2 * The simplest AC-3 encoder
3 * Copyright (c) 2000 Fabrice Bellard
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 * @file libavcodec/ac3enc.c
24 * The simplest AC-3 encoder.
27 //#define DEBUG_BITALLOC
28 #include "libavutil/crc.h"
30 #include "get_bits.h" // for ff_reverse
33 #include "audioconvert.h"
35 typedef struct AC3EncodeContext
{
40 const uint8_t *channel_map
;
42 unsigned int sample_rate
;
43 unsigned int bitstream_id
;
44 unsigned int frame_size_min
; /* minimum frame size in case rounding is necessary */
45 unsigned int frame_size
; /* current frame size in words */
46 unsigned int bits_written
;
47 unsigned int samples_written
;
49 unsigned int frame_size_code
;
50 unsigned int sr_code
; /* frequency */
51 unsigned int channel_mode
;
53 unsigned int bitstream_mode
;
54 short last_samples
[AC3_MAX_CHANNELS
][256];
55 unsigned int chbwcod
[AC3_MAX_CHANNELS
];
56 int nb_coefs
[AC3_MAX_CHANNELS
];
58 /* bitrate allocation control */
59 int slow_gain_code
, slow_decay_code
, fast_decay_code
, db_per_bit_code
, floor_code
;
60 AC3BitAllocParameters bit_alloc
;
61 int coarse_snr_offset
;
62 int fast_gain_code
[AC3_MAX_CHANNELS
];
63 int fine_snr_offset
[AC3_MAX_CHANNELS
];
64 /* mantissa encoding */
65 int mant1_cnt
, mant2_cnt
, mant4_cnt
;
68 static int16_t costab
[64];
69 static int16_t sintab
[64];
70 static int16_t xcos1
[128];
71 static int16_t xsin1
[128];
74 #define N (1 << MDCT_NBITS)
76 /* new exponents are sent if their Norm 1 exceed this number */
77 #define EXP_DIFF_THRESHOLD 1000
79 static inline int16_t fix15(float a
)
82 v
= (int)(a
* (float)(1 << 15));
90 typedef struct IComplex
{
94 static av_cold
void fft_init(int ln
)
101 for(i
=0;i
<(n
/2);i
++) {
102 alpha
= 2 * M_PI
* (float)i
/ (float)n
;
103 costab
[i
] = fix15(cos(alpha
));
104 sintab
[i
] = fix15(sin(alpha
));
109 #define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \
116 pre = (bx + ax) >> 1;\
117 pim = (by + ay) >> 1;\
118 qre = (bx - ax) >> 1;\
119 qim = (by - ay) >> 1;\
122 #define CMUL(pre, pim, are, aim, bre, bim) \
124 pre = (MUL16(are, bre) - MUL16(aim, bim)) >> 15;\
125 pim = (MUL16(are, bim) + MUL16(bre, aim)) >> 15;\
129 /* do a 2^n point complex fft on 2^ln points. */
130 static void fft(IComplex
*z
, int ln
)
134 register IComplex
*p
,*q
;
141 int k
= ff_reverse
[j
] >> (8 - ln
);
143 FFSWAP(IComplex
, z
[k
], z
[j
]);
151 BF(p
[0].re
, p
[0].im
, p
[1].re
, p
[1].im
,
152 p
[0].re
, p
[0].im
, p
[1].re
, p
[1].im
);
161 BF(p
[0].re
, p
[0].im
, p
[2].re
, p
[2].im
,
162 p
[0].re
, p
[0].im
, p
[2].re
, p
[2].im
);
163 BF(p
[1].re
, p
[1].im
, p
[3].re
, p
[3].im
,
164 p
[1].re
, p
[1].im
, p
[3].im
, -p
[3].re
);
176 for (j
= 0; j
< nblocks
; ++j
) {
178 BF(p
->re
, p
->im
, q
->re
, q
->im
,
179 p
->re
, p
->im
, q
->re
, q
->im
);
183 for(l
= nblocks
; l
< np2
; l
+= nblocks
) {
184 CMUL(tmp_re
, tmp_im
, costab
[l
], -sintab
[l
], q
->re
, q
->im
);
185 BF(p
->re
, p
->im
, q
->re
, q
->im
,
186 p
->re
, p
->im
, tmp_re
, tmp_im
);
193 nblocks
= nblocks
>> 1;
194 nloops
= nloops
<< 1;
195 } while (nblocks
!= 0);
198 /* do a 512 point mdct */
199 static void mdct512(int32_t *out
, int16_t *in
)
201 int i
, re
, im
, re1
, im1
;
205 /* shift to simplify computations */
207 rot
[i
] = -in
[i
+ 3*N
/4];
209 rot
[i
] = in
[i
- N
/4];
213 re
= ((int)rot
[2*i
] - (int)rot
[N
-1-2*i
]) >> 1;
214 im
= -((int)rot
[N
/2+2*i
] - (int)rot
[N
/2-1-2*i
]) >> 1;
215 CMUL(x
[i
].re
, x
[i
].im
, re
, im
, -xcos1
[i
], xsin1
[i
]);
218 fft(x
, MDCT_NBITS
- 2);
224 CMUL(re1
, im1
, re
, im
, xsin1
[i
], xcos1
[i
]);
226 out
[N
/2-1-2*i
] = re1
;
230 /* XXX: use another norm ? */
