More cosmetics
[FFMpeg-mirror/DVCPRO-HD.git] / libavcodec / ac3enc.c
blob86ccb49822c436ae66c6dff201e645daf61cdc2b
1 /*
2 * The simplest AC3 encoder
3 * Copyright (c) 2000 Fabrice Bellard.
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 /**
23 * @file ac3enc.c
24 * The simplest AC3 encoder.
26 //#define DEBUG
27 //#define DEBUG_BITALLOC
28 #include "libavutil/crc.h"
29 #include "avcodec.h"
30 #include "bitstream.h"
31 #include "ac3.h"
33 typedef struct AC3EncodeContext {
34 PutBitContext pb;
35 int nb_channels;
36 int nb_all_channels;
37 int lfe_channel;
38 int bit_rate;
39 unsigned int sample_rate;
40 unsigned int bitstream_id;
41 unsigned int frame_size_min; /* minimum frame size in case rounding is necessary */
42 unsigned int frame_size; /* current frame size in words */
43 unsigned int bits_written;
44 unsigned int samples_written;
45 int sr_shift;
46 unsigned int frame_size_code;
47 unsigned int sr_code; /* frequency */
48 unsigned int channel_mode;
49 int lfe;
50 unsigned int bitstream_mode;
51 short last_samples[AC3_MAX_CHANNELS][256];
52 unsigned int chbwcod[AC3_MAX_CHANNELS];
53 int nb_coefs[AC3_MAX_CHANNELS];
55 /* bitrate allocation control */
56 int slow_gain_code, slow_decay_code, fast_decay_code, db_per_bit_code, floor_code;
57 AC3BitAllocParameters bit_alloc;
58 int coarse_snr_offset;
59 int fast_gain_code[AC3_MAX_CHANNELS];
60 int fine_snr_offset[AC3_MAX_CHANNELS];
61 /* mantissa encoding */
62 int mant1_cnt, mant2_cnt, mant4_cnt;
63 } AC3EncodeContext;
65 static int16_t costab[64];
66 static int16_t sintab[64];
67 static int16_t xcos1[128];
68 static int16_t xsin1[128];
70 #define MDCT_NBITS 9
71 #define N (1 << MDCT_NBITS)
73 /* new exponents are sent if their Norm 1 exceed this number */
74 #define EXP_DIFF_THRESHOLD 1000
76 static inline int16_t fix15(float a)
78 int v;
79 v = (int)(a * (float)(1 << 15));
80 if (v < -32767)
81 v = -32767;
82 else if (v > 32767)
83 v = 32767;
84 return v;
87 typedef struct IComplex {
88 short re,im;
89 } IComplex;
91 static void fft_init(int ln)
93 int i, n;
94 float alpha;
96 n = 1 << ln;
98 for(i=0;i<(n/2);i++) {
99 alpha = 2 * M_PI * (float)i / (float)n;
100 costab[i] = fix15(cos(alpha));
101 sintab[i] = fix15(sin(alpha));
105 /* butter fly op */
106 #define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \
108 int ax, ay, bx, by;\
109 bx=pre1;\
110 by=pim1;\
111 ax=qre1;\
112 ay=qim1;\
113 pre = (bx + ax) >> 1;\
114 pim = (by + ay) >> 1;\
115 qre = (bx - ax) >> 1;\
116 qim = (by - ay) >> 1;\
119 #define MUL16(a,b) ((a) * (b))
121 #define CMUL(pre, pim, are, aim, bre, bim) \
123 pre = (MUL16(are, bre) - MUL16(aim, bim)) >> 15;\
124 pim = (MUL16(are, bim) + MUL16(bre, aim)) >> 15;\
128 /* do a 2^n point complex fft on 2^ln points. */
129 static void fft(IComplex *z, int ln)
131 int j, l, np, np2;
132 int nblocks, nloops;
133 register IComplex *p,*q;
134 int tmp_re, tmp_im;
136 np = 1 << ln;
138 /* reverse */
139 for(j=0;j<np;j++) {
140 int k = ff_reverse[j] >> (8 - ln);
141 if (k < j)
142 FFSWAP(IComplex, z[k], z[j]);
145 /* pass 0 */
147 p=&z[0];
148 j=(np >> 1);
149 do {
150 BF(p[0].re, p[0].im, p[1].re, p[1].im,
151 p[0].re, p[0].im, p[1].re, p[1].im);
152 p+=2;
153 } while (--j != 0);
155 /* pass 1 */
157 p=&z[0];
158 j=np >> 2;
159 do {
160 BF(p[0].re, p[0].im, p[2].re, p[2].im,
161 p[0].re, p[0].im, p[2].re, p[2].im);
162 BF(p[1].re, p[1].im, p[3].re, p[3].im,
163 p[1].re, p[1].im, p[3].im, -p[3].re);
164 p+=4;
165 } while (--j != 0);
167 /* pass 2 .. ln-1 */
169 nblocks = np >> 3;
170 nloops = 1 << 2;
171 np2 = np >> 1;
172 do {
173 p = z;
174 q = z + nloops;
175 for (j = 0; j < nblocks; ++j) {
177 BF(p->re, p->im, q->re, q->im,
178 p->re, p->im, q->re, q->im);
180 p++;
181 q++;
182 for(l = nblocks; l < np2; l += nblocks) {
183 CMUL(tmp_re, tmp_im, costab[l], -sintab[l], q->re, q->im);
184 BF(p->re, p->im, q->re, q->im,
185 p->re, p->im, tmp_re, tmp_im);
186 p++;
187 q++;
189 p += nloops;
190 q += nloops;
192 nblocks = nblocks >> 1;
193 nloops = nloops << 1;
194 } while (nblocks != 0);
197 /* do a 512 point mdct */
198 static void mdct512(int32_t *out, int16_t *in)
200 int i, re, im, re1, im1;
201 int16_t rot[N];
202 IComplex x[N/4];
204 /* shift to simplify computations */
205 for(i=0;i<N/4;i++)
206 rot[i] = -in[i + 3*N/4];
