dlls/winemp3.acm/layer3.c

   1 /*
   2  * Mpeg Layer-3 audio decoder
   3  * --------------------------
   4  * copyright (c) 1995,1996,1997 by Michael Hipp.
   5  * All rights reserved. See also 'README'
   6  * This file has been copied from mpglib.
   7  *
   8  * This library is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * This library is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with this library; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  21  */
  22
  23 #include "config.h"
  24 #include <stdlib.h>
  25 #include "wine/debug.h"
  26 #include "mpg123.h"
  27 #include "mpglib.h"
  28 #include "huffman.h"
  29
  30 #define MPEG1
  31 WINE_DEFAULT_DEBUG_CHANNEL(mpeg3);
  32
  33 static real ispow[8207];
  34 static real aa_ca[8],aa_cs[8];
  35 static real COS1[12][6];
  36 static real win[4][36];
  37 static real win1[4][36];
  38 static real gainpow2[256+118+4];
  39 static real COS9[9];
  40 static real COS6_1,COS6_2;
  41 static real tfcos36[9];
  42 static real tfcos12[3];
  43
  44 struct bandInfoStruct {
  45   short longIdx[23];
  46   short longDiff[22];
  47   short shortIdx[14];
  48   short shortDiff[13];
  49 };
  50
  51 static int longLimit[9][23];
  52 static int shortLimit[9][14];
  53
  54 static const struct bandInfoStruct bandInfo[9] = {
  55
  56 /* MPEG 1.0 */
  57  { {0,4,8,12,16,20,24,30,36,44,52,62,74, 90,110,134,162,196,238,288,342,418,576},
  58    {4,4,4,4,4,4,6,6,8, 8,10,12,16,20,24,28,34,42,50,54, 76,158},
  59    {0,4*3,8*3,12*3,16*3,22*3,30*3,40*3,52*3,66*3, 84*3,106*3,136*3,192*3},
  60    {4,4,4,4,6,8,10,12,14,18,22,30,56} } ,
  61
  62  { {0,4,8,12,16,20,24,30,36,42,50,60,72, 88,106,128,156,190,230,276,330,384,576},
  63    {4,4,4,4,4,4,6,6,6, 8,10,12,16,18,22,28,34,40,46,54, 54,192},
  64    {0,4*3,8*3,12*3,16*3,22*3,28*3,38*3,50*3,64*3, 80*3,100*3,126*3,192*3},
  65    {4,4,4,4,6,6,10,12,14,16,20,26,66} } ,
  66
  67  { {0,4,8,12,16,20,24,30,36,44,54,66,82,102,126,156,194,240,296,364,448,550,576} ,
  68    {4,4,4,4,4,4,6,6,8,10,12,16,20,24,30,38,46,56,68,84,102, 26} ,
  69    {0,4*3,8*3,12*3,16*3,22*3,30*3,42*3,58*3,78*3,104*3,138*3,180*3,192*3} ,
  70    {4,4,4,4,6,8,12,16,20,26,34,42,12} }  ,
  71
  72 /* MPEG 2.0 */
  73  { {0,6,12,18,24,30,36,44,54,66,80,96,116,140,168,200,238,284,336,396,464,522,576},
  74    {6,6,6,6,6,6,8,10,12,14,16,20,24,28,32,38,46,52,60,68,58,54 } ,
  75    {0,4*3,8*3,12*3,18*3,24*3,32*3,42*3,56*3,74*3,100*3,132*3,174*3,192*3} ,
  76    {4,4,4,6,6,8,10,14,18,26,32,42,18 } } ,
  77
  78  { {0,6,12,18,24,30,36,44,54,66,80,96,114,136,162,194,232,278,330,394,464,540,576},
  79    {6,6,6,6,6,6,8,10,12,14,16,18,22,26,32,38,46,52,64,70,76,36 } ,
  80    {0,4*3,8*3,12*3,18*3,26*3,36*3,48*3,62*3,80*3,104*3,136*3,180*3,192*3} ,
  81    {4,4,4,6,8,10,12,14,18,24,32,44,12 } } ,
  82
  83  { {0,6,12,18,24,30,36,44,54,66,80,96,116,140,168,200,238,284,336,396,464,522,576},
  84    {6,6,6,6,6,6,8,10,12,14,16,20,24,28,32,38,46,52,60,68,58,54 },
  85    {0,4*3,8*3,12*3,18*3,26*3,36*3,48*3,62*3,80*3,104*3,134*3,174*3,192*3},
  86    {4,4,4,6,8,10,12,14,18,24,30,40,18 } } ,
  87 /* MPEG 2.5 */
  88  { {0,6,12,18,24,30,36,44,54,66,80,96,116,140,168,200,238,284,336,396,464,522,576} ,
  89    {6,6,6,6,6,6,8,10,12,14,16,20,24,28,32,38,46,52,60,68,58,54},
  90    {0,12,24,36,54,78,108,144,186,240,312,402,522,576},
  91    {4,4,4,6,8,10,12,14,18,24,30,40,18} },
  92  { {0,6,12,18,24,30,36,44,54,66,80,96,116,140,168,200,238,284,336,396,464,522,576} ,
  93    {6,6,6,6,6,6,8,10,12,14,16,20,24,28,32,38,46,52,60,68,58,54},
  94    {0,12,24,36,54,78,108,144,186,240,312,402,522,576},
  95    {4,4,4,6,8,10,12,14,18,24,30,40,18} },
  96  { {0,12,24,36,48,60,72,88,108,132,160,192,232,280,336,400,476,566,568,570,572,574,576},
  97    {12,12,12,12,12,12,16,20,24,28,32,40,48,56,64,76,90,2,2,2,2,2},
  98    {0, 24, 48, 72,108,156,216,288,372,480,486,492,498,576},
  99    {8,8,8,12,16,20,24,28,36,2,2,2,26} } ,
 100 };
 101
 102 static int mapbuf0[9][152];
 103 static int mapbuf1[9][156];
 104 static int mapbuf2[9][44];
 105 static int *map[9][3];
 106 static int *mapend[9][3];
 107
 108 static unsigned int n_slen2[512]; /* MPEG 2.0 slen for 'normal' mode */
 109 static unsigned int i_slen2[256]; /* MPEG 2.0 slen for intensity stereo */
 110
 111 static real tan1_1[16],tan2_1[16],tan1_2[16],tan2_2[16];
 112 static real pow1_1[2][16],pow2_1[2][16],pow1_2[2][16],pow2_2[2][16];
 113
 114 /*
 115  * init tables for layer-3
 116  */
 117 void init_layer3(int down_sample_sblimit)
 118 {
 119   int i,j,k,l;
 120
 121   for(i=-256;i<118+4;i++)
 122     gainpow2[i+256] = pow((double)2.0,-0.25 * (double) (i+210) );
 123
 124   for(i=0;i<8207;i++)
 125     ispow[i] = pow((double)i,(double)4.0/3.0);
 126
 127   for (i=0;i<8;i++)
 128   {
 129     static const double Ci[8]={-0.6,-0.535,-0.33,-0.185,-0.095,-0.041,-0.0142,-0.0037};
 130     double sq=sqrt(1.0+Ci[i]*Ci[i]);
 131     aa_cs[i] = 1.0/sq;
 132     aa_ca[i] = Ci[i]/sq;
 133   }
 134
 135   for(i=0;i<18;i++)
 136   {
 137     win[0][i]    = win[1][i]    = 0.5 * sin( M_PI / 72.0 * (double) (2*(i+0) +1) ) / cos ( M_PI * (double) (2*(i+0) +19) / 72.0 );
 138     win[0][i+18] = win[3][i+18] = 0.5 * sin( M_PI / 72.0 * (double) (2*(i+18)+1) ) / cos ( M_PI * (double) (2*(i+18)+19) / 72.0 );
 139   }
 140   for(i=0;i<6;i++)
 141   {
 142     win[1][i+18] = 0.5 / cos ( M_PI * (double) (2*(i+18)+19) / 72.0 );
 143     win[3][i+12] = 0.5 / cos ( M_PI * (double) (2*(i+12)+19) / 72.0 );
 144     win[1][i+24] = 0.5 * sin( M_PI / 24.0 * (double) (2*i+13) ) / cos ( M_PI * (double) (2*(i+24)+19) / 72.0 );
 145     win[1][i+30] = win[3][i] = 0.0;
 146     win[3][i+6 ] = 0.5 * sin( M_PI / 24.0 * (double) (2*i+1) )  / cos ( M_PI * (double) (2*(i+6 )+19) / 72.0 );
 147   }
 148
 149   for(i=0;i<9;i++)
 150     COS9[i] = cos( M_PI / 18.0 * (double) i);
 151
 152   for(i=0;i<9;i++)
 153     tfcos36[i] = 0.5 / cos ( M_PI * (double) (i*2+1) / 36.0 );
 154   for(i=0;i<3;i++)
