2 * The simplest AC-3 encoder
3 * Copyright (c) 2000 Fabrice Bellard
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 * @file libavcodec/ac3enc.c
24 * The simplest AC-3 encoder.
27 //#define DEBUG_BITALLOC
28 #include "libavutil/crc.h"
30 #include "libavutil/common.h" /* for av_reverse */
33 #include "audioconvert.h"
35 typedef struct AC3EncodeContext {
40 const uint8_t *channel_map;
42 unsigned int sample_rate;
43 unsigned int bitstream_id;
44 unsigned int frame_size_min; /* minimum frame size in case rounding is necessary */
45 unsigned int frame_size; /* current frame size in words */
46 unsigned int bits_written;
47 unsigned int samples_written;
49 unsigned int frame_size_code;
50 unsigned int sr_code; /* frequency */
51 unsigned int channel_mode;
53 unsigned int bitstream_mode;
54 short last_samples[AC3_MAX_CHANNELS][256];
55 unsigned int chbwcod[AC3_MAX_CHANNELS];
56 int nb_coefs[AC3_MAX_CHANNELS];
58 /* bitrate allocation control */
59 int slow_gain_code, slow_decay_code, fast_decay_code, db_per_bit_code, floor_code;
60 AC3BitAllocParameters bit_alloc;
61 int coarse_snr_offset;
62 int fast_gain_code[AC3_MAX_CHANNELS];
63 int fine_snr_offset[AC3_MAX_CHANNELS];
64 /* mantissa encoding */
65 int mant1_cnt, mant2_cnt, mant4_cnt;
68 static int16_t costab[64];
69 static int16_t sintab[64];
70 static int16_t xcos1[128];
71 static int16_t xsin1[128];
74 #define N (1 << MDCT_NBITS)
76 /* new exponents are sent if their Norm 1 exceed this number */
77 #define EXP_DIFF_THRESHOLD 1000
79 static inline int16_t fix15(float a)
82 v = (int)(a * (float)(1 << 15));
90 typedef struct IComplex {
94 static av_cold void fft_init(int ln)
101 for(i=0;i<(n/2);i++) {
102 alpha = 2 * M_PI * (float)i / (float)n;
103 costab[i] = fix15(cos(alpha));
104 sintab[i] = fix15(sin(alpha));
109 #define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \
116 pre = (bx + ax) >> 1;\
117 pim = (by + ay) >> 1;\
118 qre = (bx - ax) >> 1;\
119 qim = (by - ay) >> 1;\
122 #define CMUL(pre, pim, are, aim, bre, bim) \
124 pre = (MUL16(are, bre) - MUL16(aim, bim)) >> 15;\
125 pim = (MUL16(are, bim) + MUL16(bre, aim)) >> 15;\
129 /* do a 2^n point complex fft on 2^ln points. */
130 static void fft(IComplex *z, int ln)
134 register IComplex *p,*q;
141 int k = av_reverse[j] >> (8 - ln);
143 FFSWAP(IComplex, z[k], z[j]);
151 BF(p[0].re, p[0].im, p[1].re, p[1].im,
152 p[0].re, p[0].im, p[1].re, p[1].im);
161 BF(p[0].re, p[0].im, p[2].re, p[2].im,
162 p[0].re, p[0].im, p[2].re, p[2].im);
163 BF(p[1].re, p[1].im, p[3].re, p[3].im,
164 p[1].re, p[1].im, p[3].im, -p[3].re);
176 for (j = 0; j < nblocks; ++j) {
178 BF(p->re, p->im, q->re, q->im,
179 p->re, p->im, q->re, q->im);
183 for(l = nblocks; l < np2; l += nblocks) {
184 CMUL(tmp_re, tmp_im, costab[l], -sintab[l], q->re, q->im);
185 BF(p->re, p->im, q->re, q->im,
186 p->re, p->im, tmp_re, tmp_im);
193 nblocks = nblocks >> 1;
194 nloops = nloops << 1;
195 } while (nblocks != 0);
198 /* do a 512 point mdct */
199 static void mdct512(int32_t *out, int16_t *in)
201 int i, re, im, re1, im1;
205 /* shift to simplify computations */
207 rot[i] = -in[i + 3*N/4];
209 rot[i] = in[i - N/4];
213 re = ((int)rot[2*i] - (int)rot[N-1-2*i]) >> 1;
214 im = -((int)rot[N/2+2*i] - (int)rot[N/2-1-2*i]) >> 1;
215 CMUL(x[i].re, x[i].im, re, im, -xcos1[i], xsin1[i]);
218 fft(x, MDCT_NBITS - 2);
224 CMUL(re1, im1, re, im, xsin1[i], xcos1[i]);
226 out[N/2-1-2*i] = re1;
230 /* XXX: use another norm ? */
231 static int calc_exp_diff(uint8_t *exp1, uint8_t *exp2, int n)
