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"
40 /** Maximum number of exponent groups. +1 for separate DC exponent. */
41 #define AC3_MAX_EXP_GROUPS 85
43 /** Scale a float value by 2^bits and convert to an integer. */
44 #define SCALE_FLOAT(a, bits) lrintf((a) * (float)(1 << (bits)))
46 /** Scale a float value by 2^15, convert to an integer, and clip to range -32767..32767. */
47 #define FIX15(a) av_clip(SCALE_FLOAT(a, 15), -32767, 32767)
52 * Used in fixed-point MDCT calculation.
54 typedef struct IComplex {
58 typedef struct AC3MDCTContext {
59 const int16_t *window; ///< MDCT window function
60 int nbits; ///< log2(transform size)
61 int16_t *costab; ///< FFT cos table
62 int16_t *sintab; ///< FFT sin table
63 int16_t *xcos1; ///< MDCT cos table
64 int16_t *xsin1; ///< MDCT sin table
65 int16_t *rot_tmp; ///< temp buffer for pre-rotated samples
66 IComplex *cplx_tmp; ///< temp buffer for complex pre-rotated samples
70 * Data for a single audio block.
72 typedef struct AC3Block {
73 uint8_t **bap; ///< bit allocation pointers (bap)
74 int32_t **mdct_coef; ///< MDCT coefficients
75 uint8_t **exp; ///< original exponents
76 uint8_t **grouped_exp; ///< grouped exponents
77 int16_t **psd; ///< psd per frequency bin
78 int16_t **band_psd; ///< psd per critical band
79 int16_t **mask; ///< masking curve
80 uint16_t **qmant; ///< quantized mantissas
81 uint8_t exp_strategy[AC3_MAX_CHANNELS]; ///< exponent strategies
82 int8_t exp_shift[AC3_MAX_CHANNELS]; ///< exponent shift values
86 * AC-3 encoder private context.
88 typedef struct AC3EncodeContext {
89 PutBitContext pb; ///< bitstream writer context
91 AC3MDCTContext mdct; ///< MDCT context
93 AC3Block blocks[AC3_MAX_BLOCKS]; ///< per-block info
95 int bitstream_id; ///< bitstream id (bsid)
96 int bitstream_mode; ///< bitstream mode (bsmod)
98 int bit_rate; ///< target bit rate, in bits-per-second
99 int sample_rate; ///< sampling frequency, in Hz
101 int frame_size_min; ///< minimum frame size in case rounding is necessary
102 int frame_size; ///< current frame size in bytes
103 int frame_size_code; ///< frame size code (frmsizecod)
105 int bits_written; ///< bit count (used to avg. bitrate)
106 int samples_written; ///< sample count (used to avg. bitrate)
108 int fbw_channels; ///< number of full-bandwidth channels (nfchans)
109 int channels; ///< total number of channels (nchans)
110 int lfe_on; ///< indicates if there is an LFE channel (lfeon)
111 int lfe_channel; ///< channel index of the LFE channel
112 int channel_mode; ///< channel mode (acmod)
113 const uint8_t *channel_map; ///< channel map used to reorder channels
115 int cutoff; ///< user-specified cutoff frequency, in Hz
116 int bandwidth_code[AC3_MAX_CHANNELS]; ///< bandwidth code (0 to 60) (chbwcod)
117 int nb_coefs[AC3_MAX_CHANNELS];
119 /* bitrate allocation control */
120 int slow_gain_code; ///< slow gain code (sgaincod)
121 int slow_decay_code; ///< slow decay code (sdcycod)
122 int fast_decay_code; ///< fast decay code (fdcycod)
123 int db_per_bit_code; ///< dB/bit code (dbpbcod)
124 int floor_code; ///< floor code (floorcod)
125 AC3BitAllocParameters bit_alloc; ///< bit allocation parameters
126 int coarse_snr_offset; ///< coarse SNR offsets (csnroffst)
127 int fast_gain_code[AC3_MAX_CHANNELS]; ///< fast gain codes (signal-to-mask ratio) (fgaincod)
128 int fine_snr_offset[AC3_MAX_CHANNELS]; ///< fine SNR offsets (fsnroffst)
129 int frame_bits_fixed; ///< number of non-coefficient bits for fixed parameters
130 int frame_bits; ///< all frame bits except exponents and mantissas
131 int exponent_bits; ///< number of bits used for exponents
133 /* mantissa encoding */
134 int mant1_cnt, mant2_cnt, mant4_cnt; ///< mantissa counts for bap=1,2,4
135 uint16_t *qmant1_ptr, *qmant2_ptr, *qmant4_ptr; ///< mantissa pointers for bap=1,2,4
137 int16_t **planar_samples;
139 uint8_t *bap1_buffer;
140 int32_t *mdct_coef_buffer;
142 uint8_t *grouped_exp_buffer;
144 int16_t *band_psd_buffer;
145 int16_t *mask_buffer;
146 uint16_t *qmant_buffer;
148 DECLARE_ALIGNED(16, int16_t, windowed_samples)[AC3_WINDOW_SIZE];
153 * LUT for number of exponent groups.
154 * exponent_group_tab[exponent strategy-1][number of coefficients]
156 static uint8_t exponent_group_tab[3][256];
160 * List of supported channel layouts.
162 static const int64_t ac3_channel_layouts[] = {
166 AV_CH_LAYOUT_SURROUND,
169 AV_CH_LAYOUT_4POINT0,
170 AV_CH_LAYOUT_5POINT0,
171 AV_CH_LAYOUT_5POINT0_BACK,
172 (AV_CH_LAYOUT_MONO | AV_CH_LOW_FREQUENCY),
173 (AV_CH_LAYOUT_STEREO | AV_CH_LOW_FREQUENCY),
174 (AV_CH_LAYOUT_2_1 | AV_CH_LOW_FREQUENCY),
175 (AV_CH_LAYOUT_SURROUND | AV_CH_LOW_FREQUENCY),
176 (AV_CH_LAYOUT_2_2 | AV_CH_LOW_FREQUENCY),
177 (AV_CH_LAYOUT_QUAD | AV_CH_LOW_FREQUENCY),
178 (AV_CH_LAYOUT_4POINT0 | AV_CH_LOW_FREQUENCY),
179 AV_CH_LAYOUT_5POINT1,
180 AV_CH_LAYOUT_5POINT1_BACK,
186 * Adjust the frame size to make the average bit rate match the target bit rate.
187 * This is only needed for 11025, 22050, and 44100 sample rates.
189 static void adjust_frame_size(AC3EncodeContext *s)
191 while (s->bits_written >= s->bit_rate && s->samples_written >= s->sample_rate) {
192 s->bits_written -= s->bit_rate;
193 s->samples_written -= s->sample_rate;
195 s->frame_size = s->frame_size_min +
196 2 * (s->bits_written * s->sample_rate < s->samples_written * s->bit_rate);
197 s->bits_written += s->frame_size * 8;
198 s->samples_written += AC3_FRAME_SIZE;
203 * Deinterleave input samples.
204 * Channels are reordered from FFmpeg's default order to AC-3 order.
206 static void deinterleave_input_samples(AC3EncodeContext *s,
207 const int16_t *samples)
211 /* deinterleave and remap input samples */
212 for (ch = 0; ch < s->channels; ch++) {
216 /* copy last 256 samples of previous frame to the start of the current frame */
217 memcpy(&s->planar_samples[ch][0], &s->planar_samples[ch][AC3_FRAME_SIZE],
218 AC3_BLOCK_SIZE * sizeof(s->planar_samples[0][0]));
222 sptr = samples + s->channel_map[ch];
223 for (i = AC3_BLOCK_SIZE; i < AC3_FRAME_SIZE+AC3_BLOCK_SIZE; i++) {
224 s->planar_samples[ch][i] = *sptr;
232 * Finalize MDCT and free allocated memory.
