Book a Demo!
CoCalc Logo Icon
StoreFeaturesDocsShareSupportNewsAboutPoliciesSign UpSign In
Download
52866 views
1
/*****************************************************************************
2
* ratecontrol.c: ratecontrol
3
*****************************************************************************
4
* Copyright (C) 2005-2016 x264 project
5
*
6
* Authors: Loren Merritt <[email protected]>
7
* Michael Niedermayer <[email protected]>
8
* Gabriel Bouvigne <[email protected]>
9
* Fiona Glaser <[email protected]>
10
* M�ns Rullg�rd <[email protected]>
11
*
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.
16
*
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.
21
*
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.
25
*
26
* This program is also available under a commercial proprietary license.
27
* For more information, contact us at [email protected].
28
*****************************************************************************/
29
30
#define _ISOC99_SOURCE
31
#undef NDEBUG // always check asserts, the speed effect is far too small to disable them
32
33
#include "common/common.h"
34
#include "ratecontrol.h"
35
#include "me.h"
36
37
typedef struct
38
{
39
int pict_type;
40
int frame_type;
41
int kept_as_ref;
42
double qscale;
43
int mv_bits;
44
int tex_bits;
45
int misc_bits;
46
double expected_bits; /* total expected bits up to the current frame (current one excluded) */
47
double expected_vbv;
48
double new_qscale;
49
float new_qp;
50
int i_count;
51
int p_count;
52
int s_count;
53
float blurred_complexity;
54
char direct_mode;
55
int16_t weight[3][2];
56
int16_t i_weight_denom[2];
57
int refcount[16];
58
int refs;
59
int64_t i_duration;
60
int64_t i_cpb_duration;
61
int out_num;
62
} ratecontrol_entry_t;
63
64
typedef struct
65
{
66
float coeff_min;
67
float coeff;
68
float count;
69
float decay;
70
float offset;
71
} predictor_t;
72
73
struct x264_ratecontrol_t
74
{
75
/* constants */
76
int b_abr;
77
int b_2pass;
78
int b_vbv;
79
int b_vbv_min_rate;
80
double fps;
81
double bitrate;
82
double rate_tolerance;
83
double qcompress;
84
int nmb; /* number of macroblocks in a frame */
85
int qp_constant[3];
86
87
/* current frame */
88
ratecontrol_entry_t *rce;
89
float qpm; /* qp for current macroblock: precise float for AQ */
90
float qpa_rc; /* average of macroblocks' qp before aq */
91
float qpa_rc_prev;
92
int qpa_aq; /* average of macroblocks' qp after aq */
93
int qpa_aq_prev;
94
float qp_novbv; /* QP for the current frame if 1-pass VBV was disabled. */
95
96
/* VBV stuff */
97
double buffer_size;
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). */
106
107
/* ABR stuff */
108
int last_satd;
109
double last_rceq;
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 */
114
double cbr_decay;
115
double short_term_cplxsum;
116
double short_term_cplxcount;
117
double rate_factor_constant;
118
double ip_offset;
119
double pb_offset;
120
121
/* 2pass stuff */
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;
128
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 */
131
ratecontrol_entry_t **entry_out;
132
double last_qscale;
133
double last_qscale_for[3]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
134
int last_non_b_pict_type;
135
double accum_p_qp; /* for determining I-frame quant */
136
double accum_p_norm;
137
double last_accum_p_norm;
138
double lmin[3]; /* min qscale by frame type */
139
double lmax[3];
140
double lstep; /* max change (multiply) in qscale per frame */
141
struct
142
{
143
uint16_t *qp_buffer[2]; /* Global buffers for converting MB-tree quantizer data. */
144
int qpbuf_pos; /* In order to handle pyramid reordering, QP buffer acts as a stack.
145
* This value is the current position (0 or 1). */
146
int src_mb_count;
147
148
/* For rescaling */
149
int rescale_enabled;
150
float *scale_buffer[2]; /* Intermediate buffers */
151
int filtersize[2]; /* filter size (H/V) */
152
float *coeffs[2];
153
int *pos[2];
154
int srcdim[2]; /* Source dimensions (W/H) */
155
} mbtree;
156
157
/* MBRC stuff */
158
float frame_size_estimated; /* Access to this variable must be atomic: double is
159
* not atomic on all arches we care about */
160
double frame_size_maximum; /* Maximum frame size due to MinCR */
161
double frame_size_planned;
162
double slice_size_planned;
163
predictor_t *row_pred;
164
predictor_t row_preds[3][2];
165
predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
166
int bframes; /* # consecutive B-frames before this P-frame */
167
int bframe_bits; /* total cost of those frames */
168
169
int i_zones;
170
x264_zone_t *zones;
171
x264_zone_t *prev_zone;
172
173
/* hrd stuff */
174
int initial_cpb_removal_delay;
175
int initial_cpb_removal_delay_offset;
176
double nrt_first_access_unit; /* nominal removal time */
177
double previous_cpb_final_arrival_time;
178
uint64_t hrd_multiply_denom;
179
};
180
181
182
static int parse_zones( x264_t *h );
183
static int init_pass2(x264_t *);
184
static float rate_estimate_qscale( x264_t *h );
185
static int update_vbv( x264_t *h, int bits );
186
static void update_vbv_plan( x264_t *h, int overhead );
187
static float predict_size( predictor_t *p, float q, float var );
188
static void update_predictor( predictor_t *p, float q, float var, float bits );
189
190
#define CMP_OPT_FIRST_PASS( opt, param_val )\
191
{\
192
if( ( p = strstr( opts, opt "=" ) ) && sscanf( p, opt "=%d" , &i ) && param_val != i )\
193
{\
194
x264_log( h, X264_LOG_ERROR, "different " opt " setting than first pass (%d vs %d)\n", param_val, i );\
195
return -1;\
196
}\
197
}
198
199
/* Terminology:
200
* qp = h.264's quantizer
201
* qscale = linearized quantizer = Lagrange multiplier
202
*/
203
static inline float qp2qscale( float qp )
204
{
205
return 0.85f * powf( 2.0f, ( qp - (12.0f + QP_BD_OFFSET) ) / 6.0f );
206
}
207
static inline float qscale2qp( float qscale )
208
{
209
return (12.0f + QP_BD_OFFSET) + 6.0f * log2f( qscale/0.85f );
210
}
211
212
/* Texture bitrate is not quite inversely proportional to qscale,
213
* probably due the the changing number of SKIP blocks.
214
* MV bits level off at about qp<=12, because the lambda used
215
* for motion estimation is constant there. */
216
static inline double qscale2bits( ratecontrol_entry_t *rce, double qscale )
217
{
218
if( qscale<0.1 )
219
qscale = 0.1;
220
return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
221
+ rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
222
+ rce->misc_bits;
223
}
224
225
static ALWAYS_INLINE uint32_t ac_energy_var( uint64_t sum_ssd, int shift, x264_frame_t *frame, int i, int b_store )
226
{
227
uint32_t sum = sum_ssd;
228
uint32_t ssd = sum_ssd >> 32;
229
if( b_store )
230
{
231
frame->i_pixel_sum[i] += sum;
232
frame->i_pixel_ssd[i] += ssd;
233
}
234
return ssd - ((uint64_t)sum * sum >> shift);
235
}
236
237
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
{
239
int height = b_chroma ? 16>>CHROMA_V_SHIFT : 16;
240
int stride = frame->i_stride[i];
241
int offset = b_field
242
? 16 * mb_x + height * (mb_y&~1) * stride + (mb_y&1) * stride
243
: 16 * mb_x + height * mb_y * stride;
244
stride <<= b_field;
245
if( b_chroma )
246
{
247
ALIGNED_ARRAY_16( pixel, pix,[FENC_STRIDE*16] );
248
int chromapix = h->luma2chroma_pixel[PIXEL_16x16];
249
int shift = 7 - CHROMA_V_SHIFT;
250
251
h->mc.load_deinterleave_chroma_fenc( pix, frame->plane[1] + offset, stride, height );
252
return ac_energy_var( h->pixf.var[chromapix]( pix, FENC_STRIDE ), shift, frame, 1, b_store )
253
+ ac_energy_var( h->pixf.var[chromapix]( pix+FENC_STRIDE/2, FENC_STRIDE ), shift, frame, 2, b_store );
254
}
255
else
256
return ac_energy_var( h->pixf.var[PIXEL_16x16]( frame->plane[i] + offset, stride ), 8, frame, i, b_store );
257
}
258
259
// Find the total AC energy of the block in all planes.
260
static NOINLINE uint32_t x264_ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
261
{
262
/* This function contains annoying hacks because GCC has a habit of reordering emms
263
* and putting it after floating point ops. As a result, we put the emms at the end of the
264
* function and make sure that its always called before the float math. Noinline makes
265
* sure no reordering goes on. */
266
uint32_t var;
267
x264_prefetch_fenc( h, frame, mb_x, mb_y );
268
if( h->mb.b_adaptive_mbaff )
269
{
270
/* We don't know the super-MB mode we're going to pick yet, so
271
* simply try both and pick the lower of the two. */
272
uint32_t var_interlaced, var_progressive;
273
var_interlaced = ac_energy_plane( h, mb_x, mb_y, frame, 0, 0, 1, 1 );
274
var_progressive = ac_energy_plane( h, mb_x, mb_y, frame, 0, 0, 0, 0 );
275
if( CHROMA444 )
276
{
277
var_interlaced += ac_energy_plane( h, mb_x, mb_y, frame, 1, 0, 1, 1 );
278
var_progressive += ac_energy_plane( h, mb_x, mb_y, frame, 1, 0, 0, 0 );
279
var_interlaced += ac_energy_plane( h, mb_x, mb_y, frame, 2, 0, 1, 1 );
280
var_progressive += ac_energy_plane( h, mb_x, mb_y, frame, 2, 0, 0, 0 );
281
}
282
else
283
{
284
var_interlaced += ac_energy_plane( h, mb_x, mb_y, frame, 1, 1, 1, 1 );
285
var_progressive += ac_energy_plane( h, mb_x, mb_y, frame, 1, 1, 0, 0 );
286
}
287
var = X264_MIN( var_interlaced, var_progressive );
288
}
289
else
290
{
291
var = ac_energy_plane( h, mb_x, mb_y, frame, 0, 0, PARAM_INTERLACED, 1 );
292
if( CHROMA444 )
293
{
294
var += ac_energy_plane( h, mb_x, mb_y, frame, 1, 0, PARAM_INTERLACED, 1 );
295
var += ac_energy_plane( h, mb_x, mb_y, frame, 2, 0, PARAM_INTERLACED, 1 );
296
}
297
else
298
var += ac_energy_plane( h, mb_x, mb_y, frame, 1, 1, PARAM_INTERLACED, 1 );
299
}
300
x264_emms();
301
return var;
302
}
303
304
void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame, float *quant_offsets )
305
{
306
/* Initialize frame stats */
307
for( int i = 0; i < 3; i++ )
308
{
309
frame->i_pixel_sum[i] = 0;
310
frame->i_pixel_ssd[i] = 0;
311
}
312
313
/* Degenerate cases */
314
if( h->param.rc.i_aq_mode == X264_AQ_NONE || h->param.rc.f_aq_strength == 0 )
315
{
316
/* Need to init it anyways for MB tree */
317
if( h->param.rc.i_aq_mode && h->param.rc.f_aq_strength == 0 )
318
{
319
if( quant_offsets )
320
{
321
for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
322
frame->f_qp_offset[mb_xy] = frame->f_qp_offset_aq[mb_xy] = quant_offsets[mb_xy];
323
if( h->frames.b_have_lowres )
324
for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
325
frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8( frame->f_qp_offset[mb_xy] );
326
}
327
else
328
{
329
memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
330
memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
331
if( h->frames.b_have_lowres )
332
for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
333
frame->i_inv_qscale_factor[mb_xy] = 256;
334
}
335
}
336
/* Need variance data for weighted prediction */
337
if( h->param.analyse.i_weighted_pred )
338
{
339
for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
340
for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
341
x264_ac_energy_mb( h, mb_x, mb_y, frame );
342
}
343
else
344
return;
345
}
346
/* Actual adaptive quantization */
347
else
348
{
349
/* constants chosen to result in approximately the same overall bitrate as without AQ.
