mp3lib/dct64_altivec.c

   1
   2 /*
   3  * Discrete Cosine Tansform (DCT) for Altivec
   4  * Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org>
   5  * based upon code from "mp3lib/dct64.c"
   6  * This file is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public License
   8  */
   9
  10 #define real float
  11
  12 #include <stdio.h>
  13 #include "mpg123.h"
  14
  15 #ifdef HAVE_ALTIVEC_H
  16 #include <altivec.h>
  17 #endif
  18
  19 // used to build registers permutation vectors (vcprm)
  20 // the 's' are for words in the _s_econd vector
  21 #define WORD_0 0x00,0x01,0x02,0x03
  22 #define WORD_1 0x04,0x05,0x06,0x07
  23 #define WORD_2 0x08,0x09,0x0a,0x0b
  24 #define WORD_3 0x0c,0x0d,0x0e,0x0f
  25 #define WORD_s0 0x10,0x11,0x12,0x13
  26 #define WORD_s1 0x14,0x15,0x16,0x17
  27 #define WORD_s2 0x18,0x19,0x1a,0x1b
  28 #define WORD_s3 0x1c,0x1d,0x1e,0x1f
  29
  30 #ifdef __APPLE_CC__
  31 #define AVV(x...) (x)
  32 #else
  33 #define AVV(x...) {x}
  34 #endif
  35
  36 #define vcprm(a,b,c,d) (const vector unsigned char)AVV(WORD_ ## a, WORD_ ## b, WORD_ ## c, WORD_ ## d)
  37 #define vcii(a,b,c,d) (const vector float)AVV(FLOAT_ ## a, FLOAT_ ## b, FLOAT_ ## c, FLOAT_ ## d)
  38
  39 #define FOUROF(a) AVV(a,a,a,a)
  40
  41 // vcprmle is used to keep the same index as in the SSE version.
  42 // it's the same as vcprm, with the index inversed
  43 // ('le' is Little Endian)
  44 #define vcprmle(a,b,c,d) vcprm(d,c,b,a)
  45
  46 // used to build inverse/identity vectors (vcii)
  47 // n is _n_egative, p is _p_ositive
  48 #define FLOAT_n -1.
  49 #define FLOAT_p 1.
  50
  51 void dct64_altivec(real *a,real *b,real *c)
  52 {
  53   real __attribute__ ((aligned(16))) b1[0x20];
  54   real __attribute__ ((aligned(16))) b2[0x20];
  55
  56   real *out0 = a;
  57   real *out1 = b;
  58   real *samples = c;
  59
  60   const vector float vczero = (const vector float)FOUROF(0.);
  61   const vector unsigned char reverse = (const vector unsigned char)vcprm(3,2,1,0);
  62
  63
  64   if (((unsigned long)b1 & 0x0000000F) ||
  65       ((unsigned long)b2 & 0x0000000F))
  66
  67   {
  68     printf("MISALIGNED:\t%p\t%p\t%p\t%p\t%p\n",
  69            b1, b2, a, b, samples);
  70   }
  71
  72
  73 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
  74
  75   {
  76     register real *costab = mp3lib_pnts[0];
  77
  78     b1[0x00] = samples[0x00] + samples[0x1F];
  79     b1[0x01] = samples[0x01] + samples[0x1E];
  80     b1[0x02] = samples[0x02] + samples[0x1D];
  81     b1[0x03] = samples[0x03] + samples[0x1C];
  82     b1[0x04] = samples[0x04] + samples[0x1B];
  83     b1[0x05] = samples[0x05] + samples[0x1A];
  84     b1[0x06] = samples[0x06] + samples[0x19];
  85     b1[0x07] = samples[0x07] + samples[0x18];
  86     b1[0x08] = samples[0x08] + samples[0x17];
  87     b1[0x09] = samples[0x09] + samples[0x16];
  88     b1[0x0A] = samples[0x0A] + samples[0x15];
  89     b1[0x0B] = samples[0x0B] + samples[0x14];
  90     b1[0x0C] = samples[0x0C] + samples[0x13];