207 for(i=N/4;i<N;i++)
208 rot[i] = in[i - N/4];
210 /* pre rotation */
211 for(i=0;i<N/4;i++) {
212 re = ((int)rot[2*i] - (int)rot[N-1-2*i]) >> 1;
213 im = -((int)rot[N/2+2*i] - (int)rot[N/2-1-2*i]) >> 1;
214 CMUL(x[i].re, x[i].im, re, im, -xcos1[i], xsin1[i]);
217 fft(x, MDCT_NBITS - 2);
219 /* post rotation */
220 for(i=0;i<N/4;i++) {
221 re = x[i].re;
222 im = x[i].im;
223 CMUL(re1, im1, re, im, xsin1[i], xcos1[i]);
224 out[2*i] = im1;
225 out[N/2-1-2*i] = re1;
229 /* XXX: use another norm ? */
230 static int calc_exp_diff(uint8_t *exp1, uint8_t *exp2, int n)
232 int sum, i;
233 sum = 0;
234 for(i=0;i<n;i++) {
235 sum += abs(exp1[i] - exp2[i]);
237 return sum;
240 static void compute_exp_strategy(uint8_t exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS],
241 uint8_t exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
242 int ch, int is_lfe)
244 int i, j;
245 int exp_diff;
247 /* estimate if the exponent variation & decide if they should be
248 reused in the next frame */
249 exp_strategy[0][ch] = EXP_NEW;
250 for(i=1;i<NB_BLOCKS;i++) {
251 exp_diff = calc_exp_diff(exp[i][ch], exp[i-1][ch], N/2);
252 #ifdef DEBUG
253 av_log(NULL, AV_LOG_DEBUG, "exp_diff=%d\n", exp_diff);
254 #endif
255 if (exp_diff > EXP_DIFF_THRESHOLD)
256 exp_strategy[i][ch] = EXP_NEW;
257 else
258 exp_strategy[i][ch] = EXP_REUSE;
260 if (is_lfe)
261 return;
263 /* now select the encoding strategy type : if exponents are often
264 recoded, we use a coarse encoding */
265 i = 0;
266 while (i < NB_BLOCKS) {
267 j = i + 1;
268 while (j < NB_BLOCKS && exp_strategy[j][ch] == EXP_REUSE)
269 j++;
270 switch(j - i) {
271 case 1:
272 exp_strategy[i][ch] = EXP_D45;
273 break;
274 case 2:
275 case 3:
276 exp_strategy[i][ch] = EXP_D25;
277 break;
278 default:
279 exp_strategy[i][ch] = EXP_D15;
280 break;
282 i = j;
286 /* set exp[i] to min(exp[i], exp1[i]) */
287 static void exponent_min(uint8_t exp[N/2], uint8_t exp1[N/2], int n)
289 int i;
291 for(i=0;i<n;i++) {
292 if (exp1[i] < exp[i])
293 exp[i] = exp1[i];
297 /* update the exponents so that they are the ones the decoder will
298 decode. Return the number of bits used to code the exponents */
299 static int encode_exp(uint8_t encoded_exp[N/2],
300 uint8_t exp[N/2],
301 int nb_exps,
302 int exp_strategy)
304 int group_size, nb_groups, i, j, k, exp_min;
305 uint8_t exp1[N/2];
307 switch(exp_strategy) {
308 case EXP_D15:
309 group_size = 1;
310 break;
311 case EXP_D25:
312 group_size = 2;
313 break;
314 default:
315 case EXP_D45:
316 group_size = 4;
317 break;
319 nb_groups = ((nb_exps + (group_size * 3) - 4) / (3 * group_size)) * 3;
321 /* for each group, compute the minimum exponent */
322 exp1[0] = exp[0]; /* DC exponent is handled separately */
323 k = 1;
324 for(i=1;i<=nb_groups;i++) {
325 exp_min = exp[k];
326 assert(exp_min >= 0 && exp_min <= 24);
327 for(j=1;j<group_size;j++) {
328 if (exp[k+j] < exp_min)
329 exp_min = exp[k+j];
331 exp1[i] = exp_min;
332 k += group_size;
335 /* constraint for DC exponent */
336 if (exp1[0] > 15)
337 exp1[0] = 15;
339 /* Decrease the delta between each groups to within 2
340 * so that they can be differentially encoded */
341 for (i=1;i<=nb_groups;i++)
342 exp1[i] = FFMIN(exp1[i], exp1[i-1] + 2);
343 for (i=nb_groups-1;i>=0;i--)
344 exp1[i] = FFMIN(exp1[i], exp1[i+1] + 2);
346 /* now we have the exponent values the decoder will see */
347 encoded_exp[0] = exp1[0];
348 k = 1;
349 for(i=1;i<=nb_groups;i++) {
350 for(j=0;j<group_size;j++) {
351 encoded_exp[k+j] = exp1[i];
353 k += group_size;
356 #if defined(DEBUG)
357 av_log(NULL, AV_LOG_DEBUG, "exponents: strategy=%d\n", exp_strategy);
358 for(i=0;i<=nb_groups * group_size;i++) {
359 av_log(NULL, AV_LOG_DEBUG, "%d ", encoded_exp[i]);
361 av_log(NULL, AV_LOG_DEBUG, "\n");
362 #endif
364 return 4 + (nb_groups / 3) * 7;
367 /* return the size in bits taken by the mantissa */
368 static int compute_mantissa_size(AC3EncodeContext *s, uint8_t *m, int nb_coefs)
370 int bits, mant, i;
372 bits = 0;
373 for(i=0;i<nb_coefs;i++) {
374 mant = m[i];
375 switch(mant) {
376 case 0:
377 /* nothing */
378 break;
379 case 1:
380 /* 3 mantissa in 5 bits */
381 if (s->mant1_cnt == 0)
382 bits += 5;
383 if (++s->mant1_cnt == 3)
384 s->mant1_cnt = 0;
385 break;
386 case 2:
387 /* 3 mantissa in 7 bits */