 155     tfcos12[i] = 0.5 / cos ( M_PI * (double) (i*2+1) / 12.0 );
 156
 157   COS6_1 = cos( M_PI / 6.0 * (double) 1);
 158   COS6_2 = cos( M_PI / 6.0 * (double) 2);
 159
 160   for(i=0;i<12;i++)
 161   {
 162     win[2][i]  = 0.5 * sin( M_PI / 24.0 * (double) (2*i+1) ) / cos ( M_PI * (double) (2*i+7) / 24.0 );
 163     for(j=0;j<6;j++)
 164       COS1[i][j] = cos( M_PI / 24.0 * (double) ((2*i+7)*(2*j+1)) );
 165   }
 166
 167   for(j=0;j<4;j++) {
 168     static const int len[4] = { 36,36,12,36 };
 169     for(i=0;i<len[j];i+=2)
 170       win1[j][i] = + win[j][i];
 171     for(i=1;i<len[j];i+=2)
 172       win1[j][i] = - win[j][i];
 173   }
 174
 175   for(i=0;i<16;i++)
 176   {
 177     double t = tan( (double) i * M_PI / 12.0 );
 178     tan1_1[i] = t / (1.0+t);
 179     tan2_1[i] = 1.0 / (1.0 + t);
 180     tan1_2[i] = M_SQRT2 * t / (1.0+t);
 181     tan2_2[i] = M_SQRT2 / (1.0 + t);
 182
 183     for(j=0;j<2;j++) {
 184       double base = pow(2.0,-0.25*(j+1.0));
 185       double p1=1.0,p2=1.0;
 186       if(i > 0) {
 187         if( i & 1 )
 188           p1 = pow(base,(i+1.0)*0.5);
 189         else
 190           p2 = pow(base,i*0.5);
 191       }
 192       pow1_1[j][i] = p1;
 193       pow2_1[j][i] = p2;
 194       pow1_2[j][i] = M_SQRT2 * p1;
 195       pow2_2[j][i] = M_SQRT2 * p2;
 196     }
 197   }
 198
 199   for(j=0;j<9;j++)
 200   {
 201    const struct bandInfoStruct *bi = &bandInfo[j];
 202    int *mp;
 203    int cb,lwin;
 204    const short *bdf;
 205
 206    mp = map[j][0] = mapbuf0[j];
 207    bdf = bi->longDiff;
 208    for(i=0,cb = 0; cb < 8 ; cb++,i+=*bdf++) {
 209      *mp++ = (*bdf) >> 1;
 210      *mp++ = i;
 211      *mp++ = 3;
 212      *mp++ = cb;
 213    }
 214    bdf = bi->shortDiff+3;
 215    for(cb=3;cb<13;cb++) {
 216      int l = (*bdf++) >> 1;
 217      for(lwin=0;lwin<3;lwin++) {
 218        *mp++ = l;
 219        *mp++ = i + lwin;
 220        *mp++ = lwin;
 221        *mp++ = cb;
 222      }
 223      i += 6*l;
 224    }
 225    mapend[j][0] = mp;
 226
 227    mp = map[j][1] = mapbuf1[j];
 228    bdf = bi->shortDiff+0;
 229    for(i=0,cb=0;cb<13;cb++) {
 230      int l = (*bdf++) >> 1;
 231      for(lwin=0;lwin<3;lwin++) {
 232        *mp++ = l;
 233        *mp++ = i + lwin;
 234        *mp++ = lwin;
 235        *mp++ = cb;
 236      }
 237      i += 6*l;
 238    }
 239    mapend[j][1] = mp;
 240
 241    mp = map[j][2] = mapbuf2[j];
 242    bdf = bi->longDiff;
 243    for(cb = 0; cb < 22 ; cb++) {
 244      *mp++ = (*bdf++) >> 1;
 245      *mp++ = cb;
 246    }
 247    mapend[j][2] = mp;
 248
 249   }
 250
 251   for(j=0;j<9;j++) {
 252     for(i=0;i<23;i++) {
 253       longLimit[j][i] = (bandInfo[j].longIdx[i] - 1 + 8) / 18 + 1;
 254       if(longLimit[j][i] > (down_sample_sblimit) )
 255         longLimit[j][i] = down_sample_sblimit;
 256     }
 257     for(i=0;i<14;i++) {
 258       shortLimit[j][i] = (bandInfo[j].shortIdx[i] - 1) / 18 + 1;
 259       if(shortLimit[j][i] > (down_sample_sblimit) )
 260         shortLimit[j][i] = down_sample_sblimit;
 261     }
 262   }
 263
 264   for(i=0;i<5;i++) {
 265     for(j=0;j<6;j++) {
 266       for(k=0;k<6;k++) {
 267         int n = k + j * 6 + i * 36;
 268         i_slen2[n] = i|(j<<3)|(k<<6)|(3<<12);
 269       }
 270     }
 271   }
 272   for(i=0;i<4;i++) {
 273     for(j=0;j<4;j++) {
 274       for(k=0;k<4;k++) {
 275         int n = k + j * 4 + i * 16;
 276         i_slen2[n+180] = i|(j<<3)|(k<<6)|(4<<12);
 277       }
 278     }
 279   }
 280   for(i=0;i<4;i++) {
 281     for(j=0;j<3;j++) {
 282       int n = j + i * 3;
 283       i_slen2[n+244] = i|(j<<3) | (5<<12);
 284       n_slen2[n+500] = i|(j<<3) | (2<<12) | (1<<15);
 285     }
 286   }
 287
 288   for(i=0;i<5;i++) {
 289     for(j=0;j<5;j++) {
 290       for(k=0;k<4;k++) {
 291         for(l=0;l<4;l++) {
 292           int n = l + k * 4 + j * 16 + i * 80;
 293           n_slen2[n] = i|(j<<3)|(k<<6)|(l<<9)|(0<<12);
 294         }
 295       }
 296     }
 297   }
 298   for(i=0;i<5;i++) {
 299     for(j=0;j<5;j++) {
 300       for(k=0;k<4;k++) {
 301         int n = k + j * 4 + i * 20;
 302         n_slen2[n+400] = i|(j<<3)|(k<<6)|(1<<12);
 303       }
 304     }
 305   }
 306 }
 307
 308 /*
 309  * read additional side information
 310  */
 311 #ifdef MPEG1
 312 static int III_get_side_info_1(struct III_sideinfo *si,int stereo,
 313  int ms_stereo,long sfreq,int single)
 314 {
 315    int ch, gr;
 316    int powdiff = (single == 3) ? 4 : 0;
 317
 318    si->main_data_begin = getbits(9);
 319    if (stereo == 1)
 320      si->private_bits = getbits_fast(5);
 321    else
 322      si->private_bits = getbits_fast(3);
 323
 324    for (ch=0; ch<stereo; ch++) {
 325        si->ch[ch].gr[0].scfsi = -1;
 326        si->ch[ch].gr[1].scfsi = getbits_fast(4);
 327    }
 328
 329    for (gr=0; gr<2; gr++)
 330    {
 331      for (ch=0; ch<stereo; ch++)
 332      {
 333        register struct gr_info_s *gr_info = &(si->ch[ch].gr[gr]);
 334
 335        gr_info->part2_3_length = getbits(12);
 336        gr_info->big_values = getbits_fast(9);
 337        if(gr_info->big_values > 288) {
 338           FIXME("big_values (%d) too large!\n", gr_info->big_values);
 339           gr_info->big_values = 288;
 340        }
 341        gr_info->pow2gain = gainpow2+256 - getbits_fast(8) + powdiff;
 342        if(ms_stereo)
 343          gr_info->pow2gain += 2;
 344        gr_info->scalefac_compress = getbits_fast(4);
 345 /* window-switching flag == 1 for block_Type != 0 .. and block-type == 0 -> win-sw-flag = 0 */
 346        if(get1bit())
 347        {
 348          int i;
 349          gr_info->block_type = getbits_fast(2);
 350          gr_info->mixed_block_flag = get1bit();
 351          gr_info->table_select[0] = getbits_fast(5);
 352          gr_info->table_select[1] = getbits_fast(5);
 353          /*
 354           * table_select[2] not needed, because there is no region2,
 355           * but to satisfy some verifications tools we set it either.