  91     b1[0x0D] = samples[0x0D] + samples[0x12];
  92     b1[0x0E] = samples[0x0E] + samples[0x11];
  93     b1[0x0F] = samples[0x0F] + samples[0x10];
  94     b1[0x10] = (samples[0x0F] - samples[0x10]) * costab[0xF];
  95     b1[0x11] = (samples[0x0E] - samples[0x11]) * costab[0xE];
  96     b1[0x12] = (samples[0x0D] - samples[0x12]) * costab[0xD];
  97     b1[0x13] = (samples[0x0C] - samples[0x13]) * costab[0xC];
  98     b1[0x14] = (samples[0x0B] - samples[0x14]) * costab[0xB];
  99     b1[0x15] = (samples[0x0A] - samples[0x15]) * costab[0xA];
 100     b1[0x16] = (samples[0x09] - samples[0x16]) * costab[0x9];
 101     b1[0x17] = (samples[0x08] - samples[0x17]) * costab[0x8];
 102     b1[0x18] = (samples[0x07] - samples[0x18]) * costab[0x7];
 103     b1[0x19] = (samples[0x06] - samples[0x19]) * costab[0x6];
 104     b1[0x1A] = (samples[0x05] - samples[0x1A]) * costab[0x5];
 105     b1[0x1B] = (samples[0x04] - samples[0x1B]) * costab[0x4];
 106     b1[0x1C] = (samples[0x03] - samples[0x1C]) * costab[0x3];
 107     b1[0x1D] = (samples[0x02] - samples[0x1D]) * costab[0x2];
 108     b1[0x1E] = (samples[0x01] - samples[0x1E]) * costab[0x1];
 109     b1[0x1F] = (samples[0x00] - samples[0x1F]) * costab[0x0];
 110
 111   }
 112   {
 113     register real *costab = mp3lib_pnts[1];
 114
 115     b2[0x00] = b1[0x00] + b1[0x0F];
 116     b2[0x01] = b1[0x01] + b1[0x0E];
 117     b2[0x02] = b1[0x02] + b1[0x0D];
 118     b2[0x03] = b1[0x03] + b1[0x0C];
 119     b2[0x04] = b1[0x04] + b1[0x0B];
 120     b2[0x05] = b1[0x05] + b1[0x0A];
 121     b2[0x06] = b1[0x06] + b1[0x09];
 122     b2[0x07] = b1[0x07] + b1[0x08];
 123     b2[0x08] = (b1[0x07] - b1[0x08]) * costab[7];
 124     b2[0x09] = (b1[0x06] - b1[0x09]) * costab[6];
 125     b2[0x0A] = (b1[0x05] - b1[0x0A]) * costab[5];
 126     b2[0x0B] = (b1[0x04] - b1[0x0B]) * costab[4];
 127     b2[0x0C] = (b1[0x03] - b1[0x0C]) * costab[3];
 128     b2[0x0D] = (b1[0x02] - b1[0x0D]) * costab[2];
 129     b2[0x0E] = (b1[0x01] - b1[0x0E]) * costab[1];
 130     b2[0x0F] = (b1[0x00] - b1[0x0F]) * costab[0];
 131     b2[0x10] = b1[0x10] + b1[0x1F];
 132     b2[0x11] = b1[0x11] + b1[0x1E];
 133     b2[0x12] = b1[0x12] + b1[0x1D];
 134     b2[0x13] = b1[0x13] + b1[0x1C];
 135     b2[0x14] = b1[0x14] + b1[0x1B];
 136     b2[0x15] = b1[0x15] + b1[0x1A];
 137     b2[0x16] = b1[0x16] + b1[0x19];
 138     b2[0x17] = b1[0x17] + b1[0x18];
 139     b2[0x18] = (b1[0x18] - b1[0x17]) * costab[7];
 140     b2[0x19] = (b1[0x19] - b1[0x16]) * costab[6];
 141     b2[0x1A] = (b1[0x1A] - b1[0x15]) * costab[5];
 142     b2[0x1B] = (b1[0x1B] - b1[0x14]) * costab[4];
 143     b2[0x1C] = (b1[0x1C] - b1[0x13]) * costab[3];
 144     b2[0x1D] = (b1[0x1D] - b1[0x12]) * costab[2];
 145     b2[0x1E] = (b1[0x1E] - b1[0x11]) * costab[1];
 146     b2[0x1F] = (b1[0x1F] - b1[0x10]) * costab[0];
 147
 148   }
 149
 150   {
 151     register real *costab = mp3lib_pnts[2];
 152
 153     b1[0x00] = b2[0x00] + b2[0x07];
 154     b1[0x01] = b2[0x01] + b2[0x06];