236 sum += abs(exp1[i] - exp2[i]);
241 static void compute_exp_strategy(uint8_t exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS],
242 uint8_t exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
248 /* estimate if the exponent variation & decide if they should be
249 reused in the next frame */
250 exp_strategy[0][ch] = EXP_NEW;
251 for(i=1;i<NB_BLOCKS;i++) {
252 exp_diff = calc_exp_diff(exp[i][ch], exp[i-1][ch], N/2);
253 dprintf(NULL, "exp_diff=%d\n", exp_diff);
254 if (exp_diff > EXP_DIFF_THRESHOLD)
255 exp_strategy[i][ch] = EXP_NEW;
257 exp_strategy[i][ch] = EXP_REUSE;
262 /* now select the encoding strategy type : if exponents are often
263 recoded, we use a coarse encoding */
265 while (i < NB_BLOCKS) {
267 while (j < NB_BLOCKS && exp_strategy[j][ch] == EXP_REUSE)
271 exp_strategy[i][ch] = EXP_D45;
275 exp_strategy[i][ch] = EXP_D25;
278 exp_strategy[i][ch] = EXP_D15;
285 /* set exp[i] to min(exp[i], exp1[i]) */
286 static void exponent_min(uint8_t exp[N/2], uint8_t exp1[N/2], int n)
291 if (exp1[i] < exp[i])
296 /* update the exponents so that they are the ones the decoder will
297 decode. Return the number of bits used to code the exponents */
298 static int encode_exp(uint8_t encoded_exp[N/2],
303 int group_size, nb_groups, i, j, k, exp_min;
306 switch(exp_strategy) {
318 nb_groups = ((nb_exps + (group_size * 3) - 4) / (3 * group_size)) * 3;
320 /* for each group, compute the minimum exponent */
321 exp1[0] = exp[0]; /* DC exponent is handled separately */
323 for(i=1;i<=nb_groups;i++) {
325 assert(exp_min >= 0 && exp_min <= 24);
326 for(j=1;j<group_size;j++) {
327 if (exp[k+j] < exp_min)
334 /* constraint for DC exponent */
338 /* Decrease the delta between each groups to within 2
339 * so that they can be differentially encoded */
340 for (i=1;i<=nb_groups;i++)
341 exp1[i] = FFMIN(exp1[i], exp1[i-1] + 2);
342 for (i=nb_groups-1;i>=0;i--)
343 exp1[i] = FFMIN(exp1[i], exp1[i+1] + 2);
345 /* now we have the exponent values the decoder will see */
346 encoded_exp[0] = exp1[0];
348 for(i=1;i<=nb_groups;i++) {
349 for(j=0;j<group_size;j++) {
350 encoded_exp[k+j] = exp1[i];
356 av_log(NULL, AV_LOG_DEBUG, "exponents: strategy=%d\n", exp_strategy);
357 for(i=0;i<=nb_groups * group_size;i++) {
358 av_log(NULL, AV_LOG_DEBUG, "%d ", encoded_exp[i]);
360 av_log(NULL, AV_LOG_DEBUG, "\n");
363 return 4 + (nb_groups / 3) * 7;
366 /* return the size in bits taken by the mantissa */
367 static int compute_mantissa_size(AC3EncodeContext *s, uint8_t *m, int nb_coefs)
372 for(i=0;i<nb_coefs;i++) {
379 /* 3 mantissa in 5 bits */
380 if (s->mant1_cnt == 0)
382 if (++s->mant1_cnt == 3)
386 /* 3 mantissa in 7 bits */
387 if (s->mant2_cnt == 0)
389 if (++s->mant2_cnt == 3)
396 /* 2 mantissa in 7 bits */
397 if (s->mant4_cnt == 0)
399 if (++s->mant4_cnt == 2)
417 static void bit_alloc_masking(AC3EncodeContext *s,
418 uint8_t encoded_exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
419 uint8_t exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS],
420 int16_t psd[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
421 int16_t mask[NB_BLOCKS][AC3_MAX_CHANNELS][50])
424 int16_t band_psd[NB_BLOCKS][AC3_MAX_CHANNELS][50];
426 for(blk=0; blk<NB_BLOCKS; blk++) {
427 for(ch=0;ch<s->nb_all_channels;ch++) {
428 if(exp_strategy[blk][ch] == EXP_REUSE) {
429 memcpy(psd[blk][ch], psd[blk-1][ch], (N/2)*sizeof(int16_t));
430 memcpy(mask[blk][ch], mask[blk-1][ch], 50*sizeof(int16_t));
432 ff_ac3_bit_alloc_calc_psd(encoded_exp[blk][ch], 0,
