hvirtual/quicktime/libavcodec/ac3enc.c

   1 /*
   2  * The simplest AC3 encoder
   3  * Copyright (c) 2000 Fabrice Bellard.
   4  *
   5  * This library is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU Lesser General Public
   7  * License as published by the Free Software Foundation; either
   8  * version 2 of the License, or (at your option) any later version.
   9  *
  10  * This library is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * Lesser General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU Lesser General Public
  16  * License along with this library; if not, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  18  */
  19 //#define DEBUG
  20 //#define DEBUG_BITALLOC
  21 #include "avcodec.h"
  22
  23 #include "ac3.h"
  24
  25 typedef struct AC3EncodeContext {
  26     PutBitContext pb;
  27     int nb_channels;
  28     int nb_all_channels;
  29     int lfe_channel;
  30     int bit_rate;
  31     int sample_rate;
  32     int bsid;
  33     int frame_size_min; /* minimum frame size in case rounding is necessary */
  34     int frame_size; /* current frame size in words */
  35     int halfratecod;
  36     int frmsizecod;
  37     int fscod; /* frequency */
  38     int acmod;
  39     int lfe;
  40     int bsmod;
  41     short last_samples[AC3_MAX_CHANNELS][256];
  42     int chbwcod[AC3_MAX_CHANNELS];
  43     int nb_coefs[AC3_MAX_CHANNELS];
  44
  45     /* bitrate allocation control */
  46     int sgaincod, sdecaycod, fdecaycod, dbkneecod, floorcod;
  47     AC3BitAllocParameters bit_alloc;
  48     int csnroffst;
  49     int fgaincod[AC3_MAX_CHANNELS];
  50     int fsnroffst[AC3_MAX_CHANNELS];
  51     /* mantissa encoding */
  52     int mant1_cnt, mant2_cnt, mant4_cnt;
  53 } AC3EncodeContext;
  54
  55 #include "ac3tab.h"
  56
  57 #define MDCT_NBITS 9
  58 #define N         (1 << MDCT_NBITS)
  59
  60 /* new exponents are sent if their Norm 1 exceed this number */
  61 #define EXP_DIFF_THRESHOLD 1000
  62
  63 static void fft_init(int ln);
  64 static void ac3_crc_init(void);
  65
  66 static inline INT16 fix15(float a)
  67 {
  68     int v;
  69     v = (int)(a * (float)(1 << 15));
  70     if (v < -32767)
  71         v = -32767;
  72     else if (v > 32767)
  73         v = 32767;
  74     return v;
  75 }
  76
  77 static inline int calc_lowcomp1(int a, int b0, int b1)
  78 {
  79     if ((b0 + 256) == b1) {
  80         a = 384 ;
  81     } else if (b0 > b1) {
  82         a = a - 64;
  83         if (a < 0) a=0;
  84     }
  85     return a;
  86 }
  87
  88 static inline int calc_lowcomp(int a, int b0, int b1, int bin)
  89 {
  90     if (bin < 7) {
  91         if ((b0 + 256) == b1) {
  92             a = 384 ;
  93         } else if (b0 > b1) {
  94             a = a - 64;
  95             if (a < 0) a=0;
  96         }
  97     } else if (bin < 20) {
  98         if ((b0 + 256) == b1) {
  99             a = 320 ;
 100         } else if (b0 > b1) {
 101             a= a - 64;
 102             if (a < 0) a=0;
 103         }
 104     } else {
 105         a = a - 128;
 106         if (a < 0) a=0;
 107     }
 108     return a;
 109 }
 110
 111 /* AC3 bit allocation. The algorithm is the one described in the AC3
 112    spec. */
 113 void ac3_parametric_bit_allocation(AC3BitAllocParameters *s, UINT8 *bap,
 114                                    INT8 *exp, int start, int end,
 115                                    int snroffset, int fgain, int is_lfe,
 116                                    int deltbae,int deltnseg,
 117                                    UINT8 *deltoffst, UINT8 *deltlen, UINT8 *deltba)
 118 {
 119     int bin,i,j,k,end1,v,v1,bndstrt,bndend,lowcomp,begin;
 120     int fastleak,slowleak,address,tmp;
 121     INT16 psd[256]; /* scaled exponents */
 122     INT16 bndpsd[50]; /* interpolated exponents */
 123     INT16 excite[50]; /* excitation */
 124     INT16 mask[50];   /* masking value */
 125
 126     /* exponent mapping to PSD */
 127     for(bin=start;bin<end;bin++) {
 128         psd[bin]=(3072 - (exp[bin] << 7));
 129     }
 130
 131     /* PSD integration */
 132     j=start;
 133     k=masktab[start];
 134     do {
 135         v=psd[j];
 136         j++;
 137         end1=bndtab[k+1];
 138         if (end1 > end) end1=end;
 139         for(i=j;i<end1;i++) {
 140             int c,adr;
 141             /* logadd */
 142             v1=psd[j];
 143             c=v-v1;
 144             if (c >= 0) {
 145                 adr=c >> 1;
 146                 if (adr > 255) adr=255;
 147                 v=v + latab[adr];
 148             } else {
 149                 adr=(-c) >> 1;
 150                 if (adr > 255) adr=255;
 151                 v=v1 + latab[adr];
 152             }
 153             j++;
 154         }
 155         bndpsd[k]=v;
 156         k++;
 157     } while (end > bndtab[k]);
 158
 159     /* excitation function */
 160     bndstrt = masktab[start];
 161     bndend = masktab[end-1] + 1;
 162
 163     if (bndstrt == 0) {
 164         lowcomp = 0;
 165         lowcomp = calc_lowcomp1(lowcomp, bndpsd[0], bndpsd[1]) ;
 166         excite[0] = bndpsd[0] - fgain - lowcomp ;
 167         lowcomp = calc_lowcomp1(lowcomp, bndpsd[1], bndpsd[2]) ;
 168         excite[1] = bndpsd[1] - fgain - lowcomp ;
 169         begin = 7 ;
 170         for (bin = 2; bin < 7; bin++) {
 171             if (!(is_lfe && bin == 6))
 172                 lowcomp = calc_lowcomp1(lowcomp, bndpsd[bin], bndpsd[bin+1]) ;
 173             fastleak = bndpsd[bin] - fgain ;
 174             slowleak = bndpsd[bin] - s->sgain ;
 175             excite[bin] = fastleak - lowcomp ;
 176             if (!(is_lfe && bin == 6)) {
 177                 if (bndpsd[bin] <= bndpsd[bin+1]) {
 178                     begin = bin + 1 ;
 179                     break ;
 180                 }
 181             }
 182         }
 183
 184         end1=bndend;
 185         if (end1 > 22) end1=22;
 186
 187         for (bin = begin; bin < end1; bin++) {
 188             if (!(is_lfe && bin == 6))
 189                 lowcomp = calc_lowcomp(lowcomp, bndpsd[bin], bndpsd[bin+1], bin) ;
 190
 191             fastleak -= s->fdecay ;
 192             v = bndpsd[bin] - fgain;
 193             if (fastleak < v) fastleak = v;
 194
 195             slowleak -= s->sdecay ;
 196             v = bndpsd[bin] - s->sgain;
