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
24 * The simplest AC-3 encoder.
29 #include "libavcore/audioconvert.h"
30 #include "libavutil/crc.h"
34 #include "audioconvert.h"
38 #define MDCT_SAMPLES (1 << MDCT_NBITS)
40 /** Scale a float value by 2^bits and convert to an integer. */
41 #define SCALE_FLOAT(a, bits) lrintf((a) * (float)(1 << (bits)))
43 /** Scale a float value by 2^15, convert to an integer, and clip to int16_t range. */
44 #define FIX15(a) av_clip_int16(SCALE_FLOAT(a, 15))
49 * Used in fixed-point MDCT calculation.
51 typedef struct IComplex {
56 * AC-3 encoder private context.
58 typedef struct AC3EncodeContext {
59 PutBitContext pb; ///< bitstream writer context
61 int bitstream_id; ///< bitstream id (bsid)
62 int bitstream_mode; ///< bitstream mode (bsmod)
64 int bit_rate; ///< target bit rate, in bits-per-second
65 int sample_rate; ///< sampling frequency, in Hz
67 int frame_size_min; ///< minimum frame size in case rounding is necessary
68 int frame_size; ///< current frame size in words
69 int frame_size_code; ///< frame size code (frmsizecod)
70 int bits_written; ///< bit count (used to avg. bitrate)
71 int samples_written; ///< sample count (used to avg. bitrate)
73 int fbw_channels; ///< number of full-bandwidth channels (nfchans)
74 int channels; ///< total number of channels (nchans)
75 int lfe_on; ///< indicates if there is an LFE channel (lfeon)
76 int lfe_channel; ///< channel index of the LFE channel
77 int channel_mode; ///< channel mode (acmod)
78 const uint8_t *channel_map; ///< channel map used to reorder channels
80 int bandwidth_code[AC3_MAX_CHANNELS]; ///< bandwidth code (0 to 60) (chbwcod)
81 int nb_coefs[AC3_MAX_CHANNELS];
83 /* bitrate allocation control */
84 int slow_gain_code; ///< slow gain code (sgaincod)
85 int slow_decay_code; ///< slow decay code (sdcycod)
86 int fast_decay_code; ///< fast decay code (fdcycod)
87 int db_per_bit_code; ///< dB/bit code (dbpbcod)
88 int floor_code; ///< floor code (floorcod)
89 AC3BitAllocParameters bit_alloc; ///< bit allocation parameters
90 int coarse_snr_offset; ///< coarse SNR offsets (csnroffst)
91 int fast_gain_code[AC3_MAX_CHANNELS]; ///< fast gain codes (signal-to-mask ratio) (fgaincod)
92 int fine_snr_offset[AC3_MAX_CHANNELS]; ///< fine SNR offsets (fsnroffst)
94 /* mantissa encoding */
95 int mant1_cnt, mant2_cnt, mant4_cnt; ///< mantissa counts for bap=1,2,4
97 int16_t last_samples[AC3_MAX_CHANNELS][AC3_BLOCK_SIZE]; ///< last 256 samples from previous frame
101 /** MDCT and FFT tables */
102 static int16_t costab[64];
103 static int16_t sintab[64];
104 static int16_t xcos1[128];
105 static int16_t xsin1[128];
109 * Initialize FFT tables.
110 * @param ln log2(FFT size)
112 static av_cold void fft_init(int ln)
120 for (i = 0; i < n2; i++) {
121 alpha = 2.0 * M_PI * i / n;
122 costab[i] = FIX15(cos(alpha));
123 sintab[i] = FIX15(sin(alpha));
129 * Initialize MDCT tables.
130 * @param nbits log2(MDCT size)
132 static av_cold void mdct_init(int nbits)
141 for (i = 0; i < n4; i++) {
142 float alpha = 2.0 * M_PI * (i + 1.0 / 8.0) / n;
143 xcos1[i] = FIX15(-cos(alpha));
144 xsin1[i] = FIX15(-sin(alpha));
150 #define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \
152 int ax, ay, bx, by; \
157 pre = (bx + ax) >> 1; \
158 pim = (by + ay) >> 1; \
159 qre = (bx - ax) >> 1; \
160 qim = (by - ay) >> 1; \
164 /** Complex multiply */
165 #define CMUL(pre, pim, are, aim, bre, bim) \
167 pre = (MUL16(are, bre) - MUL16(aim, bim)) >> 15; \
168 pim = (MUL16(are, bim) + MUL16(bre, aim)) >> 15; \
173 * Calculate a 2^n point complex FFT on 2^ln points.