434 psd[blk][ch], band_psd[blk][ch]);
435 ff_ac3_bit_alloc_calc_mask(&s->bit_alloc, band_psd[blk][ch],
437 ff_ac3_fast_gain_tab[s->fast_gain_code[ch]],
438 ch == s->lfe_channel,
439 DBA_NONE, 0, NULL, NULL, NULL,
446 static int bit_alloc(AC3EncodeContext *s,
447 int16_t mask[NB_BLOCKS][AC3_MAX_CHANNELS][50],
448 int16_t psd[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
449 uint8_t bap[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
450 int frame_bits, int coarse_snr_offset, int fine_snr_offset)
455 snr_offset = (((coarse_snr_offset - 15) << 4) + fine_snr_offset) << 2;
458 for(i=0;i<NB_BLOCKS;i++) {
462 for(ch=0;ch<s->nb_all_channels;ch++) {
463 ff_ac3_bit_alloc_calc_bap(mask[i][ch], psd[i][ch], 0,
464 s->nb_coefs[ch], snr_offset,
465 s->bit_alloc.floor, ff_ac3_bap_tab,
467 frame_bits += compute_mantissa_size(s, bap[i][ch],
472 printf("csnr=%d fsnr=%d frame_bits=%d diff=%d\n",
473 coarse_snr_offset, fine_snr_offset, frame_bits,
474 16 * s->frame_size - ((frame_bits + 7) & ~7));
476 return 16 * s->frame_size - frame_bits;
481 static int compute_bit_allocation(AC3EncodeContext *s,
482 uint8_t bap[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
483 uint8_t encoded_exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
484 uint8_t exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS],
488 int coarse_snr_offset, fine_snr_offset;
489 uint8_t bap1[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
490 int16_t psd[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
491 int16_t mask[NB_BLOCKS][AC3_MAX_CHANNELS][50];
492 static const int frame_bits_inc[8] = { 0, 0, 2, 2, 2, 4, 2, 4 };
494 /* init default parameters */
495 s->slow_decay_code = 2;
496 s->fast_decay_code = 1;
497 s->slow_gain_code = 1;
498 s->db_per_bit_code = 2;
500 for(ch=0;ch<s->nb_all_channels;ch++)
501 s->fast_gain_code[ch] = 4;
503 /* compute real values */
504 s->bit_alloc.sr_code = s->sr_code;
505 s->bit_alloc.sr_shift = s->sr_shift;
506 s->bit_alloc.slow_decay = ff_ac3_slow_decay_tab[s->slow_decay_code] >> s->sr_shift;
507 s->bit_alloc.fast_decay = ff_ac3_fast_decay_tab[s->fast_decay_code] >> s->sr_shift;
508 s->bit_alloc.slow_gain = ff_ac3_slow_gain_tab[s->slow_gain_code];
509 s->bit_alloc.db_per_bit = ff_ac3_db_per_bit_tab[s->db_per_bit_code];
510 s->bit_alloc.floor = ff_ac3_floor_tab[s->floor_code];
514 // if (s->channel_mode == 2)
516 frame_bits += frame_bits_inc[s->channel_mode];
519 for(i=0;i<NB_BLOCKS;i++) {
520 frame_bits += s->nb_channels * 2 + 2; /* blksw * c, dithflag * c, dynrnge, cplstre */
521 if (s->channel_mode == AC3_CHMODE_STEREO) {
522 frame_bits++; /* rematstr */
523 if(i==0) frame_bits += 4;
525 frame_bits += 2 * s->nb_channels; /* chexpstr[2] * c */
527 frame_bits++; /* lfeexpstr */
528 for(ch=0;ch<s->nb_channels;ch++) {
529 if (exp_strategy[i][ch] != EXP_REUSE)
530 frame_bits += 6 + 2; /* chbwcod[6], gainrng[2] */
532 frame_bits++; /* baie */
533 frame_bits++; /* snr */
534 frame_bits += 2; /* delta / skip */
536 frame_bits++; /* cplinu for block 0 */
538 /* sdcycod[2], fdcycod[2], sgaincod[2], dbpbcod[2], floorcod[3] */
540 /* (fsnoffset[4] + fgaincod[4]) * c */
541 frame_bits += 2*4 + 3 + 6 + s->nb_all_channels * (4 + 3);
543 /* auxdatae, crcrsv */
549 /* calculate psd and masking curve before doing bit allocation */
550 bit_alloc_masking(s, encoded_exp, exp_strategy, psd, mask);
552 /* now the big work begins : do the bit allocation. Modify the snr
553 offset until we can pack everything in the requested frame size */