350
* FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
351
float strength;
352
float avg_adj = 0.f;
353
float bias_strength = 0.f;
354
355
if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE || h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE_BIASED )
356
{
357
float bit_depth_correction = 1.f / (1 << (2*(BIT_DEPTH-8)));
358
float avg_adj_pow2 = 0.f;
359
for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
360
for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
361
{
362
uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
363
float qp_adj = powf( energy * bit_depth_correction + 1, 0.125f );
364
frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
365
avg_adj += qp_adj;
366
avg_adj_pow2 += qp_adj * qp_adj;
367
}
368
avg_adj /= h->mb.i_mb_count;
369
avg_adj_pow2 /= h->mb.i_mb_count;
370
strength = h->param.rc.f_aq_strength * avg_adj;
371
avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - 14.f) / avg_adj;
372
bias_strength = h->param.rc.f_aq_strength;
373
}
374
else
375
strength = h->param.rc.f_aq_strength * 1.0397f;
376
377
for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
378
for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
379
{
380
float qp_adj;
381
int mb_xy = mb_x + mb_y*h->mb.i_mb_stride;
382
if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE_BIASED )
383
{
384
qp_adj = frame->f_qp_offset[mb_xy];
385
qp_adj = strength * (qp_adj - avg_adj) + bias_strength * (1.f - 14.f / (qp_adj * qp_adj));
386
}
387
else if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
388
{
389
qp_adj = frame->f_qp_offset[mb_xy];
390
qp_adj = strength * (qp_adj - avg_adj);
391
}
392
else
393
{
394
uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
395
qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - (14.427f + 2*(BIT_DEPTH-8)));
396
}
397
if( quant_offsets )
398
qp_adj += quant_offsets[mb_xy];
399
frame->f_qp_offset[mb_xy] =
400
frame->f_qp_offset_aq[mb_xy] = qp_adj;
401
if( h->frames.b_have_lowres )
402
frame->i_inv_qscale_factor[mb_xy] = x264_exp2fix8(qp_adj);
403
}
404
}
405
406
/* Remove mean from SSD calculation */
407
for( int i = 0; i < 3; i++ )
408
{
409
uint64_t ssd = frame->i_pixel_ssd[i];
410
uint64_t sum = frame->i_pixel_sum[i];
411
int width = 16*h->mb.i_mb_width >> (i && CHROMA_H_SHIFT);
412
int height = 16*h->mb.i_mb_height >> (i && CHROMA_V_SHIFT);
413
frame->i_pixel_ssd[i] = ssd - (sum * sum + width * height / 2) / (width * height);
414
}
415
}
416
417
static int x264_macroblock_tree_rescale_init( x264_t *h, x264_ratecontrol_t *rc )
418
{
419
/* Use fractional QP array dimensions to compensate for edge padding */
420
float srcdim[2] = {rc->mbtree.srcdim[0] / 16.f, rc->mbtree.srcdim[1] / 16.f};
421
float dstdim[2] = { h->param.i_width / 16.f, h->param.i_height / 16.f};
422
int srcdimi[2] = {ceil(srcdim[0]), ceil(srcdim[1])};
423
int dstdimi[2] = {ceil(dstdim[0]), ceil(dstdim[1])};
424
if( PARAM_INTERLACED )
425
{
426
srcdimi[1] = (srcdimi[1]+1)&~1;
427
dstdimi[1] = (dstdimi[1]+1)&~1;
428
}
429
430
rc->mbtree.src_mb_count = srcdimi[0] * srcdimi[1];
431
432
CHECKED_MALLOC( rc->mbtree.qp_buffer[0], rc->mbtree.src_mb_count * sizeof(uint16_t) );
433
if( h->param.i_bframe_pyramid && h->param.rc.b_stat_read )
434
CHECKED_MALLOC( rc->mbtree.qp_buffer[1], rc->mbtree.src_mb_count * sizeof(uint16_t) );
435
rc->mbtree.qpbuf_pos = -1;
436
437
/* No rescaling to do */
438
if( srcdimi[0] == dstdimi[0] && srcdimi[1] == dstdimi[1] )
439
return 0;
440
441
rc->mbtree.rescale_enabled = 1;
442
443
/* Allocate intermediate scaling buffers */
444
CHECKED_MALLOC( rc->mbtree.scale_buffer[0], srcdimi[0] * srcdimi[1] * sizeof(float) );
445
CHECKED_MALLOC( rc->mbtree.scale_buffer[1], dstdimi[0] * srcdimi[1] * sizeof(float) );
446
447
/* Allocate and calculate resize filter parameters and coefficients */
448
for( int i = 0; i < 2; i++ )
449
{
450
if( srcdim[i] > dstdim[i] ) // downscale
451
rc->mbtree.filtersize[i] = 1 + (2 * srcdimi[i] + dstdimi[i] - 1) / dstdimi[i];
452
else // upscale
453
rc->mbtree.filtersize[i] = 3;
454
455
CHECKED_MALLOC( rc->mbtree.coeffs[i], rc->mbtree.filtersize[i] * dstdimi[i] * sizeof(float) );
456
CHECKED_MALLOC( rc->mbtree.pos[i], dstdimi[i] * sizeof(int) );
457
458
/* Initialize filter coefficients */
459
float inc = srcdim[i] / dstdim[i];
460
float dmul = inc > 1.f ? dstdim[i] / srcdim[i] : 1.f;
461
float dstinsrc = 0.5f * inc - 0.5f;
462
int filtersize = rc->mbtree.filtersize[i];
463
for( int j = 0; j < dstdimi[i]; j++ )
464
{
465
int pos = dstinsrc - (filtersize - 2.f) * 0.5f;
466
float sum = 0.0;
467
rc->mbtree.pos[i][j] = pos;
468
for( int k = 0; k < filtersize; k++ )
469
{
470
float d = fabs( pos + k - dstinsrc ) * dmul;
471
float coeff = X264_MAX( 1.f - d, 0 );
472
rc->mbtree.coeffs[i][j * filtersize + k] = coeff;
473
sum += coeff;
474
}
475
sum = 1.0f / sum;
476
for( int k = 0; k < filtersize; k++ )
477
rc->mbtree.coeffs[i][j * filtersize + k] *= sum;
478
dstinsrc += inc;
479
}
480
}
481
482
/* Write back actual qp array dimensions */
483
rc->mbtree.srcdim[0] = srcdimi[0];
484
rc->mbtree.srcdim[1] = srcdimi[1];
485
return 0;
486
fail:
487
return -1;
488
}
489
490
static void x264_macroblock_tree_rescale_destroy( x264_ratecontrol_t *rc )
491
{
492
for( int i = 0; i < 2; i++ )
493
{
494
x264_free( rc->mbtree.qp_buffer[i] );
495
x264_free( rc->mbtree.scale_buffer[i] );
496
x264_free( rc->mbtree.coeffs[i] );
497
x264_free( rc->mbtree.pos[i] );
498
}
499
}
500
501
static ALWAYS_INLINE float tapfilter( float *src, int pos, int max, int stride, float *coeff, int filtersize )
502
{
503
float sum = 0.f;
504
for( int i = 0; i < filtersize; i++, pos++ )
505
sum += src[x264_clip3( pos, 0, max-1 )*stride] * coeff[i];
506
return sum;
507
}
508
509
static void x264_macroblock_tree_rescale( x264_t *h, x264_ratecontrol_t *rc, float *dst )
510
{
511
float *input, *output;
512
int filtersize, stride, height;
513
514
/* H scale first */
515
input = rc->mbtree.scale_buffer[0];
516
output = rc->mbtree.scale_buffer[1];
517
filtersize = rc->mbtree.filtersize[0];
518
stride = rc->mbtree.srcdim[0];
519
height = rc->mbtree.srcdim[1];
520
for( int y = 0; y < height; y++, input += stride, output += h->mb.i_mb_width )
521
{
522
float *coeff = rc->mbtree.coeffs[0];
523
for( int x = 0; x < h->mb.i_mb_width; x++, coeff+=filtersize )
524
output[x] = tapfilter( input, rc->mbtree.pos[0][x], stride, 1, coeff, filtersize );
525
}
526
527
/* V scale next */
528
input = rc->mbtree.scale_buffer[1];
529
output = dst;
530
filtersize = rc->mbtree.filtersize[1];
531
stride = h->mb.i_mb_width;
532
height = rc->mbtree.srcdim[1];
533
for( int x = 0; x < h->mb.i_mb_width; x++, input++, output++ )
534
{
535
float *coeff = rc->mbtree.coeffs[1];
536
for( int y = 0; y < h->mb.i_mb_height; y++, coeff+=filtersize )
537
output[y*stride] = tapfilter( input, rc->mbtree.pos[1][y], height, stride, coeff, filtersize );
538
}
539
}
540
541
int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame, float *quant_offsets )
542
{
543
x264_ratecontrol_t *rc = h->rc;
544
uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
545
546
if( rc->entry[frame->i_frame].kept_as_ref )
547
{
548
uint8_t i_type;
549
if( rc->mbtree.qpbuf_pos < 0 )
550
{
551
do
552
{
553
rc->mbtree.qpbuf_pos++;
554
555
if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
556
goto fail;
557
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 )
558
goto fail;
559
560
if( i_type != i_type_actual && rc->mbtree.qpbuf_pos == 1 )
561
{
562
x264_log( h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type, i_type_actual );
563
return -1;
564
}
565
} while( i_type != i_type_actual );
566
}
567
568
float *dst = rc->mbtree.rescale_enabled ? rc->mbtree.scale_buffer[0] : frame->f_qp_offset;
569
for( int i = 0; i < rc->mbtree.src_mb_count; i++ )
570
{
571
int16_t qp_fix8 = endian_fix16( rc->mbtree.qp_buffer[rc->mbtree.qpbuf_pos][i] );
572
dst[i] = qp_fix8 * (1.f/256.f);
573
}
574
if( rc->mbtree.rescale_enabled )
575
x264_macroblock_tree_rescale( h, rc, frame->f_qp_offset );
576
if( h->frames.b_have_lowres )
577
for( int i = 0; i < h->mb.i_mb_count; i++ )
578
frame->i_inv_qscale_factor[i] = x264_exp2fix8( frame->f_qp_offset[i] );
579
rc->mbtree.qpbuf_pos--;
580
}
581
else
582
x264_stack_align( x264_adaptive_quant_frame, h, frame, quant_offsets );
583
return 0;
584
fail:
585
x264_log( h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n" );
586
return -1;
587
}
588
589
int x264_reference_build_list_optimal( x264_t *h )
590
{
591
ratecontrol_entry_t *rce = h->rc->rce;
592
x264_frame_t *frames[16];
593
x264_weight_t weights[16][3];
594
int refcount[16];
595
596
if( rce->refs != h->i_ref[0] )
597
return -1;
598
599
memcpy( frames, h->fref[0], sizeof(frames) );
600
memcpy( refcount, rce->refcount, sizeof(refcount) );
601
memcpy( weights, h->fenc->weight, sizeof(weights) );
602
memset( &h->fenc->weight[1][0], 0, sizeof(x264_weight_t[15][3]) );
603
604
/* For now don't reorder ref 0; it seems to lower quality
605
in most cases due to skips. */
606
for( int ref = 1; ref < h->i_ref[0]; ref++ )
607
{
608
int max = -1;
609
int bestref = 1;
610
611
for( int i = 1; i < h->i_ref[0]; i++ )
612
/* Favor lower POC as a tiebreaker. */
613
COPY2_IF_GT( max, refcount[i], bestref, i );
614
615
/* FIXME: If there are duplicates from frames other than ref0 then it is possible
616
* that the optimal ordering doesnt place every duplicate. */
617
618
refcount[bestref] = -1;
619
h->fref[0][ref] = frames[bestref];
620
memcpy( h->fenc->weight[ref], weights[bestref], sizeof(weights[bestref]) );
621
}
622
623
return 0;
624
}
625
626
static char *x264_strcat_filename( char *input, char *suffix )
627
{
628
char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
629
if( !output )
630
return NULL;
631
strcpy( output, input );
632
strcat( output, suffix );
633
return output;
634
}
635
636
void x264_ratecontrol_init_reconfigurable( x264_t *h, int b_init )
637
{
638
x264_ratecontrol_t *rc = h->rc;
639
if( !b_init && rc->b_2pass )
640
return;
641
642
if( h->param.rc.i_rc_method == X264_RC_CRF )
643
{
644
/* Arbitrary rescaling to make CRF somewhat similar to QP.
645
* Try to compensate for MB-tree's effects as well. */
646
double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
647
double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
648
rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
649
/ qp2qscale( h->param.rc.f_rf_constant + mbtree_offset + QP_BD_OFFSET );
650
}
651
652
if( h->param.rc.i_vbv_max_bitrate > 0 && h->param.rc.i_vbv_buffer_size > 0 )
653
{
654
/* We don't support changing the ABR bitrate right now,
655
so if the stream starts as CBR, keep it CBR. */
656
if( rc->b_vbv_min_rate )
657
h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
658
659
if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
660
{
661
h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
662
x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
663
h->param.rc.i_vbv_buffer_size );
664
}
665
666
int kilobit_size = h->param.i_avcintra_class ? 1024 : 1000;
667
int vbv_buffer_size = h->param.rc.i_vbv_buffer_size * kilobit_size;
668
int vbv_max_bitrate = h->param.rc.i_vbv_max_bitrate * kilobit_size;
669
670
/* Init HRD */
671
if( h->param.i_nal_hrd && b_init )
672
{
673
h->sps->vui.hrd.i_cpb_cnt = 1;
674
h->sps->vui.hrd.b_cbr_hrd = h->param.i_nal_hrd == X264_NAL_HRD_CBR;
675
h->sps->vui.hrd.i_time_offset_length = 0;
676
677
#define BR_SHIFT 6
678
#define CPB_SHIFT 4
679
680
// normalize HRD size and rate to the value / scale notation
681
h->sps->vui.hrd.i_bit_rate_scale = x264_clip3( x264_ctz( vbv_max_bitrate ) - BR_SHIFT, 0, 15 );
682
h->sps->vui.hrd.i_bit_rate_value = vbv_max_bitrate >> ( h->sps->vui.hrd.i_bit_rate_scale + BR_SHIFT );
683
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 );
684
h->sps->vui.hrd.i_cpb_size_scale = x264_clip3( x264_ctz( vbv_buffer_size ) - CPB_SHIFT, 0, 15 );
685
h->sps->vui.hrd.i_cpb_size_value = vbv_buffer_size >> ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
686
h->sps->vui.hrd.i_cpb_size_unscaled = h->sps->vui.hrd.i_cpb_size_value << ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
687
688
#undef CPB_SHIFT
689
#undef BR_SHIFT
690
691
// arbitrary
692
#define MAX_DURATION 0.5
693
694
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 );
695
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;
696
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
698
h->sps->vui.hrd.i_initial_cpb_removal_delay_length = 2 + x264_clip3( 32 - x264_clz( max_delay ), 4, 22 );
699
h->sps->vui.hrd.i_cpb_removal_delay_length = x264_clip3( 32 - x264_clz( max_cpb_output_delay ), 4, 31 );
700
h->sps->vui.hrd.i_dpb_output_delay_length = x264_clip3( 32 - x264_clz( max_dpb_output_delay ), 4, 31 );
701
702
#undef MAX_DURATION
703
704
vbv_buffer_size = h->sps->vui.hrd.i_cpb_size_unscaled;
705
vbv_max_bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
706
}
707
else if( h->param.i_nal_hrd && !b_init )
708
{
709
x264_log( h, X264_LOG_WARNING, "VBV parameters cannot be changed when NAL HRD is in use\n" );
710
return;
711
}
712
h->sps->vui.hrd.i_bit_rate_unscaled = vbv_max_bitrate;
713
h->sps->vui.hrd.i_cpb_size_unscaled = vbv_buffer_size;
714
715
if( rc->b_vbv_min_rate )
716
rc->bitrate = (double)h->param.rc.i_bitrate * kilobit_size;
717
rc->buffer_rate = vbv_max_bitrate / rc->fps;
718
rc->vbv_max_rate = vbv_max_bitrate;
719
