dlls/winemp3.acm/dct64_i386.c

   1 /*
   2  * Copyright (c) Michael Hipp and other authors of the mpglib project.
   3  *
   4  * Discrete Cosine Transform (DCT) for subband synthesis
   5  * optimized for machines with no auto-increment.
   6  * The performance is highly compiler dependent. Maybe
   7  * the dct64.c version for 'normal' processor may be faster
   8  * even for Intel processors.
   9  *
  10  * This library is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * This library is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with this library; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  23  */
  24
  25 #include "mpg123.h"
  26
  27 static void dct64_1(real *out0,real *out1,real *b1,real *b2,real *samples)
  28 {
  29
  30  {
  31   register real *costab = pnts[0];
  32
  33   b1[0x00] = samples[0x00] + samples[0x1F];
  34   b1[0x1F] = (samples[0x00] - samples[0x1F]) * costab[0x0];
  35
  36   b1[0x01] = samples[0x01] + samples[0x1E];
  37   b1[0x1E] = (samples[0x01] - samples[0x1E]) * costab[0x1];
  38
  39   b1[0x02] = samples[0x02] + samples[0x1D];
  40   b1[0x1D] = (samples[0x02] - samples[0x1D]) * costab[0x2];
  41
  42   b1[0x03] = samples[0x03] + samples[0x1C];
  43   b1[0x1C] = (samples[0x03] - samples[0x1C]) * costab[0x3];
  44
  45   b1[0x04] = samples[0x04] + samples[0x1B];
  46   b1[0x1B] = (samples[0x04] - samples[0x1B]) * costab[0x4];
  47
  48   b1[0x05] = samples[0x05] + samples[0x1A];
  49   b1[0x1A] = (samples[0x05] - samples[0x1A]) * costab[0x5];
  50
  51   b1[0x06] = samples[0x06] + samples[0x19];
  52   b1[0x19] = (samples[0x06] - samples[0x19]) * costab[0x6];
  53
  54   b1[0x07] = samples[0x07] + samples[0x18];
  55   b1[0x18] = (samples[0x07] - samples[0x18]) * costab[0x7];
  56
  57   b1[0x08] = samples[0x08] + samples[0x17];
  58   b1[0x17] = (samples[0x08] - samples[0x17]) * costab[0x8];
  59
  60   b1[0x09] = samples[0x09] + samples[0x16];
  61   b1[0x16] = (samples[0x09] - samples[0x16]) * costab[0x9];
  62
  63   b1[0x0A] = samples[0x0A] + samples[0x15];
  64   b1[0x15] = (samples[0x0A] - samples[0x15]) * costab[0xA];
  65
  66   b1[0x0B] = samples[0x0B] + samples[0x14];
  67   b1[0x14] = (samples[0x0B] - samples[0x14]) * costab[0xB];
  68
  69   b1[0x0C] = samples[0x0C] + samples[0x13];
  70   b1[0x13] = (samples[0x0C] - samples[0x13]) * costab[0xC];
  71
  72   b1[0x0D] = samples[0x0D] + samples[0x12];
  73   b1[0x12] = (samples[0x0D] - samples[0x12]) * costab[0xD];
  74
  75   b1[0x0E] = samples[0x0E] + samples[0x11];
  76   b1[0x11] = (samples[0x0E] - samples[0x11]) * costab[0xE];
  77
  78   b1[0x0F] = samples[0x0F] + samples[0x10];
  79   b1[0x10] = (samples[0x0F] - samples[0x10]) * costab[0xF];
  80  }
  81
  82
  83  {
  84   register real *costab = pnts[1];
  85
  86   b2[0x00] = b1[0x00] + b1[0x0F];