 197             if (slowleak < v) slowleak = v;
 198
 199             v=fastleak - lowcomp;
 200             if (slowleak > v) v=slowleak;
 201
 202             excite[bin] = v;
 203         }
 204         begin = 22;
 205     } else {
 206         /* coupling channel */
 207         begin = bndstrt;
 208
 209         fastleak = (s->cplfleak << 8) + 768;
 210         slowleak = (s->cplsleak << 8) + 768;
 211     }
 212
 213     for (bin = begin; bin < bndend; bin++) {
 214         fastleak -= s->fdecay ;
 215         v = bndpsd[bin] - fgain;
 216         if (fastleak < v) fastleak = v;
 217         slowleak -= s->sdecay ;
 218         v = bndpsd[bin] - s->sgain;
 219         if (slowleak < v) slowleak = v;
 220
 221         v=fastleak;
 222         if (slowleak > v) v = slowleak;
 223         excite[bin] = v;
 224     }
 225
 226     /* compute masking curve */
 227
 228     for (bin = bndstrt; bin < bndend; bin++) {
 229         v1 = excite[bin];
 230         tmp = s->dbknee - bndpsd[bin];
 231         if (tmp > 0) {
 232             v1 += tmp >> 2;
 233         }
 234         v=hth[bin >> s->halfratecod][s->fscod];
 235         if (v1 > v) v=v1;
 236         mask[bin] = v;
 237     }
 238
 239     /* delta bit allocation */
 240
 241     if (deltbae == 0 || deltbae == 1) {
 242         int band, seg, delta;
 243         band = 0 ;
 244         for (seg = 0; seg < deltnseg; seg++) {
 245             band += deltoffst[seg] ;
 246             if (deltba[seg] >= 4) {
 247                 delta = (deltba[seg] - 3) << 7;
 248             } else {
 249                 delta = (deltba[seg] - 4) << 7;
 250             }
 251             for (k = 0; k < deltlen[seg]; k++) {
 252                 mask[band] += delta ;
 253                 band++ ;
 254             }
 255         }
 256     }
 257
 258     /* compute bit allocation */
 259
 260     i = start ;
 261     j = masktab[start] ;
 262     do {
 263         v=mask[j];
 264         v -= snroffset ;
 265         v -= s->floor ;
 266         if (v < 0) v = 0;
 267         v &= 0x1fe0 ;
 268         v += s->floor ;
 269
 270         end1=bndtab[j] + bndsz[j];
 271         if (end1 > end) end1=end;
 272
 273         for (k = i; k < end1; k++) {
 274             address = (psd[i] - v) >> 5 ;
 275             if (address < 0) address=0;
 276             else if (address > 63) address=63;
 277             bap[i] = baptab[address];
 278             i++;
 279         }
 280     } while (end > bndtab[j++]) ;
 281 }
 282
 283 typedef struct IComplex {
 284     short re,im;
 285 } IComplex;
 286
 287 static void fft_init(int ln)
 288 {
 289     int i, j, m, n;
 290     float alpha;
 291
 292     n = 1 << ln;
 293
 294     for(i=0;i<(n/2);i++) {
 295         alpha = 2 * M_PI * (float)i / (float)n;
 296         costab[i] = fix15(cos(alpha));
 297         sintab[i] = fix15(sin(alpha));
 298     }
 299
 300     for(i=0;i<n;i++) {
 301         m=0;
 302         for(j=0;j<ln;j++) {
 303             m |= ((i >> j) & 1) << (ln-j-1);
 304         }
 305         fft_rev[i]=m;
 306     }
 307 }
 308
 309 /* butter fly op */
 310 #define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \
 311 {\
 312   int ax, ay, bx, by;\
 313   bx=pre1;\
 314   by=pim1;\
 315   ax=qre1;\
 316   ay=qim1;\
 317   pre = (bx + ax) >> 1;\
 318   pim = (by + ay) >> 1;\
 319   qre = (bx - ax) >> 1;\
 320   qim = (by - ay) >> 1;\
 321 }
 322
 323 #define MUL16(a,b) ((a) * (b))
 324
 325 #define CMUL(pre, pim, are, aim, bre, bim) \
 326 {\
 327    pre = (MUL16(are, bre) - MUL16(aim, bim)) >> 15;\
 328    pim = (MUL16(are, bim) + MUL16(bre, aim)) >> 15;\
 329 }
 330
 331
 332 /* do a 2^n point complex fft on 2^ln points. */
 333 static void fft(IComplex *z, int ln)
 334 {
 335     int j, l, np, np2;
 336     int nblocks, nloops;
 337     register IComplex *p,*q;
 338     int tmp_re, tmp_im;
 339
 340     np = 1 << ln;
 341
 342     /* reverse */
 343     for(j=0;j<np;j++) {
 344         int k;
 345         IComplex tmp;
 346         k = fft_rev[j];
 347         if (k < j) {
 348             tmp = z[k];
 349             z[k] = z[j];
 350             z[j] = tmp;
 351         }
 352     }
 353
 354     /* pass 0 */
 355
 356     p=&z[0];
 357     j=(np >> 1);
 358     do {
 359         BF(p[0].re, p[0].im, p[1].re, p[1].im,
 360            p[0].re, p[0].im, p[1].re, p[1].im);
 361         p+=2;
 362     } while (--j != 0);
 363
 364     /* pass 1 */
 365
 366     p=&z[0];
 367     j=np >> 2;
 368     do {
 369         BF(p[0].re, p[0].im, p[2].re, p[2].im,
 370            p[0].re, p[0].im, p[2].re, p[2].im);
 371         BF(p[1].re, p[1].im, p[3].re, p[3].im,
 372            p[1].re, p[1].im, p[3].im, -p[3].re);
 373         p+=4;
 374     } while (--j != 0);
 375
 376     /* pass 2 .. ln-1 */
 377
 378     nblocks = np >> 3;
 379     nloops = 1 << 2;
 380     np2 = np >> 1;
 381     do {
 382         p = z;
 383         q = z + nloops;
 384         for (j = 0; j < nblocks; ++j) {
 385
 386             BF(p->re, p->im, q->re, q->im,
 387                p->re, p->im, q->re, q->im);
 388
 389             p++;
 390             q++;
 391             for(l = nblocks; l < np2; l += nblocks) {
 392                 CMUL(tmp_re, tmp_im, costab[l], -sintab[l], q->re, q->im);
 393                 BF(p->re, p->im, q->re, q->im,
 394                    p->re, p->im, tmp_re, tmp_im);
 395                 p++;
 396                 q++;
 397             }
 398             p += nloops;
 399             q += nloops;
 400         }
 401         nblocks = nblocks >> 1;
 402         nloops = nloops << 1;
 403     } while (nblocks != 0);
 404 }
 405
 406 /* do a 512 point mdct */
 407 static void mdct512(INT32 *out, INT16 *in)
 408 {
 409     int i, re, im, re1, im1;
 410     INT16 rot[N];
 411     IComplex x[N/4];
 412
 413     /* shift to simplify computations */
 414     for(i=0;i<N/4;i++)
 415         rot[i] = -in[i + 3*N/4];
 416     for(i=N/4;i<N;i++)
 417         rot[i] = in[i - N/4];
 418
 419     /* pre rotation */
 420     for(i=0;i<N/4;i++) {
 421         re = ((int)rot[2*i] - (int)rot[N-1-2*i]) >> 1;
 422         im = -((int)rot[N/2+2*i] - (int)rot[N/2-1-2*i]) >> 1;
 423         CMUL(x[i].re, x[i].im, re, im, -xcos1[i], xsin1[i]);
 424     }
 425
 426     fft(x, MDCT_NBITS - 2);
 427
 428     /* post rotation */
