2 * The simplest mpeg audio layer 2 encoder
3 * Copyright (c) 2000, 2001 Fabrice Bellard.
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., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 * The simplest mpeg audio layer 2 encoder.
26 #include "mpegaudio.h"
28 /* currently, cannot change these constants (need to modify
29 quantization stage) */
32 #define MUL(a,b) (((int64_t)(a) * (int64_t)(b)) >> FRAC_BITS)
33 #define FIX(a) ((int)((a) * (1 << FRAC_BITS)))
35 #define SAMPLES_BUF_SIZE 4096
37 typedef struct MpegAudioContext {
41 int lsf; /* 1 if mpeg2 low bitrate selected */
42 int bitrate_index; /* bit rate */
44 int frame_size; /* frame size, in bits, without padding */
45 int64_t nb_samples; /* total number of samples encoded */
46 /* padding computation */
47 int frame_frac, frame_frac_incr, do_padding;
48 short samples_buf[MPA_MAX_CHANNELS][SAMPLES_BUF_SIZE]; /* buffer for filter */
49 int samples_offset[MPA_MAX_CHANNELS]; /* offset in samples_buf */
50 int sb_samples[MPA_MAX_CHANNELS][3][12][SBLIMIT];
51 unsigned char scale_factors[MPA_MAX_CHANNELS][SBLIMIT][3]; /* scale factors */
52 /* code to group 3 scale factors */
53 unsigned char scale_code[MPA_MAX_CHANNELS][SBLIMIT];
54 int sblimit; /* number of used subbands */
55 const unsigned char *alloc_table;
58 /* define it to use floats in quantization (I don't like floats !) */
61 #include "mpegaudiotab.h"
63 static int MPA_encode_init(AVCodecContext *avctx)
65 MpegAudioContext *s = avctx->priv_data;
66 int freq = avctx->sample_rate;
67 int bitrate = avctx->bit_rate;
68 int channels = avctx->channels;
74 bitrate = bitrate / 1000;
75 s->nb_channels = channels;
77 s->bit_rate = bitrate * 1000;
78 avctx->frame_size = MPA_FRAME_SIZE;
83 if (mpa_freq_tab[i] == freq)
85 if ((mpa_freq_tab[i] / 2) == freq) {
94 /* encoding bitrate & frequency */
96 if (mpa_bitrate_tab[s->lsf][1][i] == bitrate)
101 s->bitrate_index = i;
103 /* compute total header size & pad bit */
105 a = (float)(bitrate * 1000 * MPA_FRAME_SIZE) / (freq * 8.0);
106 s->frame_size = ((int)a) * 8;
108 /* frame fractional size to compute padding */
110 s->frame_frac_incr = (int)((a - floor(a)) * 65536.0);
112 /* select the right allocation table */
113 table = l2_select_table(bitrate, s->nb_channels, freq, s->lsf);
115 /* number of used subbands */
116 s->sblimit = sblimit_table[table];
117 s->alloc_table = alloc_tables[table];
120 av_log(AV_LOG_DEBUG, "%d kb/s, %d Hz, frame_size=%d bits, table=%d, padincr=%x\n",
121 bitrate, freq, s->frame_size, table, s->frame_frac_incr);
124 for(i=0;i<s->nb_channels;i++)
125 s->samples_offset[i] = 0;
131 v = (v + (1 << (16 - WFRAC_BITS - 1))) >> (16 - WFRAC_BITS);
137 filter_bank[512 - i] = v;
141 v = (int)(pow(2.0, (3 - i) / 3.0) * (1 << 20));
144 scale_factor_table[i] = v;
146 scale_factor_inv_table[i] = pow(2.0, -(3 - i) / 3.0) / (float)(1 << 20);
149 scale_factor_shift[i] = 21 - P - (i / 3);
150 scale_factor_mult[i] = (1 << P) * pow(2.0, (i % 3) / 3.0);
165 scale_diff_table[i] = v;
174 total_quant_bits[i] = 12 * v;
177 avctx->coded_frame= avcodec_alloc_frame();
178 avctx->coded_frame->key_frame= 1;
183 /* 32 point floating point IDCT without 1/sqrt(2) coef zero scaling */
184 static void idct32(int *out, int *tab)
188 const int *xp = costab32;
190 for(j=31;j>=3;j-=2) tab[j] += tab[j - 2];
229 x3 = MUL(t[16], FIX(SQRT2*0.5));
