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 #ifdef CONFIG_ENCODERS
410 // Heavily modified Levinson-Durbin algorithm which
411 // copes better with quantization, and calculates the
412 // actual whitened result as it goes.
414 static void modified_levinson_durbin(int *window, int window_entries,
415 int *out, int out_entries, int channels, int *tap_quant)
418 int *state = av_mallocz(4* window_entries);
420 memcpy(state, window, 4* window_entries);
422 for (i = 0; i < out_entries; i++)
424 int step = (i+1)*channels, k, j;
425 double xx = 0.0, xy = 0.0;
427 int *x_ptr = &(window[step]), *state_ptr = &(state[0]);
428 j = window_entries - step;
429 for (;j>=0;j--,x_ptr++,state_ptr++)
431 double x_value = *x_ptr, state_value = *state_ptr;
432 xx += state_value*state_value;
433 xy += x_value*state_value;
436 for (j = 0; j <= (window_entries - step); j++);
438 double stepval = window[step+j], stateval = window[j];
439 // xx += (double)window[j]*(double)window[j];
440 // xy += (double)window[step+j]*(double)window[j];
441 xx += stateval*stateval;
442 xy += stepval*stateval;
448 k = (int)(floor(-xy/xx * (double)LATTICE_FACTOR / (double)(tap_quant[i]) + 0.5));
450 if (k > (LATTICE_FACTOR/tap_quant[i]))
451 k = LATTICE_FACTOR/tap_quant[i];
452 if (-k > (LATTICE_FACTOR/tap_quant[i]))
453 k = -(LATTICE_FACTOR/tap_quant[i]);
459 x_ptr = &(window[step]);
460 state_ptr = &(state[0]);
461 j = window_entries - step;
462 for (;j>=0;j--,x_ptr++,state_ptr++)
464 int x_value = *x_ptr, state_value = *state_ptr;
465 *x_ptr = x_value + shift_down(k*state_value,LATTICE_SHIFT);
466 *state_ptr = state_value + shift_down(k*x_value, LATTICE_SHIFT);
469 for (j=0; j <= (window_entries - step); j++)
471 int stepval = window[step+j], stateval=state[j];
472 window[step+j] += shift_down(k * stateval, LATTICE_SHIFT);
473 state[j] += shift_down(k * stepval, LATTICE_SHIFT);
480 #endif /* CONFIG_ENCODERS */
482 static int samplerate_table[] =
483 { 44100, 22050, 11025, 96000, 48000, 32000, 24000, 16000, 8000 };
485 #ifdef CONFIG_ENCODERS
487 static inline int code_samplerate(int samplerate)
491 case 44100: return 0;
492 case 22050: return 1;
493 case 11025: return 2;
494 case 96000: return 3;
495 case 48000: return 4;
496 case 32000: return 5;
497 case 24000: return 6;
498 case 16000: return 7;
504 static int sonic_encode_init(AVCodecContext *avctx)
506 SonicContext *s = avctx->priv_data;
510 if (avctx->channels > MAX_CHANNELS)
512 av_log(avctx, AV_LOG_ERROR, "Only mono and stereo streams are supported by now\n");
513 return -1; /* only stereo or mono for now */
516 if (avctx->channels == 2)
517 s->decorrelation = MID_SIDE;
519 if (avctx->codec->id == CODEC_ID_SONIC_LS)
524 s->quantization = 0.0;
530 s->quantization = 1.0;
534 if ((s->num_taps < 32) || (s->num_taps > 1024) ||
535 ((s->num_taps>>5)<<5 != s->num_taps))
537 av_log(avctx, AV_LOG_ERROR, "Invalid number of taps\n");
542 s->tap_quant = av_mallocz(4* s->num_taps);
543 for (i = 0; i < s->num_taps; i++)
544 s->tap_quant[i] = (int)(sqrt(i+1));