231 static int calc_exp_diff(uint8_t *exp1
, uint8_t *exp2
, int n
)
236 sum
+= abs(exp1
[i
] - exp2
[i
]);
241 static void compute_exp_strategy(uint8_t exp_strategy
[NB_BLOCKS
][AC3_MAX_CHANNELS
],
242 uint8_t exp
[NB_BLOCKS
][AC3_MAX_CHANNELS
][N
/2],
248 /* estimate if the exponent variation & decide if they should be
249 reused in the next frame */
250 exp_strategy
[0][ch
] = EXP_NEW
;
251 for(i
=1;i
<NB_BLOCKS
;i
++) {
252 exp_diff
= calc_exp_diff(exp
[i
][ch
], exp
[i
-1][ch
], N
/2);
254 av_log(NULL
, AV_LOG_DEBUG
, "exp_diff=%d\n", exp_diff
);
256 if (exp_diff
> EXP_DIFF_THRESHOLD
)
257 exp_strategy
[i
][ch
] = EXP_NEW
;
259 exp_strategy
[i
][ch
] = EXP_REUSE
;
264 /* now select the encoding strategy type : if exponents are often
265 recoded, we use a coarse encoding */
267 while (i
< NB_BLOCKS
) {
269 while (j
< NB_BLOCKS
&& exp_strategy
[j
][ch
] == EXP_REUSE
)
273 exp_strategy
[i
][ch
] = EXP_D45
;
277 exp_strategy
[i
][ch
] = EXP_D25
;
280 exp_strategy
[i
][ch
] = EXP_D15
;
287 /* set exp[i] to min(exp[i], exp1[i]) */
288 static void exponent_min(uint8_t exp
[N
/2], uint8_t exp1
[N
/2], int n
)
293 if (exp1
[i
] < exp
[i
])
298 /* update the exponents so that they are the ones the decoder will
299 decode. Return the number of bits used to code the exponents */
300 static int encode_exp(uint8_t encoded_exp
[N
/2],
305 int group_size
, nb_groups
, i
, j
, k
, exp_min
;
308 switch(exp_strategy
) {
320 nb_groups
= ((nb_exps
+ (group_size
* 3) - 4) / (3 * group_size
)) * 3;
322 /* for each group, compute the minimum exponent */
323 exp1
[0] = exp
[0]; /* DC exponent is handled separately */
325 for(i
=1;i
<=nb_groups
;i
++) {
327 assert(exp_min
>= 0 && exp_min
<= 24);
328 for(j
=1;j
<group_size
;j
++) {
329 if (exp
[k
+j
] < exp_min
)
336 /* constraint for DC exponent */
340 /* Decrease the delta between each groups to within 2
341 * so that they can be differentially encoded */
342 for (i
=1;i
<=nb_groups
;i
++)
343 exp1
[i
] = FFMIN(exp1
[i
], exp1
[i
-1] + 2);
344 for (i
=nb_groups
-1;i
>=0;i
--)
345 exp1
[i
] = FFMIN(exp1
[i
], exp1
[i
+1] + 2);
347 /* now we have the exponent values the decoder will see */
348 encoded_exp
[0] = exp1
[0];
350 for(i
=1;i
<=nb_groups
;i
++) {
351 for(j
=0;j
<group_size
;j
++) {
352 encoded_exp
[k
+j
] = exp1
[i
];
358 av_log(NULL
, AV_LOG_DEBUG
, "exponents: strategy=%d\n", exp_strategy
);
359 for(i
=0;i
<=nb_groups
* group_size
;i
++) {
360 av_log(NULL
, AV_LOG_DEBUG
, "%d ", encoded_exp
[i
]);
362 av_log(NULL
, AV_LOG_DEBUG
, "\n");
365 return 4 + (nb_groups
/ 3) * 7;
368 /* return the size in bits taken by the mantissa */
369 static int compute_mantissa_size(AC3EncodeContext
*s
, uint8_t *m
, int nb_coefs
)
374 for(i
=0;i
<nb_coefs
;i
++) {
381 /* 3 mantissa in 5 bits */
382 if (s
->mant1_cnt
== 0)
384 if (++s
->mant1_cnt
== 3)
388 /* 3 mantissa in 7 bits */
389 if (s
->mant2_cnt
== 0)
391 if (++s
->mant2_cnt
== 3)
398 /* 2 mantissa in 7 bits */
399 if (s
->mant4_cnt
== 0)
401 if (++s
->mant4_cnt
== 2)
419 static void bit_alloc_masking(AC3EncodeContext
*s
,
420 uint8_t encoded_exp
[NB_BLOCKS
][AC3_MAX_CHANNELS
][N
/2],
421 uint8_t exp_strategy
[NB_BLOCKS
][AC3_MAX_CHANNELS
],
422 int16_t psd
[NB_BLOCKS
][AC3_MAX_CHANNELS
][N
/2],
423 int16_t mask
[NB_BLOCKS
][AC3_MAX_CHANNELS
][50])
426 int16_t band_psd
[NB_BLOCKS
][AC3_MAX_CHANNELS
][50];
428 for(blk
=0; blk
<NB_BLOCKS
; blk
++) {
429 for(ch
=0;ch
<s
->nb_all_channels
;ch
++) {
430 if(exp_strategy
[blk
][ch
] == EXP_REUSE
) {
431 memcpy(psd
[blk
][ch
], psd
[blk
-1][ch
], (N
/2)*sizeof(int16_t));
432 memcpy(mask
[blk
][ch
], mask
[blk
-1][ch
], 50*sizeof(int16_t));