174 * @param z complex input/output samples
175 * @param ln log2(FFT size)
177 static void fft(IComplex *z, int ln)
181 register IComplex *p,*q;
187 for (j = 0; j < np; j++) {
188 int k = av_reverse[j] >> (8 - ln);
190 FFSWAP(IComplex, z[k], z[j]);
198 BF(p[0].re, p[0].im, p[1].re, p[1].im,
199 p[0].re, p[0].im, p[1].re, p[1].im);
208 BF(p[0].re, p[0].im, p[2].re, p[2].im,
209 p[0].re, p[0].im, p[2].re, p[2].im);
210 BF(p[1].re, p[1].im, p[3].re, p[3].im,
211 p[1].re, p[1].im, p[3].im, -p[3].re);
223 for (j = 0; j < nblocks; j++) {
224 BF(p->re, p->im, q->re, q->im,
225 p->re, p->im, q->re, q->im);
228 for(l = nblocks; l < np2; l += nblocks) {
229 CMUL(tmp_re, tmp_im, costab[l], -sintab[l], q->re, q->im);
230 BF(p->re, p->im, q->re, q->im,
231 p->re, p->im, tmp_re, tmp_im);
238 nblocks = nblocks >> 1;
239 nloops = nloops << 1;
245 * Calculate a 512-point MDCT
246 * @param out 256 output frequency coefficients
247 * @param in 512 windowed input audio samples
249 static void mdct512(int32_t *out, int16_t *in)
251 int i, re, im, re1, im1;
252 int16_t rot[MDCT_SAMPLES];
253 IComplex x[MDCT_SAMPLES/4];
255 /* shift to simplify computations */
256 for (i = 0; i < MDCT_SAMPLES/4; i++)
257 rot[i] = -in[i + 3*MDCT_SAMPLES/4];
258 for (;i < MDCT_SAMPLES; i++)
259 rot[i] = in[i - MDCT_SAMPLES/4];
262 for (i = 0; i < MDCT_SAMPLES/4; i++) {
263 re = ((int)rot[ 2*i] - (int)rot[MDCT_SAMPLES -1-2*i]) >> 1;
264 im = -((int)rot[MDCT_SAMPLES/2+2*i] - (int)rot[MDCT_SAMPLES/2-1-2*i]) >> 1;
265 CMUL(x[i].re, x[i].im, re, im, -xcos1[i], xsin1[i]);
268 fft(x, MDCT_NBITS - 2);
271 for (i = 0; i < MDCT_SAMPLES/4; i++) {
274 CMUL(re1, im1, re, im, xsin1[i], xcos1[i]);
276 out[MDCT_SAMPLES/2-1-2*i] = re1;
282 * Calculate the log2() of the maximum absolute value in an array.
283 * @param tab input array
284 * @param n number of values in the array
285 * @return log2(max(abs(tab[])))
287 static int log2_tab(int16_t *tab, int n)
292 for (i = 0; i < n; i++)
300 * Left-shift each value in an array by a specified amount.
301 * @param tab input array
302 * @param n number of values in the array
303 * @param lshift left shift amount. a negative value means right shift.
305 static void lshift_tab(int16_t *tab, int n, int lshift)
310 for(i = 0; i < n; i++)
312 } else if (lshift < 0) {
314 for (i = 0; i < n; i++)
321 * Calculate the sum of absolute differences (SAD) between 2 sets of exponents.
323 static int calc_exp_diff(uint8_t *exp1, uint8_t *exp2, int n)
327 for (i = 0; i < n; i++)
328 sum += abs(exp1[i] - exp2[i]);
334 * Exponent Difference Threshold.
335 * New exponents are sent if their SAD exceed this number.
337 #define EXP_DIFF_THRESHOLD 1000
341 * Calculate exponent strategies for all blocks in a single channel.
343 static void compute_exp_strategy(uint8_t exp_strategy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
344 uint8_t exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
350 /* estimate if the exponent variation & decide if they should be
351 reused in the next frame */
352 exp_strategy[0][ch] = EXP_NEW;
353 for (i = 1; i < AC3_MAX_BLOCKS; i++) {
354 exp_diff = calc_exp_diff(exp[i][ch], exp[i-1][ch], AC3_MAX_COEFS);
355 if (exp_diff > EXP_DIFF_THRESHOLD)
356 exp_strategy[i][ch] = EXP_NEW;
358 exp_strategy[i][ch] = EXP_REUSE;
363 /* now select the encoding strategy type : if exponents are often
364 recoded, we use a coarse encoding */
366 while (i < AC3_MAX_BLOCKS) {
368 while (j < AC3_MAX_BLOCKS && exp_strategy[j][ch] == EXP_REUSE)
371 case 1: exp_strategy[i][ch] = EXP_D45; break;
373 case 3: exp_strategy[i][ch] = EXP_D25; break;
374 default: exp_strategy[i][ch] = EXP_D15; break;
382 * Set each encoded exponent in a block to the minimum of itself and the
383 * exponent in the same frequency bin of a following block.
384 * exp[i] = min(exp[i], exp1[i]
386 static void exponent_min(uint8_t exp[AC3_MAX_COEFS], uint8_t exp1[AC3_MAX_COEFS], int n)
389 for (i = 0; i < n; i++) {
390 if (exp1[i] < exp[i])
397 * Update the exponents so that they are the ones the decoder will decode.
398 * @return the number of bits used to encode the exponents.
400 static int encode_exp(uint8_t encoded_exp[AC3_MAX_COEFS],
401 uint8_t exp[AC3_MAX_COEFS],
402 int nb_exps, int exp_strategy)
404 int group_size, nb_groups, i, j, k, exp_min;
405 uint8_t exp1[AC3_MAX_COEFS];
407 group_size = exp_strategy + (exp_strategy == EXP_D45);
408 nb_groups = ((nb_exps + (group_size * 3) - 4) / (3 * group_size)) * 3;
410 /* for each group, compute the minimum exponent */
411 exp1[0] = exp[0]; /* DC exponent is handled separately */
413 for (i = 1; i <= nb_groups; i++) {
415 assert(exp_min >= 0 && exp_min <= 24);
416 for (j = 1; j < group_size; j++) {
417 if (exp[k+j] < exp_min)
424 /* constraint for DC exponent */
428 /* decrease the delta between each groups to within 2 so that they can be
429 differentially encoded */
430 for (i = 1; i <= nb_groups; i++)
431 exp1[i] = FFMIN(exp1[i], exp1[i-1] + 2);
432 for (i = nb_groups-1; i >= 0; i--)
433 exp1[i] = FFMIN(exp1[i], exp1[i+1] + 2);
435 /* now we have the exponent values the decoder will see */
436 encoded_exp[0] = exp1[0];
438 for (i = 1; i <= nb_groups; i++) {
439 for (j = 0; j < group_size; j++)
440 encoded_exp[k+j] = exp1[i];
444 return 4 + (nb_groups / 3) * 7;
449 * Calculate the number of bits needed to encode a set of mantissas.