233 x2 = MUL(-(t[24] + t[8]), FIX(SQRT2*0.5));
234 x1 = MUL((t[8] - x2), xp[0]);
235 x2 = MUL((t[8] + x2), xp[1]);
248 xr = MUL(t[28],xp[0]);
252 xr = MUL(t[4],xp[1]);
253 t[ 4] = (t[24] - xr);
254 t[24] = (t[24] + xr);
256 xr = MUL(t[20],xp[2]);
260 xr = MUL(t[12],xp[3]);
261 t[12] = (t[16] - xr);
262 t[16] = (t[16] + xr);
267 for (i = 0; i < 4; i++) {
268 xr = MUL(tab[30-i*4],xp[0]);
269 tab[30-i*4] = (tab[i*4] - xr);
270 tab[ i*4] = (tab[i*4] + xr);
272 xr = MUL(tab[ 2+i*4],xp[1]);
273 tab[ 2+i*4] = (tab[28-i*4] - xr);
274 tab[28-i*4] = (tab[28-i*4] + xr);
276 xr = MUL(tab[31-i*4],xp[0]);
277 tab[31-i*4] = (tab[1+i*4] - xr);
278 tab[ 1+i*4] = (tab[1+i*4] + xr);
280 xr = MUL(tab[ 3+i*4],xp[1]);
281 tab[ 3+i*4] = (tab[29-i*4] - xr);
282 tab[29-i*4] = (tab[29-i*4] + xr);
290 xr = MUL(t1[0], *xp);
299 out[i] = tab[bitinv32[i]];
303 #define WSHIFT (WFRAC_BITS + 15 - FRAC_BITS)
305 static void filter(MpegAudioContext *s, int ch, short *samples, int incr)
308 int sum, offset, i, j;
313 // print_pow1(samples, 1152);
315 offset = s->samples_offset[ch];
316 out = &s->sb_samples[ch][0][0][0];
318 /* 32 samples at once */
320 s->samples_buf[ch][offset + (31 - i)] = samples[0];
325 p = s->samples_buf[ch] + offset;
329 sum = p[0*64] * q[0*64];
330 sum += p[1*64] * q[1*64];
331 sum += p[2*64] * q[2*64];
332 sum += p[3*64] * q[3*64];
333 sum += p[4*64] * q[4*64];
334 sum += p[5*64] * q[5*64];
335 sum += p[6*64] * q[6*64];
336 sum += p[7*64] * q[7*64];
341 tmp1[0] = tmp[16] >> WSHIFT;
342 for( i=1; i<=16; i++ ) tmp1[i] = (tmp[i+16]+tmp[16-i]) >> WSHIFT;
343 for( i=17; i<=31; i++ ) tmp1[i] = (tmp[i+16]-tmp[80-i]) >> WSHIFT;
347 /* advance of 32 samples */
350 /* handle the wrap around */
352 memmove(s->samples_buf[ch] + SAMPLES_BUF_SIZE - (512 - 32),
353 s->samples_buf[ch], (512 - 32) * 2);
354 offset = SAMPLES_BUF_SIZE - 512;
357 s->samples_offset[ch] = offset;
359 // print_pow(s->sb_samples, 1152);
362 static void compute_scale_factors(unsigned char scale_code[SBLIMIT],
363 unsigned char scale_factors[SBLIMIT][3],
364 int sb_samples[3][12][SBLIMIT],
367 int *p, vmax, v, n, i, j, k, code;
369 unsigned char *sf = &scale_factors[0][0];
371 for(j=0;j<sblimit;j++) {
373 /* find the max absolute value */
374 p = &sb_samples[i][0][j];
382 /* compute the scale factor index using log 2 computations */
385 /* n is the position of the MSB of vmax. now
386 use at most 2 compares to find the index */
387 index = (21 - n) * 3 - 3;
389 while (vmax <= scale_factor_table[index+1])
392 index = 0; /* very unlikely case of overflow */
395 index = 62; /* value 63 is not allowed */
399 printf("%2d:%d in=%x %x %d\n",
400 j, i, vmax, scale_factor_table[index], index);
402 /* store the scale factor */
403 assert(index >=0 && index <= 63);
407 /* compute the transmission factor : look if the scale factors
408 are close enough to each other */
409 d1 = scale_diff_table[sf[0] - sf[1] + 64];
410 d2 = scale_diff_table[sf[1] - sf[2] + 64];
412 /* handle the 25 cases */
413 switch(d1 * 5 + d2) {
445 sf[1] = sf[2] = sf[0];
450 sf[0] = sf[1] = sf[2];
456 sf[0] = sf[2] = sf[1];
462 sf[1] = sf[2] = sf[0];
469 printf("%d: %2d %2d %2d %d %d -> %d\n", j,
470 sf[0], sf[1], sf[2], d1, d2, code);
472 scale_code[j] = code;
477 /* The most important function : psycho acoustic module. In this
478 encoder there is basically none, so this is the worst you can do,
