2 * The simplest AC-3 encoder
3 * Copyright (c) 2000 Fabrice Bellard
4 * Copyright (c) 2006-2010 Justin Ruggles <justin.ruggles@gmail.com>
5 * Copyright (c) 2006-2010 Prakash Punnoor <prakash@punnoor.de>
7 * This file is part of FFmpeg.
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * FFmpeg is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 * The simplest AC-3 encoder.
31 #include "libavcore/audioconvert.h"
32 #include "libavutil/crc.h"
37 #include "audioconvert.h"
41 #define MDCT_SAMPLES (1 << MDCT_NBITS)
43 /** Maximum number of exponent groups. +1 for separate DC exponent. */
44 #define AC3_MAX_EXP_GROUPS 85
46 /** Scale a float value by 2^bits and convert to an integer. */
47 #define SCALE_FLOAT(a, bits) lrintf((a) * (float)(1 << (bits)))
49 /** Scale a float value by 2^15, convert to an integer, and clip to int16_t range. */
50 #define FIX15(a) av_clip_int16(SCALE_FLOAT(a, 15))
55 * Used in fixed-point MDCT calculation.
57 typedef struct IComplex {
61 typedef struct AC3MDCTContext {
62 AVCodecContext *avctx; ///< parent context for av_log()
63 int16_t *rot_tmp; ///< temp buffer for pre-rotated samples
64 IComplex *cplx_tmp; ///< temp buffer for complex pre-rotated samples
68 * Data for a single audio block.
70 typedef struct AC3Block {
71 uint8_t **bap; ///< bit allocation pointers (bap)
72 int32_t **mdct_coef; ///< MDCT coefficients
73 uint8_t **exp; ///< original exponents
74 uint8_t **grouped_exp; ///< grouped exponents
75 int16_t **psd; ///< psd per frequency bin
76 int16_t **band_psd; ///< psd per critical band
77 int16_t **mask; ///< masking curve
78 uint16_t **qmant; ///< quantized mantissas
79 uint8_t exp_strategy[AC3_MAX_CHANNELS]; ///< exponent strategies
80 int8_t exp_shift[AC3_MAX_CHANNELS]; ///< exponent shift values
84 * AC-3 encoder private context.
86 typedef struct AC3EncodeContext {
87 PutBitContext pb; ///< bitstream writer context
89 AC3MDCTContext mdct; ///< MDCT context
91 AC3Block blocks[AC3_MAX_BLOCKS]; ///< per-block info
93 int bitstream_id; ///< bitstream id (bsid)
94 int bitstream_mode; ///< bitstream mode (bsmod)
96 int bit_rate; ///< target bit rate, in bits-per-second
97 int sample_rate; ///< sampling frequency, in Hz
99 int frame_size_min; ///< minimum frame size in case rounding is necessary
100 int frame_size; ///< current frame size in bytes
101 int frame_size_code; ///< frame size code (frmsizecod)
102 int bits_written; ///< bit count (used to avg. bitrate)
103 int samples_written; ///< sample count (used to avg. bitrate)
105 int fbw_channels; ///< number of full-bandwidth channels (nfchans)
106 int channels; ///< total number of channels (nchans)
107 int lfe_on; ///< indicates if there is an LFE channel (lfeon)
108 int lfe_channel; ///< channel index of the LFE channel
109 int channel_mode; ///< channel mode (acmod)
110 const uint8_t *channel_map; ///< channel map used to reorder channels
112 int bandwidth_code[AC3_MAX_CHANNELS]; ///< bandwidth code (0 to 60) (chbwcod)
113 int nb_coefs[AC3_MAX_CHANNELS];
115 /* bitrate allocation control */
116 int slow_gain_code; ///< slow gain code (sgaincod)
117 int slow_decay_code; ///< slow decay code (sdcycod)
118 int fast_decay_code; ///< fast decay code (fdcycod)
119 int db_per_bit_code; ///< dB/bit code (dbpbcod)
120 int floor_code; ///< floor code (floorcod)
121 AC3BitAllocParameters bit_alloc; ///< bit allocation parameters
122 int coarse_snr_offset; ///< coarse SNR offsets (csnroffst)
123 int fast_gain_code[AC3_MAX_CHANNELS]; ///< fast gain codes (signal-to-mask ratio) (fgaincod)
124 int fine_snr_offset[AC3_MAX_CHANNELS]; ///< fine SNR offsets (fsnroffst)
125 int frame_bits_fixed; ///< number of non-coefficient bits for fixed parameters
126 int frame_bits; ///< all frame bits except exponents and mantissas
127 int exponent_bits; ///< number of bits used for exponents
129 /* mantissa encoding */
130 int mant1_cnt, mant2_cnt, mant4_cnt; ///< mantissa counts for bap=1,2,4
131 uint16_t *qmant1_ptr, *qmant2_ptr, *qmant4_ptr; ///< mantissa pointers for bap=1,2,4
133 int16_t **planar_samples;
135 uint8_t *bap1_buffer;
136 int32_t *mdct_coef_buffer;
138 uint8_t *grouped_exp_buffer;
140 int16_t *band_psd_buffer;
141 int16_t *mask_buffer;
142 uint16_t *qmant_buffer;
144 DECLARE_ALIGNED(16, int16_t, windowed_samples)[AC3_WINDOW_SIZE];
148 /** MDCT and FFT tables */
149 static int16_t costab[64];
150 static int16_t sintab[64];
151 static int16_t xcos1[128];
152 static int16_t xsin1[128];
155 * LUT for number of exponent groups.
156 * exponent_group_tab[exponent strategy-1][number of coefficients]
158 uint8_t exponent_group_tab[3][256];
162 * Adjust the frame size to make the average bit rate match the target bit rate.
163 * This is only needed for 11025, 22050, and 44100 sample rates.
165 static void adjust_frame_size(AC3EncodeContext *s)
167 while (s->bits_written >= s->bit_rate && s->samples_written >= s->sample_rate) {
168 s->bits_written -= s->bit_rate;
169 s->samples_written -= s->sample_rate;
171 s->frame_size = s->frame_size_min +
172 2 * (s->bits_written * s->sample_rate < s->samples_written * s->bit_rate);
173 s->bits_written += s->frame_size * 8;
174 s->samples_written += AC3_FRAME_SIZE;
179 * Deinterleave input samples.
180 * Channels are reordered from FFmpeg's default order to AC-3 order.
182 static void deinterleave_input_samples(AC3EncodeContext *s,
183 const int16_t *samples)
187 /* deinterleave and remap input samples */
188 for (ch = 0; ch < s->channels; ch++) {
192 /* copy last 256 samples of previous frame to the start of the current frame */
193 memcpy(&s->planar_samples[ch][0], &s->planar_samples[ch][AC3_FRAME_SIZE],
194 AC3_BLOCK_SIZE * sizeof(s->planar_samples[0][0]));
198 sptr = samples + s->channel_map[ch];
199 for (i = AC3_BLOCK_SIZE; i < AC3_FRAME_SIZE+AC3_BLOCK_SIZE; i++) {
200 s->planar_samples[ch][i] = *sptr;
208 * Finalize MDCT and free allocated memory.
210 static av_cold void mdct_end(AC3MDCTContext *mdct)
212 av_freep(&mdct->rot_tmp);
213 av_freep(&mdct->cplx_tmp);
219 * Initialize FFT tables.
