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
24 * The simplest AC3 encoder.
27 //#define DEBUG_BITALLOC
29 #include "bitstream.h"
33 typedef struct AC3EncodeContext {
39 unsigned int sample_rate;
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;
46 unsigned int frmsizecod;
47 unsigned int fscod; /* frequency */
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 sgaincod, sdecaycod, fdecaycod, dbkneecod, floorcod;
57 AC3BitAllocParameters bit_alloc;
59 int fgaincod[AC3_MAX_CHANNELS];
60 int fsnroffst[AC3_MAX_CHANNELS];
61 /* mantissa encoding */
62 int mant1_cnt, mant2_cnt, mant4_cnt;
68 #define N (1 << MDCT_NBITS)
70 /* new exponents are sent if their Norm 1 exceed this number */
71 #define EXP_DIFF_THRESHOLD 1000
73 static void fft_init(int ln);
75 static inline int16_t fix15(float a)
78 v = (int)(a * (float)(1 << 15));
86 static inline int calc_lowcomp1(int a, int b0, int b1)
88 if ((b0 + 256) == b1) {
97 static inline int calc_lowcomp(int a, int b0, int b1, int bin)
100 if ((b0 + 256) == b1) {
102 } else if (b0 > b1) {
106 } else if (bin < 20) {
107 if ((b0 + 256) == b1) {
109 } else if (b0 > b1) {
120 /* AC3 bit allocation. The algorithm is the one described in the AC3
122 void ac3_parametric_bit_allocation(AC3BitAllocParameters *s, uint8_t *bap,
123 int8_t *exp, int start, int end,
124 int snroffset, int fgain, int is_lfe,
125 int deltbae,int deltnseg,
126 uint8_t *deltoffst, uint8_t *deltlen, uint8_t *deltba)
128 int bin,i,j,k,end1,v,v1,bndstrt,bndend,lowcomp,begin;
129 int fastleak,slowleak,address,tmp;
130 int16_t psd[256]; /* scaled exponents */
131 int16_t bndpsd[50]; /* interpolated exponents */
132 int16_t excite[50]; /* excitation */
133 int16_t mask[50]; /* masking value */
135 /* exponent mapping to PSD */
136 for(bin=start;bin<end;bin++) {
137 psd[bin]=(3072 - (exp[bin] << 7));
140 /* PSD integration */
147 if (end1 > end) end1=end;
148 for(i=j;i<end1;i++) {
155 if (adr > 255) adr=255;
159 if (adr > 255) adr=255;
166 } while (end > bndtab[k]);
168 /* excitation function */
169 bndstrt = masktab[start];
170 bndend = masktab[end-1] + 1;
174 lowcomp = calc_lowcomp1(lowcomp, bndpsd[0], bndpsd[1]) ;
175 excite[0] = bndpsd[0] - fgain - lowcomp ;
176 lowcomp = calc_lowcomp1(lowcomp, bndpsd[1], bndpsd[2]) ;
177 excite[1] = bndpsd[1] - fgain - lowcomp ;
179 for (bin = 2; bin < 7; bin++) {
180 if (!(is_lfe && bin == 6))
181 lowcomp = calc_lowcomp1(lowcomp, bndpsd[bin], bndpsd[bin+1]) ;
182 fastleak = bndpsd[bin] - fgain ;
183 slowleak = bndpsd[bin] - s->sgain ;
184 excite[bin] = fastleak - lowcomp ;
185 if (!(is_lfe && bin == 6)) {
186 if (bndpsd[bin] <= bndpsd[bin+1]) {
194 if (end1 > 22) end1=22;
196 for (bin = begin; bin < end1; bin++) {
197 if (!(is_lfe && bin == 6))
198 lowcomp = calc_lowcomp(lowcomp, bndpsd[bin], bndpsd[bin+1], bin) ;
200 fastleak -= s->fdecay ;
201 v = bndpsd[bin] - fgain;
202 if (fastleak < v) fastleak = v;
204 slowleak -= s->sdecay ;
205 v = bndpsd[bin] - s->sgain;
206 if (slowleak < v) slowleak = v;
208 v=fastleak - lowcomp;
209 if (slowleak > v) v=slowleak;
215 /* coupling channel */
218 fastleak = (s->cplfleak << 8) + 768;
219 slowleak = (s->cplsleak << 8) + 768;
222 for (bin = begin; bin < bndend; bin++) {
223 fastleak -= s->fdecay ;
224 v = bndpsd[bin] - fgain;
225 if (fastleak < v) fastleak = v;
226 slowleak -= s->sdecay ;
227 v = bndpsd[bin] - s->sgain;