 429     for(i=0;i<N/4;i++) {
 430         re = x[i].re;
 431         im = x[i].im;
 432         CMUL(re1, im1, re, im, xsin1[i], xcos1[i]);
 433         out[2*i] = im1;
 434         out[N/2-1-2*i] = re1;
 435     }
 436 }
 437
 438 /* XXX: use another norm ? */
 439 static int calc_exp_diff(UINT8 *exp1, UINT8 *exp2, int n)
 440 {
 441     int sum, i;
 442     sum = 0;
 443     for(i=0;i<n;i++) {
 444         sum += abs(exp1[i] - exp2[i]);
 445     }
 446     return sum;
 447 }
 448
 449 static void compute_exp_strategy(UINT8 exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS],
 450                                  UINT8 exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
 451                                  int ch, int is_lfe)
 452 {
 453     int i, j;
 454     int exp_diff;
 455
 456     /* estimate if the exponent variation & decide if they should be
 457        reused in the next frame */
 458     exp_strategy[0][ch] = EXP_NEW;
 459     for(i=1;i<NB_BLOCKS;i++) {
 460         exp_diff = calc_exp_diff(exp[i][ch], exp[i-1][ch], N/2);
 461 #ifdef DEBUG
 462         printf("exp_diff=%d\n", exp_diff);
 463 #endif
 464         if (exp_diff > EXP_DIFF_THRESHOLD)
 465             exp_strategy[i][ch] = EXP_NEW;
 466         else
 467             exp_strategy[i][ch] = EXP_REUSE;
 468     }
 469     if (is_lfe)
 470         return;
 471
 472     /* now select the encoding strategy type : if exponents are often
 473        recoded, we use a coarse encoding */
 474     i = 0;
 475     while (i < NB_BLOCKS) {
 476         j = i + 1;
 477         while (j < NB_BLOCKS && exp_strategy[j][ch] == EXP_REUSE)
 478             j++;
 479         switch(j - i) {
 480         case 1:
 481             exp_strategy[i][ch] = EXP_D45;
 482             break;
 483         case 2:
 484         case 3:
 485             exp_strategy[i][ch] = EXP_D25;
 486             break;
 487         default:
 488             exp_strategy[i][ch] = EXP_D15;
 489             break;
 490         }
 491         i = j;
 492     }
 493 }
 494
 495 /* set exp[i] to min(exp[i], exp1[i]) */
 496 static void exponent_min(UINT8 exp[N/2], UINT8 exp1[N/2], int n)
 497 {
 498     int i;
 499
 500     for(i=0;i<n;i++) {
 501         if (exp1[i] < exp[i])
 502             exp[i] = exp1[i];
 503     }
 504 }
 505
 506 /* update the exponents so that they are the ones the decoder will
 507    decode. Return the number of bits used to code the exponents */
 508 static int encode_exp(UINT8 encoded_exp[N/2],
 509                       UINT8 exp[N/2],
 510                       int nb_exps,
 511                       int exp_strategy)
 512 {
 513     int group_size, nb_groups, i, j, k, recurse, exp_min, delta;
 514     UINT8 exp1[N/2];
 515
 516     switch(exp_strategy) {
 517     case EXP_D15:
 518         group_size = 1;
 519         break;
 520     case EXP_D25:
 521         group_size = 2;
 522         break;
 523     default:
 524     case EXP_D45:
 525         group_size = 4;
 526         break;
 527     }
 528     nb_groups = ((nb_exps + (group_size * 3) - 4) / (3 * group_size)) * 3;
 529
 530     /* for each group, compute the minimum exponent */
 531     exp1[0] = exp[0]; /* DC exponent is handled separately */
 532     k = 1;
 533     for(i=1;i<=nb_groups;i++) {
 534         exp_min = exp[k];
 535         assert(exp_min >= 0 && exp_min <= 24);
 536         for(j=1;j<group_size;j++) {
 537             if (exp[k+j] < exp_min)
 538                 exp_min = exp[k+j];
 539         }
 540         exp1[i] = exp_min;
 541         k += group_size;
 542     }
 543
 544     /* constraint for DC exponent */
 545     if (exp1[0] > 15)
 546         exp1[0] = 15;
 547
 548     /* Iterate until the delta constraints between each groups are
 549        satisfyed. I'm sure it is possible to find a better algorithm,
 550        but I am lazy */
 551     do {
 552         recurse = 0;
 553         for(i=1;i<=nb_groups;i++) {
 554             delta = exp1[i] - exp1[i-1];
 555             if (delta > 2) {
 556                 /* if delta too big, we encode a smaller exponent */
 557                 exp1[i] = exp1[i-1] + 2;
 558             } else if (delta < -2) {
 559                 /* if delta is too small, we must decrease the previous
 560                exponent, which means we must recurse */
 561                 recurse = 1;
 562                 exp1[i-1] = exp1[i] + 2;
 563             }
 564         }
 565     } while (recurse);
 566
 567     /* now we have the exponent values the decoder will see */
 568     encoded_exp[0] = exp1[0];
 569     k = 1;
 570     for(i=1;i<=nb_groups;i++) {
 571         for(j=0;j<group_size;j++) {
 572             encoded_exp[k+j] = exp1[i];
 573         }
 574         k += group_size;
 575     }
 576
 577 #if defined(DEBUG)
 578     printf("exponents: strategy=%d\n", exp_strategy);
 579     for(i=0;i<=nb_groups * group_size;i++) {
 580         printf("%d ", encoded_exp[i]);
 581     }
 582     printf("\n");
 583 #endif
 584
 585     return 4 + (nb_groups / 3) * 7;
 586 }
 587
 588 /* return the size in bits taken by the mantissa */
 589 int compute_mantissa_size(AC3EncodeContext *s, UINT8 *m, int nb_coefs)
 590 {
 591     int bits, mant, i;
 592
 593     bits = 0;
 594     for(i=0;i<nb_coefs;i++) {
 595         mant = m[i];
 596         switch(mant) {
 597         case 0:
 598             /* nothing */
 599             break;
 600         case 1:
 601             /* 3 mantissa in 5 bits */
 602             if (s->mant1_cnt == 0)
 603                 bits += 5;
 604             if (++s->mant1_cnt == 3)
 605                 s->mant1_cnt = 0;
 606             break;
 607         case 2:
 608             /* 3 mantissa in 7 bits */
 609             if (s->mant2_cnt == 0)
 610                 bits += 7;
 611             if (++s->mant2_cnt == 3)
 612                 s->mant2_cnt = 0;
 613             break;
 614         case 3:
 615             bits += 3;
 616             break;
 617         case 4:
 618             /* 2 mantissa in 7 bits */
 619             if (s->mant4_cnt == 0)
 620                 bits += 7;
 621             if (++s->mant4_cnt == 2)
 622                 s->mant4_cnt = 0;
 623             break;
 624         case 14:
 625             bits += 14;
 626             break;
 627         case 15:
 628             bits += 16;
 629             break;