546 s->channels = avctx->channels;
547 s->samplerate = avctx->sample_rate;
549 s->block_align = (int)(2048.0*s->samplerate/44100)/s->downsampling;
550 s->frame_size = s->channels*s->block_align*s->downsampling;
552 s->tail = av_mallocz(4* s->num_taps*s->channels);
555 s->tail_size = s->num_taps*s->channels;
557 s->predictor_k = av_mallocz(4 * s->num_taps);
561 for (i = 0; i < s->channels; i++)
563 s->coded_samples[i] = av_mallocz(4* s->block_align);
564 if (!s->coded_samples[i])
568 s->int_samples = av_mallocz(4* s->frame_size);
570 s->window_size = ((2*s->tail_size)+s->frame_size);
571 s->window = av_mallocz(4* s->window_size);
575 avctx->extradata = av_mallocz(16);
576 if (!avctx->extradata)
578 init_put_bits(&pb, avctx->extradata, 16*8);
580 put_bits(&pb, 2, version); // version
583 put_bits(&pb, 2, s->channels);
584 put_bits(&pb, 4, code_samplerate(s->samplerate));
586 put_bits(&pb, 1, s->lossless);
588 put_bits(&pb, 3, SAMPLE_SHIFT); // XXX FIXME: sample precision
589 put_bits(&pb, 2, s->decorrelation);
590 put_bits(&pb, 2, s->downsampling);
591 put_bits(&pb, 5, (s->num_taps >> 5)-1); // 32..1024
592 put_bits(&pb, 1, 0); // XXX FIXME: no custom tap quant table
595 avctx->extradata_size = put_bits_count(&pb)/8;
597 av_log(avctx, AV_LOG_INFO, "Sonic: ver: %d ls: %d dr: %d taps: %d block: %d frame: %d downsamp: %d\n",
598 version, s->lossless, s->decorrelation, s->num_taps, s->block_align, s->frame_size, s->downsampling);
600 avctx->coded_frame = avcodec_alloc_frame();
601 if (!avctx->coded_frame)
603 avctx->coded_frame->key_frame = 1;
604 avctx->frame_size = s->block_align*s->downsampling;
609 static int sonic_encode_close(AVCodecContext *avctx)
611 SonicContext *s = avctx->priv_data;
614 av_freep(&avctx->coded_frame);
616 for (i = 0; i < s->channels; i++)
617 av_free(s->coded_samples[i]);
619 av_free(s->predictor_k);
621 av_free(s->tap_quant);
623 av_free(s->int_samples);
628 static int sonic_encode_frame(AVCodecContext *avctx,
629 uint8_t *buf, int buf_size, void *data)
631 SonicContext *s = avctx->priv_data;
633 int i, j, ch, quant = 0, x = 0;
634 short *samples = data;
636 init_put_bits(&pb, buf, buf_size*8);
639 for (i = 0; i < s->frame_size; i++)
640 s->int_samples[i] = samples[i];
643 for (i = 0; i < s->frame_size; i++)
644 s->int_samples[i] = s->int_samples[i] << SAMPLE_SHIFT;
646 switch(s->decorrelation)
649 for (i = 0; i < s->frame_size; i += s->channels)
651 s->int_samples[i] += s->int_samples[i+1];
652 s->int_samples[i+1] -= shift(s->int_samples[i], 1);
656 for (i = 0; i < s->frame_size; i += s->channels)
657 s->int_samples[i+1] -= s->int_samples[i];
660 for (i = 0; i < s->frame_size; i += s->channels)
661 s->int_samples[i] -= s->int_samples[i+1];
665 memset(s->window, 0, 4* s->window_size);
667 for (i = 0; i < s->tail_size; i++)
668 s->window[x++] = s->tail[i];
670 for (i = 0; i < s->frame_size; i++)
671 s->window[x++] = s->int_samples[i];
673 for (i = 0; i < s->tail_size; i++)
676 for (i = 0; i < s->tail_size; i++)