228 if (slowleak < v) slowleak = v;
231 if (slowleak > v) v = slowleak;
235 /* compute masking curve */
237 for (bin = bndstrt; bin < bndend; bin++) {
239 tmp = s->dbknee - bndpsd[bin];
243 v=hth[bin >> s->halfratecod][s->fscod];
248 /* delta bit allocation */
250 if (deltbae == 0 || deltbae == 1) {
251 int band, seg, delta;
253 for (seg = 0; seg < deltnseg; seg++) {
254 band += deltoffst[seg] ;
255 if (deltba[seg] >= 4) {
256 delta = (deltba[seg] - 3) << 7;
258 delta = (deltba[seg] - 4) << 7;
260 for (k = 0; k < deltlen[seg]; k++) {
261 mask[band] += delta ;
267 /* compute bit allocation */
279 end1=bndtab[j] + bndsz[j];
280 if (end1 > end) end1=end;
282 for (k = i; k < end1; k++) {
283 address = (psd[i] - v) >> 5 ;
284 if (address < 0) address=0;
285 else if (address > 63) address=63;
286 bap[i] = baptab[address];
289 } while (end > bndtab[j++]) ;
292 typedef struct IComplex {
296 static void fft_init(int ln)
303 for(i=0;i<(n/2);i++) {
304 alpha = 2 * M_PI * (float)i / (float)n;
305 costab[i] = fix15(cos(alpha));
306 sintab[i] = fix15(sin(alpha));
312 m |= ((i >> j) & 1) << (ln-j-1);
319 #define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \
326 pre = (bx + ax) >> 1;\
327 pim = (by + ay) >> 1;\
328 qre = (bx - ax) >> 1;\
329 qim = (by - ay) >> 1;\
332 #define MUL16(a,b) ((a) * (b))
334 #define CMUL(pre, pim, are, aim, bre, bim) \
336 pre = (MUL16(are, bre) - MUL16(aim, bim)) >> 15;\
337 pim = (MUL16(are, bim) + MUL16(bre, aim)) >> 15;\
341 /* do a 2^n point complex fft on 2^ln points. */
342 static void fft(IComplex *z, int ln)
346 register IComplex *p,*q;
368 BF(p[0].re, p[0].im, p[1].re, p[1].im,
369 p[0].re, p[0].im, p[1].re, p[1].im);
378 BF(p[0].re, p[0].im, p[2].re, p[2].im,
379 p[0].re, p[0].im, p[2].re, p[2].im);
380 BF(p[1].re, p[1].im, p[3].re, p[3].im,
381 p[1].re, p[1].im, p[3].im, -p[3].re);
393 for (j = 0; j < nblocks; ++j) {
395 BF(p->re, p->im, q->re, q->im,
396 p->re, p->im, q->re, q->im);
400 for(l = nblocks; l < np2; l += nblocks) {
401 CMUL(tmp_re, tmp_im, costab[l], -sintab[l], q->re, q->im);
402 BF(p->re, p->im, q->re, q->im,
403 p->re, p->im, tmp_re, tmp_im);
410 nblocks = nblocks >> 1;
411 nloops = nloops << 1;
412 } while (nblocks != 0);
415 /* do a 512 point mdct */
416 static void mdct512(int32_t *out, int16_t *in)
418 int i, re, im, re1, im1;
422 /* shift to simplify computations */
424 rot[i] = -in[i + 3*N/4];
426 rot[i] = in[i - N/4];
430 re = ((int)rot[2*i] - (int)rot[N-1-2*i]) >> 1;
431 im = -((int)rot[N/2+2*i] - (int)rot[N/2-1-2*i]) >> 1;
432 CMUL(x[i].re, x[i].im, re, im, -xcos1[i], xsin1[i]);
435 fft(x, MDCT_NBITS - 2);
441 CMUL(re1, im1, re, im, xsin1[i], xcos1[i]);
443 out[N/2-1-2*i] = re1;
447 /* XXX: use another norm ? */
448 static int calc_exp_diff(uint8_t *exp1, uint8_t *exp2, int n)
453 sum += abs(exp1[i] - exp2[i]);
458 static void compute_exp_strategy(uint8_t exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS],
459 uint8_t exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
465 /* estimate if the exponent variation & decide if they should be
466 reused in the next frame */
467 exp_strategy[0][ch] = EXP_NEW;
468 for(i=1;i<NB_BLOCKS;i++) {
469 exp_diff = calc_exp_diff(exp[i][ch], exp[i-1][ch], N/2);
471 av_log(NULL, AV_LOG_DEBUG, "exp_diff=%d\n", exp_diff);
473 if (exp_diff > EXP_DIFF_THRESHOLD)
474 exp_strategy[i][ch] = EXP_NEW;
476 exp_strategy[i][ch] = EXP_REUSE;