 155     b1[0x02] = b2[0x02] + b2[0x05];
 156     b1[0x03] = b2[0x03] + b2[0x04];
 157     b1[0x04] = (b2[0x03] - b2[0x04]) * costab[3];
 158     b1[0x05] = (b2[0x02] - b2[0x05]) * costab[2];
 159     b1[0x06] = (b2[0x01] - b2[0x06]) * costab[1];
 160     b1[0x07] = (b2[0x00] - b2[0x07]) * costab[0];
 161     b1[0x08] = b2[0x08] + b2[0x0F];
 162     b1[0x09] = b2[0x09] + b2[0x0E];
 163     b1[0x0A] = b2[0x0A] + b2[0x0D];
 164     b1[0x0B] = b2[0x0B] + b2[0x0C];
 165     b1[0x0C] = (b2[0x0C] - b2[0x0B]) * costab[3];
 166     b1[0x0D] = (b2[0x0D] - b2[0x0A]) * costab[2];
 167     b1[0x0E] = (b2[0x0E] - b2[0x09]) * costab[1];
 168     b1[0x0F] = (b2[0x0F] - b2[0x08]) * costab[0];
 169     b1[0x10] = b2[0x10] + b2[0x17];
 170     b1[0x11] = b2[0x11] + b2[0x16];
 171     b1[0x12] = b2[0x12] + b2[0x15];
 172     b1[0x13] = b2[0x13] + b2[0x14];
 173     b1[0x14] = (b2[0x13] - b2[0x14]) * costab[3];
 174     b1[0x15] = (b2[0x12] - b2[0x15]) * costab[2];
 175     b1[0x16] = (b2[0x11] - b2[0x16]) * costab[1];
 176     b1[0x17] = (b2[0x10] - b2[0x17]) * costab[0];
 177     b1[0x18] = b2[0x18] + b2[0x1F];
 178     b1[0x19] = b2[0x19] + b2[0x1E];
 179     b1[0x1A] = b2[0x1A] + b2[0x1D];
 180     b1[0x1B] = b2[0x1B] + b2[0x1C];
 181     b1[0x1C] = (b2[0x1C] - b2[0x1B]) * costab[3];
 182     b1[0x1D] = (b2[0x1D] - b2[0x1A]) * costab[2];
 183     b1[0x1E] = (b2[0x1E] - b2[0x19]) * costab[1];
 184     b1[0x1F] = (b2[0x1F] - b2[0x18]) * costab[0];
 185   }
 186
 187 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
 188
 189   // How does it work ?
 190   // the first three passes are reproducted in the three block below
 191   // all computations are done on a 4 elements vector
 192   // 'reverse' is a special perumtation vector used to reverse
 193   // the order of the elements inside a vector.
 194   // note that all loads/stores to b1 (b2) between passes 1 and 2 (2 and 3)
 195   // have been removed, all elements are stored inside b1vX (b2vX)
 196   {
 197     register vector float
 198       b1v0, b1v1, b1v2, b1v3,
 199       b1v4, b1v5, b1v6, b1v7;
 200     register vector float
 201       temp1, temp2;
 202
 203     {
 204       register real *costab = mp3lib_pnts[0];
 205
 206       register vector float
 207         samplesv1, samplesv2, samplesv3, samplesv4,
 208         samplesv5, samplesv6, samplesv7, samplesv8,
 209         samplesv9;
 210       register vector unsigned char samples_perm = vec_lvsl(0, samples);
 211       register vector float costabv1, costabv2, costabv3, costabv4, costabv5;
 212       register vector unsigned char costab_perm = vec_lvsl(0, costab);
 213
 214       samplesv1 = vec_ld(0, samples);
 215       samplesv2 = vec_ld(16, samples);
 216       samplesv1 = vec_perm(samplesv1, samplesv2, samples_perm);
 217       samplesv3 = vec_ld(32, samples);
 218       samplesv2 = vec_perm(samplesv2, samplesv3, samples_perm);
 219       samplesv4 = vec_ld(48, samples);
 220       samplesv3 = vec_perm(samplesv3, samplesv4, samples_perm);
 221       samplesv5 = vec_ld(64, samples);
