2 * The simplest mpeg audio layer 2 encoder
3 * Copyright (c) 2000, 2001 Gerard Lantau.
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program 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
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 #include "mpegaudio.h"
23 /* currently, cannot change these constants (need to modify
24 quantization stage) */
27 #define MUL(a,b) (((INT64)(a) * (INT64)(b)) >> FRAC_BITS)
28 #define FIX(a) ((int)((a) * (1 << FRAC_BITS)))
30 #define SAMPLES_BUF_SIZE 4096
32 typedef struct MpegAudioContext {
36 int lsf; /* 1 if mpeg2 low bitrate selected */
37 int bitrate_index; /* bit rate */
39 int frame_size; /* frame size, in bits, without padding */
40 INT64 nb_samples; /* total number of samples encoded */
41 /* padding computation */
42 int frame_frac, frame_frac_incr, do_padding;
43 short samples_buf[MPA_MAX_CHANNELS][SAMPLES_BUF_SIZE]; /* buffer for filter */
44 int samples_offset[MPA_MAX_CHANNELS]; /* offset in samples_buf */
45 int sb_samples[MPA_MAX_CHANNELS][3][12][SBLIMIT];
46 unsigned char scale_factors[MPA_MAX_CHANNELS][SBLIMIT][3]; /* scale factors */
47 /* code to group 3 scale factors */
48 unsigned char scale_code[MPA_MAX_CHANNELS][SBLIMIT];
49 int sblimit; /* number of used subbands */
50 const unsigned char *alloc_table;
53 /* define it to use floats in quantization (I don't like floats !) */
56 #include "mpegaudiotab.h"
58 int MPA_encode_init(AVCodecContext *avctx)
60 MpegAudioContext *s = avctx->priv_data;
61 int freq = avctx->sample_rate;
62 int bitrate = avctx->bit_rate;
63 int channels = avctx->channels;
69 bitrate = bitrate / 1000;
70 s->nb_channels = channels;
72 s->bit_rate = bitrate * 1000;
73 avctx->frame_size = MPA_FRAME_SIZE;
74 avctx->key_frame = 1; /* always key frame */
79 if (mpa_freq_tab[i] == freq)
81 if ((mpa_freq_tab[i] / 2) == freq) {
90 /* encoding bitrate & frequency */
92 if (mpa_bitrate_tab[s->lsf][1][i] == bitrate)
99 /* compute total header size & pad bit */
101 a = (float)(bitrate * 1000 * MPA_FRAME_SIZE) / (freq * 8.0);
102 s->frame_size = ((int)a) * 8;
104 /* frame fractional size to compute padding */
106 s->frame_frac_incr = (int)((a - floor(a)) * 65536.0);
108 /* select the right allocation table */
109 table = l2_select_table(bitrate, s->nb_channels, freq, s->lsf);
111 /* number of used subbands */
112 s->sblimit = sblimit_table[table];
113 s->alloc_table = alloc_tables[table];
116 printf("%d kb/s, %d Hz, frame_size=%d bits, table=%d, padincr=%x\n",
117 bitrate, freq, s->frame_size, table, s->frame_frac_incr);
120 for(i=0;i<s->nb_channels;i++)
121 s->samples_offset[i] = 0;
127 v = (v + (1 << (16 - WFRAC_BITS - 1))) >> (16 - WFRAC_BITS);
133 filter_bank[512 - i] = v;
137 v = (int)(pow(2.0, (3 - i) / 3.0) * (1 << 20));
140 scale_factor_table[i] = v;
142 scale_factor_inv_table[i] = pow(2.0, -(3 - i) / 3.0) / (float)(1 << 20);
145 scale_factor_shift[i] = 21 - P - (i / 3);
146 scale_factor_mult[i] = (1 << P) * pow(2.0, (i % 3) / 3.0);
161 scale_diff_table[i] = v;
170 total_quant_bits[i] = 12 * v;
176 /* 32 point floating point IDCT without 1/sqrt(2) coef zero scaling */
177 static void idct32(int *out, int *tab)
181 const int *xp = costab32;
183 for(j=31;j>=3;j-=2) tab[j] += tab[j - 2];
222 x3 = MUL(t[16], FIX(SQRT2*0.5));
226 x2 = MUL(-(t[24] + t[8]), FIX(SQRT2*0.5));