234 static av_cold void mdct_end(AC3MDCTContext *mdct)
237 av_freep(&mdct->costab);
238 av_freep(&mdct->sintab);
239 av_freep(&mdct->xcos1);
240 av_freep(&mdct->xsin1);
241 av_freep(&mdct->rot_tmp);
242 av_freep(&mdct->cplx_tmp);
247 * Initialize FFT tables.
248 * @param ln log2(FFT size)
250 static av_cold int fft_init(AVCodecContext *avctx, AC3MDCTContext *mdct, int ln)
258 FF_ALLOC_OR_GOTO(avctx, mdct->costab, n2 * sizeof(*mdct->costab), fft_alloc_fail);
259 FF_ALLOC_OR_GOTO(avctx, mdct->sintab, n2 * sizeof(*mdct->sintab), fft_alloc_fail);
261 for (i = 0; i < n2; i++) {
262 alpha = 2.0 * M_PI * i / n;
263 mdct->costab[i] = FIX15(cos(alpha));
264 mdct->sintab[i] = FIX15(sin(alpha));
270 return AVERROR(ENOMEM);
275 * Initialize MDCT tables.
276 * @param nbits log2(MDCT size)
278 static av_cold int mdct_init(AVCodecContext *avctx, AC3MDCTContext *mdct,
288 ret = fft_init(avctx, mdct, nbits - 2);
292 mdct->window = ff_ac3_window;
294 FF_ALLOC_OR_GOTO(avctx, mdct->xcos1, n4 * sizeof(*mdct->xcos1), mdct_alloc_fail);
295 FF_ALLOC_OR_GOTO(avctx, mdct->xsin1, n4 * sizeof(*mdct->xsin1), mdct_alloc_fail);
296 FF_ALLOC_OR_GOTO(avctx, mdct->rot_tmp, n * sizeof(*mdct->rot_tmp), mdct_alloc_fail);
297 FF_ALLOC_OR_GOTO(avctx, mdct->cplx_tmp, n4 * sizeof(*mdct->cplx_tmp), mdct_alloc_fail);
299 for (i = 0; i < n4; i++) {
300 float alpha = 2.0 * M_PI * (i + 1.0 / 8.0) / n;
301 mdct->xcos1[i] = FIX15(-cos(alpha));
302 mdct->xsin1[i] = FIX15(-sin(alpha));
308 return AVERROR(ENOMEM);
313 #define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \
315 int ax, ay, bx, by; \
320 pre = (bx + ax) >> 1; \
321 pim = (by + ay) >> 1; \
322 qre = (bx - ax) >> 1; \
323 qim = (by - ay) >> 1; \
327 /** Complex multiply */
328 #define CMUL(pre, pim, are, aim, bre, bim) \
330 pre = (MUL16(are, bre) - MUL16(aim, bim)) >> 15; \
331 pim = (MUL16(are, bim) + MUL16(bre, aim)) >> 15; \
336 * Calculate a 2^n point complex FFT on 2^ln points.
337 * @param z complex input/output samples
338 * @param ln log2(FFT size)
340 static void fft(AC3MDCTContext *mdct, IComplex *z, int ln)
344 register IComplex *p,*q;
350 for (j = 0; j < np; j++) {
351 int k = av_reverse[j] >> (8 - ln);
353 FFSWAP(IComplex, z[k], z[j]);
361 BF(p[0].re, p[0].im, p[1].re, p[1].im,
362 p[0].re, p[0].im, p[1].re, p[1].im);
371 BF(p[0].re, p[0].im, p[2].re, p[2].im,
372 p[0].re, p[0].im, p[2].re, p[2].im);
373 BF(p[1].re, p[1].im, p[3].re, p[3].im,
374 p[1].re, p[1].im, p[3].im, -p[3].re);
386 for (j = 0; j < nblocks; j++) {
387 BF(p->re, p->im, q->re, q->im,
388 p->re, p->im, q->re, q->im);
391 for(l = nblocks; l < np2; l += nblocks) {
392 CMUL(tmp_re, tmp_im, mdct->costab[l], -mdct->sintab[l], q->re, q->im);
393 BF(p->re, p->im, q->re, q->im,
394 p->re, p->im, tmp_re, tmp_im);
401 nblocks = nblocks >> 1;
402 nloops = nloops << 1;
408 * Calculate a 512-point MDCT
409 * @param out 256 output frequency coefficients
410 * @param in 512 windowed input audio samples
412 static void mdct512(AC3MDCTContext *mdct, int32_t *out, int16_t *in)
414 int i, re, im, n, n2, n4;
415 int16_t *rot = mdct->rot_tmp;
416 IComplex *x = mdct->cplx_tmp;
418 n = 1 << mdct->nbits;
422 /* shift to simplify computations */
423 for (i = 0; i <n4; i++)
424 rot[i] = -in[i + 3*n4];
425 memcpy(&rot[n4], &in[0], 3*n4*sizeof(*in));
428 for (i = 0; i < n4; i++) {
429 re = ((int)rot[ 2*i] - (int)rot[ n-1-2*i]) >> 1;
430 im = -((int)rot[n2+2*i] - (int)rot[n2-1-2*i]) >> 1;
431 CMUL(x[i].re, x[i].im, re, im, -mdct->xcos1[i], mdct->xsin1[i]);
434 fft(mdct, x, mdct->nbits - 2);
437 for (i = 0; i < n4; i++) {
440 CMUL(out[n2-1-2*i], out[2*i], re, im, mdct->xsin1[i], mdct->xcos1[i]);
446 * Apply KBD window to input samples prior to MDCT.
448 static void apply_window(int16_t *output, const int16_t *input,
449 const int16_t *window, int n)
454 for (i = 0; i < n2; i++) {
455 output[i] = MUL16(input[i], window[i]) >> 15;
456 output[n-i-1] = MUL16(input[n-i-1], window[i]) >> 15;
462 * Calculate the log2() of the maximum absolute value in an array.
463 * @param tab input array
464 * @param n number of values in the array
465 * @return log2(max(abs(tab[])))
467 static int log2_tab(int16_t *tab, int n)
472 for (i = 0; i < n; i++)
480 * Left-shift each value in an array by a specified amount.
481 * @param tab input array
482 * @param n number of values in the array
483 * @param lshift left shift amount. a negative value means right shift.
485 static void lshift_tab(int16_t *tab, int n, int lshift)
490 for (i = 0; i < n; i++)
492 } else if (lshift < 0) {
494 for (i = 0; i < n; i++)
501 * Normalize the input samples to use the maximum available precision.
502 * This assumes signed 16-bit input samples. Exponents are reduced by 9 to
503 * match the 24-bit internal precision for MDCT coefficients.
505 * @return exponent shift
507 static int normalize_samples(AC3EncodeContext *s)
509 int v = 14 - log2_tab(s->windowed_samples, AC3_WINDOW_SIZE);
511 lshift_tab(s->windowed_samples, AC3_WINDOW_SIZE, v);
517 * Apply the MDCT to input samples to generate frequency coefficients.
518 * This applies the KBD window and normalizes the input to reduce precision
519 * loss due to fixed-point calculations.
521 static void apply_mdct(AC3EncodeContext *s)
525 for (ch = 0; ch < s->channels; ch++) {
526 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
527 AC3Block *block = &s->blocks[blk];
528 const int16_t *input_samples = &s->planar_samples[ch][blk * AC3_BLOCK_SIZE];
530 apply_window(s->windowed_samples, input_samples, s->mdct.window, AC3_WINDOW_SIZE);
532 block->exp_shift[ch] = normalize_samples(s);
534 mdct512(&s->mdct, block->mdct_coef[ch], s->windowed_samples);
541 * Initialize exponent tables.
543 static av_cold void exponent_init(AC3EncodeContext *s)
546 for (i = 73; i < 256; i++) {
547 exponent_group_tab[0][i] = (i - 1) / 3;
548 exponent_group_tab[1][i] = (i + 2) / 6;
549 exponent_group_tab[2][i] = (i + 8) / 12;
552 exponent_group_tab[0][7] = 2;
557 * Extract exponents from the MDCT coefficients.