  87   b2[0x0F] = (b1[0x00] - b1[0x0F]) * costab[0];
  88   b2[0x01] = b1[0x01] + b1[0x0E];
  89   b2[0x0E] = (b1[0x01] - b1[0x0E]) * costab[1];
  90   b2[0x02] = b1[0x02] + b1[0x0D];
  91   b2[0x0D] = (b1[0x02] - b1[0x0D]) * costab[2];
  92   b2[0x03] = b1[0x03] + b1[0x0C];
  93   b2[0x0C] = (b1[0x03] - b1[0x0C]) * costab[3];
  94   b2[0x04] = b1[0x04] + b1[0x0B];
  95   b2[0x0B] = (b1[0x04] - b1[0x0B]) * costab[4];
  96   b2[0x05] = b1[0x05] + b1[0x0A];
  97   b2[0x0A] = (b1[0x05] - b1[0x0A]) * costab[5];
  98   b2[0x06] = b1[0x06] + b1[0x09];
  99   b2[0x09] = (b1[0x06] - b1[0x09]) * costab[6];
 100   b2[0x07] = b1[0x07] + b1[0x08];
 101   b2[0x08] = (b1[0x07] - b1[0x08]) * costab[7];
 102
 103   b2[0x10] = b1[0x10] + b1[0x1F];
 104   b2[0x1F] = (b1[0x1F] - b1[0x10]) * costab[0];
 105   b2[0x11] = b1[0x11] + b1[0x1E];
 106   b2[0x1E] = (b1[0x1E] - b1[0x11]) * costab[1];
 107   b2[0x12] = b1[0x12] + b1[0x1D];
 108   b2[0x1D] = (b1[0x1D] - b1[0x12]) * costab[2];
 109   b2[0x13] = b1[0x13] + b1[0x1C];
 110   b2[0x1C] = (b1[0x1C] - b1[0x13]) * costab[3];
 111   b2[0x14] = b1[0x14] + b1[0x1B];
 112   b2[0x1B] = (b1[0x1B] - b1[0x14]) * costab[4];
 113   b2[0x15] = b1[0x15] + b1[0x1A];
 114   b2[0x1A] = (b1[0x1A] - b1[0x15]) * costab[5];
 115   b2[0x16] = b1[0x16] + b1[0x19];
 116   b2[0x19] = (b1[0x19] - b1[0x16]) * costab[6];
 117   b2[0x17] = b1[0x17] + b1[0x18];
 118   b2[0x18] = (b1[0x18] - b1[0x17]) * costab[7];
 119  }
 120
 121  {
 122   register real *costab = pnts[2];
 123
 124   b1[0x00] = b2[0x00] + b2[0x07];
 125   b1[0x07] = (b2[0x00] - b2[0x07]) * costab[0];
 126   b1[0x01] = b2[0x01] + b2[0x06];
 127   b1[0x06] = (b2[0x01] - b2[0x06]) * costab[1];
 128   b1[0x02] = b2[0x02] + b2[0x05];
 129   b1[0x05] = (b2[0x02] - b2[0x05]) * costab[2];
 130   b1[0x03] = b2[0x03] + b2[0x04];
 131   b1[0x04] = (b2[0x03] - b2[0x04]) * costab[3];
 132
 133   b1[0x08] = b2[0x08] + b2[0x0F];
 134   b1[0x0F] = (b2[0x0F] - b2[0x08]) * costab[0];
 135   b1[0x09] = b2[0x09] + b2[0x0E];
 136   b1[0x0E] = (b2[0x0E] - b2[0x09]) * costab[1];
 137   b1[0x0A] = b2[0x0A] + b2[0x0D];
 138   b1[0x0D] = (b2[0x0D] - b2[0x0A]) * costab[2];
 139   b1[0x0B] = b2[0x0B] + b2[0x0C];
 140   b1[0x0C] = (b2[0x0C] - b2[0x0B]) * costab[3];
 141
 142   b1[0x10] = b2[0x10] + b2[0x17];
 143   b1[0x17] = (b2[0x10] - b2[0x17]) * costab[0];
 144   b1[0x11] = b2[0x11] + b2[0x16];
 145   b1[0x16] = (b2[0x11] - b2[0x16]) * costab[1];
 146   b1[0x12] = b2[0x12] + b2[0x15];
 147   b1[0x15] = (b2[0x12] - b2[0x15]) * costab[2];
 148   b1[0x13] = b2[0x13] + b2[0x14];
 149   b1[0x14] = (b2[0x13] - b2[0x14]) * costab[3];
 150
 151   b1[0x18] = b2[0x18] + b2[0x1F];
 152   b1[0x1F] = (b2[0x1F] - b2[0x18]) * costab[0];
 153   b1[0x19] = b2[0x19] + b2[0x1E];
 154   b1[0x1E] = (b2[0x1E] - b2[0x19]) * costab[1];
 155   b1[0x1A] = b2[0x1A] + b2[0x1D];