451 static int compute_mantissa_size(AC3EncodeContext *s, uint8_t *m, int nb_coefs)
456 for (i = 0; i < nb_coefs; i++) {
463 /* 3 mantissa in 5 bits */
464 if (s->mant1_cnt == 0)
466 if (++s->mant1_cnt == 3)
470 /* 3 mantissa in 7 bits */
471 if (s->mant2_cnt == 0)
473 if (++s->mant2_cnt == 3)
480 /* 2 mantissa in 7 bits */
481 if (s->mant4_cnt == 0)
483 if (++s->mant4_cnt == 2)
502 * Calculate masking curve based on the final exponents.
503 * Also calculate the power spectral densities to use in future calculations.
505 static void bit_alloc_masking(AC3EncodeContext *s,
506 uint8_t encoded_exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
507 uint8_t exp_strategy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
508 int16_t psd[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
509 int16_t mask[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][50])
512 int16_t band_psd[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][50];
514 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
515 for (ch = 0; ch < s->channels; ch++) {
516 if(exp_strategy[blk][ch] == EXP_REUSE) {
517 memcpy(psd[blk][ch], psd[blk-1][ch], AC3_MAX_COEFS*sizeof(psd[0][0][0]));
518 memcpy(mask[blk][ch], mask[blk-1][ch], 50*sizeof(mask[0][0][0]));
520 ff_ac3_bit_alloc_calc_psd(encoded_exp[blk][ch], 0,
522 psd[blk][ch], band_psd[blk][ch]);
523 ff_ac3_bit_alloc_calc_mask(&s->bit_alloc, band_psd[blk][ch],
525 ff_ac3_fast_gain_tab[s->fast_gain_code[ch]],
526 ch == s->lfe_channel,
527 DBA_NONE, 0, NULL, NULL, NULL,
536 * Run the bit allocation with a given SNR offset.
537 * This calculates the bit allocation pointers that will be used to determine
538 * the quantization of each mantissa.
539 * @return the number of remaining bits (positive or negative) if the given
540 * SNR offset is used to quantize the mantissas.
542 static int bit_alloc(AC3EncodeContext *s,
543 int16_t mask[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][50],
544 int16_t psd[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
545 uint8_t bap[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
546 int frame_bits, int coarse_snr_offset, int fine_snr_offset)
551 snr_offset = (((coarse_snr_offset - 15) << 4) + fine_snr_offset) << 2;
553 for (i = 0; i < AC3_MAX_BLOCKS; i++) {
557 for (ch = 0; ch < s->channels; ch++) {
558 ff_ac3_bit_alloc_calc_bap(mask[i][ch], psd[i][ch], 0,
559 s->nb_coefs[ch], snr_offset,
560 s->bit_alloc.floor, ff_ac3_bap_tab,
562 frame_bits += compute_mantissa_size(s, bap[i][ch], s->nb_coefs[ch]);
565 return 16 * s->frame_size - frame_bits;
572 * Perform bit allocation search.
573 * Finds the SNR offset value that maximizes quality and fits in the specified
574 * frame size. Output is the SNR offset and a set of bit allocation pointers
575 * used to quantize the mantissas.
