1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
|
/*
* AAC encoder psychoacoustic model
* Copyright (C) 2008 Konstantin Shishkov
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
/**
* @file
* AAC encoder psychoacoustic model
*/
#include "libavutil/attributes.h"
#include "libavutil/ffmath.h"
#include "libavutil/mem.h"
#include "avcodec.h"
#include "aac.h"
#include "psymodel.h"
/***********************************
* TODOs:
* try other bitrate controlling mechanism (maybe use ratecontrol.c?)
* control quality for quality-based output
**********************************/
/**
* constants for 3GPP AAC psychoacoustic model
* @{
*/
#define PSY_3GPP_THR_SPREAD_HI 1.5f // spreading factor for low-to-hi threshold spreading (15 dB/Bark)
#define PSY_3GPP_THR_SPREAD_LOW 3.0f // spreading factor for hi-to-low threshold spreading (30 dB/Bark)
/* spreading factor for low-to-hi energy spreading, long block, > 22kbps/channel (20dB/Bark) */
#define PSY_3GPP_EN_SPREAD_HI_L1 2.0f
/* spreading factor for low-to-hi energy spreading, long block, <= 22kbps/channel (15dB/Bark) */
#define PSY_3GPP_EN_SPREAD_HI_L2 1.5f
/* spreading factor for low-to-hi energy spreading, short block (15 dB/Bark) */
#define PSY_3GPP_EN_SPREAD_HI_S 1.5f
/* spreading factor for hi-to-low energy spreading, long block (30dB/Bark) */
#define PSY_3GPP_EN_SPREAD_LOW_L 3.0f
/* spreading factor for hi-to-low energy spreading, short block (20dB/Bark) */
#define PSY_3GPP_EN_SPREAD_LOW_S 2.0f
#define PSY_3GPP_RPEMIN 0.01f
#define PSY_3GPP_RPELEV 2.0f
#define PSY_3GPP_C1 3.0f /* log2(8) */
#define PSY_3GPP_C2 1.3219281f /* log2(2.5) */
#define PSY_3GPP_C3 0.55935729f /* 1 - C2 / C1 */
#define PSY_SNR_1DB 7.9432821e-1f /* -1dB */
#define PSY_SNR_25DB 3.1622776e-3f /* -25dB */
#define PSY_3GPP_SAVE_SLOPE_L -0.46666667f
#define PSY_3GPP_SAVE_SLOPE_S -0.36363637f
#define PSY_3GPP_SAVE_ADD_L -0.84285712f
#define PSY_3GPP_SAVE_ADD_S -0.75f
#define PSY_3GPP_SPEND_SLOPE_L 0.66666669f
#define PSY_3GPP_SPEND_SLOPE_S 0.81818181f
#define PSY_3GPP_SPEND_ADD_L -0.35f
#define PSY_3GPP_SPEND_ADD_S -0.26111111f
#define PSY_3GPP_CLIP_LO_L 0.2f
#define PSY_3GPP_CLIP_LO_S 0.2f
#define PSY_3GPP_CLIP_HI_L 0.95f
#define PSY_3GPP_CLIP_HI_S 0.75f
#define PSY_3GPP_AH_THR_LONG 0.5f
#define PSY_3GPP_AH_THR_SHORT 0.63f
#define PSY_PE_FORGET_SLOPE 511
enum {
PSY_3GPP_AH_NONE,
PSY_3GPP_AH_INACTIVE,
PSY_3GPP_AH_ACTIVE
};
#define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f)
#define PSY_3GPP_PE_TO_BITS(bits) ((bits) / 1.18f)
/* LAME psy model constants */
#define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order
#define AAC_BLOCK_SIZE_LONG 1024 ///< long block size
#define AAC_BLOCK_SIZE_SHORT 128 ///< short block size
#define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence
#define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block
/**
* @}
*/
/**
* information for single band used by 3GPP TS26.403-inspired psychoacoustic model
*/
typedef struct AacPsyBand{
float energy; ///< band energy
float thr; ///< energy threshold
float thr_quiet; ///< threshold in quiet
float nz_lines; ///< number of non-zero spectral lines
float active_lines; ///< number of active spectral lines
float pe; ///< perceptual entropy
float pe_const; ///< constant part of the PE calculation
