratecontrol_init_reconfigurable( x264_t *h, int b_init )
637 x264_ratecontrol_t *rc = h->rc;
638 if( !b_init && rc->b_2pass )
641 if( h->param.rc.i_rc_method == X264_RC_CRF )
643 /* Arbitrary rescaling to make CRF somewhat similar to QP.
644 * Try to compensate for MB-tree's effects as well. */
645 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
646 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
647 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
648 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset + QP_BD_OFFSET );
651 if( h->param.rc.i_vbv_max_bitrate > 0 && h->param.rc.i_vbv_buffer_size > 0 )
653 /* We don't support changing the ABR bitrate right now,
654 so if the stream starts as CBR, keep it CBR. */
655 if( rc->b_vbv_min_rate )
656 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
658 if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
660 h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
661 x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
662 h->param.rc.i_vbv_buffer_size );
665 int kilobit_size = h->param.i_avcintra_class ? 1024 : 1000;
666 int vbv_buffer_size = h->param.rc.i_vbv_buffer_size * kilobit_size;
667 int vbv_max_bitrate = h->param.rc.i_vbv_max_bitrate * kilobit_size;
670 if( h->param.i_nal_hrd && b_init )
672 h->sps->vui.hrd.i_cpb_cnt = 1;
673 h->sps->vui.hrd.b_cbr_hrd = h->param.i_nal_hrd == X264_NAL_HRD_CBR;
674 h->sps->vui.hrd.i_time_offset_length = 0;
679 // normalize HRD size and rate to the value / scale notation
680 h->sps->vui.hrd.i_bit_rate_scale = x264_clip3( x264_ctz( vbv_max_bitrate ) - BR_SHIFT, 0, 15 );
681 h->sps->vui.hrd.i_bit_rate_value = vbv_max_bitrate >> ( h->sps->vui.hrd.i_bit_rate_scale + BR_SHIFT );
682 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 );
683 h->sps->vui.hrd.i_cpb_size_scale = x264_clip3( x264_ctz( vbv_buffer_size ) - CPB_SHIFT, 0, 15 );
684 h->sps->vui.hrd.i_cpb_size_value = vbv_buffer_size >> ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
685 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 );
691 #define MAX_DURATION 0.5
693 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 );
694 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;
695 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);
697 h->sps->vui.hrd.i_initial_cpb_removal_delay_length = 2 + x264_clip3( 32 - x264_clz( max_delay ), 4, 22 );
698 h->sps->vui.hrd.i_cpb_removal_delay_length = x264_clip3( 32 - x264_clz( max_cpb_output_delay ), 4, 31 );
699 h->sps->vui.hrd.i_dpb_output_delay_length = x264_clip3( 32 - x264_clz( max_dpb_output_delay ), 4, 31 );
703 vbv_buffer_size = h->sps->vui.hrd.i_cpb_size_unscaled;
704 vbv_max_bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
706 else if( h->param.i_nal_hrd && !b_init )
708 x264_log( h, X264_LOG_WARNING, "VBV parameters cannot be changed when NAL HRD is in use\n" );
711 h->sps->vui.hrd.i_bit_rate_unscaled = vbv_max_bitrate;
712 h->sps->vui.hrd.i_cpb_size_unscaled = vbv_buffer_size;
714 if( rc->b_vbv_min_rate )
715 rc->bitrate = (double)h->param.rc.i_bitrate * kilobit_size;
716 rc->buffer_rate = vbv_max_bitrate / rc->fps;
717 rc->vbv_max_rate = vbv_max_bitrate;
718 rc->buffer_size = vbv_buffer_size;
719 rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
720 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
721 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
722 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.f_rf_constant_max )
724 rc->rate_factor_max_increment = h->param.rc.f_rf_constant_max - h->param.rc.f_rf_constant;
725 if( rc->rate_factor_max_increment <= 0 )
727 x264_log( h, X264_LOG_WARNING, "CRF max must be greater than CRF\n" );
728 rc->rate_factor_max_increment = 0;
733 if( h->param.rc.f_vbv_buffer_init > 1. )
734 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 );
735 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);
736 rc->buffer_fill_final =
737 rc->buffer_fill_final_min = rc->buffer_size * h->param.rc.f_vbv_buffer_init * h->sps->vui.i_time_scale;
739 rc->b_vbv_min_rate = !rc->b_2pass
740 && h->param.rc.i_rc_method == X264_RC_ABR
741 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
746 int x264_ratecontrol_new( x264_t *h )
748 x264_ratecontrol_t *rc;
752 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
755 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
756 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
758 /* FIXME: use integers */
759 if( h->param.i_fps_num > 0 && h->param.i_fps_den > 0 )
760 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
764 if( h->param.rc.b_mb_tree )
766 h->param.rc.f_pb_factor = 1;
770 rc->qcompress = h->param.rc.f_qcompress;
772 rc->bitrate = h->param.rc.i_bitrate * (h->param.i_avcintra_class ? 1024. : 1000.);
773 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
774 rc->nmb = h->mb.i_mb_count;
775 rc->last_non_b_pict_type = -1;
778 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
780 x264_log( h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n" );
784 x264_ratecontrol_init_reconfigurable( h, 1 );
786 if( h->param.i_nal_hrd )
788 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale;
789 uint64_t num = 90000;
790 x264_reduce_fraction64( &num, &denom );
791 rc->hrd_multiply_denom = 90000 / num;
793 double bits_required = log2( 90000 / rc->hrd_multiply_denom )
794 + log2( h->sps->vui.i_time_scale )
795 + log2( h->sps->vui.hrd.i_cpb_size_unscaled );
796 if( bits_required >= 63 )
798 x264_log( h, X264_LOG_ERROR, "HRD with very large timescale and bufsize not supported\n" );
803 if( rc->rate_tolerance < 0.01 )
805 x264_log( h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n" );
806 rc->rate_tolerance = 0.01;
809 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
813 /* FIXME ABR_INIT_QP is actually used only in CRF */
814 #define ABR_INIT_QP (( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 ) + QP_BD_OFFSET)
815 rc->accum_p_norm = .01;
816 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
817 /* estimated ratio that produces a reasonable QP for the first I-frame */
818 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
819 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
820 rc->last_non_b_pict_type = SLICE_TYPE_I;
823 rc->ip_offset = 6.0 * log2f( h->param.rc.f_ip_factor );
824 rc->pb_offset = 6.0 * log2f( h->param.rc.f_pb_factor );
825 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
826 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, QP_MAX );
827 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, QP_MAX );
828 h->mb.ip_offset = rc->ip_offset + 0.5;
830 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
831 rc->last_qscale = qp2qscale( 26 );
832 int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
833 CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
834 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
835 static const float pred_coeff_table[3] = { 1.0, 1.0, 1.5 };
836 for( int i = 0; i < 3; i++ )
838 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
839 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
840 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
841 for( int j = 0; j < num_preds; j++ )
843 rc->pred[i+j*5].coeff_min = pred_coeff_table[i] / 2;
844 rc->pred[i+j*5].coeff = pred_coeff_table[i];
845 rc->pred[i+j*5].count = 1.0;
846 rc->pred[i+j*5].decay = 0.5;
847 rc->pred[i+j*5].offset = 0.0;
849 for( int j = 0; j < 2; j++ )
851 rc->row_preds[i][j].coeff_min = .25 / 4;
852 rc->row_preds[i][j].coeff = .25;
853 rc->row_preds[i][j].count = 1.0;
854 rc->row_preds[i][j].decay = 0.5;
855 rc->row_preds[i][j].offset = 0.0;
858 rc->pred_b_from_p->coeff_min = 0.5 / 2;
859 rc->pred_b_from_p->coeff = 0.5;
860 rc->pred_b_from_p->count = 1.0;
861 rc->pred_b_from_p->decay = 0.5;
862 rc->pred_b_from_p->offset = 0.0;
864 if( parse_zones( h ) < 0 )
866 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
870 /* Load stat file and init 2pass algo */
871 if( h->param.rc.b_stat_read )
873 char *p, *stats_in, *stats_buf;
875 /* read 1st pass stats */
876 assert( h->param.rc.psz_stat_in );
877 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
880 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n" );
883 if( h->param.rc.b_mb_tree )
885 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
886 if( !mbtree_stats_in )
888 rc->p_mbtree_stat_file_in = x264_fopen( mbtree_stats_in, "rb" );
889 x264_free( mbtree_stats_in );
890 if( !rc->p_mbtree_stat_file_in )
892 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n" );