481 /* now select the encoding strategy type : if exponents are often
482 recoded, we use a coarse encoding */
484 while (i < NB_BLOCKS) {
486 while (j < NB_BLOCKS && exp_strategy[j][ch] == EXP_REUSE)
490 exp_strategy[i][ch] = EXP_D45;
494 exp_strategy[i][ch] = EXP_D25;
497 exp_strategy[i][ch] = EXP_D15;
504 /* set exp[i] to min(exp[i], exp1[i]) */
505 static void exponent_min(uint8_t exp[N/2], uint8_t exp1[N/2], int n)
510 if (exp1[i] < exp[i])
515 /* update the exponents so that they are the ones the decoder will
516 decode. Return the number of bits used to code the exponents */
517 static int encode_exp(uint8_t encoded_exp[N/2],
522 int group_size, nb_groups, i, j, k, exp_min;
525 switch(exp_strategy) {
537 nb_groups = ((nb_exps + (group_size * 3) - 4) / (3 * group_size)) * 3;
539 /* for each group, compute the minimum exponent */
540 exp1[0] = exp[0]; /* DC exponent is handled separately */
542 for(i=1;i<=nb_groups;i++) {
544 assert(exp_min >= 0 && exp_min <= 24);
545 for(j=1;j<group_size;j++) {
546 if (exp[k+j] < exp_min)
553 /* constraint for DC exponent */
557 /* Decrease the delta between each groups to within 2
558 * so that they can be differentially encoded */
559 for (i=1;i<=nb_groups;i++)
560 exp1[i] = FFMIN(exp1[i], exp1[i-1] + 2);
561 for (i=nb_groups-1;i>=0;i--)
562 exp1[i] = FFMIN(exp1[i], exp1[i+1] + 2);
564 /* now we have the exponent values the decoder will see */
565 encoded_exp[0] = exp1[0];
567 for(i=1;i<=nb_groups;i++) {
568 for(j=0;j<group_size;j++) {
569 encoded_exp[k+j] = exp1[i];
575 av_log(NULL, AV_LOG_DEBUG, "exponents: strategy=%d\n", exp_strategy);
576 for(i=0;i<=nb_groups * group_size;i++) {
577 av_log(NULL, AV_LOG_DEBUG, "%d ", encoded_exp[i]);
579 av_log(NULL, AV_LOG_DEBUG, "\n");
582 return 4 + (nb_groups / 3) * 7;
585 /* return the size in bits taken by the mantissa */
586 static int compute_mantissa_size(AC3EncodeContext *s, uint8_t *m, int nb_coefs)
591 for(i=0;i<nb_coefs;i++) {
598 /* 3 mantissa in 5 bits */
599 if (s->mant1_cnt == 0)
601 if (++s->mant1_cnt == 3)
605 /* 3 mantissa in 7 bits */
606 if (s->mant2_cnt == 0)
608 if (++s->mant2_cnt == 3)
615 /* 2 mantissa in 7 bits */
616 if (s->mant4_cnt == 0)
618 if (++s->mant4_cnt == 2)
636 static int bit_alloc(AC3EncodeContext *s,
637 uint8_t bap[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
638 uint8_t encoded_exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
639 uint8_t exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS],
640 int frame_bits, int csnroffst, int fsnroffst)
645 for(i=0;i<NB_BLOCKS;i++) {
649 for(ch=0;ch<s->nb_all_channels;ch++) {
650 ac3_parametric_bit_allocation(&s->bit_alloc,
651 bap[i][ch], (int8_t *)encoded_exp[i][ch],
653 (((csnroffst-15) << 4) +
655 fgaintab[s->fgaincod[ch]],
656 ch == s->lfe_channel,
657 2, 0, NULL, NULL, NULL);
658 frame_bits += compute_mantissa_size(s, bap[i][ch],
663 printf("csnr=%d fsnr=%d frame_bits=%d diff=%d\n",
664 csnroffst, fsnroffst, frame_bits,
665 16 * s->frame_size - ((frame_bits + 7) & ~7));
667 return 16 * s->frame_size - frame_bits;
672 static int compute_bit_allocation(AC3EncodeContext *s,
673 uint8_t bap[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
674 uint8_t encoded_exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
675 uint8_t exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS],
679 int csnroffst, fsnroffst;