558 * This takes into account the normalization that was done to the input samples
559 * by adjusting the exponents by the exponent shift values.
561 static void extract_exponents(AC3EncodeContext *s)
565 for (ch = 0; ch < s->channels; ch++) {
566 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
567 AC3Block *block = &s->blocks[blk];
568 for (i = 0; i < AC3_MAX_COEFS; i++) {
570 int v = abs(block->mdct_coef[ch][i]);
574 e = 23 - av_log2(v) + block->exp_shift[ch];
577 block->mdct_coef[ch][i] = 0;
580 block->exp[ch][i] = e;
588 * Exponent Difference Threshold.
589 * New exponents are sent if their SAD exceed this number.
591 #define EXP_DIFF_THRESHOLD 1000
595 * Calculate exponent strategies for all blocks in a single channel.
597 static void compute_exp_strategy_ch(AC3EncodeContext *s, uint8_t *exp_strategy,
603 /* estimate if the exponent variation & decide if they should be
604 reused in the next frame */
605 exp_strategy[0] = EXP_NEW;
606 for (blk = 1; blk < AC3_MAX_BLOCKS; blk++) {
607 exp_diff = s->dsp.sad[0](NULL, exp[blk], exp[blk-1], 16, 16);
608 if (exp_diff > EXP_DIFF_THRESHOLD)
609 exp_strategy[blk] = EXP_NEW;
611 exp_strategy[blk] = EXP_REUSE;
615 /* now select the encoding strategy type : if exponents are often
616 recoded, we use a coarse encoding */
618 while (blk < AC3_MAX_BLOCKS) {
620 while (blk1 < AC3_MAX_BLOCKS && exp_strategy[blk1] == EXP_REUSE)
622 switch (blk1 - blk) {
623 case 1: exp_strategy[blk] = EXP_D45; break;
625 case 3: exp_strategy[blk] = EXP_D25; break;
626 default: exp_strategy[blk] = EXP_D15; break;
634 * Calculate exponent strategies for all channels.
635 * Array arrangement is reversed to simplify the per-channel calculation.
637 static void compute_exp_strategy(AC3EncodeContext *s)
639 uint8_t *exp1[AC3_MAX_CHANNELS][AC3_MAX_BLOCKS];
640 uint8_t exp_str1[AC3_MAX_CHANNELS][AC3_MAX_BLOCKS];
643 for (ch = 0; ch < s->fbw_channels; ch++) {
644 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
645 exp1[ch][blk] = s->blocks[blk].exp[ch];
646 exp_str1[ch][blk] = s->blocks[blk].exp_strategy[ch];
649 compute_exp_strategy_ch(s, exp_str1[ch], exp1[ch]);
651 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++)
652 s->blocks[blk].exp_strategy[ch] = exp_str1[ch][blk];
656 s->blocks[0].exp_strategy[ch] = EXP_D15;
657 for (blk = 1; blk < AC3_MAX_BLOCKS; blk++)
658 s->blocks[blk].exp_strategy[ch] = EXP_REUSE;
664 * Set each encoded exponent in a block to the minimum of itself and the
665 * exponent in the same frequency bin of a following block.
666 * exp[i] = min(exp[i], exp1[i]
668 static void exponent_min(uint8_t *exp, uint8_t *exp1, int n)
671 for (i = 0; i < n; i++) {
672 if (exp1[i] < exp[i])
679 * Update the exponents so that they are the ones the decoder will decode.
681 static void encode_exponents_blk_ch(uint8_t *exp, int nb_exps, int exp_strategy)
685 nb_groups = exponent_group_tab[exp_strategy-1][nb_exps] * 3;
687 /* for each group, compute the minimum exponent */
688 switch(exp_strategy) {
690 for (i = 1, k = 1; i <= nb_groups; i++) {
691 uint8_t exp_min = exp[k];
692 if (exp[k+1] < exp_min)
699 for (i = 1, k = 1; i <= nb_groups; i++) {
700 uint8_t exp_min = exp[k];
701 if (exp[k+1] < exp_min)
703 if (exp[k+2] < exp_min)
705 if (exp[k+3] < exp_min)
713 /* constraint for DC exponent */
717 /* decrease the delta between each groups to within 2 so that they can be
718 differentially encoded */
719 for (i = 1; i <= nb_groups; i++)
720 exp[i] = FFMIN(exp[i], exp[i-1] + 2);
723 exp[i] = FFMIN(exp[i], exp[i+1] + 2);
725 /* now we have the exponent values the decoder will see */
726 switch (exp_strategy) {
728 for (i = nb_groups, k = nb_groups * 2; i > 0; i--) {
729 uint8_t exp1 = exp[i];
735 for (i = nb_groups, k = nb_groups * 4; i > 0; i--) {
736 exp[k] = exp[k-1] = exp[k-2] = exp[k-3] = exp[i];
745 * Encode exponents from original extracted form to what the decoder will see.
746 * This copies and groups exponents based on exponent strategy and reduces
747 * deltas between adjacent exponent groups so that they can be differentially
750 static void encode_exponents(AC3EncodeContext *s)
752 int blk, blk1, blk2, ch;
753 AC3Block *block, *block1, *block2;
755 for (ch = 0; ch < s->channels; ch++) {
757 block = &s->blocks[0];
758 while (blk < AC3_MAX_BLOCKS) {
761 /* for the EXP_REUSE case we select the min of the exponents */
762 while (blk1 < AC3_MAX_BLOCKS && block1->exp_strategy[ch] == EXP_REUSE) {
763 exponent_min(block->exp[ch], block1->exp[ch], s->nb_coefs[ch]);
767 encode_exponents_blk_ch(block->exp[ch], s->nb_coefs[ch],
768 block->exp_strategy[ch]);
769 /* copy encoded exponents for reuse case */
771 for (blk2 = blk+1; blk2 < blk1; blk2++, block2++) {
772 memcpy(block2->exp[ch], block->exp[ch],
773 s->nb_coefs[ch] * sizeof(uint8_t));
784 * 3 delta-encoded exponents are in each 7-bit group. The number of groups
785 * varies depending on exponent strategy and bandwidth.
787 static void group_exponents(AC3EncodeContext *s)
790 int group_size, nb_groups, bit_count;
792 int delta0, delta1, delta2;
796 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
797 AC3Block *block = &s->blocks[blk];
798 for (ch = 0; ch < s->channels; ch++) {
799 if (block->exp_strategy[ch] == EXP_REUSE) {
802 group_size = block->exp_strategy[ch] + (block->exp_strategy[ch] == EXP_D45);
803 nb_groups = exponent_group_tab[block->exp_strategy[ch]-1][s->nb_coefs[ch]];
804 bit_count += 4 + (nb_groups * 7);
809 block->grouped_exp[ch][0] = exp1;
811 /* remaining exponents are delta encoded */
812 for (i = 1; i <= nb_groups; i++) {
813 /* merge three delta in one code */
817 delta0 = exp1 - exp0 + 2;
822 delta1 = exp1 - exp0 + 2;
827 delta2 = exp1 - exp0 + 2;
829 block->grouped_exp[ch][i] = ((delta0 * 5 + delta1) * 5) + delta2;
834 s->exponent_bits = bit_count;
839 * Calculate final exponents from the supplied MDCT coefficients and exponent shift.
840 * Extract exponents from MDCT coefficients, calculate exponent strategies,
841 * and encode final exponents.
843 static void process_exponents(AC3EncodeContext *s)
845 extract_exponents(s);
847 compute_exp_strategy(s);
856 * Count frame bits that are based solely on fixed parameters.
857 * This only has to be run once when the encoder is initialized.