388 if (s->mant2_cnt == 0)
389 bits += 7;
390 if (++s->mant2_cnt == 3)
391 s->mant2_cnt = 0;
392 break;
393 case 3:
394 bits += 3;
395 break;
396 case 4:
397 /* 2 mantissa in 7 bits */
398 if (s->mant4_cnt == 0)
399 bits += 7;
400 if (++s->mant4_cnt == 2)
401 s->mant4_cnt = 0;
402 break;
403 case 14:
404 bits += 14;
405 break;
406 case 15:
407 bits += 16;
408 break;
409 default:
410 bits += mant - 1;
411 break;
414 return bits;
418 static void bit_alloc_masking(AC3EncodeContext *s,
419 uint8_t encoded_exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
420 uint8_t exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS],
421 int16_t psd[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
422 int16_t mask[NB_BLOCKS][AC3_MAX_CHANNELS][50])
424 int blk, ch;
425 int16_t band_psd[NB_BLOCKS][AC3_MAX_CHANNELS][50];
427 for(blk=0; blk<NB_BLOCKS; blk++) {
428 for(ch=0;ch<s->nb_all_channels;ch++) {
429 if(exp_strategy[blk][ch] == EXP_REUSE) {
430 memcpy(psd[blk][ch], psd[blk-1][ch], (N/2)*sizeof(int16_t));
431 memcpy(mask[blk][ch], mask[blk-1][ch], 50*sizeof(int16_t));
432 } else {
433 ff_ac3_bit_alloc_calc_psd(encoded_exp[blk][ch], 0,
434 s->nb_coefs[ch],
435 psd[blk][ch], band_psd[blk][ch]);
436 ff_ac3_bit_alloc_calc_mask(&s->bit_alloc, band_psd[blk][ch],
437 0, s->nb_coefs[ch],
438 ff_ac3_fast_gain_tab[s->fast_gain_code[ch]],
439 ch == s->lfe_channel,
440 DBA_NONE, 0, NULL, NULL, NULL,
441 mask[blk][ch]);
447 static int bit_alloc(AC3EncodeContext *s,
448 int16_t mask[NB_BLOCKS][AC3_MAX_CHANNELS][50],
449 int16_t psd[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
450 uint8_t bap[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
451 int frame_bits, int coarse_snr_offset, int fine_snr_offset)
453 int i, ch;
454 int snr_offset;
456 snr_offset = (((coarse_snr_offset - 15) << 4) + fine_snr_offset) << 2;
458 /* compute size */
459 for(i=0;i<NB_BLOCKS;i++) {
460 s->mant1_cnt = 0;
461 s->mant2_cnt = 0;
462 s->mant4_cnt = 0;
463 for(ch=0;ch<s->nb_all_channels;ch++) {
464 ff_ac3_bit_alloc_calc_bap(mask[i][ch], psd[i][ch], 0,
465 s->nb_coefs[ch], snr_offset,
466 s->bit_alloc.floor, ff_ac3_bap_tab,
467 bap[i][ch]);
468 frame_bits += compute_mantissa_size(s, bap[i][ch],
469 s->nb_coefs[ch]);
472 #if 0
473 printf("csnr=%d fsnr=%d frame_bits=%d diff=%d\n",
474 coarse_snr_offset, fine_snr_offset, frame_bits,
475 16 * s->frame_size - ((frame_bits + 7) & ~7));
476 #endif
477 return 16 * s->frame_size - frame_bits;
480 #define SNR_INC1 4
482 static int compute_bit_allocation(AC3EncodeContext *s,
483 uint8_t bap[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
484 uint8_t encoded_exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
485 uint8_t exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS],
486 int frame_bits)
488 int i, ch;
489 int coarse_snr_offset, fine_snr_offset;
490 uint8_t bap1[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
491 int16_t psd[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
492 int16_t mask[NB_BLOCKS][AC3_MAX_CHANNELS][50];
493 static const int frame_bits_inc[8] = { 0, 0, 2, 2, 2, 4, 2, 4 };
495 /* init default parameters */
496 s->slow_decay_code = 2;
497 s->fast_decay_code = 1;
498 s->slow_gain_code = 1;
499 s->db_per_bit_code = 2;
500 s->floor_code = 4;
501 for(ch=0;ch<s->nb_all_channels;ch++)
502 s->fast_gain_code[ch] = 4;
504 /* compute real values */
505 s->bit_alloc.sr_code = s->sr_code;
506 s->bit_alloc.sr_shift = s->sr_shift;
507 s->bit_alloc.slow_decay = ff_ac3_slow_decay_tab[s->slow_decay_code] >> s->sr_shift;
508 s->bit_alloc.fast_decay = ff_ac3_fast_decay_tab[s->fast_decay_code] >> s->sr_shift;
509 s->bit_alloc.slow_gain = ff_ac3_slow_gain_tab[s->slow_gain_code];
510 s->bit_alloc.db_per_bit = ff_ac3_db_per_bit_tab[s->db_per_bit_code];
511 s->bit_alloc.floor = ff_ac3_floor_tab[s->floor_code];
513 /* header size */
514 frame_bits += 65;
515 // if (s->channel_mode == 2)
516 // frame_bits += 2;
517 frame_bits += frame_bits_inc[s->channel_mode];
519 /* audio blocks */
520 for(i=0;i<NB_BLOCKS;i++) {
521 frame_bits += s->nb_channels * 2 + 2; /* blksw * c, dithflag * c, dynrnge, cplstre */
522 if (s->channel_mode == AC3_CHMODE_STEREO) {
523 frame_bits++; /* rematstr */
524 if(i==0) frame_bits += 4;
526 frame_bits += 2 * s->nb_channels; /* chexpstr[2] * c */