577 static int compute_bit_allocation(AC3EncodeContext *s,
578 uint8_t bap[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
579 uint8_t encoded_exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS],
580 uint8_t exp_strategy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS],
584 int coarse_snr_offset, fine_snr_offset;
585 uint8_t bap1[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS];
586 int16_t psd[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS];
587 int16_t mask[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][50];
588 static const int frame_bits_inc[8] = { 0, 0, 2, 2, 2, 4, 2, 4 };
590 /* init default parameters */
591 s->slow_decay_code = 2;
592 s->fast_decay_code = 1;
593 s->slow_gain_code = 1;
594 s->db_per_bit_code = 2;
596 for (ch = 0; ch < s->channels; ch++)
597 s->fast_gain_code[ch] = 4;
599 /* compute real values */
600 s->bit_alloc.slow_decay = ff_ac3_slow_decay_tab[s->slow_decay_code] >> s->bit_alloc.sr_shift;
601 s->bit_alloc.fast_decay = ff_ac3_fast_decay_tab[s->fast_decay_code] >> s->bit_alloc.sr_shift;
602 s->bit_alloc.slow_gain = ff_ac3_slow_gain_tab[s->slow_gain_code];
603 s->bit_alloc.db_per_bit = ff_ac3_db_per_bit_tab[s->db_per_bit_code];
604 s->bit_alloc.floor = ff_ac3_floor_tab[s->floor_code];
608 // if (s->channel_mode == 2)
610 frame_bits += frame_bits_inc[s->channel_mode];
613 for (i = 0; i < AC3_MAX_BLOCKS; i++) {
614 frame_bits += s->fbw_channels * 2 + 2; /* blksw * c, dithflag * c, dynrnge, cplstre */
615 if (s->channel_mode == AC3_CHMODE_STEREO) {
616 frame_bits++; /* rematstr */
620 frame_bits += 2 * s->fbw_channels; /* chexpstr[2] * c */
622 frame_bits++; /* lfeexpstr */
623 for (ch = 0; ch < s->fbw_channels; ch++) {
624 if (exp_strategy[i][ch] != EXP_REUSE)
625 frame_bits += 6 + 2; /* chbwcod[6], gainrng[2] */
627 frame_bits++; /* baie */
628 frame_bits++; /* snr */
629 frame_bits += 2; /* delta / skip */
631 frame_bits++; /* cplinu for block 0 */
633 /* sdcycod[2], fdcycod[2], sgaincod[2], dbpbcod[2], floorcod[3] */
635 /* (fsnoffset[4] + fgaincod[4]) * c */
636 frame_bits += 2*4 + 3 + 6 + s->channels * (4 + 3);
638 /* auxdatae, crcrsv */
644 /* calculate psd and masking curve before doing bit allocation */
645 bit_alloc_masking(s, encoded_exp, exp_strategy, psd, mask);
647 /* now the big work begins : do the bit allocation. Modify the snr
648 offset until we can pack everything in the requested frame size */
650 coarse_snr_offset = s->coarse_snr_offset;
651 while (coarse_snr_offset >= 0 &&
652 bit_alloc(s, mask, psd, bap, frame_bits, coarse_snr_offset, 0) < 0)
653 coarse_snr_offset -= SNR_INC1;
654 if (coarse_snr_offset < 0) {
655 av_log(NULL, AV_LOG_ERROR, "Bit allocation failed. Try increasing the bitrate.\n");
658 while (coarse_snr_offset + SNR_INC1 <= 63 &&
659 bit_alloc(s, mask, psd, bap1, frame_bits,
660 coarse_snr_offset + SNR_INC1, 0) >= 0) {
661 coarse_snr_offset += SNR_INC1;
662 memcpy(bap, bap1, sizeof(bap1));
664 while (coarse_snr_offset + 1 <= 63 &&
665 bit_alloc(s, mask, psd, bap1, frame_bits, coarse_snr_offset + 1, 0) >= 0) {
667 memcpy(bap, bap1, sizeof(bap1));
671 while (fine_snr_offset + SNR_INC1 <= 15 &&
672 bit_alloc(s, mask, psd, bap1, frame_bits,
673 coarse_snr_offset, fine_snr_offset + SNR_INC1) >= 0) {
674 fine_snr_offset += SNR_INC1;
675 memcpy(bap, bap1, sizeof(bap1));
677 while (fine_snr_offset + 1 <= 15 &&
678 bit_alloc(s, mask, psd, bap1, frame_bits,
679 coarse_snr_offset, fine_snr_offset + 1) >= 0) {
681 memcpy(bap, bap1, sizeof(bap1));
684 s->coarse_snr_offset = coarse_snr_offset;
685 for (ch = 0; ch < s->channels; ch++)
686 s->fine_snr_offset[ch] = fine_snr_offset;
693 * Write the AC-3 frame header to the output bitstream.
695 static void output_frame_header(AC3EncodeContext *s, unsigned char *frame)
697 init_put_bits(&s->pb, frame, AC3_MAX_CODED_FRAME_SIZE);
699 put_bits(&s->pb, 16, 0x0b77); /* frame header */
700 put_bits(&s->pb, 16, 0); /* crc1: will be filled later */
701 put_bits(&s->pb, 2, s->bit_alloc.sr_code);
702 put_bits(&s->pb, 6, s->frame_size_code + (s->frame_size - s->frame_size_min));
703 put_bits(&s->pb, 5, s->bitstream_id);
704 put_bits(&s->pb, 3, s->bitstream_mode);
705 put_bits(&s->pb, 3, s->channel_mode);
706 if ((s->channel_mode & 0x01) && s->channel_mode != AC3_CHMODE_MONO)
707 put_bits(&s->pb, 2, 1); /* XXX -4.5 dB */
708 if (s->channel_mode & 0x04)
709 put_bits(&s->pb, 2, 1); /* XXX -6 dB */
710 if (s->channel_mode == AC3_CHMODE_STEREO)
711 put_bits(&s->pb, 2, 0); /* surround not indicated */
712 put_bits(&s->pb, 1, s->lfe_on); /* LFE */
713 put_bits(&s->pb, 5, 31); /* dialog norm: -31 db */
714 put_bits(&s->pb, 1, 0); /* no compression control word */
715 put_bits(&s->pb, 1, 0); /* no lang code */
716 put_bits(&s->pb, 1, 0); /* no audio production info */
717 put_bits(&s->pb, 1, 0); /* no copyright */
718 put_bits(&s->pb, 1, 1); /* original bitstream */
719 put_bits(&s->pb, 1, 0); /* no time code 1 */
720 put_bits(&s->pb, 1, 0); /* no time code 2 */
721 put_bits(&s->pb, 1, 0); /* no additional bit stream info */
726 * Symmetric quantization on 'levels' levels.
728 static inline int sym_quant(int c, int e, int levels)
733 v = (levels * (c << e)) >> 24;
735 v = (levels >> 1) + v;
737 v = (levels * ((-c) << e)) >> 24;
739 v = (levels >> 1) - v;
741 assert (v >= 0 && v < levels);
747 * Asymmetric quantization on 2^qbits levels.