897 /* check whether 1st pass options were compatible with current options */
898 if( strncmp( stats_buf, "#options:", 9 ) )
900 x264_log( h, X264_LOG_ERROR, "options list in stats file not valid\n" );
904 float res_factor, res_factor_bits;
908 char *opts = stats_buf;
909 stats_in = strchr( stats_buf, '\n' );
914 if( sscanf( opts, "#options: %dx%d", &i, &j ) != 2 )
916 x264_log( h, X264_LOG_ERROR, "resolution specified in stats file not valid\n" );
919 else if( h->param.rc.b_mb_tree )
921 rc->mbtree.srcdim[0] = i;
922 rc->mbtree.srcdim[1] = j;
924 res_factor = (float)h->param.i_width * h->param.i_height / (i*j);
925 /* Change in bits relative to resolution isn't quite linear on typical sources,
926 * so we'll at least try to roughly approximate this effect. */
927 res_factor_bits = powf( res_factor, 0.7 );
929 if( !( p = strstr( opts, "timebase=" ) ) || sscanf( p, "timebase=%u/%u", &k, &l ) != 2 )
931 x264_log( h, X264_LOG_ERROR, "timebase specified in stats file not valid\n" );
934 if( k != h->param.i_timebase_num || l != h->param.i_timebase_den )
936 x264_log( h, X264_LOG_ERROR, "timebase mismatch with 1st pass (%u/%u vs %u/%u)\n",
937 h->param.i_timebase_num, h->param.i_timebase_den, k, l );
941 CMP_OPT_FIRST_PASS( "bitdepth", BIT_DEPTH );
942 CMP_OPT_FIRST_PASS( "weightp", X264_MAX( 0, h->param.analyse.i_weighted_pred ) );
943 CMP_OPT_FIRST_PASS( "bframes", h->param.i_bframe );
944 CMP_OPT_FIRST_PASS( "b_pyramid", h->param.i_bframe_pyramid );
945 CMP_OPT_FIRST_PASS( "intra_refresh", h->param.b_intra_refresh );
946 CMP_OPT_FIRST_PASS( "open_gop", h->param.b_open_gop );
947 CMP_OPT_FIRST_PASS( "bluray_compat", h->param.b_bluray_compat );
949 if( (p = strstr( opts, "interlaced=" )) )
951 char *current = h->param.b_interlaced ? h->param.b_tff ? "tff" : "bff" : h->param.b_fake_interlaced ? "fake" : "0";
953 sscanf( p, "interlaced=%4s", buf );
954 if( strcmp( current, buf ) )
956 x264_log( h, X264_LOG_ERROR, "different interlaced setting than first pass (%s vs %s)\n", current, buf );
961 if( (p = strstr( opts, "keyint=" )) )
964 char buf[13] = "infinite ";
965 if( h->param.i_keyint_max != X264_KEYINT_MAX_INFINITE )
966 sprintf( buf, "%d ", h->param.i_keyint_max );
967 if( strncmp( p, buf, strlen(buf) ) )
969 x264_log( h, X264_LOG_ERROR, "different keyint setting than first pass (%.*s vs %.*s)\n",
970 strlen(buf)-1, buf, strcspn(p, " "), p );
975 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
976 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
978 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
980 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
981 h->mb.b_direct_auto_write = 1;
984 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
985 h->param.i_bframe_adaptive = i;
986 else if( h->param.i_bframe )
988 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
992 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 ) )
993 h->param.rc.i_lookahead = i;
996 /* find number of pics */
999 for( num_entries = -1; p; num_entries++ )
1000 p = strchr( p + 1, ';' );
1003 x264_log( h, X264_LOG_ERROR, "empty stats file\n" );
1006 rc->num_entries = num_entries;
1008 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
1010 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
1011 h->param.i_frame_total, rc->num_entries );
1013 if( h->param.i_frame_total > rc->num_entries )
1015 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
1016 h->param.i_frame_total, rc->num_entries );
1020 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
1022 /* init all to skipped p frames */
1023 for( int i = 0; i < rc->num_entries; i++ )
1025 ratecontrol_entry_t *rce = &rc->entry[i];
1026 rce->pict_type = SLICE_TYPE_P;
1027 rce->qscale = rce->new_qscale = qp2qscale( 20 );
1028 rce->misc_bits = rc->nmb + 10;
1034 double total_qp_aq = 0;
1035 for( int i = 0; i < rc->num_entries; i++ )
1037 ratecontrol_entry_t *rce;
1045 next= strchr(p, ';');
1047 *next++ = 0; //sscanf is unbelievably slow on long strings
1048 e = sscanf( p, " in:%d ", &frame_number );
1050 if( frame_number < 0 || frame_number >= rc->num_entries )
1052 x264_log( h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i );
1055 rce = &rc->entry[frame_number];
1056 rce->direct_mode = 0;
1058 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",
1059 &pict_type, &rce->i_duration, &rce->i_cpb_duration, &qp_rc, &qp_aq, &rce->tex_bits,
1060 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
1061 &rce->s_count, &rce->direct_mode );
1062 rce->tex_bits *= res_factor_bits;
1063 rce->mv_bits *= res_factor_bits;
1064 rce->misc_bits *= res_factor_bits;
1065 rce->i_count *= res_factor;
1066 rce->p_count *= res_factor;
1067 rce->s_count *= res_factor;
1069 p = strstr( p, "ref:" );
1073 for( ref = 0; ref < 16; ref++ )
1075 if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
1077 p = strchr( p+1, ' ' );
1084 rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
1085 char *w = strchr( p, 'w' );
1088 int count = sscanf( w, "w:%hd,%hd,%hd,%hd,%hd,%hd,%hd,%hd",
1089 &rce->i_weight_denom[0], &rce->weight[0][0], &rce->weight[0][1],
1090 &rce->i_weight_denom[1], &rce->weight[1][0], &rce->weight[1][1],
1091 &rce->weight[2][0], &rce->weight[2][1] );
1093 rce->i_weight_denom[1] = -1;
1094 else if ( count != 8 )
1095 rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
1098 if( pict_type != 'b' )
1099 rce->kept_as_ref = 1;
1103 rce->frame_type = X264_TYPE_IDR;
1104 rce->pict_type = SLICE_TYPE_I;
1107 rce->frame_type = X264_TYPE_I;
1108 rce->pict_type = SLICE_TYPE_I;
1111 rce->frame_type = X264_TYPE_P;
1112 rce->pict_type = SLICE_TYPE_P;
1115 rce->frame_type = X264_TYPE_BREF;
1116 rce->pict_type = SLICE_TYPE_B;
1119 rce->frame_type = X264_TYPE_B;
1120 rce->pict_type = SLICE_TYPE_B;
1122 default: e = -1; break;
1127 x264_log( h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e );
1130 rce->qscale = qp2qscale( qp_rc );
1131 total_qp_aq += qp_aq;
1134 if( !h->param.b_stitchable )
1135 h->pps->i_pic_init_qp = SPEC_QP( (int)(total_qp_aq / rc->num_entries + 0.5) );
1137 x264_free( stats_buf );
1139 if( h->param.rc.i_rc_method == X264_RC_ABR )
1141 if( init_pass2( h ) < 0 )
1143 } /* else we're using constant quant, so no need to run the bitrate allocation */
1146 /* Open output file */
1147 /* If input and output files are the same, output to a temp file
1148 * and move it to the real name only when it's complete */
1149 if( h->param.rc.b_stat_write )
1152 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
1153 if( !rc->psz_stat_file_tmpname )
1156 rc->p_stat_file_out = x264_fopen( rc->psz_stat_file_tmpname, "wb" );
1157 if( rc->p_stat_file_out == NULL )
1159 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n" );
1163 p = x264_param2string( &h->param, 1 );
1165 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
1167 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
1169 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
1170 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
1171 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
1174 rc->p_mbtree_stat_file_out = x264_fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
1175 if( rc->p_mbtree_stat_file_out == NULL )
1177 x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n" );
1183 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
1185 if( !h->param.rc.b_stat_read )
1187 rc->mbtree.srcdim[0] = h->param.i_width;
1188 rc->mbtree.srcdim[1] = h->param.i_height;
1190 if( x264_macroblock_tree_rescale_init( h, rc ) < 0 )
1194 for( int i = 0; i<h->param.i_threads; i++ )
1196 h->thread[i]->rc = rc+i;
1200 h->thread[i]->param = h->param;
1201 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
1202 h->thread[i]->mb.ip_offset = h->mb.ip_offset;
1211 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
1214 char *tok, UNUSED *saveptr=NULL;
1216 z->f_bitrate_factor = 1;
1217 if( 3 <= sscanf(p, "%d,%d,q=%d%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
1219 else if( 3 <= sscanf(p, "%d,%d,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
1221 else if( 2 <= sscanf(p, "%d,%d%n", &z->i_start, &z->i_end, &len) )
1225 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
1231 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
1232 memcpy( z->param, &h->param, sizeof(x264_param_t) );
1233 z->param->param_free = x264_free;
1234 while( (tok = strtok_r( p, ",", &saveptr )) )
1236 char *val = strchr( tok, '=' );
1242 if( x264_param_parse( z->param, tok, val ) )
1244 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