527 if (s->lfe)
528 frame_bits++; /* lfeexpstr */
529 for(ch=0;ch<s->nb_channels;ch++) {
530 if (exp_strategy[i][ch] != EXP_REUSE)
531 frame_bits += 6 + 2; /* chbwcod[6], gainrng[2] */
533 frame_bits++; /* baie */
534 frame_bits++; /* snr */
535 frame_bits += 2; /* delta / skip */
537 frame_bits++; /* cplinu for block 0 */
538 /* bit alloc info */
539 /* sdcycod[2], fdcycod[2], sgaincod[2], dbpbcod[2], floorcod[3] */
540 /* csnroffset[6] */
541 /* (fsnoffset[4] + fgaincod[4]) * c */
542 frame_bits += 2*4 + 3 + 6 + s->nb_all_channels * (4 + 3);
544 /* auxdatae, crcrsv */
545 frame_bits += 2;
547 /* CRC */
548 frame_bits += 16;
550 /* calculate psd and masking curve before doing bit allocation */
551 bit_alloc_masking(s, encoded_exp, exp_strategy, psd, mask);
553 /* now the big work begins : do the bit allocation. Modify the snr
554 offset until we can pack everything in the requested frame size */
556 coarse_snr_offset = s->coarse_snr_offset;
557 while (coarse_snr_offset >= 0 &&
558 bit_alloc(s, mask, psd, bap, frame_bits, coarse_snr_offset, 0) < 0)
559 coarse_snr_offset -= SNR_INC1;
560 if (coarse_snr_offset < 0) {
561 av_log(NULL, AV_LOG_ERROR, "Bit allocation failed. Try increasing the bitrate.\n");
562 return -1;
564 while ((coarse_snr_offset + SNR_INC1) <= 63 &&
565 bit_alloc(s, mask, psd, bap1, frame_bits,
566 coarse_snr_offset + SNR_INC1, 0) >= 0) {
567 coarse_snr_offset += SNR_INC1;
568 memcpy(bap, bap1, sizeof(bap1));
570 while ((coarse_snr_offset + 1) <= 63 &&
571 bit_alloc(s, mask, psd, bap1, frame_bits, coarse_snr_offset + 1, 0) >= 0) {
572 coarse_snr_offset++;
573 memcpy(bap, bap1, sizeof(bap1));
576 fine_snr_offset = 0;
577 while ((fine_snr_offset + SNR_INC1) <= 15 &&
578 bit_alloc(s, mask, psd, bap1, frame_bits,
579 coarse_snr_offset, fine_snr_offset + SNR_INC1) >= 0) {
580 fine_snr_offset += SNR_INC1;
581 memcpy(bap, bap1, sizeof(bap1));
583 while ((fine_snr_offset + 1) <= 15 &&
584 bit_alloc(s, mask, psd, bap1, frame_bits,
585 coarse_snr_offset, fine_snr_offset + 1) >= 0) {
586 fine_snr_offset++;
587 memcpy(bap, bap1, sizeof(bap1));
590 s->coarse_snr_offset = coarse_snr_offset;
591 for(ch=0;ch<s->nb_all_channels;ch++)
592 s->fine_snr_offset[ch] = fine_snr_offset;
593 #if defined(DEBUG_BITALLOC)
595 int j;
597 for(i=0;i<6;i++) {
598 for(ch=0;ch<s->nb_all_channels;ch++) {
599 printf("Block #%d Ch%d:\n", i, ch);
600 printf("bap=");
601 for(j=0;j<s->nb_coefs[ch];j++) {
602 printf("%d ",bap[i][ch][j]);
604 printf("\n");
608 #endif
609 return 0;
612 static av_cold int AC3_encode_init(AVCodecContext *avctx)
614 int freq = avctx->sample_rate;
615 int bitrate = avctx->bit_rate;
616 int channels = avctx->channels;
617 AC3EncodeContext *s = avctx->priv_data;
618 int i, j, ch;
619 float alpha;
620 int bw_code;
621 static const uint8_t channel_mode_defs[6] = {
622 0x01, /* C */
623 0x02, /* L R */
624 0x03, /* L C R */
625 0x06, /* L R SL SR */
626 0x07, /* L C R SL SR */
627 0x07, /* L C R SL SR (+LFE) */
630 avctx->frame_size = AC3_FRAME_SIZE;
632 ac3_common_init();
634 /* number of channels */
635 if (channels < 1 || channels > 6)
636 return -1;
637 s->channel_mode = channel_mode_defs[channels - 1];
638 s->lfe = (channels == 6) ? 1 : 0;
639 s->nb_all_channels = channels;
640 s->nb_channels = channels > 5 ? 5 : channels;
641 s->lfe_channel = s->lfe ? 5 : -1;
643 /* frequency */
644 for(i=0;i<3;i++) {
645 for(j=0;j<3;j++)
646 if ((ff_ac3_sample_rate_tab[j] >> i) == freq)
647 goto found;
649 return -1;
650 found:
651 s->sample_rate = freq;
652 s->sr_shift = i;
653 s->sr_code = j;
654 s->bitstream_id = 8 + s->sr_shift;
655 s->bitstream_mode = 0; /* complete main audio service */
657 /* bitrate & frame size */
658 for(i=0;i<19;i++) {
659 if ((ff_ac3_bitrate_tab[i] >> s->sr_shift)*1000 == bitrate)
660 break;
662 if (i == 19)
663 return -1;
664 s->bit_rate = bitrate;
665 s->frame_size_code = i << 1;
666 s->frame_size_min = ff_ac3_frame_size_tab[s->frame_size_code][s->sr_code];
667 s->bits_written = 0;
668 s->samples_written = 0;
669 s->frame_size = s->frame_size_min;
671 /* bit allocation init */
672 if(avctx->cutoff) {
673 /* calculate bandwidth based on user-specified cutoff frequency */
674 int cutoff = av_clip(avctx->cutoff, 1, s->sample_rate >> 1);
675 int fbw_coeffs = cutoff * 512 / s->sample_rate;