rc->buffer_size = vbv_buffer_size;
720
rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
721
rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
722
* 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
723
if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.f_rf_constant_max )
724
{
725
rc->rate_factor_max_increment = h->param.rc.f_rf_constant_max - h->param.rc.f_rf_constant;
726
if( rc->rate_factor_max_increment <= 0 )
727
{
728
x264_log( h, X264_LOG_WARNING, "CRF max must be greater than CRF\n" );
729
rc->rate_factor_max_increment = 0;
730
}
731
}
732
if( b_init )
733
{
734
if( h->param.rc.f_vbv_buffer_init > 1. )
735
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 );
736
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);
737
rc->buffer_fill_final =
738
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 = 1;
740
rc->b_vbv_min_rate = !rc->b_2pass
741
&& h->param.rc.i_rc_method == X264_RC_ABR
742
&& h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
743
}
744
}
745
}
746
747
int x264_ratecontrol_new( x264_t *h )
748
{
749
x264_ratecontrol_t *rc;
750
751
x264_emms();
752
753
CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
754
rc = h->rc;
755
756
rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
757
rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
758
759
/* FIXME: use integers */
760
if( h->param.i_fps_num > 0 && h->param.i_fps_den > 0 )
761
rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
762
else
763
rc->fps = 25.0;
764
765
if( h->param.rc.b_mb_tree )
766
{
767
h->param.rc.f_pb_factor = 1;
768
rc->qcompress = 1;
769
}
770
else
771
rc->qcompress = h->param.rc.f_qcompress;
772
773
rc->bitrate = h->param.rc.i_bitrate * (h->param.i_avcintra_class ? 1024. : 1000.);
774
rc->rate_tolerance = h->param.rc.f_rate_tolerance;
775
rc->nmb = h->mb.i_mb_count;
776
rc->last_non_b_pict_type = -1;
777
rc->cbr_decay = 1.0;
778
779
if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
780
{
781
x264_log( h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n" );
782
return -1;
783
}
784
785
x264_ratecontrol_init_reconfigurable( h, 1 );
786
787
if( h->param.i_nal_hrd )
788
{
789
uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale;
790
uint64_t num = 90000;
791
x264_reduce_fraction64( &num, &denom );
792
rc->hrd_multiply_denom = 90000 / num;
793
794
double bits_required = log2( 90000 / rc->hrd_multiply_denom )
795
+ log2( h->sps->vui.i_time_scale )
796
+ log2( h->sps->vui.hrd.i_cpb_size_unscaled );
797
if( bits_required >= 63 )
798
{
799
x264_log( h, X264_LOG_ERROR, "HRD with very large timescale and bufsize not supported\n" );
800
return -1;
801
}
802
}
803
804
if( rc->rate_tolerance < 0.01 )
805
{
806
x264_log( h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n" );
807
rc->rate_tolerance = 0.01;
808
}
809
810
h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
811
812
if( rc->b_abr )
813
{
814
/* FIXME ABR_INIT_QP is actually used only in CRF */
815
#define ABR_INIT_QP (( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 ) + QP_BD_OFFSET)
816
rc->accum_p_norm = .01;
817
rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
818
/* estimated ratio that produces a reasonable QP for the first I-frame */
819
rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
820
rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
821
rc->last_non_b_pict_type = SLICE_TYPE_I;
822
}
823
824
rc->ip_offset = 6.0 * log2f( h->param.rc.f_ip_factor );
825
rc->pb_offset = 6.0 * log2f( h->param.rc.f_pb_factor );
826
rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
827
rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, QP_MAX );
828
rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, QP_MAX );
829
h->mb.ip_offset = rc->ip_offset + 0.5;
830
831
rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
832
rc->last_qscale = qp2qscale( 26 + QP_BD_OFFSET );
833
int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
834
CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
835
CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
836
static const float pred_coeff_table[3] = { 1.0, 1.0, 1.5 };
837
for( int i = 0; i < 3; i++ )
838
{
839
rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
840
rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
841
rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
842
for( int j = 0; j < num_preds; j++ )
843
{
844
rc->pred[i+j*5].coeff_min = pred_coeff_table[i] / 2;
845
rc->pred[i+j*5].coeff = pred_coeff_table[i];
846
rc->pred[i+j*5].count = 1.0;
847
rc->pred[i+j*5].decay = 0.5;
848
rc->pred[i+j*5].offset = 0.0;
849
}
850
for( int j = 0; j < 2; j++ )
851
{
852
rc->row_preds[i][j].coeff_min = .25 / 4;
853
rc->row_preds[i][j].coeff = .25;
854
rc->row_preds[i][j].count = 1.0;
855
rc->row_preds[i][j].decay = 0.5;
856
rc->row_preds[i][j].offset = 0.0;
857
}
858
}
859
rc->pred_b_from_p->coeff_min = 0.5 / 2;
860
rc->pred_b_from_p->coeff = 0.5;
861
rc->pred_b_from_p->count = 1.0;
862
rc->pred_b_from_p->decay = 0.5;
863
rc->pred_b_from_p->offset = 0.0;
864
865
if( parse_zones( h ) < 0 )
866
{
867
x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
868
return -1;
869
}
870
871
/* Load stat file and init 2pass algo */
872
if( h->param.rc.b_stat_read )
873
{
874
char *p, *stats_in, *stats_buf;
875
876
/* read 1st pass stats */
877
assert( h->param.rc.psz_stat_in );
878
stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
879
if( !stats_buf )
880
{
881
x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n" );
882
return -1;
883
}
884
if( h->param.rc.b_mb_tree )
885
{
886
char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
887
if( !mbtree_stats_in )
888
return -1;
889
rc->p_mbtree_stat_file_in = x264_fopen( mbtree_stats_in, "rb" );
890
x264_free( mbtree_stats_in );
891
if( !rc->p_mbtree_stat_file_in )
892
{
893
x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n" );
894
return -1;
895
}
896
}
897
898
/* check whether 1st pass options were compatible with current options */
899
if( strncmp( stats_buf, "#options:", 9 ) )
900
{
901
x264_log( h, X264_LOG_ERROR, "options list in stats file not valid\n" );
902
return -1;
903
}
904
905
float res_factor, res_factor_bits;
906
{
907
int i, j;
908
uint32_t k, l;
909
char *opts = stats_buf;
910
stats_in = strchr( stats_buf, '\n' );
911
if( !stats_in )
912
return -1;
913
*stats_in = '\0';
914
stats_in++;
915
if( sscanf( opts, "#options: %dx%d", &i, &j ) != 2 )
916
{
917
x264_log( h, X264_LOG_ERROR, "resolution specified in stats file not valid\n" );
918
return -1;
919
}
920
else if( h->param.rc.b_mb_tree )
921
{
922
rc->mbtree.srcdim[0] = i;
923
rc->mbtree.srcdim[1] = j;
924
}
925
res_factor = (float)h->param.i_width * h->param.i_height / (i*j);
926
/* Change in bits relative to resolution isn't quite linear on typical sources,
927
* so we'll at least try to roughly approximate this effect. */
928
res_factor_bits = powf( res_factor, 0.7 );
929
930
if( !( p = strstr( opts, "timebase=" ) ) || sscanf( p, "timebase=%u/%u", &k, &l ) != 2 )
931
{
932
x264_log( h, X264_LOG_ERROR, "timebase specified in stats file not valid\n" );
933
return -1;
934
}
935
if( k != h->param.i_timebase_num || l != h->param.i_timebase_den )
936
{
937
x264_log( h, X264_LOG_ERROR, "timebase mismatch with 1st pass (%u/%u vs %u/%u)\n",
938
h->param.i_timebase_num, h->param.i_timebase_den, k, l );
939
return -1;
940
}
941
942
CMP_OPT_FIRST_PASS( "bitdepth", BIT_DEPTH );
943
CMP_OPT_FIRST_PASS( "weightp", X264_MAX( 0, h->param.analyse.i_weighted_pred ) );
944
CMP_OPT_FIRST_PASS( "bframes", h->param.i_bframe );
945
CMP_OPT_FIRST_PASS( "b_pyramid", h->param.i_bframe_pyramid );
946
CMP_OPT_FIRST_PASS( "intra_refresh", h->param.b_intra_refresh );
947
CMP_OPT_FIRST_PASS( "open_gop", h->param.b_open_gop );
948
CMP_OPT_FIRST_PASS( "bluray_compat", h->param.b_bluray_compat );
949
950
if( (p = strstr( opts, "interlaced=" )) )
951
{
952
char *current = h->param.b_interlaced ? h->param.b_tff ? "tff" : "bff" : h->param.b_fake_interlaced ? "fake" : "0";
953
char buf[5];
954
sscanf( p, "interlaced=%4s", buf );
955
if( strcmp( current, buf ) )
956
{
957
x264_log( h, X264_LOG_ERROR, "different interlaced setting than first pass (%s vs %s)\n", current, buf );
958
return -1;
959
}
960
}
961
962
if( (p = strstr( opts, "keyint=" )) )
963
{
964
p += 7;
965
char buf[13] = "infinite ";
966
if( h->param.i_keyint_max != X264_KEYINT_MAX_INFINITE )
967
sprintf( buf, "%d ", h->param.i_keyint_max );
968
if( strncmp( p, buf, strlen(buf) ) )
969
{
970
x264_log( h, X264_LOG_ERROR, "different keyint setting than first pass (%.*s vs %.*s)\n",
971
strlen(buf)-1, buf, strcspn(p, " "), p );
972
return -1;
973
}
974
}
975
976
if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
977
x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
978
979
if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
980
{
981
x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
982
h->mb.b_direct_auto_write = 1;
983
}
984
985
if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
986
h->param.i_bframe_adaptive = i;
987
else if( h->param.i_bframe )
988
{
989
x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
990
return -1;
991
}
992
993
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 ) )
994
h->param.rc.i_lookahead = i;
995
}
996
997
/* find number of pics */
998
p = stats_in;
999
int num_entries;
1000
for( num_entries = -1; p; num_entries++ )
1001
p = strchr( p + 1, ';' );
1002
if( !num_entries )
1003
{
1004
x264_log( h, X264_LOG_ERROR, "empty stats file\n" );
1005
return -1;
1006
}
1007
rc->num_entries = num_entries;
1008
1009
if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
1010
{
1011
x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
1012
h->param.i_frame_total, rc->num_entries );
1013
}
1014
if( h->param.i_frame_total > rc->num_entries )
1015
{
1016
x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
1017
h->param.i_frame_total, rc->num_entries );
1018
return -1;
1019
}
1020
1021
CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
1022
CHECKED_MALLOC( rc->entry_out, rc->num_entries * sizeof(ratecontrol_entry_t*) );
1023
1024
/* init all to skipped p frames */
1025
for( int i = 0; i < rc->num_entries; i++ )
1026
{
1027
ratecontrol_entry_t *rce = &rc->entry[i];
1028
rce->pict_type = SLICE_TYPE_P;
1029
rce->qscale = rce->new_qscale = qp2qscale( 20 + QP_BD_OFFSET );
1030
rce->misc_bits = rc->nmb + 10;
1031
rce->new_qp = 0;
1032
rc->entry_out[i] = rce;
1033
}
1034
1035
/* read stats */
1036
p = stats_in;
1037
double total_qp_aq = 0;
1038
for( int i = 0; i < rc->num_entries; i++ )
1039
{
1040
ratecontrol_entry_t *rce;
1041
int frame_number = 0;
1042
int frame_out_number = 0;
1043
char pict_type = 0;
1044
int e;
1045
char *next;
1046
float qp_rc, qp_aq;
1047
int ref;
1048
1049
next= strchr(p, ';');
1050
if( next )
1051
*next++ = 0; //sscanf is unbelievably slow on long strings
1052
e = sscanf( p, " in:%d out:%d ", &frame_number, &frame_out_number );
1053
1054
if( frame_number < 0 || frame_number >= rc->num_entries )
1055
{
1056
x264_log( h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i );
1057
return -1;
1058
}
1059
if( frame_out_number < 0 || frame_out_number >= rc->num_entries )
1060
{
1061
x264_log( h, X264_LOG_ERROR, "bad frame output number (%d) at stats line %d\n", frame_out_number, i );
1062
return -1;
1063
}
1064
rce = &rc->entry[frame_number];
1065
rc->entry_out[frame_out_number] = rce;
1066
rce->direct_mode = 0;
1067
1068
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",
1069
&pict_type, &rce->i_duration, &rce->i_cpb_duration, &qp_rc, &qp_aq, &rce->tex_bits,
1070
&rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
1071
&rce->s_count, &rce->direct_mode );
1072
rce->tex_bits *= res_factor_bits;
1073
rce->mv_bits *= res_factor_bits;
1074
rce->misc_bits *= res_factor_bits;
1075
rce->i_count *= res_factor;
1076
rce->p_count *= res_factor;
1077
rce->s_count *= res_factor;
1078
1079
p = strstr( p, "ref:" );
1080
if( !p )
1081
goto parse_error;
1082
p += 4;
1083
for( ref = 0; ref < 16; ref++ )
1084
{
1085
if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
1086
break;
1087
p = strchr( p+1, ' ' );
1088
if( !p )
1089
goto parse_error;
1090
}
1091
rce->refs = ref;
1092
1093
/* find weights */
1094
rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
1095
char *w = strchr( p, 'w' );
1096
if( w )
1097
{
1098
int count = sscanf( w, "w:%hd,%hd,%hd,%hd,%hd,%hd,%hd,%hd",
1099
&rce->i_weight_denom[0], &rce->weight[0][0], &rce->weight[0][1],
1100
&rce->i_weight_denom[1], &rce->weight[1][0], &rce->weight[1][1],
1101
&rce->weight[2][0], &rce->weight[2][1] );
1102
if( count == 3 )
1103
rce->i_weight_denom[1] = -1;
1104
else if ( count != 8 )
1105
rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
1106
}
1107
1108
if( pict_type != 'b' )
1109
rce->kept_as_ref = 1;
1110
switch( pict_type )
1111
{
1112
case 'I':
1113
rce->frame_type = X264_TYPE_IDR;
1114
rce->pict_type = SLICE_TYPE_I;
1115
break;
1116
case 'i':
1117
rce->frame_type = X264_TYPE_I;
1118
rce->pict_type = SLICE_TYPE_I;
1119
break;
1120
case 'P':
1121
rce->frame_type = X264_TYPE_P;
1122
rce->pict_type = SLICE_TYPE_P;
1123
break;
1124
case 'B':
1125
rce->frame_type = X264_TYPE_BREF;
1126
rce->pict_type = SLICE_TYPE_B;
1127
break;
1128
case 'b':
1129
rce->frame_type = X264_TYPE_B;
1130
rce->pict_type = SLICE_TYPE_B;
1131
break;
1132
default: e = -1; break;
1133
