src/lib/dl/ext/lame/dct64_i386.c

   1 /*
   2  * dct64_i368.c
   3  *
   4  * Copyright (C) 1999-2010 The L.A.M.E. project
   5  *
   6  * Initially written by Michael Hipp, see also AUTHORS and README.
   7  *
   8  * This library is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Library General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2 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  * Library General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Library General Public
  19  * License along with this library; if not, write to the
  20  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  21  * Boston, MA 02111-1307, USA.
  22  *
  23  *
  24  * Discrete Cosine Tansform (DCT) for subband synthesis
  25  * optimized for machines with no auto-increment.
  26  * The performance is highly compiler dependend. Maybe
  27  * the dct64.c version for 'normal' processor may be faster
  28  * even for Intel processors.
  29  */
  30
  31 /* $Id: dct64_i386.c,v 1.14 2010/03/22 14:30:19 robert Exp $ */
  32
  33 #ifdef HAVE_CONFIG_H
  34 #include <config.h>
  35 #endif
  36
  37 #include "dct64_i386.h"
  38 #include "tabinit.h"
  39
  40 #ifdef WITH_DMALLOC
  41 #include <dmalloc.h>
  42 #endif
  43
  44 static void
  45 dct64_1(real * out0, real * out1, real * b1, real * b2, real * samples)
  46 {
  47
  48     {
  49         real   *costab = pnts[0];
  50
  51         b1[0x00] = samples[0x00] + samples[0x1F];
  52         b1[0x1F] = (samples[0x00] - samples[0x1F]) * costab[0x0];
  53
  54         b1[0x01] = samples[0x01] + samples[0x1E];
  55         b1[0x1E] = (samples[0x01] - samples[0x1E]) * costab[0x1];
  56
  57         b1[0x02] = samples[0x02] + samples[0x1D];
  58         b1[0x1D] = (samples[0x02] - samples[0x1D]) * costab[0x2];
  59
  60         b1[0x03] = samples[0x03] + samples[0x1C];
  61         b1[0x1C] = (samples[0x03] - samples[0x1C]) * costab[0x3];
  62
  63         b1[0x04] = samples[0x04] + samples[0x1B];
  64         b1[0x1B] = (samples[0x04] - samples[0x1B]) * costab[0x4];
  65
  66         b1[0x05] = samples[0x05] + samples[0x1A];
  67         b1[0x1A] = (samples[0x05] - samples[0x1A]) * costab[0x5];
  68
  69         b1[0x06] = samples[0x06] + samples[0x19];
  70         b1[0x19] = (samples[0x06] - samples[0x19]) * costab[0x6];
  71
  72         b1[0x07] = samples[0x07] + samples[0x18];
  73         b1[0x18] = (samples[0x07] - samples[0x18]) * costab[0x7];
  74
  75         b1[0x08] = samples[0x08] + samples[0x17];
  76         b1[0x17] = (samples[0x08] - samples[0x17]) * costab[0x8];
  77
  78         b1[0x09] = samples[0x09] + samples[0x16];