676 bw_code = av_clip((fbw_coeffs - 73) / 3, 0, 60);
677 } else {
678 /* use default bandwidth setting */
679 /* XXX: should compute the bandwidth according to the frame
680 size, so that we avoid annoying high frequency artifacts */
681 bw_code = 50;
683 for(ch=0;ch<s->nb_channels;ch++) {
684 /* bandwidth for each channel */
685 s->chbwcod[ch] = bw_code;
686 s->nb_coefs[ch] = bw_code * 3 + 73;
688 if (s->lfe) {
689 s->nb_coefs[s->lfe_channel] = 7; /* fixed */
691 /* initial snr offset */
692 s->coarse_snr_offset = 40;
694 /* mdct init */
695 fft_init(MDCT_NBITS - 2);
696 for(i=0;i<N/4;i++) {
697 alpha = 2 * M_PI * (i + 1.0 / 8.0) / (float)N;
698 xcos1[i] = fix15(-cos(alpha));
699 xsin1[i] = fix15(-sin(alpha));
702 avctx->coded_frame= avcodec_alloc_frame();
703 avctx->coded_frame->key_frame= 1;
705 return 0;
708 /* output the AC3 frame header */
709 static void output_frame_header(AC3EncodeContext *s, unsigned char *frame)
711 init_put_bits(&s->pb, frame, AC3_MAX_CODED_FRAME_SIZE);
713 put_bits(&s->pb, 16, 0x0b77); /* frame header */
714 put_bits(&s->pb, 16, 0); /* crc1: will be filled later */
715 put_bits(&s->pb, 2, s->sr_code);
716 put_bits(&s->pb, 6, s->frame_size_code + (s->frame_size - s->frame_size_min));
717 put_bits(&s->pb, 5, s->bitstream_id);
718 put_bits(&s->pb, 3, s->bitstream_mode);
719 put_bits(&s->pb, 3, s->channel_mode);
720 if ((s->channel_mode & 0x01) && s->channel_mode != AC3_CHMODE_MONO)
721 put_bits(&s->pb, 2, 1); /* XXX -4.5 dB */
722 if (s->channel_mode & 0x04)
723 put_bits(&s->pb, 2, 1); /* XXX -6 dB */
724 if (s->channel_mode == AC3_CHMODE_STEREO)
725 put_bits(&s->pb, 2, 0); /* surround not indicated */
726 put_bits(&s->pb, 1, s->lfe); /* LFE */
727 put_bits(&s->pb, 5, 31); /* dialog norm: -31 db */
728 put_bits(&s->pb, 1, 0); /* no compression control word */
729 put_bits(&s->pb, 1, 0); /* no lang code */
730 put_bits(&s->pb, 1, 0); /* no audio production info */
731 put_bits(&s->pb, 1, 0); /* no copyright */
732 put_bits(&s->pb, 1, 1); /* original bitstream */
733 put_bits(&s->pb, 1, 0); /* no time code 1 */
734 put_bits(&s->pb, 1, 0); /* no time code 2 */
735 put_bits(&s->pb, 1, 0); /* no additional bit stream info */
738 /* symetric quantization on 'levels' levels */
739 static inline int sym_quant(int c, int e, int levels)
741 int v;
743 if (c >= 0) {
744 v = (levels * (c << e)) >> 24;
745 v = (v + 1) >> 1;
746 v = (levels >> 1) + v;
747 } else {
748 v = (levels * ((-c) << e)) >> 24;
749 v = (v + 1) >> 1;
750 v = (levels >> 1) - v;
752 assert (v >= 0 && v < levels);
753 return v;
756 /* asymetric quantization on 2^qbits levels */
757 static inline int asym_quant(int c, int e, int qbits)
759 int lshift, m, v;
761 lshift = e + qbits - 24;
762 if (lshift >= 0)
763 v = c << lshift;
764 else
765 v = c >> (-lshift);
766 /* rounding */
767 v = (v + 1) >> 1;
768 m = (1 << (qbits-1));
769 if (v >= m)
770 v = m - 1;
771 assert(v >= -m);
772 return v & ((1 << qbits)-1);
775 /* Output one audio block. There are NB_BLOCKS audio blocks in one AC3
776 frame */
777 static void output_audio_block(AC3EncodeContext *s,
778 uint8_t exp_strategy[AC3_MAX_CHANNELS],
779 uint8_t encoded_exp[AC3_MAX_CHANNELS][N/2],
780 uint8_t bap[AC3_MAX_CHANNELS][N/2],
781 int32_t mdct_coefs[AC3_MAX_CHANNELS][N/2],
782 int8_t global_exp[AC3_MAX_CHANNELS],
783 int block_num)
785 int ch, nb_groups, group_size, i, baie, rbnd;
786 uint8_t *p;
787 uint16_t qmant[AC3_MAX_CHANNELS][N/2];
788 int exp0, exp1;
789 int mant1_cnt, mant2_cnt, mant4_cnt;
790 uint16_t *qmant1_ptr, *qmant2_ptr, *qmant4_ptr;
791 int delta0, delta1, delta2;
793 for(ch=0;ch<s->nb_channels;ch++)
794 put_bits(&s->pb, 1, 0); /* 512 point MDCT */
795 for(ch=0;ch<s->nb_channels;ch++)
796 put_bits(&s->pb, 1, 1); /* no dither */
797 put_bits(&s->pb, 1, 0); /* no dynamic range */
798 if (block_num == 0) {
799 /* for block 0, even if no coupling, we must say it. This is a
800 waste of bit :-) */
801 put_bits(&s->pb, 1, 1); /* coupling strategy present */
802 put_bits(&s->pb, 1, 0); /* no coupling strategy */
803 } else {
804 put_bits(&s->pb, 1, 0); /* no new coupling strategy */
807 if (s->channel_mode == AC3_CHMODE_STEREO)
809 if(block_num==0)
811 /* first block must define rematrixing (rematstr) */
812 put_bits(&s->pb, 1, 1);
