annotate src/fftw-3.3.3/dft/scalar/codelets/t2_32.c @ 95:89f5e221ed7b

Add FFTW3
author Chris Cannam <cannam@all-day-breakfast.com>
date Wed, 20 Mar 2013 15:35:50 +0000
parents
children
rev   line source
cannam@95 1 /*
cannam@95 2 * Copyright (c) 2003, 2007-11 Matteo Frigo
cannam@95 3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
cannam@95 4 *
cannam@95 5 * This program is free software; you can redistribute it and/or modify
cannam@95 6 * it under the terms of the GNU General Public License as published by
cannam@95 7 * the Free Software Foundation; either version 2 of the License, or
cannam@95 8 * (at your option) any later version.
cannam@95 9 *
cannam@95 10 * This program is distributed in the hope that it will be useful,
cannam@95 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@95 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
cannam@95 13 * GNU General Public License for more details.
cannam@95 14 *
cannam@95 15 * You should have received a copy of the GNU General Public License
cannam@95 16 * along with this program; if not, write to the Free Software
cannam@95 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
cannam@95 18 *
cannam@95 19 */
cannam@95 20
cannam@95 21 /* This file was automatically generated --- DO NOT EDIT */
cannam@95 22 /* Generated on Sun Nov 25 07:36:00 EST 2012 */
cannam@95 23
cannam@95 24 #include "codelet-dft.h"
cannam@95 25
cannam@95 26 #ifdef HAVE_FMA
cannam@95 27
cannam@95 28 /* Generated by: ../../../genfft/gen_twiddle.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 32 -name t2_32 -include t.h */
cannam@95 29
cannam@95 30 /*
cannam@95 31 * This function contains 488 FP additions, 350 FP multiplications,
cannam@95 32 * (or, 236 additions, 98 multiplications, 252 fused multiply/add),
cannam@95 33 * 181 stack variables, 7 constants, and 128 memory accesses
cannam@95 34 */
cannam@95 35 #include "t.h"
cannam@95 36
cannam@95 37 static void t2_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
cannam@95 38 {
cannam@95 39 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
cannam@95 40 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
cannam@95 41 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
cannam@95 42 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
cannam@95 43 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
cannam@95 44 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
cannam@95 45 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