680 uint8_t bap1[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
681 static int frame_bits_inc[8] = { 0, 0, 2, 2, 2, 4, 2, 4 };
683 /* init default parameters */
689 for(ch=0;ch<s->nb_all_channels;ch++)
692 /* compute real values */
693 s->bit_alloc.fscod = s->fscod;
694 s->bit_alloc.halfratecod = s->halfratecod;
695 s->bit_alloc.sdecay = sdecaytab[s->sdecaycod] >> s->halfratecod;
696 s->bit_alloc.fdecay = fdecaytab[s->fdecaycod] >> s->halfratecod;
697 s->bit_alloc.sgain = sgaintab[s->sgaincod];
698 s->bit_alloc.dbknee = dbkneetab[s->dbkneecod];
699 s->bit_alloc.floor = floortab[s->floorcod];
703 // if (s->acmod == 2)
705 frame_bits += frame_bits_inc[s->acmod];
708 for(i=0;i<NB_BLOCKS;i++) {
709 frame_bits += s->nb_channels * 2 + 2; /* blksw * c, dithflag * c, dynrnge, cplstre */
711 frame_bits++; /* rematstr */
712 if(i==0) frame_bits += 4;
714 frame_bits += 2 * s->nb_channels; /* chexpstr[2] * c */
716 frame_bits++; /* lfeexpstr */
717 for(ch=0;ch<s->nb_channels;ch++) {
718 if (exp_strategy[i][ch] != EXP_REUSE)
719 frame_bits += 6 + 2; /* chbwcod[6], gainrng[2] */
721 frame_bits++; /* baie */
722 frame_bits++; /* snr */
723 frame_bits += 2; /* delta / skip */
725 frame_bits++; /* cplinu for block 0 */
727 /* sdcycod[2], fdcycod[2], sgaincod[2], dbpbcod[2], floorcod[3] */
729 /* (fsnoffset[4] + fgaincod[4]) * c */
730 frame_bits += 2*4 + 3 + 6 + s->nb_all_channels * (4 + 3);
732 /* auxdatae, crcrsv */
738 /* now the big work begins : do the bit allocation. Modify the snr
739 offset until we can pack everything in the requested frame size */
741 csnroffst = s->csnroffst;
742 while (csnroffst >= 0 &&
743 bit_alloc(s, bap, encoded_exp, exp_strategy, frame_bits, csnroffst, 0) < 0)
744 csnroffst -= SNR_INC1;
746 av_log(NULL, AV_LOG_ERROR, "Bit allocation failed, try increasing the bitrate, -ab 384 for example!\n");
749 while ((csnroffst + SNR_INC1) <= 63 &&
750 bit_alloc(s, bap1, encoded_exp, exp_strategy, frame_bits,
751 csnroffst + SNR_INC1, 0) >= 0) {
752 csnroffst += SNR_INC1;
753 memcpy(bap, bap1, sizeof(bap1));
755 while ((csnroffst + 1) <= 63 &&
756 bit_alloc(s, bap1, encoded_exp, exp_strategy, frame_bits, csnroffst + 1, 0) >= 0) {
758 memcpy(bap, bap1, sizeof(bap1));
762 while ((fsnroffst + SNR_INC1) <= 15 &&
763 bit_alloc(s, bap1, encoded_exp, exp_strategy, frame_bits,
764 csnroffst, fsnroffst + SNR_INC1) >= 0) {
765 fsnroffst += SNR_INC1;
766 memcpy(bap, bap1, sizeof(bap1));
768 while ((fsnroffst + 1) <= 15 &&
769 bit_alloc(s, bap1, encoded_exp, exp_strategy, frame_bits,
770 csnroffst, fsnroffst + 1) >= 0) {
772 memcpy(bap, bap1, sizeof(bap1));
775 s->csnroffst = csnroffst;
776 for(ch=0;ch<s->nb_all_channels;ch++)
777 s->fsnroffst[ch] = fsnroffst;
778 #if defined(DEBUG_BITALLOC)
783 for(ch=0;ch<s->nb_all_channels;ch++) {
784 printf("Block #%d Ch%d:\n", i, ch);
786 for(j=0;j<s->nb_coefs[ch];j++) {
787 printf("%d ",bap[i][ch][j]);
797 void ac3_common_init(void)
800 /* compute bndtab and masktab from bandsz */
806 for(j=0;j<v;j++) masktab[k++]=i;
813 static int AC3_encode_init(AVCodecContext *avctx)
815 int freq = avctx->sample_rate;
816 int bitrate = avctx->bit_rate;
817 int channels = avctx->channels;
818 AC3EncodeContext *s = avctx->priv_data;
821 static const uint8_t acmod_defs[6] = {
825 0x06, /* L R SL SR */
826 0x07, /* L C R SL SR */
827 0x07, /* L C R SL SR (+LFE) */
830 avctx->frame_size = AC3_FRAME_SIZE;