float norm_fac; ///< normalization factor for linearization
int avoid_holes; ///< hole avoidance flag
}AacPsyBand;
/**
* single/pair channel context for psychoacoustic model
*/
typedef struct AacPsyChannel{
AacPsyBand band[128]; ///< bands information
AacPsyBand prev_band[128]; ///< bands information from the previous frame
float win_energy; ///< sliding average of channel energy
float iir_state[2]; ///< hi-pass IIR filter state
uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence)
enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame
/* LAME psy model specific members */
float attack_threshold; ///< attack threshold for this channel
float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS];
int prev_attack; ///< attack value for the last short block in the previous sequence
}AacPsyChannel;
/**
* psychoacoustic model frame type-dependent coefficients
*/
typedef struct AacPsyCoeffs{
float ath; ///< absolute threshold of hearing per bands
float barks; ///< Bark value for each spectral band in long frame
float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame
float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame
float min_snr; ///< minimal SNR
}AacPsyCoeffs;
/**
* 3GPP TS26.403-inspired psychoacoustic model specific data
*/
typedef struct AacPsyContext{
int chan_bitrate; ///< bitrate per channel
int frame_bits; ///< average bits per frame
int fill_level; ///< bit reservoir fill level
struct {
float min; ///< minimum allowed PE for bit factor calculation
float max; ///< maximum allowed PE for bit factor calculation
float previous; ///< allowed PE of the previous frame
float correction; ///< PE correction factor
} pe;
AacPsyCoeffs psy_coef[2][64];
AacPsyChannel *ch;
float global_quality; ///< normalized global quality taken from avctx
}AacPsyContext;
/**
* LAME psy model preset struct
*/
typedef struct PsyLamePreset {
int quality; ///< Quality to map the rest of the vaules to.
/* This is overloaded to be both kbps per channel in ABR mode, and
* requested quality in constant quality mode.
*/
float st_lrm; ///< short threshold for L, R, and M channels
} PsyLamePreset;
/**
* LAME psy model preset table for ABR
*/
static const PsyLamePreset psy_abr_map[] = {
/* TODO: Tuning. These were taken from LAME. */
/* kbps/ch st_lrm */
{ 8, 6.60},
{ 16, 6.60},
{ 24, 6.60},
{ 32, 6.60},
{ 40, 6.60},
{ 48, 6.60},
{ 56, 6.60},
{ 64, 6.40},
{ 80, 6.00},
{ 96, 5.60},
{112, 5.20},
{128, 5.20},
{160, 5.20}
};
/**
* LAME psy model preset table for constant quality
*/
static const PsyLamePreset psy_vbr_map[] = {
/* vbr_q st_lrm */
{ 0, 4.20},
{ 1, 4.20},
{ 2, 4.20},
{ 3, 4.20},
{ 4, 4.20},
{ 5, 4.20},
{ 6, 4.20},
{ 7, 4.20},
{ 8, 4.20},
{ 9, 4.20},
{10, 4.20}
};
/**
* LAME psy model FIR coefficient table
*/
static const float psy_fir_coeffs[] = {
-8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
-3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
-5.52212e-17 * 2, -0.313819 * 2
};
/**
* Calculate the ABR attack threshold from the above LAME psymodel table.
*/
static float lame_calc_attack_threshold(int bitrate)
{
/* Assume max bitrate to start with */
int lower_range = 12, upper_range = 12;
int lower_range_kbps = psy_abr_map[12].quality;
int upper_range_kbps = psy_abr_map[12].quality;
int i;
/* Determine which bitrates the value specified falls between.