cannam@95 46 {
cannam@95 47 INT m;
cannam@95 48 for (m = mb, W = W + (mb * 8); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 8, MAKE_VOLATILE_STRIDE(64, rs)) {
cannam@95 49 E T9A, T9z;
cannam@95 50 {
cannam@95 51 E T2, T8, T3, T6, Te, Tr, T18, T4, Ta, Tz, T1n, T10, Ti, T5, Tc;
cannam@95 52 T2 = W[0];
cannam@95 53 T8 = W[4];
cannam@95 54 T3 = W[2];
cannam@95 55 T6 = W[3];
cannam@95 56 Te = W[6];
cannam@95 57 Tr = T2 * T8;
cannam@95 58 T18 = T3 * T8;
cannam@95 59 T4 = T2 * T3;
cannam@95 60 Ta = T2 * T6;
cannam@95 61 Tz = T3 * Te;
cannam@95 62 T1n = T8 * Te;
cannam@95 63 T10 = T2 * Te;
cannam@95 64 Ti = W[7];
cannam@95 65 T5 = W[1];
cannam@95 66 Tc = W[5];
cannam@95 67 {
cannam@95 68 E T34, T31, T2X, T2T, Tq, T46, T8H, T97, TH, T98, T4b, T8D, TZ, T7f, T4j;
cannam@95 69 E T6t, T1g, T7g, T4q, T6u, T4z, T6x, T1J, T7m, T7l, T8d, T6y, T4G, T2k, T7o;
cannam@95 70 E T7r, T8e, T6A, T4O, T6B, T4V, T6P, T5E, T7L, T3G, T6M, T61, T8n, T7I, T6I;
cannam@95 71 E T55, T7A, T2N, T6F, T5s, T8i, T7x, T5L, T62, T43, T7J, T5S, T63, T7O, T8o;
cannam@95 72 E T2U, T2R, T2V, T57, T3a, T5h, T2Y, T32, T35;
cannam@95 73 {
cannam@95 74 E T1K, T23, T1N, T26, T2b, T1U, T3C, T3j, T3z, T3f, T1R, T29, TR, Th, T2J;
cannam@95 75 E T2F, Td, TP, T3r, T3n, T2w, T2s, T3Q, T3M, T1Z, T1V, T2g, T2c;
cannam@95 76 {
cannam@95 77 E T11, T1C, TM, Tb, TJ, T7, T1o, T19, T1w, T1F, T15, T1s, T1d, T1z, TW;
cannam@95 78 E TS, Ty, T48, TG, T4a;
cannam@95 79 {
cannam@95 80 E T1, TA, Ts, TE, Tw, Tn, Tj, T8G, Tk, To, T14;
cannam@95 81 T1 = ri[0];
cannam@95 82 TA = FMA(T6, Ti, Tz);
cannam@95 83 T1K = FNMS(T6, Ti, Tz);
cannam@95 84 T14 = T2 * Ti;
cannam@95 85 {
cannam@95 86 E T1r, TD, T1c, Tv;
cannam@95 87 T1r = T8 * Ti;
cannam@95 88 TD = T3 * Ti;
cannam@95 89 T11 = FNMS(T5, Ti, T10);
cannam@95 90 T1C = FMA(T5, Ti, T10);
cannam@95 91 TM = FMA(T5, T3, Ta);
cannam@95 92 Tb = FNMS(T5, T3, Ta);
cannam@95 93 TJ = FNMS(T5, T6, T4);
cannam@95 94 T7 = FMA(T5, T6, T4);
cannam@95 95 T1o = FMA(Tc, Ti, T1n);
cannam@95 96 T23 = FMA(T6, Tc, T18);
cannam@95 97 T19 = FNMS(T6, Tc, T18);
cannam@95 98 T1w = FNMS(T5, Tc, Tr);
cannam@95 99 Ts = FMA(T5, Tc, Tr);
cannam@95 100 T1c = T3 * Tc;
cannam@95 101 Tv = T2 * Tc;
cannam@95 102 T1F = FNMS(T5, Te, T14);
cannam@95 103 T15 = FMA(T5, Te, T14);
cannam@95 104 T1s = FNMS(Tc, Te, T1r);
cannam@95 105 T1N = FMA(T6, Te, TD);
cannam@95 106 TE = FNMS(T6, Te, TD);
cannam@95 107 {
cannam@95 108 E T1T, T3i, T3e, T1Q;