 222       samplesv4 = vec_perm(samplesv4, samplesv5, samples_perm);
 223       samplesv6 = vec_ld(80, samples);
 224       samplesv5 = vec_perm(samplesv5, samplesv6, samples_perm);
 225       samplesv7 = vec_ld(96, samples);
 226       samplesv6 = vec_perm(samplesv6, samplesv7, samples_perm);
 227       samplesv8 = vec_ld(112, samples);
 228       samplesv7 = vec_perm(samplesv7, samplesv8, samples_perm);
 229       samplesv9 = vec_ld(128, samples);
 230       samplesv8 = vec_perm(samplesv8, samplesv9, samples_perm);
 231
 232       temp1 = vec_add(samplesv1,
 233                       vec_perm(samplesv8, samplesv8, reverse));
 234       //vec_st(temp1, 0, b1);
 235       b1v0 = temp1;
 236       temp1 = vec_add(samplesv2,
 237                       vec_perm(samplesv7, samplesv7, reverse));
 238       //vec_st(temp1, 16, b1);
 239       b1v1 = temp1;
 240       temp1 = vec_add(samplesv3,
 241                       vec_perm(samplesv6, samplesv6, reverse));
 242       //vec_st(temp1, 32, b1);
 243       b1v2 = temp1;
 244       temp1 = vec_add(samplesv4,
 245                       vec_perm(samplesv5, samplesv5, reverse));
 246       //vec_st(temp1, 48, b1);
 247       b1v3 = temp1;
 248
 249       costabv1 = vec_ld(0, costab);
 250       costabv2 = vec_ld(16, costab);
 251       costabv1 = vec_perm(costabv1, costabv2, costab_perm);
 252       costabv3 = vec_ld(32, costab);
 253       costabv2 = vec_perm(costabv2, costabv3, costab_perm);
 254       costabv4 = vec_ld(48, costab);
 255       costabv3 = vec_perm(costabv3, costabv4, costab_perm);
 256       costabv5 = vec_ld(64, costab);
 257       costabv4 = vec_perm(costabv4, costabv5, costab_perm);
 258
 259       temp1 = vec_sub(vec_perm(samplesv4, samplesv4, reverse),
 260                       samplesv5);
 261       temp2 = vec_madd(temp1,
 262                        vec_perm(costabv4, costabv4, reverse),
 263                        vczero);
 264       //vec_st(temp2, 64, b1);
 265       b1v4 = temp2;
 266
 267       temp1 = vec_sub(vec_perm(samplesv3, samplesv3, reverse),
 268                       samplesv6);
 269       temp2 = vec_madd(temp1,
 270                        vec_perm(costabv3, costabv3, reverse),
 271                        vczero);
 272       //vec_st(temp2, 80, b1);
 273       b1v5 = temp2;
 274       temp1 = vec_sub(vec_perm(samplesv2, samplesv2, reverse),
 275                       samplesv7);
 276       temp2 = vec_madd(temp1,
 277                        vec_perm(costabv2, costabv2, reverse),
 278                        vczero);
 279       //vec_st(temp2, 96, b1);
 280       b1v6 = temp2;
 281
 282       temp1 = vec_sub(vec_perm(samplesv1, samplesv1, reverse),
 283                       samplesv8);
 284       temp2 = vec_madd(temp1,
 285                        vec_perm(costabv1, costabv1, reverse),
 286                        vczero);
 287       //vec_st(temp2, 112, b1);
 288       b1v7 = temp2;
 289
 290     }
 291
 292     {
 293       register vector float
 294         b2v0, b2v1, b2v2, b2v3,