227 x1 = MUL((t[8] - x2), xp[0]);
228 x2 = MUL((t[8] + x2), xp[1]);
241 xr = MUL(t[28],xp[0]);
245 xr = MUL(t[4],xp[1]);
246 t[ 4] = (t[24] - xr);
247 t[24] = (t[24] + xr);
249 xr = MUL(t[20],xp[2]);
253 xr = MUL(t[12],xp[3]);
254 t[12] = (t[16] - xr);
255 t[16] = (t[16] + xr);
260 for (i = 0; i < 4; i++) {
261 xr = MUL(tab[30-i*4],xp[0]);
262 tab[30-i*4] = (tab[i*4] - xr);
263 tab[ i*4] = (tab[i*4] + xr);
265 xr = MUL(tab[ 2+i*4],xp[1]);
266 tab[ 2+i*4] = (tab[28-i*4] - xr);
267 tab[28-i*4] = (tab[28-i*4] + xr);
269 xr = MUL(tab[31-i*4],xp[0]);
270 tab[31-i*4] = (tab[1+i*4] - xr);
271 tab[ 1+i*4] = (tab[1+i*4] + xr);
273 xr = MUL(tab[ 3+i*4],xp[1]);
274 tab[ 3+i*4] = (tab[29-i*4] - xr);
275 tab[29-i*4] = (tab[29-i*4] + xr);
283 xr = MUL(t1[0], *xp);
292 out[i] = tab[bitinv32[i]];
296 #define WSHIFT (WFRAC_BITS + 15 - FRAC_BITS)
298 static void filter(MpegAudioContext *s, int ch, short *samples, int incr)
301 int sum, offset, i, j;
306 // print_pow1(samples, 1152);
308 offset = s->samples_offset[ch];
309 out = &s->sb_samples[ch][0][0][0];
311 /* 32 samples at once */
313 s->samples_buf[ch][offset + (31 - i)] = samples[0];
318 p = s->samples_buf[ch] + offset;
322 sum = p[0*64] * q[0*64];
323 sum += p[1*64] * q[1*64];
324 sum += p[2*64] * q[2*64];
325 sum += p[3*64] * q[3*64];
326 sum += p[4*64] * q[4*64];
327 sum += p[5*64] * q[5*64];
328 sum += p[6*64] * q[6*64];
329 sum += p[7*64] * q[7*64];
334 tmp1[0] = tmp[16] >> WSHIFT;
335 for( i=1; i<=16; i++ ) tmp1[i] = (tmp[i+16]+tmp[16-i]) >> WSHIFT;
336 for( i=17; i<=31; i++ ) tmp1[i] = (tmp[i+16]-tmp[80-i]) >> WSHIFT;
340 /* advance of 32 samples */
343 /* handle the wrap around */
345 memmove(s->samples_buf[ch] + SAMPLES_BUF_SIZE - (512 - 32),
346 s->samples_buf[ch], (512 - 32) * 2);
347 offset = SAMPLES_BUF_SIZE - 512;
350 s->samples_offset[ch] = offset;
352 // print_pow(s->sb_samples, 1152);
355 static void compute_scale_factors(unsigned char scale_code[SBLIMIT],
356 unsigned char scale_factors[SBLIMIT][3],
357 int sb_samples[3][12][SBLIMIT],
360 int *p, vmax, v, n, i, j, k, code;
362 unsigned char *sf = &scale_factors[0][0];
364 for(j=0;j<sblimit;j++) {
366 /* find the max absolute value */
367 p = &sb_samples[i][0][j];
375 /* compute the scale factor index using log 2 computations */
378 /* n is the position of the MSB of vmax. now
379 use at most 2 compares to find the index */
380 index = (21 - n) * 3 - 3;
382 while (vmax <= scale_factor_table[index+1])
385 index = 0; /* very unlikely case of overflow */
388 index = 62; /* value 63 is not allowed */
392 printf("%2d:%d in=%x %x %d\n",
393 j, i, vmax, scale_factor_table[index], index);
395 /* store the scale factor */
396 assert(index >=0 && index <= 63);
400 /* compute the transmission factor : look if the scale factors
401 are close enough to each other */
402 d1 = scale_diff_table[sf[0] - sf[1] + 64];
403 d2 = scale_diff_table[sf[1] - sf[2] + 64];
405 /* handle the 25 cases */
406 switch(d1 * 5 + d2) {
438 sf[1] = sf[2] = sf[0];
443 sf[0] = sf[1] = sf[2];
449 sf[0] = sf[2] = sf[1];
455 sf[1] = sf[2] = sf[0];
462 printf("%d: %2d %2d %2d %d %d -> %d\n", j,
463 sf[0], sf[1], sf[2], d1, d2, code);
465 scale_code[j] = code;
470 /* The most important function : psycho acoustic module. In this
471 encoder there is basically none, so this is the worst you can do,