cannam@95 109 T1T = TJ * Tc;
cannam@95 110 T3i = TJ * Ti;
cannam@95 111 T3e = TJ * Te;
cannam@95 112 T1Q = TJ * T8;
cannam@95 113 {
cannam@95 114 E Tg, T2I, T2E, T9;
cannam@95 115 Tg = T7 * Tc;
cannam@95 116 T2I = T7 * Ti;
cannam@95 117 T2E = T7 * Te;
cannam@95 118 T9 = T7 * T8;
cannam@95 119 {
cannam@95 120 E T3q, T3m, T2v, T2r;
cannam@95 121 T3q = T19 * Ti;
cannam@95 122 T3m = T19 * Te;
cannam@95 123 T2v = T1w * Ti;
cannam@95 124 T2r = T1w * Te;
cannam@95 125 {
cannam@95 126 E T2W, T2S, T3P, T3L;
cannam@95 127 T2W = T23 * Ti;
cannam@95 128 T2S = T23 * Te;
cannam@95 129 T3P = Ts * Ti;
cannam@95 130 T3L = Ts * Te;
cannam@95 131 T26 = FNMS(T6, T8, T1c);
cannam@95 132 T1d = FMA(T6, T8, T1c);
cannam@95 133 T1z = FMA(T5, T8, Tv);
cannam@95 134 Tw = FNMS(T5, T8, Tv);
cannam@95 135 T2b = FNMS(TM, T8, T1T);
cannam@95 136 T1U = FMA(TM, T8, T1T);
cannam@95 137 T3C = FNMS(TM, Te, T3i);
cannam@95 138 T3j = FMA(TM, Te, T3i);
cannam@95 139 T3z = FMA(TM, Ti, T3e);
cannam@95 140 T3f = FNMS(TM, Ti, T3e);
cannam@95 141 T1R = FNMS(TM, Tc, T1Q);
cannam@95 142 T29 = FMA(TM, Tc, T1Q);
cannam@95 143 TR = FNMS(Tb, T8, Tg);
cannam@95 144 Th = FMA(Tb, T8, Tg);
cannam@95 145 T34 = FMA(Tb, Te, T2I);
cannam@95 146 T2J = FNMS(Tb, Te, T2I);
cannam@95 147 T31 = FNMS(Tb, Ti, T2E);
cannam@95 148 T2F = FMA(Tb, Ti, T2E);
cannam@95 149 Td = FNMS(Tb, Tc, T9);
cannam@95 150 TP = FMA(Tb, Tc, T9);
cannam@95 151 T2X = FNMS(T26, Te, T2W);
cannam@95 152 T2T = FMA(T26, Ti, T2S);
cannam@95 153 T3r = FNMS(T1d, Te, T3q);
cannam@95 154 T3n = FMA(T1d, Ti, T3m);
cannam@95 155 T2w = FNMS(T1z, Te, T2v);
cannam@95 156 T2s = FMA(T1z, Ti, T2r);
cannam@95 157 T3Q = FNMS(Tw, Te, T3P);
cannam@95 158 T3M = FMA(Tw, Ti, T3L);
cannam@95 159 {
cannam@95 160 E T1Y, T1S, T2f, T2a;
cannam@95 161 T1Y = T1R * Ti;
cannam@95 162 T1S = T1R * Te;
cannam@95 163 T2f = T29 * Ti;
cannam@95 164 T2a = T29 * Te;
cannam@95 165 {
cannam@95 166 E Tm, Tf, TV, TQ;
cannam@95 167 Tm = Td * Ti;
cannam@95 168 Tf = Td * Te;
cannam@95 169 TV = TP * Ti;
cannam@95 170 TQ = TP * Te;
cannam@95 171 T1Z = FNMS(T1U, Te, T1Y);
cannam@95 172 T1V = FMA(T1U, Ti, T1S);
cannam@95 173 T2g = FNMS(T2b, Te, T2f);
cannam@95 174 T2c = FMA(T2b, Ti, T2a);
cannam@95 175 Tn = FNMS(Th, Te, Tm);
cannam@95 176 Tj = FMA(Th, Ti, Tf);
cannam@95 177 TW = FNMS(TR, Te, TV);
cannam@95 178 TS = FMA(TR, Ti, TQ);
cannam@95 179 T8G = ii[0];
cannam@95 180 }
cannam@95 181 }
cannam@95 182 }
cannam@95 183 }
cannam@95 184 }
cannam@95 185 }