 630         default:
 631             bits += mant - 1;
 632             break;
 633         }
 634     }
 635     return bits;
 636 }
 637
 638
 639 static int bit_alloc(AC3EncodeContext *s,
 640                      UINT8 bap[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
 641                      UINT8 encoded_exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
 642                      UINT8 exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS],
 643                      int frame_bits, int csnroffst, int fsnroffst)
 644 {
 645     int i, ch;
 646
 647     /* compute size */
 648     for(i=0;i<NB_BLOCKS;i++) {
 649         s->mant1_cnt = 0;
 650         s->mant2_cnt = 0;
 651         s->mant4_cnt = 0;
 652         for(ch=0;ch<s->nb_all_channels;ch++) {
 653             ac3_parametric_bit_allocation(&s->bit_alloc,
 654                                           bap[i][ch], (INT8 *)encoded_exp[i][ch],
 655                                           0, s->nb_coefs[ch],
 656                                           (((csnroffst-15) << 4) +
 657                                            fsnroffst) << 2,
 658                                           fgaintab[s->fgaincod[ch]],
 659                                           ch == s->lfe_channel,
 660                                           2, 0, NULL, NULL, NULL);
 661             frame_bits += compute_mantissa_size(s, bap[i][ch],
 662                                                  s->nb_coefs[ch]);
 663         }
 664     }
 665 #if 0
 666     printf("csnr=%d fsnr=%d frame_bits=%d diff=%d\n",
 667            csnroffst, fsnroffst, frame_bits,
 668            16 * s->frame_size - ((frame_bits + 7) & ~7));
 669 #endif
 670     return 16 * s->frame_size - frame_bits;
 671 }
 672
 673 #define SNR_INC1 4
 674
 675 static int compute_bit_allocation(AC3EncodeContext *s,
 676                                   UINT8 bap[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
 677                                   UINT8 encoded_exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2],
 678                                   UINT8 exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS],
 679                                   int frame_bits)
 680 {
 681     int i, ch;
 682     int csnroffst, fsnroffst;
 683     UINT8 bap1[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
 684     static int frame_bits_inc[8] = { 0, 0, 2, 2, 2, 4, 2, 4 };
 685
 686     /* init default parameters */
 687     s->sdecaycod = 2;
 688     s->fdecaycod = 1;
 689     s->sgaincod = 1;
 690     s->dbkneecod = 2;
 691     s->floorcod = 4;
 692     for(ch=0;ch<s->nb_all_channels;ch++)
 693         s->fgaincod[ch] = 4;
 694
 695     /* compute real values */
 696     s->bit_alloc.fscod = s->fscod;
 697     s->bit_alloc.halfratecod = s->halfratecod;
 698     s->bit_alloc.sdecay = sdecaytab[s->sdecaycod] >> s->halfratecod;
 699     s->bit_alloc.fdecay = fdecaytab[s->fdecaycod] >> s->halfratecod;
 700     s->bit_alloc.sgain = sgaintab[s->sgaincod];
 701     s->bit_alloc.dbknee = dbkneetab[s->dbkneecod];
 702     s->bit_alloc.floor = floortab[s->floorcod];
 703
 704     /* header size */
 705     frame_bits += 65;
 706     // if (s->acmod == 2)
 707     //    frame_bits += 2;
 708     frame_bits += frame_bits_inc[s->acmod];
 709
 710     /* audio blocks */
 711     for(i=0;i<NB_BLOCKS;i++) {
 712         frame_bits += s->nb_channels * 2 + 2; /* blksw * c, dithflag * c, dynrnge, cplstre */
 713         if (s->acmod == 2)
 714             frame_bits++; /* rematstr */
 715         frame_bits += 2 * s->nb_channels; /* chexpstr[2] * c */
 716         if (s->lfe)
 717             frame_bits++; /* lfeexpstr */
 718         for(ch=0;ch<s->nb_channels;ch++) {
 719             if (exp_strategy[i][ch] != EXP_REUSE)
 720                 frame_bits += 6 + 2; /* chbwcod[6], gainrng[2] */
 721         }
 722         frame_bits++; /* baie */
 723         frame_bits++; /* snr */
 724         frame_bits += 2; /* delta / skip */
 725     }
 726     frame_bits++; /* cplinu for block 0 */
 727     /* bit alloc info */
 728     /* sdcycod[2], fdcycod[2], sgaincod[2], dbpbcod[2], floorcod[3] */
 729     /* csnroffset[6] */
 730     /* (fsnoffset[4] + fgaincod[4]) * c */
 731     frame_bits += 2*4 + 3 + 6 + s->nb_all_channels * (4 + 3);
 732
 733     /* CRC */
 734     frame_bits += 16;
 735
 736     /* now the big work begins : do the bit allocation. Modify the snr
 737        offset until we can pack everything in the requested frame size */
 738
 739     csnroffst = s->csnroffst;
 740     while (csnroffst >= 0 &&
 741            bit_alloc(s, bap, encoded_exp, exp_strategy, frame_bits, csnroffst, 0) < 0)
 742         csnroffst -= SNR_INC1;
 743     if (csnroffst < 0) {
 744         fprintf(stderr, "Yack, Error !!!\n");
 745         return -1;
 746     }
 747     while ((csnroffst + SNR_INC1) <= 63 &&
 748            bit_alloc(s, bap1, encoded_exp, exp_strategy, frame_bits,
 749                      csnroffst + SNR_INC1, 0) >= 0) {
 750         csnroffst += SNR_INC1;
 751         memcpy(bap, bap1, sizeof(bap1));
 752     }
 753     while ((csnroffst + 1) <= 63 &&
 754            bit_alloc(s, bap1, encoded_exp, exp_strategy, frame_bits, csnroffst + 1, 0) >= 0) {
 755         csnroffst++;
 756         memcpy(bap, bap1, sizeof(bap1));
 757     }
 758
 759     fsnroffst = 0;
 760     while ((fsnroffst + SNR_INC1) <= 15 &&
 761            bit_alloc(s, bap1, encoded_exp, exp_strategy, frame_bits,
 762                      csnroffst, fsnroffst + SNR_INC1) >= 0) {
 763         fsnroffst += SNR_INC1;
 764         memcpy(bap, bap1, sizeof(bap1));
 765     }
 766     while ((fsnroffst + 1) <= 15 &&
 767            bit_alloc(s, bap1, encoded_exp, exp_strategy, frame_bits,
 768                      csnroffst, fsnroffst + 1) >= 0) {
 769         fsnroffst++;
 770         memcpy(bap, bap1, sizeof(bap1));
 771     }
 772
 773     s->csnroffst = csnroffst;
 774     for(ch=0;ch<s->nb_all_channels;ch++)
 775         s->fsnroffst[ch] = fsnroffst;
 776 #if defined(DEBUG_BITALLOC)
 777     {
 778         int j;
 779
 780         for(i=0;i<6;i++) {
 781             for(ch=0;ch<s->nb_all_channels;ch++) {
 782                 printf("Block #%d Ch%d:\n", i, ch);
 783                 printf("bap=");
 784                 for(j=0;j<s->nb_coefs[ch];j++) {