479 but also this is the simpler. */
480 static void psycho_acoustic_model(MpegAudioContext *s, short smr[SBLIMIT])
484 for(i=0;i<s->sblimit;i++) {
485 smr[i] = (int)(fixed_smr[i] * 10);
490 #define SB_NOTALLOCATED 0
491 #define SB_ALLOCATED 1
494 /* Try to maximize the smr while using a number of bits inferior to
495 the frame size. I tried to make the code simpler, faster and
496 smaller than other encoders :-) */
497 static void compute_bit_allocation(MpegAudioContext *s,
498 short smr1[MPA_MAX_CHANNELS][SBLIMIT],
499 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT],
502 int i, ch, b, max_smr, max_ch, max_sb, current_frame_size, max_frame_size;
504 short smr[MPA_MAX_CHANNELS][SBLIMIT];
505 unsigned char subband_status[MPA_MAX_CHANNELS][SBLIMIT];
506 const unsigned char *alloc;
508 memcpy(smr, smr1, s->nb_channels * sizeof(short) * SBLIMIT);
509 memset(subband_status, SB_NOTALLOCATED, s->nb_channels * SBLIMIT);
510 memset(bit_alloc, 0, s->nb_channels * SBLIMIT);
512 /* compute frame size and padding */
513 max_frame_size = s->frame_size;
514 s->frame_frac += s->frame_frac_incr;
515 if (s->frame_frac >= 65536) {
516 s->frame_frac -= 65536;
523 /* compute the header + bit alloc size */
524 current_frame_size = 32;
525 alloc = s->alloc_table;
526 for(i=0;i<s->sblimit;i++) {
528 current_frame_size += incr * s->nb_channels;
532 /* look for the subband with the largest signal to mask ratio */
535 max_smr = 0x80000000;
536 for(ch=0;ch<s->nb_channels;ch++) {
537 for(i=0;i<s->sblimit;i++) {
538 if (smr[ch][i] > max_smr && subband_status[ch][i] != SB_NOMORE) {
539 max_smr = smr[ch][i];
546 printf("current=%d max=%d max_sb=%d alloc=%d\n",
547 current_frame_size, max_frame_size, max_sb,
553 /* find alloc table entry (XXX: not optimal, should use
555 alloc = s->alloc_table;
556 for(i=0;i<max_sb;i++) {
557 alloc += 1 << alloc[0];
560 if (subband_status[max_ch][max_sb] == SB_NOTALLOCATED) {
561 /* nothing was coded for this band: add the necessary bits */
562 incr = 2 + nb_scale_factors[s->scale_code[max_ch][max_sb]] * 6;
563 incr += total_quant_bits[alloc[1]];
565 /* increments bit allocation */
566 b = bit_alloc[max_ch][max_sb];
567 incr = total_quant_bits[alloc[b + 1]] -
568 total_quant_bits[alloc[b]];
571 if (current_frame_size + incr <= max_frame_size) {
572 /* can increase size */
573 b = ++bit_alloc[max_ch][max_sb];
574 current_frame_size += incr;
575 /* decrease smr by the resolution we added */
576 smr[max_ch][max_sb] = smr1[max_ch][max_sb] - quant_snr[alloc[b]];
577 /* max allocation size reached ? */
578 if (b == ((1 << alloc[0]) - 1))
579 subband_status[max_ch][max_sb] = SB_NOMORE;
581 subband_status[max_ch][max_sb] = SB_ALLOCATED;
583 /* cannot increase the size of this subband */
584 subband_status[max_ch][max_sb] = SB_NOMORE;
587 *padding = max_frame_size - current_frame_size;
588 assert(*padding >= 0);
591 for(i=0;i<s->sblimit;i++) {
592 printf("%d ", bit_alloc[i]);
599 * Output the mpeg audio layer 2 frame. Note how the code is small
600 * compared to other encoders :-)
602 static void encode_frame(MpegAudioContext *s,
603 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT],
606 int i, j, k, l, bit_alloc_bits, b, ch;
609 PutBitContext *p = &s->pb;
613 put_bits(p, 12, 0xfff);
614 put_bits(p, 1, 1 - s->lsf); /* 1 = mpeg1 ID, 0 = mpeg2 lsf ID */
615 put_bits(p, 2, 4-2); /* layer 2 */