cannam@95 186 }
cannam@95 187 Tk = ri[WS(rs, 16)];
cannam@95 188 To = ii[WS(rs, 16)];
cannam@95 189 {
cannam@95 190 E Tt, Tx, Tu, T47, TB, TF, TC, T49;
cannam@95 191 {
cannam@95 192 E Tl, T8E, Tp, T8F;
cannam@95 193 Tt = ri[WS(rs, 8)];
cannam@95 194 Tx = ii[WS(rs, 8)];
cannam@95 195 Tl = Tj * Tk;
cannam@95 196 T8E = Tj * To;
cannam@95 197 Tu = Ts * Tt;
cannam@95 198 T47 = Ts * Tx;
cannam@95 199 Tp = FMA(Tn, To, Tl);
cannam@95 200 T8F = FNMS(Tn, Tk, T8E);
cannam@95 201 TB = ri[WS(rs, 24)];
cannam@95 202 TF = ii[WS(rs, 24)];
cannam@95 203 Tq = T1 + Tp;
cannam@95 204 T46 = T1 - Tp;
cannam@95 205 T8H = T8F + T8G;
cannam@95 206 T97 = T8G - T8F;
cannam@95 207 TC = TA * TB;
cannam@95 208 T49 = TA * TF;
cannam@95 209 }
cannam@95 210 Ty = FMA(Tw, Tx, Tu);
cannam@95 211 T48 = FNMS(Tw, Tt, T47);
cannam@95 212 TG = FMA(TE, TF, TC);
cannam@95 213 T4a = FNMS(TE, TB, T49);
cannam@95 214 }
cannam@95 215 }
cannam@95 216 {
cannam@95 217 E TT, TX, TO, T4f, TU, T4g;
cannam@95 218 {
cannam@95 219 E TK, TN, TL, T4e;
cannam@95 220 TK = ri[WS(rs, 4)];
cannam@95 221 TN = ii[WS(rs, 4)];
cannam@95 222 TH = Ty + TG;
cannam@95 223 T98 = Ty - TG;
cannam@95 224 T4b = T48 - T4a;
cannam@95 225 T8D = T48 + T4a;
cannam@95 226 TL = TJ * TK;
cannam@95 227 T4e = TJ * TN;
cannam@95 228 TT = ri[WS(rs, 20)];
cannam@95 229 TX = ii[WS(rs, 20)];
cannam@95 230 TO = FMA(TM, TN, TL);
cannam@95 231 T4f = FNMS(TM, TK, T4e);
cannam@95 232 TU = TS * TT;
cannam@95 233 T4g = TS * TX;
cannam@95 234 }
cannam@95 235 {
cannam@95 236 E T17, T4m, T1a, T1e, T4d, T4i;
cannam@95 237 {
cannam@95 238 E T12, T16, TY, T4h, T13, T4l;
cannam@95 239 T12 = ri[WS(rs, 28)];
cannam@95 240 T16 = ii[WS(rs, 28)];
cannam@95 241 TY = FMA(TW, TX, TU);
cannam@95 242 T4h = FNMS(TW, TT, T4g);
cannam@95 243 T13 = T11 * T12;
cannam@95 244 T4l = T11 * T16;
cannam@95 245 TZ = TO + TY;
cannam@95 246 T4d = TO - TY;
cannam@95 247 T7f = T4f + T4h;
cannam@95 248 T4i = T4f - T4h;
cannam@95 249 T17 = FMA(T15, T16, T13);
cannam@95 250 T4m = FNMS(T15, T12, T4l);
cannam@95 251 }
cannam@95 252 T4j = T4d + T4i;
cannam@95 253 T6t = T4i - T4d;
cannam@95 254 T1a = ri[WS(rs, 12)];
cannam@95 255 T1e = ii[WS(rs, 12)];
cannam@95 256 {
cannam@95 257 E T1m, T4u, T1H, T4E, T1x, T1A, T1u, T4w, T1y, T4B;
cannam@95 258 {
cannam@95 259 E T1D, T1G, T1E, T4D;
cannam@95 260 {
cannam@95 261 E T1f, T4o, T4k, T4p;
cannam@95 262 {
cannam@95 263 E T1j, T1l, T1b, T4n, T1k, T4t;
cannam@95 264 T1j = ri[WS(rs, 2)];