}
1134
if( e < 14 )
1135
{
1136
parse_error:
1137
x264_log( h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e );
1138
return -1;
1139
}
1140
rce->qscale = qp2qscale( qp_rc );
1141
total_qp_aq += qp_aq;
1142
p = next;
1143
}
1144
if( !h->param.b_stitchable )
1145
h->pps->i_pic_init_qp = SPEC_QP( (int)(total_qp_aq / rc->num_entries + 0.5) );
1146
1147
x264_free( stats_buf );
1148
1149
if( h->param.rc.i_rc_method == X264_RC_ABR )
1150
{
1151
if( init_pass2( h ) < 0 )
1152
return -1;
1153
} /* else we're using constant quant, so no need to run the bitrate allocation */
1154
}
1155
1156
/* Open output file */
1157
/* If input and output files are the same, output to a temp file
1158
* and move it to the real name only when it's complete */
1159
if( h->param.rc.b_stat_write )
1160
{
1161
char *p;
1162
rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
1163
if( !rc->psz_stat_file_tmpname )
1164
return -1;
1165
1166
rc->p_stat_file_out = x264_fopen( rc->psz_stat_file_tmpname, "wb" );
1167
if( rc->p_stat_file_out == NULL )
1168
{
1169
x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n" );
1170
return -1;
1171
}
1172
1173
p = x264_param2string( &h->param, 1 );
1174
if( p )
1175
fprintf( rc->p_stat_file_out, "#options: %s\n", p );
1176
x264_free( p );
1177
if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
1178
{
1179
rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
1180
rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
1181
if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
1182
return -1;
1183
1184
rc->p_mbtree_stat_file_out = x264_fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
1185
if( rc->p_mbtree_stat_file_out == NULL )
1186
{
1187
x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n" );
1188
return -1;
1189
}
1190
}
1191
}
1192
1193
if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
1194
{
1195
if( !h->param.rc.b_stat_read )
1196
{
1197
rc->mbtree.srcdim[0] = h->param.i_width;
1198
rc->mbtree.srcdim[1] = h->param.i_height;
1199
}
1200
if( x264_macroblock_tree_rescale_init( h, rc ) < 0 )
1201
return -1;
1202
}
1203
1204
for( int i = 0; i<h->param.i_threads; i++ )
1205
{
1206
h->thread[i]->rc = rc+i;
1207
if( i )
1208
{
1209
rc[i] = rc[0];
1210
h->thread[i]->param = h->param;
1211
h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
1212
h->thread[i]->mb.ip_offset = h->mb.ip_offset;
1213
}
1214
}
1215
1216
return 0;
1217
fail:
1218
return -1;
1219
}
1220
1221
static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
1222
{
1223
int len = 0;
1224
char *tok, UNUSED *saveptr=NULL;
1225
z->param = NULL;
1226
z->f_bitrate_factor = 1;
1227
if( 3 <= sscanf(p, "%d,%d,q=%d%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
1228
z->b_force_qp = 1;
1229
else if( 3 <= sscanf(p, "%d,%d,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
1230
z->b_force_qp = 0;
1231
else if( 2 <= sscanf(p, "%d,%d%n", &z->i_start, &z->i_end, &len) )
1232
z->b_force_qp = 0;
1233
else
1234
{
1235
x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
1236
return -1;
1237
}
1238
p += len;
1239
if( !*p )
1240
return 0;
1241
CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
1242
memcpy( z->param, &h->param, sizeof(x264_param_t) );
1243
z->param->param_free = x264_free;
1244
while( (tok = strtok_r( p, ",", &saveptr )) )
1245
{
1246
char *val = strchr( tok, '=' );
1247
if( val )
1248
{
1249
*val = '\0';
1250
val++;
1251
}
1252
if( x264_param_parse( z->param, tok, val ) )
1253
{
1254
x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
1255
return -1;
1256
}
1257
p = NULL;
1258
}
1259
return 0;
1260
fail:
1261
return -1;
1262
}
1263
1264
static int parse_zones( x264_t *h )
1265
{
1266
x264_ratecontrol_t *rc = h->rc;
1267
if( h->param.rc.psz_zones && !h->param.rc.i_zones )
1268
{
1269
char *psz_zones, *p;
1270
CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
1271
strcpy( psz_zones, h->param.rc.psz_zones );
1272
h->param.rc.i_zones = 1;
1273
for( p = psz_zones; *p; p++ )
1274
h->param.rc.i_zones += (*p == '/');
1275
CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
1276
p = psz_zones;
1277
for( int i = 0; i < h->param.rc.i_zones; i++ )
1278
{
1279
int i_tok = strcspn( p, "/" );
1280
p[i_tok] = 0;
1281
if( parse_zone( h, &h->param.rc.zones[i], p ) )
1282
return -1;
1283
p += i_tok + 1;
1284
}
1285
x264_free( psz_zones );
1286
}
1287
1288
if( h->param.rc.i_zones > 0 )
1289
{
1290
for( int i = 0; i < h->param.rc.i_zones; i++ )
1291
{
1292
x264_zone_t z = h->param.rc.zones[i];
1293
if( z.i_start < 0 || z.i_start > z.i_end )
1294
{
1295
x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
1296
z.i_start, z.i_end );
1297
return -1;
1298
}
1299
else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
1300
{
1301
x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
1302
z.f_bitrate_factor );
1303
return -1;
1304
}
1305
}
1306
1307
rc->i_zones = h->param.rc.i_zones + 1;
1308
CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
1309
memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
1310
1311
// default zone to fall back to if none of the others match
1312
rc->zones[0].i_start = 0;
1313
rc->zones[0].i_end = INT_MAX;
1314
rc->zones[0].b_force_qp = 0;
1315
rc->zones[0].f_bitrate_factor = 1;
1316
CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
1317
memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
1318
for( int i = 1; i < rc->i_zones; i++ )
1319
{
1320
if( !rc->zones[i].param )
1321
rc->zones[i].param = rc->zones[0].param;
1322
}
1323
}
1324
1325
return 0;
1326
fail:
1327
return -1;
1328
}
1329
1330
static x264_zone_t *get_zone( x264_t *h, int frame_num )
1331
{
1332
for( int i = h->rc->i_zones - 1; i >= 0; i-- )
1333
{
1334
x264_zone_t *z = &h->rc->zones[i];
1335
if( frame_num >= z->i_start && frame_num <= z->i_end )
1336
return z;
1337
}
1338
return NULL;
1339
}
1340
1341
void x264_ratecontrol_summary( x264_t *h )
1342
{
1343
x264_ratecontrol_t *rc = h->rc;
1344
if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
1345
{
1346
double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
1347
double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
1348
x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
1349
qscale2qp( pow( base_cplx, 1 - rc->qcompress )
1350
* rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset - QP_BD_OFFSET );
1351
}
1352
}
1353
1354
void x264_ratecontrol_delete( x264_t *h )
1355
{
1356
x264_ratecontrol_t *rc = h->rc;
1357
int b_regular_file;
1358
1359
if( rc->p_stat_file_out )
1360
{
1361
b_regular_file = x264_is_regular_file( rc->p_stat_file_out );
1362
fclose( rc->p_stat_file_out );
1363
if( h->i_frame >= rc->num_entries && b_regular_file )
1364
if( x264_rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
1365
{
1366
x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1367
rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
1368
}
1369
x264_free( rc->psz_stat_file_tmpname );
1370
}
1371
if( rc->p_mbtree_stat_file_out )
1372
{
1373
b_regular_file = x264_is_regular_file( rc->p_mbtree_stat_file_out );
1374
fclose( rc->p_mbtree_stat_file_out );
1375
if( h->i_frame >= rc->num_entries && b_regular_file )
1376
if( x264_rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
1377
{
1378
x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1379
rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
1380
}
1381
x264_free( rc->psz_mbtree_stat_file_tmpname );
1382
x264_free( rc->psz_mbtree_stat_file_name );
1383
}
1384
if( rc->p_mbtree_stat_file_in )
1385
fclose( rc->p_mbtree_stat_file_in );
1386
x264_free( rc->pred );
1387
x264_free( rc->pred_b_from_p );
1388
x264_free( rc->entry );
1389
x264_free( rc->entry_out );
1390
x264_macroblock_tree_rescale_destroy( rc );
1391
if( rc->zones )
1392
{
1393
x264_free( rc->zones[0].param );
1394
for( int i = 1; i < rc->i_zones; i++ )
1395
if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
1396
rc->zones[i].param->param_free( rc->zones[i].param );
1397
x264_free( rc->zones );
1398
}
1399
x264_free( rc );
1400
}
1401
1402
static void accum_p_qp_update( x264_t *h, float qp )
1403
{
1404
x264_ratecontrol_t *rc = h->rc;
1405
rc->accum_p_qp *= .95;
1406
rc->accum_p_norm *= .95;
1407
rc->accum_p_norm += 1;
1408
if( h->sh.i_type == SLICE_TYPE_I )
1409
rc->accum_p_qp += qp + rc->ip_offset;
1410
else
1411
rc->accum_p_qp += qp;
1412
}
1413
1414
/* Before encoding a frame, choose a QP for it */
1415
void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
1416
{
1417
x264_ratecontrol_t *rc = h->rc;
1418
ratecontrol_entry_t *rce = NULL;
1419
x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
1420
float q;
1421
1422
x264_emms();
1423
1424
if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
1425
x264_encoder_reconfig_apply( h, zone->param );
1426
rc->prev_zone = zone;
1427
1428
if( h->param.rc.b_stat_read )
1429
{
1430
int frame = h->fenc->i_frame;
1431
assert( frame >= 0 && frame < rc->num_entries );
1432
rce = h->rc->rce = &h->rc->entry[frame];
1433
1434
if( h->sh.i_type == SLICE_TYPE_B
1435
&& h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
1436
{
1437
h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
1438
h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
1439
}
1440
}
1441
1442
if( rc->b_vbv )
1443
{
1444
memset( h->fdec->i_row_bits, 0, h->mb.i_mb_height * sizeof(int) );
1445
memset( h->fdec->f_row_qp, 0, h->mb.i_mb_height * sizeof(float) );
1446
memset( h->fdec->f_row_qscale, 0, h->mb.i_mb_height * sizeof(float) );
1447
rc->row_pred = rc->row_preds[h->sh.i_type];
1448
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;
1449
update_vbv_plan( h, overhead );
1450
1451
const x264_level_t *l = x264_levels;
1452
while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
1453
l++;
1454
1455
int mincr = l->mincr;
1456
1457
if( h->param.b_bluray_compat )
1458
mincr = 4;
1459
1460
/* Profiles above High don't require minCR, so just set the maximum to a large value. */
1461
if( h->sps->i_profile_idc > PROFILE_HIGH )
1462
rc->frame_size_maximum = 1e9;
1463
else
1464
{
1465
/* The spec has a bizarre special case for the first frame. */
1466
if( h->i_frame == 0 )
1467
{
1468
//384 * ( Max( PicSizeInMbs, fR * MaxMBPS ) + MaxMBPS * ( tr( 0 ) - tr,n( 0 ) ) ) / MinCR
1469
double fr = 1. / 172;
1470
int pic_size_in_mbs = h->mb.i_mb_width * h->mb.i_mb_height;
1471
rc->frame_size_maximum = 384 * BIT_DEPTH * X264_MAX( pic_size_in_mbs, fr*l->mbps ) / mincr;
1472
}
1473
else
1474
{
1475
//384 * MaxMBPS * ( tr( n ) - tr( n - 1 ) ) / MinCR
1476
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;
1477
}
1478
}
1479
}
1480
1481
if( h->sh.i_type != SLICE_TYPE_B )
1482
rc->bframes = h->fenc->i_bframes;
1483
1484
if( rc->b_abr )
1485
{
1486
q = qscale2qp( rate_estimate_qscale( h ) );
1487
}
1488
else if( rc->b_2pass )
1489
{
1490
rce->new_qscale = rate_estimate_qscale( h );
1491
q = qscale2qp( rce->new_qscale );
1492
}
1493
else /* CQP */
1494
{
1495
if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
1496
q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
1497
else
1498
q = rc->qp_constant[ h->sh.i_type ];
1499
1500
if( zone )
1501
{
1502
if( zone->b_force_qp )
1503
q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1504
else
1505
q -= 6*log2f( zone->f_bitrate_factor );
1506
}
1507
}
1508
if( i_force_qp != X264_QP_AUTO )
1509
q = i_force_qp - 1;
1510
1511
q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1512
1513
rc->qpa_rc = rc->qpa_rc_prev =
1514
rc->qpa_aq = rc->qpa_aq_prev = 0;
1515
h->fdec->f_qp_avg_rc =
1516
h->fdec->f_qp_avg_aq =
1517
rc->qpm = q;
1518
if( rce )
1519
rce->new_qp = q;
1520
1521
accum_p_qp_update( h, rc->qpm );
1522
1523
if( h->sh.i_type != SLICE_TYPE_B )
1524
rc->last_non_b_pict_type = h->sh.i_type;
1525
}
1526
1527
static float predict_row_size( x264_t *h, int y, float qscale )
1528
{
1529
/* average between two predictors:
1530
* absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1531
x264_ratecontrol_t *rc = h->rc;
1532
float pred_s = predict_size( &rc->row_pred[0], qscale, h->fdec->i_row_satd[y] );
1533
if( h->sh.i_type == SLICE_TYPE_I || qscale >= h->fref[0][0]->f_row_qscale[y] )
1534
{
1535
if( h->sh.i_type == SLICE_TYPE_P
1536
&& h->fref[0][0]->i_type == h->fdec->i_type
1537
&& h->fref[0][0]->f_row_qscale[y] > 0
1538
&& h->fref[0][0]->i_row_satd[y] > 0
1539
&& (abs(h->fref[0][0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1540
{
1541
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]
1542
* h->fref[0][0]->f_row_qscale[y] / qscale;
1543
return (pred_s + pred_t) * 0.5f;
1544
}
1545
return pred_s;
1546
}
1547
/* Our QP is lower than the reference! */
1548
else
1549
{
1550
float pred_intra = predict_size( &rc->row_pred[1], qscale, h->fdec->i_row_satds[0][0][y] );
1551
/* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1552
return pred_intra + pred_s;
1553
}
1554
}
1555
1556
static int row_bits_so_far( x264_t *h, int y )
1557
{
1558
int bits = 0;
1559
for( int i = h->i_threadslice_start; i <= y; i++ )
1560
bits += h->fdec->i_row_bits[i];
1561
return bits;
1562
}
1563
1564
static float predict_row_size_to_end( x264_t *h, int y, float qp )
1565
{
1566
float qscale = qp2qscale( qp );
1567
float bits = 0;
1568
for( int i = y+1; i < h->i_threadslice_end; i++ )
1569
bits += predict_row_size( h, i, qscale );
1570
return bits;
1571
}
1572
1573
/* TODO:
1574
* eliminate all use of qp in row ratecontrol: make it entirely qscale-based.