677 s->tail[i] = s->int_samples[s->frame_size - s->tail_size + i];
680 modified_levinson_durbin(s->window, s->window_size,
681 s->predictor_k, s->num_taps, s->channels, s->tap_quant);
682 if (intlist_write(&pb, s->predictor_k, s->num_taps, 0) < 0)
685 for (ch = 0; ch < s->channels; ch++)
688 for (i = 0; i < s->block_align; i++)
691 for (j = 0; j < s->downsampling; j++, x += s->channels)
693 s->coded_samples[ch][i] = sum;
697 // simple rate control code
700 double energy1 = 0.0, energy2 = 0.0;
701 for (ch = 0; ch < s->channels; ch++)
703 for (i = 0; i < s->block_align; i++)
705 double sample = s->coded_samples[ch][i];
706 energy2 += sample*sample;
707 energy1 += fabs(sample);
711 energy2 = sqrt(energy2/(s->channels*s->block_align));
712 energy1 = sqrt(2.0)*energy1/(s->channels*s->block_align);
714 // increase bitrate when samples are like a gaussian distribution
715 // reduce bitrate when samples are like a two-tailed exponential distribution
717 if (energy2 > energy1)
718 energy2 += (energy2-energy1)*RATE_VARIATION;
720 quant = (int)(BASE_QUANT*s->quantization*energy2/SAMPLE_FACTOR);
721 // av_log(avctx, AV_LOG_DEBUG, "quant: %d energy: %f / %f\n", quant, energy1, energy2);
728 set_ue_golomb(&pb, quant);
730 quant *= SAMPLE_FACTOR;
733 // write out coded samples
734 for (ch = 0; ch < s->channels; ch++)
737 for (i = 0; i < s->block_align; i++)
738 s->coded_samples[ch][i] = divide(s->coded_samples[ch][i], quant);
740 if (intlist_write(&pb, s->coded_samples[ch], s->block_align, 1) < 0)
744 // av_log(avctx, AV_LOG_DEBUG, "used bytes: %d\n", (put_bits_count(&pb)+7)/8);
747 return (put_bits_count(&pb)+7)/8;
749 #endif //CONFIG_ENCODERS
751 #ifdef CONFIG_DECODERS
752 static int sonic_decode_init(AVCodecContext *avctx)
754 SonicContext *s = avctx->priv_data;
758 s->channels = avctx->channels;
759 s->samplerate = avctx->sample_rate;
761 if (!avctx->extradata)
763 av_log(avctx, AV_LOG_ERROR, "No mandatory headers present\n");
767 init_get_bits(&gb, avctx->extradata, avctx->extradata_size);
769 version = get_bits(&gb, 2);
772 av_log(avctx, AV_LOG_ERROR, "Unsupported Sonic version, please report\n");
778 s->channels = get_bits(&gb, 2);
779 s->samplerate = samplerate_table[get_bits(&gb, 4)];
780 av_log(avctx, AV_LOG_INFO, "Sonicv2 chans: %d samprate: %d\n",
781 s->channels, s->samplerate);
784 if (s->channels > MAX_CHANNELS)
786 av_log(avctx, AV_LOG_ERROR, "Only mono and stereo streams are supported by now\n");
790 s->lossless = get_bits1(&gb);
792 skip_bits(&gb, 3); // XXX FIXME
793 s->decorrelation = get_bits(&gb, 2);
795 s->downsampling = get_bits(&gb, 2);
796 s->num_taps = (get_bits(&gb, 5)+1)<<5;
797 if (get_bits1(&gb)) // XXX FIXME
798 av_log(avctx, AV_LOG_INFO, "Custom quant table\n");
800 s->block_align = (int)(2048.0*(s->samplerate/44100))/s->downsampling;
801 s->frame_size = s->channels*s->block_align*s->downsampling;
802 // avctx->frame_size = s->block_align;
804 av_log(avctx, AV_LOG_INFO, "Sonic: ver: %d ls: %d dr: %d taps: %d block: %d frame: %d downsamp: %d\n",