859 static void count_frame_bits_fixed(AC3EncodeContext *s)
861 static const int frame_bits_inc[8] = { 0, 0, 2, 2, 2, 4, 2, 4 };
866 * no dynamic range codes
867 * no channel coupling
869 * bit allocation parameters do not change between blocks
870 * SNR offsets do not change between blocks
871 * no delta bit allocation
878 frame_bits += frame_bits_inc[s->channel_mode];
881 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
882 frame_bits += s->fbw_channels * 2 + 2; /* blksw * c, dithflag * c, dynrnge, cplstre */
883 if (s->channel_mode == AC3_CHMODE_STEREO) {
884 frame_bits++; /* rematstr */
888 frame_bits += 2 * s->fbw_channels; /* chexpstr[2] * c */
890 frame_bits++; /* lfeexpstr */
891 frame_bits++; /* baie */
892 frame_bits++; /* snr */
893 frame_bits += 2; /* delta / skip */
895 frame_bits++; /* cplinu for block 0 */
897 /* sdcycod[2], fdcycod[2], sgaincod[2], dbpbcod[2], floorcod[3] */
899 /* (fsnoffset[4] + fgaincod[4]) * c */
900 frame_bits += 2*4 + 3 + 6 + s->channels * (4 + 3);
902 /* auxdatae, crcrsv */
908 s->frame_bits_fixed = frame_bits;
913 * Initialize bit allocation.
914 * Set default parameter codes and calculate parameter values.
916 static void bit_alloc_init(AC3EncodeContext *s)
920 /* init default parameters */
921 s->slow_decay_code = 2;
922 s->fast_decay_code = 1;
923 s->slow_gain_code = 1;
924 s->db_per_bit_code = 3;
926 for (ch = 0; ch < s->channels; ch++)
927 s->fast_gain_code[ch] = 4;
929 /* initial snr offset */
930 s->coarse_snr_offset = 40;
932 /* compute real values */
933 /* currently none of these values change during encoding, so we can just
934 set them once at initialization */
935 s->bit_alloc.slow_decay = ff_ac3_slow_decay_tab[s->slow_decay_code] >> s->bit_alloc.sr_shift;
936 s->bit_alloc.fast_decay = ff_ac3_fast_decay_tab[s->fast_decay_code] >> s->bit_alloc.sr_shift;
937 s->bit_alloc.slow_gain = ff_ac3_slow_gain_tab[s->slow_gain_code];
938 s->bit_alloc.db_per_bit = ff_ac3_db_per_bit_tab[s->db_per_bit_code];
939 s->bit_alloc.floor = ff_ac3_floor_tab[s->floor_code];
941 count_frame_bits_fixed(s);
946 * Count the bits used to encode the frame, minus exponents and mantissas.
947 * Bits based on fixed parameters have already been counted, so now we just
948 * have to add the bits based on parameters that change during encoding.
950 static void count_frame_bits(AC3EncodeContext *s)
955 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
956 uint8_t *exp_strategy = s->blocks[blk].exp_strategy;
957 for (ch = 0; ch < s->fbw_channels; ch++) {
958 if (exp_strategy[ch] != EXP_REUSE)
959 frame_bits += 6 + 2; /* chbwcod[6], gainrng[2] */
962 s->frame_bits = s->frame_bits_fixed + frame_bits;
967 * Calculate the number of bits needed to encode a set of mantissas.
969 static int compute_mantissa_size(int mant_cnt[5], uint8_t *bap, int nb_coefs)
974 for (i = 0; i < nb_coefs; i++) {
977 // bap=1 to bap=4 will be counted in compute_mantissa_size_final
979 } else if (b <= 13) {
980 // bap=5 to bap=13 use (bap-1) bits
983 // bap=14 uses 14 bits and bap=15 uses 16 bits
984 bits += (b == 14) ? 14 : 16;
992 * Finalize the mantissa bit count by adding in the grouped mantissas.
994 static int compute_mantissa_size_final(int mant_cnt[5])
996 // bap=1 : 3 mantissas in 5 bits
997 int bits = (mant_cnt[1] / 3) * 5;
998 // bap=2 : 3 mantissas in 7 bits
999 // bap=4 : 2 mantissas in 7 bits
1000 bits += ((mant_cnt[2] / 3) + (mant_cnt[4] >> 1)) * 7;
1001 // bap=3 : each mantissa is 3 bits
1002 bits += mant_cnt[3] * 3;
1008 * Calculate masking curve based on the final exponents.
1009 * Also calculate the power spectral densities to use in future calculations.
1011 static void bit_alloc_masking(AC3EncodeContext *s)
1015 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
1016 AC3Block *block = &s->blocks[blk];
1017 for (ch = 0; ch < s->channels; ch++) {
1018 /* We only need psd and mask for calculating bap.
1019 Since we currently do not calculate bap when exponent
1020 strategy is EXP_REUSE we do not need to calculate psd or mask. */
1021 if (block->exp_strategy[ch] != EXP_REUSE) {
1022 ff_ac3_bit_alloc_calc_psd(block->exp[ch], 0,
1024 block->psd[ch], block->band_psd[ch]);
1025 ff_ac3_bit_alloc_calc_mask(&s->bit_alloc, block->band_psd[ch],
1027 ff_ac3_fast_gain_tab[s->fast_gain_code[ch]],
1028 ch == s->lfe_channel,
1029 DBA_NONE, 0, NULL, NULL, NULL,
1038 * Ensure that bap for each block and channel point to the current bap_buffer.
1039 * They may have been switched during the bit allocation search.
1041 static void reset_block_bap(AC3EncodeContext *s)
1044 if (s->blocks[0].bap[0] == s->bap_buffer)
1046 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
1047 for (ch = 0; ch < s->channels; ch++) {
1048 s->blocks[blk].bap[ch] = &s->bap_buffer[AC3_MAX_COEFS * (blk * s->channels + ch)];
1055 * Run the bit allocation with a given SNR offset.
1056 * This calculates the bit allocation pointers that will be used to determine
1057 * the quantization of each mantissa.
1058 * @return the number of bits needed for mantissas if the given SNR offset is
1061 static int bit_alloc(AC3EncodeContext *s, int snr_offset)
1067 snr_offset = (snr_offset - 240) << 2;
1071 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
1072 AC3Block *block = &s->blocks[blk];
1073 // initialize grouped mantissa counts. these are set so that they are
1074 // padded to the next whole group size when bits are counted in
1075 // compute_mantissa_size_final
1076 mant_cnt[0] = mant_cnt[3] = 0;
1077 mant_cnt[1] = mant_cnt[2] = 2;
1079 for (ch = 0; ch < s->channels; ch++) {
1080 /* Currently the only bit allocation parameters which vary across
1081 blocks within a frame are the exponent values. We can take
1082 advantage of that by reusing the bit allocation pointers
1083 whenever we reuse exponents. */
1084 if (block->exp_strategy[ch] == EXP_REUSE) {
1085 memcpy(block->bap[ch], s->blocks[blk-1].bap[ch], AC3_MAX_COEFS);
1087 ff_ac3_bit_alloc_calc_bap(block->mask[ch], block->psd[ch], 0,
1088 s->nb_coefs[ch], snr_offset,
1089 s->bit_alloc.floor, ff_ac3_bap_tab,
1092 mantissa_bits += compute_mantissa_size(mant_cnt, block->bap[ch], s->nb_coefs[ch]);
1094 mantissa_bits += compute_mantissa_size_final(mant_cnt);
1096 return mantissa_bits;
1101 * Constant bitrate bit allocation search.
1102 * Find the largest SNR offset that will allow data to fit in the frame.