1575
* make this function stop being needlessly O(N^2)
1576
* update more often than once per row? */
1577
int x264_ratecontrol_mb( x264_t *h, int bits )
1578
{
1579
x264_ratecontrol_t *rc = h->rc;
1580
const int y = h->mb.i_mb_y;
1581
1582
h->fdec->i_row_bits[y] += bits;
1583
rc->qpa_aq += h->mb.i_qp;
1584
1585
if( h->mb.i_mb_x != h->mb.i_mb_width - 1 )
1586
return 0;
1587
1588
x264_emms();
1589
rc->qpa_rc += rc->qpm * h->mb.i_mb_width;
1590
1591
if( !rc->b_vbv )
1592
return 0;
1593
1594
float qscale = qp2qscale( rc->qpm );
1595
h->fdec->f_row_qp[y] = rc->qpm;
1596
h->fdec->f_row_qscale[y] = qscale;
1597
1598
update_predictor( &rc->row_pred[0], qscale, h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1599
if( h->sh.i_type != SLICE_TYPE_I && rc->qpm < h->fref[0][0]->f_row_qp[y] )
1600
update_predictor( &rc->row_pred[1], qscale, h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1601
1602
/* update ratecontrol per-mbpair in MBAFF */
1603
if( SLICE_MBAFF && !(y&1) )
1604
return 0;
1605
1606
/* FIXME: We don't currently support the case where there's a slice
1607
* boundary in between. */
1608
int can_reencode_row = h->sh.i_first_mb <= ((h->mb.i_mb_y - SLICE_MBAFF) * h->mb.i_mb_stride);
1609
1610
/* tweak quality based on difference from predicted size */
1611
float prev_row_qp = h->fdec->f_row_qp[y];
1612
float qp_absolute_max = h->param.rc.i_qp_max;
1613
if( rc->rate_factor_max_increment )
1614
qp_absolute_max = X264_MIN( qp_absolute_max, rc->qp_novbv + rc->rate_factor_max_increment );
1615
float qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, qp_absolute_max );
1616
float qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1617
float step_size = 0.5f;
1618
float slice_size_planned = h->param.b_sliced_threads ? rc->slice_size_planned : rc->frame_size_planned;
1619
float bits_so_far = row_bits_so_far( h, y );
1620
float max_frame_error = x264_clip3f( 1.0 / h->mb.i_mb_height, 0.05, 0.25 );
1621
float max_frame_size = rc->frame_size_maximum - rc->frame_size_maximum * max_frame_error;
1622
max_frame_size = X264_MIN( max_frame_size, rc->buffer_fill - rc->buffer_rate * max_frame_error );
1623
float size_of_other_slices = 0;
1624
if( h->param.b_sliced_threads )
1625
{
1626
float size_of_other_slices_planned = 0;
1627
for( int i = 0; i < h->param.i_threads; i++ )
1628
if( h != h->thread[i] )
1629
{
1630
size_of_other_slices += h->thread[i]->rc->frame_size_estimated;
1631
size_of_other_slices_planned += h->thread[i]->rc->slice_size_planned;
1632
}
1633
float weight = rc->slice_size_planned / rc->frame_size_planned;
1634
size_of_other_slices = (size_of_other_slices - size_of_other_slices_planned) * weight + size_of_other_slices_planned;
1635
}
1636
if( y < h->i_threadslice_end-1 )
1637
{
1638
/* B-frames shouldn't use lower QP than their reference frames. */
1639
if( h->sh.i_type == SLICE_TYPE_B )
1640
{
1641
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] ) );
1642
rc->qpm = X264_MAX( rc->qpm, qp_min );
1643
}
1644
1645
float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1646
buffer_left_planned = X264_MAX( buffer_left_planned, 0.f );
1647
/* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1648
float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
1649
float b1 = bits_so_far + predict_row_size_to_end( h, y, rc->qpm ) + size_of_other_slices;
1650
float trust_coeff = x264_clip3f( bits_so_far / slice_size_planned, 0.0, 1.0 );
1651
1652
/* Don't increase the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1653
/* area at the top of the frame was measured inaccurately. */
1654
if( trust_coeff < 0.05f )
1655
qp_max = qp_absolute_max = prev_row_qp;
1656
1657
if( h->sh.i_type != SLICE_TYPE_I )
1658
rc_tol *= 0.5f;
1659
1660
if( !rc->b_vbv_min_rate )
1661
qp_min = X264_MAX( qp_min, rc->qp_novbv );
1662
1663
while( rc->qpm < qp_max
1664
&& ((b1 > rc->frame_size_planned + rc_tol) ||
1665
(b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv) ||
1666
(b1 > rc->buffer_fill - buffer_left_planned * 0.5f)) )
1667
{
1668
rc->qpm += step_size;
1669
b1 = bits_so_far + predict_row_size_to_end( h, y, rc->qpm ) + size_of_other_slices;
1670
}
1671
1672
float b_max = b1 + ((rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 0.90f - b1) * trust_coeff;
1673
rc->qpm -= step_size;
1674
float b2 = bits_so_far + predict_row_size_to_end( h, y, rc->qpm ) + size_of_other_slices;
1675
while( rc->qpm > qp_min && rc->qpm < prev_row_qp
1676
&& (rc->qpm > h->fdec->f_row_qp[0] || rc->single_frame_vbv)
1677
&& (b2 < max_frame_size)
1678
&& ((b2 < rc->frame_size_planned * 0.8f) || (b2 < b_max)) )
1679
{
1680
b1 = b2;
1681
rc->qpm -= step_size;
1682
b2 = bits_so_far + predict_row_size_to_end( h, y, rc->qpm ) + size_of_other_slices;
1683
}
1684
rc->qpm += step_size;
1685
1686
/* avoid VBV underflow or MinCR violation */
1687
while( rc->qpm < qp_absolute_max && (b1 > max_frame_size) )
1688
{
1689
rc->qpm += step_size;
1690
b1 = bits_so_far + predict_row_size_to_end( h, y, rc->qpm ) + size_of_other_slices;
1691
}
1692
1693
h->rc->frame_size_estimated = b1 - size_of_other_slices;
1694
1695
/* If the current row was large enough to cause a large QP jump, try re-encoding it. */
1696
if( rc->qpm > qp_max && prev_row_qp < qp_max && can_reencode_row )
1697
{
1698
/* Bump QP to halfway in between... close enough. */
1699
rc->qpm = x264_clip3f( (prev_row_qp + rc->qpm)*0.5f, prev_row_qp + 1.0f, qp_max );
1700
rc->qpa_rc = rc->qpa_rc_prev;
1701
rc->qpa_aq = rc->qpa_aq_prev;
1702
h->fdec->i_row_bits[y] = 0;
1703
h->fdec->i_row_bits[y-SLICE_MBAFF] = 0;
1704
return -1;
1705
}
1706
}
1707
else
1708
{
1709
h->rc->frame_size_estimated = bits_so_far;
1710
1711
/* Last-ditch attempt: if the last row of the frame underflowed the VBV,
1712
* try again. */
1713
if( rc->qpm < qp_max && can_reencode_row
1714
&& (h->rc->frame_size_estimated + size_of_other_slices > X264_MIN( rc->frame_size_maximum, rc->buffer_fill )) )
1715
{
1716
rc->qpm = qp_max;
1717
rc->qpa_rc = rc->qpa_rc_prev;
1718
rc->qpa_aq = rc->qpa_aq_prev;
1719
h->fdec->i_row_bits[y] = 0;
1720
h->fdec->i_row_bits[y-SLICE_MBAFF] = 0;
1721
return -1;
1722
}
1723
}
1724
1725
rc->qpa_rc_prev = rc->qpa_rc;
1726
rc->qpa_aq_prev = rc->qpa_aq;
1727
1728
return 0;
1729
}
1730
1731
int x264_ratecontrol_qp( x264_t *h )
1732
{
1733
x264_emms();
1734
return x264_clip3( h->rc->qpm + 0.5f, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1735
}
1736
1737
int x264_ratecontrol_mb_qp( x264_t *h )
1738
{
1739
x264_emms();
1740
float qp = h->rc->qpm;
1741
if( h->param.rc.i_aq_mode )
1742
{
1743
/* MB-tree currently doesn't adjust quantizers in unreferenced frames. */
1744
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];
1745
/* Scale AQ's effect towards zero in emergency mode. */
1746
if( qp > QP_MAX_SPEC )
1747
qp_offset *= (QP_MAX - qp) / (QP_MAX - QP_MAX_SPEC);
1748
qp += qp_offset;
1749
}
1750
return x264_clip3( qp + 0.5f, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1751
}
1752
1753
/* In 2pass, force the same frame types as in the 1st pass */
1754
int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1755
{
1756
x264_ratecontrol_t *rc = h->rc;
1757
if( h->param.rc.b_stat_read )
1758
{
1759
if( frame_num >= rc->num_entries )
1760
{
1761
/* We could try to initialize everything required for ABR and
1762
* adaptive B-frames, but that would be complicated.
1763
* So just calculate the average QP used so far. */
1764
h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24 + QP_BD_OFFSET
1765
: 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1766
rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, QP_MAX );
1767
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 );
1768
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 );
1769
1770
x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries );
1771
x264_log( h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant );
1772
if( h->param.i_bframe_adaptive )
1773
x264_log( h, X264_LOG_ERROR, "disabling adaptive B-frames\n" );
1774
1775
for( int i = 0; i < h->param.i_threads; i++ )
1776
{
1777
h->thread[i]->rc->b_abr = 0;
1778
h->thread[i]->rc->b_2pass = 0;
1779
h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1780
h->thread[i]->param.rc.b_stat_read = 0;
1781
h->thread[i]->param.i_bframe_adaptive = 0;
1782
h->thread[i]->param.i_scenecut_threshold = 0;
1783
h->thread[i]->param.rc.b_mb_tree = 0;
1784
if( h->thread[i]->param.i_bframe > 1 )
1785
h->thread[i]->param.i_bframe = 1;
1786
}
1787
return X264_TYPE_AUTO;
1788
}
1789
return rc->entry[frame_num].frame_type;
1790
}
1791
else
1792
return X264_TYPE_AUTO;
1793
}
1794
1795
void x264_ratecontrol_set_weights( x264_t *h, x264_frame_t *frm )
1796
{
1797
ratecontrol_entry_t *rce = &h->rc->entry[frm->i_frame];
1798
if( h->param.analyse.i_weighted_pred <= 0 )
1799
return;
1800
1801
if( rce->i_weight_denom[0] >= 0 )
1802
SET_WEIGHT( frm->weight[0][0], 1, rce->weight[0][0], rce->i_weight_denom[0], rce->weight[0][1] );
1803
1804
if( rce->i_weight_denom[1] >= 0 )
1805
{
1806
SET_WEIGHT( frm->weight[0][1], 1, rce->weight[1][0], rce->i_weight_denom[1], rce->weight[1][1] );
1807
SET_WEIGHT( frm->weight[0][2], 1, rce->weight[2][0], rce->i_weight_denom[1], rce->weight[2][1] );
1808
}
1809
}
1810
1811
/* After encoding one frame, save stats and update ratecontrol state */
1812
int x264_ratecontrol_end( x264_t *h, int bits, int *filler )
1813
{
1814
x264_ratecontrol_t *rc = h->rc;
1815
const int *mbs = h->stat.frame.i_mb_count;
1816
1817
x264_emms();
1818
1819
h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1820
h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1821
h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1822
for( int i = B_DIRECT; i < B_8x8; i++ )
1823
h->stat.frame.i_mb_count_p += mbs[i];
1824
1825
h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1826
h->fdec->f_qp_avg_aq = (float)rc->qpa_aq / h->mb.i_mb_count;
1827
h->fdec->f_crf_avg = h->param.rc.f_rf_constant + h->fdec->f_qp_avg_rc - rc->qp_novbv;
1828
1829
if( h->param.rc.b_stat_write )
1830
{
1831
char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1832
: h->sh.i_type==SLICE_TYPE_P ? 'P'
1833
: h->fenc->b_kept_as_ref ? 'B' : 'b';
1834
int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1835
int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1836
char c_direct = h->mb.b_direct_auto_write ?
1837
( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1838
dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1839
: '-';
1840
if( fprintf( rc->p_stat_file_out,
1841
"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:",
1842
h->fenc->i_frame, h->i_frame,
1843
c_type, h->fenc->i_duration,
1844
h->fenc->i_cpb_duration,
1845
rc->qpa_rc, h->fdec->f_qp_avg_aq,
1846
h->stat.frame.i_tex_bits,
1847
h->stat.frame.i_mv_bits,
1848
h->stat.frame.i_misc_bits,
1849
h->stat.frame.i_mb_count_i,
1850
h->stat.frame.i_mb_count_p,
1851
h->stat.frame.i_mb_count_skip,
1852
c_direct) < 0 )
1853
goto fail;
1854
1855
/* Only write information for reference reordering once. */
1856
int use_old_stats = h->param.rc.b_stat_read && rc->rce->refs > 1;
1857
for( int i = 0; i < (use_old_stats ? rc->rce->refs : h->i_ref[0]); i++ )
1858
{
1859
int refcount = use_old_stats ? rc->rce->refcount[i]
1860
: PARAM_INTERLACED ? h->stat.frame.i_mb_count_ref[0][i*2]
1861
+ h->stat.frame.i_mb_count_ref[0][i*2+1]
1862
: h->stat.frame.i_mb_count_ref[0][i];
1863
if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1864
goto fail;
1865
}
1866
1867
if( h->param.analyse.i_weighted_pred >= X264_WEIGHTP_SIMPLE && h->sh.weight[0][0].weightfn )
1868
{
1869
if( fprintf( rc->p_stat_file_out, "w:%d,%d,%d",
1870
h->sh.weight[0][0].i_denom, h->sh.weight[0][0].i_scale, h->sh.weight[0][0].i_offset ) < 0 )
1871
goto fail;
1872
if( h->sh.weight[0][1].weightfn || h->sh.weight[0][2].weightfn )
1873
{
1874
if( fprintf( rc->p_stat_file_out, ",%d,%d,%d,%d,%d ",
1875
h->sh.weight[0][1].i_denom, h->sh.weight[0][1].i_scale, h->sh.weight[0][1].i_offset,
1876
h->sh.weight[0][2].i_scale, h->sh.weight[0][2].i_offset ) < 0 )
1877
goto fail;
1878
}
1879
else if( fprintf( rc->p_stat_file_out, " " ) < 0 )
1880
goto fail;
1881
}
1882
1883
if( fprintf( rc->p_stat_file_out, ";\n") < 0 )
1884
goto fail;
1885
1886
/* Don't re-write the data in multi-pass mode. */
1887
if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1888
{
1889
uint8_t i_type = h->sh.i_type;
1890
/* Values are stored as big-endian FIX8.8 */
1891
for( int i = 0; i < h->mb.i_mb_count; i++ )
1892
rc->mbtree.qp_buffer[0][i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1893
if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1894
goto fail;
1895
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 )
1896
goto fail;
1897
}
1898
}
1899
1900
if( rc->b_abr )
1901
{
1902
if( h->sh.i_type != SLICE_TYPE_B )
1903
rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / rc->last_rceq;
1904
else
1905
{
1906
/* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1907
* Not perfectly accurate with B-refs, but good enough. */
1908
rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / (rc->last_rceq * fabs( h->param.rc.f_pb_factor ));
1909
}
1910
rc->cplxr_sum *= rc->cbr_decay;
1911
rc->wanted_bits_window += h->fenc->f_duration * rc->bitrate;
1912
rc->wanted_bits_window *= rc->cbr_decay;
1913
}
1914
1915
if( rc->b_2pass )
1916
rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale( rc->rce->new_qp ) );
1917
1918
if( h->mb.b_variable_qp )
1919
{
1920
if( h->sh.i_type == SLICE_TYPE_B )
1921
{
1922
rc->bframe_bits += bits;
1923
if( h->fenc->b_last_minigop_bframe )
1924
{
1925
update_predictor( rc->pred_b_from_p, qp2qscale( rc->qpa_rc ),
1926
h->fref[1][h->i_ref[1]-1]->i_satd, rc->bframe_bits / rc->bframes );
1927
rc->bframe_bits = 0;
1928
}
1929
}
1930
}
1931
1932
*filler = update_vbv( h, bits );
1933
rc->filler_bits_sum += *filler * 8;
1934
1935