 356           */
 357          gr_info->table_select[2] = 0;
 358          for(i=0;i<3;i++)
 359            gr_info->full_gain[i] = gr_info->pow2gain + (getbits_fast(3)<<3);
 360
 361          if(gr_info->block_type == 0) {
 362            FIXME("Blocktype == 0 and window-switching == 1 not allowed.\n");
 363            return 0;
 364          }
 365          /* region_count/start parameters are implicit in this case. */
 366          gr_info->region1start = 36>>1;
 367          gr_info->region2start = 576>>1;
 368        }
 369        else
 370        {
 371          int i,r0c,r1c;
 372          for (i=0; i<3; i++)
 373            gr_info->table_select[i] = getbits_fast(5);
 374          r0c = getbits_fast(4);
 375          r1c = getbits_fast(3);
 376          gr_info->region1start = bandInfo[sfreq].longIdx[r0c+1] >> 1 ;
 377          gr_info->region2start = bandInfo[sfreq].longIdx[r0c+1+r1c+1] >> 1;
 378          gr_info->block_type = 0;
 379          gr_info->mixed_block_flag = 0;
 380        }
 381        gr_info->preflag = get1bit();
 382        gr_info->scalefac_scale = get1bit();
 383        gr_info->count1table_select = get1bit();
 384      }
 385    }
 386    return !0;
 387 }
 388 #endif
 389
 390 /*
 391  * Side Info for MPEG 2.0 / LSF
 392  */
 393 static int III_get_side_info_2(struct III_sideinfo *si,int stereo,
 394  int ms_stereo,long sfreq,int single)
 395 {
 396    int ch;
 397    int powdiff = (single == 3) ? 4 : 0;
 398
 399    si->main_data_begin = getbits(8);
 400    if (stereo == 1)
 401      si->private_bits = get1bit();
 402    else
 403      si->private_bits = getbits_fast(2);
 404
 405    for (ch=0; ch<stereo; ch++)
 406    {
 407        register struct gr_info_s *gr_info = &(si->ch[ch].gr[0]);
 408
 409        gr_info->part2_3_length = getbits(12);
 410        gr_info->big_values = getbits_fast(9);
 411        if(gr_info->big_values > 288) {
 412          FIXME("big_values(%d) too large!\n", gr_info->big_values);
 413          gr_info->big_values = 288;
 414        }
 415        gr_info->pow2gain = gainpow2+256 - getbits_fast(8) + powdiff;
 416        if(ms_stereo)
 417          gr_info->pow2gain += 2;
 418        gr_info->scalefac_compress = getbits(9);
 419 /* window-switching flag == 1 for block_Type != 0 .. and block-type == 0 -> win-sw-flag = 0 */
 420        if(get1bit())
 421        {
 422          int i;
 423          gr_info->block_type = getbits_fast(2);
 424          gr_info->mixed_block_flag = get1bit();
 425          gr_info->table_select[0] = getbits_fast(5);
 426          gr_info->table_select[1] = getbits_fast(5);
 427          /*
 428           * table_select[2] not needed, because there is no region2,
 429           * but to satisfy some verifications tools we set it either.
 430           */
 431          gr_info->table_select[2] = 0;
 432          for(i=0;i<3;i++)
 433            gr_info->full_gain[i] = gr_info->pow2gain + (getbits_fast(3)<<3);
 434
 435          if(gr_info->block_type == 0) {
 436            FIXME("Blocktype == 0 and window-switching == 1 not allowed.\n");
 437            return 0;
 438          }
 439          /* region_count/start parameters are implicit in this case. */
 440 /* check this again! */
 441          if(gr_info->block_type == 2)
 442            gr_info->region1start = 36>>1;
 443          else if(sfreq == 8)
 444 /* check this for 2.5 and sfreq=8 */
 445            gr_info->region1start = 108>>1;
 446          else
 447            gr_info->region1start = 54>>1;
 448          gr_info->region2start = 576>>1;
 449        }
 450        else
 451        {
 452          int i,r0c,r1c;
 453          for (i=0; i<3; i++)
 454            gr_info->table_select[i] = getbits_fast(5);
 455          r0c = getbits_fast(4);
 456          r1c = getbits_fast(3);
 457          gr_info->region1start = bandInfo[sfreq].longIdx[r0c+1] >> 1 ;
 458          gr_info->region2start = bandInfo[sfreq].longIdx[r0c+1+r1c+1] >> 1;
 459          gr_info->block_type = 0;
 460          gr_info->mixed_block_flag = 0;
 461        }
 462        gr_info->scalefac_scale = get1bit();
 463        gr_info->count1table_select = get1bit();
 464    }
 465    return !0;
 466 }
 467
 468 /*
 469  * read scalefactors
 470  */
 471 #ifdef MPEG1
 472 static int III_get_scale_factors_1(int *scf,struct gr_info_s *gr_info)
 473 {
 474    static const unsigned char slen[2][16] = {
 475      {0, 0, 0, 0, 3, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4},
 476      {0, 1, 2, 3, 0, 1, 2, 3, 1, 2, 3, 1, 2, 3, 2, 3}
 477    };
 478    int numbits;
 479    int num0 = slen[0][gr_info->scalefac_compress];
 480    int num1 = slen[1][gr_info->scalefac_compress];
 481
 482     if (gr_info->block_type == 2) {
 483       int i=18;
 484       numbits = (num0 + num1) * 18;
 485
 486       if (gr_info->mixed_block_flag) {
 487          for (i=8;i;i--)
 488            *scf++ = getbits_fast(num0);
 489          i = 9;
 490          numbits -= num0; /* num0 * 17 + num1 * 18 */
 491       }
 492
 493       for (;i;i--)
 494         *scf++ = getbits_fast(num0);
 495       for (i = 18; i; i--)
 496         *scf++ = getbits_fast(num1);
 497       *scf++ = 0; *scf++ = 0; *scf++ = 0; /* short[13][0..2] = 0 */
 498     }
 499     else {
 500       int i;
 501       int scfsi = gr_info->scfsi;
 502
 503       if(scfsi < 0) { /* scfsi < 0 => granule == 0 */
 504          for(i=11;i;i--)
 505            *scf++ = getbits_fast(num0);
 506          for(i=10;i;i--)
 507            *scf++ = getbits_fast(num1);
 508          numbits = (num0 + num1) * 10 + num0;
 509          *scf++ = 0;
 510       }
 511       else {
 512         numbits = 0;
 513         if(!(scfsi & 0x8)) {
 514           for (i=0;i<6;i++)
 515             *scf++ = getbits_fast(num0);
 516           numbits += num0 * 6;
 517         }
 518         else {
 519           scf += 6;
 520         }
 521
 522         if(!(scfsi & 0x4)) {
 523           for (i=0;i<5;i++)
 524             *scf++ = getbits_fast(num0);
 525           numbits += num0 * 5;
 526         }
 527         else {
 528           scf += 5;
 529         }
 530
 531         if(!(scfsi & 0x2)) {
 532           for(i=0;i<5;i++)
 533             *scf++ = getbits_fast(num1);
 534           numbits += num1 * 5;
 535         }
 536         else {
 537           scf += 5;
 538         }
 539
 540         if(!(scfsi & 0x1)) {
 541           for (i=0;i<5;i++)
 542             *scf++ = getbits_fast(num1);
 543           numbits += num1 * 5;
 544         }
 545         else {