 156   b1[0x1D] = (b2[0x1D] - b2[0x1A]) * costab[2];
 157   b1[0x1B] = b2[0x1B] + b2[0x1C];
 158   b1[0x1C] = (b2[0x1C] - b2[0x1B]) * costab[3];
 159  }
 160
 161  {
 162   register real const cos0 = pnts[3][0];
 163   register real const cos1 = pnts[3][1];
 164
 165   b2[0x00] = b1[0x00] + b1[0x03];
 166   b2[0x03] = (b1[0x00] - b1[0x03]) * cos0;
 167   b2[0x01] = b1[0x01] + b1[0x02];
 168   b2[0x02] = (b1[0x01] - b1[0x02]) * cos1;
 169
 170   b2[0x04] = b1[0x04] + b1[0x07];
 171   b2[0x07] = (b1[0x07] - b1[0x04]) * cos0;
 172   b2[0x05] = b1[0x05] + b1[0x06];
 173   b2[0x06] = (b1[0x06] - b1[0x05]) * cos1;
 174
 175   b2[0x08] = b1[0x08] + b1[0x0B];
 176   b2[0x0B] = (b1[0x08] - b1[0x0B]) * cos0;
 177   b2[0x09] = b1[0x09] + b1[0x0A];
 178   b2[0x0A] = (b1[0x09] - b1[0x0A]) * cos1;
 179
 180   b2[0x0C] = b1[0x0C] + b1[0x0F];
 181   b2[0x0F] = (b1[0x0F] - b1[0x0C]) * cos0;
 182   b2[0x0D] = b1[0x0D] + b1[0x0E];
 183   b2[0x0E] = (b1[0x0E] - b1[0x0D]) * cos1;
 184
 185   b2[0x10] = b1[0x10] + b1[0x13];
 186   b2[0x13] = (b1[0x10] - b1[0x13]) * cos0;
 187   b2[0x11] = b1[0x11] + b1[0x12];
 188   b2[0x12] = (b1[0x11] - b1[0x12]) * cos1;
 189
 190   b2[0x14] = b1[0x14] + b1[0x17];
 191   b2[0x17] = (b1[0x17] - b1[0x14]) * cos0;
 192   b2[0x15] = b1[0x15] + b1[0x16];
 193   b2[0x16] = (b1[0x16] - b1[0x15]) * cos1;
 194
 195   b2[0x18] = b1[0x18] + b1[0x1B];
 196   b2[0x1B] = (b1[0x18] - b1[0x1B]) * cos0;
 197   b2[0x19] = b1[0x19] + b1[0x1A];
 198   b2[0x1A] = (b1[0x19] - b1[0x1A]) * cos1;
 199
 200   b2[0x1C] = b1[0x1C] + b1[0x1F];
 201   b2[0x1F] = (b1[0x1F] - b1[0x1C]) * cos0;
 202   b2[0x1D] = b1[0x1D] + b1[0x1E];
 203   b2[0x1E] = (b1[0x1E] - b1[0x1D]) * cos1;
 204  }
 205
 206  {
 207   register real const cos0 = pnts[4][0];
 208
 209   b1[0x00] = b2[0x00] + b2[0x01];
 210   b1[0x01] = (b2[0x00] - b2[0x01]) * cos0;
 211   b1[0x02] = b2[0x02] + b2[0x03];
 212   b1[0x03] = (b2[0x03] - b2[0x02]) * cos0;
 213   b1[0x02] += b1[0x03];
 214
 215   b1[0x04] = b2[0x04] + b2[0x05];
 216   b1[0x05] = (b2[0x04] - b2[0x05]) * cos0;
 217   b1[0x06] = b2[0x06] + b2[0x07];
 218   b1[0x07] = (b2[0x07] - b2[0x06]) * cos0;
 219   b1[0x06] += b1[0x07];
 220   b1[0x04] += b1[0x06];
 221   b1[0x06] += b1[0x05];
 222   b1[0x05] += b1[0x07];
 223
 224   b1[0x08] = b2[0x08] + b2[0x09];
 225   b1[0x09] = (b2[0x08] - b2[0x09]) * cos0;
 226   b1[0x0A] = b2[0x0A] + b2[0x0B];
 227   b1[0x0B] = (b2[0x0B] - b2[0x0A]) * cos0;
 228   b1[0x0A] += b1[0x0B];
 229
 230   b1[0x0C] = b2[0x0C] + b2[0x0D];
 231   b1[0x0D] = (b2[0x0C] - b2[0x0D]) * cos0;
 232   b1[0x0E] = b2[0x0E] + b2[0x0F];
 233   b1[0x0F] = (b2[0x0F] - b2[0x0E]) * cos0;
 234   b1[0x0E] += b1[0x0F];
 235   b1[0x0C] += b1[0x0E];
 236   b1[0x0E] += b1[0x0D];
 237   b1[0x0D] += b1[0x0F];
 238
 239   b1[0x10] = b2[0x10] + b2[0x11];
 240   b1[0x11] = (b2[0x10] - b2[0x11]) * cos0;