 785                     printf("%d ",bap[i][ch][j]);
 786                 }
 787                 printf("\n");
 788             }
 789         }
 790     }
 791 #endif
 792     return 0;
 793 }
 794
 795 void ac3_common_init(void)
 796 {
 797     int i, j, k, l, v;
 798     /* compute bndtab and masktab from bandsz */
 799     k = 0;
 800     l = 0;
 801     for(i=0;i<50;i++) {
 802         bndtab[i] = l;
 803         v = bndsz[i];
 804         for(j=0;j<v;j++) masktab[k++]=i;
 805         l += v;
 806     }
 807     bndtab[50] = 0;
 808 }
 809
 810
 811 static int AC3_encode_init(AVCodecContext *avctx)
 812 {
 813     int freq = avctx->sample_rate;
 814     int bitrate = avctx->bit_rate;
 815     int channels = avctx->channels;
 816     AC3EncodeContext *s = avctx->priv_data;
 817     int i, j, ch;
 818     float alpha;
 819     static const UINT8 acmod_defs[6] = {
 820         0x01, /* C */
 821         0x02, /* L R */
 822         0x03, /* L C R */
 823         0x06, /* L R SL SR */
 824         0x07, /* L C R SL SR */
 825         0x07, /* L C R SL SR (+LFE) */
 826     };
 827
 828     avctx->frame_size = AC3_FRAME_SIZE;
 829
 830     /* number of channels */
 831     if (channels < 1 || channels > 6)
 832         return -1;
 833     s->acmod = acmod_defs[channels - 1];
 834     s->lfe = (channels == 6) ? 1 : 0;
 835     s->nb_all_channels = channels;
 836     s->nb_channels = channels > 5 ? 5 : channels;
 837     s->lfe_channel = s->lfe ? 5 : -1;
 838
 839     /* frequency */
 840     for(i=0;i<3;i++) {
 841         for(j=0;j<3;j++)
 842             if ((ac3_freqs[j] >> i) == freq)
 843                 goto found;
 844     }
 845     return -1;
 846  found:
 847     s->sample_rate = freq;
 848     s->halfratecod = i;
 849     s->fscod = j;
 850     s->bsid = 8 + s->halfratecod;
 851     s->bsmod = 0; /* complete main audio service */
 852
 853     /* bitrate & frame size */
 854     bitrate /= 1000;
 855     for(i=0;i<19;i++) {
 856         if ((ac3_bitratetab[i] >> s->halfratecod) == bitrate)
 857             break;
 858     }
 859     if (i == 19)
 860         return -1;
 861     s->bit_rate = bitrate;
 862     s->frmsizecod = i << 1;
 863     s->frame_size_min = (bitrate * 1000 * AC3_FRAME_SIZE) / (freq * 16);
 864     /* for now we do not handle fractional sizes */
 865     s->frame_size = s->frame_size_min;
 866
 867     /* bit allocation init */
 868     for(ch=0;ch<s->nb_channels;ch++) {
 869         /* bandwidth for each channel */
 870         /* XXX: should compute the bandwidth according to the frame
 871            size, so that we avoid anoying high freq artefacts */
 872         s->chbwcod[ch] = 50; /* sample bandwidth as mpeg audio layer 2 table 0 */
 873         s->nb_coefs[ch] = ((s->chbwcod[ch] + 12) * 3) + 37;
 874     }
 875     if (s->lfe) {
 876         s->nb_coefs[s->lfe_channel] = 7; /* fixed */
 877     }
 878     /* initial snr offset */
 879     s->csnroffst = 40;
 880
 881     ac3_common_init();
 882
 883     /* mdct init */
 884     fft_init(MDCT_NBITS - 2);
 885     for(i=0;i<N/4;i++) {
 886         alpha = 2 * M_PI * (i + 1.0 / 8.0) / (float)N;
 887         xcos1[i] = fix15(-cos(alpha));
 888         xsin1[i] = fix15(-sin(alpha));
 889     }
 890
 891     ac3_crc_init();
 892
 893     avctx->coded_frame= avcodec_alloc_frame();
 894     avctx->coded_frame->key_frame= 1;
 895
 896     return 0;
 897 }
 898
 899 /* output the AC3 frame header */
 900 static void output_frame_header(AC3EncodeContext *s, unsigned char *frame)
 901 {
 902     init_put_bits(&s->pb, frame, AC3_MAX_CODED_FRAME_SIZE, NULL, NULL);
 903
 904     put_bits(&s->pb, 16, 0x0b77); /* frame header */
 905     put_bits(&s->pb, 16, 0); /* crc1: will be filled later */
 906     put_bits(&s->pb, 2, s->fscod);
 907     put_bits(&s->pb, 6, s->frmsizecod + (s->frame_size - s->frame_size_min));
 908     put_bits(&s->pb, 5, s->bsid);
 909     put_bits(&s->pb, 3, s->bsmod);
 910     put_bits(&s->pb, 3, s->acmod);
 911     if ((s->acmod & 0x01) && s->acmod != 0x01)
 912         put_bits(&s->pb, 2, 1); /* XXX -4.5 dB */
 913     if (s->acmod & 0x04)
 914         put_bits(&s->pb, 2, 1); /* XXX -6 dB */
 915     if (s->acmod == 0x02)
 916         put_bits(&s->pb, 2, 0); /* surround not indicated */
 917     put_bits(&s->pb, 1, s->lfe); /* LFE */
 918     put_bits(&s->pb, 5, 31); /* dialog norm: -31 db */
 919     put_bits(&s->pb, 1, 0); /* no compression control word */
 920     put_bits(&s->pb, 1, 0); /* no lang code */
 921     put_bits(&s->pb, 1, 0); /* no audio production info */
 922     put_bits(&s->pb, 1, 0); /* no copyright */
 923     put_bits(&s->pb, 1, 1); /* original bitstream */
 924     put_bits(&s->pb, 1, 0); /* no time code 1 */
 925     put_bits(&s->pb, 1, 0); /* no time code 2 */
 926     put_bits(&s->pb, 1, 0); /* no addtional bit stream info */
 927 }
 928
 929 /* symetric quantization on 'levels' levels */
 930 static inline int sym_quant(int c, int e, int levels)
 931 {
 932     int v;
 933
 934     if (c >= 0) {
 935         v = (levels * (c << e)) >> 24;
 936         v = (v + 1) >> 1;
 937         v = (levels >> 1) + v;
 938     } else {
 939         v = (levels * ((-c) << e)) >> 24;
 940         v = (v + 1) >> 1;
 941         v = (levels >> 1) - v;
 942     }
 943     assert (v >= 0 && v < levels);
 944     return v;
 945 }
 946
 947 /* asymetric quantization on 2^qbits levels */
 948 static inline int asym_quant(int c, int e, int qbits)
 949 {
 950     int lshift, m, v;
 951
 952     lshift = e + qbits - 24;
 953     if (lshift >= 0)
 954         v = c << lshift;
 955     else
 956         v = c >> (-lshift);
 957     /* rounding */
 958     v = (v + 1) >> 1;
 959     m = (1 << (qbits-1));
 960     if (v >= m)
 961         v = m - 1;
 962     assert(v >= -m);
 963     return v & ((1 << qbits)-1);
 964 }
 965
 966 /* Output one audio block. There are NB_BLOCKS audio blocks in one AC3
 967    frame */
 968 static void output_audio_block(AC3EncodeContext *s,
 969                                UINT8 exp_strategy[AC3_MAX_CHANNELS],
 970                                UINT8 encoded_exp[AC3_MAX_CHANNELS][N/2],
 971                                UINT8 bap[AC3_MAX_CHANNELS][N/2],
 972                                INT32 mdct_coefs[AC3_MAX_CHANNELS][N/2],
 973                                INT8 global_exp[AC3_MAX_CHANNELS],