* If the loop ends without breaking our above assumption of 320kbps was correct.
*/
for (i = 1; i < 13; i++) {
if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) {
upper_range = i;
upper_range_kbps = psy_abr_map[i ].quality;
lower_range = i - 1;
lower_range_kbps = psy_abr_map[i - 1].quality;
break; /* Upper range found */
}
}
/* Determine which range the value specified is closer to */
if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps))
return psy_abr_map[lower_range].st_lrm;
return psy_abr_map[upper_range].st_lrm;
}
/**
* LAME psy model specific initialization
*/
static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx)
{
int i, j;
for (i = 0; i < avctx->ch_layout.nb_channels; i++) {
AacPsyChannel *pch = &ctx->ch[i];
if (avctx->flags & AV_CODEC_FLAG_QSCALE)
pch->attack_threshold = psy_vbr_map[av_clip(avctx->global_quality / FF_QP2LAMBDA, 0, 10)].st_lrm;
else
pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->ch_layout.nb_channels / 1000);
for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++)
pch->prev_energy_subshort[j] = 10.0f;
}
}
/**
* Calculate Bark value for given line.
*/
static av_cold float calc_bark(float f)
{
return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f));
}
#define ATH_ADD 4
/**
* Calculate ATH value for given frequency.
* Borrowed from Lame.
*/
static av_cold float ath(float f, float add)
{
f /= 1000.0f;
return 3.64 * pow(f, -0.8)
- 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4))
+ 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7))
+ (0.6 + 0.04 * add) * 0.001 * f * f * f * f;
}
static av_cold int psy_3gpp_init(FFPsyContext *ctx) {
AacPsyContext *pctx;
float bark;
int i, j, g, start;
float prev, minscale, minath, minsnr, pe_min;
int chan_bitrate = ctx->avctx->bit_rate / ((ctx->avctx->flags & AV_CODEC_FLAG_QSCALE) ? 2.0f : ctx->avctx->ch_layout.nb_channels);
const int bandwidth = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx);
const float num_bark = calc_bark((float)bandwidth);
if (bandwidth <= 0)
return AVERROR(EINVAL);
ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
if (!ctx->model_priv_data)
return AVERROR(ENOMEM);
pctx = ctx->model_priv_data;
pctx->global_quality = (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) * 0.01f;
if (ctx->avctx->flags & AV_CODEC_FLAG_QSCALE) {
/* Use the target average bitrate to compute spread parameters */
chan_bitrate = (int)(chan_bitrate / 120.0 * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120));
}
pctx->chan_bitrate = chan_bitrate;
pctx->frame_bits = FFMIN(2560, chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate);
pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
ctx->bitres.size = 6144 - pctx->frame_bits;
ctx->bitres.size -= ctx->bitres.size % 8;
pctx->fill_level = ctx->bitres.size;
minath = ath(3410 - 0.733 * ATH_ADD, ATH_ADD);
for (j = 0; j < 2; j++) {
AacPsyCoeffs *coeffs = pctx->psy_coef[j];
const uint8_t *band_sizes = ctx->bands[j];
float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) / ctx->avctx->sample_rate;
/* reference encoder uses 2.4% here instead of 60% like the spec says */
float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark;
float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L;
/* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */
float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1;
i = 0;
prev = 0.0;
for (g = 0; g < ctx->num_bands[j]; g++) {
i += band_sizes[g];
bark = calc_bark((i-1) * line_to_frequency);
coeffs[g].barks = (bark + prev) / 2.0;
prev = bark;
}
for (g = 0; g < ctx->num_bands[j] - 1; g++) {
AacPsyCoeffs *coeff = &coeffs[g];
float bark_width = coeffs[g+1].barks - coeffs->barks;
coeff->spread_low[0] = ff_exp10(-bark_width * PSY_3GPP_THR_SPREAD_LOW);
coeff->spread_hi [0] = ff_exp10(-bark_width * PSY_3GPP_THR_SPREAD_HI);
coeff->spread_low[1] = ff_exp10(-bark_width * en_spread_low);
coeff->spread_hi [1] = ff_exp10(-bark_width * en_spread_hi);
pe_min = bark_pe * bark_width;
minsnr = exp2(pe_min / band_sizes[g]) - 1.5f;
coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB);
}
start = 0;
for (g = 0; g < ctx->num_bands[j]; g++) {
minscale = ath(start * line_to_frequency, ATH_ADD);
for (i = 1; i < band_sizes[g]; i++)
minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD));
coeffs[g].ath = minscale - minath;
start += band_sizes[g];
}
}
pctx->ch = av_calloc(ctx->avctx->ch_layout.nb_channels, sizeof(*pctx->ch));
if (!pctx->ch) {
av_freep(&ctx->model_priv_data);
return AVERROR(ENOMEM);
}
lame_window_init(pctx, ctx->avctx);
return 0;
}
/**
* IIR filter used in block switching decision
*/
static float iir_filter(int in, float state[2])
{
float ret;
ret = 0.7548f * (in - state[0]) + 0.5095f * state[1];
state[0] = in;
state[1] = ret;
return ret;
}
/**
* window grouping information stored as bits (0 - new group, 1 - group continues)
*/
static const uint8_t window_grouping[9] = {
0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
};
/**
* Tell encoder which window types to use.
* @see 3GPP TS26.403 5.4.1 "Blockswitching"
*/
static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx,
const int16_t *audio,
const int16_t *la,
int channel, int prev_type)
{
int i, j;
int br = ((AacPsyContext*)ctx->model_priv_data)->chan_bitrate;
int attack_ratio = br <= 16000 ? 18 : 10;
AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
AacPsyChannel *pch = &pctx->ch[channel];
uint8_t grouping = 0;
int next_type = pch->next_window_seq;
FFPsyWindowInfo wi = { { 0 } };
if (la) {
float s[8], v;
int switch_to_eight = 0;
float sum = 0.0, sum2 = 0.0;
int attack_n = 0;
int stay_short = 0;
for (i = 0; i < 8; i++) {
for (j = 0; j < 128; j++) {
v = iir_filter(la[i*128+j], pch->iir_state);
sum += v*v;
}
s[i] = sum;
sum2 += sum;
}
for (i = 0; i < 8; i++) {
if (s[i] > pch->win_energy * attack_ratio) {
attack_n = i + 1;
switch_to_eight = 1;
break;
}
}
pch->win_energy = pch->win_energy*7/8 + sum2/64;
wi.window_type[1] = prev_type;
switch (prev_type) {
case ONLY_LONG_SEQUENCE:
wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
break;
case LONG_START_SEQUENCE:
wi.window_type[0] = EIGHT_SHORT_SEQUENCE;
grouping = pch->next_grouping;
next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
break;
case LONG_STOP_SEQUENCE:
wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
break;
case EIGHT_SHORT_SEQUENCE:
stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight;
wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0;
next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
break;
}
pch->next_grouping = window_grouping[attack_n];
pch->next_window_seq = next_type;
} else {
for (i = 0; i < 3; i++)
wi.window_type[i] = prev_type;
grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0;
}
wi.window_shape = 1;
if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
wi.num_windows = 1;
wi.grouping[0] = 1;
} else {
int lastgrp = 0;
wi.num_windows = 8;
for (i = 0; i < 8; i++) {
if (!((grouping >> i) & 1))
lastgrp = i;
wi.grouping[lastgrp]++;
}
}
return wi;
}
/* 5.6.1.2 "Calculation of Bit Demand" */
static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size,
int short_window)
{
const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L;
const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L;
const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L;
const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L;
const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L;
const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L;
float clipped_pe, bit_save, bit_spend, bit_factor, fill_level, forgetful_min_pe;
ctx->fill_level += ctx->frame_bits - bits;
ctx->fill_level = av_clip(ctx->fill_level, 0, size);
fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high);
clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max);
bit_save = (fill_level + bitsave_add) * bitsave_slope;
assert(bit_save <= 0.3f && bit_save >= -0.05000001f);
bit_spend = (fill_level + bitspend_add) * bitspend_slope;
assert(bit_spend <= 0.5f && bit_spend >= -0.1f);
/* The bit factor graph in the spec is obviously incorrect.