220 * @param ln log2(FFT size)
222 static av_cold void fft_init(int ln)
230 for (i = 0; i < n2; i++) {
231 alpha = 2.0 * M_PI * i / n;
232 costab[i] = FIX15(cos(alpha));
233 sintab[i] = FIX15(sin(alpha));
239 * Initialize MDCT tables.
240 * @param nbits log2(MDCT size)
242 static av_cold int mdct_init(AC3MDCTContext *mdct, int nbits)
251 FF_ALLOC_OR_GOTO(mdct->avctx, mdct->rot_tmp, n * sizeof(*mdct->rot_tmp),
253 FF_ALLOC_OR_GOTO(mdct->avctx, mdct->cplx_tmp, n4 * sizeof(*mdct->cplx_tmp),
256 for (i = 0; i < n4; i++) {
257 float alpha = 2.0 * M_PI * (i + 1.0 / 8.0) / n;
258 xcos1[i] = FIX15(-cos(alpha));
259 xsin1[i] = FIX15(-sin(alpha));
264 return AVERROR(ENOMEM);
269 #define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \
271 int ax, ay, bx, by; \
276 pre = (bx + ax) >> 1; \
277 pim = (by + ay) >> 1; \
278 qre = (bx - ax) >> 1; \
279 qim = (by - ay) >> 1; \
283 /** Complex multiply */
284 #define CMUL(pre, pim, are, aim, bre, bim) \
286 pre = (MUL16(are, bre) - MUL16(aim, bim)) >> 15; \
287 pim = (MUL16(are, bim) + MUL16(bre, aim)) >> 15; \
292 * Calculate a 2^n point complex FFT on 2^ln points.
293 * @param z complex input/output samples
294 * @param ln log2(FFT size)
296 static void fft(IComplex *z, int ln)
300 register IComplex *p,*q;
306 for (j = 0; j < np; j++) {
307 int k = av_reverse[j] >> (8 - ln);
309 FFSWAP(IComplex, z[k], z[j]);
317 BF(p[0].re, p[0].im, p[1].re, p[1].im,
318 p[0].re, p[0].im, p[1].re, p[1].im);
327 BF(p[0].re, p[0].im, p[2].re, p[2].im,
328 p[0].re, p[0].im, p[2].re, p[2].im);
329 BF(p[1].re, p[1].im, p[3].re, p[3].im,
330 p[1].re, p[1].im, p[3].im, -p[3].re);
342 for (j = 0; j < nblocks; j++) {
343 BF(p->re, p->im, q->re, q->im,
344 p->re, p->im, q->re, q->im);
347 for(l = nblocks; l < np2; l += nblocks) {
348 CMUL(tmp_re, tmp_im, costab[l], -sintab[l], q->re, q->im);
349 BF(p->re, p->im, q->re, q->im,
350 p->re, p->im, tmp_re, tmp_im);
357 nblocks = nblocks >> 1;
358 nloops = nloops << 1;
364 * Calculate a 512-point MDCT
365 * @param out 256 output frequency coefficients
366 * @param in 512 windowed input audio samples
368 static void mdct512(AC3MDCTContext *mdct, int32_t *out, int16_t *in)
371 int16_t *rot = mdct->rot_tmp;
372 IComplex *x = mdct->cplx_tmp;
374 /* shift to simplify computations */
375 for (i = 0; i < MDCT_SAMPLES/4; i++)
376 rot[i] = -in[i + 3*MDCT_SAMPLES/4];
377 memcpy(&rot[MDCT_SAMPLES/4], &in[0], 3*MDCT_SAMPLES/4*sizeof(*in));
380 for (i = 0; i < MDCT_SAMPLES/4; i++) {
381 re = ((int)rot[ 2*i] - (int)rot[MDCT_SAMPLES -1-2*i]) >> 1;
382 im = -((int)rot[MDCT_SAMPLES/2+2*i] - (int)rot[MDCT_SAMPLES/2-1-2*i]) >> 1;
383 CMUL(x[i].re, x[i].im, re, im, -xcos1[i], xsin1[i]);
386 fft(x, MDCT_NBITS - 2);
389 for (i = 0; i < MDCT_SAMPLES/4; i++) {
392 CMUL(out[MDCT_SAMPLES/2-1-2*i], out[2*i], re, im, xsin1[i], xcos1[i]);
398 * Apply KBD window to input samples prior to MDCT.
400 static void apply_window(int16_t *output, const int16_t *input,
401 const int16_t *window, int n)
406 for (i = 0; i < n2; i++) {
407 output[i] = MUL16(input[i], window[i]) >> 15;
408 output[n-i-1] = MUL16(input[n-i-1], window[i]) >> 15;
414 * Calculate the log2() of the maximum absolute value in an array.
415 * @param tab input array
416 * @param n number of values in the array
417 * @return log2(max(abs(tab[])))
419 static int log2_tab(int16_t *tab, int n)
424 for (i = 0; i < n; i++)
432 * Left-shift each value in an array by a specified amount.
433 * @param tab input array
434 * @param n number of values in the array
435 * @param lshift left shift amount. a negative value means right shift.
437 static void lshift_tab(int16_t *tab, int n, int lshift)
442 for (i = 0; i < n; i++)
444 } else if (lshift < 0) {
446 for (i = 0; i < n; i++)
453 * Normalize the input samples to use the maximum available precision.
454 * This assumes signed 16-bit input samples. Exponents are reduced by 9 to
455 * match the 24-bit internal precision for MDCT coefficients.
457 * @return exponent shift
459 static int normalize_samples(AC3EncodeContext *s)
461 int v = 14 - log2_tab(s->windowed_samples, AC3_WINDOW_SIZE);
463 lshift_tab(s->windowed_samples, AC3_WINDOW_SIZE, v);
469 * Apply the MDCT to input samples to generate frequency coefficients.
470 * This applies the KBD window and normalizes the input to reduce precision
471 * loss due to fixed-point calculations.
473 static void apply_mdct(AC3EncodeContext *s)
477 for (ch = 0; ch < s->channels; ch++) {
478 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
479 AC3Block *block = &s->blocks[blk];
480 const int16_t *input_samples = &s->planar_samples[ch][blk * AC3_BLOCK_SIZE];
482 apply_window(s->windowed_samples, input_samples, ff_ac3_window, AC3_WINDOW_SIZE);
484 block->exp_shift[ch] = normalize_samples(s);
486 mdct512(&s->mdct, block->mdct_coef[ch], s->windowed_samples);
493 * Initialize exponent tables.
495 static av_cold void exponent_init(AC3EncodeContext *s)
498 for (i = 73; i < 256; i++) {
499 exponent_group_tab[0][i] = (i - 1) / 3;
500 exponent_group_tab[1][i] = (i + 2) / 6;
501 exponent_group_tab[2][i] = (i + 8) / 12;
507 * Extract exponents from the MDCT coefficients.
508 * This takes into account the normalization that was done to the input samples
509 * by adjusting the exponents by the exponent shift values.
511 static void extract_exponents(AC3EncodeContext *s)
515 for (ch = 0; ch < s->channels; ch++) {
516 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
517 AC3Block *block = &s->blocks[blk];
518 for (i = 0; i < AC3_MAX_COEFS; i++) {
520 int v = abs(block->mdct_coef[ch][i]);
524 e = 23 - av_log2(v) + block->exp_shift[ch];
527 block->mdct_coef[ch][i] = 0;
530 block->exp[ch][i] = e;
538 * Exponent Difference Threshold.