1254 static int parse_zones( x264_t *h )
1256 x264_ratecontrol_t *rc = h->rc;
1257 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
1259 char *psz_zones, *p;
1260 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
1261 strcpy( psz_zones, h->param.rc.psz_zones );
1262 h->param.rc.i_zones = 1;
1263 for( p = psz_zones; *p; p++ )
1264 h->param.rc.i_zones += (*p == '/');
1265 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
1267 for( int i = 0; i < h->param.rc.i_zones; i++ )
1269 int i_tok = strcspn( p, "/" );
1271 if( parse_zone( h, &h->param.rc.zones[i], p ) )
1275 x264_free( psz_zones );
1278 if( h->param.rc.i_zones > 0 )
1280 for( int i = 0; i < h->param.rc.i_zones; i++ )
1282 x264_zone_t z = h->param.rc.zones[i];
1283 if( z.i_start < 0 || z.i_start > z.i_end )
1285 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
1286 z.i_start, z.i_end );
1289 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
1291 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
1292 z.f_bitrate_factor );
1297 rc->i_zones = h->param.rc.i_zones + 1;
1298 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
1299 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
1301 // default zone to fall back to if none of the others match
1302 rc->zones[0].i_start = 0;
1303 rc->zones[0].i_end = INT_MAX;
1304 rc->zones[0].b_force_qp = 0;
1305 rc->zones[0].f_bitrate_factor = 1;
1306 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
1307 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
1308 for( int i = 1; i < rc->i_zones; i++ )
1310 if( !rc->zones[i].param )
1311 rc->zones[i].param = rc->zones[0].param;
1320 static x264_zone_t *get_zone( x264_t *h, int frame_num )
1322 for( int i = h->rc->i_zones - 1; i >= 0; i-- )
1324 x264_zone_t *z = &h->rc->zones[i];
1325 if( frame_num >= z->i_start && frame_num <= z->i_end )
1331 void x264_ratecontrol_summary( x264_t *h )
1333 x264_ratecontrol_t *rc = h->rc;
1334 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
1336 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
1337 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
1338 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
1339 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
1340 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset - QP_BD_OFFSET );
1344 void x264_ratecontrol_delete( x264_t *h )
1346 x264_ratecontrol_t *rc = h->rc;
1349 if( rc->p_stat_file_out )
1351 b_regular_file = x264_is_regular_file( rc->p_stat_file_out );
1352 fclose( rc->p_stat_file_out );
1353 if( h->i_frame >= rc->num_entries && b_regular_file )
1354 if( x264_rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
1356 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1357 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
1359 x264_free( rc->psz_stat_file_tmpname );
1361 if( rc->p_mbtree_stat_file_out )
1363 b_regular_file = x264_is_regular_file( rc->p_mbtree_stat_file_out );
1364 fclose( rc->p_mbtree_stat_file_out );
1365 if( h->i_frame >= rc->num_entries && b_regular_file )
1366 if( x264_rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
1368 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1369 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
1371 x264_free( rc->psz_mbtree_stat_file_tmpname );
1372 x264_free( rc->psz_mbtree_stat_file_name );
1374 if( rc->p_mbtree_stat_file_in )
1375 fclose( rc->p_mbtree_stat_file_in );
1376 x264_free( rc->pred );
1377 x264_free( rc->pred_b_from_p );
1378 x264_free( rc->entry );
1379 x264_macroblock_tree_rescale_destroy( rc );
1382 x264_free( rc->zones[0].param );
1383 for( int i = 1; i < rc->i_zones; i++ )
1384 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
1385 rc->zones[i].param->param_free( rc->zones[i].param );
1386 x264_free( rc->zones );
1391 static void accum_p_qp_update( x264_t *h, float qp )
1393 x264_ratecontrol_t *rc = h->rc;
1394 rc->accum_p_qp *= .95;
1395 rc->accum_p_norm *= .95;
1396 rc->accum_p_norm += 1;
1397 if( h->sh.i_type == SLICE_TYPE_I )
1398 rc->accum_p_qp += qp + rc->ip_offset;
1400 rc->accum_p_qp += qp;
1403 /* Before encoding a frame, choose a QP for it */
1404 void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
1406 x264_ratecontrol_t *rc = h->rc;
1407 ratecontrol_entry_t *rce = NULL;
1408 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
1413 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
1414 x264_encoder_reconfig_apply( h, zone->param );
1415 rc->prev_zone = zone;
1417 if( h->param.rc.b_stat_read )
1419 int frame = h->fenc->i_frame;
1420 assert( frame >= 0 && frame < rc->num_entries );
1421 rce = h->rc->rce = &h->rc->entry[frame];
1423 if( h->sh.i_type == SLICE_TYPE_B
1424 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
1426 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
1427 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
1433 memset( h->fdec->i_row_bits, 0, h->mb.i_mb_height * sizeof(int) );
1434 memset( h->fdec->f_row_qp, 0, h->mb.i_mb_height * sizeof(float) );
1435 memset( h->fdec->f_row_qscale, 0, h->mb.i_mb_height * sizeof(float) );
1436 rc->row_pred = rc->row_preds[h->sh.i_type];
1437 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;
1438 update_vbv_plan( h, overhead );
1440 const x264_level_t *l = x264_levels;
1441 while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
1444 int mincr = l->mincr;
1446 if( h->param.b_bluray_compat )
1449 /* Profiles above High don't require minCR, so just set the maximum to a large value. */
1450 if( h->sps->i_profile_idc > PROFILE_HIGH )
1451 rc->frame_size_maximum = 1e9;
1454 /* The spec has a bizarre special case for the first frame. */
1455 if( h->i_frame == 0 )
1457 //384 * ( Max( PicSizeInMbs, fR * MaxMBPS ) + MaxMBPS * ( tr( 0 ) - tr,n( 0 ) ) ) / MinCR
1458 double fr = 1. / 172;
1459 int pic_size_in_mbs = h->mb.i_mb_width * h->mb.i_mb_height;
1460 rc->frame_size_maximum = 384 * BIT_DEPTH * X264_MAX( pic_size_in_mbs, fr*l->mbps ) / mincr;
1464 //384 * MaxMBPS * ( tr( n ) - tr( n - 1 ) ) / MinCR
1465 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;
1470 if( h->sh.i_type != SLICE_TYPE_B )
1471 rc->bframes = h->fenc->i_bframes;
1475 q = qscale2qp( rate_estimate_qscale( h ) );
1477 else if( rc->b_2pass )
1479 rce->new_qscale = rate_estimate_qscale( h );
1480 q = qscale2qp( rce->new_qscale );
1484 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
1485 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
1487 q = rc->qp_constant[ h->sh.i_type ];
1491 if( zone->b_force_qp )
1492 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1494 q -= 6*log2f( zone->f_bitrate_factor );
1497 if( i_force_qp != X264_QP_AUTO )
1500 q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1502 rc->qpa_rc = rc->qpa_rc_prev =
1503 rc->qpa_aq = rc->qpa_aq_prev = 0;
1504 rc->qp = x264_clip3( q + 0.5f, 0, QP_MAX );
1505 h->fdec->f_qp_avg_rc =
1506 h->fdec->f_qp_avg_aq =
1509 rce->new_qp = rc->qp;
1511 accum_p_qp_update( h, rc->qpm );
1513 if( h->sh.i_type != SLICE_TYPE_B )
1514 rc->last_non_b_pict_type = h->sh.i_type;
1517 static float predict_row_size( x264_t *h, int y, float qscale )
1519 /* average between two predictors:
1520 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1521 x264_ratecontrol_t *rc = h->rc;
1522 float pred_s = predict_size( &rc->row_pred[0], qscale, h->fdec->i_row_satd[y] );
1523 if( h->sh.i_type == SLICE_TYPE_I || qscale >= h->fref[0][0]->f_row_qscale[y] )
1525 if( h->sh.i_type == SLICE_TYPE_P
1526 && h->fref[0][0]->i_type == h->fdec->i_type
1527 && h->fref[0][0]->f_row_qscale[y] > 0
1528 && h->fref[0][0]->i_row_satd[y] > 0
1529 && (abs(h->fref[0][0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1531 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]
1532 * h->fref[0][0]->f_row_qscale[y] / qscale;
1533 return (pred_s + pred_t) * 0.5f;
1537 /* Our QP is lower than the reference! */
1540 float pred_intra = predict_size( &rc->row_pred[1], qscale, h->fdec->i_row_satds[0][0][y] );
1541 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1542 return pred_intra + pred_s;
1546 static int row_bits_so_far( x264_t *h, int y )
1549 for( int i = h->i_threadslice_start; i <= y; i++ )
1550 bits += h->fdec->i_row_bits[i];
1554 static float predict_row_size_sum( x264_t *h, int y, float qp )
1556 float qscale = qp2qscale( qp );
1557 float bits = row_bits_so_far( h, y );
1558 for( int i = y+1; i < h->i_threadslice_end; i++ )
1559 bits += predict_row_size( h, i, qscale );
1564 * eliminate all use of qp in row ratecontrol: make it entirely qscale-based.
