2
* Simple free lossless/lossy audio codec
3
* Copyright (c) 2004 Alex Beregszaszi
5
* This library is free software; you can redistribute it and/or
6
* modify it under the terms of the GNU Lesser General Public
7
* License as published by the Free Software Foundation; either
8
* version 2 of the License, or (at your option) any later version.
10
* This library is distributed in the hope that it will be useful,
11
* but WITHOUT ANY WARRANTY; without even the implied warranty of
12
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13
* Lesser General Public License for more details.
15
* You should have received a copy of the GNU Lesser General Public
16
* License along with this library; if not, write to the Free Software
17
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20
#include "bitstream.h"
25
* Simple free lossless/lossy audio codec
26
* Based on Paul Francis Harrison's Bonk (http://www.logarithmic.net/pfh/bonk)
27
* Written and designed by Alex Beregszaszi
30
* - CABAC put/get_symbol
31
* - independent quantizer for channels
32
* - >2 channels support
33
* - more decorrelation types
34
* - more tap_quant tests
35
* - selectable intlist writers/readers (bonk-style, golomb, cabac)
38
#define MAX_CHANNELS 2
44
typedef struct SonicContext {
45
int lossless, decorrelation;
47
int num_taps, downsampling;
50
int channels, samplerate, block_align, frame_size;
54
int *coded_samples[MAX_CHANNELS];
64
int *predictor_state[MAX_CHANNELS];
67
#define LATTICE_SHIFT 10
68
#define SAMPLE_SHIFT 4
69
#define LATTICE_FACTOR (1 << LATTICE_SHIFT)
70
#define SAMPLE_FACTOR (1 << SAMPLE_SHIFT)
72
#define BASE_QUANT 0.6
73
#define RATE_VARIATION 3.0
75
static inline int divide(int a, int b)
78
return -( (-a + b/2)/b );
83
static inline int shift(int a,int b)
85
return (a+(1<<(b-1))) >> b;
88
static inline int shift_down(int a,int b)
90
return (a>>b)+((a<0)?1:0);
94
static inline int intlist_write(PutBitContext *pb, int *buf, int entries, int base_2_part)
98
for (i = 0; i < entries; i++)
99
set_se_golomb(pb, buf[i]);
104
static inline int intlist_read(GetBitContext *gb, int *buf, int entries, int base_2_part)
108
for (i = 0; i < entries; i++)
109
buf[i] = get_se_golomb(gb);
116
#define ADAPT_LEVEL 8
118
static int bits_to_store(uint64_t x)
130
static void write_uint_max(PutBitContext *pb, unsigned int value, unsigned int max)
137
bits = bits_to_store(max);
139
for (i = 0; i < bits-1; i++)
140
put_bits(pb, 1, value & (1 << i));
142
if ( (value | (1 << (bits-1))) <= max)
143
put_bits(pb, 1, value & (1 << (bits-1)));
146
static unsigned int read_uint_max(GetBitContext *gb, int max)
148
int i, bits, value = 0;
153
bits = bits_to_store(max);
155
for (i = 0; i < bits-1; i++)
159
if ( (value | (1<<(bits-1))) <= max)
161
value += 1 << (bits-1);
166
static int intlist_write(PutBitContext *pb, int *buf, int entries, int base_2_part)
168
int i, j, x = 0, low_bits = 0, max = 0;
169
int step = 256, pos = 0, dominant = 0, any = 0;
172
copy = av_mallocz(4* entries);
180
for (i = 0; i < entries; i++)
181
energy += abs(buf[i]);
183
low_bits = bits_to_store(energy / (entries * 2));
187