555 coarse_snr_offset = s->coarse_snr_offset;
556 while (coarse_snr_offset >= 0 &&
557 bit_alloc(s, mask, psd, bap, frame_bits, coarse_snr_offset, 0) < 0)
558 coarse_snr_offset -= SNR_INC1;
559 if (coarse_snr_offset < 0) {
560 av_log(NULL, AV_LOG_ERROR, "Bit allocation failed. Try increasing the bitrate.\n");
563 while ((coarse_snr_offset + SNR_INC1) <= 63 &&
564 bit_alloc(s, mask, psd, bap1, frame_bits,
565 coarse_snr_offset + SNR_INC1, 0) >= 0) {
566 coarse_snr_offset += SNR_INC1;
567 memcpy(bap, bap1, sizeof(bap1));
569 while ((coarse_snr_offset + 1) <= 63 &&
570 bit_alloc(s, mask, psd, bap1, frame_bits, coarse_snr_offset + 1, 0) >= 0) {
572 memcpy(bap, bap1, sizeof(bap1));
576 while ((fine_snr_offset + SNR_INC1) <= 15 &&
577 bit_alloc(s, mask, psd, bap1, frame_bits,
578 coarse_snr_offset, fine_snr_offset + SNR_INC1) >= 0) {
579 fine_snr_offset += SNR_INC1;
580 memcpy(bap, bap1, sizeof(bap1));
582 while ((fine_snr_offset + 1) <= 15 &&
583 bit_alloc(s, mask, psd, bap1, frame_bits,
584 coarse_snr_offset, fine_snr_offset + 1) >= 0) {
586 memcpy(bap, bap1, sizeof(bap1));
589 s->coarse_snr_offset = coarse_snr_offset;
590 for(ch=0;ch<s->nb_all_channels;ch++)
591 s->fine_snr_offset[ch] = fine_snr_offset;
592 #if defined(DEBUG_BITALLOC)
597 for(ch=0;ch<s->nb_all_channels;ch++) {
598 printf("Block #%d Ch%d:\n", i, ch);
600 for(j=0;j<s->nb_coefs[ch];j++) {
601 printf("%d ",bap[i][ch][j]);
611 static av_cold int set_channel_info(AC3EncodeContext *s, int channels,
612 int64_t *channel_layout)
616 if (channels < 1 || channels > AC3_MAX_CHANNELS)
618 if ((uint64_t)*channel_layout > 0x7FF)
620 ch_layout = *channel_layout;
622 ch_layout = avcodec_guess_channel_layout(channels, CODEC_ID_AC3, NULL);
623 if (avcodec_channel_layout_num_channels(ch_layout) != channels)
626 s->lfe = !!(ch_layout & CH_LOW_FREQUENCY);
627 s->nb_all_channels = channels;
628 s->nb_channels = channels - s->lfe;
629 s->lfe_channel = s->lfe ? s->nb_channels : -1;
631 ch_layout -= CH_LOW_FREQUENCY;
634 case CH_LAYOUT_MONO: s->channel_mode = AC3_CHMODE_MONO; break;
635 case CH_LAYOUT_STEREO: s->channel_mode = AC3_CHMODE_STEREO; break;
636 case CH_LAYOUT_SURROUND: s->channel_mode = AC3_CHMODE_3F; break;
637 case CH_LAYOUT_2_1: s->channel_mode = AC3_CHMODE_2F1R; break;
638 case CH_LAYOUT_4POINT0: s->channel_mode = AC3_CHMODE_3F1R; break;
640 case CH_LAYOUT_2_2: s->channel_mode = AC3_CHMODE_2F2R; break;
641 case CH_LAYOUT_5POINT0:
642 case CH_LAYOUT_5POINT0_BACK: s->channel_mode = AC3_CHMODE_3F2R; break;
647 s->channel_map = ff_ac3_enc_channel_map[s->channel_mode][s->lfe];
648 *channel_layout = ch_layout;
650 *channel_layout |= CH_LOW_FREQUENCY;
655 static av_cold int AC3_encode_init(AVCodecContext *avctx)
657 int freq = avctx->sample_rate;
658 int bitrate = avctx->bit_rate;
659 AC3EncodeContext *s = avctx->priv_data;
664 avctx->frame_size = AC3_FRAME_SIZE;
668 if (!avctx->channel_layout) {
669 av_log(avctx, AV_LOG_WARNING, "No channel layout specified. The "
670 "encoder will guess the layout, but it "
671 "might be incorrect.\n");
673 if (set_channel_info(s, avctx->channels, &avctx->channel_layout)) {
674 av_log(avctx, AV_LOG_ERROR, "invalid channel layout\n");
681 if ((ff_ac3_sample_rate_tab[j] >> i) == freq)
686 s->sample_rate = freq;
689 s->bitstream_id = 8 + s->sr_shift;
690 s->bitstream_mode = 0; /* complete main audio service */
692 /* bitrate & frame size */
694 if ((ff_ac3_bitrate_tab[i] >> s->sr_shift)*1000 == bitrate)