539 * New exponents are sent if their SAD exceed this number.
541 #define EXP_DIFF_THRESHOLD 1000
545 * Calculate exponent strategies for all blocks in a single channel.
547 static void compute_exp_strategy_ch(AC3EncodeContext *s, uint8_t *exp_strategy, uint8_t **exp)
552 /* estimate if the exponent variation & decide if they should be
553 reused in the next frame */
554 exp_strategy[0] = EXP_NEW;
555 for (blk = 1; blk < AC3_MAX_BLOCKS; blk++) {
556 exp_diff = s->dsp.sad[0](NULL, exp[blk], exp[blk-1], 16, 16);
557 if (exp_diff > EXP_DIFF_THRESHOLD)
558 exp_strategy[blk] = EXP_NEW;
560 exp_strategy[blk] = EXP_REUSE;
563 /* now select the encoding strategy type : if exponents are often
564 recoded, we use a coarse encoding */
566 while (blk < AC3_MAX_BLOCKS) {
568 while (blk1 < AC3_MAX_BLOCKS && exp_strategy[blk1] == EXP_REUSE)
570 switch (blk1 - blk) {
571 case 1: exp_strategy[blk] = EXP_D45; break;
573 case 3: exp_strategy[blk] = EXP_D25; break;
574 default: exp_strategy[blk] = EXP_D15; break;
582 * Calculate exponent strategies for all channels.
583 * Array arrangement is reversed to simplify the per-channel calculation.
585 static void compute_exp_strategy(AC3EncodeContext *s)
587 uint8_t *exp1[AC3_MAX_CHANNELS][AC3_MAX_BLOCKS];
588 uint8_t exp_str1[AC3_MAX_CHANNELS][AC3_MAX_BLOCKS];
591 for (ch = 0; ch < s->fbw_channels; ch++) {
592 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
593 exp1[ch][blk] = s->blocks[blk].exp[ch];
594 exp_str1[ch][blk] = s->blocks[blk].exp_strategy[ch];
597 compute_exp_strategy_ch(s, exp_str1[ch], exp1[ch]);
599 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++)
600 s->blocks[blk].exp_strategy[ch] = exp_str1[ch][blk];
604 s->blocks[0].exp_strategy[ch] = EXP_D15;
605 for (blk = 1; blk < AC3_MAX_BLOCKS; blk++)
606 s->blocks[blk].exp_strategy[ch] = EXP_REUSE;
612 * Set each encoded exponent in a block to the minimum of itself and the
613 * exponent in the same frequency bin of a following block.
614 * exp[i] = min(exp[i], exp1[i]
616 static void exponent_min(uint8_t *exp, uint8_t *exp1, int n)
619 for (i = 0; i < n; i++) {
620 if (exp1[i] < exp[i])
627 * Update the exponents so that they are the ones the decoder will decode.
629 static void encode_exponents_blk_ch(uint8_t *exp,
630 int nb_exps, int exp_strategy)
634 nb_groups = exponent_group_tab[exp_strategy-1][nb_exps] * 3;
636 /* for each group, compute the minimum exponent */
637 switch(exp_strategy) {
639 for (i = 1, k = 1; i <= nb_groups; i++) {
640 uint8_t exp_min = exp[k];
641 if (exp[k+1] < exp_min)
648 for (i = 1, k = 1; i <= nb_groups; i++) {
649 uint8_t exp_min = exp[k];
650 if (exp[k+1] < exp_min)
652 if (exp[k+2] < exp_min)
654 if (exp[k+3] < exp_min)
662 /* constraint for DC exponent */
666 /* decrease the delta between each groups to within 2 so that they can be
667 differentially encoded */
668 for (i = 1; i <= nb_groups; i++)
669 exp[i] = FFMIN(exp[i], exp[i-1] + 2);
672 exp[i] = FFMIN(exp[i], exp[i+1] + 2);
674 /* now we have the exponent values the decoder will see */
675 switch (exp_strategy) {
677 for (i = nb_groups, k = nb_groups * 2; i > 0; i--) {
678 uint8_t exp1 = exp[i];
684 for (i = nb_groups, k = nb_groups * 4; i > 0; i--) {
685 exp[k] = exp[k-1] = exp[k-2] = exp[k-3] = exp[i];
694 * Encode exponents from original extracted form to what the decoder will see.
695 * This copies and groups exponents based on exponent strategy and reduces
696 * deltas between adjacent exponent groups so that they can be differentially
699 static void encode_exponents(AC3EncodeContext *s)
701 int blk, blk1, blk2, ch;
702 AC3Block *block, *block1, *block2;
704 for (ch = 0; ch < s->channels; ch++) {
706 block = &s->blocks[0];
707 while (blk < AC3_MAX_BLOCKS) {
710 /* for the EXP_REUSE case we select the min of the exponents */
711 while (blk1 < AC3_MAX_BLOCKS && block1->exp_strategy[ch] == EXP_REUSE) {
712 exponent_min(block->exp[ch], block1->exp[ch], s->nb_coefs[ch]);
716 encode_exponents_blk_ch(block->exp[ch], s->nb_coefs[ch],
717 block->exp_strategy[ch]);
718 /* copy encoded exponents for reuse case */
720 for (blk2 = blk+1; blk2 < blk1; blk2++, block2++) {
721 memcpy(block2->exp[ch], block->exp[ch],
722 s->nb_coefs[ch] * sizeof(uint8_t));
733 * 3 delta-encoded exponents are in each 7-bit group. The number of groups
734 * varies depending on exponent strategy and bandwidth.
736 static void group_exponents(AC3EncodeContext *s)
739 int group_size, nb_groups, bit_count;
741 int delta0, delta1, delta2;
745 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
746 AC3Block *block = &s->blocks[blk];
747 for (ch = 0; ch < s->channels; ch++) {
748 if (block->exp_strategy[ch] == EXP_REUSE) {
751 group_size = block->exp_strategy[ch] + (block->exp_strategy[ch] == EXP_D45);
752 nb_groups = exponent_group_tab[block->exp_strategy[ch]-1][s->nb_coefs[ch]];
753 bit_count += 4 + (nb_groups * 7);
758 block->grouped_exp[ch][0] = exp1;
760 /* remaining exponents are delta encoded */
761 for (i = 1; i <= nb_groups; i++) {
762 /* merge three delta in one code */
766 delta0 = exp1 - exp0 + 2;
771 delta1 = exp1 - exp0 + 2;
776 delta2 = exp1 - exp0 + 2;
778 block->grouped_exp[ch][i] = ((delta0 * 5 + delta1) * 5) + delta2;
783 s->exponent_bits = bit_count;
788 * Calculate final exponents from the supplied MDCT coefficients and exponent shift.
789 * Extract exponents from MDCT coefficients, calculate exponent strategies,
790 * and encode final exponents.
792 static void process_exponents(AC3EncodeContext *s)
794 extract_exponents(s);
796 compute_exp_strategy(s);
805 * Count frame bits that are based solely on fixed parameters.
806 * This only has to be run once when the encoder is initialized.