 546            scf += 5;
 547         }
 548         *scf++ = 0;  /* no l[21] in original sources */
 549       }
 550     }
 551     return numbits;
 552 }
 553 #endif
 554
 555
 556 static int III_get_scale_factors_2(int *scf,struct gr_info_s *gr_info,int i_stereo)
 557 {
 558   const unsigned char *pnt;
 559   int i,j;
 560   unsigned int slen;
 561   int n = 0;
 562   int numbits = 0;
 563
 564   static const unsigned char stab[3][6][4] = {
 565    { { 6, 5, 5,5 } , { 6, 5, 7,3 } , { 11,10,0,0} ,
 566      { 7, 7, 7,0 } , { 6, 6, 6,3 } , {  8, 8,5,0} } ,
 567    { { 9, 9, 9,9 } , { 9, 9,12,6 } , { 18,18,0,0} ,
 568      {12,12,12,0 } , {12, 9, 9,6 } , { 15,12,9,0} } ,
 569    { { 6, 9, 9,9 } , { 6, 9,12,6 } , { 15,18,0,0} ,
 570      { 6,15,12,0 } , { 6,12, 9,6 } , {  6,18,9,0} } };
 571
 572   if(i_stereo) /* i_stereo AND second channel -> do_layer3() checks this */
 573     slen = i_slen2[gr_info->scalefac_compress>>1];
 574   else
 575     slen = n_slen2[gr_info->scalefac_compress];
 576
 577   gr_info->preflag = (slen>>15) & 0x1;
 578
 579   n = 0;
 580   if( gr_info->block_type == 2 ) {
 581     n++;
 582     if(gr_info->mixed_block_flag)
 583       n++;
 584   }
 585
 586   pnt = stab[n][(slen>>12)&0x7];
 587
 588   for(i=0;i<4;i++) {
 589     int num = slen & 0x7;
 590     slen >>= 3;
 591     if(num) {
 592       for(j=0;j<(int)(pnt[i]);j++)
 593         *scf++ = getbits_fast(num);
 594       numbits += pnt[i] * num;
 595     }
 596     else {
 597       for(j=0;j<(int)(pnt[i]);j++)
 598         *scf++ = 0;
 599     }
 600   }
 601
 602   n = (n << 1) + 1;
 603   for(i=0;i<n;i++)
 604     *scf++ = 0;
 605
 606   return numbits;
 607 }
 608
 609 static const int pretab1[22] = {0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,2,2,3,3,3,2,0};
 610 static const int pretab2[22] = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
 611
 612 static int III_dequantize_sample(real xr[SBLIMIT][SSLIMIT],int *scf,
 613    struct gr_info_s *gr_info,int sfreq,int part2bits)
 614 {
 615   int shift = 1 + gr_info->scalefac_scale;
 616   real *xrpnt = (real *) xr;
 617   int l[3],l3;
 618   int part2remain = gr_info->part2_3_length - part2bits;
 619   int *me;
 620
 621   {
 622     int bv       = gr_info->big_values;
 623     int region1  = gr_info->region1start;
 624     int region2  = gr_info->region2start;
 625
 626     l3 = ((576>>1)-bv)>>1;
 627 /*
 628  * we may lose the 'odd' bit here !!
 629  * check this later again
 630  */
 631     if(bv <= region1) {
 632       l[0] = bv; l[1] = 0; l[2] = 0;
 633     }
 634     else {
 635       l[0] = region1;
 636       if(bv <= region2) {
 637         l[1] = bv - l[0];  l[2] = 0;
 638       }
 639       else {
 640         l[1] = region2 - l[0]; l[2] = bv - region2;
 641       }
 642     }
 643   }
 644
 645   if(gr_info->block_type == 2) {
 646     /*
 647      * decoding with short or mixed mode BandIndex table
 648      */
 649     int i,max[4];
 650     int step=0,lwin=0,cb=0;
 651     register real v = 0.0;
 652     register int *m,mc;
 653
 654     if(gr_info->mixed_block_flag) {
 655       max[3] = -1;
 656       max[0] = max[1] = max[2] = 2;
 657       m = map[sfreq][0];
 658       me = mapend[sfreq][0];
 659     }
 660     else {
 661       max[0] = max[1] = max[2] = max[3] = -1;
 662       /* max[3] not really needed in this case */
 663       m = map[sfreq][1];
 664       me = mapend[sfreq][1];
 665     }
 666
 667     mc = 0;
 668     for(i=0;i<2;i++) {
 669       int lp = l[i];
 670       const struct newhuff *h = ht+gr_info->table_select[i];
 671       for(;lp;lp--,mc--) {
 672         register int x,y;
 673         if( (!mc) ) {
 674           mc = *m++;
 675           xrpnt = ((real *) xr) + (*m++);
 676           lwin = *m++;
 677           cb = *m++;
 678           if(lwin == 3) {
 679             v = gr_info->pow2gain[(*scf++) << shift];
 680             step = 1;
 681           }
 682           else {
 683             v = gr_info->full_gain[lwin][(*scf++) << shift];
 684             step = 3;
 685           }
 686         }
 687         {
 688           register const short *val = h->table;
 689           while((y=*val++)<0) {
 690             if (get1bit())
 691               val -= y;
 692             part2remain--;
 693           }
 694           x = y >> 4;
 695           y &= 0xf;
 696         }
 697         if(x == 15) {
 698           max[lwin] = cb;
 699           part2remain -= h->linbits+1;
 700           x += getbits(h->linbits);
 701           if(get1bit())
 702             *xrpnt = -ispow[x] * v;
 703           else
 704             *xrpnt =  ispow[x] * v;
 705         }
 706         else if(x) {
 707           max[lwin] = cb;
 708           if(get1bit())
 709             *xrpnt = -ispow[x] * v;
 710           else
 711             *xrpnt =  ispow[x] * v;
 712           part2remain--;
 713         }
 714         else
 715           *xrpnt = 0.0;
 716         xrpnt += step;
 717         if(y == 15) {
 718           max[lwin] = cb;
 719           part2remain -= h->linbits+1;
 720           y += getbits(h->linbits);
 721           if(get1bit())
 722             *xrpnt = -ispow[y] * v;
 723           else
 724             *xrpnt =  ispow[y] * v;
 725         }
 726         else if(y) {
 727           max[lwin] = cb;
 728           if(get1bit())
 729             *xrpnt = -ispow[y] * v;
 730           else
 731             *xrpnt =  ispow[y] * v;
 732           part2remain--;
 733         }
 734         else
 735           *xrpnt = 0.0;
 736         xrpnt += step;
 737       }
 738     }
 739     for(;l3 && (part2remain > 0);l3--) {
 740       const struct newhuff *h = htc+gr_info->count1table_select;
 741       const short *val = h->table;
 742       short a;
 743
 744       while((a=*val++)<0) {
 745         part2remain--;
 746         if(part2remain < 0) {
 747           part2remain++;
 748           a = 0;
 749           break;
 750         }
 751         if (get1bit())
 752           val -= a;
 753       }
 754
 755       for(i=0;i<4;i++) {
 756         if(!(i & 1)) {
 757           if(!mc) {
 758             mc = *m++;
 759             xrpnt = ((real *) xr) + (*m++);
 760             lwin = *m++;
 761             cb = *m++;
 762             if(lwin == 3) {
 763               v = gr_info->pow2gain[(*scf++) << shift];
 764               step = 1;
 765             }
 766             else {
 767               v = gr_info->full_gain[lwin][(*scf++) << shift];
 768               step = 3;
 769             }