* bit_spend + ((bit_spend - bit_spend))...
* The reference encoder subtracts everything from 1, but also seems incorrect.
* 1 - bit_save + ((bit_spend + bit_save))...
* Hopefully below is correct.
*/
bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min);
/* NOTE: The reference encoder attempts to center pe max/min around the current pe.
* Here we do that by slowly forgetting pe.min when pe stays in a range that makes
* it unlikely (ie: above the mean)
*/
ctx->pe.max = FFMAX(pe, ctx->pe.max);
forgetful_min_pe = ((ctx->pe.min * PSY_PE_FORGET_SLOPE)
+ FFMAX(ctx->pe.min, pe * (pe / ctx->pe.max))) / (PSY_PE_FORGET_SLOPE + 1);
ctx->pe.min = FFMIN(pe, forgetful_min_pe);
/* NOTE: allocate a minimum of 1/8th average frame bits, to avoid
* reservoir starvation from producing zero-bit frames
*/
return FFMIN(
ctx->frame_bits * bit_factor,
FFMAX(ctx->frame_bits + size - bits, ctx->frame_bits / 8));
}
static float calc_pe_3gpp(AacPsyBand *band)
{
float pe, a;
band->pe = 0.0f;
band->pe_const = 0.0f;
band->active_lines = 0.0f;
if (band->energy > band->thr) {
a = log2f(band->energy);
pe = a - log2f(band->thr);
band->active_lines = band->nz_lines;
if (pe < PSY_3GPP_C1) {
pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2;
a = a * PSY_3GPP_C3 + PSY_3GPP_C2;
band->active_lines *= PSY_3GPP_C3;
}
band->pe = pe * band->nz_lines;
band->pe_const = a * band->nz_lines;
}
return band->pe;
}
static float calc_reduction_3gpp(float a, float desired_pe, float pe,
float active_lines)
{
float thr_avg, reduction;
if(active_lines == 0.0)
return 0;
thr_avg = exp2f((a - pe) / (4.0f * active_lines));
reduction = exp2f((a - desired_pe) / (4.0f * active_lines)) - thr_avg;
return FFMAX(reduction, 0.0f);
}
static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr,
float reduction)
{
float thr = band->thr;
if (band->energy > thr) {
thr = sqrtf(thr);
thr = sqrtf(thr) + reduction;
thr *= thr;
thr *= thr;
/* This deviates from the 3GPP spec to match the reference encoder.
* It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands
* that have hole avoidance on (active or inactive). It always reduces the
* threshold of bands with hole avoidance off.
*/
if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) {
thr = FFMAX(band->thr, band->energy * min_snr);
band->avoid_holes = PSY_3GPP_AH_ACTIVE;
}
}
return thr;
}
#ifndef calc_thr_3gpp
static void calc_thr_3gpp(const FFPsyWindowInfo *wi, const int num_bands, AacPsyChannel *pch,
const uint8_t *band_sizes, const float *coefs, const int cutoff)
{
int i, w, g;
int start = 0, wstart = 0;
for (w = 0; w < wi->num_windows*16; w += 16) {
wstart = 0;
for (g = 0; g < num_bands; g++) {
AacPsyBand *band = &pch->band[w+g];
float form_factor = 0.0f;
float Temp;
band->energy = 0.0f;
if (wstart < cutoff) {
for (i = 0; i < band_sizes[g]; i++) {
band->energy += coefs[start+i] * coefs[start+i];
form_factor += sqrtf(fabs(coefs[start+i]));
}
}
Temp = band->energy > 0 ? sqrtf((float)band_sizes[g] / band->energy) : 0;
band->thr = band->energy * 0.001258925f;
band->nz_lines = form_factor * sqrtf(Temp);
start += band_sizes[g];
wstart += band_sizes[g];
}
}
}
#endif /* calc_thr_3gpp */
#ifndef psy_hp_filter
static void psy_hp_filter(const float *firbuf, float *hpfsmpl, const float *psy_fir_coeffs)
{
int i, j;
for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
float sum1, sum2;
sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2];
sum2 = 0.0;
for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]);
sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]);
}
/* NOTE: The LAME psymodel expects it's input in the range -32768 to 32768.