1565 * make this function stop being needlessly O(N^2)
1566 * update more often than once per row? */
1567 int x264_ratecontrol_mb( x264_t *h, int bits )
1569 x264_ratecontrol_t *rc = h->rc;
1570 const int y = h->mb.i_mb_y;
1572 h->fdec->i_row_bits[y] += bits;
1573 rc->qpa_aq += h->mb.i_qp;
1575 if( h->mb.i_mb_x != h->mb.i_mb_width - 1 )
1579 rc->qpa_rc += rc->qpm * h->mb.i_mb_width;
1584 float qscale = qp2qscale( rc->qpm );
1585 h->fdec->f_row_qp[y] = rc->qpm;
1586 h->fdec->f_row_qscale[y] = qscale;
1588 update_predictor( &rc->row_pred[0], qscale, h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1589 if( h->sh.i_type == SLICE_TYPE_P && rc->qpm < h->fref[0][0]->f_row_qp[y] )
1590 update_predictor( &rc->row_pred[1], qscale, h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1592 /* update ratecontrol per-mbpair in MBAFF */
1593 if( SLICE_MBAFF && !(y&1) )
1596 /* FIXME: We don't currently support the case where there's a slice
1597 * boundary in between. */
1598 int can_reencode_row = h->sh.i_first_mb <= ((h->mb.i_mb_y - SLICE_MBAFF) * h->mb.i_mb_stride);
1600 /* tweak quality based on difference from predicted size */
1601 float prev_row_qp = h->fdec->f_row_qp[y];
1602 float qp_absolute_max = h->param.rc.i_qp_max;
1603 if( rc->rate_factor_max_increment )
1604 qp_absolute_max = X264_MIN( qp_absolute_max, rc->qp_novbv + rc->rate_factor_max_increment );
1605 float qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, qp_absolute_max );
1606 float qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1607 float step_size = 0.5f;
1608 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1609 float slice_size_planned = h->param.b_sliced_threads ? rc->slice_size_planned : rc->frame_size_planned;
1610 float max_frame_error = X264_MAX( 0.05f, 1.0f / h->mb.i_mb_height );
1611 float size_of_other_slices = 0;
1612 if( h->param.b_sliced_threads )
1614 float size_of_other_slices_planned = 0;
1615 for( int i = 0; i < h->param.i_threads; i++ )
1616 if( h != h->thread[i] )
1618 size_of_other_slices += h->thread[i]->rc->frame_size_estimated;
1619 size_of_other_slices_planned += h->thread[i]->rc->slice_size_planned;
1621 float weight = rc->slice_size_planned / rc->frame_size_planned;
1622 size_of_other_slices = (size_of_other_slices - size_of_other_slices_planned) * weight + size_of_other_slices_planned;
1624 if( y < h->i_threadslice_end-1 )
1626 /* B-frames shouldn't use lower QP than their reference frames. */
1627 if( h->sh.i_type == SLICE_TYPE_B )
1629 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] ) );
1630 rc->qpm = X264_MAX( rc->qpm, qp_min );
1633 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1634 float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
1635 float b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1637 /* Don't increase the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1638 /* area at the top of the frame was measured inaccurately. */
1639 if( row_bits_so_far( h, y ) < 0.05f * slice_size_planned )
1640 qp_max = qp_absolute_max = prev_row_qp;
1642 if( h->sh.i_type != SLICE_TYPE_I )
1645 if( !rc->b_vbv_min_rate )
1646 qp_min = X264_MAX( qp_min, rc->qp_novbv );
1648 while( rc->qpm < qp_max
1649 && ((b1 > rc->frame_size_planned + rc_tol) ||
1650 (rc->buffer_fill - b1 < buffer_left_planned * 0.5f) ||
1651 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1653 rc->qpm += step_size;
1654 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1657 while( rc->qpm > qp_min
1658 && (rc->qpm > h->fdec->f_row_qp[0] || rc->single_frame_vbv)
1659 && ((b1 < rc->frame_size_planned * 0.8f && rc->qpm <= prev_row_qp)
1660 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1f) )
1662 rc->qpm -= step_size;
1663 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1666 /* avoid VBV underflow or MinCR violation */
1667 while( (rc->qpm < qp_absolute_max)
1668 && ((rc->buffer_fill - b1 < rc->buffer_rate * max_frame_error) ||
1669 (rc->frame_size_maximum - b1 < rc->frame_size_maximum * max_frame_error)))
1671 rc->qpm += step_size;
1672 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1675 h->rc->frame_size_estimated = b1 - size_of_other_slices;
1677 /* If the current row was large enough to cause a large QP jump, try re-encoding it. */
1678 if( rc->qpm > qp_max && prev_row_qp < qp_max && can_reencode_row )
1680 /* Bump QP to halfway in between... close enough. */
1681 rc->qpm = x264_clip3f( (prev_row_qp + rc->qpm)*0.5f, prev_row_qp + 1.0f, qp_max );
1682 rc->qpa_rc = rc->qpa_rc_prev;
1683 rc->qpa_aq = rc->qpa_aq_prev;
1684 h->fdec->i_row_bits[y] = 0;
1685 h->fdec->i_row_bits[y-SLICE_MBAFF] = 0;
1691 h->rc->frame_size_estimated = predict_row_size_sum( h, y, rc->qpm );
1693 /* Last-ditch attempt: if the last row of the frame underflowed the VBV,
1695 if( (h->rc->frame_size_estimated + size_of_other_slices) > (rc->buffer_fill - rc->buffer_rate * max_frame_error) &&
1696 rc->qpm < qp_max && can_reencode_row )
1699 rc->qpa_rc = rc->qpa_rc_prev;
1700 rc->qpa_aq = rc->qpa_aq_prev;
1701 h->fdec->i_row_bits[y] = 0;
1702 h->fdec->i_row_bits[y-SLICE_MBAFF] = 0;
1707 rc->qpa_rc_prev = rc->qpa_rc;
1708 rc->qpa_aq_prev = rc->qpa_aq;
1713 int x264_ratecontrol_qp( x264_t *h )
1716 return x264_clip3( h->rc->qpm + 0.5f, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1719 int x264_ratecontrol_mb_qp( x264_t *h )
1722 float qp = h->rc->qpm;
1723 if( h->param.rc.i_aq_mode )
1725 /* MB-tree currently doesn't adjust quantizers in unreferenced frames. */
1726 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];
1727 /* Scale AQ's effect towards zero in emergency mode. */
1728 if( qp > QP_MAX_SPEC )
1729 qp_offset *= (QP_MAX - qp) / (QP_MAX - QP_MAX_SPEC);
1732 return x264_clip3( qp + 0.5f, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1735 /* In 2pass, force the same frame types as in the 1st pass */
1736 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1738 x264_ratecontrol_t *rc = h->rc;
1739 if( h->param.rc.b_stat_read )
1741 if( frame_num >= rc->num_entries )
1743 /* We could try to initialize everything required for ABR and
1744 * adaptive B-frames, but that would be complicated.
1745 * So just calculate the average QP used so far. */
1746 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24 + QP_BD_OFFSET
1747 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1748 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, QP_MAX );
1749 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 );
1750 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 );
1752 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries );
1753 x264_log( h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant );
1754 if( h->param.i_bframe_adaptive )
1755 x264_log( h, X264_LOG_ERROR, "disabling adaptive B-frames\n" );
1757 for( int i = 0; i < h->param.i_threads; i++ )
1759 h->thread[i]->rc->b_abr = 0;
1760 h->thread[i]->rc->b_2pass = 0;
1761 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1762 h->thread[i]->param.rc.b_stat_read = 0;
1763 h->thread[i]->param.i_bframe_adaptive = 0;
1764 h->thread[i]->param.i_scenecut_threshold = 0;
1765 h->thread[i]->param.rc.b_mb_tree = 0;
1766 if( h->thread[i]->param.i_bframe > 1 )
1767 h->thread[i]->param.i_bframe = 1;
1769 return X264_TYPE_AUTO;
1771 return rc->entry[frame_num].frame_type;
1774 return X264_TYPE_AUTO;
1777 void x264_ratecontrol_set_weights( x264_t *h, x264_frame_t *frm )
1779 ratecontrol_entry_t *rce = &h->rc->entry[frm->i_frame];
1780 if( h->param.analyse.i_weighted_pred <= 0 )
1783 if( rce->i_weight_denom[0] >= 0 )
1784 SET_WEIGHT( frm->weight[0][0], 1, rce->weight[0][0], rce->i_weight_denom[0], rce->weight[0][1] );
1786 if( rce->i_weight_denom[1] >= 0 )
1788 SET_WEIGHT( frm->weight[0][1], 1, rce->weight[1][0], rce->i_weight_denom[1], rce->weight[1][1] );
1789 SET_WEIGHT( frm->weight[0][2], 1, rce->weight[2][0], rce->i_weight_denom[1], rce->weight[2][1] );
1793 /* After encoding one frame, save stats and update ratecontrol state */
1794 int x264_ratecontrol_end( x264_t *h, int bits, int *filler )
1796 x264_ratecontrol_t *rc = h->rc;
1797 const int *mbs = h->stat.frame.i_mb_count;
1801 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1802 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1803 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1804 for( int i = B_DIRECT; i < B_8x8; i++ )
1805 h->stat.frame.i_mb_count_p += mbs[i];
1807 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1808 h->fdec->f_qp_avg_aq = (float)rc->qpa_aq / h->mb.i_mb_count;
1809 h->fdec->f_crf_avg = h->param.rc.f_rf_constant + h->fdec->f_qp_avg_rc - rc->qp_novbv;
1811 if( h->param.rc.b_stat_write )
1813 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1814 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1815 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1816 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1817 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1818 char c_direct = h->mb.b_direct_auto_write ?