749 static inline int asym_quant(int c, int e, int qbits)
753 lshift = e + qbits - 24;
760 m = (1 << (qbits-1));
764 return v & ((1 << qbits)-1);
769 * Write one audio block to the output bitstream.
771 static void output_audio_block(AC3EncodeContext *s,
772 uint8_t exp_strategy[AC3_MAX_CHANNELS],
773 uint8_t encoded_exp[AC3_MAX_CHANNELS][AC3_MAX_COEFS],
774 uint8_t bap[AC3_MAX_CHANNELS][AC3_MAX_COEFS],
775 int32_t mdct_coefs[AC3_MAX_CHANNELS][AC3_MAX_COEFS],
776 int8_t global_exp[AC3_MAX_CHANNELS],
779 int ch, nb_groups, group_size, i, baie, rbnd;
781 uint16_t qmant[AC3_MAX_CHANNELS][AC3_MAX_COEFS];
783 int mant1_cnt, mant2_cnt, mant4_cnt;
784 uint16_t *qmant1_ptr, *qmant2_ptr, *qmant4_ptr;
785 int delta0, delta1, delta2;
787 for (ch = 0; ch < s->fbw_channels; ch++)
788 put_bits(&s->pb, 1, 0); /* no block switching */
789 for (ch = 0; ch < s->fbw_channels; ch++)
790 put_bits(&s->pb, 1, 1); /* no dither */
791 put_bits(&s->pb, 1, 0); /* no dynamic range */
793 put_bits(&s->pb, 1, 1); /* coupling strategy present */
794 put_bits(&s->pb, 1, 0); /* no coupling strategy */
796 put_bits(&s->pb, 1, 0); /* no new coupling strategy */
799 if (s->channel_mode == AC3_CHMODE_STEREO) {
801 /* first block must define rematrixing (rematstr) */
802 put_bits(&s->pb, 1, 1);
804 /* dummy rematrixing rematflg(1:4)=0 */
805 for (rbnd = 0; rbnd < 4; rbnd++)
806 put_bits(&s->pb, 1, 0);
808 /* no matrixing (but should be used in the future) */
809 put_bits(&s->pb, 1, 0);
813 /* exponent strategy */
814 for (ch = 0; ch < s->fbw_channels; ch++)
815 put_bits(&s->pb, 2, exp_strategy[ch]);
818 put_bits(&s->pb, 1, exp_strategy[s->lfe_channel]);
821 for (ch = 0; ch < s->fbw_channels; ch++) {
822 if (exp_strategy[ch] != EXP_REUSE)
823 put_bits(&s->pb, 6, s->bandwidth_code[ch]);
827 for (ch = 0; ch < s->channels; ch++) {
828 if (exp_strategy[ch] == EXP_REUSE)
830 group_size = exp_strategy[ch] + (exp_strategy[ch] == EXP_D45);
831 nb_groups = (s->nb_coefs[ch] + (group_size * 3) - 4) / (3 * group_size);
836 put_bits(&s->pb, 4, exp1);
838 /* next ones are delta encoded */
839 for (i = 0; i < nb_groups; i++) {
840 /* merge three delta in one code */
844 delta0 = exp1 - exp0 + 2;
849 delta1 = exp1 - exp0 + 2;
854 delta2 = exp1 - exp0 + 2;
856 put_bits(&s->pb, 7, ((delta0 * 5 + delta1) * 5) + delta2);
859 if (ch != s->lfe_channel)
860 put_bits(&s->pb, 2, 0); /* no gain range info */
863 /* bit allocation info */
864 baie = (block_num == 0);
865 put_bits(&s->pb, 1, baie);
867 put_bits(&s->pb, 2, s->slow_decay_code);
868 put_bits(&s->pb, 2, s->fast_decay_code);
869 put_bits(&s->pb, 2, s->slow_gain_code);
870 put_bits(&s->pb, 2, s->db_per_bit_code);
871 put_bits(&s->pb, 3, s->floor_code);
875 put_bits(&s->pb, 1, baie);
877 put_bits(&s->pb, 6, s->coarse_snr_offset);
878 for (ch = 0; ch < s->channels; ch++) {
879 put_bits(&s->pb, 4, s->fine_snr_offset[ch]);
880 put_bits(&s->pb, 3, s->fast_gain_code[ch]);
884 put_bits(&s->pb, 1, 0); /* no delta bit allocation */
885 put_bits(&s->pb, 1, 0); /* no data to skip */
887 /* mantissa encoding : we use two passes to handle the grouping. A
888 one pass method may be faster, but it would necessitate to
889 modify the output stream. */
891 /* first pass: quantize */
892 mant1_cnt = mant2_cnt = mant4_cnt = 0;
893 qmant1_ptr = qmant2_ptr = qmant4_ptr = NULL;
895 for (ch = 0; ch < s->channels; ch++) {
898 for (i = 0; i < s->nb_coefs[ch]; i++) {
899 c = mdct_coefs[ch][i];
900 e = encoded_exp[ch][i] - global_exp[ch];
907 v = sym_quant(c, e, 3);
910 qmant1_ptr = &qmant[ch][i];
915 *qmant1_ptr += 3 * v;
927 v = sym_quant(c, e, 5);
930 qmant2_ptr = &qmant[ch][i];
935 *qmant2_ptr += 5 * v;
947 v = sym_quant(c, e, 7);
950 v = sym_quant(c, e, 11);
953 qmant4_ptr = &qmant[ch][i];
965 v = sym_quant(c, e, 15);
968 v = asym_quant(c, e, 14);
971 v = asym_quant(c, e, 16);
974 v = asym_quant(c, e, b - 1);
981 /* second pass : output the values */
982 for (ch = 0; ch < s->channels; ch++) {
985 for (i = 0; i < s->nb_coefs[ch]; i++) {
990 case 1: if (q != 128) put_bits(&s->pb, 5, q); break;
991 case 2: if (q != 128) put_bits(&s->pb, 7, q); break;
992 case 3: put_bits(&s->pb, 3, q); break;
993 case 4: if (q != 128) put_bits(&s->pb, 7, q); break;
994 case 14: put_bits(&s->pb, 14, q); break;
995 case 15: put_bits(&s->pb, 16, q); break;
996 default: put_bits(&s->pb, b-1, q); break;
1003 /** CRC-16 Polynomial */
1004 #define CRC16_POLY ((1 << 0) | (1 << 2) | (1 << 15) | (1 << 16))
1007 static unsigned int mul_poly(unsigned int a, unsigned int b, unsigned int poly)
1024 static unsigned int pow_poly(unsigned int a, unsigned int n, unsigned int poly)
1030 r = mul_poly(r, a, poly);
1031 a = mul_poly(a, a, poly);
1039 * Fill the end of the frame with 0's and compute the two CRCs.