814 /* dummy rematrixing rematflg(1:4)=0 */
815 for (rbnd=0;rbnd<4;rbnd++)
816 put_bits(&s->pb, 1, 0);
818 else
820 /* no matrixing (but should be used in the future) */
821 put_bits(&s->pb, 1, 0);
825 #if defined(DEBUG)
827 static int count = 0;
828 av_log(NULL, AV_LOG_DEBUG, "Block #%d (%d)\n", block_num, count++);
830 #endif
831 /* exponent strategy */
832 for(ch=0;ch<s->nb_channels;ch++) {
833 put_bits(&s->pb, 2, exp_strategy[ch]);
836 if (s->lfe) {
837 put_bits(&s->pb, 1, exp_strategy[s->lfe_channel]);
840 for(ch=0;ch<s->nb_channels;ch++) {
841 if (exp_strategy[ch] != EXP_REUSE)
842 put_bits(&s->pb, 6, s->chbwcod[ch]);
845 /* exponents */
846 for (ch = 0; ch < s->nb_all_channels; ch++) {
847 switch(exp_strategy[ch]) {
848 case EXP_REUSE:
849 continue;
850 case EXP_D15:
851 group_size = 1;
852 break;
853 case EXP_D25:
854 group_size = 2;
855 break;
856 default:
857 case EXP_D45:
858 group_size = 4;
859 break;
861 nb_groups = (s->nb_coefs[ch] + (group_size * 3) - 4) / (3 * group_size);
862 p = encoded_exp[ch];
864 /* first exponent */
865 exp1 = *p++;
866 put_bits(&s->pb, 4, exp1);
868 /* next ones are delta encoded */
869 for(i=0;i<nb_groups;i++) {
870 /* merge three delta in one code */
871 exp0 = exp1;
872 exp1 = p[0];
873 p += group_size;
874 delta0 = exp1 - exp0 + 2;
876 exp0 = exp1;
877 exp1 = p[0];
878 p += group_size;
879 delta1 = exp1 - exp0 + 2;
881 exp0 = exp1;
882 exp1 = p[0];
883 p += group_size;
884 delta2 = exp1 - exp0 + 2;
886 put_bits(&s->pb, 7, ((delta0 * 5 + delta1) * 5) + delta2);
889 if (ch != s->lfe_channel)
890 put_bits(&s->pb, 2, 0); /* no gain range info */
893 /* bit allocation info */
894 baie = (block_num == 0);
895 put_bits(&s->pb, 1, baie);
896 if (baie) {
897 put_bits(&s->pb, 2, s->slow_decay_code);
898 put_bits(&s->pb, 2, s->fast_decay_code);
899 put_bits(&s->pb, 2, s->slow_gain_code);
900 put_bits(&s->pb, 2, s->db_per_bit_code);
901 put_bits(&s->pb, 3, s->floor_code);
904 /* snr offset */
905 put_bits(&s->pb, 1, baie); /* always present with bai */
906 if (baie) {
907 put_bits(&s->pb, 6, s->coarse_snr_offset);
908 for(ch=0;ch<s->nb_all_channels;ch++) {
909 put_bits(&s->pb, 4, s->fine_snr_offset[ch]);
910 put_bits(&s->pb, 3, s->fast_gain_code[ch]);
914 put_bits(&s->pb, 1, 0); /* no delta bit allocation */
915 put_bits(&s->pb, 1, 0); /* no data to skip */
917 /* mantissa encoding : we use two passes to handle the grouping. A
918 one pass method may be faster, but it would necessitate to
919 modify the output stream. */
921 /* first pass: quantize */
922 mant1_cnt = mant2_cnt = mant4_cnt = 0;
923 qmant1_ptr = qmant2_ptr = qmant4_ptr = NULL;
925 for (ch = 0; ch < s->nb_all_channels; ch++) {
926 int b, c, e, v;
928 for(i=0;i<s->nb_coefs[ch];i++) {
929 c = mdct_coefs[ch][i];
930 e = encoded_exp[ch][i] - global_exp[ch];
931 b = bap[ch][i];
932 switch(b) {
933 case 0:
934 v = 0;
935 break;
936 case 1:
937 v = sym_quant(c, e, 3);
938 switch(mant1_cnt) {
939 case 0:
940 qmant1_ptr = &qmant[ch][i];
941 v = 9 * v;
942 mant1_cnt = 1;
943 break;
944 case 1:
945 *qmant1_ptr += 3 * v;
946 mant1_cnt = 2;
947 v = 128;
948 break;
949 default:
950 *qmant1_ptr += v;
951 mant1_cnt = 0;
952 v = 128;
953 break;
955 break;
956 case 2:
957 v = sym_quant(c, e, 5);
958 switch(mant2_cnt) {
959 case 0:
960 qmant2_ptr = &qmant[ch][i];
961 v = 25 * v;
962 mant2_cnt = 1;
963 break;
964 case 1:
965 *qmant2_ptr += 5 * v;
966 mant2_cnt = 2;
967 v = 128;
968 break;
969 default:
970 *qmant2_ptr += v;
971 mant2_cnt = 0;
972 v = 128;
973 break;
975 break;
976 case 3:
977 v = sym_quant(c, e, 7);
978 break;
979 case 4:
980 v = sym_quant(c, e, 11);
981 switch(mant4_cnt) {
982 case 0:
983 qmant4_ptr = &qmant[ch][i];
984 v = 11 * v;
985 mant4_cnt = 1;
986 break;
987 default:
988 *qmant4_ptr += v;
989 mant4_cnt = 0;
990 v = 128;
991 break;
993 break;
994 case 5:
995 v = sym_quant(c, e, 15);
996 break;
997 case 14:
998 v = asym_quant(c, e, 14);
999 break;
1000 case 15:
1001 v = asym_quant(c, e, 16);
1002 break;
1003 default:
1004 v = asym_quant(c, e, b - 1);
1005 break;
1007 qmant[ch][i] = v;
1011 /* second pass : output the values */
1012 for (ch = 0; ch < s->nb_all_channels; ch++) {
1013 int b, q;
1015 for(i=0;i<s->nb_coefs[ch];i++) {
1016 q = qmant[ch][i];
1017 b = bap[ch][i];
1018 switch(b) {
1019 case 0:
1020 break;
1021 case 1:
1022 if (q != 128)
1023 put_bits(&s->pb, 5, q);
1024 break;
1025 case 2:
1026 if (q != 128)
1027 put_bits(&s->pb, 7, q);
1028 break;
1029 case 3:
1030 put_bits(&s->pb, 3, q);
1031 break;
1032 case 4:
1033 if (q != 128)
1034 put_bits(&s->pb, 7, q);
1035 break;
1036 case 14:
1037 put_bits(&s->pb, 14, q);
1038 break;
1039 case 15:
1040 put_bits(&s->pb, 16, q);
1041 break;
1042 default:
1043 put_bits(&s->pb, b - 1, q);
1044 break;
1050 #define CRC16_POLY ((1 << 0) | (1 << 2) | (1 << 15) | (1 << 16))
1052 static unsigned int mul_poly(unsigned int a, unsigned int b, unsigned int poly)
1054 unsigned int c;
1056 c = 0;
1057 while (a) {
1058 if (a & 1)
1059 c ^= b;
1060 a = a >> 1;
1061 b = b << 1;
1062 if (b & (1 << 16))
1063 b ^= poly;
1065 return c;
1068 static unsigned int pow_poly(unsigned int a, unsigned int n, unsigned int poly)
1070 unsigned int r;
1071 r = 1;
1072 while (n) {
1073 if (n & 1)
1074 r = mul_poly(r, a, poly);
1075 a = mul_poly(a, a, poly);
1076 n >>= 1;
1078 return r;
1082 /* compute log2(max(abs(tab[]))) */
1083 static int log2_tab(int16_t *tab, int n)
1085 int i, v;
1087 v = 0;
1088 for(i=0;i<n;i++) {
1089 v |= abs(tab[i]);
1091 return av_log2(v);
1094 static void lshift_tab(int16_t *tab, int n, int lshift)
1096 int i;
1098 if (lshift > 0) {
1099 for(i=0;i<n;i++) {
1100 tab[i] <<= lshift;
1102 } else if (lshift < 0) {
1103 lshift = -lshift;
1104 for(i=0;i<n;i++) {
1105 tab[i] >>= lshift;
1110 /* fill the end of the frame and compute the two crcs */
1111 static int output_frame_end(AC3EncodeContext *s)
1113 int frame_size, frame_size_58, n, crc1, crc2, crc_inv;
1114 uint8_t *frame;
1116 frame_size = s->frame_size; /* frame size in words */
1117 /* align to 8 bits */
1118 flush_put_bits(&s->pb);
1119 /* add zero bytes to reach the frame size */
1120 frame = s->pb.buf;
1121 n = 2 * s->frame_size - (pbBufPtr(&s->pb) - frame) - 2;
1122 assert(n >= 0);
1123 if(n>0)
1124 memset(pbBufPtr(&s->pb), 0, n);
1126 /* Now we must compute both crcs : this is not so easy for crc1
1127 because it is at the beginning of the data... */
1128 frame_size_58 = (frame_size >> 1) + (frame_size >> 3);
1129 crc1 = bswap_16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1130 frame + 4, 2 * frame_size_58 - 4));
1131 /* XXX: could precompute crc_inv */
1132 crc_inv = pow_poly((CRC16_POLY >> 1), (16 * frame_size_58) - 16, CRC16_POLY);
1133 crc1 = mul_poly(crc_inv, crc1, CRC16_POLY);
1134 AV_WB16(frame+2,crc1);
1136 crc2 = bswap_16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1137 frame + 2 * frame_size_58,
1138 (frame_size - frame_size_58) * 2 - 2));
1139 AV_WB16(frame+2*frame_size-2,crc2);
1141 // printf("n=%d frame_size=%d\n", n, frame_size);
1142 return frame_size * 2;
1145 static int AC3_encode_frame(AVCodecContext *avctx,
1146 unsigned char *frame, int buf_size, void *data)
1148 AC3EncodeContext *s = avctx->priv_data;
1149 int16_t *samples = data;
1150 int i, j, k, v, ch;
1151 int16_t input_samples[N];
1152 int32_t mdct_coef[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1153 uint8_t exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1154 uint8_t exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS];
1155 uint8_t encoded_exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1156 uint8_t bap[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1157 int8_t exp_samples[NB_BLOCKS][AC3_MAX_CHANNELS];
1158 int frame_bits;
1160 frame_bits = 0;
1161 for(ch=0;ch<s->nb_all_channels;ch++) {
1162 /* fixed mdct to the six sub blocks & exponent computation */
1163 for(i=0;i<NB_BLOCKS;i++) {
1164 int16_t *sptr;
1165 int sinc;
1167 /* compute input samples */
1168 memcpy(input_samples, s->last_samples[ch], N/2 * sizeof(int16_t));
1169 sinc = s->nb_all_channels;
1170 sptr = samples + (sinc * (N/2) * i) + ch;
1171 for(j=0;j<N/2;j++) {
1172 v = *sptr;
1173 input_samples[j + N/2] = v;
1174 s->last_samples[ch][j] = v;
1175 sptr += sinc;
1178 /* apply the MDCT window */
1179 for(j=0;j<N/2;j++) {
1180 input_samples[j] = MUL16(input_samples[j],
1181 ff_ac3_window[j]) >> 15;
1182 input_samples[N-j-1] = MUL16(input_samples[N-j-1],
1183 ff_ac3_window[j]) >> 15;
1186 /* Normalize the samples to use the maximum available
1187 precision */
1188 v = 14 - log2_tab(input_samples, N);