699 s->bit_rate = bitrate;
700 s->frame_size_code = i << 1;
701 s->frame_size_min = ff_ac3_frame_size_tab[s->frame_size_code][s->sr_code];
703 s->samples_written = 0;
704 s->frame_size = s->frame_size_min;
706 /* bit allocation init */
708 /* calculate bandwidth based on user-specified cutoff frequency */
709 int cutoff = av_clip(avctx->cutoff, 1, s->sample_rate >> 1);
710 int fbw_coeffs = cutoff * 512 / s->sample_rate;
711 bw_code = av_clip((fbw_coeffs - 73) / 3, 0, 60);
713 /* use default bandwidth setting */
714 /* XXX: should compute the bandwidth according to the frame
715 size, so that we avoid annoying high frequency artifacts */
718 for(ch=0;ch<s->nb_channels;ch++) {
719 /* bandwidth for each channel */
720 s->chbwcod[ch] = bw_code;
721 s->nb_coefs[ch] = bw_code * 3 + 73;
724 s->nb_coefs[s->lfe_channel] = 7; /* fixed */
726 /* initial snr offset */
727 s->coarse_snr_offset = 40;
730 fft_init(MDCT_NBITS - 2);
732 alpha = 2 * M_PI * (i + 1.0 / 8.0) / (float)N;
733 xcos1[i] = fix15(-cos(alpha));
734 xsin1[i] = fix15(-sin(alpha));
737 avctx->coded_frame= avcodec_alloc_frame();
738 avctx->coded_frame->key_frame= 1;
743 /* output the AC-3 frame header */
744 static void output_frame_header(AC3EncodeContext *s, unsigned char *frame)
746 init_put_bits(&s->pb, frame, AC3_MAX_CODED_FRAME_SIZE);
748 put_bits(&s->pb, 16, 0x0b77); /* frame header */
749 put_bits(&s->pb, 16, 0); /* crc1: will be filled later */
750 put_bits(&s->pb, 2, s->sr_code);
751 put_bits(&s->pb, 6, s->frame_size_code + (s->frame_size - s->frame_size_min));
752 put_bits(&s->pb, 5, s->bitstream_id);
753 put_bits(&s->pb, 3, s->bitstream_mode);
754 put_bits(&s->pb, 3, s->channel_mode);
755 if ((s->channel_mode & 0x01) && s->channel_mode != AC3_CHMODE_MONO)
756 put_bits(&s->pb, 2, 1); /* XXX -4.5 dB */
757 if (s->channel_mode & 0x04)
758 put_bits(&s->pb, 2, 1); /* XXX -6 dB */
759 if (s->channel_mode == AC3_CHMODE_STEREO)
760 put_bits(&s->pb, 2, 0); /* surround not indicated */
761 put_bits(&s->pb, 1, s->lfe); /* LFE */
762 put_bits(&s->pb, 5, 31); /* dialog norm: -31 db */
763 put_bits(&s->pb, 1, 0); /* no compression control word */
764 put_bits(&s->pb, 1, 0); /* no lang code */
765 put_bits(&s->pb, 1, 0); /* no audio production info */
766 put_bits(&s->pb, 1, 0); /* no copyright */
767 put_bits(&s->pb, 1, 1); /* original bitstream */
768 put_bits(&s->pb, 1, 0); /* no time code 1 */
769 put_bits(&s->pb, 1, 0); /* no time code 2 */
770 put_bits(&s->pb, 1, 0); /* no additional bit stream info */
773 /* symetric quantization on 'levels' levels */
774 static inline int sym_quant(int c, int e, int levels)
779 v = (levels * (c << e)) >> 24;
781 v = (levels >> 1) + v;
783 v = (levels * ((-c) << e)) >> 24;
785 v = (levels >> 1) - v;
787 assert (v >= 0 && v < levels);
791 /* asymetric quantization on 2^qbits levels */
792 static inline int asym_quant(int c, int e, int qbits)
796 lshift = e + qbits - 24;
803 m = (1 << (qbits-1));
807 return v & ((1 << qbits)-1);
810 /* Output one audio block. There are NB_BLOCKS audio blocks in one AC-3
812 static void output_audio_block(AC3EncodeContext *s,
813 uint8_t exp_strategy[AC3_MAX_CHANNELS],
814 uint8_t encoded_exp[AC3_MAX_CHANNELS][N/2],
815 uint8_t bap[AC3_MAX_CHANNELS][N/2],
816 int32_t mdct_coefs[AC3_MAX_CHANNELS][N/2],
817 int8_t global_exp[AC3_MAX_CHANNELS],
820 int ch, nb_groups, group_size, i, baie, rbnd;
822 uint16_t qmant[AC3_MAX_CHANNELS][N/2];
824 int mant1_cnt, mant2_cnt, mant4_cnt;
825 uint16_t *qmant1_ptr, *qmant2_ptr, *qmant4_ptr;