1819 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1820 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1822 if( fprintf( rc->p_stat_file_out,
1823 "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:",
1824 h->fenc->i_frame, h->i_frame,
1825 c_type, h->fenc->i_duration,
1826 h->fenc->i_cpb_duration,
1827 rc->qpa_rc, h->fdec->f_qp_avg_aq,
1828 h->stat.frame.i_tex_bits,
1829 h->stat.frame.i_mv_bits,
1830 h->stat.frame.i_misc_bits,
1831 h->stat.frame.i_mb_count_i,
1832 h->stat.frame.i_mb_count_p,
1833 h->stat.frame.i_mb_count_skip,
1837 /* Only write information for reference reordering once. */
1838 int use_old_stats = h->param.rc.b_stat_read && rc->rce->refs > 1;
1839 for( int i = 0; i < (use_old_stats ? rc->rce->refs : h->i_ref[0]); i++ )
1841 int refcount = use_old_stats ? rc->rce->refcount[i]
1842 : PARAM_INTERLACED ? h->stat.frame.i_mb_count_ref[0][i*2]
1843 + h->stat.frame.i_mb_count_ref[0][i*2+1]
1844 : h->stat.frame.i_mb_count_ref[0][i];
1845 if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1849 if( h->param.analyse.i_weighted_pred >= X264_WEIGHTP_SIMPLE && h->sh.weight[0][0].weightfn )
1851 if( fprintf( rc->p_stat_file_out, "w:%d,%d,%d",
1852 h->sh.weight[0][0].i_denom, h->sh.weight[0][0].i_scale, h->sh.weight[0][0].i_offset ) < 0 )
1854 if( h->sh.weight[0][1].weightfn || h->sh.weight[0][2].weightfn )
1856 if( fprintf( rc->p_stat_file_out, ",%d,%d,%d,%d,%d ",
1857 h->sh.weight[0][1].i_denom, h->sh.weight[0][1].i_scale, h->sh.weight[0][1].i_offset,
1858 h->sh.weight[0][2].i_scale, h->sh.weight[0][2].i_offset ) < 0 )
1861 else if( fprintf( rc->p_stat_file_out, " " ) < 0 )
1865 if( fprintf( rc->p_stat_file_out, ";\n") < 0 )
1868 /* Don't re-write the data in multi-pass mode. */
1869 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1871 uint8_t i_type = h->sh.i_type;
1872 /* Values are stored as big-endian FIX8.8 */
1873 for( int i = 0; i < h->mb.i_mb_count; i++ )
1874 rc->mbtree.qp_buffer[0][i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1875 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1877 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 )
1884 if( h->sh.i_type != SLICE_TYPE_B )
1885 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / rc->last_rceq;
1888 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1889 * Not perfectly accurate with B-refs, but good enough. */
1890 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / (rc->last_rceq * fabs( h->param.rc.f_pb_factor ));
1892 rc->cplxr_sum *= rc->cbr_decay;
1893 rc->wanted_bits_window += h->fenc->f_duration * rc->bitrate;
1894 rc->wanted_bits_window *= rc->cbr_decay;
1898 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale( rc->rce->new_qp ) );
1900 if( h->mb.b_variable_qp )
1902 if( h->sh.i_type == SLICE_TYPE_B )
1904 rc->bframe_bits += bits;
1905 if( h->fenc->b_last_minigop_bframe )
1907 update_predictor( rc->pred_b_from_p, qp2qscale( rc->qpa_rc ),
1908 h->fref[1][h->i_ref[1]-1]->i_satd, rc->bframe_bits / rc->bframes );
1909 rc->bframe_bits = 0;
1914 *filler = update_vbv( h, bits );
1915 rc->filler_bits_sum += *filler * 8;
1917 if( h->sps->vui.b_nal_hrd_parameters_present )
1919 if( h->fenc->i_frame == 0 )
1921 // access unit initialises the HRD
1922 h->fenc->hrd_timing.cpb_initial_arrival_time = 0;
1923 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1924 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1925 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit = (double)rc->initial_cpb_removal_delay / 90000;
1929 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit + (double)(h->fenc->i_cpb_delay - h->i_cpb_delay_pir_offset) *
1930 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1932 if( h->fenc->b_keyframe )
1934 rc->nrt_first_access_unit = h->fenc->hrd_timing.cpb_removal_time;
1935 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1936 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1939 double cpb_earliest_arrival_time = h->fenc->hrd_timing.cpb_removal_time - (double)rc->initial_cpb_removal_delay / 90000;
1940 if( !h->fenc->b_keyframe )
1941 cpb_earliest_arrival_time -= (double)rc->initial_cpb_removal_delay_offset / 90000;
1943 if( h->sps->vui.hrd.b_cbr_hrd )
1944 h->fenc->hrd_timing.cpb_initial_arrival_time = rc->previous_cpb_final_arrival_time;
1946 h->fenc->hrd_timing.cpb_initial_arrival_time = X264_MAX( rc->previous_cpb_final_arrival_time, cpb_earliest_arrival_time );
1948 int filler_bits = *filler ? X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), *filler )*8 : 0;
1950 h->fenc->hrd_timing.cpb_final_arrival_time = rc->previous_cpb_final_arrival_time = h->fenc->hrd_timing.cpb_initial_arrival_time +
1951 (double)(bits + filler_bits) / h->sps->vui.hrd.i_bit_rate_unscaled;
1953 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 +
1954 h->fenc->hrd_timing.cpb_removal_time;
1959 x264_log( h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n" );
1963 /****************************************************************************
1965 ***************************************************************************/
1968 * modify the bitrate curve from pass1 for one frame
1970 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1972 x264_ratecontrol_t *rcc= h->rc;
1973 x264_zone_t *zone = get_zone( h, frame_num );
1975 if( h->param.rc.b_mb_tree )
1977 double timescale = (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1978 q = pow( BASE_FRAME_DURATION / CLIP_DURATION(rce->i_duration * timescale), 1 - h->param.rc.f_qcompress );
1981 q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1983 // avoid NaN's in the rc_eq
1984 if( !isfinite(q) || rce->tex_bits + rce->mv_bits == 0 )
1985 q = rcc->last_qscale_for[rce->pict_type];
1990 rcc->last_qscale = q;
1995 if( zone->b_force_qp )
1996 q = qp2qscale( zone->i_qp );
1998 q /= zone->f_bitrate_factor;
2004 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q, int frame_num)
2006 x264_ratecontrol_t *rcc = h->rc;
2007 const int pict_type = rce->pict_type;
2008 x264_zone_t *zone = get_zone( h, frame_num );
2010 // force I/B quants as a function of P quants
2011 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
2012 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
2013 if( pict_type == SLICE_TYPE_I )
2016 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2017 double ip_factor = fabs( h->param.rc.f_ip_factor );
2018 /* don't apply ip_factor if the following frame is also I */
2019 if( rcc->accum_p_norm <= 0 )
2021 else if( h->param.rc.f_ip_factor < 0 )
2023 else if( rcc->accum_p_norm >= 1 )
2026 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
2028 else if( pict_type == SLICE_TYPE_B )
2030 if( h->param.rc.f_pb_factor > 0 )
2032 if( !rce->kept_as_ref )
2033 q *= fabs( h->param.rc.f_pb_factor );
2035 else if( pict_type == SLICE_TYPE_P
2036 && rcc->last_non_b_pict_type == SLICE_TYPE_P
2037 && rce->tex_bits == 0 )
2042 /* last qscale / qdiff stuff */
2043 if( rcc->last_non_b_pict_type == pict_type &&
2044 (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1) )
2046 double last_q = rcc->last_qscale_for[pict_type];
2047 double max_qscale = last_q * rcc->lstep;
2048 double min_qscale = last_q / rcc->lstep;
2050 if ( q > max_qscale ) q = max_qscale;
2051 else if( q < min_qscale ) q = min_qscale;
2054 rcc->last_qscale_for[pict_type] = q;
2055 if( pict_type != SLICE_TYPE_B )
2056 rcc->last_non_b_pict_type = pict_type;
2057 if( pict_type == SLICE_TYPE_I )
2059 rcc->last_accum_p_norm = rcc->accum_p_norm;
2060 rcc->accum_p_norm = 0;
2061 rcc->accum_p_qp = 0;
2063 if( pict_type == SLICE_TYPE_P )
2065 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
2066 rcc->accum_p_qp = mask * (qscale2qp( q ) + rcc->accum_p_qp);
2067 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
2072 if( zone->b_force_qp )
2073 q = qp2qscale( zone->i_qp );
2075 q /= zone->f_bitrate_factor;
2081 static float predict_size( predictor_t *p, float q, float var )
2083 return (p->coeff*var + p->offset) / (q*p->count);
2086 static void update_predictor( predictor_t *p, float q, float var, float bits )
2091 float old_coeff = p->coeff / p->count;
2092 float new_coeff = X264_MAX( bits*q / var, p->coeff_min );
2093 float new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
2094 float new_offset = bits*q - new_coeff_clipped * var;
2095 if( new_offset >= 0 )
2096 new_coeff = new_coeff_clipped;
2099 p->count *= p->decay;
2100 p->coeff *= p->decay;
2101 p->offset *= p->decay;
2103 p->coeff += new_coeff;
2104 p->offset += new_offset;
2107 // update VBV after encoding a frame
2108 static int update_vbv( x264_t *h, int bits )
2111 int bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
2112 x264_ratecontrol_t *rcc = h->rc;
2113 x264_ratecontrol_t *rct = h->thread[0]->rc;
2114 int64_t buffer_size = (int64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
2116 if( rcc->last_satd >= h->mb.i_mb_count )
2117 update_predictor( &rct->pred[h->sh.i_type], qp2qscale( rcc->qpa_rc ), rcc->last_satd, bits );
2122 uint64_t buffer_diff = (uint64_t)bits * h->sps->vui.i_time_scale;
2123 rct->buffer_fill_final -= buffer_diff;
2124 rct->buffer_fill_final_min -= buffer_diff;
2126 if( rct->buffer_fill_final_min < 0 )
2128 double underflow = (double)rct->buffer_fill_final_min / h->sps->vui.i_time_scale;
2129 if( rcc->rate_factor_max_increment && rcc->qpm >= rcc->qp_novbv + rcc->rate_factor_max_increment )
2130 x264_log( h, X264_LOG_DEBUG, "VBV underflow due to CRF-max (frame %d, %.0f bits)\n", h->i_frame, underflow );
2132 x264_log( h, X264_LOG_WARNING, "VBV underflow (frame %d, %.0f bits)\n", h->i_frame, underflow );
2133 rct->buffer_fill_final =
2134 rct->buffer_fill_final_min = 0;
2137 if( h->param.i_avcintra_class )
2138 buffer_diff = buffer_size;
2140 buffer_diff = (uint64_t)bitrate * h->sps->vui.i_num_units_in_tick * h->fenc->i_cpb_duration;
2141 rct->buffer_fill_final += buffer_diff;
2142 rct->buffer_fill_final_min += buffer_diff;
2144 if( rct->buffer_fill_final > buffer_size )
2146 if( h->param.rc.b_filler )
2148 int64_t scale = (int64_t)h->sps->vui.i_time_scale * 8;
2149 filler = (rct->buffer_fill_final - buffer_size + scale - 1) / scale;
2150 bits = h->param.i_avcintra_class ? filler * 8 : X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), filler ) * 8;
2151 buffer_diff = (uint64_t)bits * h->sps->vui.i_time_scale;
2152 rct->buffer_fill_final -= buffer_diff;
2153 rct->buffer_fill_final_min -= buffer_diff;
2157 rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, buffer_size );
2158 rct->buffer_fill_final_min = X264_MIN( rct->buffer_fill_final_min, buffer_size );
2165 void x264_hrd_fullness( x264_t *h )
2167 x264_ratecontrol_t *rct = h->thread[0]->rc;
2168 uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale / rct->hrd_multiply_denom;
2169 uint64_t cpb_state = rct->buffer_fill_final;
2170 uint64_t cpb_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
2171 uint64_t multiply_factor = 90000 / rct->hrd_multiply_denom;