1104 static int cbr_bit_allocation(AC3EncodeContext *s)
1108 int snr_offset, snr_incr;
1110 bits_left = 8 * s->frame_size - (s->frame_bits + s->exponent_bits);
1112 snr_offset = s->coarse_snr_offset << 4;
1114 while (snr_offset >= 0 &&
1115 bit_alloc(s, snr_offset) > bits_left) {
1119 return AVERROR(EINVAL);
1121 FFSWAP(uint8_t *, s->bap_buffer, s->bap1_buffer);
1122 for (snr_incr = 64; snr_incr > 0; snr_incr >>= 2) {
1123 while (snr_offset + 64 <= 1023 &&
1124 bit_alloc(s, snr_offset + snr_incr) <= bits_left) {
1125 snr_offset += snr_incr;
1126 FFSWAP(uint8_t *, s->bap_buffer, s->bap1_buffer);
1129 FFSWAP(uint8_t *, s->bap_buffer, s->bap1_buffer);
1132 s->coarse_snr_offset = snr_offset >> 4;
1133 for (ch = 0; ch < s->channels; ch++)
1134 s->fine_snr_offset[ch] = snr_offset & 0xF;
1141 * Downgrade exponent strategies to reduce the bits used by the exponents.
1142 * This is a fallback for when bit allocation fails with the normal exponent
1143 * strategies. Each time this function is run it only downgrades the
1144 * strategy in 1 channel of 1 block.
1145 * @return non-zero if downgrade was unsuccessful
1147 static int downgrade_exponents(AC3EncodeContext *s)
1151 for (ch = 0; ch < s->fbw_channels; ch++) {
1152 for (blk = AC3_MAX_BLOCKS-1; blk >= 0; blk--) {
1153 if (s->blocks[blk].exp_strategy[ch] == EXP_D15) {
1154 s->blocks[blk].exp_strategy[ch] = EXP_D25;
1159 for (ch = 0; ch < s->fbw_channels; ch++) {
1160 for (blk = AC3_MAX_BLOCKS-1; blk >= 0; blk--) {
1161 if (s->blocks[blk].exp_strategy[ch] == EXP_D25) {
1162 s->blocks[blk].exp_strategy[ch] = EXP_D45;
1167 for (ch = 0; ch < s->fbw_channels; ch++) {
1168 /* block 0 cannot reuse exponents, so only downgrade D45 to REUSE if
1169 the block number > 0 */
1170 for (blk = AC3_MAX_BLOCKS-1; blk > 0; blk--) {
1171 if (s->blocks[blk].exp_strategy[ch] > EXP_REUSE) {
1172 s->blocks[blk].exp_strategy[ch] = EXP_REUSE;
1182 * Reduce the bandwidth to reduce the number of bits used for a given SNR offset.
1183 * This is a second fallback for when bit allocation still fails after exponents
1184 * have been downgraded.
1185 * @return non-zero if bandwidth reduction was unsuccessful
1187 static int reduce_bandwidth(AC3EncodeContext *s, int min_bw_code)
1191 if (s->bandwidth_code[0] > min_bw_code) {
1192 for (ch = 0; ch < s->fbw_channels; ch++) {
1193 s->bandwidth_code[ch]--;
1194 s->nb_coefs[ch] = s->bandwidth_code[ch] * 3 + 73;
1203 * Perform bit allocation search.
1204 * Finds the SNR offset value that maximizes quality and fits in the specified
1205 * frame size. Output is the SNR offset and a set of bit allocation pointers
1206 * used to quantize the mantissas.
1208 static int compute_bit_allocation(AC3EncodeContext *s)
1212 count_frame_bits(s);
1214 bit_alloc_masking(s);
1216 ret = cbr_bit_allocation(s);
1218 /* fallback 1: downgrade exponents */
1219 if (!downgrade_exponents(s)) {
1220 extract_exponents(s);
1221 encode_exponents(s);
1223 ret = compute_bit_allocation(s);
1227 /* fallback 2: reduce bandwidth */
1228 /* only do this if the user has not specified a specific cutoff
1230 if (!s->cutoff && !reduce_bandwidth(s, 0)) {
1231 process_exponents(s);
1232 ret = compute_bit_allocation(s);
1236 /* fallbacks were not enough... */
1245 * Symmetric quantization on 'levels' levels.
1247 static inline int sym_quant(int c, int e, int levels)
1252 v = (levels * (c << e)) >> 24;
1254 v = (levels >> 1) + v;
1256 v = (levels * ((-c) << e)) >> 24;
1258 v = (levels >> 1) - v;
1260 assert(v >= 0 && v < levels);
1266 * Asymmetric quantization on 2^qbits levels.
1268 static inline int asym_quant(int c, int e, int qbits)
1272 lshift = e + qbits - 24;
1279 m = (1 << (qbits-1));
1283 return v & ((1 << qbits)-1);
1288 * Quantize a set of mantissas for a single channel in a single block.
1290 static void quantize_mantissas_blk_ch(AC3EncodeContext *s, int32_t *mdct_coef,
1291 int8_t exp_shift, uint8_t *exp,
1292 uint8_t *bap, uint16_t *qmant, int n)
1296 for (i = 0; i < n; i++) {
1298 int c = mdct_coef[i];
1299 int e = exp[i] - exp_shift;
1306 v = sym_quant(c, e, 3);
1307 switch (s->mant1_cnt) {
1309 s->qmant1_ptr = &qmant[i];
1314 *s->qmant1_ptr += 3 * v;
1319 *s->qmant1_ptr += v;
1326 v = sym_quant(c, e, 5);
1327 switch (s->mant2_cnt) {
1329 s->qmant2_ptr = &qmant[i];
1334 *s->qmant2_ptr += 5 * v;
1339 *s->qmant2_ptr += v;
1346 v = sym_quant(c, e, 7);
1349 v = sym_quant(c, e, 11);
1350 switch (s->mant4_cnt) {
1352 s->qmant4_ptr = &qmant[i];
1357 *s->qmant4_ptr += v;
1364 v = sym_quant(c, e, 15);
1367 v = asym_quant(c, e, 14);
1370 v = asym_quant(c, e, 16);
1373 v = asym_quant(c, e, b - 1);
1382 * Quantize mantissas using coefficients, exponents, and bit allocation pointers.
1384 static void quantize_mantissas(AC3EncodeContext *s)
1389 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
1390 AC3Block *block = &s->blocks[blk];
1391 s->mant1_cnt = s->mant2_cnt = s->mant4_cnt = 0;
1392 s->qmant1_ptr = s->qmant2_ptr = s->qmant4_ptr = NULL;
1394 for (ch = 0; ch < s->channels; ch++) {
1395 quantize_mantissas_blk_ch(s, block->mdct_coef[ch], block->exp_shift[ch],
1396 block->exp[ch], block->bap[ch],
1397 block->qmant[ch], s->nb_coefs[ch]);
1404 * Write the AC-3 frame header to the output bitstream.
1406 static void output_frame_header(AC3EncodeContext *s)
1408 put_bits(&s->pb, 16, 0x0b77); /* frame header */
1409 put_bits(&s->pb, 16, 0); /* crc1: will be filled later */
1410 put_bits(&s->pb, 2, s->bit_alloc.sr_code);
1411 put_bits(&s->pb, 6, s->frame_size_code + (s->frame_size - s->frame_size_min) / 2);
1412 put_bits(&s->pb, 5, s->bitstream_id);
1413 put_bits(&s->pb, 3, s->bitstream_mode);
1414 put_bits(&s->pb, 3, s->channel_mode);
1415 if ((s->channel_mode & 0x01) && s->channel_mode != AC3_CHMODE_MONO)
1416 put_bits(&s->pb, 2, 1); /* XXX -4.5 dB */
1417 if (s->channel_mode & 0x04)
1418 put_bits(&s->pb, 2, 1); /* XXX -6 dB */
1419 if (s->channel_mode == AC3_CHMODE_STEREO)
1420 put_bits(&s->pb, 2, 0); /* surround not indicated */
1421 put_bits(&s->pb, 1, s->lfe_on); /* LFE */
1422 put_bits(&s->pb, 5, 31); /* dialog norm: -31 db */
1423 put_bits(&s->pb, 1, 0); /* no compression control word */
1424 put_bits(&s->pb, 1, 0); /* no lang code */
1425 put_bits(&s->pb, 1, 0); /* no audio production info */
1426 put_bits(&s->pb, 1, 0); /* no copyright */
1427 put_bits(&s->pb, 1, 1); /* original bitstream */
1428 put_bits(&s->pb, 1, 0); /* no time code 1 */
1429 put_bits(&s->pb, 1, 0); /* no time code 2 */
1430 put_bits(&s->pb, 1, 0); /* no additional bit stream info */
1435 * Write one audio block to the output bitstream.