434 ff_ac3_bit_alloc_calc_psd(encoded_exp
[blk
][ch
], 0,
436 psd
[blk
][ch
], band_psd
[blk
][ch
]);
437 ff_ac3_bit_alloc_calc_mask(&s
->bit_alloc
, band_psd
[blk
][ch
],
439 ff_ac3_fast_gain_tab
[s
->fast_gain_code
[ch
]],
440 ch
== s
->lfe_channel
,
441 DBA_NONE
, 0, NULL
, NULL
, NULL
,
448 static int bit_alloc(AC3EncodeContext
*s
,
449 int16_t mask
[NB_BLOCKS
][AC3_MAX_CHANNELS
][50],
450 int16_t psd
[NB_BLOCKS
][AC3_MAX_CHANNELS
][N
/2],
451 uint8_t bap
[NB_BLOCKS
][AC3_MAX_CHANNELS
][N
/2],
452 int frame_bits
, int coarse_snr_offset
, int fine_snr_offset
)
457 snr_offset
= (((coarse_snr_offset
- 15) << 4) + fine_snr_offset
) << 2;
460 for(i
=0;i
<NB_BLOCKS
;i
++) {
464 for(ch
=0;ch
<s
->nb_all_channels
;ch
++) {
465 ff_ac3_bit_alloc_calc_bap(mask
[i
][ch
], psd
[i
][ch
], 0,
466 s
->nb_coefs
[ch
], snr_offset
,
467 s
->bit_alloc
.floor
, ff_ac3_bap_tab
,
469 frame_bits
+= compute_mantissa_size(s
, bap
[i
][ch
],
474 printf("csnr=%d fsnr=%d frame_bits=%d diff=%d\n",
475 coarse_snr_offset
, fine_snr_offset
, frame_bits
,
476 16 * s
->frame_size
- ((frame_bits
+ 7) & ~7));
478 return 16 * s
->frame_size
- frame_bits
;
483 static int compute_bit_allocation(AC3EncodeContext
*s
,
484 uint8_t bap
[NB_BLOCKS
][AC3_MAX_CHANNELS
][N
/2],
485 uint8_t encoded_exp
[NB_BLOCKS
][AC3_MAX_CHANNELS
][N
/2],
486 uint8_t exp_strategy
[NB_BLOCKS
][AC3_MAX_CHANNELS
],
490 int coarse_snr_offset
, fine_snr_offset
;
491 uint8_t bap1
[NB_BLOCKS
][AC3_MAX_CHANNELS
][N
/2];
492 int16_t psd
[NB_BLOCKS
][AC3_MAX_CHANNELS
][N
/2];
493 int16_t mask
[NB_BLOCKS
][AC3_MAX_CHANNELS
][50];
494 static const int frame_bits_inc
[8] = { 0, 0, 2, 2, 2, 4, 2, 4 };
496 /* init default parameters */
497 s
->slow_decay_code
= 2;
498 s
->fast_decay_code
= 1;
499 s
->slow_gain_code
= 1;
500 s
->db_per_bit_code
= 2;
502 for(ch
=0;ch
<s
->nb_all_channels
;ch
++)
503 s
->fast_gain_code
[ch
] = 4;
505 /* compute real values */
506 s
->bit_alloc
.sr_code
= s
->sr_code
;
507 s
->bit_alloc
.sr_shift
= s
->sr_shift
;
508 s
->bit_alloc
.slow_decay
= ff_ac3_slow_decay_tab
[s
->slow_decay_code
] >> s
->sr_shift
;
509 s
->bit_alloc
.fast_decay
= ff_ac3_fast_decay_tab
[s
->fast_decay_code
] >> s
->sr_shift
;
510 s
->bit_alloc
.slow_gain
= ff_ac3_slow_gain_tab
[s
->slow_gain_code
];
511 s
->bit_alloc
.db_per_bit
= ff_ac3_db_per_bit_tab
[s
->db_per_bit_code
];
512 s
->bit_alloc
.floor
= ff_ac3_floor_tab
[s
->floor_code
];
516 // if (s->channel_mode == 2)
518 frame_bits
+= frame_bits_inc
[s
->channel_mode
];
521 for(i
=0;i
<NB_BLOCKS
;i
++) {
522 frame_bits
+= s
->nb_channels
* 2 + 2; /* blksw * c, dithflag * c, dynrnge, cplstre */
523 if (s
->channel_mode
== AC3_CHMODE_STEREO
) {
524 frame_bits
++; /* rematstr */
525 if(i
==0) frame_bits
+= 4;
527 frame_bits
+= 2 * s
->nb_channels
; /* chexpstr[2] * c */
529 frame_bits
++; /* lfeexpstr */
530 for(ch
=0;ch
<s
->nb_channels
;ch
++) {
531 if (exp_strategy
[i
][ch
] != EXP_REUSE
)
532 frame_bits
+= 6 + 2; /* chbwcod[6], gainrng[2] */
534 frame_bits
++; /* baie */
535 frame_bits
++; /* snr */
536 frame_bits
+= 2; /* delta / skip */
538 frame_bits
++; /* cplinu for block 0 */
540 /* sdcycod[2], fdcycod[2], sgaincod[2], dbpbcod[2], floorcod[3] */
542 /* (fsnoffset[4] + fgaincod[4]) * c */
543 frame_bits
+= 2*4 + 3 + 6 + s
->nb_all_channels
* (4 + 3);
545 /* auxdatae, crcrsv */
551 /* calculate psd and masking curve before doing bit allocation */
552 bit_alloc_masking(s
, encoded_exp
, exp_strategy
, psd
, mask
);
554 /* now the big work begins : do the bit allocation. Modify the snr
555 offset until we can pack everything in the requested frame size */
557 coarse_snr_offset
= s