 974                                int block_num)
 975 {
 976     int ch, nb_groups, group_size, i, baie;
 977     UINT8 *p;
 978     UINT16 qmant[AC3_MAX_CHANNELS][N/2];
 979     int exp0, exp1;
 980     int mant1_cnt, mant2_cnt, mant4_cnt;
 981     UINT16 *qmant1_ptr, *qmant2_ptr, *qmant4_ptr;
 982     int delta0, delta1, delta2;
 983
 984     for(ch=0;ch<s->nb_channels;ch++)
 985         put_bits(&s->pb, 1, 0); /* 512 point MDCT */
 986     for(ch=0;ch<s->nb_channels;ch++)
 987         put_bits(&s->pb, 1, 1); /* no dither */
 988     put_bits(&s->pb, 1, 0); /* no dynamic range */
 989     if (block_num == 0) {
 990         /* for block 0, even if no coupling, we must say it. This is a
 991            waste of bit :-) */
 992         put_bits(&s->pb, 1, 1); /* coupling strategy present */
 993         put_bits(&s->pb, 1, 0); /* no coupling strategy */
 994     } else {
 995         put_bits(&s->pb, 1, 0); /* no new coupling strategy */
 996     }
 997
 998     if (s->acmod == 2) {
 999         put_bits(&s->pb, 1, 0); /* no matrixing (but should be used in the future) */
1000     }
1001
1002 #if defined(DEBUG)
1003     {
1004         static int count = 0;
1005         printf("Block #%d (%d)\n", block_num, count++);
1006     }
1007 #endif
1008     /* exponent strategy */
1009     for(ch=0;ch<s->nb_channels;ch++) {
1010         put_bits(&s->pb, 2, exp_strategy[ch]);
1011     }
1012
1013     if (s->lfe) {
1014         put_bits(&s->pb, 1, exp_strategy[s->lfe_channel]);
1015     }
1016
1017     for(ch=0;ch<s->nb_channels;ch++) {
1018         if (exp_strategy[ch] != EXP_REUSE)
1019             put_bits(&s->pb, 6, s->chbwcod[ch]);
1020     }
1021
1022     /* exponents */
1023     for (ch = 0; ch < s->nb_all_channels; ch++) {
1024         switch(exp_strategy[ch]) {
1025         case EXP_REUSE:
1026             continue;
1027         case EXP_D15:
1028             group_size = 1;
1029             break;
1030         case EXP_D25:
1031             group_size = 2;
1032             break;
1033         default:
1034         case EXP_D45:
1035             group_size = 4;
1036             break;
1037         }
1038         nb_groups = (s->nb_coefs[ch] + (group_size * 3) - 4) / (3 * group_size);
1039         p = encoded_exp[ch];
1040
1041         /* first exponent */
1042         exp1 = *p++;
1043         put_bits(&s->pb, 4, exp1);
1044
1045         /* next ones are delta encoded */
1046         for(i=0;i<nb_groups;i++) {
1047             /* merge three delta in one code */
1048             exp0 = exp1;
1049             exp1 = p[0];
1050             p += group_size;
1051             delta0 = exp1 - exp0 + 2;
1052
1053             exp0 = exp1;
1054             exp1 = p[0];
1055             p += group_size;
1056             delta1 = exp1 - exp0 + 2;
1057
1058             exp0 = exp1;
1059             exp1 = p[0];
1060             p += group_size;
1061             delta2 = exp1 - exp0 + 2;
1062
1063             put_bits(&s->pb, 7, ((delta0 * 5 + delta1) * 5) + delta2);
1064         }
1065
1066         if (ch != s->lfe_channel)
1067             put_bits(&s->pb, 2, 0); /* no gain range info */
1068     }
1069
1070     /* bit allocation info */
1071     baie = (block_num == 0);
1072     put_bits(&s->pb, 1, baie);
1073     if (baie) {
1074         put_bits(&s->pb, 2, s->sdecaycod);
1075         put_bits(&s->pb, 2, s->fdecaycod);
1076         put_bits(&s->pb, 2, s->sgaincod);
1077         put_bits(&s->pb, 2, s->dbkneecod);
1078         put_bits(&s->pb, 3, s->floorcod);
1079     }
1080
1081     /* snr offset */
1082     put_bits(&s->pb, 1, baie); /* always present with bai */
1083     if (baie) {
1084         put_bits(&s->pb, 6, s->csnroffst);
1085         for(ch=0;ch<s->nb_all_channels;ch++) {
1086             put_bits(&s->pb, 4, s->fsnroffst[ch]);
1087             put_bits(&s->pb, 3, s->fgaincod[ch]);
1088         }
1089     }
1090
1091     put_bits(&s->pb, 1, 0); /* no delta bit allocation */
1092     put_bits(&s->pb, 1, 0); /* no data to skip */
1093
1094     /* mantissa encoding : we use two passes to handle the grouping. A
1095        one pass method may be faster, but it would necessitate to
1096        modify the output stream. */
1097
1098     /* first pass: quantize */
1099     mant1_cnt = mant2_cnt = mant4_cnt = 0;
1100     qmant1_ptr = qmant2_ptr = qmant4_ptr = NULL;
1101
1102     for (ch = 0; ch < s->nb_all_channels; ch++) {
1103         int b, c, e, v;
1104
1105         for(i=0;i<s->nb_coefs[ch];i++) {
1106             c = mdct_coefs[ch][i];
1107             e = encoded_exp[ch][i] - global_exp[ch];
1108             b = bap[ch][i];
1109             switch(b) {
1110             case 0:
1111                 v = 0;
1112                 break;
1113             case 1:
1114                 v = sym_quant(c, e, 3);
1115                 switch(mant1_cnt) {
1116                 case 0:
1117                     qmant1_ptr = &qmant[ch][i];
1118                     v = 9 * v;
1119                     mant1_cnt = 1;
1120                     break;
1121                 case 1:
1122                     *qmant1_ptr += 3 * v;
1123                     mant1_cnt = 2;
1124                     v = 128;
1125                     break;
1126                 default:
1127                     *qmant1_ptr += v;
1128                     mant1_cnt = 0;
1129                     v = 128;
1130                     break;
1131                 }
1132                 break;
1133             case 2:
1134                 v = sym_quant(c, e, 5);
1135                 switch(mant2_cnt) {
1136                 case 0:
1137                     qmant2_ptr = &qmant[ch][i];
1138                     v = 25 * v;
1139                     mant2_cnt = 1;
1140                     break;
1141                 case 1:
1142                     *qmant2_ptr += 5 * v;
1143                     mant2_cnt = 2;
1144                     v = 128;
1145                     break;
1146                 default:
1147                     *qmant2_ptr += v;
1148                     mant2_cnt = 0;
1149                     v = 128;
1150                     break;
1151                 }
1152                 break;
1153             case 3:
1154                 v = sym_quant(c, e, 7);
1155                 break;
1156             case 4:
1157                 v = sym_quant(c, e, 11);
1158                 switch(mant4_cnt) {