1437 static void output_audio_block(AC3EncodeContext *s, int block_num)
1439 int ch, i, baie, rbnd;
1440 AC3Block *block = &s->blocks[block_num];
1442 /* block switching */
1443 for (ch = 0; ch < s->fbw_channels; ch++)
1444 put_bits(&s->pb, 1, 0);
1447 for (ch = 0; ch < s->fbw_channels; ch++)
1448 put_bits(&s->pb, 1, 1);
1450 /* dynamic range codes */
1451 put_bits(&s->pb, 1, 0);
1453 /* channel coupling */
1455 put_bits(&s->pb, 1, 1); /* coupling strategy present */
1456 put_bits(&s->pb, 1, 0); /* no coupling strategy */
1458 put_bits(&s->pb, 1, 0); /* no new coupling strategy */
1461 /* stereo rematrixing */
1462 if (s->channel_mode == AC3_CHMODE_STEREO) {
1464 /* first block must define rematrixing (rematstr) */
1465 put_bits(&s->pb, 1, 1);
1467 /* dummy rematrixing rematflg(1:4)=0 */
1468 for (rbnd = 0; rbnd < 4; rbnd++)
1469 put_bits(&s->pb, 1, 0);
1471 /* no matrixing (but should be used in the future) */
1472 put_bits(&s->pb, 1, 0);
1476 /* exponent strategy */
1477 for (ch = 0; ch < s->fbw_channels; ch++)
1478 put_bits(&s->pb, 2, block->exp_strategy[ch]);
1480 put_bits(&s->pb, 1, block->exp_strategy[s->lfe_channel]);
1483 for (ch = 0; ch < s->fbw_channels; ch++) {
1484 if (block->exp_strategy[ch] != EXP_REUSE)
1485 put_bits(&s->pb, 6, s->bandwidth_code[ch]);
1489 for (ch = 0; ch < s->channels; ch++) {
1492 if (block->exp_strategy[ch] == EXP_REUSE)
1496 put_bits(&s->pb, 4, block->grouped_exp[ch][0]);
1498 /* exponent groups */
1499 nb_groups = exponent_group_tab[block->exp_strategy[ch]-1][s->nb_coefs[ch]];
1500 for (i = 1; i <= nb_groups; i++)
1501 put_bits(&s->pb, 7, block->grouped_exp[ch][i]);
1503 /* gain range info */
1504 if (ch != s->lfe_channel)
1505 put_bits(&s->pb, 2, 0);
1508 /* bit allocation info */
1509 baie = (block_num == 0);
1510 put_bits(&s->pb, 1, baie);
1512 put_bits(&s->pb, 2, s->slow_decay_code);
1513 put_bits(&s->pb, 2, s->fast_decay_code);
1514 put_bits(&s->pb, 2, s->slow_gain_code);
1515 put_bits(&s->pb, 2, s->db_per_bit_code);
1516 put_bits(&s->pb, 3, s->floor_code);
1520 put_bits(&s->pb, 1, baie);
1522 put_bits(&s->pb, 6, s->coarse_snr_offset);
1523 for (ch = 0; ch < s->channels; ch++) {
1524 put_bits(&s->pb, 4, s->fine_snr_offset[ch]);
1525 put_bits(&s->pb, 3, s->fast_gain_code[ch]);
1529 put_bits(&s->pb, 1, 0); /* no delta bit allocation */
1530 put_bits(&s->pb, 1, 0); /* no data to skip */
1533 for (ch = 0; ch < s->channels; ch++) {
1535 for (i = 0; i < s->nb_coefs[ch]; i++) {
1536 q = block->qmant[ch][i];
1537 b = block->bap[ch][i];
1540 case 1: if (q != 128) put_bits(&s->pb, 5, q); break;
1541 case 2: if (q != 128) put_bits(&s->pb, 7, q); break;
1542 case 3: put_bits(&s->pb, 3, q); break;
1543 case 4: if (q != 128) put_bits(&s->pb, 7, q); break;
1544 case 14: put_bits(&s->pb, 14, q); break;
1545 case 15: put_bits(&s->pb, 16, q); break;
1546 default: put_bits(&s->pb, b-1, q); break;
1553 /** CRC-16 Polynomial */
1554 #define CRC16_POLY ((1 << 0) | (1 << 2) | (1 << 15) | (1 << 16))
1557 static unsigned int mul_poly(unsigned int a, unsigned int b, unsigned int poly)
1574 static unsigned int pow_poly(unsigned int a, unsigned int n, unsigned int poly)
1580 r = mul_poly(r, a, poly);
1581 a = mul_poly(a, a, poly);
1589 * Fill the end of the frame with 0's and compute the two CRCs.
1591 static void output_frame_end(AC3EncodeContext *s)
1593 const AVCRC *crc_ctx = av_crc_get_table(AV_CRC_16_ANSI);
1594 int frame_size_58, pad_bytes, crc1, crc2_partial, crc2, crc_inv;
1597 frame_size_58 = ((s->frame_size >> 2) + (s->frame_size >> 4)) << 1;
1599 /* pad the remainder of the frame with zeros */
1600 flush_put_bits(&s->pb);
1602 pad_bytes = s->frame_size - (put_bits_ptr(&s->pb) - frame) - 2;
1603 assert(pad_bytes >= 0);
1605 memset(put_bits_ptr(&s->pb), 0, pad_bytes);
1608 /* this is not so easy because it is at the beginning of the data... */
1609 crc1 = av_bswap16(av_crc(crc_ctx, 0, frame + 4, frame_size_58 - 4));
1610 crc_inv = s->crc_inv[s->frame_size > s->frame_size_min];
1611 crc1 = mul_poly(crc_inv, crc1, CRC16_POLY);
1612 AV_WB16(frame + 2, crc1);
1615 crc2_partial = av_crc(crc_ctx, 0, frame + frame_size_58,
1616 s->frame_size - frame_size_58 - 3);
1617 crc2 = av_crc(crc_ctx, crc2_partial, frame + s->frame_size - 3, 1);
1618 /* ensure crc2 does not match sync word by flipping crcrsv bit if needed */
1619 if (crc2 == 0x770B) {
1620 frame[s->frame_size - 3] ^= 0x1;
1621 crc2 = av_crc(crc_ctx, crc2_partial, frame + s->frame_size - 3, 1);
1623 crc2 = av_bswap16(crc2);
1624 AV_WB16(frame + s->frame_size - 2, crc2);
1629 * Write the frame to the output bitstream.
1631 static void output_frame(AC3EncodeContext *s, unsigned char *frame)
1635 init_put_bits(&s->pb, frame, AC3_MAX_CODED_FRAME_SIZE);
1637 output_frame_header(s);
1639 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++)
1640 output_audio_block(s, blk);
1642 output_frame_end(s);
1647 * Encode a single AC-3 frame.
1649 static int ac3_encode_frame(AVCodecContext *avctx, unsigned char *frame,
1650 int buf_size, void *data)
1652 AC3EncodeContext *s = avctx->priv_data;
1653 const int16_t *samples = data;
1656 if (s->bit_alloc.sr_code == 1)
1657 adjust_frame_size(s);
1659 deinterleave_input_samples(s, samples);
1663 process_exponents(s);
1665 ret = compute_bit_allocation(s);
1667 av_log(avctx, AV_LOG_ERROR, "Bit allocation failed. Try increasing the bitrate.\n");
1671 quantize_mantissas(s);
1673 output_frame(s, frame);
1675 return s->frame_size;
1680 * Finalize encoding and free any memory allocated by the encoder.