 295         b2v4, b2v5, b2v6, b2v7;
 296       {
 297         register real *costab = mp3lib_pnts[1];
 298         register vector float costabv1r, costabv2r, costabv1, costabv2, costabv3;
 299         register vector unsigned char costab_perm = vec_lvsl(0, costab);
 300
 301         costabv1 = vec_ld(0, costab);
 302         costabv2 = vec_ld(16, costab);
 303         costabv1 = vec_perm(costabv1, costabv2, costab_perm);
 304         costabv3  = vec_ld(32, costab);
 305         costabv2 = vec_perm(costabv2, costabv3 , costab_perm);
 306         costabv1r = vec_perm(costabv1, costabv1, reverse);
 307         costabv2r = vec_perm(costabv2, costabv2, reverse);
 308
 309         temp1 = vec_add(b1v0, vec_perm(b1v3, b1v3, reverse));
 310         //vec_st(temp1, 0, b2);
 311         b2v0 = temp1;
 312         temp1 = vec_add(b1v1, vec_perm(b1v2, b1v2, reverse));
 313         //vec_st(temp1, 16, b2);
 314         b2v1 = temp1;
 315         temp2 = vec_sub(vec_perm(b1v1, b1v1, reverse), b1v2);
 316         temp1 = vec_madd(temp2, costabv2r, vczero);
 317         //vec_st(temp1, 32, b2);
 318         b2v2 = temp1;
 319         temp2 = vec_sub(vec_perm(b1v0, b1v0, reverse), b1v3);
 320         temp1 = vec_madd(temp2, costabv1r, vczero);
 321         //vec_st(temp1, 48, b2);
 322         b2v3 = temp1;
 323         temp1 = vec_add(b1v4, vec_perm(b1v7, b1v7, reverse));
 324         //vec_st(temp1, 64, b2);
 325         b2v4 = temp1;
 326         temp1 = vec_add(b1v5, vec_perm(b1v6, b1v6, reverse));
 327         //vec_st(temp1, 80, b2);
 328         b2v5 = temp1;
 329         temp2 = vec_sub(b1v6, vec_perm(b1v5, b1v5, reverse));
 330         temp1 = vec_madd(temp2, costabv2r, vczero);
 331         //vec_st(temp1, 96, b2);
 332         b2v6 = temp1;
 333         temp2 = vec_sub(b1v7, vec_perm(b1v4, b1v4, reverse));
 334         temp1 = vec_madd(temp2, costabv1r, vczero);
 335         //vec_st(temp1, 112, b2);
 336         b2v7 = temp1;
 337       }
 338
 339       {
 340         register real *costab = mp3lib_pnts[2];
 341
 342
 343         vector float costabv1r, costabv1, costabv2;
 344         vector unsigned char costab_perm = vec_lvsl(0, costab);
 345
 346         costabv1 = vec_ld(0, costab);
 347         costabv2 = vec_ld(16, costab);
 348         costabv1 = vec_perm(costabv1, costabv2, costab_perm);
 349         costabv1r = vec_perm(costabv1, costabv1, reverse);
 350
 351         temp1 = vec_add(b2v0, vec_perm(b2v1, b2v1, reverse));
 352         vec_st(temp1, 0, b1);
 353         temp2 = vec_sub(vec_perm(b2v0, b2v0, reverse), b2v1);
 354         temp1 = vec_madd(temp2, costabv1r, vczero);
 355         vec_st(temp1, 16, b1);
 356
 357         temp1 = vec_add(b2v2, vec_perm(b2v3, b2v3, reverse));
 358         vec_st(temp1, 32, b1);
 359         temp2 = vec_sub(b2v3, vec_perm(b2v2, b2v2, reverse));
 360         temp1 = vec_madd(temp2, costabv1r, vczero);
 361         vec_st(temp1, 48, b1);
 362
 363         temp1 = vec_add(b2v4, vec_perm(b2v5, b2v5, reverse));