->coarse_snr_offset
;
558 while (coarse_snr_offset
>= 0 &&
559 bit_alloc(s
, mask
, psd
, bap
, frame_bits
, coarse_snr_offset
, 0) < 0)
560 coarse_snr_offset
-= SNR_INC1
;
561 if (coarse_snr_offset
< 0) {
562 av_log(NULL
, AV_LOG_ERROR
, "Bit allocation failed. Try increasing the bitrate.\n");
565 while ((coarse_snr_offset
+ SNR_INC1
) <= 63 &&
566 bit_alloc(s
, mask
, psd
, bap1
, frame_bits
,
567 coarse_snr_offset
+ SNR_INC1
, 0) >= 0) {
568 coarse_snr_offset
+= SNR_INC1
;
569 memcpy(bap
, bap1
, sizeof(bap1
));
571 while ((coarse_snr_offset
+ 1) <= 63 &&
572 bit_alloc(s
, mask
, psd
, bap1
, frame_bits
, coarse_snr_offset
+ 1, 0) >= 0) {
574 memcpy(bap
, bap1
, sizeof(bap1
));
578 while ((fine_snr_offset
+ SNR_INC1
) <= 15 &&
579 bit_alloc(s
, mask
, psd
, bap1
, frame_bits
,
580 coarse_snr_offset
, fine_snr_offset
+ SNR_INC1
) >= 0) {
581 fine_snr_offset
+= SNR_INC1
;
582 memcpy(bap
, bap1
, sizeof(bap1
));
584 while ((fine_snr_offset
+ 1) <= 15 &&
585 bit_alloc(s
, mask
, psd
, bap1
, frame_bits
,
586 coarse_snr_offset
, fine_snr_offset
+ 1) >= 0) {
588 memcpy(bap
, bap1
, sizeof(bap1
));
591 s
->coarse_snr_offset
= coarse_snr_offset
;
592 for(ch
=0;ch
<s
->nb_all_channels
;ch
++)
593 s
->fine_snr_offset
[ch
] = fine_snr_offset
;
594 #if defined(DEBUG_BITALLOC)
599 for(ch
=0;ch
<s
->nb_all_channels
;ch
++) {
600 printf("Block #%d Ch%d:\n", i
, ch
);
602 for(j
=0;j
<s
->nb_coefs
[ch
];j
++) {
603 printf("%d ",bap
[i
][ch
][j
]);
613 static av_cold
int set_channel_info(AC3EncodeContext
*s
, int channels
,
614 int64_t *channel_layout
)
618 if (channels
< 1 || channels
> AC3_MAX_CHANNELS
)
620 if ((uint64_t)*channel_layout
> 0x7FF)
622 ch_layout
= *channel_layout
;
624 ch_layout
= avcodec_guess_channel_layout(channels
, CODEC_ID_AC3
, NULL
);
625 if (avcodec_channel_layout_num_channels(ch_layout
) != channels
)
628 s
->lfe
= !!(ch_layout
& CH_LOW_FREQUENCY
);
629 s
->nb_all_channels
= channels
;
630 s
->nb_channels
= channels
- s
->lfe
;
631 s
->lfe_channel
= s
->lfe
? s
->nb_channels
: -1;
633 ch_layout
-= CH_LOW_FREQUENCY
;
636 case CH_LAYOUT_MONO
: s
->channel_mode
= AC3_CHMODE_MONO
; break;
637 case CH_LAYOUT_STEREO
: s
->channel_mode
= AC3_CHMODE_STEREO
; break;
638 case CH_LAYOUT_SURROUND
: s
->channel_mode
= AC3_CHMODE_3F
; break;
639 case CH_LAYOUT_2_1
: s
->channel_mode
= AC3_CHMODE_2F1R
; break;
640 case CH_LAYOUT_4POINT0
: s
->channel_mode
= AC3_CHMODE_3F1R
; break;
642 case CH_LAYOUT_2_2
: s
->channel_mode
= AC3_CHMODE_2F2R
; break;
643 case CH_LAYOUT_5POINT0
:
644 case CH_LAYOUT_5POINT0_BACK
: s
->channel_mode
= AC3_CHMODE_3F2R
; break;
649 s
->channel_map
= ff_ac3_enc_channel_map
[s
->channel_mode
][s
->lfe
];
650 *channel_layout
= ch_layout
;
652 *channel_layout
|= CH_LOW_FREQUENCY
;
657 static av_cold
int AC3_encode_init(AVCodecContext
*avctx
)
659 int freq
= avctx
->sample_rate
;
660 int bitrate
= avctx
->bit_rate
;
661 AC3EncodeContext
*s
= avctx
->priv_data
;
666 avctx
->frame_size
= AC3_FRAME_SIZE
;
670 if (!avctx
->channel_layout
) {
671 av_log(avctx
, AV_LOG_WARNING
, "No channel layout specified. The "
672 "encoder will guess the layout, but it "
673 "might be incorrect.\n");
675 if (set_channel_info(s
, avctx
->channels
, &avctx
->channel_layout
)) {
676 av_log(avctx
, AV_LOG_ERROR
, "invalid channel layout\n");
683 if ((ff_ac3_sample_rate_tab
[j
] >> i
) == freq
)
688 s
->sample_rate
= freq
;
691 s
->bitstream_id
= 8 + s
->sr_shift
;
692 s
->bitstream_mode
= 0; /* complete main audio service */
694 /* bitrate & frame size */
696 if ((ff_ac3_bitrate_tab
[i
] >> s
->sr_shift
)*1000 == bitrate
)
701 s
->bit_rate
= bitrate
;
702 s
->frame_size_code