472 but also this is the simpler. */
473 static void psycho_acoustic_model(MpegAudioContext *s, short smr[SBLIMIT])
477 for(i=0;i<s->sblimit;i++) {
478 smr[i] = (int)(fixed_smr[i] * 10);
483 #define SB_NOTALLOCATED 0
484 #define SB_ALLOCATED 1
487 /* Try to maximize the smr while using a number of bits inferior to
488 the frame size. I tried to make the code simpler, faster and
489 smaller than other encoders :-) */
490 static void compute_bit_allocation(MpegAudioContext *s,
491 short smr1[MPA_MAX_CHANNELS][SBLIMIT],
492 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT],
495 int i, ch, b, max_smr, max_ch, max_sb, current_frame_size, max_frame_size;
497 short smr[MPA_MAX_CHANNELS][SBLIMIT];
498 unsigned char subband_status[MPA_MAX_CHANNELS][SBLIMIT];
499 const unsigned char *alloc;
501 memcpy(smr, smr1, s->nb_channels * sizeof(short) * SBLIMIT);
502 memset(subband_status, SB_NOTALLOCATED, s->nb_channels * SBLIMIT);
503 memset(bit_alloc, 0, s->nb_channels * SBLIMIT);
505 /* compute frame size and padding */
506 max_frame_size = s->frame_size;
507 s->frame_frac += s->frame_frac_incr;
508 if (s->frame_frac >= 65536) {
509 s->frame_frac -= 65536;
516 /* compute the header + bit alloc size */
517 current_frame_size = 32;
518 alloc = s->alloc_table;
519 for(i=0;i<s->sblimit;i++) {
521 current_frame_size += incr * s->nb_channels;
525 /* look for the subband with the largest signal to mask ratio */
528 max_smr = 0x80000000;
529 for(ch=0;ch<s->nb_channels;ch++) {
530 for(i=0;i<s->sblimit;i++) {
531 if (smr[ch][i] > max_smr && subband_status[ch][i] != SB_NOMORE) {
532 max_smr = smr[ch][i];
539 printf("current=%d max=%d max_sb=%d alloc=%d\n",
540 current_frame_size, max_frame_size, max_sb,
546 /* find alloc table entry (XXX: not optimal, should use
548 alloc = s->alloc_table;
549 for(i=0;i<max_sb;i++) {
550 alloc += 1 << alloc[0];
553 if (subband_status[max_ch][max_sb] == SB_NOTALLOCATED) {
554 /* nothing was coded for this band: add the necessary bits */
555 incr = 2 + nb_scale_factors[s->scale_code[max_ch][max_sb]] * 6;
556 incr += total_quant_bits[alloc[1]];
558 /* increments bit allocation */
559 b = bit_alloc[max_ch][max_sb];
560 incr = total_quant_bits[alloc[b + 1]] -
561 total_quant_bits[alloc[b]];
564 if (current_frame_size + incr <= max_frame_size) {
565 /* can increase size */
566 b = ++bit_alloc[max_ch][max_sb];
567 current_frame_size += incr;
568 /* decrease smr by the resolution we added */
569 smr[max_ch][max_sb] = smr1[max_ch][max_sb] - quant_snr[alloc[b]];
570 /* max allocation size reached ? */
571 if (b == ((1 << alloc[0]) - 1))
572 subband_status[max_ch][max_sb] = SB_NOMORE;
574 subband_status[max_ch][max_sb] = SB_ALLOCATED;
576 /* cannot increase the size of this subband */
577 subband_status[max_ch][max_sb] = SB_NOMORE;
580 *padding = max_frame_size - current_frame_size;
581 assert(*padding >= 0);
584 for(i=0;i<s->sblimit;i++) {
585 printf("%d ", bit_alloc[i]);
592 * Output the mpeg audio layer 2 frame. Note how the code is small
593 * compared to other encoders :-)
595 static void encode_frame(MpegAudioContext *s,
596 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT],
599 int i, j, k, l, bit_alloc_bits, b, ch;
602 PutBitContext *p = &s->pb;
606 put_bits(p, 12, 0xfff);
607 put_bits(p, 1, 1 - s->lsf); /* 1 = mpeg1 ID, 0 = mpeg2 lsf ID */
