mp3lib/dct64_i386.c

   1 /*
   2  * Modified for use with MPlayer, for details see the CVS changelog at
   3  * http://www.mplayerhq.hu/cgi-bin/cvsweb.cgi/main/
   4  * $Id$
   5  */
   6
   7 /*
   8  * Discrete Cosine Tansform (DCT) for subband synthesis
   9  * optimized for machines with no auto-increment.
  10  * The performance is highly compiler dependend. Maybe
  11  * the mpg123_dct64.c version for 'normal' processor may be faster
  12  * even for Intel processors.
  13  */
  14
  15 #include "mpg123.h"
  16
  17 static void mpg123_dct64_1(real * out0, real * out1, real * b1, real * b2, real * samples)
  18 {
  19
  20         {
  21                 register real *costab = mp3lib_pnts[0];
  22
  23                 b1[0x00] = samples[0x00] + samples[0x1F];
  24                 b1[0x1F] = (samples[0x00] - samples[0x1F]) * costab[0x0];
  25
  26                 b1[0x01] = samples[0x01] + samples[0x1E];
  27                 b1[0x1E] = (samples[0x01] - samples[0x1E]) * costab[0x1];
  28
  29                 b1[0x02] = samples[0x02] + samples[0x1D];
  30                 b1[0x1D] = (samples[0x02] - samples[0x1D]) * costab[0x2];
  31
  32                 b1[0x03] = samples[0x03] + samples[0x1C];
  33                 b1[0x1C] = (samples[0x03] - samples[0x1C]) * costab[0x3];
  34
  35                 b1[0x04] = samples[0x04] + samples[0x1B];
  36                 b1[0x1B] = (samples[0x04] - samples[0x1B]) * costab[0x4];
  37
  38                 b1[0x05] = samples[0x05] + samples[0x1A];
  39                 b1[0x1A] = (samples[0x05] - samples[0x1A]) * costab[0x5];
  40
  41                 b1[0x06] = samples[0x06] + samples[0x19];
  42                 b1[0x19] = (samples[0x06] - samples[0x19]) * costab[0x6];
  43
  44                 b1[0x07] = samples[0x07] + samples[0x18];
  45                 b1[0x18] = (samples[0x07] - samples[0x18]) * costab[0x7];
  46
  47                 b1[0x08] = samples[0x08] + samples[0x17];
  48                 b1[0x17] = (samples[0x08] - samples[0x17]) * costab[0x8];
  49
  50                 b1[0x09] = samples[0x09] + samples[0x16];
  51                 b1[0x16] = (samples[0x09] - samples[0x16]) * costab[0x9];
  52
  53                 b1[0x0A] = samples[0x0A] + samples[0x15];
  54                 b1[0x15] = (samples[0x0A] - samples[0x15]) * costab[0xA];
  55
  56                 b1[0x0B] = samples[0x0B] + samples[0x14];
  57                 b1[0x14] = (samples[0x0B] - samples[0x14]) * costab[0xB];
  58
  59                 b1[0x0C] = samples[0x0C] + samples[0x13];
  60                 b1[0x13] = (samples[0x0C] - samples[0x13]) * costab[0xC];
  61
  62                 b1[0x0D] = samples[0x0D] + samples[0x12];
  63                 b1[0x12] = (samples[0x0D] - samples[0x12]) * costab[0xD];
  64
  65                 b1[0x0E] = samples[0x0E] + samples[0x11];