put_bits(pb, 4, low_bits);
190
for (i = 0; i < entries; i++)
192
put_bits(pb, low_bits, abs(buf[i]));
193
copy[i] = abs(buf[i]) >> low_bits;
198
bits = av_mallocz(4* entries*max);
205
for (i = 0; i <= max; i++)
207
for (j = 0; j < entries; j++)
209
bits[x++] = copy[j] > i;
215
int steplet = step >> 8;
217
if (pos + steplet > x)
220
for (i = 0; i < steplet; i++)
221
if (bits[i+pos] != dominant)
224
put_bits(pb, 1, any);
229
step += step / ADAPT_LEVEL;
235
while (((pos + interloper) < x) && (bits[pos + interloper] == dominant))
239
write_uint_max(pb, interloper, (step >> 8) - 1);
241
pos += interloper + 1;
242
step -= step / ADAPT_LEVEL;
248
dominant = !dominant;
253
for (i = 0; i < entries; i++)
255
put_bits(pb, 1, buf[i] < 0);
263
static int intlist_read(GetBitContext *gb, int *buf, int entries, int base_2_part)
265
int i, low_bits = 0, x = 0;
266
int n_zeros = 0, step = 256, dominant = 0;
267
int pos = 0, level = 0;
268
int *bits = av_mallocz(4* entries);
275
low_bits = get_bits(gb, 4);
278
for (i = 0; i < entries; i++)
279
buf[i] = get_bits(gb, low_bits);
282
// av_log(NULL, AV_LOG_INFO, "entries: %d, low bits: %d\n", entries, low_bits);
284
while (n_zeros < entries)
286
int steplet = step >> 8;
290
for (i = 0; i < steplet; i++)
291
bits[x++] = dominant;
296
step += step / ADAPT_LEVEL;
300
int actual_run = read_uint_max(gb, steplet-1);
302
// av_log(NULL, AV_LOG_INFO, "actual run: %d\n", actual_run);
304
for (i = 0; i < actual_run; i++)
305
bits[x++] = dominant;
307
bits[x++] = !dominant;
310
n_zeros += actual_run;
314
step -= step / ADAPT_LEVEL;
320
dominant = !dominant;
324
// reconstruct unsigned values
326
for (i = 0; n_zeros < entries; i++)
333
level += 1 << low_bits;
336
if (buf[pos] >= level)
343
buf[pos] += 1 << low_bits;
352
for (i = 0; i < entries; i++)
353
if (buf[i] && get_bits1(gb))
356
// av_log(NULL, AV_LOG_INFO, "zeros: %d pos: %d\n", n_zeros, pos);
362
static void predictor_init_state(int *k, int *state, int order)
366
for (i = order-2; i >= 0; i--)
368
int j, p, x = state[i];
370
for (j = 0, p = i+1; p < order; j++,p++)
372
int tmp = x + shift_down(k[j] * state[p], LATTICE_SHIFT);
373
state[p] += shift_down(k[j]*x, LATTICE_SHIFT);
379
static int predictor_calc_error(int *k, int *state, int order, int error)
381
int i, x = error - shift_down(k[order-1] * state[order-1], LATTICE_SHIFT);
384
int *k_ptr = &(k[order-2]),
385
*state_ptr = &(state[order-2]);
386
for (i = order-2; i >= 0; i--, k_ptr--, state_ptr--)
388
int k_value = *k_ptr, state_value = *state_ptr;
389
x -= shift_down(k_value * state_value, LATTICE_SHIFT);
390
state_ptr[1] = state_value + shift_down(k_value * x, LATTICE_SHIFT);
393
for (i = order-2; i >= 0; i--)
395
x -= shift_down(k[i] * state[i], LATTICE_SHIFT);
396
state[i+1] = state[i] + shift_down(k[i] * x, LATTICE_SHIFT);
400
// don't drift too far, to avoid overflows
401
if (x > (SAMPLE_FACTOR<<16)) x = (SAMPLE_FACTOR<<16);
402
if (x < -(SAMPLE_FACTOR<<16)) x = -(SAMPLE_FACTOR<<16);
409
// Heavily modified Levinson-Durbin algorithm which
410
// copes better with quantization, and calculates the
411
// actual whitened result as it goes.