808 static void count_frame_bits_fixed(AC3EncodeContext *s)
810 static const int frame_bits_inc[8] = { 0, 0, 2, 2, 2, 4, 2, 4 };
815 * no dynamic range codes
816 * no channel coupling
818 * bit allocation parameters do not change between blocks
819 * SNR offsets do not change between blocks
820 * no delta bit allocation
827 frame_bits += frame_bits_inc[s->channel_mode];
830 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
831 frame_bits += s->fbw_channels * 2 + 2; /* blksw * c, dithflag * c, dynrnge, cplstre */
832 if (s->channel_mode == AC3_CHMODE_STEREO) {
833 frame_bits++; /* rematstr */
837 frame_bits += 2 * s->fbw_channels; /* chexpstr[2] * c */
839 frame_bits++; /* lfeexpstr */
840 frame_bits++; /* baie */
841 frame_bits++; /* snr */
842 frame_bits += 2; /* delta / skip */
844 frame_bits++; /* cplinu for block 0 */
846 /* sdcycod[2], fdcycod[2], sgaincod[2], dbpbcod[2], floorcod[3] */
848 /* (fsnoffset[4] + fgaincod[4]) * c */
849 frame_bits += 2*4 + 3 + 6 + s->channels * (4 + 3);
851 /* auxdatae, crcrsv */
857 s->frame_bits_fixed = frame_bits;
862 * Initialize bit allocation.
863 * Set default parameter codes and calculate parameter values.
865 static void bit_alloc_init(AC3EncodeContext *s)
869 /* init default parameters */
870 s->slow_decay_code = 2;
871 s->fast_decay_code = 1;
872 s->slow_gain_code = 1;
873 s->db_per_bit_code = 2;
875 for (ch = 0; ch < s->channels; ch++)
876 s->fast_gain_code[ch] = 4;
878 /* initial snr offset */
879 s->coarse_snr_offset = 40;
881 /* compute real values */
882 /* currently none of these values change during encoding, so we can just
883 set them once at initialization */
884 s->bit_alloc.slow_decay = ff_ac3_slow_decay_tab[s->slow_decay_code] >> s->bit_alloc.sr_shift;
885 s->bit_alloc.fast_decay = ff_ac3_fast_decay_tab[s->fast_decay_code] >> s->bit_alloc.sr_shift;
886 s->bit_alloc.slow_gain = ff_ac3_slow_gain_tab[s->slow_gain_code];
887 s->bit_alloc.db_per_bit = ff_ac3_db_per_bit_tab[s->db_per_bit_code];
888 s->bit_alloc.floor = ff_ac3_floor_tab[s->floor_code];
890 count_frame_bits_fixed(s);
895 * Count the bits used to encode the frame, minus exponents and mantissas.
896 * Bits based on fixed parameters have already been counted, so now we just
897 * have to add the bits based on parameters that change during encoding.
899 static void count_frame_bits(AC3EncodeContext *s)
904 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
905 uint8_t *exp_strategy = s->blocks[blk].exp_strategy;
906 for (ch = 0; ch < s->fbw_channels; ch++) {
907 if (exp_strategy[ch] != EXP_REUSE)
908 frame_bits += 6 + 2; /* chbwcod[6], gainrng[2] */
911 s->frame_bits = s->frame_bits_fixed + frame_bits;
916 * Calculate the number of bits needed to encode a set of mantissas.
918 static int compute_mantissa_size(int mant_cnt[5], uint8_t *bap, int nb_coefs)
923 for (i = 0; i < nb_coefs; i++) {
926 // bap=1 to bap=4 will be counted in compute_mantissa_size_final
928 } else if (b <= 13) {
929 // bap=5 to bap=13 use (bap-1) bits
932 // bap=14 uses 14 bits and bap=15 uses 16 bits
933 bits += (b == 14) ? 14 : 16;
941 * Finalize the mantissa bit count by adding in the grouped mantissas.
943 static int compute_mantissa_size_final(int mant_cnt[5])
945 // bap=1 : 3 mantissas in 5 bits
946 int bits = (mant_cnt[1] / 3) * 5;
947 // bap=2 : 3 mantissas in 7 bits
948 // bap=4 : 2 mantissas in 7 bits
949 bits += ((mant_cnt[2] / 3) + (mant_cnt[4] >> 1)) * 7;
950 // bap=3 : each mantissa is 3 bits
951 bits += mant_cnt[3] * 3;
957 * Calculate masking curve based on the final exponents.
958 * Also calculate the power spectral densities to use in future calculations.
960 static void bit_alloc_masking(AC3EncodeContext *s)
964 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
965 AC3Block *block = &s->blocks[blk];
966 for (ch = 0; ch < s->channels; ch++) {
967 if (block->exp_strategy[ch] == EXP_REUSE) {
968 AC3Block *block1 = &s->blocks[blk-1];
969 memcpy(block->psd[ch], block1->psd[ch], AC3_MAX_COEFS*sizeof(block->psd[0][0]));
970 memcpy(block->mask[ch], block1->mask[ch], AC3_CRITICAL_BANDS*sizeof(block->mask[0][0]));
972 ff_ac3_bit_alloc_calc_psd(block->exp[ch], 0,
974 block->psd[ch], block->band_psd[ch]);
975 ff_ac3_bit_alloc_calc_mask(&s->bit_alloc, block->band_psd[ch],
977 ff_ac3_fast_gain_tab[s->fast_gain_code[ch]],
978 ch == s->lfe_channel,
979 DBA_NONE, 0, NULL, NULL, NULL,
988 * Ensure that bap for each block and channel point to the current bap_buffer.
989 * They may have been switched during the bit allocation search.
991 static void reset_block_bap(AC3EncodeContext *s)
994 if (s->blocks[0].bap[0] == s->bap_buffer)
996 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
997 for (ch = 0; ch < s->channels; ch++) {
998 s->blocks[blk].bap[ch] = &s->bap_buffer[AC3_MAX_COEFS * (blk * s->channels + ch)];
1005 * Run the bit allocation with a given SNR offset.
1006 * This calculates the bit allocation pointers that will be used to determine
1007 * the quantization of each mantissa.
1008 * @return the number of bits needed for mantissas if the given SNR offset is
1011 static int bit_alloc(AC3EncodeContext *s,
1018 snr_offset = (snr_offset - 240) << 2;
1022 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
1023 AC3Block *block = &s->blocks[blk];
1024 // initialize grouped mantissa counts. these are set so that they are
1025 // padded to the next whole group size when bits are counted in
1026 // compute_mantissa_size_final
1027 mant_cnt[0] = mant_cnt[3] = 0;
1028 mant_cnt[1] = mant_cnt[2] = 2;
1030 for (ch = 0; ch < s->channels; ch++) {
1031 /* Currently the only bit allocation parameters which vary across
1032 blocks within a frame are the exponent values. We can take
1033 advantage of that by reusing the bit allocation pointers
1034 whenever we reuse exponents. */
1035 if (block->exp_strategy[ch] == EXP_REUSE) {
1036 memcpy(block->bap[ch], s->blocks[blk-1].bap[ch], AC3_MAX_COEFS);
1038 ff_ac3_bit_alloc_calc_bap(block->mask[ch], block->psd[ch], 0,
1039 s->nb_coefs[ch], snr_offset,
1040 s->bit_alloc.floor, ff_ac3_bap_tab,
1043 mantissa_bits += compute_mantissa_size(mant_cnt, block->bap[ch], s->nb_coefs[ch]);
1045 mantissa_bits += compute_mantissa_size_final(mant_cnt);
1047 return mantissa_bits;
1052 * Constant bitrate bit allocation search.