= i
<< 1;
703 s
->frame_size_min
= ff_ac3_frame_size_tab
[s
->frame_size_code
][s
->sr_code
];
705 s
->samples_written
= 0;
706 s
->frame_size
= s
->frame_size_min
;
708 /* bit allocation init */
710 /* calculate bandwidth based on user-specified cutoff frequency */
711 int cutoff
= av_clip(avctx
->cutoff
, 1, s
->sample_rate
>> 1);
712 int fbw_coeffs
= cutoff
* 512 / s
->sample_rate
;
713 bw_code
= av_clip((fbw_coeffs
- 73) / 3, 0, 60);
715 /* use default bandwidth setting */
716 /* XXX: should compute the bandwidth according to the frame
717 size, so that we avoid annoying high frequency artifacts */
720 for(ch
=0;ch
<s
->nb_channels
;ch
++) {
721 /* bandwidth for each channel */
722 s
->chbwcod
[ch
] = bw_code
;
723 s
->nb_coefs
[ch
] = bw_code
* 3 + 73;
726 s
->nb_coefs
[s
->lfe_channel
] = 7; /* fixed */
728 /* initial snr offset */
729 s
->coarse_snr_offset
= 40;
732 fft_init(MDCT_NBITS
- 2);
734 alpha
= 2 * M_PI
* (i
+ 1.0 / 8.0) / (float)N
;
735 xcos1
[i
] = fix15(-cos(alpha
));
736 xsin1
[i
] = fix15(-sin(alpha
));
739 avctx
->coded_frame
= avcodec_alloc_frame();
740 avctx
->coded_frame
->key_frame
= 1;
745 /* output the AC-3 frame header */
746 static void output_frame_header(AC3EncodeContext
*s
, unsigned char *frame
)
748 init_put_bits(&s
->pb
, frame
, AC3_MAX_CODED_FRAME_SIZE
);
750 put_bits(&s
->pb
, 16, 0x0b77); /* frame header */
751 put_bits(&s
->pb
, 16, 0); /* crc1: will be filled later */
752 put_bits(&s
->pb
, 2, s
->sr_code
);
753 put_bits(&s
->pb
, 6, s
->frame_size_code
+ (s
->frame_size
- s
->frame_size_min
));
754 put_bits(&s
->pb
, 5, s
->bitstream_id
);
755 put_bits(&s
->pb
, 3, s
->bitstream_mode
);
756 put_bits(&s
->pb
, 3, s
->channel_mode
);
757 if ((s
->channel_mode
& 0x01) && s
->channel_mode
!= AC3_CHMODE_MONO
)
758 put_bits(&s
->pb
, 2, 1); /* XXX -4.5 dB */
759 if (s
->channel_mode
& 0x04)
760 put_bits(&s
->pb
, 2, 1); /* XXX -6 dB */
761 if (s
->channel_mode
== AC3_CHMODE_STEREO
)
762 put_bits(&s
->pb
, 2, 0); /* surround not indicated */
763 put_bits(&s
->pb
, 1, s
->lfe
); /* LFE */
764 put_bits(&s
->pb
, 5, 31); /* dialog norm: -31 db */
765 put_bits(&s
->pb
, 1, 0); /* no compression control word */
766 put_bits(&s
->pb
, 1, 0); /* no lang code */
767 put_bits(&s
->pb
, 1, 0); /* no audio production info */
768 put_bits(&s
->pb
, 1, 0); /* no copyright */
769 put_bits(&s
->pb
, 1, 1); /* original bitstream */
770 put_bits(&s
->pb
, 1, 0); /* no time code 1 */
771 put_bits(&s
->pb
, 1, 0); /* no time code 2 */
772 put_bits(&s
->pb
, 1, 0); /* no additional bit stream info */
775 /* symetric quantization on 'levels' levels */
776 static inline int sym_quant(int c
, int e
, int levels
)
781 v
= (levels
* (c
<< e
)) >> 24;
783 v
= (levels
>> 1) + v
;
785 v
= (levels
* ((-c
) << e
)) >> 24;
787 v
= (levels
>> 1) - v
;
789 assert (v
>= 0 && v
< levels
);
793 /* asymetric quantization on 2^qbits levels */
794 static inline int asym_quant(int c
, int e
, int qbits
)
798 lshift
= e
+ qbits
- 24;
805 m
= (1 << (qbits
-1));
809 return v
& ((1 << qbits
)-1);
812 /* Output one audio block. There are NB_BLOCKS audio blocks in one AC-3
814 static void output_audio_block(AC3EncodeContext
*s
,
815 uint8_t exp_strategy
[AC3_MAX_CHANNELS
],
816 uint8_t encoded_exp
[AC3_MAX_CHANNELS
][N
/2],
817 uint8_t bap
[AC3_MAX_CHANNELS
][N
/2],
818 int32_t mdct_coefs
[AC3_MAX_CHANNELS
][N
/2],
819 int8_t global_exp
[AC3_MAX_CHANNELS
],
822 int ch
, nb_groups
, group_size
, i
, baie
, rbnd
;
824 uint16_t qmant
[AC3_MAX_CHANNELS
][N
/2];
826 int mant1_cnt
, mant2_cnt
, mant4_cnt
;
827 uint16_t *qmant1_ptr
, *qmant2_ptr
, *qmant4_ptr
;
828 int delta0
, delta1