805 version, s->lossless, s->decorrelation, s->num_taps, s->block_align, s->frame_size, s->downsampling);
808 s->tap_quant = av_mallocz(4* s->num_taps);
809 for (i = 0; i < s->num_taps; i++)
810 s->tap_quant[i] = (int)(sqrt(i+1));
812 s->predictor_k = av_mallocz(4* s->num_taps);
814 for (i = 0; i < s->channels; i++)
816 s->predictor_state[i] = av_mallocz(4* s->num_taps);
817 if (!s->predictor_state[i])
821 for (i = 0; i < s->channels; i++)
823 s->coded_samples[i] = av_mallocz(4* s->block_align);
824 if (!s->coded_samples[i])
827 s->int_samples = av_mallocz(4* s->frame_size);
832 static int sonic_decode_close(AVCodecContext *avctx)
834 SonicContext *s = avctx->priv_data;
837 av_free(s->int_samples);
838 av_free(s->tap_quant);
839 av_free(s->predictor_k);
841 for (i = 0; i < s->channels; i++)
843 av_free(s->predictor_state[i]);
844 av_free(s->coded_samples[i]);
850 static int sonic_decode_frame(AVCodecContext *avctx,
851 void *data, int *data_size,
852 uint8_t *buf, int buf_size)
854 SonicContext *s = avctx->priv_data;
857 short *samples = data;
859 if (buf_size == 0) return 0;
861 // av_log(NULL, AV_LOG_INFO, "buf_size: %d\n", buf_size);
863 init_get_bits(&gb, buf, buf_size*8);
865 intlist_read(&gb, s->predictor_k, s->num_taps, 0);
868 for (i = 0; i < s->num_taps; i++)
869 s->predictor_k[i] *= s->tap_quant[i];
874 quant = get_ue_golomb(&gb) * SAMPLE_FACTOR;
876 // av_log(NULL, AV_LOG_INFO, "quant: %d\n", quant);
878 for (ch = 0; ch < s->channels; ch++)
882 predictor_init_state(s->predictor_k, s->predictor_state[ch], s->num_taps);
884 intlist_read(&gb, s->coded_samples[ch], s->block_align, 1);
886 for (i = 0; i < s->block_align; i++)
888 for (j = 0; j < s->downsampling - 1; j++)
890 s->int_samples[x] = predictor_calc_error(s->predictor_k, s->predictor_state[ch], s->num_taps, 0);
894 s->int_samples[x] = predictor_calc_error(s->predictor_k, s->predictor_state[ch], s->num_taps, s->coded_samples[ch][i] * quant);
898 for (i = 0; i < s->num_taps; i++)
899 s->predictor_state[ch][i] = s->int_samples[s->frame_size - s->channels + ch - i*s->channels];
902 switch(s->decorrelation)
905 for (i = 0; i < s->frame_size; i += s->channels)
907 s->int_samples[i+1] += shift(s->int_samples[i], 1);
908 s->int_samples[i] -= s->int_samples[i+1];
912 for (i = 0; i < s->frame_size; i += s->channels)
913 s->int_samples[i+1] += s->int_samples[i];
916 for (i = 0; i < s->frame_size; i += s->channels)
917 s->int_samples[i] += s->int_samples[i+1];
922 for (i = 0; i < s->frame_size; i++)
923 s->int_samples[i] = shift(s->int_samples[i], SAMPLE_SHIFT);
926 for (i = 0; i < s->frame_size; i++)
928 if (s->int_samples[i] > 32767)
930 else if (s->int_samples[i] < -32768)
933 samples[i] = s->int_samples[i];
938 *data_size = s->frame_size * 2;
940 return (get_bits_count(&gb)+7)/8;
944 #ifdef CONFIG_ENCODERS
945 AVCodec sonic_encoder = {
949 sizeof(SonicContext),
956 AVCodec sonic_ls_encoder = {
960 sizeof(SonicContext),
968 #ifdef CONFIG_DECODERS
969 AVCodec sonic_decoder = {
973 sizeof(SonicContext),