1053 * Find the largest SNR offset that will allow data to fit in the frame.
1055 static int cbr_bit_allocation(AC3EncodeContext *s)
1061 bits_left = 8 * s->frame_size - (s->frame_bits + s->exponent_bits);
1063 snr_offset = s->coarse_snr_offset << 4;
1065 while (snr_offset >= 0 &&
1066 bit_alloc(s, snr_offset) > bits_left) {
1070 return AVERROR(EINVAL);
1072 FFSWAP(uint8_t *, s->bap_buffer, s->bap1_buffer);
1073 while (snr_offset + 64 <= 1023 &&
1074 bit_alloc(s, snr_offset + 64) <= bits_left) {
1076 FFSWAP(uint8_t *, s->bap_buffer, s->bap1_buffer);
1078 while (snr_offset + 16 <= 1023 &&
1079 bit_alloc(s, snr_offset + 16) <= bits_left) {
1081 FFSWAP(uint8_t *, s->bap_buffer, s->bap1_buffer);
1083 while (snr_offset + 4 <= 1023 &&
1084 bit_alloc(s, snr_offset + 4) <= bits_left) {
1086 FFSWAP(uint8_t *, s->bap_buffer, s->bap1_buffer);
1088 while (snr_offset + 1 <= 1023 &&
1089 bit_alloc(s, snr_offset + 1) <= bits_left) {
1091 FFSWAP(uint8_t *, s->bap_buffer, s->bap1_buffer);
1093 FFSWAP(uint8_t *, s->bap_buffer, s->bap1_buffer);
1096 s->coarse_snr_offset = snr_offset >> 4;
1097 for (ch = 0; ch < s->channels; ch++)
1098 s->fine_snr_offset[ch] = snr_offset & 0xF;
1105 * Perform bit allocation search.
1106 * Finds the SNR offset value that maximizes quality and fits in the specified
1107 * frame size. Output is the SNR offset and a set of bit allocation pointers
1108 * used to quantize the mantissas.
1110 static int compute_bit_allocation(AC3EncodeContext *s)
1112 count_frame_bits(s);
1114 bit_alloc_masking(s);
1116 return cbr_bit_allocation(s);
1121 * Symmetric quantization on 'levels' levels.
1123 static inline int sym_quant(int c, int e, int levels)
1128 v = (levels * (c << e)) >> 24;
1130 v = (levels >> 1) + v;
1132 v = (levels * ((-c) << e)) >> 24;
1134 v = (levels >> 1) - v;
1136 assert(v >= 0 && v < levels);
1142 * Asymmetric quantization on 2^qbits levels.
1144 static inline int asym_quant(int c, int e, int qbits)
1148 lshift = e + qbits - 24;
1155 m = (1 << (qbits-1));
1159 return v & ((1 << qbits)-1);
1164 * Quantize a set of mantissas for a single channel in a single block.
1166 static void quantize_mantissas_blk_ch(AC3EncodeContext *s,
1167 int32_t *mdct_coef, int8_t exp_shift,
1168 uint8_t *exp, uint8_t *bap,
1169 uint16_t *qmant, int n)
1173 for (i = 0; i < n; i++) {
1175 int c = mdct_coef[i];
1176 int e = exp[i] - exp_shift;
1183 v = sym_quant(c, e, 3);
1184 switch (s->mant1_cnt) {
1186 s->qmant1_ptr = &qmant[i];
1191 *s->qmant1_ptr += 3 * v;
1196 *s->qmant1_ptr += v;
1203 v = sym_quant(c, e, 5);
1204 switch (s->mant2_cnt) {
1206 s->qmant2_ptr = &qmant[i];
1211 *s->qmant2_ptr += 5 * v;
1216 *s->qmant2_ptr += v;
1223 v = sym_quant(c, e, 7);
1226 v = sym_quant(c, e, 11);
1227 switch (s->mant4_cnt) {
1229 s->qmant4_ptr = &qmant[i];
1234 *s->qmant4_ptr += v;
1241 v = sym_quant(c, e, 15);
1244 v = asym_quant(c, e, 14);
1247 v = asym_quant(c, e, 16);
1250 v = asym_quant(c, e, b - 1);
1259 * Quantize mantissas using coefficients, exponents, and bit allocation pointers.
1261 static void quantize_mantissas(AC3EncodeContext *s)
1266 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
1267 AC3Block *block = &s->blocks[blk];
1268 s->mant1_cnt = s->mant2_cnt = s->mant4_cnt = 0;
1269 s->qmant1_ptr = s->qmant2_ptr = s->qmant4_ptr = NULL;
1271 for (ch = 0; ch < s->channels; ch++) {
1272 quantize_mantissas_blk_ch(s, block->mdct_coef[ch], block->exp_shift[ch],
1273 block->exp[ch], block->bap[ch],
1274 block->qmant[ch], s->nb_coefs[ch]);
1281 * Write the AC-3 frame header to the output bitstream.
1283 static void output_frame_header(AC3EncodeContext *s)
1285 put_bits(&s->pb, 16, 0x0b77); /* frame header */
1286 put_bits(&s->pb, 16, 0); /* crc1: will be filled later */
1287 put_bits(&s->pb, 2, s->bit_alloc.sr_code);
1288 put_bits(&s->pb, 6, s->frame_size_code + (s->frame_size - s->frame_size_min) / 2);
1289 put_bits(&s->pb, 5, s->bitstream_id);
1290 put_bits(&s->pb, 3, s->bitstream_mode);
1291 put_bits(&s->pb, 3, s->channel_mode);
1292 if ((s->channel_mode & 0x01) && s->channel_mode != AC3_CHMODE_MONO)
1293 put_bits(&s->pb, 2, 1); /* XXX -4.5 dB */
1294 if (s->channel_mode & 0x04)
1295 put_bits(&s->pb, 2, 1); /* XXX -6 dB */
1296 if (s->channel_mode == AC3_CHMODE_STEREO)
1297 put_bits(&s->pb, 2, 0); /* surround not indicated */
1298 put_bits(&s->pb, 1, s->lfe_on); /* LFE */
1299 put_bits(&s->pb, 5, 31); /* dialog norm: -31 db */
1300 put_bits(&s->pb, 1, 0); /* no compression control word */
1301 put_bits(&s->pb, 1, 0); /* no lang code */
1302 put_bits(&s->pb, 1, 0); /* no audio production info */
1303 put_bits(&s->pb, 1, 0); /* no copyright */
1304 put_bits(&s->pb, 1, 1); /* original bitstream */
1305 put_bits(&s->pb, 1, 0); /* no time code 1 */
1306 put_bits(&s->pb, 1, 0); /* no time code 2 */
1307 put_bits(&s->pb, 1, 0); /* no additional bit stream info */
1312 * Write one audio block to the output bitstream.