, delta2
;
830 for(ch
=0;ch
<s
->nb_channels
;ch
++)
831 put_bits(&s
->pb
, 1, 0); /* 512 point MDCT */
832 for(ch
=0;ch
<s
->nb_channels
;ch
++)
833 put_bits(&s
->pb
, 1, 1); /* no dither */
834 put_bits(&s
->pb
, 1, 0); /* no dynamic range */
835 if (block_num
== 0) {
836 /* for block 0, even if no coupling, we must say it. This is a
838 put_bits(&s
->pb
, 1, 1); /* coupling strategy present */
839 put_bits(&s
->pb
, 1, 0); /* no coupling strategy */
841 put_bits(&s
->pb
, 1, 0); /* no new coupling strategy */
844 if (s
->channel_mode
== AC3_CHMODE_STEREO
)
848 /* first block must define rematrixing (rematstr) */
849 put_bits(&s
->pb
, 1, 1);
851 /* dummy rematrixing rematflg(1:4)=0 */
852 for (rbnd
=0;rbnd
<4;rbnd
++)
853 put_bits(&s
->pb
, 1, 0);
857 /* no matrixing (but should be used in the future) */
858 put_bits(&s
->pb
, 1, 0);
864 static int count
= 0;
865 av_log(NULL
, AV_LOG_DEBUG
, "Block #%d (%d)\n", block_num
, count
++);
868 /* exponent strategy */
869 for(ch
=0;ch
<s
->nb_channels
;ch
++) {
870 put_bits(&s
->pb
, 2, exp_strategy
[ch
]);
874 put_bits(&s
->pb
, 1, exp_strategy
[s
->lfe_channel
]);
877 for(ch
=0;ch
<s
->nb_channels
;ch
++) {
878 if (exp_strategy
[ch
] != EXP_REUSE
)
879 put_bits(&s
->pb
, 6, s
->chbwcod
[ch
]);
883 for (ch
= 0; ch
< s
->nb_all_channels
; ch
++) {
884 switch(exp_strategy
[ch
]) {
898 nb_groups
= (s
->nb_coefs
[ch
] + (group_size
* 3) - 4) / (3 * group_size
);
903 put_bits(&s
->pb
, 4, exp1
);
905 /* next ones are delta encoded */
906 for(i
=0;i
<nb_groups
;i
++) {
907 /* merge three delta in one code */
911 delta0
= exp1
- exp0
+ 2;
916 delta1
= exp1
- exp0
+ 2;
921 delta2
= exp1
- exp0
+ 2;
923 put_bits(&s
->pb
, 7, ((delta0
* 5 + delta1
) * 5) + delta2
);
926 if (ch
!= s
->lfe_channel
)
927 put_bits(&s
->pb
, 2, 0); /* no gain range info */
930 /* bit allocation info */
931 baie
= (block_num
== 0);
932 put_bits(&s
->pb
, 1, baie
);
934 put_bits(&s
->pb
, 2, s
->slow_decay_code
);
935 put_bits(&s
->pb
, 2, s
->fast_decay_code
);
936 put_bits(&s
->pb
, 2, s
->slow_gain_code
);
937 put_bits(&s
->pb
, 2, s
->db_per_bit_code
);
938 put_bits(&s
->pb
, 3, s
->floor_code
);
942 put_bits(&s
->pb
, 1, baie
); /* always present with bai */
944 put_bits(&s
->pb
, 6, s
->coarse_snr_offset
);
945 for(ch
=0;ch
<s
->nb_all_channels
;ch
++) {
946 put_bits(&s
->pb
, 4, s
->fine_snr_offset
[ch
]);
947 put_bits(&s
->pb
, 3, s
->fast_gain_code
[ch
]);
951 put_bits(&s
->pb
, 1, 0); /* no delta bit allocation */
952 put_bits(&s
->pb
, 1, 0); /* no data to skip */
954 /* mantissa encoding : we use two passes to handle the grouping. A
955 one pass method may be faster, but it would necessitate to
956 modify the output stream. */
958 /* first pass: quantize */
959 mant1_cnt
= mant2_cnt
= mant4_cnt
= 0;
960 qmant1_ptr
= qmant2_ptr
= qmant4_ptr
= NULL
;
962 for (ch
= 0; ch
< s
->nb_all_channels
; ch
++) {
965 for(i
=0;i
<s
->nb_coefs
[ch
];i
++) {
966 c
= mdct_coefs
[ch
][i
];
967 e
= encoded_exp
[ch
][i
] - global_exp
[ch
];
974 v
= sym_quant(c
, e
, 3);
977 qmant1_ptr
= &qmant
[ch
][i
];
982 *qmant1_ptr
+= 3 * v
;
994 v
= sym_quant(c
, e
, 5);
997 qmant2_ptr
= &qmant
[ch
][i
];
1002 *qmant2_ptr
+= 5 * v
;
1014 v
= sym_quant(c
, e
, 7);
1017 v
= sym_quant(c
, e
, 11);
1020 qmant4_ptr
= &qmant
[ch
][i
];
1032 v
= sym_quant(c
, e
, 15);
1035 v
= asym_quant(c
, e
, 14);
1038 v
= asym_quant(c
, e
, 16);
1041 v
= asym_quant(c
, e
, b
- 1);
1048 /* second pass : output the values */
1049 for (ch
= 0; ch
< s
->nb_all_channels
; ch
++) {
1052 for(i
=0;i
<s
->nb_coefs
[ch
];i
++) {
1060 put_bits(&s
->pb
, 5, q
);
1064 put_bits(&s
->pb
, 7, q
);
1067 put_bits(&s