2173 if( rct->buffer_fill_final < 0 || rct->buffer_fill_final > (int64_t)cpb_size )
2175 x264_log( h, X264_LOG_WARNING, "CPB %s: %.0f bits in a %.0f-bit buffer\n",
2176 rct->buffer_fill_final < 0 ? "underflow" : "overflow",
2177 (double)rct->buffer_fill_final / h->sps->vui.i_time_scale, (double)cpb_size / h->sps->vui.i_time_scale );
2180 h->initial_cpb_removal_delay = (multiply_factor * cpb_state) / denom;
2181 h->initial_cpb_removal_delay_offset = (multiply_factor * cpb_size) / denom - h->initial_cpb_removal_delay;
2183 int64_t decoder_buffer_fill = h->initial_cpb_removal_delay * denom / multiply_factor;
2184 rct->buffer_fill_final_min = X264_MIN( rct->buffer_fill_final_min, decoder_buffer_fill );
2187 // provisionally update VBV according to the planned size of all frames currently in progress
2188 static void update_vbv_plan( x264_t *h, int overhead )
2190 x264_ratecontrol_t *rcc = h->rc;
2191 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final_min / h->sps->vui.i_time_scale;
2192 if( h->i_thread_frames > 1 )
2194 int j = h->rc - h->thread[0]->rc;
2195 for( int i = 1; i < h->i_thread_frames; i++ )
2197 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
2198 double bits = t->rc->frame_size_planned;
2199 if( !t->b_thread_active )
2201 bits = X264_MAX(bits, t->rc->frame_size_estimated);
2202 rcc->buffer_fill -= bits;
2203 rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
2204 rcc->buffer_fill += t->rc->buffer_rate;
2205 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
2208 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
2209 rcc->buffer_fill -= overhead;
2212 // apply VBV constraints and clip qscale to between lmin and lmax
2213 static double clip_qscale( x264_t *h, int pict_type, double q )
2215 x264_ratecontrol_t *rcc = h->rc;
2216 double lmin = rcc->lmin[pict_type];
2217 double lmax = rcc->lmax[pict_type];
2218 if( rcc->rate_factor_max_increment )
2219 lmax = X264_MIN( lmax, qp2qscale( rcc->qp_novbv + rcc->rate_factor_max_increment ) );
2222 /* B-frames are not directly subject to VBV,
2223 * since they are controlled by the P-frames' QPs. */
2225 if( rcc->b_vbv && rcc->last_satd > 0 )
2227 double fenc_cpb_duration = (double)h->fenc->i_cpb_duration *
2228 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2229 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
2230 * the lookahead overflow and such that the buffer is in a reasonable state
2231 * by the end of the lookahead. */
2232 if( h->param.rc.i_lookahead )
2236 /* Avoid an infinite loop. */
2237 for( int iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
2240 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2241 double buffer_fill_cur = rcc->buffer_fill - cur_bits;
2243 double total_duration = 0;
2244 double last_duration = fenc_cpb_duration;
2245 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
2246 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
2247 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
2249 /* Loop over the planned future frames. */
2250 for( int j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
2252 total_duration += last_duration;
2253 buffer_fill_cur += rcc->vbv_max_rate * last_duration;
2254 int i_type = h->fenc->i_planned_type[j];
2255 int i_satd = h->fenc->i_planned_satd[j];
2256 if( i_type == X264_TYPE_AUTO )
2258 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
2259 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
2260 buffer_fill_cur -= cur_bits;
2261 last_duration = h->fenc->f_planned_cpb_duration[j];
2263 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
2264 target_fill = X264_MIN( rcc->buffer_fill + total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.5 );
2265 if( buffer_fill_cur < target_fill )
2271 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
2272 target_fill = x264_clip3f( rcc->buffer_fill - total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
2273 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
2282 /* Fallback to old purely-reactive algorithm: no lookahead. */
2285 if( ( pict_type == SLICE_TYPE_P ||
2286 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
2287 rcc->buffer_fill/rcc->buffer_size < 0.5 )
2289 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
2292 /* Now a hard threshold to make sure the frame fits in VBV.
2293 * This one is mostly for I-frames. */
2294 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2295 /* For small VBVs, allow the frame to use up the entire VBV. */
2296 double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
2297 /* For single-frame VBVs, request that the frame use up the entire VBV. */
2298 double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
2300 if( bits > rcc->buffer_fill/max_fill_factor )
2302 double qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
2306 if( bits < rcc->buffer_rate/min_fill_factor )
2308 double qf = x264_clip3f( bits*min_fill_factor/rcc->buffer_rate, 0.001, 1.0 );
2311 q = X264_MAX( q0, q );
2314 /* Check B-frame complexity, and use up any bits that would
2315 * overflow before the next P-frame. */
2316 if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
2318 int nb = rcc->bframes;
2319 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2320 double pbbits = bits;
2321 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
2323 double bframe_cpb_duration = 0;
2324 double minigop_cpb_duration;
2325 for( int i = 0; i < nb; i++ )
2326 bframe_cpb_duration += h->fenc->f_planned_cpb_duration[i];
2328 if( bbits * nb > bframe_cpb_duration * rcc->vbv_max_rate )
2330 pbbits += nb * bbits;
2332 minigop_cpb_duration = bframe_cpb_duration + fenc_cpb_duration;
2333 space = rcc->buffer_fill + minigop_cpb_duration*rcc->vbv_max_rate - rcc->buffer_size;
2334 if( pbbits < space )
2336 q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
2338 q = X264_MAX( q0/2, q );
2341 /* Apply MinCR and buffer fill restrictions */
2342 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2343 double frame_size_maximum = X264_MIN( rcc->frame_size_maximum, X264_MAX( rcc->buffer_fill, 0.001 ) );
2344 if( bits > frame_size_maximum )
2345 q *= bits / frame_size_maximum;
2347 if( !rcc->b_vbv_min_rate )
2348 q = X264_MAX( q0, q );
2353 else if( rcc->b_2pass )
2355 double min2 = log( lmin );
2356 double max2 = log( lmax );
2357 q = (log(q) - min2)/(max2-min2) - 0.5;
2358 q = 1.0/(1.0 + exp( -4*q ));
2359 q = q*(max2-min2) + min2;
2363 return x264_clip3f( q, lmin, lmax );
2366 // update qscale for 1 frame based on actual bits used so far
2367 static float rate_estimate_qscale( x264_t *h )
2370 x264_ratecontrol_t *rcc = h->rc;
2371 ratecontrol_entry_t rce = {0};
2372 int pict_type = h->sh.i_type;
2373 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
2374 + h->stat.i_frame_size[SLICE_TYPE_P]
2375 + h->stat.i_frame_size[SLICE_TYPE_B])
2376 - rcc->filler_bits_sum;
2381 if( pict_type != rce.pict_type )
2383 x264_log( h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
2384 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type] );
2388 if( pict_type == SLICE_TYPE_B )
2390 /* B-frames don't have independent ratecontrol, but rather get the
2391 * average QP of the two adjacent P-frames + an offset */
2393 int i0 = IS_X264_TYPE_I(h->fref_nearest[0]->i_type);
2394 int i1 = IS_X264_TYPE_I(h->fref_nearest[1]->i_type);
2395 int dt0 = abs(h->fenc->i_poc - h->fref_nearest[0]->i_poc);
2396 int dt1 = abs(h->fenc->i_poc - h->fref_nearest[1]->i_poc);
2397 float q0 = h->fref_nearest[0]->f_qp_avg_rc;
2398 float q1 = h->fref_nearest[1]->f_qp_avg_rc;
2400 if( h->fref_nearest[0]->i_type == X264_TYPE_BREF )
2401 q0 -= rcc->pb_offset/2;
2402 if( h->fref_nearest[1]->i_type == X264_TYPE_BREF )
2403 q1 -= rcc->pb_offset/2;
2406 q = (q0 + q1) / 2 + rcc->ip_offset;
2412 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
2414 if( h->fenc->b_kept_as_ref )
2415 q += rcc->pb_offset/2;
2417 q += rcc->pb_offset;
2419 if( rcc->b_2pass && rcc->b_vbv )
2420 rcc->frame_size_planned = qscale2bits( &rce, qp2qscale( q ) );
2422 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, qp2qscale( q ), h->fref[1][h->i_ref[1]-1]->i_satd );
2423 /* Limit planned size by MinCR */
2425 rcc->frame_size_planned = X264_MIN( rcc->frame_size_planned, rcc->frame_size_maximum );
2426 h->rc->frame_size_estimated = rcc->frame_size_planned;
2430 rcc->last_satd = x264_rc_analyse_slice( h );
2432 return qp2qscale( q );
2436 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
2440 double lmin = rcc->lmin[pict_type];
2441 double lmax = rcc->lmax[pict_type];
2443 int64_t predicted_bits = total_bits;
2447 if( h->i_thread_frames > 1 )
2449 int j = h->rc - h->thread[0]->rc;
2450 for( int i = 1; i < h->i_thread_frames; i++ )
2452 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
2453 double bits = t->rc->frame_size_planned;
2454 if( !t->b_thread_active )
2456 bits = X264_MAX(bits, t->rc->frame_size_estimated);
2457 predicted_bits += (int64_t)bits;
2463 if( h->i_frame < h->i_thread_frames )
2464 predicted_bits += (int64_t)h->i_frame * rcc->bitrate / rcc->fps;
2466 predicted_bits += (int64_t)(h->i_thread_frames - 1) * rcc->bitrate / rcc->fps;
2469 /* Adjust ABR buffer based on distance to the end of the video. */
2470 if( rcc->num_entries > h->i_frame )
2472 double final_bits = rcc->entry[rcc->num_entries-1].expected_bits;
2473 double video_pos = rce.expected_bits / final_bits;
2474 double scale_factor = sqrt( (1 - video_pos) * rcc->num_entries );
2475 abr_buffer *= 0.5 * X264_MAX( scale_factor, 0.5 );
2478 diff = predicted_bits - (int64_t)rce.expected_bits;
2480 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
2481 if( ((h->i_frame + 1 - h->i_thread_frames) >= rcc->fps) &&
2482 (rcc->expected_bits_sum > 0))
2484 /* Adjust quant based on the difference between
2485 * achieved and expected bitrate so far */
2486 double cur_time = (double)h->i_frame / rcc->num_entries;
2487 double w = x264_clip3f( cur_time*100, 0.0, 1.0 );
2488 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
2490 rcc->qp_novbv = qscale2qp( q );
2493 /* Do not overflow vbv */
2494 double expected_size = qscale2bits( &rce, q );
2495 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2496 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
2497 double qmax = q*(2 - expected_fullness);
2498 double size_constraint = 1 + expected_fullness;
2499 qmax = X264_MAX( qmax, rce.new_qscale );
2500 if( expected_fullness < .05 )
2502 qmax = X264_MIN(qmax, lmax);
2503 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
2504 ((expected_vbv < 0) && (q < lmax)))
2507 expected_size = qscale2bits(&rce, q);
2508 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2510 rcc->last_satd = x264_rc_analyse_slice( h );
2512 q = x264_clip3f( q, lmin, lmax );
2514 else /* 1pass ABR */
2516 /* Calculate the quantizer which would have produced the desired
2517 * average bitrate if it had been applied to all frames so far.