1314 static void output_audio_block(AC3EncodeContext *s,
1317 int ch, i, baie, rbnd;
1318 AC3Block *block = &s->blocks[block_num];
1320 /* block switching */
1321 for (ch = 0; ch < s->fbw_channels; ch++)
1322 put_bits(&s->pb, 1, 0);
1325 for (ch = 0; ch < s->fbw_channels; ch++)
1326 put_bits(&s->pb, 1, 1);
1328 /* dynamic range codes */
1329 put_bits(&s->pb, 1, 0);
1331 /* channel coupling */
1333 put_bits(&s->pb, 1, 1); /* coupling strategy present */
1334 put_bits(&s->pb, 1, 0); /* no coupling strategy */
1336 put_bits(&s->pb, 1, 0); /* no new coupling strategy */
1339 /* stereo rematrixing */
1340 if (s->channel_mode == AC3_CHMODE_STEREO) {
1342 /* first block must define rematrixing (rematstr) */
1343 put_bits(&s->pb, 1, 1);
1345 /* dummy rematrixing rematflg(1:4)=0 */
1346 for (rbnd = 0; rbnd < 4; rbnd++)
1347 put_bits(&s->pb, 1, 0);
1349 /* no matrixing (but should be used in the future) */
1350 put_bits(&s->pb, 1, 0);
1354 /* exponent strategy */
1355 for (ch = 0; ch < s->fbw_channels; ch++)
1356 put_bits(&s->pb, 2, block->exp_strategy[ch]);
1358 put_bits(&s->pb, 1, block->exp_strategy[s->lfe_channel]);
1361 for (ch = 0; ch < s->fbw_channels; ch++) {
1362 if (block->exp_strategy[ch] != EXP_REUSE)
1363 put_bits(&s->pb, 6, s->bandwidth_code[ch]);
1367 for (ch = 0; ch < s->channels; ch++) {
1370 if (block->exp_strategy[ch] == EXP_REUSE)
1374 put_bits(&s->pb, 4, block->grouped_exp[ch][0]);
1376 /* exponent groups */
1377 nb_groups = exponent_group_tab[block->exp_strategy[ch]-1][s->nb_coefs[ch]];
1378 for (i = 1; i <= nb_groups; i++)
1379 put_bits(&s->pb, 7, block->grouped_exp[ch][i]);
1381 /* gain range info */
1382 if (ch != s->lfe_channel)
1383 put_bits(&s->pb, 2, 0);
1386 /* bit allocation info */
1387 baie = (block_num == 0);
1388 put_bits(&s->pb, 1, baie);
1390 put_bits(&s->pb, 2, s->slow_decay_code);
1391 put_bits(&s->pb, 2, s->fast_decay_code);
1392 put_bits(&s->pb, 2, s->slow_gain_code);
1393 put_bits(&s->pb, 2, s->db_per_bit_code);
1394 put_bits(&s->pb, 3, s->floor_code);
1398 put_bits(&s->pb, 1, baie);
1400 put_bits(&s->pb, 6, s->coarse_snr_offset);
1401 for (ch = 0; ch < s->channels; ch++) {
1402 put_bits(&s->pb, 4, s->fine_snr_offset[ch]);
1403 put_bits(&s->pb, 3, s->fast_gain_code[ch]);
1407 put_bits(&s->pb, 1, 0); /* no delta bit allocation */
1408 put_bits(&s->pb, 1, 0); /* no data to skip */
1411 for (ch = 0; ch < s->channels; ch++) {
1413 for (i = 0; i < s->nb_coefs[ch]; i++) {
1414 q = block->qmant[ch][i];
1415 b = block->bap[ch][i];
1418 case 1: if (q != 128) put_bits(&s->pb, 5, q); break;
1419 case 2: if (q != 128) put_bits(&s->pb, 7, q); break;
1420 case 3: put_bits(&s->pb, 3, q); break;
1421 case 4: if (q != 128) put_bits(&s->pb, 7, q); break;
1422 case 14: put_bits(&s->pb, 14, q); break;
1423 case 15: put_bits(&s->pb, 16, q); break;
1424 default: put_bits(&s->pb, b-1, q); break;
1431 /** CRC-16 Polynomial */
1432 #define CRC16_POLY ((1 << 0) | (1 << 2) | (1 << 15) | (1 << 16))
1435 static unsigned int mul_poly(unsigned int a, unsigned int b, unsigned int poly)
1452 static unsigned int pow_poly(unsigned int a, unsigned int n, unsigned int poly)
1458 r = mul_poly(r, a, poly);
1459 a = mul_poly(a, a, poly);
1467 * Fill the end of the frame with 0's and compute the two CRCs.
1469 static void output_frame_end(AC3EncodeContext *s)
1471 int frame_size, frame_size_58, pad_bytes, crc1, crc2, crc_inv;
1474 frame_size = s->frame_size;
1475 frame_size_58 = ((frame_size >> 2) + (frame_size >> 4)) << 1;
1477 /* pad the remainder of the frame with zeros */
1478 flush_put_bits(&s->pb);
1480 pad_bytes = s->frame_size - (put_bits_ptr(&s->pb) - frame) - 2;
1481 assert(pad_bytes >= 0);
1483 memset(put_bits_ptr(&s->pb), 0, pad_bytes);
1486 /* this is not so easy because it is at the beginning of the data... */
1487 crc1 = av_bswap16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1488 frame + 4, frame_size_58 - 4));
1489 /* XXX: could precompute crc_inv */
1490 crc_inv = pow_poly((CRC16_POLY >> 1), (8 * frame_size_58) - 16, CRC16_POLY);
1491 crc1 = mul_poly(crc_inv, crc1, CRC16_POLY);
1492 AV_WB16(frame + 2, crc1);
1495 crc2 = av_bswap16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1496 frame + frame_size_58,
1497 frame_size - frame_size_58 - 2));
1498 AV_WB16(frame + frame_size - 2, crc2);
1503 * Write the frame to the output bitstream.
1505 static void output_frame(AC3EncodeContext *s,
1506 unsigned char *frame)
1510 init_put_bits(&s->pb, frame, AC3_MAX_CODED_FRAME_SIZE);
1512 output_frame_header(s);
1514 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++)
1515 output_audio_block(s, blk);
1517 output_frame_end(s);
1522 * Encode a single AC-3 frame.
1524 static int ac3_encode_frame(AVCodecContext *avctx,
1525 unsigned char *frame, int buf_size, void *data)
1527 AC3EncodeContext *s = avctx->priv_data;
1528 const int16_t *samples = data;
1531 if (s->bit_alloc.sr_code == 1)
1532 adjust_frame_size(s);
1534 deinterleave_input_samples(s, samples);
1538 process_exponents(s);
1540 ret = compute_bit_allocation(s);
1542 av_log(avctx, AV_LOG_ERROR, "Bit allocation failed. Try increasing the bitrate.\n");
1546 quantize_mantissas(s);
1548 output_frame(s, frame);
1550 return s->frame_size;
1555 * Finalize encoding and free any memory allocated by the encoder.
1557 static av_cold int ac3_encode_close(AVCodecContext *avctx)
1560 AC3EncodeContext *s = avctx->priv_data;
1562 for (ch = 0; ch < s->channels; ch++)
1563 av_freep(&s->planar_samples[ch]);
1564 av_freep(&s->planar_samples);
1565 av_freep(&s->bap_buffer);
1566 av_freep(&s->bap1_buffer);
1567 av_freep(&s->mdct_coef_buffer);
1568 av_freep(&s->exp_buffer);
1569 av_freep(&s->grouped_exp_buffer);
1570 av_freep(&s->psd_buffer);
1571 av_freep(&s->band_psd_buffer);
1572 av_freep(&s->mask_buffer);
1573 av_freep(&s->qmant_buffer);
1574 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
1575 AC3Block *block = &s->blocks[blk];
1576 av_freep(&block->bap);
1577 av_freep(&block->mdct_coef);
1578 av_freep(&block->exp);
1579 av_freep(&block->grouped_exp);
1580 av_freep(&block->psd);
1581 av_freep(&block->band_psd);
1582 av_freep(&block->mask);
1583 av_freep(&block->qmant);
1588 av_freep(&avctx->coded_frame);
1594 * Set channel information during initialization.