->pb
, 3, q
);
1071 put_bits(&s
->pb
, 7, q
);
1074 put_bits(&s
->pb
, 14, q
);
1077 put_bits(&s
->pb
, 16, q
);
1080 put_bits(&s
->pb
, b
- 1, q
);
1087 #define CRC16_POLY ((1 << 0) | (1 << 2) | (1 << 15) | (1 << 16))
1089 static unsigned int mul_poly(unsigned int a
, unsigned int b
, unsigned int poly
)
1105 static unsigned int pow_poly(unsigned int a
, unsigned int n
, unsigned int poly
)
1111 r
= mul_poly(r
, a
, poly
);
1112 a
= mul_poly(a
, a
, poly
);
1119 /* compute log2(max(abs(tab[]))) */
1120 static int log2_tab(int16_t *tab
, int n
)
1131 static void lshift_tab(int16_t *tab
, int n
, int lshift
)
1139 } else if (lshift
< 0) {
1147 /* fill the end of the frame and compute the two crcs */
1148 static int output_frame_end(AC3EncodeContext
*s
)
1150 int frame_size
, frame_size_58
, n
, crc1
, crc2
, crc_inv
;
1153 frame_size
= s
->frame_size
; /* frame size in words */
1154 /* align to 8 bits */
1155 flush_put_bits(&s
->pb
);
1156 /* add zero bytes to reach the frame size */
1158 n
= 2 * s
->frame_size
- (put_bits_ptr(&s
->pb
) - frame
) - 2;
1161 memset(put_bits_ptr(&s
->pb
), 0, n
);
1163 /* Now we must compute both crcs : this is not so easy for crc1
1164 because it is at the beginning of the data... */
1165 frame_size_58
= (frame_size
>> 1) + (frame_size
>> 3);
1166 crc1
= bswap_16(av_crc(av_crc_get_table(AV_CRC_16_ANSI
), 0,
1167 frame
+ 4, 2 * frame_size_58
- 4));
1168 /* XXX: could precompute crc_inv */
1169 crc_inv
= pow_poly((CRC16_POLY
>> 1), (16 * frame_size_58
) - 16, CRC16_POLY
);
1170 crc1
= mul_poly(crc_inv
, crc1
, CRC16_POLY
);
1171 AV_WB16(frame
+2,crc1
);
1173 crc2
= bswap_16(av_crc(av_crc_get_table(AV_CRC_16_ANSI
), 0,
1174 frame
+ 2 * frame_size_58
,
1175 (frame_size
- frame_size_58
) * 2 - 2));
1176 AV_WB16(frame
+2*frame_size
-2,crc2
);
1178 // printf("n=%d frame_size=%d\n", n, frame_size);
1179 return frame_size
* 2;
1182 static int AC3_encode_frame(AVCodecContext
*avctx
,
1183 unsigned char *frame
, int buf_size
, void *data
)
1185 AC3EncodeContext
*s
= avctx
->priv_data
;
1186 int16_t *samples
= data
;
1188 int16_t input_samples
[N
];
1189 int32_t mdct_coef
[NB_BLOCKS
][AC3_MAX_CHANNELS
][N
/2];
1190 uint8_t exp
[NB_BLOCKS
][AC3_MAX_CHANNELS
][N
/2];
1191 uint8_t exp_strategy
[NB_BLOCKS
][AC3_MAX_CHANNELS
];
1192 uint8_t encoded_exp
[NB_BLOCKS
][AC3_MAX_CHANNELS
][N
/2];
1193 uint8_t bap
[NB_BLOCKS
][AC3_MAX_CHANNELS
][N
/2];
1194 int8_t exp_samples
[NB_BLOCKS
][AC3_MAX_CHANNELS
];
1198 for(ch
=0;ch
<s
->nb_all_channels
;ch
++) {
1199 int ich
= s
->channel_map
[ch
];
1200 /* fixed mdct to the six sub blocks & exponent computation */
1201 for(i
=0;i
<NB_BLOCKS
;i
++) {
1205 /* compute input samples */
1206 memcpy(input_samples
, s
->last_samples
[ich
], N
/2 * sizeof(int16_t));
1207 sinc
= s
->nb_all_channels
;
1208 sptr
= samples
+ (sinc
* (N
/2) * i
) + ich
;
1209 for(j
=0;j
<N
/2;j
++) {
1211 input_samples
[j
+ N
/2] = v
;
1212 s
->last_samples
[ich
][j
] = v
;
1216 /* apply the MDCT window */
1217 for(j
=0;j
<N
/2;j
++) {
1218 input_samples
[j
] = MUL16(input_samples
[j
],
1219 ff_ac3_window
[j
]) >> 15;
1220 input_samples
[N
-j
-1] = MUL16(input_samples
[N
-j
-1],
1221 ff_ac3_window
[j
]) >> 15;
1224 /* Normalize the samples to use the maximum available
1226 v
= 14 - log2_tab(input_samples
, N
);
1229 exp_samples
[i
][ch
] = v
- 9;
1230 lshift_tab(input_samples
, N
, v
);
1233 mdct512(mdct_coef
[i
][ch
], input_samples
);
1235 /* compute "exponents". We take into account the
1236 normalization there */
1237 for(j
=0;j
<N
/2;j
++) {
1239 v
= abs(mdct_coef
[i
][ch
][j
]);
1243 e
= 23 - av_log2(v
) + exp_samples
[i
][ch
];
1246 mdct_coef