 364         vec_st(temp1, 64, b1);
 365         temp2 = vec_sub(vec_perm(b2v4, b2v4, reverse), b2v5);
 366         temp1 = vec_madd(temp2, costabv1r, vczero);
 367         vec_st(temp1, 80, b1);
 368
 369         temp1 = vec_add(b2v6, vec_perm(b2v7, b2v7, reverse));
 370         vec_st(temp1, 96, b1);
 371         temp2 = vec_sub(b2v7, vec_perm(b2v6, b2v6, reverse));
 372         temp1 = vec_madd(temp2, costabv1r, vczero);
 373         vec_st(temp1, 112, b1);
 374
 375       }
 376     }
 377   }
 378
 379 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
 380
 381   {
 382     register real const cos0 = mp3lib_pnts[3][0];
 383     register real const cos1 = mp3lib_pnts[3][1];
 384
 385     b2[0x00] = b1[0x00] + b1[0x03];
 386     b2[0x01] = b1[0x01] + b1[0x02];
 387     b2[0x02] = (b1[0x01] - b1[0x02]) * cos1;
 388     b2[0x03] = (b1[0x00] - b1[0x03]) * cos0;
 389     b2[0x04] = b1[0x04] + b1[0x07];
 390     b2[0x05] = b1[0x05] + b1[0x06];
 391     b2[0x06] = (b1[0x06] - b1[0x05]) * cos1;
 392     b2[0x07] = (b1[0x07] - b1[0x04]) * cos0;
 393     b2[0x08] = b1[0x08] + b1[0x0B];
 394     b2[0x09] = b1[0x09] + b1[0x0A];
 395     b2[0x0A] = (b1[0x09] - b1[0x0A]) * cos1;
 396     b2[0x0B] = (b1[0x08] - b1[0x0B]) * cos0;
 397     b2[0x0C] = b1[0x0C] + b1[0x0F];
 398     b2[0x0D] = b1[0x0D] + b1[0x0E];
 399     b2[0x0E] = (b1[0x0E] - b1[0x0D]) * cos1;
 400     b2[0x0F] = (b1[0x0F] - b1[0x0C]) * cos0;
 401     b2[0x10] = b1[0x10] + b1[0x13];
 402     b2[0x11] = b1[0x11] + b1[0x12];
 403     b2[0x12] = (b1[0x11] - b1[0x12]) * cos1;
 404     b2[0x13] = (b1[0x10] - b1[0x13]) * cos0;
 405     b2[0x14] = b1[0x14] + b1[0x17];
 406     b2[0x15] = b1[0x15] + b1[0x16];
 407     b2[0x16] = (b1[0x16] - b1[0x15]) * cos1;
 408     b2[0x17] = (b1[0x17] - b1[0x14]) * cos0;
 409     b2[0x18] = b1[0x18] + b1[0x1B];
 410     b2[0x19] = b1[0x19] + b1[0x1A];
 411     b2[0x1A] = (b1[0x19] - b1[0x1A]) * cos1;
 412     b2[0x1B] = (b1[0x18] - b1[0x1B]) * cos0;
 413     b2[0x1C] = b1[0x1C] + b1[0x1F];
 414     b2[0x1D] = b1[0x1D] + b1[0x1E];
 415     b2[0x1E] = (b1[0x1E] - b1[0x1D]) * cos1;
 416     b2[0x1F] = (b1[0x1F] - b1[0x1C]) * cos0;
 417   }
 418
 419   {
 420     register real const cos0 = mp3lib_pnts[4][0];
 421
 422     b1[0x00] = b2[0x00] + b2[0x01];
 423     b1[0x01] = (b2[0x00] - b2[0x01]) * cos0;
 424     b1[0x02] = b2[0x02] + b2[0x03];
 425     b1[0x03] = (b2[0x03] - b2[0x02]) * cos0;
 426     b1[0x02] += b1[0x03];
 427
 428     b1[0x04] = b2[0x04] + b2[0x05];
 429     b1[0x05] = (b2[0x04] - b2[0x05]) * cos0;
 430     b1[0x06] = b2[0x06] + b2[0x07];
 431     b1[0x07] = (b2[0x07] - b2[0x06]) * cos0;
 432     b1[0x06] += b1[0x07];
 433     b1[0x04] += b1[0x06];
 434     b1[0x06] += b1[0x05];
 435     b1[0x05] += b1[0x07];
 436
 437     b1[0x08] = b2[0x08] + b2[0x09];
 438     b1[0x09] = (b2[0x08] - b2[0x09]) * cos0;
 439     b1[0x0A] = b2[0x0A] + b2[0x0B];
 440     b1[0x0B] = (b2[0x0B] - b2[0x0A]) * cos0;
 441     b1[0x0A] += b1[0x0B];
 442
 443     b1[0x0C] = b2[0x0C] + b2[0x0D];