 241   b1[0x12] = b2[0x12] + b2[0x13];
 242   b1[0x13] = (b2[0x13] - b2[0x12]) * cos0;
 243   b1[0x12] += b1[0x13];
 244
 245   b1[0x14] = b2[0x14] + b2[0x15];
 246   b1[0x15] = (b2[0x14] - b2[0x15]) * cos0;
 247   b1[0x16] = b2[0x16] + b2[0x17];
 248   b1[0x17] = (b2[0x17] - b2[0x16]) * cos0;
 249   b1[0x16] += b1[0x17];
 250   b1[0x14] += b1[0x16];
 251   b1[0x16] += b1[0x15];
 252   b1[0x15] += b1[0x17];
 253
 254   b1[0x18] = b2[0x18] + b2[0x19];
 255   b1[0x19] = (b2[0x18] - b2[0x19]) * cos0;
 256   b1[0x1A] = b2[0x1A] + b2[0x1B];
 257   b1[0x1B] = (b2[0x1B] - b2[0x1A]) * cos0;
 258   b1[0x1A] += b1[0x1B];
 259
 260   b1[0x1C] = b2[0x1C] + b2[0x1D];
 261   b1[0x1D] = (b2[0x1C] - b2[0x1D]) * cos0;
 262   b1[0x1E] = b2[0x1E] + b2[0x1F];
 263   b1[0x1F] = (b2[0x1F] - b2[0x1E]) * cos0;
 264   b1[0x1E] += b1[0x1F];
 265   b1[0x1C] += b1[0x1E];
 266   b1[0x1E] += b1[0x1D];
 267   b1[0x1D] += b1[0x1F];
 268  }
 269
 270  out0[0x10*16] = b1[0x00];
 271  out0[0x10*12] = b1[0x04];
 272  out0[0x10* 8] = b1[0x02];
 273  out0[0x10* 4] = b1[0x06];
 274  out0[0x10* 0] = b1[0x01];
 275  out1[0x10* 0] = b1[0x01];
 276  out1[0x10* 4] = b1[0x05];
 277  out1[0x10* 8] = b1[0x03];
 278  out1[0x10*12] = b1[0x07];
 279
 280  b1[0x08] += b1[0x0C];
 281  out0[0x10*14] = b1[0x08];
 282  b1[0x0C] += b1[0x0a];
 283  out0[0x10*10] = b1[0x0C];
 284  b1[0x0A] += b1[0x0E];
 285  out0[0x10* 6] = b1[0x0A];
 286  b1[0x0E] += b1[0x09];
 287  out0[0x10* 2] = b1[0x0E];
 288  b1[0x09] += b1[0x0D];
 289  out1[0x10* 2] = b1[0x09];
 290  b1[0x0D] += b1[0x0B];
 291  out1[0x10* 6] = b1[0x0D];
 292  b1[0x0B] += b1[0x0F];
 293  out1[0x10*10] = b1[0x0B];
 294  out1[0x10*14] = b1[0x0F];
 295
 296  b1[0x18] += b1[0x1C];
 297  out0[0x10*15] = b1[0x10] + b1[0x18];
 298  out0[0x10*13] = b1[0x18] + b1[0x14];
 299  b1[0x1C] += b1[0x1a];
 300  out0[0x10*11] = b1[0x14] + b1[0x1C];
 301  out0[0x10* 9] = b1[0x1C] + b1[0x12];
 302  b1[0x1A] += b1[0x1E];
 303  out0[0x10* 7] = b1[0x12] + b1[0x1A];
 304  out0[0x10* 5] = b1[0x1A] + b1[0x16];
 305  b1[0x1E] += b1[0x19];
 306  out0[0x10* 3] = b1[0x16] + b1[0x1E];
 307  out0[0x10* 1] = b1[0x1E] + b1[0x11];
 308  b1[0x19] += b1[0x1D];
 309  out1[0x10* 1] = b1[0x11] + b1[0x19];
 310  out1[0x10* 3] = b1[0x19] + b1[0x15];
 311  b1[0x1D] += b1[0x1B];
 312  out1[0x10* 5] = b1[0x15] + b1[0x1D];
 313  out1[0x10* 7] = b1[0x1D] + b1[0x13];
 314  b1[0x1B] += b1[0x1F];
 315  out1[0x10* 9] = b1[0x13] + b1[0x1B];
 316  out1[0x10*11] = b1[0x1B] + b1[0x17];
 317  out1[0x10*13] = b1[0x17] + b1[0x1F];
 318  out1[0x10*15] = b1[0x1F];
 319 }
 320
 321 /*
 322  * the call via dct64 is a trick to force GCC to use
 323  * (new) registers for the b1,b2 pointer to the bufs[xx] field
 324  */
 325 void dct64(real *a,real *b,real *c)
 326 {
 327   real bufs[0x40];
 328   dct64_1(a,b,bufs,bufs+0x20,c);
 329 }