413
static void modified_levinson_durbin(int *window, int window_entries,
414
int *out, int out_entries, int channels, int *tap_quant)
417
int *state = av_mallocz(4* window_entries);
419
memcpy(state, window, 4* window_entries);
421
for (i = 0; i < out_entries; i++)
423
int step = (i+1)*channels, k, j;
424
double xx = 0.0, xy = 0.0;
426
int *x_ptr = &(window[step]), *state_ptr = &(state[0]);
427
j = window_entries - step;
428
for (;j>=0;j--,x_ptr++,state_ptr++)
430
double x_value = *x_ptr, state_value = *state_ptr;
431
xx += state_value*state_value;
432
xy += x_value*state_value;
435
for (j = 0; j <= (window_entries - step); j++);
437
double stepval = window[step+j], stateval = window[j];
438
// xx += (double)window[j]*(double)window[j];
439
// xy += (double)window[step+j]*(double)window[j];
440
xx += stateval*stateval;
441
xy += stepval*stateval;
447
k = (int)(floor(-xy/xx * (double)LATTICE_FACTOR / (double)(tap_quant[i]) + 0.5));
449
if (k > (LATTICE_FACTOR/tap_quant[i]))
450
k = LATTICE_FACTOR/tap_quant[i];
451
if (-k > (LATTICE_FACTOR/tap_quant[i]))
452
k = -(LATTICE_FACTOR/tap_quant[i]);
458
x_ptr = &(window[step]);
459
state_ptr = &(state[0]);
460
j = window_entries - step;
461
for (;j>=0;j--,x_ptr++,state_ptr++)
463
int x_value = *x_ptr, state_value = *state_ptr;
464
*x_ptr = x_value + shift_down(k*state_value,LATTICE_SHIFT);
465
*state_ptr = state_value + shift_down(k*x_value, LATTICE_SHIFT);
468
for (j=0; j <= (window_entries - step); j++)
470
int stepval = window[step+j], stateval=state[j];
471
window[step+j] += shift_down(k * stateval, LATTICE_SHIFT);
472
state[j] += shift_down(k * stepval, LATTICE_SHIFT);
480
static int samplerate_table[] =
481
{ 44100, 22050, 11025, 96000, 48000, 32000, 24000, 16000, 8000 };
483
#ifdef CONFIG_ENCODERS
485
static inline int code_samplerate(int samplerate)
489
case 44100: return 0;
490
case 22050: return 1;
491
case 11025: return 2;
492
case 96000: return 3;
493
case 48000: return 4;
494
case 32000: return 5;
495
case 24000: return 6;
496
case 16000: return 7;
502
static int sonic_encode_init(AVCodecContext *avctx)
504
SonicContext *s = avctx->priv_data;
508
if (avctx->channels > MAX_CHANNELS)
510
av_log(avctx, AV_LOG_ERROR, "Only mono and stereo streams are supported by now\n");
511
return -1; /* only stereo or mono for now */
514
if (avctx->channels == 2)
515
s->decorrelation = MID_SIDE;
517
if (avctx->codec->id == CODEC_ID_SONIC_LS)
522
s->quantization = 0.0;
528
s->quantization = 1.0;
532
if ((s->num_taps < 32) || (s->num_taps > 1024) ||
533
((s->num_taps>>5)<<5 != s->num_taps))
535
av_log(avctx, AV_LOG_ERROR, "Invalid number of taps\n");
540
s->tap_quant = av_mallocz(4* s->num_taps);
541
for (i = 0; i < s->num_taps; i++)
542
s->tap_quant[i] = (int)(sqrt(i+1));
544
s->channels = avctx->channels;
545
s->samplerate = avctx->sample_rate;
547
s->block_align = (int)(2048.0*s->samplerate/44100)/s->downsampling;
548