1596 static av_cold int set_channel_info(AC3EncodeContext *s, int channels,
1597 int64_t *channel_layout)
1601 if (channels < 1 || channels > AC3_MAX_CHANNELS)
1602 return AVERROR(EINVAL);
1603 if ((uint64_t)*channel_layout > 0x7FF)
1604 return AVERROR(EINVAL);
1605 ch_layout = *channel_layout;
1607 ch_layout = avcodec_guess_channel_layout(channels, CODEC_ID_AC3, NULL);
1608 if (av_get_channel_layout_nb_channels(ch_layout) != channels)
1609 return AVERROR(EINVAL);
1611 s->lfe_on = !!(ch_layout & AV_CH_LOW_FREQUENCY);
1612 s->channels = channels;
1613 s->fbw_channels = channels - s->lfe_on;
1614 s->lfe_channel = s->lfe_on ? s->fbw_channels : -1;
1616 ch_layout -= AV_CH_LOW_FREQUENCY;
1618 switch (ch_layout) {
1619 case AV_CH_LAYOUT_MONO: s->channel_mode = AC3_CHMODE_MONO; break;
1620 case AV_CH_LAYOUT_STEREO: s->channel_mode = AC3_CHMODE_STEREO; break;
1621 case AV_CH_LAYOUT_SURROUND: s->channel_mode = AC3_CHMODE_3F; break;
1622 case AV_CH_LAYOUT_2_1: s->channel_mode = AC3_CHMODE_2F1R; break;
1623 case AV_CH_LAYOUT_4POINT0: s->channel_mode = AC3_CHMODE_3F1R; break;
1624 case AV_CH_LAYOUT_QUAD:
1625 case AV_CH_LAYOUT_2_2: s->channel_mode = AC3_CHMODE_2F2R; break;
1626 case AV_CH_LAYOUT_5POINT0:
1627 case AV_CH_LAYOUT_5POINT0_BACK: s->channel_mode = AC3_CHMODE_3F2R; break;
1629 return AVERROR(EINVAL);
1632 s->channel_map = ff_ac3_enc_channel_map[s->channel_mode][s->lfe_on];
1633 *channel_layout = ch_layout;
1635 *channel_layout |= AV_CH_LOW_FREQUENCY;
1641 static av_cold int validate_options(AVCodecContext *avctx, AC3EncodeContext *s)
1645 /* validate channel layout */
1646 if (!avctx->channel_layout) {
1647 av_log(avctx, AV_LOG_WARNING, "No channel layout specified. The "
1648 "encoder will guess the layout, but it "
1649 "might be incorrect.\n");
1651 ret = set_channel_info(s, avctx->channels, &avctx->channel_layout);
1653 av_log(avctx, AV_LOG_ERROR, "invalid channel layout\n");
1657 /* validate sample rate */
1658 for (i = 0; i < 9; i++) {
1659 if ((ff_ac3_sample_rate_tab[i / 3] >> (i % 3)) == avctx->sample_rate)
1663 av_log(avctx, AV_LOG_ERROR, "invalid sample rate\n");
1664 return AVERROR(EINVAL);
1666 s->sample_rate = avctx->sample_rate;
1667 s->bit_alloc.sr_shift = i % 3;
1668 s->bit_alloc.sr_code = i / 3;
1670 /* validate bit rate */
1671 for (i = 0; i < 19; i++) {
1672 if ((ff_ac3_bitrate_tab[i] >> s->bit_alloc.sr_shift)*1000 == avctx->bit_rate)
1676 av_log(avctx, AV_LOG_ERROR, "invalid bit rate\n");
1677 return AVERROR(EINVAL);
1679 s->bit_rate = avctx->bit_rate;
1680 s->frame_size_code = i << 1;
1687 * Set bandwidth for all channels.
1688 * The user can optionally supply a cutoff frequency. Otherwise an appropriate
1689 * default value will be used.
1691 static av_cold void set_bandwidth(AC3EncodeContext *s, int cutoff)
1696 /* calculate bandwidth based on user-specified cutoff frequency */
1698 cutoff = av_clip(cutoff, 1, s->sample_rate >> 1);
1699 fbw_coeffs = cutoff * 2 * AC3_MAX_COEFS / s->sample_rate;
1700 bw_code = av_clip((fbw_coeffs - 73) / 3, 0, 60);
1702 /* use default bandwidth setting */
1703 /* XXX: should compute the bandwidth according to the frame
1704 size, so that we avoid annoying high frequency artifacts */
1708 /* set number of coefficients for each channel */
1709 for (ch = 0; ch < s->fbw_channels; ch++) {
1710 s->bandwidth_code[ch] = bw_code;
1711 s->nb_coefs[ch] = bw_code * 3 + 73;
1714 s->nb_coefs[s->lfe_channel] = 7; /* LFE channel always has 7 coefs */
1718 static av_cold int allocate_buffers(AVCodecContext *avctx)
1721 AC3EncodeContext *s = avctx->priv_data;
1723 FF_ALLOC_OR_GOTO(avctx, s->planar_samples, s->channels * sizeof(*s->planar_samples),
1725 for (ch = 0; ch < s->channels; ch++) {
1726 FF_ALLOCZ_OR_GOTO(avctx, s->planar_samples[ch],
1727 (AC3_FRAME_SIZE+AC3_BLOCK_SIZE) * sizeof(**s->planar_samples),
1730 FF_ALLOC_OR_GOTO(avctx, s->bap_buffer, AC3_MAX_BLOCKS * s->channels *
1731 AC3_MAX_COEFS * sizeof(*s->bap_buffer), alloc_fail);
1732 FF_ALLOC_OR_GOTO(avctx, s->bap1_buffer, AC3_MAX_BLOCKS * s->channels *
1733 AC3_MAX_COEFS * sizeof(*s->bap1_buffer), alloc_fail);
1734 FF_ALLOC_OR_GOTO(avctx, s->mdct_coef_buffer, AC3_MAX_BLOCKS * s->channels *
1735 AC3_MAX_COEFS * sizeof(*s->mdct_coef_buffer), alloc_fail);
1736 FF_ALLOC_OR_GOTO(avctx, s->exp_buffer, AC3_MAX_BLOCKS * s->channels *
1737 AC3_MAX_COEFS * sizeof(*s->exp_buffer), alloc_fail);
1738 FF_ALLOC_OR_GOTO(avctx, s->grouped_exp_buffer, AC3_MAX_BLOCKS * s->channels *
1739 128 * sizeof(*s->grouped_exp_buffer), alloc_fail);
1740 FF_ALLOC_OR_GOTO(avctx, s->psd_buffer, AC3_MAX_BLOCKS * s->channels *
1741 AC3_MAX_COEFS * sizeof(*s->psd_buffer), alloc_fail);
1742 FF_ALLOC_OR_GOTO(avctx, s->band_psd_buffer, AC3_MAX_BLOCKS * s->channels *
1743 64 * sizeof(*s->band_psd_buffer), alloc_fail);
1744 FF_ALLOC_OR_GOTO(avctx, s->mask_buffer, AC3_MAX_BLOCKS * s->channels *