826 int delta0, delta1, delta2;
828 for(ch=0;ch<s->nb_channels;ch++)
829 put_bits(&s->pb, 1, 0); /* 512 point MDCT */
830 for(ch=0;ch<s->nb_channels;ch++)
831 put_bits(&s->pb, 1, 1); /* no dither */
832 put_bits(&s->pb, 1, 0); /* no dynamic range */
833 if (block_num == 0) {
834 /* for block 0, even if no coupling, we must say it. This is a
836 put_bits(&s->pb, 1, 1); /* coupling strategy present */
837 put_bits(&s->pb, 1, 0); /* no coupling strategy */
839 put_bits(&s->pb, 1, 0); /* no new coupling strategy */
842 if (s->channel_mode == AC3_CHMODE_STEREO)
846 /* first block must define rematrixing (rematstr) */
847 put_bits(&s->pb, 1, 1);
849 /* dummy rematrixing rematflg(1:4)=0 */
850 for (rbnd=0;rbnd<4;rbnd++)
851 put_bits(&s->pb, 1, 0);
855 /* no matrixing (but should be used in the future) */
856 put_bits(&s->pb, 1, 0);
862 static int count = 0;
863 av_log(NULL, AV_LOG_DEBUG, "Block #%d (%d)\n", block_num, count++);
866 /* exponent strategy */
867 for(ch=0;ch<s->nb_channels;ch++) {
868 put_bits(&s->pb, 2, exp_strategy[ch]);
872 put_bits(&s->pb, 1, exp_strategy[s->lfe_channel]);
875 for(ch=0;ch<s->nb_channels;ch++) {
876 if (exp_strategy[ch] != EXP_REUSE)
877 put_bits(&s->pb, 6, s->chbwcod[ch]);
881 for (ch = 0; ch < s->nb_all_channels; ch++) {
882 switch(exp_strategy[ch]) {
896 nb_groups = (s->nb_coefs[ch] + (group_size * 3) - 4) / (3 * group_size);
901 put_bits(&s->pb, 4, exp1);
903 /* next ones are delta encoded */
904 for(i=0;i<nb_groups;i++) {
905 /* merge three delta in one code */
909 delta0 = exp1 - exp0 + 2;
914 delta1 = exp1 - exp0 + 2;
919 delta2 = exp1 - exp0 + 2;
921 put_bits(&s->pb, 7, ((delta0 * 5 + delta1) * 5) + delta2);
924 if (ch != s->lfe_channel)
925 put_bits(&s->pb, 2, 0); /* no gain range info */
928 /* bit allocation info */
929 baie = (block_num == 0);
930 put_bits(&s->pb, 1, baie);
932 put_bits(&s->pb, 2, s->slow_decay_code);
933 put_bits(&s->pb, 2, s->fast_decay_code);
934 put_bits(&s->pb, 2, s->slow_gain_code);
935 put_bits(&s->pb, 2, s->db_per_bit_code);
936 put_bits(&s->pb, 3, s->floor_code);
940 put_bits(&s->pb, 1, baie); /* always present with bai */
942 put_bits(&s->pb, 6, s->coarse_snr_offset);
943 for(ch=0;ch<s->nb_all_channels;ch++) {
944 put_bits(&s->pb, 4, s->fine_snr_offset[ch]);
945 put_bits(&s->pb, 3, s->fast_gain_code[ch]);
949 put_bits(&s->pb, 1, 0); /* no delta bit allocation */
950 put_bits(&s->pb, 1, 0); /* no data to skip */
952 /* mantissa encoding : we use two passes to handle the grouping. A
953 one pass method may be faster, but it would necessitate to
954 modify the output stream. */
956 /* first pass: quantize */
957 mant1_cnt = mant2_cnt = mant4_cnt = 0;
958 qmant1_ptr = qmant2_ptr = qmant4_ptr = NULL;
960 for (ch = 0; ch < s->nb_all_channels; ch++) {
963 for(i=0;i<s->nb_coefs[ch];i++) {
964 c = mdct_coefs[ch][i];
965 e = encoded_exp[ch][i] - global_exp[ch];
972 v = sym_quant(c, e, 3);
975 qmant1_ptr = &qmant[ch][i];
980 *qmant1_ptr += 3 * v;
992 v = sym_quant(c, e, 5);
995 qmant2_ptr = &qmant[ch][i];
1000 *qmant2_ptr += 5 * v;
1012 v = sym_quant(c, e, 7);
1015 v = sym_quant(c, e, 11);
1018 qmant4_ptr = &qmant[ch][i];
1030 v = sym_quant(c, e, 15);
1033 v = asym_quant(c, e, 14);
1036 v = asym_quant(c, e, 16);
1039 v = asym_quant(c, e, b - 1);
1046 /* second pass : output the values */
1047 for (ch = 0; ch < s->nb_all_channels; ch++) {
1050 for(i=0;i<s->nb_coefs[ch];i++) {
1058 put_bits(&s->pb, 5, q);
1062 put_bits(&s->pb, 7, q);
1065 put_bits(&s->pb, 3, q);