1682 static av_cold int ac3_encode_close(AVCodecContext *avctx)
1685 AC3EncodeContext *s = avctx->priv_data;
1687 for (ch = 0; ch < s->channels; ch++)
1688 av_freep(&s->planar_samples[ch]);
1689 av_freep(&s->planar_samples);
1690 av_freep(&s->bap_buffer);
1691 av_freep(&s->bap1_buffer);
1692 av_freep(&s->mdct_coef_buffer);
1693 av_freep(&s->exp_buffer);
1694 av_freep(&s->grouped_exp_buffer);
1695 av_freep(&s->psd_buffer);
1696 av_freep(&s->band_psd_buffer);
1697 av_freep(&s->mask_buffer);
1698 av_freep(&s->qmant_buffer);
1699 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
1700 AC3Block *block = &s->blocks[blk];
1701 av_freep(&block->bap);
1702 av_freep(&block->mdct_coef);
1703 av_freep(&block->exp);
1704 av_freep(&block->grouped_exp);
1705 av_freep(&block->psd);
1706 av_freep(&block->band_psd);
1707 av_freep(&block->mask);
1708 av_freep(&block->qmant);
1713 av_freep(&avctx->coded_frame);
1719 * Set channel information during initialization.
1721 static av_cold int set_channel_info(AC3EncodeContext *s, int channels,
1722 int64_t *channel_layout)
1726 if (channels < 1 || channels > AC3_MAX_CHANNELS)
1727 return AVERROR(EINVAL);
1728 if ((uint64_t)*channel_layout > 0x7FF)
1729 return AVERROR(EINVAL);
1730 ch_layout = *channel_layout;
1732 ch_layout = avcodec_guess_channel_layout(channels, CODEC_ID_AC3, NULL);
1733 if (av_get_channel_layout_nb_channels(ch_layout) != channels)
1734 return AVERROR(EINVAL);
1736 s->lfe_on = !!(ch_layout & AV_CH_LOW_FREQUENCY);
1737 s->channels = channels;
1738 s->fbw_channels = channels - s->lfe_on;
1739 s->lfe_channel = s->lfe_on ? s->fbw_channels : -1;
1741 ch_layout -= AV_CH_LOW_FREQUENCY;
1743 switch (ch_layout) {
1744 case AV_CH_LAYOUT_MONO: s->channel_mode = AC3_CHMODE_MONO; break;
1745 case AV_CH_LAYOUT_STEREO: s->channel_mode = AC3_CHMODE_STEREO; break;
1746 case AV_CH_LAYOUT_SURROUND: s->channel_mode = AC3_CHMODE_3F; break;
1747 case AV_CH_LAYOUT_2_1: s->channel_mode = AC3_CHMODE_2F1R; break;
1748 case AV_CH_LAYOUT_4POINT0: s->channel_mode = AC3_CHMODE_3F1R; break;
1749 case AV_CH_LAYOUT_QUAD:
1750 case AV_CH_LAYOUT_2_2: s->channel_mode = AC3_CHMODE_2F2R; break;
1751 case AV_CH_LAYOUT_5POINT0:
1752 case AV_CH_LAYOUT_5POINT0_BACK: s->channel_mode = AC3_CHMODE_3F2R; break;
1754 return AVERROR(EINVAL);
1757 s->channel_map = ff_ac3_enc_channel_map[s->channel_mode][s->lfe_on];
1758 *channel_layout = ch_layout;
1760 *channel_layout |= AV_CH_LOW_FREQUENCY;
1766 static av_cold int validate_options(AVCodecContext *avctx, AC3EncodeContext *s)
1770 /* validate channel layout */
1771 if (!avctx->channel_layout) {
1772 av_log(avctx, AV_LOG_WARNING, "No channel layout specified. The "
1773 "encoder will guess the layout, but it "
1774 "might be incorrect.\n");
1776 ret = set_channel_info(s, avctx->channels, &avctx->channel_layout);
1778 av_log(avctx, AV_LOG_ERROR, "invalid channel layout\n");
1782 /* validate sample rate */
1783 for (i = 0; i < 9; i++) {
1784 if ((ff_ac3_sample_rate_tab[i / 3] >> (i % 3)) == avctx->sample_rate)
1788 av_log(avctx, AV_LOG_ERROR, "invalid sample rate\n");
1789 return AVERROR(EINVAL);
1791 s->sample_rate = avctx->sample_rate;
1792 s->bit_alloc.sr_shift = i % 3;
1793 s->bit_alloc.sr_code = i / 3;
1795 /* validate bit rate */
1796 for (i = 0; i < 19; i++) {
1797 if ((ff_ac3_bitrate_tab[i] >> s->bit_alloc.sr_shift)*1000 == avctx->bit_rate)
1801 av_log(avctx, AV_LOG_ERROR, "invalid bit rate\n");
1802 return AVERROR(EINVAL);
1804 s->bit_rate = avctx->bit_rate;
1805 s->frame_size_code = i << 1;
1807 /* validate cutoff */
1808 if (avctx->cutoff < 0) {
1809 av_log(avctx, AV_LOG_ERROR, "invalid cutoff frequency\n");
1810 return AVERROR(EINVAL);
1812 s->cutoff = avctx->cutoff;
1813 if (s->cutoff > (s->sample_rate >> 1))
1814 s->cutoff = s->sample_rate >> 1;
1821 * Set bandwidth for all channels.
1822 * The user can optionally supply a cutoff frequency. Otherwise an appropriate
1823 * default value will be used.
1825 static av_cold void set_bandwidth(AC3EncodeContext *s)
1830 /* calculate bandwidth based on user-specified cutoff frequency */
1832 fbw_coeffs = s->cutoff * 2 * AC3_MAX_COEFS / s->sample_rate;
1833 bw_code = av_clip((fbw_coeffs - 73) / 3, 0, 60);
1835 /* use default bandwidth setting */
1836 /* XXX: should compute the bandwidth according to the frame
1837 size, so that we avoid annoying high frequency artifacts */
1841 /* set number of coefficients for each channel */
1842 for (ch = 0; ch < s->fbw_channels; ch++) {
1843 s->bandwidth_code[ch] = bw_code;
1844 s->nb_coefs[ch] = bw_code * 3 + 73;
1847 s->nb_coefs[s->lfe_channel] = 7; /* LFE channel always has 7 coefs */
1851 static av_cold int allocate_buffers(AVCodecContext *avctx)
1854 AC3EncodeContext *s = avctx->priv_data;
1856 FF_ALLOC_OR_GOTO(avctx, s->planar_samples, s->channels * sizeof(*s->planar_samples),
1858 for (ch = 0; ch < s->channels; ch++) {
1859 FF_ALLOCZ_OR_GOTO(avctx, s->planar_samples[ch],
1860 (AC3_FRAME_SIZE+AC3_BLOCK_SIZE) * sizeof(**s->planar_samples),
1863 FF_ALLOC_OR_GOTO(avctx, s->bap_buffer, AC3_MAX_BLOCKS * s->channels *
1864 AC3_MAX_COEFS * sizeof(*s->bap_buffer), alloc_fail);
1865 FF_ALLOC_OR_GOTO(avctx, s->bap1_buffer, AC3_MAX_BLOCKS * s->channels *
1866 AC3_MAX_COEFS * sizeof(*s->bap1_buffer), alloc_fail);
1867 FF_ALLOC_OR_GOTO(avctx, s->mdct_coef_buffer, AC3_MAX_BLOCKS * s->channels *
1868 AC3_MAX_COEFS * sizeof(*s->mdct_coef_buffer), alloc_fail);
1869 FF_ALLOC_OR_GOTO(avctx, s->exp_buffer, AC3_MAX_BLOCKS * s->channels *