s->frame_size = s->channels*s->block_align*s->downsampling;
550
s->tail = av_mallocz(4* s->num_taps*s->channels);
553
s->tail_size = s->num_taps*s->channels;
555
s->predictor_k = av_mallocz(4 * s->num_taps);
559
for (i = 0; i < s->channels; i++)
561
s->coded_samples[i] = av_mallocz(4* s->block_align);
562
if (!s->coded_samples[i])
566
s->int_samples = av_mallocz(4* s->frame_size);
568
s->window_size = ((2*s->tail_size)+s->frame_size);
569
s->window = av_mallocz(4* s->window_size);
573
avctx->extradata = av_mallocz(16);
574
if (!avctx->extradata)
576
init_put_bits(&pb, avctx->extradata, 16*8);
578
put_bits(&pb, 2, version); // version
581
put_bits(&pb, 2, s->channels);
582
put_bits(&pb, 4, code_samplerate(s->samplerate));
584
put_bits(&pb, 1, s->lossless);
586
put_bits(&pb, 3, SAMPLE_SHIFT); // XXX FIXME: sample precision
587
put_bits(&pb, 2, s->decorrelation);
588
put_bits(&pb, 2, s->downsampling);
589
put_bits(&pb, 5, (s->num_taps >> 5)-1); // 32..1024
590
put_bits(&pb, 1, 0); // XXX FIXME: no custom tap quant table
593
avctx->extradata_size = put_bits_count(&pb)/8;
595
av_log(avctx, AV_LOG_INFO, "Sonic: ver: %d ls: %d dr: %d taps: %d block: %d frame: %d downsamp: %d\n",
596
version, s->lossless, s->decorrelation, s->num_taps, s->block_align, s->frame_size, s->downsampling);
598
avctx->coded_frame = avcodec_alloc_frame();
599
if (!avctx->coded_frame)
601
avctx->coded_frame->key_frame = 1;
602
avctx->frame_size = s->block_align*s->downsampling;
607
static int sonic_encode_close(AVCodecContext *avctx)
609
SonicContext *s = avctx->priv_data;
612
av_freep(&avctx->coded_frame);
614
for (i = 0; i < s->channels; i++)
615
av_free(s->coded_samples[i]);
617
av_free(s->predictor_k);
619
av_free(s->tap_quant);
621
av_free(s->int_samples);
626
static int sonic_encode_frame(AVCodecContext *avctx,
627
uint8_t *buf, int buf_size, void *data)
629
SonicContext *s = avctx->priv_data;
631
int i, j, ch, quant = 0, x = 0;
632
short *samples = data;
634
init_put_bits(&pb, buf, buf_size*8);
637
for (i = 0; i < s->frame_size; i++)
638
s->int_samples[i] = samples[i];
641
for (i = 0; i < s->frame_size; i++)
642
s->int_samples[i] = s->int_samples[i] << SAMPLE_SHIFT;
644
switch(s->decorrelation)
647
for (i = 0; i < s->frame_size; i += s->channels)
649
s->int_samples[i] += s->int_samples[i+1];
650
s->int_samples[i+1] -= shift(s->int_samples[i], 1);
654
for (i = 0; i < s->frame_size; i += s->channels)
655
s->int_samples[i+1] -= s->int_samples[i];
658
for (i = 0; i < s->frame_size; i += s->channels)
659
s->int_samples[i] -= s->int_samples[i+1];
663
memset(s->window, 0, 4* s->window_size);
665
for (i = 0; i < s->tail_size; i++)
666
s->window[x++] = s->tail[i];
668
for (i = 0; i < s->frame_size; i++)
669
s->window[x++] = s->int_samples[i];
671
for (i = 0; i < s->tail_size; i++)
674
for (i = 0; i < s->tail_size; i++)
675
s->tail[i] = s->int_samples[s->frame_size - s->tail_size + i];
678
modified_levinson_durbin(s->window, s->window_size,
679