1069 put_bits(&s->pb, 7, q);
1072 put_bits(&s->pb, 14, q);
1075 put_bits(&s->pb, 16, q);
1078 put_bits(&s->pb, b - 1, q);
1085 #define CRC16_POLY ((1 << 0) | (1 << 2) | (1 << 15) | (1 << 16))
1087 static unsigned int mul_poly(unsigned int a, unsigned int b, unsigned int poly)
1103 static unsigned int pow_poly(unsigned int a, unsigned int n, unsigned int poly)
1109 r = mul_poly(r, a, poly);
1110 a = mul_poly(a, a, poly);
1117 /* compute log2(max(abs(tab[]))) */
1118 static int log2_tab(int16_t *tab, int n)
1129 static void lshift_tab(int16_t *tab, int n, int lshift)
1137 } else if (lshift < 0) {
1145 /* fill the end of the frame and compute the two crcs */
1146 static int output_frame_end(AC3EncodeContext *s)
1148 int frame_size, frame_size_58, n, crc1, crc2, crc_inv;
1151 frame_size = s->frame_size; /* frame size in words */
1152 /* align to 8 bits */
1153 flush_put_bits(&s->pb);
1154 /* add zero bytes to reach the frame size */
1156 n = 2 * s->frame_size - (put_bits_ptr(&s->pb) - frame) - 2;
1159 memset(put_bits_ptr(&s->pb), 0, n);
1161 /* Now we must compute both crcs : this is not so easy for crc1
1162 because it is at the beginning of the data... */
1163 frame_size_58 = (frame_size >> 1) + (frame_size >> 3);
1164 crc1 = bswap_16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1165 frame + 4, 2 * frame_size_58 - 4));
1166 /* XXX: could precompute crc_inv */
1167 crc_inv = pow_poly((CRC16_POLY >> 1), (16 * frame_size_58) - 16, CRC16_POLY);
1168 crc1 = mul_poly(crc_inv, crc1, CRC16_POLY);
1169 AV_WB16(frame+2,crc1);
1171 crc2 = bswap_16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1172 frame + 2 * frame_size_58,
1173 (frame_size - frame_size_58) * 2 - 2));
1174 AV_WB16(frame+2*frame_size-2,crc2);
1176 // printf("n=%d frame_size=%d\n", n, frame_size);
1177 return frame_size * 2;
1180 static int AC3_encode_frame(AVCodecContext *avctx,
1181 unsigned char *frame, int buf_size, void *data)
1183 AC3EncodeContext *s = avctx->priv_data;
1184 int16_t *samples = data;
1186 int16_t input_samples[N];
1187 int32_t mdct_coef[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1188 uint8_t exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1189 uint8_t exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS];
1190 uint8_t encoded_exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1191 uint8_t bap[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1192 int8_t exp_samples[NB_BLOCKS][AC3_MAX_CHANNELS];
1196 for(ch=0;ch<s->nb_all_channels;ch++) {
1197 int ich = s->channel_map[ch];
1198 /* fixed mdct to the six sub blocks & exponent computation */
1199 for(i=0;i<NB_BLOCKS;i++) {
1203 /* compute input samples */
1204 memcpy(input_samples, s->last_samples[ich], N/2 * sizeof(int16_t));
1205 sinc = s->nb_all_channels;
1206 sptr = samples + (sinc * (N/2) * i) + ich;
1207 for(j=0;j<N/2;j++) {
1209 input_samples[j + N/2] = v;
1210 s->last_samples[ich][j] = v;
1214 /* apply the MDCT window */
1215 for(j=0;j<N/2;j++) {
1216 input_samples[j] = MUL16(input_samples[j],
1217 ff_ac3_window[j]) >> 15;
1218 input_samples[N-j-1] = MUL16(input_samples[N-j-1],
1219 ff_ac3_window[j]) >> 15;
1222 /* Normalize the samples to use the maximum available
1224 v = 14 - log2_tab(input_samples, N);
1227 exp_samples[i][ch] = v - 9;
1228 lshift_tab(input_samples, N, v);
1231 mdct512(mdct_coef[i][ch], input_samples);
1233 /* compute "exponents". We take into account the
1234 normalization there */
1235 for(j=0;j<N/2;j++) {
1237 v = abs(mdct_coef[i][ch][j]);