832 /* number of channels */
833 if (channels < 1 || channels > 6)
835 s->acmod = acmod_defs[channels - 1];
836 s->lfe = (channels == 6) ? 1 : 0;
837 s->nb_all_channels = channels;
838 s->nb_channels = channels > 5 ? 5 : channels;
839 s->lfe_channel = s->lfe ? 5 : -1;
844 if ((ac3_freqs[j] >> i) == freq)
849 s->sample_rate = freq;
852 s->bsid = 8 + s->halfratecod;
853 s->bsmod = 0; /* complete main audio service */
855 /* bitrate & frame size */
858 if ((ac3_bitratetab[i] >> s->halfratecod) == bitrate)
863 s->bit_rate = bitrate;
864 s->frmsizecod = i << 1;
865 s->frame_size_min = (bitrate * 1000 * AC3_FRAME_SIZE) / (freq * 16);
867 s->samples_written = 0;
868 s->frame_size = s->frame_size_min;
870 /* bit allocation init */
871 for(ch=0;ch<s->nb_channels;ch++) {
872 /* bandwidth for each channel */
873 /* XXX: should compute the bandwidth according to the frame
874 size, so that we avoid anoying high freq artefacts */
875 s->chbwcod[ch] = 50; /* sample bandwidth as mpeg audio layer 2 table 0 */
876 s->nb_coefs[ch] = ((s->chbwcod[ch] + 12) * 3) + 37;
879 s->nb_coefs[s->lfe_channel] = 7; /* fixed */
881 /* initial snr offset */
887 fft_init(MDCT_NBITS - 2);
889 alpha = 2 * M_PI * (i + 1.0 / 8.0) / (float)N;
890 xcos1[i] = fix15(-cos(alpha));
891 xsin1[i] = fix15(-sin(alpha));
894 avctx->coded_frame= avcodec_alloc_frame();
895 avctx->coded_frame->key_frame= 1;
900 /* output the AC3 frame header */
901 static void output_frame_header(AC3EncodeContext *s, unsigned char *frame)
903 init_put_bits(&s->pb, frame, AC3_MAX_CODED_FRAME_SIZE);
905 put_bits(&s->pb, 16, 0x0b77); /* frame header */
906 put_bits(&s->pb, 16, 0); /* crc1: will be filled later */
907 put_bits(&s->pb, 2, s->fscod);
908 put_bits(&s->pb, 6, s->frmsizecod + (s->frame_size - s->frame_size_min));
909 put_bits(&s->pb, 5, s->bsid);
910 put_bits(&s->pb, 3, s->bsmod);
911 put_bits(&s->pb, 3, s->acmod);
912 if ((s->acmod & 0x01) && s->acmod != 0x01)
913 put_bits(&s->pb, 2, 1); /* XXX -4.5 dB */
915 put_bits(&s->pb, 2, 1); /* XXX -6 dB */
916 if (s->acmod == 0x02)
917 put_bits(&s->pb, 2, 0); /* surround not indicated */
918 put_bits(&s->pb, 1, s->lfe); /* LFE */
919 put_bits(&s->pb, 5, 31); /* dialog norm: -31 db */
920 put_bits(&s->pb, 1, 0); /* no compression control word */
921 put_bits(&s->pb, 1, 0); /* no lang code */
922 put_bits(&s->pb, 1, 0); /* no audio production info */
923 put_bits(&s->pb, 1, 0); /* no copyright */
924 put_bits(&s->pb, 1, 1); /* original bitstream */
925 put_bits(&s->pb, 1, 0); /* no time code 1 */
926 put_bits(&s->pb, 1, 0); /* no time code 2 */
927 put_bits(&s->pb, 1, 0); /* no addtional bit stream info */
930 /* symetric quantization on 'levels' levels */
931 static inline int sym_quant(int c, int e, int levels)
936 v = (levels * (c << e)) >> 24;
938 v = (levels >> 1) + v;
940 v = (levels * ((-c) << e)) >> 24;
942 v = (levels >> 1) - v;
944 assert (v >= 0 && v < levels);
948 /* asymetric quantization on 2^qbits levels */
949 static inline int asym_quant(int c, int e, int qbits)
953 lshift = e + qbits - 24;
960 m = (1 << (qbits-1));
964 return v & ((1 << qbits)-1);
967 /* Output one audio block. There are NB_BLOCKS audio blocks in one AC3
969 static void output_audio_block(AC3EncodeContext *s,
970 uint8_t exp_strategy[AC3_MAX_CHANNELS],
971 uint8_t encoded_exp[AC3_MAX_CHANNELS][N/2],
972 uint8_t bap[AC3_MAX_CHANNELS][N/2],