  79         b1[0x16] = (samples[0x09] - samples[0x16]) * costab[0x9];
  80
  81         b1[0x0A] = samples[0x0A] + samples[0x15];
  82         b1[0x15] = (samples[0x0A] - samples[0x15]) * costab[0xA];
  83
  84         b1[0x0B] = samples[0x0B] + samples[0x14];
  85         b1[0x14] = (samples[0x0B] - samples[0x14]) * costab[0xB];
  86
  87         b1[0x0C] = samples[0x0C] + samples[0x13];
  88         b1[0x13] = (samples[0x0C] - samples[0x13]) * costab[0xC];
  89
  90         b1[0x0D] = samples[0x0D] + samples[0x12];
  91         b1[0x12] = (samples[0x0D] - samples[0x12]) * costab[0xD];
  92
  93         b1[0x0E] = samples[0x0E] + samples[0x11];
  94         b1[0x11] = (samples[0x0E] - samples[0x11]) * costab[0xE];
  95
  96         b1[0x0F] = samples[0x0F] + samples[0x10];
  97         b1[0x10] = (samples[0x0F] - samples[0x10]) * costab[0xF];
  98     }
  99
 100
 101     {
 102         real   *costab = pnts[1];
 103
 104         b2[0x00] = b1[0x00] + b1[0x0F];
 105         b2[0x0F] = (b1[0x00] - b1[0x0F]) * costab[0];
 106         b2[0x01] = b1[0x01] + b1[0x0E];
 107         b2[0x0E] = (b1[0x01] - b1[0x0E]) * costab[1];
 108         b2[0x02] = b1[0x02] + b1[0x0D];
 109         b2[0x0D] = (b1[0x02] - b1[0x0D]) * costab[2];
 110         b2[0x03] = b1[0x03] + b1[0x0C];
 111         b2[0x0C] = (b1[0x03] - b1[0x0C]) * costab[3];
 112         b2[0x04] = b1[0x04] + b1[0x0B];
 113         b2[0x0B] = (b1[0x04] - b1[0x0B]) * costab[4];
 114         b2[0x05] = b1[0x05] + b1[0x0A];
 115         b2[0x0A] = (b1[0x05] - b1[0x0A]) * costab[5];
 116         b2[0x06] = b1[0x06] + b1[0x09];
 117         b2[0x09] = (b1[0x06] - b1[0x09]) * costab[6];
 118         b2[0x07] = b1[0x07] + b1[0x08];
 119         b2[0x08] = (b1[0x07] - b1[0x08]) * costab[7];
 120
 121         b2[0x10] = b1[0x10] + b1[0x1F];
 122         b2[0x1F] = (b1[0x1F] - b1[0x10]) * costab[0];
 123         b2[0x11] = b1[0x11] + b1[0x1E];
 124         b2[0x1E] = (b1[0x1E] - b1[0x11]) * costab[1];
 125         b2[0x12] = b1[0x12] + b1[0x1D];
 126         b2[0x1D] = (b1[0x1D] - b1[0x12]) * costab[2];
 127         b2[0x13] = b1[0x13] + b1[0x1C];
 128         b2[0x1C] = (b1[0x1C] - b1[0x13]) * costab[3];
 129         b2[0x14] = b1[0x14] + b1[0x1B];
 130         b2[0x1B] = (b1[0x1B] - b1[0x14]) * costab[4];
 131         b2[0x15] = b1[0x15] + b1[0x1A];
 132         b2[0x1A] = (b1[0x1A] - b1[0x15]) * costab[5];
 133         b2[0x16] = b1[0x16] + b1[0x19];
 134         b2[0x19] = (b1[0x19] - b1[0x16]) * costab[6];
 135         b2[0x17] = b1[0x17] + b1[0x18];
 136         b2[0x18] = (b1[0x18] - b1[0x17]) * costab[7];
 137     }
 138
 139     {
 140         real   *costab = pnts[2];