if( h->sps->vui.b_nal_hrd_parameters_present )
1936
{
1937
if( h->fenc->i_frame == 0 )
1938
{
1939
// access unit initialises the HRD
1940
h->fenc->hrd_timing.cpb_initial_arrival_time = 0;
1941
rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1942
rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1943
h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit = (double)rc->initial_cpb_removal_delay / 90000;
1944
}
1945
else
1946
{
1947
h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit + (double)(h->fenc->i_cpb_delay - h->i_cpb_delay_pir_offset) *
1948
h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1949
1950
if( h->fenc->b_keyframe )
1951
{
1952
rc->nrt_first_access_unit = h->fenc->hrd_timing.cpb_removal_time;
1953
rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1954
rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1955
}
1956
1957
double cpb_earliest_arrival_time = h->fenc->hrd_timing.cpb_removal_time - (double)rc->initial_cpb_removal_delay / 90000;
1958
if( !h->fenc->b_keyframe )
1959
cpb_earliest_arrival_time -= (double)rc->initial_cpb_removal_delay_offset / 90000;
1960
1961
if( h->sps->vui.hrd.b_cbr_hrd )
1962
h->fenc->hrd_timing.cpb_initial_arrival_time = rc->previous_cpb_final_arrival_time;
1963
else
1964
h->fenc->hrd_timing.cpb_initial_arrival_time = X264_MAX( rc->previous_cpb_final_arrival_time, cpb_earliest_arrival_time );
1965
}
1966
int filler_bits = *filler ? X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), *filler )*8 : 0;
1967
// Equation C-6
1968
h->fenc->hrd_timing.cpb_final_arrival_time = rc->previous_cpb_final_arrival_time = h->fenc->hrd_timing.cpb_initial_arrival_time +
1969
(double)(bits + filler_bits) / h->sps->vui.hrd.i_bit_rate_unscaled;
1970
1971
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 +
1972
h->fenc->hrd_timing.cpb_removal_time;
1973
}
1974
1975
return 0;
1976
fail:
1977
x264_log( h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n" );
1978
return -1;
1979
}
1980
1981
/****************************************************************************
1982
* 2 pass functions
1983
***************************************************************************/
1984
1985
/**
1986
* modify the bitrate curve from pass1 for one frame
1987
*/
1988
static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1989
{
1990
x264_ratecontrol_t *rcc= h->rc;
1991
x264_zone_t *zone = get_zone( h, frame_num );
1992
double q;
1993
if( h->param.rc.b_mb_tree )
1994
{
1995
double timescale = (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1996
q = pow( BASE_FRAME_DURATION / CLIP_DURATION(rce->i_duration * timescale), 1 - h->param.rc.f_qcompress );
1997
}
1998
else
1999
q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
2000
2001
// avoid NaN's in the rc_eq
2002
if( !isfinite(q) || rce->tex_bits + rce->mv_bits == 0 )
2003
q = rcc->last_qscale_for[rce->pict_type];
2004
else
2005
{
2006
rcc->last_rceq = q;
2007
q /= rate_factor;
2008
rcc->last_qscale = q;
2009
}
2010
2011
if( zone )
2012
{
2013
if( zone->b_force_qp )
2014
q = qp2qscale( zone->i_qp );
2015
else
2016
q /= zone->f_bitrate_factor;
2017
}
2018
2019
return q;
2020
}
2021
2022
static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q, int frame_num)
2023
{
2024
x264_ratecontrol_t *rcc = h->rc;
2025
const int pict_type = rce->pict_type;
2026
x264_zone_t *zone = get_zone( h, frame_num );
2027
2028
// force I/B quants as a function of P quants
2029
const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
2030
const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
2031
if( pict_type == SLICE_TYPE_I )
2032
{
2033
double iq = q;
2034
double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2035
double ip_factor = fabs( h->param.rc.f_ip_factor );
2036
/* don't apply ip_factor if the following frame is also I */
2037
if( rcc->accum_p_norm <= 0 )
2038
q = iq;
2039
else if( h->param.rc.f_ip_factor < 0 )
2040
q = iq / ip_factor;
2041
else if( rcc->accum_p_norm >= 1 )
2042
q = pq / ip_factor;
2043
else
2044
q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
2045
}
2046
else if( pict_type == SLICE_TYPE_B )
2047
{
2048
if( h->param.rc.f_pb_factor > 0 )
2049
q = last_non_b_q;
2050
if( !rce->kept_as_ref )
2051
q *= fabs( h->param.rc.f_pb_factor );
2052
}
2053
else if( pict_type == SLICE_TYPE_P
2054
&& rcc->last_non_b_pict_type == SLICE_TYPE_P
2055
&& rce->tex_bits == 0 )
2056
{
2057
q = last_p_q;
2058
}
2059
2060
/* last qscale / qdiff stuff */
2061
if( rcc->last_non_b_pict_type == pict_type &&
2062
(pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1) )
2063
{
2064
double last_q = rcc->last_qscale_for[pict_type];
2065
double max_qscale = last_q * rcc->lstep;
2066
double min_qscale = last_q / rcc->lstep;
2067
2068
if ( q > max_qscale ) q = max_qscale;
2069
else if( q < min_qscale ) q = min_qscale;
2070
}
2071
2072
rcc->last_qscale_for[pict_type] = q;
2073
if( pict_type != SLICE_TYPE_B )
2074
rcc->last_non_b_pict_type = pict_type;
2075
if( pict_type == SLICE_TYPE_I )
2076
{
2077
rcc->last_accum_p_norm = rcc->accum_p_norm;
2078
rcc->accum_p_norm = 0;
2079
rcc->accum_p_qp = 0;
2080
}
2081
if( pict_type == SLICE_TYPE_P )
2082
{
2083
float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
2084
rcc->accum_p_qp = mask * (qscale2qp( q ) + rcc->accum_p_qp);
2085
rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
2086
}
2087
2088
if( zone )
2089
{
2090
if( zone->b_force_qp )
2091
q = qp2qscale( zone->i_qp );
2092
else
2093
q /= zone->f_bitrate_factor;
2094
}
2095
2096
return q;
2097
}
2098
2099
static float predict_size( predictor_t *p, float q, float var )
2100
{
2101
return (p->coeff*var + p->offset) / (q*p->count);
2102
}
2103
2104
static void update_predictor( predictor_t *p, float q, float var, float bits )
2105
{
2106
float range = 1.5;
2107
if( var < 10 )
2108
return;
2109
float old_coeff = p->coeff / p->count;
2110
float old_offset = p->offset / p->count;
2111
float new_coeff = X264_MAX( (bits*q - old_offset) / var, p->coeff_min );
2112
float new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
2113
float new_offset = bits*q - new_coeff_clipped * var;
2114
if( new_offset >= 0 )
2115
new_coeff = new_coeff_clipped;
2116
else
2117
new_offset = 0;
2118
p->count *= p->decay;
2119
p->coeff *= p->decay;
2120
p->offset *= p->decay;
2121
p->count ++;
2122
p->coeff += new_coeff;
2123
p->offset += new_offset;
2124
}
2125
2126
// update VBV after encoding a frame
2127
static int update_vbv( x264_t *h, int bits )
2128
{
2129
int filler = 0;
2130
int bitrate = h->sps->vui.hrd.i_bit_rate_unscaled;
2131
x264_ratecontrol_t *rcc = h->rc;
2132
x264_ratecontrol_t *rct = h->thread[0]->rc;
2133
int64_t buffer_size = (int64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
2134
2135
if( rcc->last_satd >= h->mb.i_mb_count )
2136
update_predictor( &rct->pred[h->sh.i_type], qp2qscale( rcc->qpa_rc ), rcc->last_satd, bits );
2137
2138
if( !rcc->b_vbv )
2139
return filler;
2140
2141
uint64_t buffer_diff = (uint64_t)bits * h->sps->vui.i_time_scale;
2142
rct->buffer_fill_final -= buffer_diff;
2143
rct->buffer_fill_final_min -= buffer_diff;
2144
2145
if( rct->buffer_fill_final_min < 0 )
2146
{
2147
double underflow = (double)rct->buffer_fill_final_min / h->sps->vui.i_time_scale;
2148
if( rcc->rate_factor_max_increment && rcc->qpm >= rcc->qp_novbv + rcc->rate_factor_max_increment )
2149
x264_log( h, X264_LOG_DEBUG, "VBV underflow due to CRF-max (frame %d, %.0f bits)\n", h->i_frame, underflow );
2150
else
2151
x264_log( h, X264_LOG_WARNING, "VBV underflow (frame %d, %.0f bits)\n", h->i_frame, underflow );
2152
rct->buffer_fill_final =
2153
rct->buffer_fill_final_min = 0;
2154
}
2155
2156
if( h->param.i_avcintra_class )
2157
buffer_diff = buffer_size;
2158
else
2159
buffer_diff = (uint64_t)bitrate * h->sps->vui.i_num_units_in_tick * h->fenc->i_cpb_duration;
2160
rct->buffer_fill_final += buffer_diff;
2161
rct->buffer_fill_final_min += buffer_diff;
2162
2163
if( rct->buffer_fill_final > buffer_size )
2164
{
2165
if( h->param.rc.b_filler )
2166
{
2167
int64_t scale = (int64_t)h->sps->vui.i_time_scale * 8;
2168
filler = (rct->buffer_fill_final - buffer_size + scale - 1) / scale;
2169
bits = h->param.i_avcintra_class ? filler * 8 : X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), filler ) * 8;
2170
buffer_diff = (uint64_t)bits * h->sps->vui.i_time_scale;
2171
rct->buffer_fill_final -= buffer_diff;
2172
rct->buffer_fill_final_min -= buffer_diff;
2173
}
2174
else
2175
{
2176
rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, buffer_size );
2177
rct->buffer_fill_final_min = X264_MIN( rct->buffer_fill_final_min, buffer_size );
2178
}
2179
}
2180
2181
return filler;
2182
}
2183
2184
void x264_hrd_fullness( x264_t *h )
2185
{
2186
x264_ratecontrol_t *rct = h->thread[0]->rc;
2187
uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale / rct->hrd_multiply_denom;
2188
uint64_t cpb_state = rct->buffer_fill_final;
2189
uint64_t cpb_size = (uint64_t)h->sps->vui.hrd.i_cpb_size_unscaled * h->sps->vui.i_time_scale;
2190
uint64_t multiply_factor = 90000 / rct->hrd_multiply_denom;
2191
2192
if( rct->buffer_fill_final < 0 || rct->buffer_fill_final > (int64_t)cpb_size )
2193
{
2194
x264_log( h, X264_LOG_WARNING, "CPB %s: %.0f bits in a %.0f-bit buffer\n",
2195
rct->buffer_fill_final < 0 ? "underflow" : "overflow",
2196
(double)rct->buffer_fill_final / h->sps->vui.i_time_scale, (double)cpb_size / h->sps->vui.i_time_scale );
2197
}
2198
2199
h->initial_cpb_removal_delay = (multiply_factor * cpb_state) / denom;
2200
h->initial_cpb_removal_delay_offset = (multiply_factor * cpb_size) / denom - h->initial_cpb_removal_delay;
2201
2202
int64_t decoder_buffer_fill = h->initial_cpb_removal_delay * denom / multiply_factor;
2203
rct->buffer_fill_final_min = X264_MIN( rct->buffer_fill_final_min, decoder_buffer_fill );
2204
}
2205
2206
// provisionally update VBV according to the planned size of all frames currently in progress
2207
static void update_vbv_plan( x264_t *h, int overhead )
2208
{
2209
x264_ratecontrol_t *rcc = h->rc;
2210
rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final_min / h->sps->vui.i_time_scale;
2211
if( h->i_thread_frames > 1 )
2212
{
2213
int j = h->rc - h->thread[0]->rc;
2214
for( int i = 1; i < h->i_thread_frames; i++ )
2215
{
2216
x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
2217
double bits = t->rc->frame_size_planned;
2218
if( !t->b_thread_active )
2219
continue;
2220
bits = X264_MAX(bits, t->rc->frame_size_estimated);
2221
rcc->buffer_fill -= bits;
2222
rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
2223
rcc->buffer_fill += t->rc->buffer_rate;
2224
rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
2225
}
2226
}
2227
rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
2228
rcc->buffer_fill -= overhead;
2229
}
2230
2231
// apply VBV constraints and clip qscale to between lmin and lmax
2232
static double clip_qscale( x264_t *h, int pict_type, double q )
2233
{
2234
x264_ratecontrol_t *rcc = h->rc;
2235
double lmin = rcc->lmin[pict_type];
2236
double lmax = rcc->lmax[pict_type];
2237
if( rcc->rate_factor_max_increment )
2238
lmax = X264_MIN( lmax, qp2qscale( rcc->qp_novbv + rcc->rate_factor_max_increment ) );
2239
double q0 = q;
2240
2241
/* B-frames are not directly subject to VBV,
2242
* since they are controlled by the P-frames' QPs. */
2243
2244
if( rcc->b_vbv && rcc->last_satd > 0 )
2245
{
2246
double fenc_cpb_duration = (double)h->fenc->i_cpb_duration *
2247
h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2248
/* Lookahead VBV: raise the quantizer as necessary such that no frames in
2249
* the lookahead overflow and such that the buffer is in a reasonable state
2250
* by the end of the lookahead. */
2251
if( h->param.rc.i_lookahead )
2252
{
2253
int terminate = 0;
2254
2255
/* Avoid an infinite loop. */
2256
for( int iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
2257
{
2258
double frame_q[3];
2259
double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2260
double buffer_fill_cur = rcc->buffer_fill - cur_bits;
2261
double target_fill;
2262
double total_duration = 0;
2263
double last_duration = fenc_cpb_duration;
2264
frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
2265
frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
2266
frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
2267
2268
/* Loop over the planned future frames. */
2269
for( int j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
2270
{
2271
total_duration += last_duration;
2272
buffer_fill_cur += rcc->vbv_max_rate * last_duration;
2273
int i_type = h->fenc->i_planned_type[j];
2274
int i_satd = h->fenc->i_planned_satd[j];
2275
if( i_type == X264_TYPE_AUTO )
2276
break;
2277
i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
2278
cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
2279
buffer_fill_cur -= cur_bits;
2280
last_duration = h->fenc->f_planned_cpb_duration[j];
2281
}
2282
/* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
2283
target_fill = X264_MIN( rcc->buffer_fill + total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.5 );
2284
if( buffer_fill_cur < target_fill )
2285
{
2286
q *= 1.01;
2287
terminate |= 1;
2288
continue;
2289
}
2290
/* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
2291
target_fill = x264_clip3f( rcc->buffer_fill - total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
2292
if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
2293
{
2294
q /= 1.01;
2295
terminate |= 2;
2296
continue;
2297
}
2298
break;
2299
}
2300
}
2301
/* Fallback to old purely-reactive algorithm: no lookahead. */
2302
else
2303
{
2304
if( ( pict_type == SLICE_TYPE_P ||
2305
( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
2306
rcc->buffer_fill/rcc->buffer_size < 0.5 )
2307
{
2308
q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
2309
}
2310
2311
/* Now a hard threshold to make sure the frame fits in VBV.