* Tuning this for normalized floats would be difficult. */
hpfsmpl[i] = (sum1 + sum2) * 32768.0f;
}
}
#endif /* psy_hp_filter */
/**
* Calculate band thresholds as suggested in 3GPP TS26.403
*/
static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel,
const float *coefs, const FFPsyWindowInfo *wi)
{
AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
AacPsyChannel *pch = &pctx->ch[channel];
int i, w, g;
float desired_bits, desired_pe, delta_pe, reduction= NAN, spread_en[128] = {0};
float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f);
const int num_bands = ctx->num_bands[wi->num_windows == 8];
const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8];
AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG;
const int bandwidth = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx);
const int cutoff = bandwidth * 2048 / wi->num_windows / ctx->avctx->sample_rate;
//calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs, cutoff);
//modify thresholds and energies - spread, threshold in quiet, pre-echo control
for (w = 0; w < wi->num_windows*16; w += 16) {
AacPsyBand *bands = &pch->band[w];
/* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */
spread_en[0] = bands[0].energy;
for (g = 1; g < num_bands; g++) {
bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]);
spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]);
}
for (g = num_bands - 2; g >= 0; g--) {
bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]);
spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]);
}
//5.4.2.4 "Threshold in quiet"
for (g = 0; g < num_bands; g++) {
AacPsyBand *band = &bands[g];
band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath);
//5.4.2.5 "Pre-echo control"
if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (!w && wi->window_type[1] == LONG_START_SEQUENCE)))
band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr,
PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
/* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */
pe += calc_pe_3gpp(band);
a += band->pe_const;
active_lines += band->active_lines;
/* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */
if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f)
band->avoid_holes = PSY_3GPP_AH_NONE;
else
band->avoid_holes = PSY_3GPP_AH_INACTIVE;
}
}
/* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */
ctx->ch[channel].entropy = pe;
if (ctx->avctx->flags & AV_CODEC_FLAG_QSCALE) {
/* (2.5 * 120) achieves almost transparent rate, and we want to give
* ample room downwards, so we make that equivalent to QSCALE=2.4
*/
desired_pe = pe * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) / (2 * 2.5f * 120.0f);
desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe));
desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping
/* PE slope smoothing */
if (ctx->bitres.bits > 0) {
desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe));
desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping
}
pctx->pe.max = FFMAX(pe, pctx->pe.max);
pctx->pe.min = FFMIN(pe, pctx->pe.min);
} else {
desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8);
desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits);
/* NOTE: PE correction is kept simple. During initial testing it had very
* little effect on the final bitrate. Probably a good idea to come
* back and do more testing later.