s->predictor_k, s->num_taps, s->channels, s->tap_quant);
680
if (intlist_write(&pb, s->predictor_k, s->num_taps, 0) < 0)
683
for (ch = 0; ch < s->channels; ch++)
686
for (i = 0; i < s->block_align; i++)
689
for (j = 0; j < s->downsampling; j++, x += s->channels)
691
s->coded_samples[ch][i] = sum;
695
// simple rate control code
698
double energy1 = 0.0, energy2 = 0.0;
699
for (ch = 0; ch < s->channels; ch++)
701
for (i = 0; i < s->block_align; i++)
703
double sample = s->coded_samples[ch][i];
704
energy2 += sample*sample;
705
energy1 += fabs(sample);
709
energy2 = sqrt(energy2/(s->channels*s->block_align));
710
energy1 = sqrt(2.0)*energy1/(s->channels*s->block_align);
712
// increase bitrate when samples are like a gaussian distribution
713
// reduce bitrate when samples are like a two-tailed exponential distribution
715
if (energy2 > energy1)
716
energy2 += (energy2-energy1)*RATE_VARIATION;
718
quant = (int)(BASE_QUANT*s->quantization*energy2/SAMPLE_FACTOR);
719
// av_log(avctx, AV_LOG_DEBUG, "quant: %d energy: %f / %f\n", quant, energy1, energy2);
726
set_ue_golomb(&pb, quant);
728
quant *= SAMPLE_FACTOR;
731
// write out coded samples
732
for (ch = 0; ch < s->channels; ch++)
735
for (i = 0; i < s->block_align; i++)
736
s->coded_samples[ch][i] = divide(s->coded_samples[ch][i], quant);
738
if (intlist_write(&pb, s->coded_samples[ch], s->block_align, 1) < 0)
742
// av_log(avctx, AV_LOG_DEBUG, "used bytes: %d\n", (put_bits_count(&pb)+7)/8);
745
return (put_bits_count(&pb)+7)/8;
747
#endif //CONFIG_ENCODERS
749
static int sonic_decode_init(AVCodecContext *avctx)
751
SonicContext *s = avctx->priv_data;
755
s->channels = avctx->channels;
756
s->samplerate = avctx->sample_rate;
758
if (!avctx->extradata)
760
av_log(avctx, AV_LOG_ERROR, "No mandatory headers present\n");
764
init_get_bits(&gb, avctx->extradata, avctx->extradata_size);
766
version = get_bits(&gb, 2);
769
av_log(avctx, AV_LOG_ERROR, "Unsupported Sonic version, please report\n");
775
s->channels = get_bits(&gb, 2);
776
s->samplerate = samplerate_table[get_bits(&gb, 4)];
777
av_log(avctx, AV_LOG_INFO, "Sonicv2 chans: %d samprate: %d\n",
778
s->channels, s->samplerate);
781
if (s->channels > MAX_CHANNELS)
783
av_log(avctx, AV_LOG_ERROR, "Only mono and stereo streams are supported by now\n");
787
s->lossless = get_bits1(&gb);
789
skip_bits(&gb, 3); // XXX FIXME
790
s->decorrelation = get_bits(&gb, 2);
792
s->downsampling = get_bits(&gb, 2);
793
s->num_taps = (get_bits(&gb, 5)+1)<<5;
794
if (get_bits1(&gb)) // XXX FIXME
795
av_log(avctx, AV_LOG_INFO, "Custom quant table\n");
797
s->block_align = (int)(2048.0*(s->samplerate/44100))/s->downsampling;
798
s->frame_size = s->channels*s->block_align*s->downsampling;
799
// avctx->frame_size = s->block_align;
801
av_log(avctx, AV_LOG_INFO, "Sonic: ver: %d ls: %d dr: %d taps: %d block: %d frame: %d downsamp: %d\n",
802
version, s->lossless, s->decorrelation, s->num_taps, s->block_align, s->frame_size, s->downsampling);