 770           }
 771           mc--;
 772         }
 773         if( (a & (0x8>>i)) ) {
 774           max[lwin] = cb;
 775           part2remain--;
 776           if(part2remain < 0) {
 777             part2remain++;
 778             break;
 779           }
 780           if(get1bit())
 781             *xrpnt = -v;
 782           else
 783             *xrpnt = v;
 784         }
 785         else
 786           *xrpnt = 0.0;
 787         xrpnt += step;
 788       }
 789     }
 790
 791     while( m < me ) {
 792       if(!mc) {
 793         mc = *m++;
 794         xrpnt = ((real *) xr) + *m++;
 795         if( (*m++) == 3)
 796           step = 1;
 797         else
 798           step = 3;
 799         m++; /* cb */
 800       }
 801       mc--;
 802       *xrpnt = 0.0;
 803       xrpnt += step;
 804       *xrpnt = 0.0;
 805       xrpnt += step;
 806 /* we could add a little opt. here:
 807  * if we finished a band for window 3 or a long band
 808  * further bands could copied in a simple loop without a
 809  * special 'map' decoding
 810  */
 811     }
 812
 813     gr_info->maxband[0] = max[0]+1;
 814     gr_info->maxband[1] = max[1]+1;
 815     gr_info->maxband[2] = max[2]+1;
 816     gr_info->maxbandl = max[3]+1;
 817
 818     {
 819       int rmax = max[0] > max[1] ? max[0] : max[1];
 820       rmax = (rmax > max[2] ? rmax : max[2]) + 1;
 821       gr_info->maxb = rmax ? shortLimit[sfreq][rmax] : longLimit[sfreq][max[3]+1];
 822     }
 823
 824   }
 825   else {
 826         /*
 827      * decoding with 'long' BandIndex table (block_type != 2)
 828      */
 829     const int *pretab = gr_info->preflag ? pretab1 : pretab2;
 830     int i,max = -1;
 831     int cb = 0;
 832     register int *m = map[sfreq][2];
 833     register real v = 0.0;
 834     register int mc = 0;
 835 #if 0
 836     me = mapend[sfreq][2];
 837 #endif
 838
 839         /*
 840      * long hash table values
 841      */
 842     for(i=0;i<3;i++) {
 843       int lp = l[i];
 844       const struct newhuff *h = ht+gr_info->table_select[i];
 845
 846       for(;lp;lp--,mc--) {
 847         int x,y;
 848
 849         if(!mc) {
 850           mc = *m++;
 851           v = gr_info->pow2gain[((*scf++) + (*pretab++)) << shift];
 852           cb = *m++;
 853         }
 854         {
 855           register const short *val = h->table;
 856           while((y=*val++)<0) {
 857             if (get1bit())
 858               val -= y;
 859             part2remain--;
 860           }
 861           x = y >> 4;
 862           y &= 0xf;
 863         }
 864         if (x == 15) {
 865           max = cb;
 866           part2remain -= h->linbits+1;
 867           x += getbits(h->linbits);
 868           if(get1bit())
 869             *xrpnt++ = -ispow[x] * v;
 870           else
 871             *xrpnt++ =  ispow[x] * v;
 872         }
 873         else if(x) {
 874           max = cb;
 875           if(get1bit())
 876             *xrpnt++ = -ispow[x] * v;
 877           else
 878             *xrpnt++ =  ispow[x] * v;
 879           part2remain--;
 880         }
 881         else
 882           *xrpnt++ = 0.0;
 883
 884         if (y == 15) {
 885           max = cb;
 886           part2remain -= h->linbits+1;
 887           y += getbits(h->linbits);
 888           if(get1bit())
 889             *xrpnt++ = -ispow[y] * v;
 890           else
 891             *xrpnt++ =  ispow[y] * v;
 892         }
 893         else if(y) {
 894           max = cb;
 895           if(get1bit())
 896             *xrpnt++ = -ispow[y] * v;
 897           else
 898             *xrpnt++ =  ispow[y] * v;
 899           part2remain--;
 900         }
 901         else
 902           *xrpnt++ = 0.0;
 903       }
 904     }
 905
 906         /*
 907      * short (count1table) values
 908      */
 909     for(;l3 && (part2remain > 0);l3--) {
 910       const struct newhuff *h = htc+gr_info->count1table_select;
 911       const short *val = h->table;
 912       short a;
 913
 914       while((a=*val++)<0) {
 915         part2remain--;
 916         if(part2remain < 0) {
 917           part2remain++;
 918           a = 0;
 919           break;
 920         }
 921         if (get1bit())
 922           val -= a;
 923       }
 924
 925       for(i=0;i<4;i++) {
 926         if(!(i & 1)) {
 927           if(!mc) {
 928             mc = *m++;
 929             cb = *m++;
 930             v = gr_info->pow2gain[((*scf++) + (*pretab++)) << shift];
 931           }
 932           mc--;
 933         }
 934         if ( (a & (0x8>>i)) ) {
 935           max = cb;
 936           part2remain--;
 937           if(part2remain < 0) {
 938             part2remain++;
 939             break;
 940           }
 941           if(get1bit())
 942             *xrpnt++ = -v;
 943           else
 944             *xrpnt++ = v;
 945         }
 946         else
 947           *xrpnt++ = 0.0;
 948       }
 949     }
 950
 951         /*
 952      * zero part
 953      */
 954     for(i=(&xr[SBLIMIT][0]-xrpnt)>>1;i;i--) {
 955       *xrpnt++ = 0.0;
 956       *xrpnt++ = 0.0;
 957     }
 958
 959     gr_info->maxbandl = max+1;
 960     gr_info->maxb = longLimit[sfreq][gr_info->maxbandl];
 961   }
 962
 963   while( part2remain > 16 ) {
 964     getbits(16); /* Dismiss stuffing Bits */
 965     part2remain -= 16;
 966   }
 967   if(part2remain > 0)
 968     getbits(part2remain);
 969   else if(part2remain < 0) {
 970     FIXME("mpg123: Can't rewind stream by %d bits!\n",-part2remain);
 971     return 1; /* -> error */
 972   }
 973   return 0;
 974 }
 975
 976 /*
 977  * III_stereo: calculate real channel values for Joint-I-Stereo-mode
 978  */
 979 static void III_i_stereo(real xr_buf[2][SBLIMIT][SSLIMIT],int *scalefac,
 980    struct gr_info_s *gr_info,int sfreq,int ms_stereo,int lsf)
 981 {
 982       real (*xr)[SBLIMIT*SSLIMIT] = (real (*)[SBLIMIT*SSLIMIT] ) xr_buf;
 983       const struct bandInfoStruct *bi = &bandInfo[sfreq];
 984       real *tab1,*tab2;
 985
 986       if(lsf) {
 987         int p = gr_info->scalefac_compress & 0x1;
 988             if(ms_stereo) {
 989           tab1 = pow1_2[p]; tab2 = pow2_2[p];
 990         }
 991         else {
 992           tab1 = pow1_1[p]; tab2 = pow2_1[p];
 993         }
 994       }
 995       else {
 996         if(ms_stereo) {
 997           tab1 = tan1_2; tab2 = tan2_2;
 998         }
 999         else {
1000           tab1 = tan1_1; tab2 = tan2_1;
1001         }
1002       }
1003
1004       if (gr_info->block_type == 2)
1005       {