1041 static int output_frame_end(AC3EncodeContext *s)
1043 int frame_size, frame_size_58, n, crc1, crc2, crc_inv;
1046 frame_size = s->frame_size; /* frame size in words */
1047 /* align to 8 bits */
1048 flush_put_bits(&s->pb);
1049 /* add zero bytes to reach the frame size */
1051 n = 2 * s->frame_size - (put_bits_ptr(&s->pb) - frame) - 2;
1054 memset(put_bits_ptr(&s->pb), 0, n);
1056 /* Now we must compute both crcs : this is not so easy for crc1
1057 because it is at the beginning of the data... */
1058 frame_size_58 = (frame_size >> 1) + (frame_size >> 3);
1060 crc1 = av_bswap16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1061 frame + 4, 2 * frame_size_58 - 4));
1063 /* XXX: could precompute crc_inv */
1064 crc_inv = pow_poly((CRC16_POLY >> 1), (16 * frame_size_58) - 16, CRC16_POLY);
1065 crc1 = mul_poly(crc_inv, crc1, CRC16_POLY);
1066 AV_WB16(frame + 2, crc1);
1068 crc2 = av_bswap16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1069 frame + 2 * frame_size_58,
1070 (frame_size - frame_size_58) * 2 - 2));
1071 AV_WB16(frame + 2*frame_size - 2, crc2);
1073 return frame_size * 2;
1078 * Encode a single AC-3 frame.
1080 static int ac3_encode_frame(AVCodecContext *avctx,
1081 unsigned char *frame, int buf_size, void *data)
1083 AC3EncodeContext *s = avctx->priv_data;
1084 const int16_t *samples = data;
1086 int16_t input_samples[AC3_WINDOW_SIZE];
1087 int32_t mdct_coef[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS];
1088 uint8_t exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS];
1089 uint8_t exp_strategy[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS];
1090 uint8_t encoded_exp[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS];
1091 uint8_t bap[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS][AC3_MAX_COEFS];
1092 int8_t exp_samples[AC3_MAX_BLOCKS][AC3_MAX_CHANNELS];
1096 for (ch = 0; ch < s->channels; ch++) {
1097 int ich = s->channel_map[ch];
1098 /* fixed mdct to the six sub blocks & exponent computation */
1099 for (i = 0; i < AC3_MAX_BLOCKS; i++) {
1100 const int16_t *sptr;
1103 /* compute input samples */
1104 memcpy(input_samples, s->last_samples[ich], AC3_BLOCK_SIZE * sizeof(int16_t));
1106 sptr = samples + (sinc * AC3_BLOCK_SIZE * i) + ich;
1107 for (j = 0; j < AC3_BLOCK_SIZE; j++) {
1109 input_samples[j + AC3_BLOCK_SIZE] = v;
1110 s->last_samples[ich][j] = v;
1114 /* apply the MDCT window */
1115 for (j = 0; j < AC3_BLOCK_SIZE; j++) {
1116 input_samples[j] = MUL16(input_samples[j],
1117 ff_ac3_window[j]) >> 15;
1118 input_samples[AC3_WINDOW_SIZE-j-1] = MUL16(input_samples[AC3_WINDOW_SIZE-j-1],
1119 ff_ac3_window[j]) >> 15;
1122 /* Normalize the samples to use the maximum available precision */
1123 v = 14 - log2_tab(input_samples, AC3_WINDOW_SIZE);
1126 exp_samples[i][ch] = v - 9;
1127 lshift_tab(input_samples, AC3_WINDOW_SIZE, v);
1130 mdct512(mdct_coef[i][ch], input_samples);
1132 /* compute "exponents". We take into account the normalization there */
1133 for (j = 0; j < AC3_MAX_COEFS; j++) {
1135 v = abs(mdct_coef[i][ch][j]);
1139 e = 23 - av_log2(v) + exp_samples[i][ch];