805
s->tap_quant = av_mallocz(4* s->num_taps);
806
for (i = 0; i < s->num_taps; i++)
807
s->tap_quant[i] = (int)(sqrt(i+1));
809
s->predictor_k = av_mallocz(4* s->num_taps);
811
for (i = 0; i < s->channels; i++)
813
s->predictor_state[i] = av_mallocz(4* s->num_taps);
814
if (!s->predictor_state[i])
818
for (i = 0; i < s->channels; i++)
820
s->coded_samples[i] = av_mallocz(4* s->block_align);
821
if (!s->coded_samples[i])
824
s->int_samples = av_mallocz(4* s->frame_size);
829
static int sonic_decode_close(AVCodecContext *avctx)
831
SonicContext *s = avctx->priv_data;
834
av_free(s->int_samples);
835
av_free(s->tap_quant);
836
av_free(s->predictor_k);
838
for (i = 0; i < s->channels; i++)
840
av_free(s->predictor_state[i]);
841
av_free(s->coded_samples[i]);
847
static int sonic_decode_frame(AVCodecContext *avctx,
848
void *data, int *data_size,
849
uint8_t *buf, int buf_size)
851
SonicContext *s = avctx->priv_data;
854
short *samples = data;
856
if (buf_size == 0) return 0;
858
// av_log(NULL, AV_LOG_INFO, "buf_size: %d\n", buf_size);
860
init_get_bits(&gb, buf, buf_size*8);
862
intlist_read(&gb, s->predictor_k, s->num_taps, 0);
865
for (i = 0; i < s->num_taps; i++)
866
s->predictor_k[i] *= s->tap_quant[i];
871
quant = get_ue_golomb(&gb) * SAMPLE_FACTOR;
873
// av_log(NULL, AV_LOG_INFO, "quant: %d\n", quant);
875
for (ch = 0; ch < s->channels; ch++)
879
predictor_init_state(s->predictor_k, s->predictor_state[ch], s->num_taps);
881
intlist_read(&gb, s->coded_samples[ch], s->block_align, 1);
883
for (i = 0; i < s->block_align; i++)
885
for (j = 0; j < s->downsampling - 1; j++)
887
s->int_samples[x] = predictor_calc_error(s->predictor_k, s->predictor_state[ch], s->num_taps, 0);
891
s->int_samples[x] = predictor_calc_error(s->predictor_k, s->predictor_state[ch], s->num_taps, s->coded_samples[ch][i] * quant);
895
for (i = 0; i < s->num_taps; i++)
896
s->predictor_state[ch][i] = s->int_samples[s->frame_size - s->channels + ch - i*s->channels];
899
switch(s->decorrelation)
902
for (i = 0; i < s->frame_size; i += s->channels)
904
s->int_samples[i+1] += shift(s->int_samples[i], 1);
905
s->int_samples[i] -= s->int_samples[i+1];
909
for (i = 0; i < s->frame_size; i += s->channels)
910
s->int_samples[i+1] += s->int_samples[i];
913
for (i = 0; i < s->frame_size; i += s->channels)
914
s->int_samples[i] += s->int_samples[i+1];
919
for (i = 0; i < s->frame_size; i++)
920
s->int_samples[i] = shift(s->int_samples[i], SAMPLE_SHIFT);
923
for (i = 0; i < s->frame_size; i++)
925
if (s->int_samples[i] > 32767)
927
else if (s->int_samples[i] < -32768)
930
samples[i] = s->int_samples[i];
935
*data_size = s->frame_size * 2;
937
return (get_bits_count(&gb)+7)/8;
940
#ifdef CONFIG_ENCODERS
941
AVCodec sonic_encoder = {
945
sizeof(SonicContext),
952
AVCodec sonic_ls_encoder = {
956
sizeof(SonicContext),
964
#ifdef CONFIG_DECODERS
965
AVCodec sonic_decoder = {
969
sizeof(SonicContext),