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
20 #include "mpegaudio.h"
22 /* currently, cannot change these constants (need to modify
23 quantization stage) */
26 #define MUL(a,b) (((INT64)(a) * (INT64)(b)) >> FRAC_BITS)
27 #define FIX(a) ((int)((a) * (1 << FRAC_BITS)))
29 #define SAMPLES_BUF_SIZE 4096
31 typedef struct MpegAudioContext {
35 int lsf; /* 1 if mpeg2 low bitrate selected */
36 int bitrate_index; /* bit rate */
38 int frame_size; /* frame size, in bits, without padding */
39 INT64 nb_samples; /* total number of samples encoded */
40 /* padding computation */
41 int frame_frac, frame_frac_incr, do_padding;
42 short samples_buf[MPA_MAX_CHANNELS][SAMPLES_BUF_SIZE]; /* buffer for filter */
43 int samples_offset[MPA_MAX_CHANNELS]; /* offset in samples_buf */
44 int sb_samples[MPA_MAX_CHANNELS][3][12][SBLIMIT];
45 unsigned char scale_factors[MPA_MAX_CHANNELS][SBLIMIT][3]; /* scale factors */
46 /* code to group 3 scale factors */
47 unsigned char scale_code[MPA_MAX_CHANNELS][SBLIMIT];
48 int sblimit; /* number of used subbands */
49 const unsigned char *alloc_table;
52 /* define it to use floats in quantization (I don't like floats !) */
55 #include "mpegaudiotab.h"
57 static int MPA_encode_init(AVCodecContext *avctx)
59 MpegAudioContext *s = avctx->priv_data;
60 int freq = avctx->sample_rate;
61 int bitrate = avctx->bit_rate;
62 int channels = avctx->channels;
68 bitrate = bitrate / 1000;
69 s->nb_channels = channels;
71 s->bit_rate = bitrate * 1000;
72 avctx->frame_size = MPA_FRAME_SIZE;
77 if (mpa_freq_tab[i] == freq)
79 if ((mpa_freq_tab[i] / 2) == freq) {
88 /* encoding bitrate & frequency */
90 if (mpa_bitrate_tab[s->lsf][1][i] == bitrate)
97 /* compute total header size & pad bit */
99 a = (float)(bitrate * 1000 * MPA_FRAME_SIZE) / (freq * 8.0);
100 s->frame_size = ((int)a) * 8;
102 /* frame fractional size to compute padding */
104 s->frame_frac_incr = (int)((a - floor(a)) * 65536.0);
106 /* select the right allocation table */
107 table = l2_select_table(bitrate, s->nb_channels, freq, s->lsf);
109 /* number of used subbands */
110 s->sblimit = sblimit_table[table];
111 s->alloc_table = alloc_tables[table];
114 printf("%d kb/s, %d Hz, frame_size=%d bits, table=%d, padincr=%x\n",
115 bitrate, freq, s->frame_size, table, s->frame_frac_incr);
118 for(i=0;i<s->nb_channels;i++)
119 s->samples_offset[i] = 0;
125 v = (v + (1 << (16 - WFRAC_BITS - 1))) >> (16 - WFRAC_BITS);
131 filter_bank[512 - i] = v;
135 v = (int)(pow(2.0, (3 - i) / 3.0) * (1 << 20));
138 scale_factor_table[i] = v;
140 scale_factor_inv_table[i] = pow(2.0, -(3 - i) / 3.0) / (float)(1 << 20);
143 scale_factor_shift[i] = 21 - P - (i / 3);
144 scale_factor_mult[i] = (1 << P) * pow(2.0, (i % 3) / 3.0);
159 scale_diff_table[i] = v;
168 total_quant_bits[i] = 12 * v;
171 avctx->coded_frame= avcodec_alloc_frame();
172 avctx->coded_frame->key_frame= 1;
177 /* 32 point floating point IDCT without 1/sqrt(2) coef zero scaling */
178 static void idct32(int *out, int *tab)
182 const int *xp = costab32;
184 for(j=31;j>=3;j-=2) tab[j] += tab[j - 2];
223 x3 = MUL(t[16], FIX(SQRT2*0.5));
227 x2 = MUL(-(t[24] + t[8]), FIX(SQRT2*0.5));
228 x1 = MUL((t[8] - x2), xp[0]);
229 x2 = MUL((t[8] + x2), xp[1]);
242 xr = MUL(t[28],xp[0]);
246 xr = MUL(t[4],xp[1]);
247 t[ 4] = (t[24] - xr);