1142 mdct_coef[i][ch][j] = 0;
1149 compute_exp_strategy(exp_strategy, exp, ch, ch == s->lfe_channel);
1151 /* compute the exponents as the decoder will see them. The
1152 EXP_REUSE case must be handled carefully : we select the
1153 min of the exponents */
1155 while (i < AC3_MAX_BLOCKS) {
1157 while (j < AC3_MAX_BLOCKS && exp_strategy[j][ch] == EXP_REUSE) {
1158 exponent_min(exp[i][ch], exp[j][ch], s->nb_coefs[ch]);
1161 frame_bits += encode_exp(encoded_exp[i][ch],
1162 exp[i][ch], s->nb_coefs[ch],
1163 exp_strategy[i][ch]);
1164 /* copy encoded exponents for reuse case */
1165 for (k = i+1; k < j; k++) {
1166 memcpy(encoded_exp[k][ch], encoded_exp[i][ch],
1167 s->nb_coefs[ch] * sizeof(uint8_t));
1173 /* adjust for fractional frame sizes */
1174 while (s->bits_written >= s->bit_rate && s->samples_written >= s->sample_rate) {
1175 s->bits_written -= s->bit_rate;
1176 s->samples_written -= s->sample_rate;
1178 s->frame_size = s->frame_size_min + (s->bits_written * s->sample_rate < s->samples_written * s->bit_rate);
1179 s->bits_written += s->frame_size * 16;
1180 s->samples_written += AC3_FRAME_SIZE;
1182 compute_bit_allocation(s, bap, encoded_exp, exp_strategy, frame_bits);
1183 /* everything is known... let's output the frame */
1184 output_frame_header(s, frame);
1186 for (i = 0; i < AC3_MAX_BLOCKS; i++) {
1187 output_audio_block(s, exp_strategy[i], encoded_exp[i],
1188 bap[i], mdct_coef[i], exp_samples[i], i);
1190 return output_frame_end(s);
1195 * Finalize encoding and free any memory allocated by the encoder.
1197 static av_cold int ac3_encode_close(AVCodecContext *avctx)
1199 av_freep(&avctx->coded_frame);
1205 * Set channel information during initialization.
1207 static av_cold int set_channel_info(AC3EncodeContext *s, int channels,
1208 int64_t *channel_layout)
1212 if (channels < 1 || channels > AC3_MAX_CHANNELS)
1214 if ((uint64_t)*channel_layout > 0x7FF)
1216 ch_layout = *channel_layout;
1218 ch_layout = avcodec_guess_channel_layout(channels, CODEC_ID_AC3, NULL);
1219 if (av_get_channel_layout_nb_channels(ch_layout) != channels)
1222 s->lfe_on = !!(ch_layout & AV_CH_LOW_FREQUENCY);
1223 s->channels = channels;
1224 s->fbw_channels = channels - s->lfe_on;
1225 s->lfe_channel = s->lfe_on ? s->fbw_channels : -1;
1227 ch_layout -= AV_CH_LOW_FREQUENCY;
1229 switch (ch_layout) {
1230 case AV_CH_LAYOUT_MONO: s->channel_mode = AC3_CHMODE_MONO; break;
1231 case AV_CH_LAYOUT_STEREO: s->channel_mode = AC3_CHMODE_STEREO; break;
1232 case AV_CH_LAYOUT_SURROUND: s->channel_mode = AC3_CHMODE_3F; break;
1233 case AV_CH_LAYOUT_2_1: s->channel_mode = AC3_CHMODE_2F1R; break;
1234 case AV_CH_LAYOUT_4POINT0: s->channel_mode = AC3_CHMODE_3F1R; break;
1235 case AV_CH_LAYOUT_QUAD:
1236 case AV_CH_LAYOUT_2_2: s->channel_mode = AC3_CHMODE_2F2R; break;
1237 case AV_CH_LAYOUT_5POINT0:
1238 case AV_CH_LAYOUT_5POINT0_BACK: s->channel_mode = AC3_CHMODE_3F2R; break;
1243 s->channel_map = ff_ac3_enc_channel_map[s->channel_mode][s->lfe_on];
1244 *channel_layout = ch_layout;
1246 *channel_layout |= AV_CH_LOW_FREQUENCY;
1253 * Initialize the encoder.