2518 * Then modulate that quant based on the current frame's complexity
2519 * relative to the average complexity so far (using the 2pass RCEQ).
2520 * Then bias the quant up or down if total size so far was far from
2522 * Result: Depending on the value of rate_tolerance, there is a
2523 * tradeoff between quality and bitrate precision. But at large
2524 * tolerances, the bit distribution approaches that of 2pass. */
2526 double wanted_bits, overflow = 1;
2528 rcc->last_satd = x264_rc_analyse_slice( h );
2529 rcc->short_term_cplxsum *= 0.5;
2530 rcc->short_term_cplxcount *= 0.5;
2531 rcc->short_term_cplxsum += rcc->last_satd / (CLIP_DURATION(h->fenc->f_duration) / BASE_FRAME_DURATION);
2532 rcc->short_term_cplxcount ++;
2534 rce.tex_bits = rcc->last_satd;
2535 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
2537 rce.p_count = rcc->nmb;
2541 rce.pict_type = pict_type;
2542 rce.i_duration = h->fenc->i_duration;
2544 if( h->param.rc.i_rc_method == X264_RC_CRF )
2546 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
2550 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
2552 /* ABR code can potentially be counterproductive in CBR, so just don't bother.
2553 * Don't run it if the frame complexity is zero either. */
2554 if( !rcc->b_vbv_min_rate && rcc->last_satd )
2556 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
2557 int i_frame_done = h->i_frame + 1 - h->i_thread_frames;
2558 double time_done = i_frame_done / rcc->fps;
2559 if( h->param.b_vfr_input && i_frame_done > 0 )
2560 time_done = ((double)(h->fenc->i_reordered_pts - h->i_reordered_pts_delay)) * h->param.i_timebase_num / h->param.i_timebase_den;
2561 wanted_bits = time_done * rcc->bitrate;
2562 if( wanted_bits > 0 )
2564 abr_buffer *= X264_MAX( 1, sqrt( time_done ) );
2565 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
2571 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
2572 /* should test _next_ pict type, but that isn't decided yet */
2573 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
2575 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2576 q /= fabs( h->param.rc.f_ip_factor );
2578 else if( h->i_frame > 0 )
2580 if( h->param.rc.i_rc_method != X264_RC_CRF )
2582 /* Asymmetric clipping, because symmetric would prevent
2583 * overflow control in areas of rapidly oscillating complexity */
2584 double lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
2585 double lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
2586 if( overflow > 1.1 && h->i_frame > 3 )
2588 else if( overflow < 0.9 )
2591 q = x264_clip3f(q, lmin, lmax);
2594 else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
2596 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
2598 rcc->qp_novbv = qscale2qp( q );
2600 //FIXME use get_diff_limited_q() ?
2601 q = clip_qscale( h, pict_type, q );
2604 rcc->last_qscale_for[pict_type] =
2605 rcc->last_qscale = q;
2607 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
2608 rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );
2610 if( rcc->b_2pass && rcc->b_vbv )
2611 rcc->frame_size_planned = qscale2bits(&rce, q);
2613 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2615 /* Always use up the whole VBV in this case. */
2616 if( rcc->single_frame_vbv )
2617 rcc->frame_size_planned = rcc->buffer_rate;
2618 /* Limit planned size by MinCR */
2620 rcc->frame_size_planned = X264_MIN( rcc->frame_size_planned, rcc->frame_size_maximum );
2621 h->rc->frame_size_estimated = rcc->frame_size_planned;
2626 static void x264_threads_normalize_predictors( x264_t *h )
2628 double totalsize = 0;
2629 for( int i = 0; i < h->param.i_threads; i++ )
2630 totalsize += h->thread[i]->rc->slice_size_planned;
2631 double factor = h->rc->frame_size_planned / totalsize;
2632 for( int i = 0; i < h->param.i_threads; i++ )
2633 h->thread[i]->rc->slice_size_planned *= factor;
2636 void x264_threads_distribute_ratecontrol( x264_t *h )
2639 x264_ratecontrol_t *rc = h->rc;
2641 float qscale = qp2qscale( rc->qpm );
2643 /* Initialize row predictors */
2644 if( h->i_frame == 0 )
2645 for( int i = 0; i < h->param.i_threads; i++ )
2647 x264_t *t = h->thread[i];
2649 memcpy( t->rc->row_preds, rc->row_preds, sizeof(rc->row_preds) );
2652 for( int i = 0; i < h->param.i_threads; i++ )
2654 x264_t *t = h->thread[i];
2656 memcpy( t->rc, rc, offsetof(x264_ratecontrol_t, row_pred) );
2657 t->rc->row_pred = t->rc->row_preds[h->sh.i_type];
2658 /* Calculate the planned slice size. */
2659 if( rc->b_vbv && rc->frame_size_planned )
2662 for( row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2663 size += h->fdec->i_row_satd[row];
2664 t->rc->slice_size_planned = predict_size( &rc->pred[h->sh.i_type + (i+1)*5], qscale, size );
2667 t->rc->slice_size_planned = 0;
2669 if( rc->b_vbv && rc->frame_size_planned )
2671 x264_threads_normalize_predictors( h );
2673 if( rc->single_frame_vbv )
2675 /* Compensate for our max frame error threshold: give more bits (proportionally) to smaller slices. */
2676 for( int i = 0; i < h->param.i_threads; i++ )
2678 x264_t *t = h->thread[i];
2679 float max_frame_error = X264_MAX( 0.05, 1.0 / (t->i_threadslice_end - t->i_threadslice_start) );
2680 t->rc->slice_size_planned += 2 * max_frame_error * rc->frame_size_planned;
2682 x264_threads_normalize_predictors( h );
2685 for( int i = 0; i < h->param.i_threads; i++ )
2686 h->thread[i]->rc->frame_size_estimated = h->thread[i]->rc->slice_size_planned;
2690 void x264_threads_merge_ratecontrol( x264_t *h )
2692 x264_ratecontrol_t *rc = h->rc;
2695 for( int i = 0; i < h->param.i_threads; i++ )
2697 x264_t *t = h->thread[i];
2698 x264_ratecontrol_t *rct = h->thread[i]->rc;
2699 if( h->param.rc.i_vbv_buffer_size )
2702 for( int row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2703 size += h->fdec->i_row_satd[row];
2704 int bits = t->stat.frame.i_mv_bits + t->stat.frame.i_tex_bits + t->stat.frame.i_misc_bits;
2705 int mb_count = (t->i_threadslice_end - t->i_threadslice_start) * h->mb.i_mb_width;
2706 update_predictor( &rc->pred[h->sh.i_type+(i+1)*5], qp2qscale( rct->qpa_rc/mb_count ), size, bits );
2710 rc->qpa_rc += rct->qpa_rc;
2711 rc->qpa_aq += rct->qpa_aq;
2715 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
2719 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
2720 /* these vars are updated in x264_ratecontrol_start()
2721 * so copy them from the context that most recently started (prev)
2722 * to the context that's about to start (cur). */
2727 COPY(last_qscale_for);
2728 COPY(last_non_b_pict_type);
2729 COPY(short_term_cplxsum);
2730 COPY(short_term_cplxcount);
2733 COPY(mbtree.qpbuf_pos);
2734 /* these vars can be updated by x264_ratecontrol_init_reconfigurable */
2739 COPY(single_frame_vbv);
2741 COPY(rate_factor_constant);
2742 COPY(rate_factor_max_increment);
2747 #define COPY(var) next->rc->var = cur->rc->var
2748 /* these vars are updated in x264_ratecontrol_end()
2749 * so copy them from the context that most recently ended (cur)
2750 * to the context that's about to end (next) */
2752 COPY(expected_bits_sum);
2753 COPY(filler_bits_sum);
2754 COPY(wanted_bits_window);
2756 COPY(initial_cpb_removal_delay);
2757 COPY(initial_cpb_removal_delay_offset);
2758 COPY(nrt_first_access_unit);
2759 COPY(previous_cpb_final_arrival_time);
2762 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
2763 /* the rest of the variables are either constant or thread-local */
2766 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
2768 /* find an interval ending on an overflow or underflow (depending on whether
2769 * we're adding or removing bits), and starting on the earliest frame that
2770 * can influence the buffer fill of that end frame. */
2771 x264_ratecontrol_t *rcc = h->rc;
2772 const double buffer_min = .1 * rcc->buffer_size;
2773 const double buffer_max = .9 * rcc->buffer_size;
2774 double fill = fills[*t0-1];
2775 double parity = over ? 1. : -1.;
2776 int start = -1, end = -1;
2777 for( int i = *t0; i < rcc->num_entries; i++ )
2779 fill += (rcc->entry[i].i_cpb_duration * rcc->vbv_max_rate * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale -
2780 qscale2bits( &rcc->entry[i], rcc->entry[i].new_qscale )) * parity;
2781 fill = x264_clip3f(fill, 0, rcc->buffer_size);
2783 if( fill <= buffer_min || i == 0 )
2789 else if( fill >= buffer_max && start >= 0 )
2794 return start >= 0 && end >= 0;
2797 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
2799 x264_ratecontrol_t *rcc = h->rc;
2800 double qscale_orig, qscale_new;
2804 for( int i = t0; i <= t1; i++ )
2806 qscale_orig = rcc->entry[i].new_qscale;
2807 qscale_orig = x264_clip3f( qscale_orig, qscale_min, qscale_max );
2808 qscale_new = qscale_orig * adjustment;
2809 qscale_new = x264_clip3f( qscale_new, qscale_min, qscale_max );
2810 rcc->entry[i].new_qscale = qscale_new;
2811 adjusted = adjusted || (qscale_new != qscale_orig);
2816 static double count_expected_bits( x264_t *h )
2818 x264_ratecontrol_t *rcc = h->rc;
2819 double expected_bits = 0;
2820 for( int i = 0; i < rcc->num_entries; i++ )
2822 ratecontrol_entry_t *rce = &rcc->entry[i];
2823 rce->expected_bits = expected_bits;
2824 expected_bits += qscale2bits( rce, rce->new_qscale );
2826 return expected_bits;
2829 static int vbv_pass2( x264_t *h, double all_available_bits )
2831 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
2832 * frames in the interval until either buffer is full at some intermediate frame or the
2833 * last frame in the interval no longer underflows. Recompute intervals and repeat.