1159                 case 0:
1160                     qmant4_ptr = &qmant[ch][i];
1161                     v = 11 * v;
1162                     mant4_cnt = 1;
1163                     break;
1164                 default:
1165                     *qmant4_ptr += v;
1166                     mant4_cnt = 0;
1167                     v = 128;
1168                     break;
1169                 }
1170                 break;
1171             case 5:
1172                 v = sym_quant(c, e, 15);
1173                 break;
1174             case 14:
1175                 v = asym_quant(c, e, 14);
1176                 break;
1177             case 15:
1178                 v = asym_quant(c, e, 16);
1179                 break;
1180             default:
1181                 v = asym_quant(c, e, b - 1);
1182                 break;
1183             }
1184             qmant[ch][i] = v;
1185         }
1186     }
1187
1188     /* second pass : output the values */
1189     for (ch = 0; ch < s->nb_all_channels; ch++) {
1190         int b, q;
1191
1192         for(i=0;i<s->nb_coefs[ch];i++) {
1193             q = qmant[ch][i];
1194             b = bap[ch][i];
1195             switch(b) {
1196             case 0:
1197                 break;
1198             case 1:
1199                 if (q != 128)
1200                     put_bits(&s->pb, 5, q);
1201                 break;
1202             case 2:
1203                 if (q != 128)
1204                     put_bits(&s->pb, 7, q);
1205                 break;
1206             case 3:
1207                 put_bits(&s->pb, 3, q);
1208                 break;
1209             case 4:
1210                 if (q != 128)
1211                     put_bits(&s->pb, 7, q);
1212                 break;
1213             case 14:
1214                 put_bits(&s->pb, 14, q);
1215                 break;
1216             case 15:
1217                 put_bits(&s->pb, 16, q);
1218                 break;
1219             default:
1220                 put_bits(&s->pb, b - 1, q);
1221                 break;
1222             }
1223         }
1224     }
1225 }
1226
1227 /* compute the ac3 crc */
1228
1229 #define CRC16_POLY ((1 << 0) | (1 << 2) | (1 << 15) | (1 << 16))
1230
1231 static void ac3_crc_init(void)
1232 {
1233     unsigned int c, n, k;
1234
1235     for(n=0;n<256;n++) {
1236         c = n << 8;
1237         for (k = 0; k < 8; k++) {
1238             if (c & (1 << 15))
1239                 c = ((c << 1) & 0xffff) ^ (CRC16_POLY & 0xffff);
1240             else
1241                 c = c << 1;
1242         }
1243         crc_table[n] = c;
1244     }
1245 }
1246
1247 static unsigned int ac3_crc(UINT8 *data, int n, unsigned int crc)
1248 {
1249     int i;
1250     for(i=0;i<n;i++) {
1251         crc = (crc_table[data[i] ^ (crc >> 8)] ^ (crc << 8)) & 0xffff;
1252     }
1253     return crc;
1254 }
1255
1256 static unsigned int mul_poly(unsigned int a, unsigned int b, unsigned int poly)
1257 {
1258     unsigned int c;
1259
1260     c = 0;
1261     while (a) {
1262         if (a & 1)
1263             c ^= b;
1264         a = a >> 1;
1265         b = b << 1;
1266         if (b & (1 << 16))
1267             b ^= poly;
1268     }
1269     return c;
1270 }
1271
1272 static unsigned int pow_poly(unsigned int a, unsigned int n, unsigned int poly)
1273 {
1274     unsigned int r;
1275     r = 1;
1276     while (n) {
1277         if (n & 1)
1278             r = mul_poly(r, a, poly);
1279         a = mul_poly(a, a, poly);
1280         n >>= 1;
1281     }
1282     return r;
1283 }
1284
1285
1286 /* compute log2(max(abs(tab[]))) */
1287 static int log2_tab(INT16 *tab, int n)
1288 {
1289     int i, v;
1290
1291     v = 0;
1292     for(i=0;i<n;i++) {
1293         v |= abs(tab[i]);
1294     }
1295     return av_log2(v);
1296 }
1297
1298 static void lshift_tab(INT16 *tab, int n, int lshift)
1299 {
1300     int i;
1301
1302     if (lshift > 0) {
1303         for(i=0;i<n;i++) {
1304             tab[i] <<= lshift;
1305         }
1306     } else if (lshift < 0) {
1307         lshift = -lshift;
1308         for(i=0;i<n;i++) {
1309             tab[i] >>= lshift;
1310         }
1311     }
1312 }
1313
1314 /* fill the end of the frame and compute the two crcs */
1315 static int output_frame_end(AC3EncodeContext *s)
1316 {
1317     int frame_size, frame_size_58, n, crc1, crc2, crc_inv;
1318     UINT8 *frame;
1319
1320     frame_size = s->frame_size; /* frame size in words */
1321     /* align to 8 bits */
1322     flush_put_bits(&s->pb);
1323     /* add zero bytes to reach the frame size */
1324     frame = s->pb.buf;
1325     n = 2 * s->frame_size - (pbBufPtr(&s->pb) - frame) - 2;
1326     assert(n >= 0);
1327     memset(pbBufPtr(&s->pb), 0, n);
1328
1329     /* Now we must compute both crcs : this is not so easy for crc1
1330        because it is at the beginning of the data... */
1331     frame_size_58 = (frame_size >> 1) + (frame_size >> 3);
1332     crc1 = ac3_crc(frame + 4, (2 * frame_size_58) - 4, 0);
1333     /* XXX: could precompute crc_inv */
1334     crc_inv = pow_poly((CRC16_POLY >> 1), (16 * frame_size_58) - 16, CRC16_POLY);
1335     crc1 = mul_poly(crc_inv, crc1, CRC16_POLY);
1336     frame[2] = crc1 >> 8;
1337     frame[3] = crc1;
1338
1339     crc2 = ac3_crc(frame + 2 * frame_size_58, (frame_size - frame_size_58) * 2 - 2, 0);
1340     frame[2*frame_size - 2] = crc2 >> 8;
1341     frame[2*frame_size - 1] = crc2;
1342
1343     //    printf("n=%d frame_size=%d\n", n, frame_size);
1344     return frame_size * 2;
1345 }
1346
1347 static int AC3_encode_frame(AVCodecContext *avctx,
1348                             unsigned char *frame, int buf_size, void *data)
1349 {
1350     AC3EncodeContext *s = avctx->priv_data;
1351     short *samples = data;
1352     int i, j, k, v, ch;
1353     INT16 input_samples[N];
1354     INT32 mdct_coef[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1355     UINT8 exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1356     UINT8 exp_strategy[NB_BLOCKS][AC3_MAX_CHANNELS];
1357     UINT8 encoded_exp[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1358     UINT8 bap[NB_BLOCKS][AC3_MAX_CHANNELS][N/2];
1359     INT8 exp_samples[NB_BLOCKS][AC3_MAX_CHANNELS];
1360     int frame_bits;
1361
1362     frame_bits = 0;
1363     for(ch=0;ch<s->nb_all_channels;ch++) {