1870 AC3_MAX_COEFS * sizeof(*s->exp_buffer), alloc_fail);
1871 FF_ALLOC_OR_GOTO(avctx, s->grouped_exp_buffer, AC3_MAX_BLOCKS * s->channels *
1872 128 * sizeof(*s->grouped_exp_buffer), alloc_fail);
1873 FF_ALLOC_OR_GOTO(avctx, s->psd_buffer, AC3_MAX_BLOCKS * s->channels *
1874 AC3_MAX_COEFS * sizeof(*s->psd_buffer), alloc_fail);
1875 FF_ALLOC_OR_GOTO(avctx, s->band_psd_buffer, AC3_MAX_BLOCKS * s->channels *
1876 64 * sizeof(*s->band_psd_buffer), alloc_fail);
1877 FF_ALLOC_OR_GOTO(avctx, s->mask_buffer, AC3_MAX_BLOCKS * s->channels *
1878 64 * sizeof(*s->mask_buffer), alloc_fail);
1879 FF_ALLOC_OR_GOTO(avctx, s->qmant_buffer, AC3_MAX_BLOCKS * s->channels *
1880 AC3_MAX_COEFS * sizeof(*s->qmant_buffer), alloc_fail);
1881 for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
1882 AC3Block *block = &s->blocks[blk];
1883 FF_ALLOC_OR_GOTO(avctx, block->bap, s->channels * sizeof(*block->bap),
1885 FF_ALLOCZ_OR_GOTO(avctx, block->mdct_coef, s->channels * sizeof(*block->mdct_coef),
1887 FF_ALLOCZ_OR_GOTO(avctx, block->exp, s->channels * sizeof(*block->exp),
1889 FF_ALLOCZ_OR_GOTO(avctx, block->grouped_exp, s->channels * sizeof(*block->grouped_exp),
1891 FF_ALLOCZ_OR_GOTO(avctx, block->psd, s->channels * sizeof(*block->psd),
1893 FF_ALLOCZ_OR_GOTO(avctx, block->band_psd, s->channels * sizeof(*block->band_psd),
1895 FF_ALLOCZ_OR_GOTO(avctx, block->mask, s->channels * sizeof(*block->mask),
1897 FF_ALLOCZ_OR_GOTO(avctx, block->qmant, s->channels * sizeof(*block->qmant),
1900 for (ch = 0; ch < s->channels; ch++) {
1901 block->bap[ch] = &s->bap_buffer [AC3_MAX_COEFS * (blk * s->channels + ch)];
1902 block->mdct_coef[ch] = &s->mdct_coef_buffer [AC3_MAX_COEFS * (blk * s->channels + ch)];
1903 block->exp[ch] = &s->exp_buffer [AC3_MAX_COEFS * (blk * s->channels + ch)];
1904 block->grouped_exp[ch] = &s->grouped_exp_buffer[128 * (blk * s->channels + ch)];
1905 block->psd[ch] = &s->psd_buffer [AC3_MAX_COEFS * (blk * s->channels + ch)];
1906 block->band_psd[ch] = &s->band_psd_buffer [64 * (blk * s->channels + ch)];
1907 block->mask[ch] = &s->mask_buffer [64 * (blk * s->channels + ch)];
1908 block->qmant[ch] = &s->qmant_buffer [AC3_MAX_COEFS * (blk * s->channels + ch)];
1914 return AVERROR(ENOMEM);
1919 * Initialize the encoder.
1921 static av_cold int ac3_encode_init(AVCodecContext *avctx)
1923 AC3EncodeContext *s = avctx->priv_data;
1924 int ret, frame_size_58;
1926 avctx->frame_size = AC3_FRAME_SIZE;
1930 ret = validate_options(avctx, s);
1934 s->bitstream_id = 8 + s->bit_alloc.sr_shift;
1935 s->bitstream_mode = 0; /* complete main audio service */
1937 s->frame_size_min = 2 * ff_ac3_frame_size_tab[s->frame_size_code][s->bit_alloc.sr_code];
1938 s->bits_written = 0;
1939 s->samples_written = 0;
1940 s->frame_size = s->frame_size_min;
1942 /* calculate crc_inv for both possible frame sizes */
1943 frame_size_58 = (( s->frame_size >> 2) + ( s->frame_size >> 4)) << 1;
1944 s->crc_inv[0] = pow_poly((CRC16_POLY >> 1), (8 * frame_size_58) - 16, CRC16_POLY);
1945 if (s->bit_alloc.sr_code == 1) {
1946 frame_size_58 = (((s->frame_size+2) >> 2) + ((s->frame_size+2) >> 4)) << 1;
1947 s->crc_inv[1] = pow_poly((CRC16_POLY >> 1), (8 * frame_size_58) - 16, CRC16_POLY);
1956 ret = mdct_init(avctx, &s->mdct, 9);
1960 ret = allocate_buffers(avctx);
1964 avctx->coded_frame= avcodec_alloc_frame();
1966 dsputil_init(&s->dsp, avctx);
1970 ac3_encode_close(avctx);
1976 /*************************************************************************/
1979 #include "libavutil/lfg.h"
1981 #define MDCT_NBITS 9
1982 #define MDCT_SAMPLES (1 << MDCT_NBITS)
1983 #define FN (MDCT_SAMPLES/4)
1986 static void fft_test(AC3MDCTContext *mdct, AVLFG *lfg)
1988 IComplex in[FN], in1[FN];
1990 float sum_re, sum_im, a;
1992 for (i = 0; i < FN; i++) {
1993 in[i].re = av_lfg_get(lfg) % 65535 - 32767;
1994 in[i].im = av_lfg_get(lfg) % 65535 - 32767;
2000 for (k = 0; k < FN; k++) {
2003 for (n = 0; n < FN; n++) {
2004 a = -2 * M_PI * (n * k) / FN;
2005 sum_re += in1[n].re * cos(a) - in1[n].im * sin(a);
2006 sum_im += in1[n].re * sin(a) + in1[n].im * cos(a);
2008 av_log(NULL, AV_LOG_DEBUG, "%3d: %6d,%6d %6.0f,%6.0f\n",
2009 k, in[k].re, in[k].im, sum_re / FN, sum_im / FN);
2014 static void mdct_test(AC3MDCTContext *mdct, AVLFG *lfg)
2016 int16_t input[MDCT_SAMPLES];
2017 int32_t output[AC3_MAX_COEFS];
2018 float input1[MDCT_SAMPLES];
2019 float output1[AC3_MAX_COEFS];
2020 float s, a, err, e, emax;
2023 for (i = 0; i < MDCT_SAMPLES; i++) {
2024 input[i] = (av_lfg_get(lfg) % 65535 - 32767) * 9 / 10;
2025 input1[i] = input[i];
2028 mdct512(mdct, output, input);
2031 for (k = 0; k < AC3_MAX_COEFS; k++) {
2033 for (n = 0; n < MDCT_SAMPLES; n++) {
2034 a = (2*M_PI*(2*n+1+MDCT_SAMPLES/2)*(2*k+1) / (4 * MDCT_SAMPLES));
2035 s += input1[n] * cos(a);
2037 output1[k] = -2 * s / MDCT_SAMPLES;
2042 for (i = 0; i < AC3_MAX_COEFS; i++) {
2043 av_log(NULL, AV_LOG_DEBUG, "%3d: %7d %7.0f\n", i, output[i], output1[i]);
2044 e = output[i] - output1[i];
2049 av_log(NULL, AV_LOG_DEBUG, "err2=%f emax=%f\n", err / AC3_MAX_COEFS, emax);
2056 AC3MDCTContext mdct;
2059 av_log_set_level(AV_LOG_DEBUG);
2060 mdct_init(&mdct, 9);
2062 fft_test(&mdct, &lfg);
2063 mdct_test(&mdct, &lfg);
2070 AVCodec ac3_encoder = {
2074 sizeof(AC3EncodeContext),
2079 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,AV_SAMPLE_FMT_NONE},
2080 .long_name = NULL_IF_CONFIG_SMALL("ATSC A/52A (AC-3)"),
2081 .channel_layouts = ac3_channel_layouts,