1255 static av_cold int ac3_encode_init(AVCodecContext *avctx)
1257 int freq = avctx->sample_rate;
1258 int bitrate = avctx->bit_rate;
1259 AC3EncodeContext *s = avctx->priv_data;
1263 avctx->frame_size = AC3_FRAME_SIZE;
1267 if (!avctx->channel_layout) {
1268 av_log(avctx, AV_LOG_WARNING, "No channel layout specified. The "
1269 "encoder will guess the layout, but it "
1270 "might be incorrect.\n");
1272 if (set_channel_info(s, avctx->channels, &avctx->channel_layout)) {
1273 av_log(avctx, AV_LOG_ERROR, "invalid channel layout\n");
1278 for (i = 0; i < 3; i++) {
1279 for (j = 0; j < 3; j++)
1280 if ((ff_ac3_sample_rate_tab[j] >> i) == freq)
1285 s->sample_rate = freq;
1286 s->bit_alloc.sr_shift = i;
1287 s->bit_alloc.sr_code = j;
1288 s->bitstream_id = 8 + s->bit_alloc.sr_shift;
1289 s->bitstream_mode = 0; /* complete main audio service */
1291 /* bitrate & frame size */
1292 for (i = 0; i < 19; i++) {
1293 if ((ff_ac3_bitrate_tab[i] >> s->bit_alloc.sr_shift)*1000 == bitrate)
1298 s->bit_rate = bitrate;
1299 s->frame_size_code = i << 1;
1300 s->frame_size_min = ff_ac3_frame_size_tab[s->frame_size_code][s->bit_alloc.sr_code];
1301 s->bits_written = 0;
1302 s->samples_written = 0;
1303 s->frame_size = s->frame_size_min;
1307 /* calculate bandwidth based on user-specified cutoff frequency */
1308 int cutoff = av_clip(avctx->cutoff, 1, s->sample_rate >> 1);
1309 int fbw_coeffs = cutoff * 2 * AC3_MAX_COEFS / s->sample_rate;
1310 bw_code = av_clip((fbw_coeffs - 73) / 3, 0, 60);
1312 /* use default bandwidth setting */
1313 /* XXX: should compute the bandwidth according to the frame
1314 size, so that we avoid annoying high frequency artifacts */
1317 for(ch=0;ch<s->fbw_channels;ch++) {
1318 /* bandwidth for each channel */
1319 s->bandwidth_code[ch] = bw_code;
1320 s->nb_coefs[ch] = bw_code * 3 + 73;
1323 s->nb_coefs[s->lfe_channel] = 7; /* LFE channel always has 7 coefs */
1325 /* initial snr offset */
1326 s->coarse_snr_offset = 40;
1330 avctx->coded_frame= avcodec_alloc_frame();
1331 avctx->coded_frame->key_frame= 1;
1338 /*************************************************************************/
1341 #include "libavutil/lfg.h"
1343 #define FN (MDCT_SAMPLES/4)
1346 static void fft_test(AVLFG *lfg)
1348 IComplex in[FN], in1[FN];
1350 float sum_re, sum_im, a;
1352 for (i = 0; i < FN; i++) {
1353 in[i].re = av_lfg_get(lfg) % 65535 - 32767;
1354 in[i].im = av_lfg_get(lfg) % 65535 - 32767;
1360 for (k = 0; k < FN; k++) {
1363 for (n = 0; n < FN; n++) {
1364 a = -2 * M_PI * (n * k) / FN;
1365 sum_re += in1[n].re * cos(a) - in1[n].im * sin(a);
1366 sum_im += in1[n].re * sin(a) + in1[n].im * cos(a);
1368 av_log(NULL, AV_LOG_DEBUG, "%3d: %6d,%6d %6.0f,%6.0f\n",
1369 k, in[k].re, in[k].im, sum_re / FN, sum_im / FN);
1374 static void mdct_test(AVLFG *lfg)
1376 int16_t input[MDCT_SAMPLES];
1377 int32_t output[AC3_MAX_COEFS];
1378 float input1[MDCT_SAMPLES];
1379 float output1[AC3_MAX_COEFS];
1380 float s, a, err, e, emax;
1383 for (i = 0; i < MDCT_SAMPLES; i++) {
1384 input[i] = (av_lfg_get(lfg) % 65535 - 32767) * 9 / 10;
1385 input1[i] = input[i];
1388 mdct512(output, input);
1391 for (k = 0; k < AC3_MAX_COEFS; k++) {
1393 for (n = 0; n < MDCT_SAMPLES; n++) {
1394 a = (2*M_PI*(2*n+1+MDCT_SAMPLES/2)*(2*k+1) / (4 * MDCT_SAMPLES));
1395 s += input1[n] * cos(a);
1397 output1[k] = -2 * s / MDCT_SAMPLES;
1402 for (i = 0; i < AC3_MAX_COEFS; i++) {
1403 av_log(NULL, AV_LOG_DEBUG, "%3d: %7d %7.0f\n", i, output[i], output1[i]);
1404 e = output[i] - output1[i];
1409 av_log(NULL, AV_LOG_DEBUG, "err2=%f emax=%f\n", err / AC3_MAX_COEFS, emax);
1417 av_log_set_level(AV_LOG_DEBUG);
1428 AVCodec ac3_encoder = {
1432 sizeof(AC3EncodeContext),
1437 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,AV_SAMPLE_FMT_NONE},
1438 .long_name = NULL_IF_CONFIG_SMALL("ATSC A/52A (AC-3)"),
1439 .channel_layouts = (const int64_t[]){
1441 AV_CH_LAYOUT_STEREO,
1443 AV_CH_LAYOUT_SURROUND,
1446 AV_CH_LAYOUT_4POINT0,
1447 AV_CH_LAYOUT_5POINT0,
1448 AV_CH_LAYOUT_5POINT0_BACK,
1449 (AV_CH_LAYOUT_MONO | AV_CH_LOW_FREQUENCY),
1450 (AV_CH_LAYOUT_STEREO | AV_CH_LOW_FREQUENCY),
1451 (AV_CH_LAYOUT_2_1 | AV_CH_LOW_FREQUENCY),
1452 (AV_CH_LAYOUT_SURROUND | AV_CH_LOW_FREQUENCY),
1453 (AV_CH_LAYOUT_2_2 | AV_CH_LOW_FREQUENCY),
1454 (AV_CH_LAYOUT_QUAD | AV_CH_LOW_FREQUENCY),
1455 (AV_CH_LAYOUT_4POINT0 | AV_CH_LOW_FREQUENCY),
1456 AV_CH_LAYOUT_5POINT1,
1457 AV_CH_LAYOUT_5POINT1_BACK,