1241 e = 23 - av_log2(v) + exp_samples[i][ch];
1244 mdct_coef[i][ch][j] = 0;
1251 compute_exp_strategy(exp_strategy, exp, ch, ch == s->lfe_channel);
1253 /* compute the exponents as the decoder will see them. The
1254 EXP_REUSE case must be handled carefully : we select the
1255 min of the exponents */
1257 while (i < NB_BLOCKS) {
1259 while (j < NB_BLOCKS && exp_strategy[j][ch] == EXP_REUSE) {
1260 exponent_min(exp[i][ch], exp[j][ch], s->nb_coefs[ch]);
1263 frame_bits += encode_exp(encoded_exp[i][ch],
1264 exp[i][ch], s->nb_coefs[ch],
1265 exp_strategy[i][ch]);
1266 /* copy encoded exponents for reuse case */
1267 for(k=i+1;k<j;k++) {
1268 memcpy(encoded_exp[k][ch], encoded_exp[i][ch],
1269 s->nb_coefs[ch] * sizeof(uint8_t));
1275 /* adjust for fractional frame sizes */
1276 while(s->bits_written >= s->bit_rate && s->samples_written >= s->sample_rate) {
1277 s->bits_written -= s->bit_rate;
1278 s->samples_written -= s->sample_rate;
1280 s->frame_size = s->frame_size_min + (s->bits_written * s->sample_rate < s->samples_written * s->bit_rate);
1281 s->bits_written += s->frame_size * 16;
1282 s->samples_written += AC3_FRAME_SIZE;
1284 compute_bit_allocation(s, bap, encoded_exp, exp_strategy, frame_bits);
1285 /* everything is known... let's output the frame */
1286 output_frame_header(s, frame);
1288 for(i=0;i<NB_BLOCKS;i++) {
1289 output_audio_block(s, exp_strategy[i], encoded_exp[i],
1290 bap[i], mdct_coef[i], exp_samples[i], i);
1292 return output_frame_end(s);
1295 static av_cold int AC3_encode_close(AVCodecContext *avctx)
1297 av_freep(&avctx->coded_frame);
1302 /*************************************************************************/
1310 IComplex in[FN], in1[FN];
1312 float sum_re, sum_im, a;
1317 in[i].re = random() % 65535 - 32767;
1318 in[i].im = random() % 65535 - 32767;
1328 a = -2 * M_PI * (n * k) / FN;
1329 sum_re += in1[n].re * cos(a) - in1[n].im * sin(a);
1330 sum_im += in1[n].re * sin(a) + in1[n].im * cos(a);
1332 printf("%3d: %6d,%6d %6.0f,%6.0f\n",
1333 k, in[k].re, in[k].im, sum_re / FN, sum_im / FN);
1337 void mdct_test(void)
1340 int32_t output[N/2];
1343 float s, a, err, e, emax;
1347 input[i] = (random() % 65535 - 32767) * 9 / 10;
1348 input1[i] = input[i];
1351 mdct512(output, input);
1354 for(k=0;k<N/2;k++) {
1357 a = (2*M_PI*(2*n+1+N/2)*(2*k+1) / (4 * N));
1358 s += input1[n] * cos(a);
1360 output1[k] = -2 * s / N;
1365 for(i=0;i<N/2;i++) {
1366 printf("%3d: %7d %7.0f\n", i, output[i], output1[i]);
1367 e = output[i] - output1[i];
1372 printf("err2=%f emax=%f\n", err / (N/2), emax);
1377 AC3EncodeContext ctx;
1378 unsigned char frame[AC3_MAX_CODED_FRAME_SIZE];
1379 short samples[AC3_FRAME_SIZE];
1382 AC3_encode_init(&ctx, 44100, 64000, 1);
1387 for(i=0;i<AC3_FRAME_SIZE;i++)
1388 samples[i] = (int)(sin(2*M_PI*i*1000.0/44100) * 10000);
1389 ret = AC3_encode_frame(&ctx, frame, samples);
1390 printf("ret=%d\n", ret);
1394 AVCodec ac3_encoder = {
1398 sizeof(AC3EncodeContext),
1403 .sample_fmts = (const enum SampleFormat[]){SAMPLE_FMT_S16,SAMPLE_FMT_NONE},
1404 .long_name = NULL_IF_CONFIG_SMALL("ATSC A/52A (AC-3)"),
1405 .channel_layouts = (const int64_t[]){
1414 CH_LAYOUT_5POINT0_BACK,
1415 (CH_LAYOUT_MONO | CH_LOW_FREQUENCY),
1416 (CH_LAYOUT_STEREO | CH_LOW_FREQUENCY),
1417 (CH_LAYOUT_2_1 | CH_LOW_FREQUENCY),
1418 (CH_LAYOUT_SURROUND | CH_LOW_FREQUENCY),
1419 (CH_LAYOUT_2_2 | CH_LOW_FREQUENCY),
1420 (CH_LAYOUT_QUAD | CH_LOW_FREQUENCY),
1421 (CH_LAYOUT_4POINT0 | CH_LOW_FREQUENCY),
1423 CH_LAYOUT_5POINT1_BACK,