2312
* This one is mostly for I-frames. */
2313
double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2314
/* For small VBVs, allow the frame to use up the entire VBV. */
2315
double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
2316
/* For single-frame VBVs, request that the frame use up the entire VBV. */
2317
double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
2318
2319
if( bits > rcc->buffer_fill/max_fill_factor )
2320
{
2321
double qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
2322
q /= qf;
2323
bits *= qf;
2324
}
2325
if( bits < rcc->buffer_rate/min_fill_factor )
2326
{
2327
double qf = x264_clip3f( bits*min_fill_factor/rcc->buffer_rate, 0.001, 1.0 );
2328
q *= qf;
2329
}
2330
q = X264_MAX( q0, q );
2331
}
2332
2333
/* Check B-frame complexity, and use up any bits that would
2334
* overflow before the next P-frame. */
2335
if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
2336
{
2337
int nb = rcc->bframes;
2338
double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2339
double pbbits = bits;
2340
double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
2341
double space;
2342
double bframe_cpb_duration = 0;
2343
double minigop_cpb_duration;
2344
for( int i = 0; i < nb; i++ )
2345
bframe_cpb_duration += h->fenc->f_planned_cpb_duration[i];
2346
2347
if( bbits * nb > bframe_cpb_duration * rcc->vbv_max_rate )
2348
nb = 0;
2349
pbbits += nb * bbits;
2350
2351
minigop_cpb_duration = bframe_cpb_duration + fenc_cpb_duration;
2352
space = rcc->buffer_fill + minigop_cpb_duration*rcc->vbv_max_rate - rcc->buffer_size;
2353
if( pbbits < space )
2354
{
2355
q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
2356
}
2357
q = X264_MAX( q0/2, q );
2358
}
2359
2360
/* Apply MinCR and buffer fill restrictions */
2361
double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2362
double frame_size_maximum = X264_MIN( rcc->frame_size_maximum, X264_MAX( rcc->buffer_fill, 0.001 ) );
2363
if( bits > frame_size_maximum )
2364
q *= bits / frame_size_maximum;
2365
2366
if( !rcc->b_vbv_min_rate )
2367
q = X264_MAX( q0, q );
2368
}
2369
2370
if( lmin==lmax )
2371
return lmin;
2372
else if( rcc->b_2pass )
2373
{
2374
double min2 = log( lmin );
2375
double max2 = log( lmax );
2376
q = (log(q) - min2)/(max2-min2) - 0.5;
2377
q = 1.0/(1.0 + exp( -4*q ));
2378
q = q*(max2-min2) + min2;
2379
return exp( q );
2380
}
2381
else
2382
return x264_clip3f( q, lmin, lmax );
2383
}
2384
2385
// update qscale for 1 frame based on actual bits used so far
2386
static float rate_estimate_qscale( x264_t *h )
2387
{
2388
float q;
2389
x264_ratecontrol_t *rcc = h->rc;
2390
ratecontrol_entry_t rce = {0};
2391
int pict_type = h->sh.i_type;
2392
int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
2393
+ h->stat.i_frame_size[SLICE_TYPE_P]
2394
+ h->stat.i_frame_size[SLICE_TYPE_B])
2395
- rcc->filler_bits_sum;
2396
2397
if( rcc->b_2pass )
2398
{
2399
rce = *rcc->rce;
2400
if( pict_type != rce.pict_type )
2401
{
2402
x264_log( h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
2403
slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type] );
2404
}
2405
}
2406
2407
if( pict_type == SLICE_TYPE_B )
2408
{
2409
/* B-frames don't have independent ratecontrol, but rather get the
2410
* average QP of the two adjacent P-frames + an offset */
2411
2412
int i0 = IS_X264_TYPE_I(h->fref_nearest[0]->i_type);
2413
int i1 = IS_X264_TYPE_I(h->fref_nearest[1]->i_type);
2414
int dt0 = abs(h->fenc->i_poc - h->fref_nearest[0]->i_poc);
2415
int dt1 = abs(h->fenc->i_poc - h->fref_nearest[1]->i_poc);
2416
float q0 = h->fref_nearest[0]->f_qp_avg_rc;
2417
float q1 = h->fref_nearest[1]->f_qp_avg_rc;
2418
2419
if( h->fref_nearest[0]->i_type == X264_TYPE_BREF )
2420
q0 -= rcc->pb_offset/2;
2421
if( h->fref_nearest[1]->i_type == X264_TYPE_BREF )
2422
q1 -= rcc->pb_offset/2;
2423
2424
if( i0 && i1 )
2425
q = (q0 + q1) / 2 + rcc->ip_offset;
2426
else if( i0 )
2427
q = q1;
2428
else if( i1 )
2429
q = q0;
2430
else
2431
q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
2432
2433
if( h->fenc->b_kept_as_ref )
2434
q += rcc->pb_offset/2;
2435
else
2436
q += rcc->pb_offset;
2437
2438
rcc->qp_novbv = q;
2439
q = qp2qscale( q );
2440
if( rcc->b_2pass )
2441
rcc->frame_size_planned = qscale2bits( &rce, q );
2442
else
2443
rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref[1][h->i_ref[1]-1]->i_satd );
2444
/* Limit planned size by MinCR */
2445
if( rcc->b_vbv )
2446
rcc->frame_size_planned = X264_MIN( rcc->frame_size_planned, rcc->frame_size_maximum );
2447
h->rc->frame_size_estimated = rcc->frame_size_planned;
2448
2449
/* For row SATDs */
2450
if( rcc->b_vbv )
2451
rcc->last_satd = x264_rc_analyse_slice( h );
2452
return q;
2453
}
2454
else
2455
{
2456
double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
2457
double predicted_bits = total_bits;
2458
if( h->i_thread_frames > 1 )
2459
{
2460
int j = h->rc - h->thread[0]->rc;
2461
for( int i = 1; i < h->i_thread_frames; i++ )
2462
{
2463
x264_t *t = h->thread[(j+i) % h->i_thread_frames];
2464
double bits = t->rc->frame_size_planned;
2465
if( !t->b_thread_active )
2466
continue;
2467
bits = X264_MAX(bits, t->rc->frame_size_estimated);
2468
predicted_bits += bits;
2469
}
2470
}
2471
2472
if( rcc->b_2pass )
2473
{
2474
double lmin = rcc->lmin[pict_type];
2475
double lmax = rcc->lmax[pict_type];
2476
double diff;
2477
2478
/* Adjust ABR buffer based on distance to the end of the video. */
2479
if( rcc->num_entries > h->i_frame )
2480
{
2481
double final_bits = rcc->entry_out[rcc->num_entries-1]->expected_bits;
2482
double video_pos = rce.expected_bits / final_bits;
2483
double scale_factor = sqrt( (1 - video_pos) * rcc->num_entries );
2484
abr_buffer *= 0.5 * X264_MAX( scale_factor, 0.5 );
2485
}
2486
2487
diff = predicted_bits - rce.expected_bits;
2488
q = rce.new_qscale;
2489
q /= x264_clip3f((abr_buffer - diff) / abr_buffer, .5, 2);
2490
if( h->i_frame >= rcc->fps && rcc->expected_bits_sum >= 1 )
2491
{
2492
/* Adjust quant based on the difference between
2493
* achieved and expected bitrate so far */
2494
double cur_time = (double)h->i_frame / rcc->num_entries;
2495
double w = x264_clip3f( cur_time*100, 0.0, 1.0 );
2496
q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
2497
}
2498
rcc->qp_novbv = qscale2qp( q );
2499
if( rcc->b_vbv )
2500
{
2501
/* Do not overflow vbv */
2502
double expected_size = qscale2bits( &rce, q );
2503
double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2504
double expected_fullness = rce.expected_vbv / rcc->buffer_size;
2505
double qmax = q*(2 - expected_fullness);
2506
double size_constraint = 1 + expected_fullness;
2507
qmax = X264_MAX( qmax, rce.new_qscale );
2508
if( expected_fullness < .05 )
2509
qmax = lmax;
2510
qmax = X264_MIN(qmax, lmax);
2511
while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
2512
((expected_vbv < 0) && (q < lmax)))
2513
{
2514
q *= 1.05;
2515
expected_size = qscale2bits(&rce, q);
2516
expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2517
}
2518
rcc->last_satd = x264_rc_analyse_slice( h );
2519
}
2520
q = x264_clip3f( q, lmin, lmax );
2521
}
2522
else /* 1pass ABR */
2523
{
2524
/* Calculate the quantizer which would have produced the desired
2525
* average bitrate if it had been applied to all frames so far.
2526
* Then modulate that quant based on the current frame's complexity
2527
* relative to the average complexity so far (using the 2pass RCEQ).
2528
* Then bias the quant up or down if total size so far was far from
2529
* the target.
2530
* Result: Depending on the value of rate_tolerance, there is a
2531
* tradeoff between quality and bitrate precision. But at large
2532
* tolerances, the bit distribution approaches that of 2pass. */
2533
2534
double wanted_bits, overflow = 1;
2535
2536
rcc->last_satd = x264_rc_analyse_slice( h );
2537
rcc->short_term_cplxsum *= 0.5;
2538
rcc->short_term_cplxcount *= 0.5;
2539
rcc->short_term_cplxsum += rcc->last_satd / (CLIP_DURATION(h->fenc->f_duration) / BASE_FRAME_DURATION);
2540
rcc->short_term_cplxcount ++;
2541
2542
rce.tex_bits = rcc->last_satd;
2543
rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
2544
rce.mv_bits = 0;
2545
rce.p_count = rcc->nmb;
2546
rce.i_count = 0;
2547
rce.s_count = 0;
2548
rce.qscale = 1;
2549
rce.pict_type = pict_type;
2550
rce.i_duration = h->fenc->i_duration;
2551
2552
if( h->param.rc.i_rc_method == X264_RC_CRF )
2553
{
2554
q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
2555
}
2556
else
2557
{
2558
q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
2559
2560
/* ABR code can potentially be counterproductive in CBR, so just don't bother.
2561
* Don't run it if the frame complexity is zero either. */
2562
if( !rcc->b_vbv_min_rate && rcc->last_satd )
2563
{
2564
// FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
2565
int i_frame_done = h->i_frame;
2566
double time_done = i_frame_done / rcc->fps;
2567
if( h->param.b_vfr_input && i_frame_done > 0 )
2568
time_done = ((double)(h->fenc->i_reordered_pts - h->i_reordered_pts_delay)) * h->param.i_timebase_num / h->param.i_timebase_den;
2569
wanted_bits = time_done * rcc->bitrate;
2570
if( wanted_bits > 0 )
2571
{
2572
abr_buffer *= X264_MAX( 1, sqrt( time_done ) );
2573
overflow = x264_clip3f( 1.0 + (predicted_bits - wanted_bits) / abr_buffer, .5, 2 );
2574
q *= overflow;
2575
}
2576
}
2577
}
2578
2579
if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
2580
/* should test _next_ pict type, but that isn't decided yet */
2581
&& rcc->last_non_b_pict_type != SLICE_TYPE_I )
2582
{
2583
q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2584
q /= fabs( h->param.rc.f_ip_factor );
2585
}
2586
else if( h->i_frame > 0 )
2587
{
2588
if( h->param.rc.i_rc_method != X264_RC_CRF )
2589
{
2590
/* Asymmetric clipping, because symmetric would prevent
2591
* overflow control in areas of rapidly oscillating complexity */
2592
double lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
2593
double lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
2594
if( overflow > 1.1 && h->i_frame > 3 )
2595
lmax *= rcc->lstep;
2596
else if( overflow < 0.9 )
2597
lmin /= rcc->lstep;
2598
2599
q = x264_clip3f(q, lmin, lmax);
2600
}
2601
}
2602
else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
2603
{
2604
q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
2605
}
2606
rcc->qp_novbv = qscale2qp( q );
2607
2608
//FIXME use get_diff_limited_q() ?