1006          int lwin,do_l = 0;
1007          if( gr_info->mixed_block_flag )
1008            do_l = 1;
1009
1010          for (lwin=0;lwin<3;lwin++) /* process each window */
1011          {
1012              /* get first band with zero values */
1013            int is_p,sb,idx,sfb = gr_info->maxband[lwin];  /* sfb is minimal 3 for mixed mode */
1014            if(sfb > 3)
1015              do_l = 0;
1016
1017            for(;sfb<12;sfb++)
1018            {
1019              is_p = scalefac[sfb*3+lwin-gr_info->mixed_block_flag]; /* scale: 0-15 */
1020              if(is_p != 7) {
1021                real t1,t2;
1022                sb = bi->shortDiff[sfb];
1023                idx = bi->shortIdx[sfb] + lwin;
1024                t1 = tab1[is_p]; t2 = tab2[is_p];
1025                for (; sb > 0; sb--,idx+=3)
1026                {
1027                  real v = xr[0][idx];
1028                  xr[0][idx] = v * t1;
1029                  xr[1][idx] = v * t2;
1030                }
1031              }
1032            }
1033
1034 #if 1
1035 /* in the original: copy 10 to 11 , here: copy 11 to 12
1036 maybe still wrong??? (copy 12 to 13?) */
1037            is_p = scalefac[11*3+lwin-gr_info->mixed_block_flag]; /* scale: 0-15 */
1038            sb = bi->shortDiff[12];
1039            idx = bi->shortIdx[12] + lwin;
1040 #else
1041            is_p = scalefac[10*3+lwin-gr_info->mixed_block_flag]; /* scale: 0-15 */
1042            sb = bi->shortDiff[11];
1043            idx = bi->shortIdx[11] + lwin;
1044 #endif
1045            if(is_p != 7)
1046            {
1047              real t1,t2;
1048              t1 = tab1[is_p]; t2 = tab2[is_p];
1049              for ( ; sb > 0; sb--,idx+=3 )
1050              {
1051                real v = xr[0][idx];
1052                xr[0][idx] = v * t1;
1053                xr[1][idx] = v * t2;
1054              }
1055            }
1056          } /* end for(lwin; .. ; . ) */
1057
1058          if (do_l)
1059          {
1060 /* also check l-part, if ALL bands in the three windows are 'empty'
1061  * and mode = mixed_mode
1062  */
1063            int sfb = gr_info->maxbandl;
1064            int idx = bi->longIdx[sfb];
1065
1066            for ( ; sfb<8; sfb++ )
1067            {
1068              int sb = bi->longDiff[sfb];
1069              int is_p = scalefac[sfb]; /* scale: 0-15 */
1070              if(is_p != 7) {
1071                real t1,t2;
1072                t1 = tab1[is_p]; t2 = tab2[is_p];
1073                for ( ; sb > 0; sb--,idx++)
1074                {
1075                  real v = xr[0][idx];
1076                  xr[0][idx] = v * t1;
1077                  xr[1][idx] = v * t2;
1078                }
1079              }
1080              else
1081                idx += sb;
1082            }
1083          }
1084       }
1085       else /* ((gr_info->block_type != 2)) */
1086       {
1087         int sfb = gr_info->maxbandl;
1088         int is_p,idx = bi->longIdx[sfb];
1089         for ( ; sfb<21; sfb++)
1090         {
1091           int sb = bi->longDiff[sfb];
1092           is_p = scalefac[sfb]; /* scale: 0-15 */
1093           if(is_p != 7) {
1094             real t1,t2;
1095             t1 = tab1[is_p]; t2 = tab2[is_p];
1096             for ( ; sb > 0; sb--,idx++)
1097             {
1098                real v = xr[0][idx];
1099                xr[0][idx] = v * t1;
1100                xr[1][idx] = v * t2;
1101             }
1102           }
1103           else
1104             idx += sb;
1105         }
1106
1107         is_p = scalefac[20]; /* copy l-band 20 to l-band 21 */
1108         if(is_p != 7)
1109         {
1110           int sb;
1111           real t1 = tab1[is_p],t2 = tab2[is_p];
1112
1113           for ( sb = bi->longDiff[21]; sb > 0; sb--,idx++ )
1114           {
1115             real v = xr[0][idx];
1116             xr[0][idx] = v * t1;
1117             xr[1][idx] = v * t2;
1118           }
1119         }
1120       } /* ... */
1121 }
1122
1123 static void III_antialias(real xr[SBLIMIT][SSLIMIT],struct gr_info_s *gr_info)
1124 {
1125    int sblim;
1126
1127    if(gr_info->block_type == 2)
1128    {
1129       if(!gr_info->mixed_block_flag)
1130         return;
1131       sblim = 1;
1132    }
1133    else {
1134      sblim = gr_info->maxb-1;
1135    }
1136
1137    /* 31 alias-reduction operations between each pair of sub-bands */
1138    /* with 8 butterflies between each pair                         */
1139
1140    {
1141      int sb;
1142      real *xr1=(real *) xr[1];
1143
1144      for(sb=sblim;sb;sb--,xr1+=10)
1145      {
1146        int ss;
1147        real *cs=aa_cs,*ca=aa_ca;
1148        real *xr2 = xr1;
1149
1150        for(ss=7;ss>=0;ss--)
1151        {       /* upper and lower butterfly inputs */
1152          register real bu = *--xr2,bd = *xr1;
1153          *xr2   = (bu * (*cs)   ) - (bd * (*ca)   );
1154          *xr1++ = (bd * (*cs++) ) + (bu * (*ca++) );
1155        }
1156      }
1157   }
1158 }
1159
1160 /*
1161  DCT inspired by Jeff Tsay's DCT from the maplay package
1162  this is an optimized version with manual unroll.
1163
1164  References:
1165  [1] S. Winograd: "On Computing the Discrete Fourier Transform",
1166      Mathematics of Computation, Volume 32, Number 141, January 1978,
1167      Pages 175-199
1168 */
1169
1170 static void dct36(real *inbuf,real *o1,real *o2,real *wintab,real *tsbuf)
1171 {
1172   {
1173     register real *in = inbuf;
1174
1175     in[17]+=in[16]; in[16]+=in[15]; in[15]+=in[14];
1176     in[14]+=in[13]; in[13]+=in[12]; in[12]+=in[11];
1177     in[11]+=in[10]; in[10]+=in[9];  in[9] +=in[8];
1178     in[8] +=in[7];  in[7] +=in[6];  in[6] +=in[5];
1179     in[5] +=in[4];  in[4] +=in[3];  in[3] +=in[2];
1180     in[2] +=in[1];  in[1] +=in[0];
1181
1182     in[17]+=in[15]; in[15]+=in[13]; in[13]+=in[11]; in[11]+=in[9];
1183     in[9] +=in[7];  in[7] +=in[5];  in[5] +=in[3];  in[3] +=in[1];
1184
1185
1186   {
1187
1188 #define MACRO0(v) { \
1189     real tmp; \
1190     out2[9+(v)] = (tmp = sum0 + sum1) * w[27+(v)]; \
1191     out2[8-(v)] = tmp * w[26-(v)];  } \
1192     sum0 -= sum1; \
1193     ts[SBLIMIT*(8-(v))] = out1[8-(v)] + sum0 * w[8-(v)]; \
1194     ts[SBLIMIT*(9+(v))] = out1[9+(v)] + sum0 * w[9+(v)];
1195 #define MACRO1(v) { \
1196         real sum0,sum1; \
1197     sum0 = tmp1a + tmp2a; \
1198         sum1 = (tmp1b + tmp2b) * tfcos36[(v)]; \
1199         MACRO0(v); }
1200 #define MACRO2(v) { \
1201     real sum0,sum1; \
1202     sum0 = tmp2a - tmp1a; \
1203     sum1 = (tmp2b - tmp1b) * tfcos36[(v)]; \