973 int32_t mdct_coefs[AC3_MAX_CHANNELS][N/2],
974 int8_t global_exp[AC3_MAX_CHANNELS],
977 int ch, nb_groups, group_size, i, baie, rbnd;
979 uint16_t qmant[AC3_MAX_CHANNELS][N/2];
981 int mant1_cnt, mant2_cnt, mant4_cnt;
982 uint16_t *qmant1_ptr, *qmant2_ptr, *qmant4_ptr;
983 int delta0, delta1, delta2;
985 for(ch=0;ch<s->nb_channels;ch++)
986 put_bits(&s->pb, 1, 0); /* 512 point MDCT */
987 for(ch=0;ch<s->nb_channels;ch++)
988 put_bits(&s->pb, 1, 1); /* no dither */
989 put_bits(&s->pb, 1, 0); /* no dynamic range */
990 if (block_num == 0) {
991 /* for block 0, even if no coupling, we must say it. This is a
993 put_bits(&s->pb, 1, 1); /* coupling strategy present */
994 put_bits(&s->pb, 1, 0); /* no coupling strategy */
996 put_bits(&s->pb, 1, 0); /* no new coupling strategy */
1003 /* first block must define rematrixing (rematstr) */
1004 put_bits(&s->pb, 1, 1);
1006 /* dummy rematrixing rematflg(1:4)=0 */
1007 for (rbnd=0;rbnd<4;rbnd++)
1008 put_bits(&s->pb, 1, 0);
1012 /* no matrixing (but should be used in the future) */
1013 put_bits(&s->pb, 1, 0);
1019 static int count = 0;
1020 av_log(NULL, AV_LOG_DEBUG, "Block #%d (%d)\n", block_num, count++);
1023 /* exponent strategy */
1024 for(ch=0;ch<s->nb_channels;ch++) {
1025 put_bits(&s->pb, 2, exp_strategy[ch]);
1029 put_bits(&s->pb, 1, exp_strategy[s->lfe_channel]);
1032 for(ch=0;ch<s->nb_channels;ch++) {
1033 if (exp_strategy[ch] != EXP_REUSE)
1034 put_bits(&s->pb, 6, s->chbwcod[ch]);
1038 for (ch = 0; ch < s->nb_all_channels; ch++) {
1039 switch(exp_strategy[ch]) {
1053 nb_groups = (s->nb_coefs[ch] + (group_size * 3) - 4) / (3 * group_size);
1054 p = encoded_exp[ch];
1056 /* first exponent */
1058 put_bits(&s->pb, 4, exp1);
1060 /* next ones are delta encoded */
1061 for(i=0;i<nb_groups;i++) {
1062 /* merge three delta in one code */
1066 delta0 = exp1 - exp0 + 2;
1071 delta1 = exp1 - exp0 + 2;
1076 delta2 = exp1 - exp0 + 2;
1078 put_bits(&s->pb, 7, ((delta0 * 5 + delta1) * 5) + delta2);
1081 if (ch != s->lfe_channel)
1082 put_bits(&s->pb, 2, 0); /* no gain range info */
1085 /* bit allocation info */
1086 baie = (block_num == 0);
1087 put_bits(&s->pb, 1, baie);
1089 put_bits(&s->pb, 2, s->sdecaycod);
1090 put_bits(&s->pb, 2, s->fdecaycod);
1091 put_bits(&s->pb, 2, s->sgaincod);
1092 put_bits(&s->pb, 2, s->dbkneecod);
1093 put_bits(&s->pb, 3, s->floorcod);
1097 put_bits(&s->pb, 1, baie); /* always present with bai */
1099 put_bits(&s->pb, 6, s->csnroffst);
1100 for(ch=0;ch<s->nb_all_channels;ch++) {
1101 put_bits(&s->pb, 4, s->fsnroffst[ch]);
1102 put_bits(&s->pb, 3, s->fgaincod[ch]);
1106 put_bits(&s->pb, 1, 0); /* no delta bit allocation */
1107 put_bits(&s->pb, 1, 0); /* no data to skip */
1109 /* mantissa encoding : we use two passes to handle the grouping. A
1110 one pass method may be faster, but it would necessitate to
1111 modify the output stream. */
1113 /* first pass: quantize */
1114 mant1_cnt = mant2_cnt = mant4_cnt = 0;
1115 qmant1_ptr = qmant2_ptr = qmant4_ptr = NULL;
1117 for (ch = 0; ch < s->nb_all_channels; ch++) {
1120 for(i=0;i<s->nb_coefs[ch];i++) {
1121 c = mdct_coefs[ch][i];
1122 e = encoded_exp[ch][i] - global_exp[ch];
1129 v = sym_quant(c, e, 3);
1132 qmant1_ptr = &qmant[ch][i];
1137 *qmant1_ptr += 3 * v;
1149 v = sym_quant(c, e, 5);
1152 qmant2_ptr = &qmant[ch][i];
1157 *qmant2_ptr += 5 * v;
1169 v = sym_quant(c, e, 7);
1172 v = sym_quant(c, e, 11);