[i
][ch
][j
] = 0;
1253 compute_exp_strategy(exp_strategy
, exp
, ch
, ch
== s
->lfe_channel
);
1255 /* compute the exponents as the decoder will see them. The
1256 EXP_REUSE case must be handled carefully : we select the
1257 min of the exponents */
1259 while (i
< NB_BLOCKS
) {
1261 while (j
< NB_BLOCKS
&& exp_strategy
[j
][ch
] == EXP_REUSE
) {
1262 exponent_min(exp
[i
][ch
], exp
[j
][ch
], s
->nb_coefs
[ch
]);
1265 frame_bits
+= encode_exp(encoded_exp
[i
][ch
],
1266 exp
[i
][ch
], s
->nb_coefs
[ch
],
1267 exp_strategy
[i
][ch
]);
1268 /* copy encoded exponents for reuse case */
1269 for(k
=i
+1;k
<j
;k
++) {
1270 memcpy(encoded_exp
[k
][ch
], encoded_exp
[i
][ch
],
1271 s
->nb_coefs
[ch
] * sizeof(uint8_t));
1277 /* adjust for fractional frame sizes */
1278 while(s
->bits_written
>= s
->bit_rate
&& s
->samples_written
>= s
->sample_rate
) {
1279 s
->bits_written
-= s
->bit_rate
;
1280 s
->samples_written
-= s
->sample_rate
;
1282 s
->frame_size
= s
->frame_size_min
+ (s
->bits_written
* s
->sample_rate
< s
->samples_written
* s
->bit_rate
);
1283 s
->bits_written
+= s
->frame_size
* 16;
1284 s
->samples_written
+= AC3_FRAME_SIZE
;
1286 compute_bit_allocation(s
, bap
, encoded_exp
, exp_strategy
, frame_bits
);
1287 /* everything is known... let's output the frame */
1288 output_frame_header(s
, frame
);
1290 for(i
=0;i
<NB_BLOCKS
;i
++) {
1291 output_audio_block(s
, exp_strategy
[i
], encoded_exp
[i
],
1292 bap
[i
], mdct_coef
[i
], exp_samples
[i
], i
);
1294 return output_frame_end(s
);
1297 static av_cold
int AC3_encode_close(AVCodecContext
*avctx
)
1299 av_freep(&avctx
->coded_frame
);
1304 /*************************************************************************/
1312 IComplex in
[FN
], in1
[FN
];
1314 float sum_re
, sum_im
, a
;
1319 in
[i
].re
= random() % 65535 - 32767;
1320 in
[i
].im
= random() % 65535 - 32767;
1330 a
= -2 * M_PI
* (n
* k
) / FN
;
1331 sum_re
+= in1
[n
].re
* cos(a
) - in1
[n
].im
* sin(a
);
1332 sum_im
+= in1
[n
].re
* sin(a
) + in1
[n
].im
* cos(a
);
1334 printf("%3d: %6d,%6d %6.0f,%6.0f\n",
1335 k
, in
[k
].re
, in
[k
].im
, sum_re
/ FN
, sum_im
/ FN
);
1339 void mdct_test(void)
1342 int32_t output
[N
/2];
1345 float s
, a
, err
, e
, emax
;
1349 input
[i
] = (random() % 65535 - 32767) * 9 / 10;
1350 input1
[i
] = input
[i
];
1353 mdct512(output
, input
);
1356 for(k
=0;k
<N
/2;k
++) {
1359 a
= (2*M_PI
*(2*n
+1+N
/2)*(2*k
+1) / (4 * N
));
1360 s
+= input1
[n
] * cos(a
);
1362 output1
[k
] = -2 * s
/ N
;
1367 for(i
=0;i
<N
/2;i
++) {
1368 printf("%3d: %7d %7.0f\n", i
, output
[i
], output1
[i
]);
1369 e
= output
[i
] - output1
[i
];
1374 printf("err2=%f emax=%f\n", err
/ (N
/2), emax
);
1379 AC3EncodeContext ctx
;
1380 unsigned char frame
[AC3_MAX_CODED_FRAME_SIZE
];
1381 short samples
[AC3_FRAME_SIZE
];
1384 AC3_encode_init(&ctx
, 44100, 64000, 1);
1389 for(i
=0;i
<AC3_FRAME_SIZE
;i
++)
1390 samples
[i
] = (int)(sin(2*M_PI
*i
*1000.0/44100) * 10000);
1391 ret
= AC3_encode_frame(&ctx
, frame
, samples
);
1392 printf("ret=%d\n", ret
);
1396 AVCodec ac3_encoder
= {
1400 sizeof(AC3EncodeContext
),
1405 .sample_fmts
= (enum SampleFormat
[]){SAMPLE_FMT_S16
,SAMPLE_FMT_NONE
},
1406 .long_name
= NULL_IF_CONFIG_SMALL("ATSC A/52A (AC-3)"),
1407 .channel_layouts
= (int64_t[]){
1416 CH_LAYOUT_5POINT0_BACK
,
1417 (CH_LAYOUT_MONO
| CH_LOW_FREQUENCY
),
1418 (CH_LAYOUT_STEREO
| CH_LOW_FREQUENCY
),
1419 (CH_LAYOUT_2_1
| CH_LOW_FREQUENCY
),
1420 (CH_LAYOUT_SURROUND
| CH_LOW_FREQUENCY
),
1421 (CH_LAYOUT_2_2
| CH_LOW_FREQUENCY
),
1422 (CH_LAYOUT_QUAD
| CH_LOW_FREQUENCY
),
1423 (CH_LAYOUT_4POINT0
| CH_LOW_FREQUENCY
),
1425 CH_LAYOUT_5POINT1_BACK
,