608 put_bits(p, 2, 4-2); /* layer 2 */
609 put_bits(p, 1, 1); /* no error protection */
610 put_bits(p, 4, s->bitrate_index);
611 put_bits(p, 2, s->freq_index);
612 put_bits(p, 1, s->do_padding); /* use padding */
613 put_bits(p, 1, 0); /* private_bit */
614 put_bits(p, 2, s->nb_channels == 2 ? MPA_STEREO : MPA_MONO);
615 put_bits(p, 2, 0); /* mode_ext */
616 put_bits(p, 1, 0); /* no copyright */
617 put_bits(p, 1, 1); /* original */
618 put_bits(p, 2, 0); /* no emphasis */
622 for(i=0;i<s->sblimit;i++) {
623 bit_alloc_bits = s->alloc_table[j];
624 for(ch=0;ch<s->nb_channels;ch++) {
625 put_bits(p, bit_alloc_bits, bit_alloc[ch][i]);
627 j += 1 << bit_alloc_bits;
631 for(i=0;i<s->sblimit;i++) {
632 for(ch=0;ch<s->nb_channels;ch++) {
633 if (bit_alloc[ch][i])
634 put_bits(p, 2, s->scale_code[ch][i]);
639 for(i=0;i<s->sblimit;i++) {
640 for(ch=0;ch<s->nb_channels;ch++) {
641 if (bit_alloc[ch][i]) {
642 sf = &s->scale_factors[ch][i][0];
643 switch(s->scale_code[ch][i]) {
645 put_bits(p, 6, sf[0]);
646 put_bits(p, 6, sf[1]);
647 put_bits(p, 6, sf[2]);
651 put_bits(p, 6, sf[0]);
652 put_bits(p, 6, sf[2]);
655 put_bits(p, 6, sf[0]);
662 /* quantization & write sub band samples */
667 for(i=0;i<s->sblimit;i++) {
668 bit_alloc_bits = s->alloc_table[j];
669 for(ch=0;ch<s->nb_channels;ch++) {
670 b = bit_alloc[ch][i];
672 int qindex, steps, m, sample, bits;
673 /* we encode 3 sub band samples of the same sub band at a time */
674 qindex = s->alloc_table[j+b];
675 steps = quant_steps[qindex];
677 sample = s->sb_samples[ch][k][l + m][i];
678 /* divide by scale factor */
682 a = (float)sample * scale_factor_inv_table[s->scale_factors[ch][i][k]];
683 q[m] = (int)((a + 1.0) * steps * 0.5);
687 int q1, e, shift, mult;
688 e = s->scale_factors[ch][i][k];
689 shift = scale_factor_shift[e];
690 mult = scale_factor_mult[e];
692 /* normalize to P bits */
694 q1 = sample << (-shift);
696 q1 = sample >> shift;
697 q1 = (q1 * mult) >> P;
698 q[m] = ((q1 + (1 << P)) * steps) >> (P + 1);
703 assert(q[m] >= 0 && q[m] < steps);
705 bits = quant_bits[qindex];
707 /* group the 3 values to save bits */
709 q[0] + steps * (q[1] + steps * q[2]));
711 printf("%d: gr1 %d\n",
712 i, q[0] + steps * (q[1] + steps * q[2]));
716 printf("%d: gr3 %d %d %d\n",
717 i, q[0], q[1], q[2]);
719 put_bits(p, bits, q[0]);
720 put_bits(p, bits, q[1]);
721 put_bits(p, bits, q[2]);
725 /* next subband in alloc table */
726 j += 1 << bit_alloc_bits;
732 for(i=0;i<padding;i++)
739 int MPA_encode_frame(AVCodecContext *avctx,
740 unsigned char *frame, int buf_size, void *data)
742 MpegAudioContext *s = avctx->priv_data;
743 short *samples = data;
744 short smr[MPA_MAX_CHANNELS][SBLIMIT];
745 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT];
748 for(i=0;i<s->nb_channels;i++) {
749 filter(s, i, samples + i, s->nb_channels);
752 for(i=0;i<s->nb_channels;i++) {
753 compute_scale_factors(s->scale_code[i], s->scale_factors[i],
754 s->sb_samples[i], s->sblimit);
756 for(i=0;i<s->nb_channels;i++) {
757 psycho_acoustic_model(s, smr[i]);
759 compute_bit_allocation(s, smr, bit_alloc, &padding);
761 init_put_bits(&s->pb, frame, MPA_MAX_CODED_FRAME_SIZE, NULL, NULL);
763 encode_frame(s, bit_alloc, padding);
765 s->nb_samples += MPA_FRAME_SIZE;
766 return pbBufPtr(&s->pb) - s->pb.buf;
770 AVCodec mp2_encoder = {
774 sizeof(MpegAudioContext),