2834 * Then do the converse to put bits back into overflow areas until target size is met */
2836 x264_ratecontrol_t *rcc = h->rc;
2838 double expected_bits = 0;
2840 double prev_bits = 0;
2842 double qscale_min = qp2qscale( h->param.rc.i_qp_min );
2843 double qscale_max = qp2qscale( h->param.rc.i_qp_max );
2845 int adj_min, adj_max;
2846 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2850 /* adjust overall stream size */
2854 prev_bits = expected_bits;
2857 { /* not first iteration */
2858 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2859 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2863 while(adj_min && find_underflow( h, fills, &t0, &t1, 1 ))
2865 adj_min = fix_underflow( h, t0, t1, adjustment, qscale_min, qscale_max );
2870 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2872 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2874 while( adj_max && find_underflow( h, fills, &t0, &t1, 0 ) )
2875 adj_max = fix_underflow( h, t0, t1, 1.001, qscale_min, qscale_max );
2877 expected_bits = count_expected_bits( h );
2878 } while( (expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2881 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2883 /* store expected vbv filling values for tracking when encoding */
2884 for( int i = 0; i < rcc->num_entries; i++ )
2885 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
2887 x264_free( fills-1 );
2893 static int init_pass2( x264_t *h )
2895 x264_ratecontrol_t *rcc = h->rc;
2896 uint64_t all_const_bits = 0;
2897 double timescale = (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2898 double duration = 0;
2899 for( int i = 0; i < rcc->num_entries; i++ )
2900 duration += rcc->entry[i].i_duration;
2901 duration *= timescale;
2902 uint64_t all_available_bits = h->param.rc.i_bitrate * 1000. * duration;
2903 double rate_factor, step_mult;
2904 double qblur = h->param.rc.f_qblur;
2905 double cplxblur = h->param.rc.f_complexity_blur;
2906 const int filter_size = (int)(qblur*4) | 1;
2907 double expected_bits;
2908 double *qscale, *blurred_qscale;
2909 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
2911 /* find total/average complexity & const_bits */
2912 for( int i = 0; i < rcc->num_entries; i++ )
2914 ratecontrol_entry_t *rce = &rcc->entry[i];
2915 all_const_bits += rce->misc_bits;
2918 if( all_available_bits < all_const_bits)
2920 x264_log( h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2921 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)) );
2925 /* Blur complexities, to reduce local fluctuation of QP.
2926 * We don't blur the QPs directly, because then one very simple frame
2927 * could drag down the QP of a nearby complex frame and give it more
2928 * bits than intended. */
2929 for( int i = 0; i < rcc->num_entries; i++ )
2931 ratecontrol_entry_t *rce = &rcc->entry[i];
2932 double weight_sum = 0;
2933 double cplx_sum = 0;
2934 double weight = 1.0;
2935 double gaussian_weight;
2936 /* weighted average of cplx of future frames */
2937 for( int j = 1; j < cplxblur*2 && j < rcc->num_entries-i; j++ )
2939 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2940 double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2941 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2942 if( weight < .0001 )
2944 gaussian_weight = weight * exp( -j*j/200.0 );
2945 weight_sum += gaussian_weight;
2946 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2948 /* weighted average of cplx of past frames */
2950 for( int j = 0; j <= cplxblur*2 && j <= i; j++ )
2952 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2953 double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2954 gaussian_weight = weight * exp( -j*j/200.0 );
2955 weight_sum += gaussian_weight;
2956 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2957 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2958 if( weight < .0001 )
2961 rce->blurred_complexity = cplx_sum / weight_sum;
2964 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2965 if( filter_size > 1 )
2966 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2968 blurred_qscale = qscale;
2970 /* Search for a factor which, when multiplied by the RCEQ values from
2971 * each frame, adds up to the desired total size.
2972 * There is no exact closed-form solution because of VBV constraints and
2973 * because qscale2bits is not invertible, but we can start with the simple
2974 * approximation of scaling the 1st pass by the ratio of bitrates.
2975 * The search range is probably overkill, but speed doesn't matter here. */
2978 for( int i = 0; i < rcc->num_entries; i++ )
2980 double q = get_qscale(h, &rcc->entry[i], 1.0, i);
2981 expected_bits += qscale2bits(&rcc->entry[i], q);
2982 rcc->last_qscale_for[rcc->entry[i].pict_type] = q;
2984 step_mult = all_available_bits / expected_bits;
2987 for( double step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2990 rate_factor += step;
2992 rcc->last_non_b_pict_type = -1;
2993 rcc->last_accum_p_norm = 1;
2994 rcc->accum_p_norm = 0;
2996 rcc->last_qscale_for[0] =
2997 rcc->last_qscale_for[1] =
2998 rcc->last_qscale_for[2] = pow( base_cplx, 1 - rcc->qcompress ) / rate_factor;
3001 for( int i = 0; i < rcc->num_entries; i++ )
3003 qscale[i] = get_qscale( h, &rcc->entry[i], rate_factor, -1 );
3004 rcc->last_qscale_for[rcc->entry[i].pict_type] = qscale[i];
3007 /* fixed I/B qscale relative to P */
3008 for( int i = rcc->num_entries-1; i >= 0; i-- )
3010 qscale[i] = get_diff_limited_q( h, &rcc->entry[i], qscale[i], i );
3011 assert(qscale[i] >= 0);
3015 if( filter_size > 1 )
3017 assert( filter_size%2 == 1 );
3018 for( int i = 0; i < rcc->num_entries; i++ )
3020 ratecontrol_entry_t *rce = &rcc->entry[i];
3021 double q = 0.0, sum = 0.0;
3023 for( int j = 0; j < filter_size; j++ )
3025 int idx = i+j-filter_size/2;
3027 double coeff = qblur==0 ? 1.0 : exp( -d*d/(qblur*qblur) );
3028 if( idx < 0 || idx >= rcc->num_entries )
3030 if( rce->pict_type != rcc->entry[idx].pict_type )
3032 q += qscale[idx] * coeff;
3035 blurred_qscale[i] = q/sum;
3039 /* find expected bits */
3040 for( int i = 0; i < rcc->num_entries; i++ )
3042 ratecontrol_entry_t *rce = &rcc->entry[i];
3043 rce->new_qscale = clip_qscale( h, rce->pict_type, blurred_qscale[i] );
3044 assert(rce->new_qscale >= 0);
3045 expected_bits += qscale2bits( rce, rce->new_qscale );
3048 if( expected_bits > all_available_bits )
3049 rate_factor -= step;
3052 x264_free( qscale );
3053 if( filter_size > 1 )
3054 x264_free( blurred_qscale );
3057 if( vbv_pass2( h, all_available_bits ) )
3059 expected_bits = count_expected_bits( h );
3061 if( fabs( expected_bits/all_available_bits - 1.0 ) > 0.01 )
3064 for( int i = 0; i < rcc->num_entries; i++ )
3065 avgq += rcc->entry[i].new_qscale;
3066 avgq = qscale2qp( avgq / rcc->num_entries );
3068 if( expected_bits > all_available_bits || !rcc->b_vbv )
3069 x264_log( h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n" );
3070 x264_log( h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
3071 (float)h->param.rc.i_bitrate,
3072 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
3074 if( expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2 )
3076 if( h->param.rc.i_qp_min > 0 )
3077 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min );
3079 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate\n" );
3081 else if( expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2 )
3083 if( h->param.rc.i_qp_max < QP_MAX )
3084 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max );
3086 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate\n");
3088 else if( !(rcc->b_2pass && rcc->b_vbv) )
3089 x264_log( h, X264_LOG_WARNING, "internal error\n" );