1745 64 * sizeof(*s->mask_buffer), alloc_fail);
1746 FF_ALLOC_OR_GOTO(avctx, s->qmant_buffer, AC3_MAX_BLOCKS * s->channels *
1747 AC3_MAX_COEFS * sizeof(*s->qmant_buffer), alloc_fail);
1748 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
1749 AC3Block *block = &s->blocks[blk];
1750 FF_ALLOC_OR_GOTO(avctx, block->bap, s->channels * sizeof(*block->bap),
1752 FF_ALLOCZ_OR_GOTO(avctx, block->mdct_coef, s->channels * sizeof(*block->mdct_coef),
1754 FF_ALLOCZ_OR_GOTO(avctx, block->exp, s->channels * sizeof(*block->exp),
1756 FF_ALLOCZ_OR_GOTO(avctx, block->grouped_exp, s->channels * sizeof(*block->grouped_exp),
1758 FF_ALLOCZ_OR_GOTO(avctx, block->psd, s->channels * sizeof(*block->psd),
1760 FF_ALLOCZ_OR_GOTO(avctx, block->band_psd, s->channels * sizeof(*block->band_psd),
1762 FF_ALLOCZ_OR_GOTO(avctx, block->mask, s->channels * sizeof(*block->mask),
1764 FF_ALLOCZ_OR_GOTO(avctx, block->qmant, s->channels * sizeof(*block->qmant),
1767 for (ch = 0; ch < s->channels; ch++) {
1768 block->bap[ch] = &s->bap_buffer [AC3_MAX_COEFS * (blk * s->channels + ch)];
1769 block->mdct_coef[ch] = &s->mdct_coef_buffer [AC3_MAX_COEFS * (blk * s->channels + ch)];
1770 block->exp[ch] = &s->exp_buffer [AC3_MAX_COEFS * (blk * s->channels + ch)];
1771 block->grouped_exp[ch] = &s->grouped_exp_buffer[128 * (blk * s->channels + ch)];
1772 block->psd[ch] = &s->psd_buffer [AC3_MAX_COEFS * (blk * s->channels + ch)];
1773 block->band_psd[ch] = &s->band_psd_buffer [64 * (blk * s->channels + ch)];
1774 block->mask[ch] = &s->mask_buffer [64 * (blk * s->channels + ch)];
1775 block->qmant[ch] = &s->qmant_buffer [AC3_MAX_COEFS * (blk * s->channels + ch)];
1781 return AVERROR(ENOMEM);
1786 * Initialize the encoder.
1788 static av_cold int ac3_encode_init(AVCodecContext *avctx)
1790 AC3EncodeContext *s = avctx->priv_data;
1793 avctx->frame_size = AC3_FRAME_SIZE;
1797 ret = validate_options(avctx, s);
1801 s->bitstream_id = 8 + s->bit_alloc.sr_shift;
1802 s->bitstream_mode = 0; /* complete main audio service */
1804 s->frame_size_min = 2 * ff_ac3_frame_size_tab[s->frame_size_code][s->bit_alloc.sr_code];
1805 s->bits_written = 0;
1806 s->samples_written = 0;
1807 s->frame_size = s->frame_size_min;
1809 set_bandwidth(s, avctx->cutoff);
1815 s->mdct.avctx = avctx;
1816 ret = mdct_init(&s->mdct, 9);
1820 ret = allocate_buffers(avctx);
1824 avctx->coded_frame= avcodec_alloc_frame();
1826 dsputil_init(&s->dsp, avctx);
1830 ac3_encode_close(avctx);
1836 /*************************************************************************/
1839 #include "libavutil/lfg.h"
1841 #define FN (MDCT_SAMPLES/4)
1844 static void fft_test(AVLFG *lfg)
1846 IComplex in[FN], in1[FN];
1848 float sum_re, sum_im, a;
1850 for (i = 0; i < FN; i++) {
1851 in[i].re = av_lfg_get(lfg) % 65535 - 32767;
1852 in[i].im = av_lfg_get(lfg) % 65535 - 32767;
1858 for (k = 0; k < FN; k++) {
1861 for (n = 0; n < FN; n++) {
1862 a = -2 * M_PI * (n * k) / FN;
1863 sum_re += in1[n].re * cos(a) - in1[n].im * sin(a);
1864 sum_im += in1[n].re * sin(a) + in1[n].im * cos(a);
1866 av_log(NULL, AV_LOG_DEBUG, "%3d: %6d,%6d %6.0f,%6.0f\n",
1867 k, in[k].re, in[k].im, sum_re / FN, sum_im / FN);
1872 static void mdct_test(AVLFG *lfg)
1874 int16_t input[MDCT_SAMPLES];
1875 int32_t output[AC3_MAX_COEFS];
1876 float input1[MDCT_SAMPLES];
1877 float output1[AC3_MAX_COEFS];
1878 float s, a, err, e, emax;
1881 for (i = 0; i < MDCT_SAMPLES; i++) {
1882 input[i] = (av_lfg_get(lfg) % 65535 - 32767) * 9 / 10;
1883 input1[i] = input[i];
1886 mdct512(output, input);
1889 for (k = 0; k < AC3_MAX_COEFS; k++) {
1891 for (n = 0; n < MDCT_SAMPLES; n++) {
1892 a = (2*M_PI*(2*n+1+MDCT_SAMPLES/2)*(2*k+1) / (4 * MDCT_SAMPLES));
1893 s += input1[n] * cos(a);
1895 output1[k] = -2 * s / MDCT_SAMPLES;
1900 for (i = 0; i < AC3_MAX_COEFS; i++) {
1901 av_log(NULL, AV_LOG_DEBUG, "%3d: %7d %7.0f\n", i, output[i], output1[i]);
1902 e = output[i] - output1[i];
1907 av_log(NULL, AV_LOG_DEBUG, "err2=%f emax=%f\n", err / AC3_MAX_COEFS, emax);
1915 av_log_set_level(AV_LOG_DEBUG);
1926 AVCodec ac3_encoder = {
1930 sizeof(AC3EncodeContext),
1935 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,AV_SAMPLE_FMT_NONE},
1936 .long_name = NULL_IF_CONFIG_SMALL("ATSC A/52A (AC-3)"),
1937 .channel_layouts = (const int64_t[]){
1939 AV_CH_LAYOUT_STEREO,
1941 AV_CH_LAYOUT_SURROUND,
1944 AV_CH_LAYOUT_4POINT0,
1945 AV_CH_LAYOUT_5POINT0,
1946 AV_CH_LAYOUT_5POINT0_BACK,
1947 (AV_CH_LAYOUT_MONO | AV_CH_LOW_FREQUENCY),
1948 (AV_CH_LAYOUT_STEREO | AV_CH_LOW_FREQUENCY),
1949 (AV_CH_LAYOUT_2_1 | AV_CH_LOW_FREQUENCY),
1950 (AV_CH_LAYOUT_SURROUND | AV_CH_LOW_FREQUENCY),
1951 (AV_CH_LAYOUT_2_2 | AV_CH_LOW_FREQUENCY),
1952 (AV_CH_LAYOUT_QUAD | AV_CH_LOW_FREQUENCY),
1953 (AV_CH_LAYOUT_4POINT0 | AV_CH_LOW_FREQUENCY),
1954 AV_CH_LAYOUT_5POINT1,
1955 AV_CH_LAYOUT_5POINT1_BACK,