1204         MACRO0(v); }
1205
1206     register const real *c = COS9;
1207     register real *out2 = o2;
1208         register real *w = wintab;
1209         register real *out1 = o1;
1210         register real *ts = tsbuf;
1211
1212     real ta33,ta66,tb33,tb66;
1213
1214     ta33 = in[2*3+0] * c[3];
1215     ta66 = in[2*6+0] * c[6];
1216     tb33 = in[2*3+1] * c[3];
1217     tb66 = in[2*6+1] * c[6];
1218
1219     {
1220       real tmp1a,tmp2a,tmp1b,tmp2b;
1221       tmp1a =             in[2*1+0] * c[1] + ta33 + in[2*5+0] * c[5] + in[2*7+0] * c[7];
1222       tmp1b =             in[2*1+1] * c[1] + tb33 + in[2*5+1] * c[5] + in[2*7+1] * c[7];
1223       tmp2a = in[2*0+0] + in[2*2+0] * c[2] + in[2*4+0] * c[4] + ta66 + in[2*8+0] * c[8];
1224       tmp2b = in[2*0+1] + in[2*2+1] * c[2] + in[2*4+1] * c[4] + tb66 + in[2*8+1] * c[8];
1225
1226       MACRO1(0);
1227       MACRO2(8);
1228     }
1229
1230     {
1231       real tmp1a,tmp2a,tmp1b,tmp2b;
1232       tmp1a = ( in[2*1+0] - in[2*5+0] - in[2*7+0] ) * c[3];
1233       tmp1b = ( in[2*1+1] - in[2*5+1] - in[2*7+1] ) * c[3];
1234       tmp2a = ( in[2*2+0] - in[2*4+0] - in[2*8+0] ) * c[6] - in[2*6+0] + in[2*0+0];
1235       tmp2b = ( in[2*2+1] - in[2*4+1] - in[2*8+1] ) * c[6] - in[2*6+1] + in[2*0+1];
1236
1237       MACRO1(1);
1238       MACRO2(7);
1239     }
1240
1241     {
1242       real tmp1a,tmp2a,tmp1b,tmp2b;
1243       tmp1a =             in[2*1+0] * c[5] - ta33 - in[2*5+0] * c[7] + in[2*7+0] * c[1];
1244       tmp1b =             in[2*1+1] * c[5] - tb33 - in[2*5+1] * c[7] + in[2*7+1] * c[1];
1245       tmp2a = in[2*0+0] - in[2*2+0] * c[8] - in[2*4+0] * c[2] + ta66 + in[2*8+0] * c[4];
1246       tmp2b = in[2*0+1] - in[2*2+1] * c[8] - in[2*4+1] * c[2] + tb66 + in[2*8+1] * c[4];
1247
1248       MACRO1(2);
1249       MACRO2(6);
1250     }
1251
1252     {
1253       real tmp1a,tmp2a,tmp1b,tmp2b;
1254       tmp1a =             in[2*1+0] * c[7] - ta33 + in[2*5+0] * c[1] - in[2*7+0] * c[5];
1255       tmp1b =             in[2*1+1] * c[7] - tb33 + in[2*5+1] * c[1] - in[2*7+1] * c[5];
1256       tmp2a = in[2*0+0] - in[2*2+0] * c[4] + in[2*4+0] * c[8] + ta66 - in[2*8+0] * c[2];
1257       tmp2b = in[2*0+1] - in[2*2+1] * c[4] + in[2*4+1] * c[8] + tb66 - in[2*8+1] * c[2];
1258
1259       MACRO1(3);
1260       MACRO2(5);
1261     }
1262
1263         {
1264                 real sum0,sum1;
1265         sum0 =  in[2*0+0] - in[2*2+0] + in[2*4+0] - in[2*6+0] + in[2*8+0];
1266         sum1 = (in[2*0+1] - in[2*2+1] + in[2*4+1] - in[2*6+1] + in[2*8+1] ) * tfcos36[4];
1267                 MACRO0(4);
1268         }
1269   }
1270
1271   }
1272 }
1273
1274 /*
1275  * new DCT12
1276  */
1277 static void dct12(real *in,real *rawout1,real *rawout2,register real *wi,register real *ts)
1278 {
1279 #define DCT12_PART1 \
1280              in5 = in[5*3];  \
1281      in5 += (in4 = in[4*3]); \
1282      in4 += (in3 = in[3*3]); \
1283      in3 += (in2 = in[2*3]); \
1284      in2 += (in1 = in[1*3]); \
1285      in1 += (in0 = in[0*3]); \
1286                              \
1287      in5 += in3; in3 += in1; \
1288                              \
1289      in2 *= COS6_1; \
1290      in3 *= COS6_1; \
1291
1292 #define DCT12_PART2 \
1293      in0 += in4 * COS6_2; \
1294                           \
1295      in4 = in0 + in2;     \
1296      in0 -= in2;          \
1297                           \
1298      in1 += in5 * COS6_2; \
1299                           \
1300      in5 = (in1 + in3) * tfcos12[0]; \
1301      in1 = (in1 - in3) * tfcos12[2]; \
1302                          \
1303      in3 = in4 + in5;    \
1304      in4 -= in5;         \
1305                          \
1306      in2 = in0 + in1;    \
1307      in0 -= in1;
1308
1309
1310    {
1311      real in0,in1,in2,in3,in4,in5;
1312      register real *out1 = rawout1;
1313      ts[SBLIMIT*0] = out1[0]; ts[SBLIMIT*1] = out1[1]; ts[SBLIMIT*2] = out1[2];
1314      ts[SBLIMIT*3] = out1[3]; ts[SBLIMIT*4] = out1[4]; ts[SBLIMIT*5] = out1[5];
1315
1316      DCT12_PART1
1317
1318      {
1319        real tmp0,tmp1 = (in0 - in4);
1320        {
1321          real tmp2 = (in1 - in5) * tfcos12[1];
1322          tmp0 = tmp1 + tmp2;
1323          tmp1 -= tmp2;
1324        }
1325        ts[(17-1)*SBLIMIT] = out1[17-1] + tmp0 * wi[11-1];
1326        ts[(12+1)*SBLIMIT] = out1[12+1] + tmp0 * wi[6+1];
1327        ts[(6 +1)*SBLIMIT] = out1[6 +1] + tmp1 * wi[1];
1328        ts[(11-1)*SBLIMIT] = out1[11-1] + tmp1 * wi[5-1];
1329      }
1330
1331      DCT12_PART2
1332
1333      ts[(17-0)*SBLIMIT] = out1[17-0] + in2 * wi[11-0];
1334      ts[(12+0)*SBLIMIT] = out1[12+0] + in2 * wi[6+0];
1335      ts[(12+2)*SBLIMIT] = out1[12+2] + in3 * wi[6+2];
1336      ts[(17-2)*SBLIMIT] = out1[17-2] + in3 * wi[11-2];
1337
1338      ts[(6+0)*SBLIMIT]  = out1[6+0] + in0 * wi[0];
1339      ts[(11-0)*SBLIMIT] = out1[11-0] + in0 * wi[5-0];
1340      ts[(6+2)*SBLIMIT]  = out1[6+2] + in4 * wi[2];
1341      ts[(11-2)*SBLIMIT] = out1[11-2] + in4 * wi[5-2];
1342   }
1343
1344   in++;
1345
1346   {
1347      real in0,in1,in2,in3,in4,in5;
1348      register real *out2 = rawout2;
1349
1350      DCT12_PART1
1351
1352      {
1353        real tmp0,tmp1 = (in0 - in4);
1354        {
1355          real tmp2 = (in1 - in5) * tfcos12[1];
1356          tmp0 = tmp1 + tmp2;
1357          tmp1 -= tmp2;
1358        }
1359        out2[5-1] = tmp0 * wi[11-1];
1360        out2[0+1] = tmp0 * wi[6+1];
1361        ts[(12+1)*SBLIMIT] += tmp1 * wi[1];
1362        ts[(17-1)*SBLIMIT] += tmp1 * wi[5-1];
1363      }
1364
1365      DCT12_PART2
1366
1367      out2[5-0] = in2 * wi[11-0];
1368      out2[0+0] = in2 * wi[6+0];
1369      out2[0+2] = in3 * wi[6+2];
1370      out2[5-2] = in3 * wi[11-2];
1371
1372      ts[(12+0)*SBLIMIT] += in0 * wi[0];
1373      ts[(17-0)*SBLIMIT] += in0 * wi[5-0];
1374      ts[(12+2)*SBLIMIT] += in4 * wi[2];
1375      ts[(17-2)*SBLIMIT] += in4 * wi[5-2];
1376   }
1377
1378   in++;
1379
1380   {
1381      real in0,in1,in2,in3,in4,in5;
1382      register real *out2 = rawout2;
1383      out2[12]=out2[13]=out2[14]=out2[15]=out2[16]=out2[17]=0.0;
1384
1385      DCT12_PART1
1386
1387      {
1388        real tmp0,tmp1 = (in0 - in4);
1389        {
1390          real tmp2 = (in1 - in5) * tfcos12[1];
1391          tmp0 = tmp1 + tmp2;
1392          tmp1 -= tmp2;
1393        }
1394        out2[11-1] = tmp0 * wi[11-1];
1395        out2[6 +1] = tmp0 * wi[6+1];
1396        out2[0+1] += tmp1 * wi[1];
1397        out2[5-1] += tmp1 * wi[5-1];
1398      }
1399
1400      DCT12_PART2