  66                 b1[0x11] = (samples[0x0E] - samples[0x11]) * costab[0xE];
  67
  68                 b1[0x0F] = samples[0x0F] + samples[0x10];
  69                 b1[0x10] = (samples[0x0F] - samples[0x10]) * costab[0xF];
  70         }
  71
  72         {
  73                 register real *costab = mp3lib_pnts[1];
  74
  75                 b2[0x00] = b1[0x00] + b1[0x0F];
  76                 b2[0x0F] = (b1[0x00] - b1[0x0F]) * costab[0];
  77                 b2[0x01] = b1[0x01] + b1[0x0E];
  78                 b2[0x0E] = (b1[0x01] - b1[0x0E]) * costab[1];
  79                 b2[0x02] = b1[0x02] + b1[0x0D];
  80                 b2[0x0D] = (b1[0x02] - b1[0x0D]) * costab[2];
  81                 b2[0x03] = b1[0x03] + b1[0x0C];
  82                 b2[0x0C] = (b1[0x03] - b1[0x0C]) * costab[3];
  83                 b2[0x04] = b1[0x04] + b1[0x0B];
  84                 b2[0x0B] = (b1[0x04] - b1[0x0B]) * costab[4];
  85                 b2[0x05] = b1[0x05] + b1[0x0A];
  86                 b2[0x0A] = (b1[0x05] - b1[0x0A]) * costab[5];
  87                 b2[0x06] = b1[0x06] + b1[0x09];
  88                 b2[0x09] = (b1[0x06] - b1[0x09]) * costab[6];
  89                 b2[0x07] = b1[0x07] + b1[0x08];
  90                 b2[0x08] = (b1[0x07] - b1[0x08]) * costab[7];
  91
  92                 b2[0x10] = b1[0x10] + b1[0x1F];
  93                 b2[0x1F] = (b1[0x1F] - b1[0x10]) * costab[0];
  94                 b2[0x11] = b1[0x11] + b1[0x1E];
  95                 b2[0x1E] = (b1[0x1E] - b1[0x11]) * costab[1];
  96                 b2[0x12] = b1[0x12] + b1[0x1D];
  97                 b2[0x1D] = (b1[0x1D] - b1[0x12]) * costab[2];
  98                 b2[0x13] = b1[0x13] + b1[0x1C];
  99                 b2[0x1C] = (b1[0x1C] - b1[0x13]) * costab[3];
 100                 b2[0x14] = b1[0x14] + b1[0x1B];
 101                 b2[0x1B] = (b1[0x1B] - b1[0x14]) * costab[4];
 102                 b2[0x15] = b1[0x15] + b1[0x1A];
 103                 b2[0x1A] = (b1[0x1A] - b1[0x15]) * costab[5];
 104                 b2[0x16] = b1[0x16] + b1[0x19];
 105                 b2[0x19] = (b1[0x19] - b1[0x16]) * costab[6];
 106                 b2[0x17] = b1[0x17] + b1[0x18];
 107                 b2[0x18] = (b1[0x18] - b1[0x17]) * costab[7];
 108         }
 109
 110         {
 111                 register real *costab = mp3lib_pnts[2];
 112
 113                 b1[0x00] = b2[0x00] + b2[0x07];
 114                 b1[0x07] = (b2[0x00] - b2[0x07]) * costab[0];
 115                 b1[0x01] = b2[0x01] + b2[0x06];
 116                 b1[0x06] = (b2[0x01] - b2[0x06]) * costab[1];
 117                 b1[0x02] = b2[0x02] + b2[0x05];
 118                 b1[0x05] = (b2[0x02] - b2[0x05]) * costab[2];
 119                 b1[0x03] = b2[0x03] + b2[0x04];
 120                 b1[0x04] = (b2[0x03] - b2[0x04]) * costab[3];
 121
 122                 b1[0x08] = b2[0x08] + b2[0x0F];