1175 qmant4_ptr = &qmant[ch][i];
1187 v = sym_quant(c, e, 15);
1190 v = asym_quant(c, e, 14);
1193 v = asym_quant(c, e, 16);
1196 v = asym_quant(c, e, b - 1);
1203 /* second pass : output the values */
1204 for (ch = 0; ch < s->nb_all_channels; ch++) {
1207 for(i=0;i<s->nb_coefs[ch];i++) {
1215 put_bits(&s->pb, 5, q);
1219 put_bits(&s->pb, 7, q);
1222 put_bits(&s->pb, 3, q);
1226 put_bits(&s->pb, 7, q);
1229 put_bits(&s->pb, 14, q);
1232 put_bits(&s->pb, 16, q);
1235 put_bits(&s->pb, b - 1, q);
1242 #define CRC16_POLY ((1 << 0) | (1 << 2) | (1 << 15) | (1 << 16))
1244 static unsigned int mul_poly(unsigned int a, unsigned int b, unsigned int poly)
1260 static unsigned int pow_poly(unsigned int a, unsigned int n, unsigned int poly)
1266 r = mul_poly(r, a, poly);
1267 a = mul_poly(a, a, poly);
1274 /* compute log2(max(abs(tab[]))) */
1275 static int log2_tab(int16_t *tab, int n)
1286 static void lshift_tab(int16_t *tab, int n, int lshift)
1294 } else if (lshift < 0) {
1302 /* fill the end of the frame and compute the two crcs */
1303 static int output_frame_end(AC3EncodeContext *s)
1305 int frame_size, frame_size_58, n, crc1, crc2, crc_inv;
1308 frame_size = s->frame_size; /* frame size in words */
1309 /* align to 8 bits */
1310 flush_put_bits(&s->pb);
1311 /* add zero bytes to reach the frame size */
1313 n = 2 * s->frame_size - (pbBufPtr(&s->pb) - frame) - 2;
1316 memset(pbBufPtr(&s->pb), 0, n);
1318 /* Now we must compute both crcs : this is not so easy for crc1
1319 because it is at the beginning of the data... */
1320 frame_size_58 = (frame_size >> 1) + (frame_size >> 3);
1321 crc1 = bswap_16(av_crc(av_crc8005, 0, frame + 4, 2 * frame_size_58 - 4));
1322 /* XXX: could precompute crc_inv */
1323 crc_inv = pow_poly((CRC16_POLY >> 1), (16 * frame_size_58) - 16, CRC16_POLY);
1324 crc1 = mul_poly(crc_inv, crc1, CRC16_POLY);
1325 frame[2] = crc1 >> 8;
1328 crc2 = bswap_16(av_crc(av_crc8005, 0, frame + 2 * frame_size_58, (frame_size - frame_size_58) * 2 - 2));
1329 frame[2*frame_size - 2] = crc2 >> 8;
1330 frame[2*frame_size - 1] = crc2;
1332 // printf("n=%d frame_size=%d\n", n, frame_size);
1333 return frame_size * 2;
1336 static int AC3_encode_frame(AVCodecContext *avctx,
1337 unsigned char *frame, int buf_size, void *data)
1339 AC3EncodeContext *s = avctx->priv_data;
1340 int16_t *samples = data;
1342 int16_t input_samples[N];
1343 int32_t mdct_coef[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1344 uint8_t exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1345 uint8_t exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS];
1346 uint8_t encoded_exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1347 uint8_t bap[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1348 int8_t exp_samples[NB_BLOCKS][AC3_MAX_CHANNELS];
1352 for(ch=0;ch<s->nb_all_channels;ch++) {
1353 /* fixed mdct to the six sub blocks & exponent computation */
1354 for(i=0;i<NB_BLOCKS;i++) {
1358 /* compute input samples */
1359 memcpy(input_samples, s->last_samples[ch], N/2 * sizeof(int16_t));
1360 sinc = s->nb_all_channels;
1361 sptr = samples + (sinc * (N/2) * i) + ch;
1362 for(j=0;j<N/2;j++) {
1364 input_samples[j + N/2] = v;
1365 s->last_samples[ch][j] = v;
1369 /* apply the MDCT window */
1370 for(j=0;j<N/2;j++) {
1371 input_samples[j] = MUL16(input_samples[j],
1372 ac3_window[j]) >> 15;