*/
if (ctx->bitres.bits > 0)
desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits),
0.85f, 1.15f);
}
pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits);
ctx->bitres.alloc = desired_bits;
if (desired_pe < pe) {
/* 5.6.1.3.4 "First Estimation of the reduction value" */
for (w = 0; w < wi->num_windows*16; w += 16) {
reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines);
pe = 0.0f;
a = 0.0f;
active_lines = 0.0f;
for (g = 0; g < num_bands; g++) {
AacPsyBand *band = &pch->band[w+g];
band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
/* recalculate PE */
pe += calc_pe_3gpp(band);
a += band->pe_const;
active_lines += band->active_lines;
}
}
/* 5.6.1.3.5 "Second Estimation of the reduction value" */
for (i = 0; i < 2; i++) {
float pe_no_ah = 0.0f, desired_pe_no_ah;
active_lines = a = 0.0f;
for (w = 0; w < wi->num_windows*16; w += 16) {
for (g = 0; g < num_bands; g++) {
AacPsyBand *band = &pch->band[w+g];
if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) {
pe_no_ah += band->pe;
a += band->pe_const;
active_lines += band->active_lines;
}
}
}
desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f);
if (active_lines > 0.0f)
reduction = calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines);
pe = 0.0f;
for (w = 0; w < wi->num_windows*16; w += 16) {
for (g = 0; g < num_bands; g++) {
AacPsyBand *band = &pch->band[w+g];
if (active_lines > 0.0f)
band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
pe += calc_pe_3gpp(band);
if (band->thr > 0.0f)
band->norm_fac = band->active_lines / band->thr;
else
band->norm_fac = 0.0f;
norm_fac += band->norm_fac;
}
}
delta_pe = desired_pe - pe;
if (fabs(delta_pe) > 0.05f * desired_pe)
break;
}
if (pe < 1.15f * desired_pe) {
/* 6.6.1.3.6 "Final threshold modification by linearization" */
norm_fac = norm_fac ? 1.0f / norm_fac : 0;
for (w = 0; w < wi->num_windows*16; w += 16) {
for (g = 0; g < num_bands; g++) {
AacPsyBand *band = &pch->band[w+g];
if (band->active_lines > 0.5f) {
float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe;
float thr = band->thr;
thr *= exp2f(delta_sfb_pe / band->active_lines);
if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE)
thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy);
band->thr = thr;
}
}
}
} else {
/* 5.6.1.3.7 "Further perceptual entropy reduction" */
g = num_bands;
while (pe > desired_pe && g--) {
for (w = 0; w < wi->num_windows*16; w+= 16) {
AacPsyBand *band = &pch->band[w+g];
if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) {
coeffs[g].min_snr = PSY_SNR_1DB;
band->thr = band->energy * PSY_SNR_1DB;
pe += band->active_lines * 1.5f - band->pe;
}
}
}
/* TODO: allow more holes (unused without mid/side) */
}
}
for (w = 0; w < wi->num_windows*16; w += 16) {
for (g = 0; g < num_bands; g++) {
AacPsyBand *band = &pch->band[w+g];
FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g];
psy_band->threshold = band->thr;
psy_band->energy = band->energy;
psy_band->spread = band->active_lines * 2.0f / band_sizes[g];
psy_band->bits = PSY_3GPP_PE_TO_BITS(band->pe);
}
}
memcpy(pch->prev_band, pch->band, sizeof(pch->band));
}
static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
const float **coeffs, const FFPsyWindowInfo *wi)
{
int ch;
FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel);
for (ch = 0; ch < group->num_ch; ch++)
psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]);
}
static av_cold void psy_3gpp_end(FFPsyContext *apc)
{
AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data;
if (pctx)
av_freep(&pctx->ch);
av_freep(&apc->model_priv_data);
}
static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
{
int blocktype = ONLY_LONG_SEQUENCE;
if (uselongblock) {
if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE)
blocktype = LONG_STOP_SEQUENCE;
} else {
blocktype = EIGHT_SHORT_SEQUENCE;
if (ctx->next_window_seq == ONLY_LONG_SEQUENCE)
ctx->next_window_seq = LONG_START_SEQUENCE;
if (ctx->next_window_seq == LONG_STOP_SEQUENCE)
ctx->next_window_seq = EIGHT_SHORT_SEQUENCE;
}
wi->window_type[0] = ctx->next_window_seq;
ctx->next_window_seq = blocktype;
}
static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio,
const float *la, int channel, int prev_type)
{
AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
AacPsyChannel *pch = &pctx->ch[channel];
int grouping = 0;
int uselongblock = 1;
int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
int i;
FFPsyWindowInfo wi = { { 0 } };
if (la) {
float hpfsmpl[AAC_BLOCK_SIZE_LONG];
const float *pf = hpfsmpl;
float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN);
int att_sum = 0;
/* LAME comment: apply high pass filter of fs/4 */
psy_hp_filter(firbuf, hpfsmpl, psy_fir_coeffs);
/* Calculate the energies of each sub-shortblock */
for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) {
energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)];
assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0);
attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)];
energy_short[0] += energy_subshort[i];
}
for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) {
const float *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS);
float p = 1.0f;
for (; pf < pfe; pf++)
p = FFMAX(p, fabsf(*pf));
pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
/* NOTE: The indexes below are [i + 3 - 2] in the LAME source.