1189 if (v < 0)
1190 v = 0;
1191 exp_samples[i][ch] = v - 9;
1192 lshift_tab(input_samples, N, v);
1194 /* do the MDCT */
1195 mdct512(mdct_coef[i][ch], input_samples);
1197 /* compute "exponents". We take into account the
1198 normalization there */
1199 for(j=0;j<N/2;j++) {
1200 int e;
1201 v = abs(mdct_coef[i][ch][j]);
1202 if (v == 0)
1203 e = 24;
1204 else {
1205 e = 23 - av_log2(v) + exp_samples[i][ch];
1206 if (e >= 24) {
1207 e = 24;
1208 mdct_coef[i][ch][j] = 0;
1211 exp[i][ch][j] = e;
1215 compute_exp_strategy(exp_strategy, exp, ch, ch == s->lfe_channel);
1217 /* compute the exponents as the decoder will see them. The
1218 EXP_REUSE case must be handled carefully : we select the
1219 min of the exponents */
1220 i = 0;
1221 while (i < NB_BLOCKS) {
1222 j = i + 1;
1223 while (j < NB_BLOCKS && exp_strategy[j][ch] == EXP_REUSE) {
1224 exponent_min(exp[i][ch], exp[j][ch], s->nb_coefs[ch]);
1225 j++;
1227 frame_bits += encode_exp(encoded_exp[i][ch],
1228 exp[i][ch], s->nb_coefs[ch],
1229 exp_strategy[i][ch]);
1230 /* copy encoded exponents for reuse case */
1231 for(k=i+1;k<j;k++) {
1232 memcpy(encoded_exp[k][ch], encoded_exp[i][ch],
1233 s->nb_coefs[ch] * sizeof(uint8_t));
1235 i = j;
1239 /* adjust for fractional frame sizes */
1240 while(s->bits_written >= s->bit_rate && s->samples_written >= s->sample_rate) {
1241 s->bits_written -= s->bit_rate;
1242 s->samples_written -= s->sample_rate;
1244 s->frame_size = s->frame_size_min + (s->bits_written * s->sample_rate < s->samples_written * s->bit_rate);
1245 s->bits_written += s->frame_size * 16;
1246 s->samples_written += AC3_FRAME_SIZE;
1248 compute_bit_allocation(s, bap, encoded_exp, exp_strategy, frame_bits);
1249 /* everything is known... let's output the frame */
1250 output_frame_header(s, frame);
1252 for(i=0;i<NB_BLOCKS;i++) {
1253 output_audio_block(s, exp_strategy[i], encoded_exp[i],
1254 bap[i], mdct_coef[i], exp_samples[i], i);
1256 return output_frame_end(s);
1259 static av_cold int AC3_encode_close(AVCodecContext *avctx)
1261 av_freep(&avctx->coded_frame);
1262 return 0;
1265 #if 0
1266 /*************************************************************************/
1267 /* TEST */
1269 #undef random
1270 #define FN (N/4)
1272 void fft_test(void)
1274 IComplex in[FN], in1[FN];
1275 int k, n, i;
1276 float sum_re, sum_im, a;
1278 /* FFT test */
1280 for(i=0;i<FN;i++) {
1281 in[i].re = random() % 65535 - 32767;
1282 in[i].im = random() % 65535 - 32767;
1283 in1[i] = in[i];
1285 fft(in, 7);
1287 /* do it by hand */
1288 for(k=0;k<FN;k++) {
1289 sum_re = 0;
1290 sum_im = 0;
1291 for(n=0;n<FN;n++) {
1292 a = -2 * M_PI * (n * k) / FN;
1293 sum_re += in1[n].re * cos(a) - in1[n].im * sin(a);
1294 sum_im += in1[n].re * sin(a) + in1[n].im * cos(a);
1296 printf("%3d: %6d,%6d %6.0f,%6.0f\n",
1297 k, in[k].re, in[k].im, sum_re / FN, sum_im / FN);
1301 void mdct_test(void)
1303 int16_t input[N];
1304 int32_t output[N/2];
1305 float input1[N];
1306 float output1[N/2];
1307 float s, a, err, e, emax;
1308 int i, k, n;
1310 for(i=0;i<N;i++) {
1311 input[i] = (random() % 65535 - 32767) * 9 / 10;
1312 input1[i] = input[i];
1315 mdct512(output, input);
1317 /* do it by hand */
1318 for(k=0;k<N/2;k++) {
1319 s = 0;
1320 for(n=0;n<N;n++) {
1321 a = (2*M_PI*(2*n+1+N/2)*(2*k+1) / (4 * N));
1322 s += input1[n] * cos(a);
1324 output1[k] = -2 * s / N;
1327 err = 0;
1328 emax = 0;
1329 for(i=0;i<N/2;i++) {
1330 printf("%3d: %7d %7.0f\n", i, output[i], output1[i]);
1331 e = output[i] - output1[i];
1332 if (e > emax)
1333 emax = e;
1334 err += e * e;
1336 printf("err2=%f emax=%f\n", err / (N/2), emax);
1339 void test_ac3(void)
1341 AC3EncodeContext ctx;
1342 unsigned char frame[AC3_MAX_CODED_FRAME_SIZE];
1343 short samples[AC3_FRAME_SIZE];
1344 int ret, i;
1346 AC3_encode_init(&ctx, 44100, 64000, 1);
1348 fft_test();
1349 mdct_test();
1351 for(i=0;i<AC3_FRAME_SIZE;i++)
1352 samples[i] = (int)(sin(2*M_PI*i*1000.0/44100) * 10000);
1353 ret = AC3_encode_frame(&ctx, frame, samples);
1354 printf("ret=%d\n", ret);
1356 #endif
1358 AVCodec ac3_encoder = {
1359 "ac3",
1360 CODEC_TYPE_AUDIO,
1361 CODEC_ID_AC3,
1362 sizeof(AC3EncodeContext),
1363 AC3_encode_init,
1364 AC3_encode_frame,
1365 AC3_encode_close,
1366 NULL,
1367 .long_name = NULL_IF_CONFIG_SMALL("ATSC A/52 / AC-3"),