616 put_bits(p, 1, 1); /* no error protection */
617 put_bits(p, 4, s->bitrate_index);
618 put_bits(p, 2, s->freq_index);
619 put_bits(p, 1, s->do_padding); /* use padding */
620 put_bits(p, 1, 0); /* private_bit */
621 put_bits(p, 2, s->nb_channels == 2 ? MPA_STEREO : MPA_MONO);
622 put_bits(p, 2, 0); /* mode_ext */
623 put_bits(p, 1, 0); /* no copyright */
624 put_bits(p, 1, 1); /* original */
625 put_bits(p, 2, 0); /* no emphasis */
629 for(i=0;i<s->sblimit;i++) {
630 bit_alloc_bits = s->alloc_table[j];
631 for(ch=0;ch<s->nb_channels;ch++) {
632 put_bits(p, bit_alloc_bits, bit_alloc[ch][i]);
634 j += 1 << bit_alloc_bits;
638 for(i=0;i<s->sblimit;i++) {
639 for(ch=0;ch<s->nb_channels;ch++) {
640 if (bit_alloc[ch][i])
641 put_bits(p, 2, s->scale_code[ch][i]);
646 for(i=0;i<s->sblimit;i++) {
647 for(ch=0;ch<s->nb_channels;ch++) {
648 if (bit_alloc[ch][i]) {
649 sf = &s->scale_factors[ch][i][0];
650 switch(s->scale_code[ch][i]) {
652 put_bits(p, 6, sf[0]);
653 put_bits(p, 6, sf[1]);
654 put_bits(p, 6, sf[2]);
658 put_bits(p, 6, sf[0]);
659 put_bits(p, 6, sf[2]);
662 put_bits(p, 6, sf[0]);
669 /* quantization & write sub band samples */
674 for(i=0;i<s->sblimit;i++) {
675 bit_alloc_bits = s->alloc_table[j];
676 for(ch=0;ch<s->nb_channels;ch++) {
677 b = bit_alloc[ch][i];
679 int qindex, steps, m, sample, bits;
680 /* we encode 3 sub band samples of the same sub band at a time */
681 qindex = s->alloc_table[j+b];
682 steps = quant_steps[qindex];
684 sample = s->sb_samples[ch][k][l + m][i];
685 /* divide by scale factor */
689 a = (float)sample * scale_factor_inv_table[s->scale_factors[ch][i][k]];
690 q[m] = (int)((a + 1.0) * steps * 0.5);
694 int q1, e, shift, mult;
695 e = s->scale_factors[ch][i][k];
696 shift = scale_factor_shift[e];
697 mult = scale_factor_mult[e];
699 /* normalize to P bits */
701 q1 = sample << (-shift);
703 q1 = sample >> shift;
704 q1 = (q1 * mult) >> P;
705 q[m] = ((q1 + (1 << P)) * steps) >> (P + 1);
710 assert(q[m] >= 0 && q[m] < steps);
712 bits = quant_bits[qindex];
714 /* group the 3 values to save bits */
716 q[0] + steps * (q[1] + steps * q[2]));
718 printf("%d: gr1 %d\n",
719 i, q[0] + steps * (q[1] + steps * q[2]));
723 printf("%d: gr3 %d %d %d\n",
724 i, q[0], q[1], q[2]);
726 put_bits(p, bits, q[0]);
727 put_bits(p, bits, q[1]);
728 put_bits(p, bits, q[2]);
732 /* next subband in alloc table */
733 j += 1 << bit_alloc_bits;
739 for(i=0;i<padding;i++)
746 static int MPA_encode_frame(AVCodecContext *avctx,
747 unsigned char *frame, int buf_size, void *data)
749 MpegAudioContext *s = avctx->priv_data;
750 short *samples = data;
751 short smr[MPA_MAX_CHANNELS][SBLIMIT];
752 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT];
755 for(i=0;i<s->nb_channels;i++) {
756 filter(s, i, samples + i, s->nb_channels);
759 for(i=0;i<s->nb_channels;i++) {
760 compute_scale_factors(s->scale_code[i], s->scale_factors[i],
761 s->sb_samples[i], s->sblimit);
763 for(i=0;i<s->nb_channels;i++) {
764 psycho_acoustic_model(s, smr[i]);
766 compute_bit_allocation(s, smr, bit_alloc, &padding);
768 init_put_bits(&s->pb, frame, MPA_MAX_CODED_FRAME_SIZE);
770 encode_frame(s, bit_alloc, padding);
772 s->nb_samples += MPA_FRAME_SIZE;
773 return pbBufPtr(&s->pb) - s->pb.buf;
776 static int MPA_encode_close(AVCodecContext *avctx)
778 av_freep(&avctx->coded_frame);
782 AVCodec mp2_encoder = {
786 sizeof(MpegAudioContext),