 123                 b1[0x0F] = (b2[0x0F] - b2[0x08]) * costab[0];
 124                 b1[0x09] = b2[0x09] + b2[0x0E];
 125                 b1[0x0E] = (b2[0x0E] - b2[0x09]) * costab[1];
 126                 b1[0x0A] = b2[0x0A] + b2[0x0D];
 127                 b1[0x0D] = (b2[0x0D] - b2[0x0A]) * costab[2];
 128                 b1[0x0B] = b2[0x0B] + b2[0x0C];
 129                 b1[0x0C] = (b2[0x0C] - b2[0x0B]) * costab[3];
 130
 131                 b1[0x10] = b2[0x10] + b2[0x17];
 132                 b1[0x17] = (b2[0x10] - b2[0x17]) * costab[0];
 133                 b1[0x11] = b2[0x11] + b2[0x16];
 134                 b1[0x16] = (b2[0x11] - b2[0x16]) * costab[1];
 135                 b1[0x12] = b2[0x12] + b2[0x15];
 136                 b1[0x15] = (b2[0x12] - b2[0x15]) * costab[2];
 137                 b1[0x13] = b2[0x13] + b2[0x14];
 138                 b1[0x14] = (b2[0x13] - b2[0x14]) * costab[3];
 139
 140                 b1[0x18] = b2[0x18] + b2[0x1F];
 141                 b1[0x1F] = (b2[0x1F] - b2[0x18]) * costab[0];
 142                 b1[0x19] = b2[0x19] + b2[0x1E];
 143                 b1[0x1E] = (b2[0x1E] - b2[0x19]) * costab[1];
 144                 b1[0x1A] = b2[0x1A] + b2[0x1D];
 145                 b1[0x1D] = (b2[0x1D] - b2[0x1A]) * costab[2];
 146                 b1[0x1B] = b2[0x1B] + b2[0x1C];
 147                 b1[0x1C] = (b2[0x1C] - b2[0x1B]) * costab[3];
 148         }
 149
 150         {
 151                 register real const cos0 = mp3lib_pnts[3][0];
 152                 register real const cos1 = mp3lib_pnts[3][1];
 153
 154                 b2[0x00] = b1[0x00] + b1[0x03];
 155                 b2[0x03] = (b1[0x00] - b1[0x03]) * cos0;
 156                 b2[0x01] = b1[0x01] + b1[0x02];
 157                 b2[0x02] = (b1[0x01] - b1[0x02]) * cos1;
 158
 159                 b2[0x04] = b1[0x04] + b1[0x07];
 160                 b2[0x07] = (b1[0x07] - b1[0x04]) * cos0;
 161                 b2[0x05] = b1[0x05] + b1[0x06];
 162                 b2[0x06] = (b1[0x06] - b1[0x05]) * cos1;
 163
 164                 b2[0x08] = b1[0x08] + b1[0x0B];
 165                 b2[0x0B] = (b1[0x08] - b1[0x0B]) * cos0;
 166                 b2[0x09] = b1[0x09] + b1[0x0A];
 167                 b2[0x0A] = (b1[0x09] - b1[0x0A]) * cos1;
 168
 169                 b2[0x0C] = b1[0x0C] + b1[0x0F];
 170                 b2[0x0F] = (b1[0x0F] - b1[0x0C]) * cos0;
 171                 b2[0x0D] = b1[0x0D] + b1[0x0E];
 172                 b2[0x0E] = (b1[0x0E] - b1[0x0D]) * cos1;
 173
 174                 b2[0x10] = b1[0x10] + b1[0x13];
 175                 b2[0x13] = (b1[0x10] - b1[0x13]) * cos0;
 176                 b2[0x11] = b1[0x11] + b1[0x12];
 177                 b2[0x12] = (b1[0x11] - b1[0x12]) * cos1;
 178
 179                 b2[0x14] = b1[0x14] + b1[0x17];
 180                 b2[0x17] = (b1[0x17] - b1[0x14]) * cos0;