248 t[24] = (t[24] + xr);
250 xr = MUL(t[20],xp[2]);
254 xr = MUL(t[12],xp[3]);
255 t[12] = (t[16] - xr);
256 t[16] = (t[16] + xr);
261 for (i = 0; i < 4; i++) {
262 xr = MUL(tab[30-i*4],xp[0]);
263 tab[30-i*4] = (tab[i*4] - xr);
264 tab[ i*4] = (tab[i*4] + xr);
266 xr = MUL(tab[ 2+i*4],xp[1]);
267 tab[ 2+i*4] = (tab[28-i*4] - xr);
268 tab[28-i*4] = (tab[28-i*4] + xr);
270 xr = MUL(tab[31-i*4],xp[0]);
271 tab[31-i*4] = (tab[1+i*4] - xr);
272 tab[ 1+i*4] = (tab[1+i*4] + xr);
274 xr = MUL(tab[ 3+i*4],xp[1]);
275 tab[ 3+i*4] = (tab[29-i*4] - xr);
276 tab[29-i*4] = (tab[29-i*4] + xr);
284 xr = MUL(t1[0], *xp);
293 out[i] = tab[bitinv32[i]];
297 #define WSHIFT (WFRAC_BITS + 15 - FRAC_BITS)
299 static void filter(MpegAudioContext *s, int ch, short *samples, int incr)
302 int sum, offset, i, j;
307 // print_pow1(samples, 1152);
309 offset = s->samples_offset[ch];
310 out = &s->sb_samples[ch][0][0][0];
312 /* 32 samples at once */
314 s->samples_buf[ch][offset + (31 - i)] = samples[0];
319 p = s->samples_buf[ch] + offset;
323 sum = p[0*64] * q[0*64];
324 sum += p[1*64] * q[1*64];
325 sum += p[2*64] * q[2*64];
326 sum += p[3*64] * q[3*64];
327 sum += p[4*64] * q[4*64];
328 sum += p[5*64] * q[5*64];
329 sum += p[6*64] * q[6*64];
330 sum += p[7*64] * q[7*64];
335 tmp1[0] = tmp[16] >> WSHIFT;
336 for( i=1; i<=16; i++ ) tmp1[i] = (tmp[i+16]+tmp[16-i]) >> WSHIFT;
337 for( i=17; i<=31; i++ ) tmp1[i] = (tmp[i+16]-tmp[80-i]) >> WSHIFT;
341 /* advance of 32 samples */
344 /* handle the wrap around */
346 memmove(s->samples_buf[ch] + SAMPLES_BUF_SIZE - (512 - 32),
347 s->samples_buf[ch], (512 - 32) * 2);
348 offset = SAMPLES_BUF_SIZE - 512;
351 s->samples_offset[ch] = offset;
353 // print_pow(s->sb_samples, 1152);
356 static void compute_scale_factors(unsigned char scale_code[SBLIMIT],
357 unsigned char scale_factors[SBLIMIT][3],
358 int sb_samples[3][12][SBLIMIT],
361 int *p, vmax, v, n, i, j, k, code;
363 unsigned char *sf = &scale_factors[0][0];
365 for(j=0;j<sblimit;j++) {
367 /* find the max absolute value */
368 p = &sb_samples[i][0][j];
376 /* compute the scale factor index using log 2 computations */
379 /* n is the position of the MSB of vmax. now
380 use at most 2 compares to find the index */
381 index = (21 - n) * 3 - 3;
383 while (vmax <= scale_factor_table[index+1])
386 index = 0; /* very unlikely case of overflow */
389 index = 62; /* value 63 is not allowed */
393 printf("%2d:%d in=%x %x %d\n",
394 j, i, vmax, scale_factor_table[index], index);
396 /* store the scale factor */
397 assert(index >=0 && index <= 63);
401 /* compute the transmission factor : look if the scale factors
402 are close enough to each other */
403 d1 = scale_diff_table[sf[0] - sf[1] + 64];
404 d2 = scale_diff_table[sf[1] - sf[2] + 64];
406 /* handle the 25 cases */
407 switch(d1 * 5 + d2) {
439 sf[1] = sf[2] = sf[0];
444 sf[0] = sf[1] = sf[2];
450 sf[0] = sf[2] = sf[1];
456 sf[1] = sf[2] = sf[0];
463 printf("%d: %2d %2d %2d %d %d -> %d\n", j,
464 sf[0], sf[1], sf[2], d1, d2, code);
466 scale_code[j] = code;
471 /* The most important function : psycho acoustic module. In this
472 encoder there is basically none, so this is the worst you can do,
473 but also this is the simpler. */
474 static void psycho_acoustic_model(MpegAudioContext *s, short smr[SBLIMIT])
478 for(i=0;i<s->sblimit;i++) {