 141
 142         b1[0x00] = b2[0x00] + b2[0x07];
 143         b1[0x07] = (b2[0x00] - b2[0x07]) * costab[0];
 144         b1[0x01] = b2[0x01] + b2[0x06];
 145         b1[0x06] = (b2[0x01] - b2[0x06]) * costab[1];
 146         b1[0x02] = b2[0x02] + b2[0x05];
 147         b1[0x05] = (b2[0x02] - b2[0x05]) * costab[2];
 148         b1[0x03] = b2[0x03] + b2[0x04];
 149         b1[0x04] = (b2[0x03] - b2[0x04]) * costab[3];
 150
 151         b1[0x08] = b2[0x08] + b2[0x0F];
 152         b1[0x0F] = (b2[0x0F] - b2[0x08]) * costab[0];
 153         b1[0x09] = b2[0x09] + b2[0x0E];
 154         b1[0x0E] = (b2[0x0E] - b2[0x09]) * costab[1];
 155         b1[0x0A] = b2[0x0A] + b2[0x0D];
 156         b1[0x0D] = (b2[0x0D] - b2[0x0A]) * costab[2];
 157         b1[0x0B] = b2[0x0B] + b2[0x0C];
 158         b1[0x0C] = (b2[0x0C] - b2[0x0B]) * costab[3];
 159
 160         b1[0x10] = b2[0x10] + b2[0x17];
 161         b1[0x17] = (b2[0x10] - b2[0x17]) * costab[0];
 162         b1[0x11] = b2[0x11] + b2[0x16];
 163         b1[0x16] = (b2[0x11] - b2[0x16]) * costab[1];
 164         b1[0x12] = b2[0x12] + b2[0x15];
 165         b1[0x15] = (b2[0x12] - b2[0x15]) * costab[2];
 166         b1[0x13] = b2[0x13] + b2[0x14];
 167         b1[0x14] = (b2[0x13] - b2[0x14]) * costab[3];
 168
 169         b1[0x18] = b2[0x18] + b2[0x1F];
 170         b1[0x1F] = (b2[0x1F] - b2[0x18]) * costab[0];
 171         b1[0x19] = b2[0x19] + b2[0x1E];
 172         b1[0x1E] = (b2[0x1E] - b2[0x19]) * costab[1];
 173         b1[0x1A] = b2[0x1A] + b2[0x1D];
 174         b1[0x1D] = (b2[0x1D] - b2[0x1A]) * costab[2];
 175         b1[0x1B] = b2[0x1B] + b2[0x1C];
 176         b1[0x1C] = (b2[0x1C] - b2[0x1B]) * costab[3];
 177     }
 178
 179     {
 180         real const cos0 = pnts[3][0];
 181         real const cos1 = pnts[3][1];
 182
 183         b2[0x00] = b1[0x00] + b1[0x03];
 184         b2[0x03] = (b1[0x00] - b1[0x03]) * cos0;
 185         b2[0x01] = b1[0x01] + b1[0x02];
 186         b2[0x02] = (b1[0x01] - b1[0x02]) * cos1;
 187
 188         b2[0x04] = b1[0x04] + b1[0x07];
 189         b2[0x07] = (b1[0x07] - b1[0x04]) * cos0;
 190         b2[0x05] = b1[0x05] + b1[0x06];
 191         b2[0x06] = (b1[0x06] - b1[0x05]) * cos1;
 192
 193         b2[0x08] = b1[0x08] + b1[0x0B];
 194         b2[0x0B] = (b1[0x08] - b1[0x0B]) * cos0;
 195         b2[0x09] = b1[0x09] + b1[0x0A];
 196         b2[0x0A] = (b1[0x09] - b1[0x0A]) * cos1;
 197
 198         b2[0x0C] = b1[0x0C] + b1[0x0F];
 199         b2[0x0F] = (b1[0x0F] - b1[0x0C]) * cos0;
 200         b2[0x0D] = b1[0x0D] + b1[0x0E];
 201         b2[0x0E] = (b1[0x0E] - b1[0x0D]) * cos1;
 202
 203         b2[0x10] = b1[0x10] + b1[0x13];