 181                 b2[0x15] = b1[0x15] + b1[0x16];
 182                 b2[0x16] = (b1[0x16] - b1[0x15]) * cos1;
 183
 184                 b2[0x18] = b1[0x18] + b1[0x1B];
 185                 b2[0x1B] = (b1[0x18] - b1[0x1B]) * cos0;
 186                 b2[0x19] = b1[0x19] + b1[0x1A];
 187                 b2[0x1A] = (b1[0x19] - b1[0x1A]) * cos1;
 188
 189                 b2[0x1C] = b1[0x1C] + b1[0x1F];
 190                 b2[0x1F] = (b1[0x1F] - b1[0x1C]) * cos0;
 191                 b2[0x1D] = b1[0x1D] + b1[0x1E];
 192                 b2[0x1E] = (b1[0x1E] - b1[0x1D]) * cos1;
 193         }
 194
 195         {
 196                 register real const cos0 = mp3lib_pnts[4][0];
 197
 198                 b1[0x00] = b2[0x00] + b2[0x01];
 199                 b1[0x01] = (b2[0x00] - b2[0x01]) * cos0;
 200                 b1[0x02] = b2[0x02] + b2[0x03];
 201                 b1[0x03] = (b2[0x03] - b2[0x02]) * cos0;
 202                 b1[0x02] += b1[0x03];
 203
 204                 b1[0x04] = b2[0x04] + b2[0x05];
 205                 b1[0x05] = (b2[0x04] - b2[0x05]) * cos0;
 206                 b1[0x06] = b2[0x06] + b2[0x07];
 207                 b1[0x07] = (b2[0x07] - b2[0x06]) * cos0;
 208                 b1[0x06] += b1[0x07];
 209                 b1[0x04] += b1[0x06];
 210                 b1[0x06] += b1[0x05];
 211                 b1[0x05] += b1[0x07];
 212
 213                 b1[0x08] = b2[0x08] + b2[0x09];
 214                 b1[0x09] = (b2[0x08] - b2[0x09]) * cos0;
 215                 b1[0x0A] = b2[0x0A] + b2[0x0B];
 216                 b1[0x0B] = (b2[0x0B] - b2[0x0A]) * cos0;
 217                 b1[0x0A] += b1[0x0B];
 218
 219                 b1[0x0C] = b2[0x0C] + b2[0x0D];
 220                 b1[0x0D] = (b2[0x0C] - b2[0x0D]) * cos0;
 221                 b1[0x0E] = b2[0x0E] + b2[0x0F];
 222                 b1[0x0F] = (b2[0x0F] - b2[0x0E]) * cos0;
 223                 b1[0x0E] += b1[0x0F];
 224                 b1[0x0C] += b1[0x0E];
 225                 b1[0x0E] += b1[0x0D];
 226                 b1[0x0D] += b1[0x0F];
 227
 228                 b1[0x10] = b2[0x10] + b2[0x11];
 229                 b1[0x11] = (b2[0x10] - b2[0x11]) * cos0;
 230                 b1[0x12] = b2[0x12] + b2[0x13];
 231                 b1[0x13] = (b2[0x13] - b2[0x12]) * cos0;
 232                 b1[0x12] += b1[0x13];
 233
 234                 b1[0x14] = b2[0x14] + b2[0x15];
 235                 b1[0x15] = (b2[0x14] - b2[0x15]) * cos0;
 236                 b1[0x16] = b2[0x16] + b2[0x17];
 237                 b1[0x17] = (b2[0x17] - b2[0x16]) * cos0;
 238                 b1[0x16] += b1[0x17];
 239                 b1[0x14] += b1[0x16];
 240                 b1[0x16] += b1[0x15];
 241                 b1[0x15] += b1[0x17];
 242
 243                 b1[0x18] = b2[0x18] + b2[0x19];
 244                 b1[0x19] = (b2[0x18] - b2[0x19]) * cos0;
 245                 b1[0x1A] = b2[0x1A] + b2[0x1B];