479 smr[i] = (int)(fixed_smr[i] * 10);
484 #define SB_NOTALLOCATED 0
485 #define SB_ALLOCATED 1
488 /* Try to maximize the smr while using a number of bits inferior to
489 the frame size. I tried to make the code simpler, faster and
490 smaller than other encoders :-) */
491 static void compute_bit_allocation(MpegAudioContext *s,
492 short smr1[MPA_MAX_CHANNELS][SBLIMIT],
493 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT],
496 int i, ch, b, max_smr, max_ch, max_sb, current_frame_size, max_frame_size;
498 short smr[MPA_MAX_CHANNELS][SBLIMIT];
499 unsigned char subband_status[MPA_MAX_CHANNELS][SBLIMIT];
500 const unsigned char *alloc;
502 memcpy(smr, smr1, s->nb_channels * sizeof(short) * SBLIMIT);
503 memset(subband_status, SB_NOTALLOCATED, s->nb_channels * SBLIMIT);
504 memset(bit_alloc, 0, s->nb_channels * SBLIMIT);
506 /* compute frame size and padding */
507 max_frame_size = s->frame_size;
508 s->frame_frac += s->frame_frac_incr;
509 if (s->frame_frac >= 65536) {
510 s->frame_frac -= 65536;
517 /* compute the header + bit alloc size */
518 current_frame_size = 32;
519 alloc = s->alloc_table;
520 for(i=0;i<s->sblimit;i++) {
522 current_frame_size += incr * s->nb_channels;
526 /* look for the subband with the largest signal to mask ratio */
529 max_smr = 0x80000000;
530 for(ch=0;ch<s->nb_channels;ch++) {
531 for(i=0;i<s->sblimit;i++) {
532 if (smr[ch][i] > max_smr && subband_status[ch][i] != SB_NOMORE) {
533 max_smr = smr[ch][i];
540 printf("current=%d max=%d max_sb=%d alloc=%d\n",
541 current_frame_size, max_frame_size, max_sb,
547 /* find alloc table entry (XXX: not optimal, should use
549 alloc = s->alloc_table;
550 for(i=0;i<max_sb;i++) {
551 alloc += 1 << alloc[0];
554 if (subband_status[max_ch][max_sb] == SB_NOTALLOCATED) {
555 /* nothing was coded for this band: add the necessary bits */
556 incr = 2 + nb_scale_factors[s->scale_code[max_ch][max_sb]] * 6;
557 incr += total_quant_bits[alloc[1]];
559 /* increments bit allocation */
560 b = bit_alloc[max_ch][max_sb];
561 incr = total_quant_bits[alloc[b + 1]] -
562 total_quant_bits[alloc[b]];
565 if (current_frame_size + incr <= max_frame_size) {
566 /* can increase size */
567 b = ++bit_alloc[max_ch][max_sb];
568 current_frame_size += incr;
569 /* decrease smr by the resolution we added */
570 smr[max_ch][max_sb] = smr1[max_ch][max_sb] - quant_snr[alloc[b]];
571 /* max allocation size reached ? */
572 if (b == ((1 << alloc[0]) - 1))
573 subband_status[max_ch][max_sb] = SB_NOMORE;
575 subband_status[max_ch][max_sb] = SB_ALLOCATED;
577 /* cannot increase the size of this subband */
578 subband_status[max_ch][max_sb] = SB_NOMORE;
581 *padding = max_frame_size - current_frame_size;
582 assert(*padding >= 0);
585 for(i=0;i<s->sblimit;i++) {
586 printf("%d ", bit_alloc[i]);
593 * Output the mpeg audio layer 2 frame. Note how the code is small
594 * compared to other encoders :-)
596 static void encode_frame(MpegAudioContext *s,
597 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT],
600 int i, j, k, l, bit_alloc_bits, b, ch;
603 PutBitContext *p = &s->pb;
607 put_bits(p, 12, 0xfff);
608 put_bits(p, 1, 1 - s->lsf); /* 1 = mpeg1 ID, 0 = mpeg2 lsf ID */
609 put_bits(p, 2, 4-2); /* layer 2 */
610 put_bits(p, 1, 1); /* no error protection */
611 put_bits(p, 4, s->bitrate_index);