1364         /* fixed mdct to the six sub blocks & exponent computation */
1365         for(i=0;i<NB_BLOCKS;i++) {
1366             INT16 *sptr;
1367             int sinc;
1368
1369             /* compute input samples */
1370             memcpy(input_samples, s->last_samples[ch], N/2 * sizeof(INT16));
1371             sinc = s->nb_all_channels;
1372             sptr = samples + (sinc * (N/2) * i) + ch;
1373             for(j=0;j<N/2;j++) {
1374                 v = *sptr;
1375                 input_samples[j + N/2] = v;
1376                 s->last_samples[ch][j] = v;
1377                 sptr += sinc;
1378             }
1379
1380             /* apply the MDCT window */
1381             for(j=0;j<N/2;j++) {
1382                 input_samples[j] = MUL16(input_samples[j],
1383                                          ac3_window[j]) >> 15;
1384                 input_samples[N-j-1] = MUL16(input_samples[N-j-1],
1385                                              ac3_window[j]) >> 15;
1386             }
1387
1388             /* Normalize the samples to use the maximum available
1389                precision */
1390             v = 14 - log2_tab(input_samples, N);
1391             if (v < 0)
1392                 v = 0;
1393             exp_samples[i][ch] = v - 8;
1394             lshift_tab(input_samples, N, v);
1395
1396             /* do the MDCT */
1397             mdct512(mdct_coef[i][ch], input_samples);
1398
1399             /* compute "exponents". We take into account the
1400                normalization there */
1401             for(j=0;j<N/2;j++) {
1402                 int e;
1403                 v = abs(mdct_coef[i][ch][j]);
1404                 if (v == 0)
1405                     e = 24;
1406                 else {
1407                     e = 23 - av_log2(v) + exp_samples[i][ch];
1408                     if (e >= 24) {
1409                         e = 24;
1410                         mdct_coef[i][ch][j] = 0;
1411                     }
1412                 }
1413                 exp[i][ch][j] = e;
1414             }
1415         }
1416
1417         compute_exp_strategy(exp_strategy, exp, ch, ch == s->lfe_channel);
1418
1419         /* compute the exponents as the decoder will see them. The
1420            EXP_REUSE case must be handled carefully : we select the
1421            min of the exponents */
1422         i = 0;
1423         while (i < NB_BLOCKS) {
1424             j = i + 1;
1425             while (j < NB_BLOCKS && exp_strategy[j][ch] == EXP_REUSE) {
1426                 exponent_min(exp[i][ch], exp[j][ch], s->nb_coefs[ch]);
1427                 j++;
1428             }
1429             frame_bits += encode_exp(encoded_exp[i][ch],
1430                                      exp[i][ch], s->nb_coefs[ch],
1431                                      exp_strategy[i][ch]);
1432             /* copy encoded exponents for reuse case */
1433             for(k=i+1;k<j;k++) {
1434                 memcpy(encoded_exp[k][ch], encoded_exp[i][ch],
1435                        s->nb_coefs[ch] * sizeof(UINT8));
1436             }
1437             i = j;
1438         }
1439     }
1440
1441     compute_bit_allocation(s, bap, encoded_exp, exp_strategy, frame_bits);
1442     /* everything is known... let's output the frame */
1443     output_frame_header(s, frame);
1444
1445     for(i=0;i<NB_BLOCKS;i++) {
1446         output_audio_block(s, exp_strategy[i], encoded_exp[i],
1447                            bap[i], mdct_coef[i], exp_samples[i], i);
1448     }
1449     return output_frame_end(s);
1450 }
1451
1452 static int AC3_encode_close(AVCodecContext *avctx)
1453 {
1454     av_freep(&avctx->coded_frame);
1455 }
1456
1457 #if 0
1458 /*************************************************************************/
1459 /* TEST */
1460
1461 #define FN (N/4)
1462
1463 void fft_test(void)
1464 {
1465     IComplex in[FN], in1[FN];
1466     int k, n, i;
1467     float sum_re, sum_im, a;
1468
1469     /* FFT test */
1470
1471     for(i=0;i<FN;i++) {
1472         in[i].re = random() % 65535 - 32767;
1473         in[i].im = random() % 65535 - 32767;
1474         in1[i] = in[i];
1475     }
1476     fft(in, 7);
1477
1478     /* do it by hand */
1479     for(k=0;k<FN;k++) {
1480         sum_re = 0;
1481         sum_im = 0;
1482         for(n=0;n<FN;n++) {
1483             a = -2 * M_PI * (n * k) / FN;
1484             sum_re += in1[n].re * cos(a) - in1[n].im * sin(a);
1485             sum_im += in1[n].re * sin(a) + in1[n].im * cos(a);
1486         }
1487         printf("%3d: %6d,%6d %6.0f,%6.0f\n",
1488                k, in[k].re, in[k].im, sum_re / FN, sum_im / FN);
1489     }
1490 }
1491
1492 void mdct_test(void)
1493 {
1494     INT16 input[N];
1495     INT32 output[N/2];
1496     float input1[N];
1497     float output1[N/2];
1498     float s, a, err, e, emax;
1499     int i, k, n;
1500
1501     for(i=0;i<N;i++) {
1502         input[i] = (random() % 65535 - 32767) * 9 / 10;
1503         input1[i] = input[i];
1504     }
1505
1506     mdct512(output, input);
1507
1508     /* do it by hand */
1509     for(k=0;k<N/2;k++) {
1510         s = 0;
1511         for(n=0;n<N;n++) {
1512             a = (2*M_PI*(2*n+1+N/2)*(2*k+1) / (4 * N));
1513             s += input1[n] * cos(a);
1514         }
1515         output1[k] = -2 * s / N;
1516     }
1517
1518     err = 0;
1519     emax = 0;
1520     for(i=0;i<N/2;i++) {
1521         printf("%3d: %7d %7.0f\n", i, output[i], output1[i]);
1522         e = output[i] - output1[i];
1523         if (e > emax)
1524             emax = e;
1525         err += e * e;
1526     }
1527     printf("err2=%f emax=%f\n", err / (N/2), emax);
1528 }
1529
1530 void test_ac3(void)
1531 {
1532     AC3EncodeContext ctx;
1533     unsigned char frame[AC3_MAX_CODED_FRAME_SIZE];
1534     short samples[AC3_FRAME_SIZE];
1535     int ret, i;
1536
1537     AC3_encode_init(&ctx, 44100, 64000, 1);
1538
1539     fft_test();
1540     mdct_test();
1541
1542     for(i=0;i<AC3_FRAME_SIZE;i++)
1543         samples[i] = (int)(sin(2*M_PI*i*1000.0/44100) * 10000);
1544     ret = AC3_encode_frame(&ctx, frame, samples);
1545     printf("ret=%d\n", ret);
1546 }
1547 #endif
1548
1549 AVCodec ac3_encoder = {
1550     "ac3",
1551     CODEC_TYPE_AUDIO,
1552     CODEC_ID_AC3,
1553     sizeof(AC3EncodeContext),
1554     AC3_encode_init,
1555     AC3_encode_frame,
1556     AC3_encode_close,
1557     NULL,
1558 };