 204         b2[0x13] = (b1[0x10] - b1[0x13]) * cos0;
 205         b2[0x11] = b1[0x11] + b1[0x12];
 206         b2[0x12] = (b1[0x11] - b1[0x12]) * cos1;
 207
 208         b2[0x14] = b1[0x14] + b1[0x17];
 209         b2[0x17] = (b1[0x17] - b1[0x14]) * cos0;
 210         b2[0x15] = b1[0x15] + b1[0x16];
 211         b2[0x16] = (b1[0x16] - b1[0x15]) * cos1;
 212
 213         b2[0x18] = b1[0x18] + b1[0x1B];
 214         b2[0x1B] = (b1[0x18] - b1[0x1B]) * cos0;
 215         b2[0x19] = b1[0x19] + b1[0x1A];
 216         b2[0x1A] = (b1[0x19] - b1[0x1A]) * cos1;
 217
 218         b2[0x1C] = b1[0x1C] + b1[0x1F];
 219         b2[0x1F] = (b1[0x1F] - b1[0x1C]) * cos0;
 220         b2[0x1D] = b1[0x1D] + b1[0x1E];
 221         b2[0x1E] = (b1[0x1E] - b1[0x1D]) * cos1;
 222     }
 223
 224     {
 225         real const cos0 = pnts[4][0];
 226
 227         b1[0x00] = b2[0x00] + b2[0x01];
 228         b1[0x01] = (b2[0x00] - b2[0x01]) * cos0;
 229         b1[0x02] = b2[0x02] + b2[0x03];
 230         b1[0x03] = (b2[0x03] - b2[0x02]) * cos0;
 231         b1[0x02] += b1[0x03];
 232
 233         b1[0x04] = b2[0x04] + b2[0x05];
 234         b1[0x05] = (b2[0x04] - b2[0x05]) * cos0;
 235         b1[0x06] = b2[0x06] + b2[0x07];
 236         b1[0x07] = (b2[0x07] - b2[0x06]) * cos0;
 237         b1[0x06] += b1[0x07];
 238         b1[0x04] += b1[0x06];
 239         b1[0x06] += b1[0x05];
 240         b1[0x05] += b1[0x07];
 241
 242         b1[0x08] = b2[0x08] + b2[0x09];
 243         b1[0x09] = (b2[0x08] - b2[0x09]) * cos0;
 244         b1[0x0A] = b2[0x0A] + b2[0x0B];
 245         b1[0x0B] = (b2[0x0B] - b2[0x0A]) * cos0;
 246         b1[0x0A] += b1[0x0B];
 247
 248         b1[0x0C] = b2[0x0C] + b2[0x0D];
 249         b1[0x0D] = (b2[0x0C] - b2[0x0D]) * cos0;
 250         b1[0x0E] = b2[0x0E] + b2[0x0F];
 251         b1[0x0F] = (b2[0x0F] - b2[0x0E]) * cos0;
 252         b1[0x0E] += b1[0x0F];
 253         b1[0x0C] += b1[0x0E];
 254         b1[0x0E] += b1[0x0D];
 255         b1[0x0D] += b1[0x0F];
 256
 257         b1[0x10] = b2[0x10] + b2[0x11];
 258         b1[0x11] = (b2[0x10] - b2[0x11]) * cos0;
 259         b1[0x12] = b2[0x12] + b2[0x13];
 260         b1[0x13] = (b2[0x13] - b2[0x12]) * cos0;
 261         b1[0x12] += b1[0x13];
 262
 263         b1[0x14] = b2[0x14] + b2[0x15];
 264         b1[0x15] = (b2[0x14] - b2[0x15]) * cos0;
 265         b1[0x16] = b2[0x16] + b2[0x17];
 266         b1[0x17] = (b2[0x17] - b2[0x16]) * cos0;
 267         b1[0x16] += b1[0x17];
 268         b1[0x14] += b1[0x16];
 269         b1[0x16] += b1[0x15];
 270         b1[0x15] += b1[0x17];
 271
 272         b1[0x18] = b2[0x18] + b2[0x19];