* Obviously the 3 and 2 have some significance, or this would be just [i + 1]
* (which is what we use here). What the 3 stands for is ambiguous, as it is both
* number of short blocks, and the number of sub-short blocks.
* It seems that LAME is comparing each sub-block to sub-block + 1 in the
* previous block.
*/
if (p > energy_subshort[i + 1])
p = p / energy_subshort[i + 1];
else if (energy_subshort[i + 1] > p * 10.0f)
p = energy_subshort[i + 1] / (p * 10.0f);
else
p = 0.0;
attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
}
/* compare energy between sub-short blocks */
for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++)
if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS])
if (attack_intensity[i] > pch->attack_threshold)
attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1;
/* should have energy change between short blocks, in order to avoid periodic signals */
/* Good samples to show the effect are Trumpet test songs */
/* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */
/* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */
for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) {
const float u = energy_short[i - 1];
const float v = energy_short[i];
const float m = FFMAX(u, v);
if (m < 40000) { /* (2) */
if (u < 1.7f * v && v < 1.7f * u) { /* (1) */
if (i == 1 && attacks[0] < attacks[i])
attacks[0] = 0;
attacks[i] = 0;
}
}
att_sum += attacks[i];
}
if (attacks[0] <= pch->prev_attack)
attacks[0] = 0;
att_sum += attacks[0];
/* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */
if (pch->prev_attack == 3 || att_sum) {
uselongblock = 0;
for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++)
if (attacks[i] && attacks[i-1])
attacks[i] = 0;
}
} else {
/* We have no lookahead info, so just use same type as the previous sequence. */
uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE);
}
lame_apply_block_type(pch, &wi, uselongblock);
wi.window_type[1] = prev_type;
if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
wi.num_windows = 1;
wi.grouping[0] = 1;
if (wi.window_type[0] == LONG_START_SEQUENCE)
wi.window_shape = 0;
else
wi.window_shape = 1;
} else {
int lastgrp = 0;
wi.num_windows = 8;
wi.window_shape = 0;
for (i = 0; i < 8; i++) {
if (!((pch->next_grouping >> i) & 1))
lastgrp = i;
wi.grouping[lastgrp]++;
}
}
/* Determine grouping, based on the location of the first attack, and save for
* the next frame.
* FIXME: Move this to analysis.
* TODO: Tune groupings depending on attack location
* TODO: Handle more than one attack in a group
*/
for (i = 0; i < 9; i++) {
if (attacks[i]) {
grouping = i;
break;
}
}
pch->next_grouping = window_grouping[grouping];
pch->prev_attack = attacks[8];
return wi;
}
const FFPsyModel ff_aac_psy_model =
{
.name = "3GPP TS 26.403-inspired model",
.init = psy_3gpp_init,
.window = psy_lame_window,
.analyze = psy_3gpp_analyze,
.end = psy_3gpp_end,
};
|