 246                 b1[0x1B] = (b2[0x1B] - b2[0x1A]) * cos0;
 247                 b1[0x1A] += b1[0x1B];
 248
 249                 b1[0x1C] = b2[0x1C] + b2[0x1D];
 250                 b1[0x1D] = (b2[0x1C] - b2[0x1D]) * cos0;
 251                 b1[0x1E] = b2[0x1E] + b2[0x1F];
 252                 b1[0x1F] = (b2[0x1F] - b2[0x1E]) * cos0;
 253                 b1[0x1E] += b1[0x1F];
 254                 b1[0x1C] += b1[0x1E];
 255                 b1[0x1E] += b1[0x1D];
 256                 b1[0x1D] += b1[0x1F];
 257         }
 258
 259         out0[0x10 * 16] = b1[0x00];
 260         out0[0x10 * 12] = b1[0x04];
 261         out0[0x10 * 8] = b1[0x02];
 262         out0[0x10 * 4] = b1[0x06];
 263         out0[0x10 * 0] = b1[0x01];
 264         out1[0x10 * 0] = b1[0x01];
 265         out1[0x10 * 4] = b1[0x05];
 266         out1[0x10 * 8] = b1[0x03];
 267         out1[0x10 * 12] = b1[0x07];
 268
 269         b1[0x08] += b1[0x0C];
 270         out0[0x10 * 14] = b1[0x08];
 271         b1[0x0C] += b1[0x0a];
 272         out0[0x10 * 10] = b1[0x0C];
 273         b1[0x0A] += b1[0x0E];
 274         out0[0x10 * 6] = b1[0x0A];
 275         b1[0x0E] += b1[0x09];
 276         out0[0x10 * 2] = b1[0x0E];
 277         b1[0x09] += b1[0x0D];
 278         out1[0x10 * 2] = b1[0x09];
 279         b1[0x0D] += b1[0x0B];
 280         out1[0x10 * 6] = b1[0x0D];
 281         b1[0x0B] += b1[0x0F];
 282         out1[0x10 * 10] = b1[0x0B];
 283         out1[0x10 * 14] = b1[0x0F];
 284
 285         b1[0x18] += b1[0x1C];
 286         out0[0x10 * 15] = b1[0x10] + b1[0x18];
 287         out0[0x10 * 13] = b1[0x18] + b1[0x14];
 288         b1[0x1C] += b1[0x1a];
 289         out0[0x10 * 11] = b1[0x14] + b1[0x1C];
 290         out0[0x10 * 9] = b1[0x1C] + b1[0x12];
 291         b1[0x1A] += b1[0x1E];
 292         out0[0x10 * 7] = b1[0x12] + b1[0x1A];
 293         out0[0x10 * 5] = b1[0x1A] + b1[0x16];
 294         b1[0x1E] += b1[0x19];
 295         out0[0x10 * 3] = b1[0x16] + b1[0x1E];
 296         out0[0x10 * 1] = b1[0x1E] + b1[0x11];
 297         b1[0x19] += b1[0x1D];
 298         out1[0x10 * 1] = b1[0x11] + b1[0x19];
 299         out1[0x10 * 3] = b1[0x19] + b1[0x15];
 300         b1[0x1D] += b1[0x1B];
 301         out1[0x10 * 5] = b1[0x15] + b1[0x1D];
 302         out1[0x10 * 7] = b1[0x1D] + b1[0x13];
 303         b1[0x1B] += b1[0x1F];
 304         out1[0x10 * 9] = b1[0x13] + b1[0x1B];
 305         out1[0x10 * 11] = b1[0x1B] + b1[0x17];
 306         out1[0x10 * 13] = b1[0x17] + b1[0x1F];
 307         out1[0x10 * 15] = b1[0x1F];
 308 }
 309
 310 /*
 311  * the call via mpg123_dct64 is a trick to force GCC to use
 312  * (new) registers for the b1,b2 pointer to the bufs[xx] field
 313  */
 314 void mpg123_dct64(real * a, real * b, real * c)
 315 {
 316         real bufs[0x40];
 317
 318         mpg123_dct64_1(a, b, bufs, bufs + 0x20, c);
 319 }