1401
1402      out2[11-0] = in2 * wi[11-0];
1403      out2[6 +0] = in2 * wi[6+0];
1404      out2[6 +2] = in3 * wi[6+2];
1405      out2[11-2] = in3 * wi[11-2];
1406
1407      out2[0+0] += in0 * wi[0];
1408      out2[5-0] += in0 * wi[5-0];
1409      out2[0+2] += in4 * wi[2];
1410      out2[5-2] += in4 * wi[5-2];
1411   }
1412 }
1413
1414 /*
1415  * III_hybrid
1416  */
1417 static void III_hybrid(real fsIn[SBLIMIT][SSLIMIT],real tsOut[SSLIMIT][SBLIMIT],
1418    int ch,struct gr_info_s *gr_info,struct mpstr *mp)
1419 {
1420    real *tspnt = (real *) tsOut;
1421    real (*block)[2][SBLIMIT*SSLIMIT] = mp->hybrid_block;
1422    int *blc = mp->hybrid_blc;
1423    real *rawout1,*rawout2;
1424    int bt;
1425    unsigned int sb = 0;
1426
1427    {
1428      int b = blc[ch];
1429      rawout1=block[b][ch];
1430      b=-b+1;
1431      rawout2=block[b][ch];
1432      blc[ch] = b;
1433    }
1434
1435
1436    if(gr_info->mixed_block_flag) {
1437      sb = 2;
1438      dct36(fsIn[0],rawout1,rawout2,win[0],tspnt);
1439      dct36(fsIn[1],rawout1+18,rawout2+18,win1[0],tspnt+1);
1440      rawout1 += 36; rawout2 += 36; tspnt += 2;
1441    }
1442
1443    bt = gr_info->block_type;
1444    if(bt == 2) {
1445      for (; sb<gr_info->maxb; sb+=2,tspnt+=2,rawout1+=36,rawout2+=36) {
1446        dct12(fsIn[sb],rawout1,rawout2,win[2],tspnt);
1447        dct12(fsIn[sb+1],rawout1+18,rawout2+18,win1[2],tspnt+1);
1448      }
1449    }
1450    else {
1451      for (; sb<gr_info->maxb; sb+=2,tspnt+=2,rawout1+=36,rawout2+=36) {
1452        dct36(fsIn[sb],rawout1,rawout2,win[bt],tspnt);
1453        dct36(fsIn[sb+1],rawout1+18,rawout2+18,win1[bt],tspnt+1);
1454      }
1455    }
1456
1457    for(;sb<SBLIMIT;sb++,tspnt++) {
1458      int i;
1459      for(i=0;i<SSLIMIT;i++) {
1460        tspnt[i*SBLIMIT] = *rawout1++;
1461        *rawout2++ = 0.0;
1462      }
1463    }
1464 }
1465
1466 /*
1467  * main layer3 handler
1468  */
1469 int do_layer3(struct frame *fr,unsigned char *pcm_sample,int *pcm_point)
1470 {
1471   int gr, ch, ss,clip=0;
1472   int scalefacs[2][39]; /* max 39 for short[13][3] mode, mixed: 38, long: 22 */
1473   struct III_sideinfo sideinfo;
1474   int stereo = fr->stereo;
1475   int single = fr->single;
1476   int ms_stereo,i_stereo;
1477   int sfreq = fr->sampling_frequency;
1478   int stereo1,granules;
1479
1480   if(stereo == 1) { /* stream is mono */
1481     stereo1 = 1;
1482     single = 0;
1483   }
1484   else if(single >= 0) /* stream is stereo, but force to mono */
1485     stereo1 = 1;
1486   else
1487     stereo1 = 2;
1488
1489   if(fr->mode == MPG_MD_JOINT_STEREO) {
1490     ms_stereo = fr->mode_ext & 0x2;
1491     i_stereo  = fr->mode_ext & 0x1;
1492   }
1493   else
1494     ms_stereo = i_stereo = 0;
1495
1496   if(fr->lsf) {
1497     granules = 1;
1498     if(!III_get_side_info_2(&sideinfo,stereo,ms_stereo,sfreq,single))
1499       return -1;
1500   }
1501   else {
1502     granules = 2;
1503 #ifdef MPEG1
1504     if(!III_get_side_info_1(&sideinfo,stereo,ms_stereo,sfreq,single))
1505       return -1;
1506 #else
1507     FIXME("Not supported\n");
1508 #endif
1509   }
1510
1511   if(set_pointer(fr->mp,sideinfo.main_data_begin) == MP3_ERR)
1512     return -1;
1513
1514   for (gr=0;gr<granules;gr++)
1515   {
1516     real hybridIn[2][SBLIMIT][SSLIMIT];
1517     real hybridOut[2][SSLIMIT][SBLIMIT];
1518
1519     {
1520       struct gr_info_s *gr_info = &(sideinfo.ch[0].gr[gr]);
1521       long part2bits;
1522       if(fr->lsf)
1523         part2bits = III_get_scale_factors_2(scalefacs[0],gr_info,0);
1524       else {
1525 #ifdef MPEG1
1526         part2bits = III_get_scale_factors_1(scalefacs[0],gr_info);
1527 #else
1528         FIXME("Not supported\n");
1529 #endif
1530       }
1531       if(III_dequantize_sample(hybridIn[0], scalefacs[0],gr_info,sfreq,part2bits))
1532         return clip;
1533     }
1534     if(stereo == 2) {
1535       struct gr_info_s *gr_info = &(sideinfo.ch[1].gr[gr]);
1536       long part2bits;
1537       if(fr->lsf)
1538         part2bits = III_get_scale_factors_2(scalefacs[1],gr_info,i_stereo);
1539       else {
1540 #ifdef MPEG1
1541         part2bits = III_get_scale_factors_1(scalefacs[1],gr_info);
1542 #else
1543         FIXME("Not supported\n");
1544 #endif
1545       }
1546
1547       if(III_dequantize_sample(hybridIn[1],scalefacs[1],gr_info,sfreq,part2bits))
1548           return clip;
1549
1550       if(ms_stereo) {
1551         int i;
1552         for(i=0;i<SBLIMIT*SSLIMIT;i++) {
1553           real tmp0,tmp1;
1554           tmp0 = ((real *) hybridIn[0])[i];
1555           tmp1 = ((real *) hybridIn[1])[i];
1556           ((real *) hybridIn[0])[i] = tmp0 + tmp1;
1557           ((real *) hybridIn[1])[i] = tmp0 - tmp1;
1558         }
1559       }
1560
1561       if(i_stereo)
1562         III_i_stereo(hybridIn,scalefacs[1],gr_info,sfreq,ms_stereo,fr->lsf);
1563
1564       if(ms_stereo || i_stereo || (single == 3) ) {
1565         if(gr_info->maxb > sideinfo.ch[0].gr[gr].maxb)
1566           sideinfo.ch[0].gr[gr].maxb = gr_info->maxb;
1567         else
1568           gr_info->maxb = sideinfo.ch[0].gr[gr].maxb;
1569       }
1570
1571       switch(single) {
1572         case 3:
1573           {
1574             register unsigned int i;
1575             register real *in0 = (real *) hybridIn[0],*in1 = (real *) hybridIn[1];
1576             for(i=0;i<SSLIMIT*gr_info->maxb;i++,in0++)
1577               *in0 = (*in0 + *in1++); /* *0.5 done by pow-scale */
1578           }
1579           break;
1580         case 1:
1581           {
1582             register unsigned int i;
1583             register real *in0 = (real *) hybridIn[0],*in1 = (real *) hybridIn[1];
1584             for(i=0;i<SSLIMIT*gr_info->maxb;i++)
1585               *in0++ = *in1++;
1586           }
1587           break;
1588       }
1589     }
1590
1591     for(ch=0;ch<stereo1;ch++) {
1592       struct gr_info_s *gr_info = &(sideinfo.ch[ch].gr[gr]);
1593       III_antialias(hybridIn[ch],gr_info);
1594       III_hybrid(hybridIn[ch], hybridOut[ch], ch,gr_info, fr->mp);
1595     }
1596
1597     for(ss=0;ss<SSLIMIT;ss++) {
1598       if(single >= 0) {
1599         clip += synth_1to1_mono(fr->mp,hybridOut[0][ss],pcm_sample,pcm_point);
1600       }
1601       else {
1602         int p1 = *pcm_point;
1603         clip += synth_1to1(fr->mp,hybridOut[0][ss],0,pcm_sample,&p1);
1604         clip += synth_1to1(fr->mp,hybridOut[1][ss],1,pcm_sample,pcm_point);
1605       }
1606     }
1607   }
1608
1609   return clip;
1610 }