1373 input_samples[N-j-1] = MUL16(input_samples[N-j-1],
1374 ac3_window[j]) >> 15;
1377 /* Normalize the samples to use the maximum available
1379 v = 14 - log2_tab(input_samples, N);
1382 exp_samples[i][ch] = v - 9;
1383 lshift_tab(input_samples, N, v);
1386 mdct512(mdct_coef[i][ch], input_samples);
1388 /* compute "exponents". We take into account the
1389 normalization there */
1390 for(j=0;j<N/2;j++) {
1392 v = abs(mdct_coef[i][ch][j]);
1396 e = 23 - av_log2(v) + exp_samples[i][ch];
1399 mdct_coef[i][ch][j] = 0;
1406 compute_exp_strategy(exp_strategy, exp, ch, ch == s->lfe_channel);
1408 /* compute the exponents as the decoder will see them. The
1409 EXP_REUSE case must be handled carefully : we select the
1410 min of the exponents */
1412 while (i < NB_BLOCKS) {
1414 while (j < NB_BLOCKS && exp_strategy[j][ch] == EXP_REUSE) {
1415 exponent_min(exp[i][ch], exp[j][ch], s->nb_coefs[ch]);
1418 frame_bits += encode_exp(encoded_exp[i][ch],
1419 exp[i][ch], s->nb_coefs[ch],
1420 exp_strategy[i][ch]);
1421 /* copy encoded exponents for reuse case */
1422 for(k=i+1;k<j;k++) {
1423 memcpy(encoded_exp[k][ch], encoded_exp[i][ch],
1424 s->nb_coefs[ch] * sizeof(uint8_t));
1430 /* adjust for fractional frame sizes */
1431 while(s->bits_written >= s->bit_rate*1000 && s->samples_written >= s->sample_rate) {
1432 s->bits_written -= s->bit_rate*1000;
1433 s->samples_written -= s->sample_rate;
1435 s->frame_size = s->frame_size_min + (s->bits_written * s->sample_rate < s->samples_written * s->bit_rate*1000);
1436 s->bits_written += s->frame_size * 16;
1437 s->samples_written += AC3_FRAME_SIZE;
1439 compute_bit_allocation(s, bap, encoded_exp, exp_strategy, frame_bits);
1440 /* everything is known... let's output the frame */
1441 output_frame_header(s, frame);
1443 for(i=0;i<NB_BLOCKS;i++) {
1444 output_audio_block(s, exp_strategy[i], encoded_exp[i],
1445 bap[i], mdct_coef[i], exp_samples[i], i);
1447 return output_frame_end(s);
1450 static int AC3_encode_close(AVCodecContext *avctx)
1452 av_freep(&avctx->coded_frame);
1457 /*************************************************************************/
1464 IComplex in[FN], in1[FN];
1466 float sum_re, sum_im, a;
1471 in[i].re = random() % 65535 - 32767;
1472 in[i].im = random() % 65535 - 32767;
1482 a = -2 * M_PI * (n * k) / FN;
1483 sum_re += in1[n].re * cos(a) - in1[n].im * sin(a);
1484 sum_im += in1[n].re * sin(a) + in1[n].im * cos(a);
1486 printf("%3d: %6d,%6d %6.0f,%6.0f\n",
1487 k, in[k].re, in[k].im, sum_re / FN, sum_im / FN);
1491 void mdct_test(void)
1494 int32_t output[N/2];
1497 float s, a, err, e, emax;
1501 input[i] = (random() % 65535 - 32767) * 9 / 10;
1502 input1[i] = input[i];
1505 mdct512(output, input);
1508 for(k=0;k<N/2;k++) {
1511 a = (2*M_PI*(2*n+1+N/2)*(2*k+1) / (4 * N));
1512 s += input1[n] * cos(a);
1514 output1[k] = -2 * s / N;
1519 for(i=0;i<N/2;i++) {
1520 printf("%3d: %7d %7.0f\n", i, output[i], output1[i]);
1521 e = output[i] - output1[i];
1526 printf("err2=%f emax=%f\n", err / (N/2), emax);
1531 AC3EncodeContext ctx;
1532 unsigned char frame[AC3_MAX_CODED_FRAME_SIZE];
1533 short samples[AC3_FRAME_SIZE];
1536 AC3_encode_init(&ctx, 44100, 64000, 1);
1541 for(i=0;i<AC3_FRAME_SIZE;i++)
1542 samples[i] = (int)(sin(2*M_PI*i*1000.0/44100) * 10000);
1543 ret = AC3_encode_frame(&ctx, frame, samples);
1544 printf("ret=%d\n", ret);
1548 AVCodec ac3_encoder = {
1552 sizeof(AC3EncodeContext),