 444     b1[0x0D] = (b2[0x0C] - b2[0x0D]) * cos0;
 445     b1[0x0E] = b2[0x0E] + b2[0x0F];
 446     b1[0x0F] = (b2[0x0F] - b2[0x0E]) * cos0;
 447     b1[0x0E] += b1[0x0F];
 448     b1[0x0C] += b1[0x0E];
 449     b1[0x0E] += b1[0x0D];
 450     b1[0x0D] += b1[0x0F];
 451
 452     b1[0x10] = b2[0x10] + b2[0x11];
 453     b1[0x11] = (b2[0x10] - b2[0x11]) * cos0;
 454     b1[0x12] = b2[0x12] + b2[0x13];
 455     b1[0x13] = (b2[0x13] - b2[0x12]) * cos0;
 456     b1[0x12] += b1[0x13];
 457
 458     b1[0x14] = b2[0x14] + b2[0x15];
 459     b1[0x15] = (b2[0x14] - b2[0x15]) * cos0;
 460     b1[0x16] = b2[0x16] + b2[0x17];
 461     b1[0x17] = (b2[0x17] - b2[0x16]) * cos0;
 462     b1[0x16] += b1[0x17];
 463     b1[0x14] += b1[0x16];
 464     b1[0x16] += b1[0x15];
 465     b1[0x15] += b1[0x17];
 466
 467     b1[0x18] = b2[0x18] + b2[0x19];
 468     b1[0x19] = (b2[0x18] - b2[0x19]) * cos0;
 469     b1[0x1A] = b2[0x1A] + b2[0x1B];
 470     b1[0x1B] = (b2[0x1B] - b2[0x1A]) * cos0;
 471     b1[0x1A] += b1[0x1B];
 472
 473     b1[0x1C] = b2[0x1C] + b2[0x1D];
 474     b1[0x1D] = (b2[0x1C] - b2[0x1D]) * cos0;
 475     b1[0x1E] = b2[0x1E] + b2[0x1F];
 476     b1[0x1F] = (b2[0x1F] - b2[0x1E]) * cos0;
 477     b1[0x1E] += b1[0x1F];
 478     b1[0x1C] += b1[0x1E];
 479     b1[0x1E] += b1[0x1D];
 480     b1[0x1D] += b1[0x1F];
 481   }
 482
 483   out0[0x10*16] = b1[0x00];
 484   out0[0x10*12] = b1[0x04];
 485   out0[0x10* 8] = b1[0x02];
 486   out0[0x10* 4] = b1[0x06];
 487   out0[0x10* 0] = b1[0x01];
 488   out1[0x10* 0] = b1[0x01];
 489   out1[0x10* 4] = b1[0x05];
 490   out1[0x10* 8] = b1[0x03];
 491   out1[0x10*12] = b1[0x07];
 492
 493   b1[0x08] += b1[0x0C];
 494   out0[0x10*14] = b1[0x08];
 495   b1[0x0C] += b1[0x0a];
 496   out0[0x10*10] = b1[0x0C];
 497   b1[0x0A] += b1[0x0E];
 498   out0[0x10* 6] = b1[0x0A];
 499   b1[0x0E] += b1[0x09];
 500   out0[0x10* 2] = b1[0x0E];
 501   b1[0x09] += b1[0x0D];
 502   out1[0x10* 2] = b1[0x09];
 503   b1[0x0D] += b1[0x0B];
 504   out1[0x10* 6] = b1[0x0D];
 505   b1[0x0B] += b1[0x0F];
 506   out1[0x10*10] = b1[0x0B];
 507   out1[0x10*14] = b1[0x0F];
 508
 509   b1[0x18] += b1[0x1C];
 510   out0[0x10*15] = b1[0x10] + b1[0x18];
 511   out0[0x10*13] = b1[0x18] + b1[0x14];
 512   b1[0x1C] += b1[0x1a];
 513   out0[0x10*11] = b1[0x14] + b1[0x1C];
 514   out0[0x10* 9] = b1[0x1C] + b1[0x12];
 515   b1[0x1A] += b1[0x1E];
 516   out0[0x10* 7] = b1[0x12] + b1[0x1A];
 517   out0[0x10* 5] = b1[0x1A] + b1[0x16];
 518   b1[0x1E] += b1[0x19];
 519   out0[0x10* 3] = b1[0x16] + b1[0x1E];
 520   out0[0x10* 1] = b1[0x1E] + b1[0x11];
 521   b1[0x19] += b1[0x1D];
 522   out1[0x10* 1] = b1[0x11] + b1[0x19];
 523   out1[0x10* 3] = b1[0x19] + b1[0x15];
 524   b1[0x1D] += b1[0x1B];
 525   out1[0x10* 5] = b1[0x15] + b1[0x1D];
 526   out1[0x10* 7] = b1[0x1D] + b1[0x13];
 527   b1[0x1B] += b1[0x1F];
 528   out1[0x10* 9] = b1[0x13] + b1[0x1B];
 529   out1[0x10*11] = b1[0x1B] + b1[0x17];
 530   out1[0x10*13] = b1[0x17] + b1[0x1F];
 531   out1[0x10*15] = b1[0x1F];
 532 }