 273         b1[0x19] = (b2[0x18] - b2[0x19]) * cos0;
 274         b1[0x1A] = b2[0x1A] + b2[0x1B];
 275         b1[0x1B] = (b2[0x1B] - b2[0x1A]) * cos0;
 276         b1[0x1A] += b1[0x1B];
 277
 278         b1[0x1C] = b2[0x1C] + b2[0x1D];
 279         b1[0x1D] = (b2[0x1C] - b2[0x1D]) * cos0;
 280         b1[0x1E] = b2[0x1E] + b2[0x1F];
 281         b1[0x1F] = (b2[0x1F] - b2[0x1E]) * cos0;
 282         b1[0x1E] += b1[0x1F];
 283         b1[0x1C] += b1[0x1E];
 284         b1[0x1E] += b1[0x1D];
 285         b1[0x1D] += b1[0x1F];
 286     }
 287
 288     out0[0x10 * 16] = b1[0x00];
 289     out0[0x10 * 12] = b1[0x04];
 290     out0[0x10 * 8] = b1[0x02];
 291     out0[0x10 * 4] = b1[0x06];
 292     out0[0x10 * 0] = b1[0x01];
 293     out1[0x10 * 0] = b1[0x01];
 294     out1[0x10 * 4] = b1[0x05];
 295     out1[0x10 * 8] = b1[0x03];
 296     out1[0x10 * 12] = b1[0x07];
 297
 298     b1[0x08] += b1[0x0C];
 299     out0[0x10 * 14] = b1[0x08];
 300     b1[0x0C] += b1[0x0a];
 301     out0[0x10 * 10] = b1[0x0C];
 302     b1[0x0A] += b1[0x0E];
 303     out0[0x10 * 6] = b1[0x0A];
 304     b1[0x0E] += b1[0x09];
 305     out0[0x10 * 2] = b1[0x0E];
 306     b1[0x09] += b1[0x0D];
 307     out1[0x10 * 2] = b1[0x09];
 308     b1[0x0D] += b1[0x0B];
 309     out1[0x10 * 6] = b1[0x0D];
 310     b1[0x0B] += b1[0x0F];
 311     out1[0x10 * 10] = b1[0x0B];
 312     out1[0x10 * 14] = b1[0x0F];
 313
 314     b1[0x18] += b1[0x1C];
 315     out0[0x10 * 15] = b1[0x10] + b1[0x18];
 316     out0[0x10 * 13] = b1[0x18] + b1[0x14];
 317     b1[0x1C] += b1[0x1a];
 318     out0[0x10 * 11] = b1[0x14] + b1[0x1C];
 319     out0[0x10 * 9] = b1[0x1C] + b1[0x12];
 320     b1[0x1A] += b1[0x1E];
 321     out0[0x10 * 7] = b1[0x12] + b1[0x1A];
 322     out0[0x10 * 5] = b1[0x1A] + b1[0x16];
 323     b1[0x1E] += b1[0x19];
 324     out0[0x10 * 3] = b1[0x16] + b1[0x1E];
 325     out0[0x10 * 1] = b1[0x1E] + b1[0x11];
 326     b1[0x19] += b1[0x1D];
 327     out1[0x10 * 1] = b1[0x11] + b1[0x19];
 328     out1[0x10 * 3] = b1[0x19] + b1[0x15];
 329     b1[0x1D] += b1[0x1B];
 330     out1[0x10 * 5] = b1[0x15] + b1[0x1D];
 331     out1[0x10 * 7] = b1[0x1D] + b1[0x13];
 332     b1[0x1B] += b1[0x1F];
 333     out1[0x10 * 9] = b1[0x13] + b1[0x1B];
 334     out1[0x10 * 11] = b1[0x1B] + b1[0x17];
 335     out1[0x10 * 13] = b1[0x17] + b1[0x1F];
 336     out1[0x10 * 15] = b1[0x1F];
 337 }
 338
 339 /*
 340  * the call via dct64 is a trick to force GCC to use
 341  * (new) registers for the b1,b2 pointer to the bufs[xx] field
 342  */
 343 void
 344 dct64(real * a, real * b, real * c)
 345 {
 346     real    bufs[0x40];
 347     dct64_1(a, b, bufs, bufs + 0x20, c);
 348 }