612 put_bits(p, 2, s->freq_index);
613 put_bits(p, 1, s->do_padding); /* use padding */
614 put_bits(p, 1, 0); /* private_bit */
615 put_bits(p, 2, s->nb_channels == 2 ? MPA_STEREO : MPA_MONO);
616 put_bits(p, 2, 0); /* mode_ext */
617 put_bits(p, 1, 0); /* no copyright */
618 put_bits(p, 1, 1); /* original */
619 put_bits(p, 2, 0); /* no emphasis */
623 for(i=0;i<s->sblimit;i++) {
624 bit_alloc_bits = s->alloc_table[j];
625 for(ch=0;ch<s->nb_channels;ch++) {
626 put_bits(p, bit_alloc_bits, bit_alloc[ch][i]);
628 j += 1 << bit_alloc_bits;
632 for(i=0;i<s->sblimit;i++) {
633 for(ch=0;ch<s->nb_channels;ch++) {
634 if (bit_alloc[ch][i])
635 put_bits(p, 2, s->scale_code[ch][i]);
640 for(i=0;i<s->sblimit;i++) {
641 for(ch=0;ch<s->nb_channels;ch++) {
642 if (bit_alloc[ch][i]) {
643 sf = &s->scale_factors[ch][i][0];
644 switch(s->scale_code[ch][i]) {
646 put_bits(p, 6, sf[0]);
647 put_bits(p, 6, sf[1]);
648 put_bits(p, 6, sf[2]);
652 put_bits(p, 6, sf[0]);
653 put_bits(p, 6, sf[2]);
656 put_bits(p, 6, sf[0]);
663 /* quantization & write sub band samples */
668 for(i=0;i<s->sblimit;i++) {
669 bit_alloc_bits = s->alloc_table[j];
670 for(ch=0;ch<s->nb_channels;ch++) {
671 b = bit_alloc[ch][i];
673 int qindex, steps, m, sample, bits;
674 /* we encode 3 sub band samples of the same sub band at a time */
675 qindex = s->alloc_table[j+b];
676 steps = quant_steps[qindex];
678 sample = s->sb_samples[ch][k][l + m][i];
679 /* divide by scale factor */
683 a = (float)sample * scale_factor_inv_table[s->scale_factors[ch][i][k]];
684 q[m] = (int)((a + 1.0) * steps * 0.5);
688 int q1, e, shift, mult;
689 e = s->scale_factors[ch][i][k];
690 shift = scale_factor_shift[e];
691 mult = scale_factor_mult[e];
693 /* normalize to P bits */
695 q1 = sample << (-shift);
697 q1 = sample >> shift;
698 q1 = (q1 * mult) >> P;
699 q[m] = ((q1 + (1 << P)) * steps) >> (P + 1);
704 assert(q[m] >= 0 && q[m] < steps);
706 bits = quant_bits[qindex];
708 /* group the 3 values to save bits */
710 q[0] + steps * (q[1] + steps * q[2]));
712 printf("%d: gr1 %d\n",
713 i, q[0] + steps * (q[1] + steps * q[2]));
717 printf("%d: gr3 %d %d %d\n",
718 i, q[0], q[1], q[2]);
720 put_bits(p, bits, q[0]);
721 put_bits(p, bits, q[1]);
722 put_bits(p, bits, q[2]);
726 /* next subband in alloc table */
727 j += 1 << bit_alloc_bits;
733 for(i=0;i<padding;i++)
740 static int MPA_encode_frame(AVCodecContext *avctx,
741 unsigned char *frame, int buf_size, void *data)
743 MpegAudioContext *s = avctx->priv_data;
744 short *samples = data;
745 short smr[MPA_MAX_CHANNELS][SBLIMIT];
746 unsigned char bit_alloc[MPA_MAX_CHANNELS][SBLIMIT];
749 for(i=0;i<s->nb_channels;i++) {
750 filter(s, i, samples + i, s->nb_channels);
753 for(i=0;i<s->nb_channels;i++) {
754 compute_scale_factors(s->scale_code[i], s->scale_factors[i],
755 s->sb_samples[i], s->sblimit);
757 for(i=0;i<s->nb_channels;i++) {
758 psycho_acoustic_model(s, smr[i]);
760 compute_bit_allocation(s, smr, bit_alloc, &padding);
762 init_put_bits(&s->pb, frame, MPA_MAX_CODED_FRAME_SIZE, NULL, NULL);
764 encode_frame(s, bit_alloc, padding);
766 s->nb_samples += MPA_FRAME_SIZE;
767 return pbBufPtr(&s->pb) - s->pb.buf;
770 static int MPA_encode_close(AVCodecContext *avctx)
772 av_freep(&avctx->coded_frame);
776 AVCodec mp2_encoder = {
780 sizeof(MpegAudioContext),