2609
q = clip_qscale( h, pict_type, q );
2610
}
2611
2612
rcc->last_qscale_for[pict_type] =
2613
rcc->last_qscale = q;
2614
2615
if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
2616
rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );
2617
2618
if( rcc->b_2pass )
2619
rcc->frame_size_planned = qscale2bits( &rce, q );
2620
else
2621
rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2622
2623
/* Always use up the whole VBV in this case. */
2624
if( rcc->single_frame_vbv )
2625
rcc->frame_size_planned = rcc->buffer_rate;
2626
/* Limit planned size by MinCR */
2627
if( rcc->b_vbv )
2628
rcc->frame_size_planned = X264_MIN( rcc->frame_size_planned, rcc->frame_size_maximum );
2629
h->rc->frame_size_estimated = rcc->frame_size_planned;
2630
return q;
2631
}
2632
}
2633
2634
static void x264_threads_normalize_predictors( x264_t *h )
2635
{
2636
double totalsize = 0;
2637
for( int i = 0; i < h->param.i_threads; i++ )
2638
totalsize += h->thread[i]->rc->slice_size_planned;
2639
double factor = h->rc->frame_size_planned / totalsize;
2640
for( int i = 0; i < h->param.i_threads; i++ )
2641
h->thread[i]->rc->slice_size_planned *= factor;
2642
}
2643
2644
void x264_threads_distribute_ratecontrol( x264_t *h )
2645
{
2646
int row;
2647
x264_ratecontrol_t *rc = h->rc;
2648
x264_emms();
2649
float qscale = qp2qscale( rc->qpm );
2650
2651
/* Initialize row predictors */
2652
if( h->i_frame == 0 )
2653
for( int i = 0; i < h->param.i_threads; i++ )
2654
{
2655
x264_t *t = h->thread[i];
2656
if( t != h )
2657
memcpy( t->rc->row_preds, rc->row_preds, sizeof(rc->row_preds) );
2658
}
2659
2660
for( int i = 0; i < h->param.i_threads; i++ )
2661
{
2662
x264_t *t = h->thread[i];
2663
if( t != h )
2664
memcpy( t->rc, rc, offsetof(x264_ratecontrol_t, row_pred) );
2665
t->rc->row_pred = t->rc->row_preds[h->sh.i_type];
2666
/* Calculate the planned slice size. */
2667
if( rc->b_vbv && rc->frame_size_planned )
2668
{
2669
int size = 0;
2670
for( row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2671
size += h->fdec->i_row_satd[row];
2672
t->rc->slice_size_planned = predict_size( &rc->pred[h->sh.i_type + (i+1)*5], qscale, size );
2673
}
2674
else
2675
t->rc->slice_size_planned = 0;
2676
}
2677
if( rc->b_vbv && rc->frame_size_planned )
2678
{
2679
x264_threads_normalize_predictors( h );
2680
2681
if( rc->single_frame_vbv )
2682
{
2683
/* Compensate for our max frame error threshold: give more bits (proportionally) to smaller slices. */
2684
for( int i = 0; i < h->param.i_threads; i++ )
2685
{
2686
x264_t *t = h->thread[i];
2687
float max_frame_error = x264_clip3f( 1.0 / (t->i_threadslice_end - t->i_threadslice_start), 0.05, 0.25 );
2688
t->rc->slice_size_planned += 2 * max_frame_error * rc->frame_size_planned;
2689
}
2690
x264_threads_normalize_predictors( h );
2691
}
2692
2693
for( int i = 0; i < h->param.i_threads; i++ )
2694
h->thread[i]->rc->frame_size_estimated = h->thread[i]->rc->slice_size_planned;
2695
}
2696
}
2697
2698
void x264_threads_merge_ratecontrol( x264_t *h )
2699
{
2700
x264_ratecontrol_t *rc = h->rc;
2701
x264_emms();
2702
2703
for( int i = 0; i < h->param.i_threads; i++ )
2704
{
2705
x264_t *t = h->thread[i];
2706
x264_ratecontrol_t *rct = h->thread[i]->rc;
2707
if( h->param.rc.i_vbv_buffer_size )
2708
{
2709
int size = 0;
2710
for( int row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2711
size += h->fdec->i_row_satd[row];
2712
int bits = t->stat.frame.i_mv_bits + t->stat.frame.i_tex_bits + t->stat.frame.i_misc_bits;
2713
int mb_count = (t->i_threadslice_end - t->i_threadslice_start) * h->mb.i_mb_width;
2714
update_predictor( &rc->pred[h->sh.i_type+(i+1)*5], qp2qscale( rct->qpa_rc/mb_count ), size, bits );
2715
}
2716
if( !i )
2717
continue;
2718
rc->qpa_rc += rct->qpa_rc;
2719
rc->qpa_aq += rct->qpa_aq;
2720
}
2721
}
2722
2723
void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
2724
{
2725
if( cur != prev )
2726
{
2727
#define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
2728
/* these vars are updated in x264_ratecontrol_start()
2729
* so copy them from the context that most recently started (prev)
2730
* to the context that's about to start (cur). */
2731
COPY(accum_p_qp);
2732
COPY(accum_p_norm);
2733
COPY(last_satd);
2734
COPY(last_rceq);
2735
COPY(last_qscale_for);
2736
COPY(last_non_b_pict_type);
2737
COPY(short_term_cplxsum);
2738
COPY(short_term_cplxcount);
2739
COPY(bframes);
2740
COPY(prev_zone);
2741
COPY(mbtree.qpbuf_pos);
2742
/* these vars can be updated by x264_ratecontrol_init_reconfigurable */
2743
COPY(bitrate);
2744
COPY(buffer_size);
2745
COPY(buffer_rate);
2746
COPY(vbv_max_rate);
2747
COPY(single_frame_vbv);
2748
COPY(cbr_decay);
2749
COPY(rate_factor_constant);
2750
COPY(rate_factor_max_increment);
2751
#undef COPY
2752
}
2753
if( cur != next )
2754
{
2755
#define COPY(var) next->rc->var = cur->rc->var
2756
/* these vars are updated in x264_ratecontrol_end()
2757
* so copy them from the context that most recently ended (cur)
2758
* to the context that's about to end (next) */
2759
COPY(cplxr_sum);
2760
COPY(expected_bits_sum);
2761
COPY(filler_bits_sum);
2762
COPY(wanted_bits_window);
2763
COPY(bframe_bits);
2764
COPY(initial_cpb_removal_delay);
2765
COPY(initial_cpb_removal_delay_offset);
2766
COPY(nrt_first_access_unit);
2767
COPY(previous_cpb_final_arrival_time);
2768
#undef COPY
2769
}
2770
//FIXME row_preds[] (not strictly necessary, but would improve prediction)
2771
/* the rest of the variables are either constant or thread-local */
2772
}
2773
2774
static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
2775
{
2776
/* find an interval ending on an overflow or underflow (depending on whether
2777
* we're adding or removing bits), and starting on the earliest frame that
2778
* can influence the buffer fill of that end frame. */
2779
x264_ratecontrol_t *rcc = h->rc;
2780
const double buffer_min = .1 * rcc->buffer_size;
2781
const double buffer_max = .9 * rcc->buffer_size;
2782
double fill = fills[*t0-1];
2783
double parity = over ? 1. : -1.;
2784
int start = -1, end = -1;
2785
for( int i = *t0; i < rcc->num_entries; i++ )
2786
{
2787
fill += (rcc->entry_out[i]->i_cpb_duration * rcc->vbv_max_rate * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale -
2788
qscale2bits( rcc->entry_out[i], rcc->entry_out[i]->new_qscale )) * parity;
2789
fill = x264_clip3f(fill, 0, rcc->buffer_size);
2790
fills[i] = fill;
2791
if( fill <= buffer_min || i == 0 )
2792
{
2793
if( end >= 0 )
2794
break;
2795
start = i;
2796
}
2797
else if( fill >= buffer_max && start >= 0 )
2798
end = i;
2799
}
2800
*t0 = start;
2801
*t1 = end;
2802
return start >= 0 && end >= 0;
2803
}
2804
2805
static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max )
2806
{
2807
x264_ratecontrol_t *rcc = h->rc;
2808
double qscale_orig, qscale_new;
2809
int adjusted = 0;
2810
if( t0 > 0 )
2811
t0++;
2812
for( int i = t0; i <= t1; i++ )
2813
{
2814
qscale_orig = rcc->entry_out[i]->new_qscale;
2815
qscale_orig = x264_clip3f( qscale_orig, qscale_min, qscale_max );
2816
qscale_new = qscale_orig * adjustment;
2817
qscale_new = x264_clip3f( qscale_new, qscale_min, qscale_max );
2818
rcc->entry_out[i]->new_qscale = qscale_new;
2819
adjusted = adjusted || (qscale_new != qscale_orig);
2820
}
2821
return adjusted;
2822
}
2823
2824
static double count_expected_bits( x264_t *h )
2825
{
2826
x264_ratecontrol_t *rcc = h->rc;
2827
double expected_bits = 0;
2828
for( int i = 0; i < rcc->num_entries; i++ )
2829
{
2830
ratecontrol_entry_t *rce = rcc->entry_out[i];
2831
rce->expected_bits = expected_bits;
2832
expected_bits += qscale2bits( rce, rce->new_qscale );
2833
}
2834
return expected_bits;
2835
}
2836
2837
static int vbv_pass2( x264_t *h, double all_available_bits )
2838
{
2839
/* for each interval of buffer_full .. underflow, uniformly increase the qp of all
2840
* frames in the interval until either buffer is full at some intermediate frame or the
2841
* last frame in the interval no longer underflows. Recompute intervals and repeat.
2842
* Then do the converse to put bits back into overflow areas until target size is met */
2843
2844
x264_ratecontrol_t *rcc = h->rc;
2845
double *fills;
2846
double expected_bits = 0;
2847
double adjustment;
2848
double prev_bits = 0;
2849
int t0, t1;
2850
double qscale_min = qp2qscale( h->param.rc.i_qp_min );
2851
double qscale_max = qp2qscale( h->param.rc.i_qp_max );
2852
int iterations = 0;
2853
int adj_min, adj_max;
2854
CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2855
2856
fills++;
2857
2858
/* adjust overall stream size */
2859
do
2860
{
2861
iterations++;
2862
prev_bits = expected_bits;
2863
2864
if( expected_bits )
2865
{ /* not first iteration */
2866
adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2867
fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2868
t0 = 0;
2869
/* fix overflows */
2870
adj_min = 1;
2871
while(adj_min && find_underflow( h, fills, &t0, &t1, 1 ))
2872
{
2873
adj_min = fix_underflow( h, t0, t1, adjustment, qscale_min, qscale_max );
2874
t0 = t1;
2875
}
2876
}
2877
2878
fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2879
t0 = 0;
2880
/* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2881
adj_max = 1;
2882
while( adj_max && find_underflow( h, fills, &t0, &t1, 0 ) )
2883
adj_max = fix_underflow( h, t0, t1, 1.001, qscale_min, qscale_max );
2884
2885
expected_bits = count_expected_bits( h );
2886
} while( (expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2887
2888
if( !adj_max )
2889
x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2890
2891
/* store expected vbv filling values for tracking when encoding */
2892
for( int i = 0; i < rcc->num_entries; i++ )
2893
rcc->entry_out[i]->expected_vbv = rcc->buffer_size - fills[i];
2894
2895
x264_free( fills-1 );
2896
return 0;
2897
fail:
2898
return -1;
2899
}
2900
2901
static int init_pass2( x264_t *h )
2902
{
2903
x264_ratecontrol_t *rcc = h->rc;
2904
uint64_t all_const_bits = 0;
2905
double timescale = (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2906
double duration = 0;
2907
for( int i = 0; i < rcc->num_entries; i++ )
2908
duration += rcc->entry[i].i_duration;
2909
duration *= timescale;
2910
uint64_t all_available_bits = h->param.rc.i_bitrate * 1000. * duration;
2911
double rate_factor, step_mult;
2912
double qblur = h->param.rc.f_qblur;
2913
double cplxblur = h->param.rc.f_complexity_blur;
2914
const int filter_size = (int)(qblur*4) | 1;
2915
double expected_bits;
2916
double *qscale, *blurred_qscale;
2917
double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
2918
2919
/* find total/average complexity & const_bits */
2920
for( int i = 0; i < rcc->num_entries; i++ )
2921
{
2922
ratecontrol_entry_t *rce = &rcc->entry[i];
2923
all_const_bits += rce->misc_bits;
2924
}
2925
2926
if( all_available_bits < all_const_bits)
2927
{
2928
x264_log( h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2929
(int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)) );
2930
return -1;
2931
}
2932
2933
/* Blur complexities, to reduce local fluctuation of QP.
2934
* We don't blur the QPs directly, because then one very simple frame
2935
* could drag down the QP of a nearby complex frame and give it more
2936
* bits than intended. */
2937
for( int i = 0; i < rcc->num_entries; i++ )
2938
{
2939
ratecontrol_entry_t *rce = &rcc->entry[i];
2940
double weight_sum = 0;
2941
double cplx_sum = 0;
2942
double weight = 1.0;
2943
double gaussian_weight;
2944
/* weighted average of cplx of future frames */
2945
for( int j = 1; j < cplxblur*2 && j < rcc->num_entries-i; j++ )
2946
{
2947
ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2948
double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2949
weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2950
if( weight < .0001 )
2951
break;
2952
gaussian_weight = weight * exp( -j*j/200.0 );
2953
weight_sum += gaussian_weight;
2954
cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2955
}
2956
/* weighted average of cplx of past frames */
2957
weight = 1.0;
2958
for( int j = 0; j <= cplxblur*2 && j <= i; j++ )
2959
{
2960
ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2961
double frame_duration = CLIP_DURATION(rcj->i_duration * timescale) / BASE_FRAME_DURATION;
2962
gaussian_weight = weight * exp( -j*j/200.0 );
2963
weight_sum += gaussian_weight;
2964
cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits) / frame_duration;
2965
weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2966
if( weight < .0001 )
2967
break;
2968
}
2969
rce->blurred_complexity = cplx_sum / weight_sum;
2970
}
2971
2972
CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2973
if( filter_size > 1 )
2974
CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2975
else
2976
blurred_qscale = qscale;
2977
2978
/* Search for a factor which, when multiplied by the RCEQ values from
2979
* each frame, adds up to the desired total size.
2980
* There is no exact closed-form solution because of VBV constraints and
2981
* because qscale2bits is not invertible, but we can start with the simple
2982
* approximation of scaling the 1st pass by the ratio of bitrates.
2983
* The search range is probably overkill, but speed doesn't matter here. */
2984
2985
expected_bits = 1;
2986
for( int i = 0; i < rcc->num_entries; i++ )
2987
{
2988
double q = get_qscale(h, &rcc->entry[i], 1.0, i);
2989
expected_bits += qscale2bits(&rcc->entry[i], q);
2990
rcc->last_qscale_for[rcc->entry[i].pict_type] = q;
2991
}
2992
step_mult = all_available_bits / expected_bits;
2993
2994
rate_factor = 0;
2995
for( double step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2996
{
2997
expected_bits = 0;
2998
rate_factor += step;
2999
3000
rcc->last_non_b_pict_type = -1;
3001
rcc->last_accum_p_norm = 1;
3002
rcc->accum_p_norm = 0;
3003
3004
rcc->last_qscale_for[0] =
3005
rcc->last_qscale_for[1] =
3006
rcc->last_qscale_for[2] = pow( base_cplx, 1 - rcc->qcompress ) / rate_factor;
3007
3008
/* find qscale */
3009
for( int i = 0; i < rcc->num_entries; i++ )
3010
{
3011
qscale[i] = get_qscale( h, &rcc->entry[i], rate_factor, -1 );
3012
rcc->last_qscale_for[rcc->entry[i].pict_type] = qscale[i];
3013
}
3014
3015
/* fixed I/B qscale relative to P */
3016
for( int i = rcc->num_entries-1; i >= 0; i-- )
3017
{
3018
qscale[i] = get_diff_limited_q( h, &rcc->entry[i], qscale[i], i );
3019
assert(qscale[i] >= 0);
3020
}
3021
3022
/* smooth curve */
3023
if( filter_size > 1 )
3024
{
3025
assert( filter_size%2 == 1 );
3026
for( int i = 0; i < rcc->num_entries; i++ )
3027
{
3028
ratecontrol_entry_t *rce = &rcc->entry[i];
3029
double q = 0.0, sum = 0.0;
3030
3031
for( int j = 0; j < filter_size; j++ )
3032
{
3033
int idx = i+j-filter_size/2;
3034
double d = idx-i;
3035
double coeff = qblur==0 ? 1.0 : exp( -d*d/(qblur*qblur) );
3036
if( idx < 0 || idx >= rcc->num_entries )
3037
continue;
3038
if( rce->pict_type != rcc->entry[idx].pict_type )
3039
continue;
3040
q += qscale[idx] * coeff;
3041
sum += coeff;
3042
}
3043
blurred_qscale[i] = q/sum;
3044
}
3045
}
3046
3047
/* find expected bits */
3048
for( int i = 0; i < rcc->num_entries; i++ )
3049
{
3050
ratecontrol_entry_t *rce = &rcc->entry[i];
3051
rce->new_qscale = clip_qscale( h, rce->pict_type, blurred_qscale[i] );
3052
assert(rce->new_qscale >= 0);
3053
expected_bits += qscale2bits( rce, rce->new_qscale );
3054
}
3055
3056
if( expected_bits > all_available_bits )
3057
rate_factor -= step;
3058
}
3059
3060
x264_free( qscale );
3061
if( filter_size > 1 )
3062
x264_free( blurred_qscale );
3063
3064
if( rcc->b_vbv )
3065
if( vbv_pass2( h, all_available_bits ) )
3066
return -1;
3067
expected_bits = count_expected_bits( h );
3068
3069
if( fabs( expected_bits/all_available_bits - 1.0 ) > 0.01 )
3070
{
3071
double avgq = 0;
3072
for( int i = 0; i < rcc->num_entries; i++ )
3073
avgq += rcc->entry[i].new_qscale;
3074
avgq = qscale2qp( avgq / rcc->num_entries );
3075
3076
if( expected_bits > all_available_bits || !rcc->b_vbv )
3077
x264_log( h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n" );
3078
x264_log( h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
3079
(float)h->param.rc.i_bitrate,
3080
expected_bits * rcc->fps / (rcc->num_entries * 1000.),
3081
avgq );
3082
if( expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2 )
3083
{
3084
if( h->param.rc.i_qp_min > 0 )
3085
x264_log( h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min );
3086
else
3087
x264_log( h, X264_LOG_WARNING, "try reducing target bitrate\n" );
3088
}
3089
else if( expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2 )
3090
{
3091
if( h->param.rc.i_qp_max < QP_MAX )
3092
x264_log( h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max );
3093
else
3094
x264_log( h, X264_LOG_WARNING, "try increasing target bitrate\n");
3095
}
3096
else if( !(rcc->b_2pass && rcc->b_vbv) )
3097
x264_log( h, X264_LOG_WARNING, "internal error\n" );
3098
}
3099
3100
return 0;
3101
fail:
3102
return -1;
3103
}
3104
3105