cannam@95 265 T1l = ii[WS(rs, 2)];
cannam@95 266 T1b = T19 * T1a;
cannam@95 267 T4n = T19 * T1e;
cannam@95 268 T1k = T7 * T1j;
cannam@95 269 T4t = T7 * T1l;
cannam@95 270 T1f = FMA(T1d, T1e, T1b);
cannam@95 271 T4o = FNMS(T1d, T1a, T4n);
cannam@95 272 T1m = FMA(Tb, T1l, T1k);
cannam@95 273 T4u = FNMS(Tb, T1j, T4t);
cannam@95 274 }
cannam@95 275 T1g = T17 + T1f;
cannam@95 276 T4k = T17 - T1f;
cannam@95 277 T7g = T4m + T4o;
cannam@95 278 T4p = T4m - T4o;
cannam@95 279 T1D = ri[WS(rs, 26)];
cannam@95 280 T1G = ii[WS(rs, 26)];
cannam@95 281 T4q = T4k - T4p;
cannam@95 282 T6u = T4k + T4p;
cannam@95 283 T1E = T1C * T1D;
cannam@95 284 T4D = T1C * T1G;
cannam@95 285 }
cannam@95 286 {
cannam@95 287 E T1p, T1t, T1q, T4v;
cannam@95 288 T1p = ri[WS(rs, 18)];
cannam@95 289 T1t = ii[WS(rs, 18)];
cannam@95 290 T1H = FMA(T1F, T1G, T1E);
cannam@95 291 T4E = FNMS(T1F, T1D, T4D);
cannam@95 292 T1q = T1o * T1p;
cannam@95 293 T4v = T1o * T1t;
cannam@95 294 T1x = ri[WS(rs, 10)];
cannam@95 295 T1A = ii[WS(rs, 10)];
cannam@95 296 T1u = FMA(T1s, T1t, T1q);
cannam@95 297 T4w = FNMS(T1s, T1p, T4v);
cannam@95 298 T1y = T1w * T1x;
cannam@95 299 T4B = T1w * T1A;
cannam@95 300 }
cannam@95 301 }
cannam@95 302 {
cannam@95 303 E T4A, T1v, T7j, T4x, T1B, T4C;
cannam@95 304 T4A = T1m - T1u;
cannam@95 305 T1v = T1m + T1u;
cannam@95 306 T7j = T4u + T4w;
cannam@95 307 T4x = T4u - T4w;
cannam@95 308 T1B = FMA(T1z, T1A, T1y);
cannam@95 309 T4C = FNMS(T1z, T1x, T4B);
cannam@95 310 {
cannam@95 311 E T1I, T4y, T4F, T7k;
cannam@95 312 T1I = T1B + T1H;
cannam@95 313 T4y = T1B - T1H;
cannam@95 314 T4F = T4C - T4E;
cannam@95 315 T7k = T4C + T4E;
cannam@95 316 T4z = T4x - T4y;
cannam@95 317 T6x = T4x + T4y;
cannam@95 318 T1J = T1v + T1I;
cannam@95 319 T7m = T1v - T1I;
cannam@95 320 T7l = T7j - T7k;
cannam@95 321 T8d = T7j + T7k;
cannam@95 322 T6y = T4A - T4F;
cannam@95 323 T4G = T4A + T4F;
cannam@95 324 }
cannam@95 325 }
cannam@95 326 }
cannam@95 327 }
cannam@95 328 }
cannam@95 329 }
cannam@95 330 {
cannam@95 331 E T5Z, T3u, T5V, T5C, T7G, T5D, T3F, T5X, T4P, T4U;
cannam@95 332 {
cannam@95 333 E T1P, T4J, T2i, T4T, T21, T4L, T28, T4R;
cannam@95 334 {
cannam@95 335 E T1L, T1O, T1W, T20;
cannam@95 336 T1L = ri[WS(rs, 30)];
cannam@95 337 T1O = ii[WS(rs, 30)];
cannam@95 338 {
cannam@95 339 E T2d, T2h, T1M, T4I, T2e, T4S;
cannam@95 340 T2d = ri[WS(rs, 22)];
cannam@95 341 T2h = ii[WS(rs, 22)];
cannam@95 342 T1M = T1K * T1L;
cannam@95 343 T4I = T1K * T1O;
cannam@95 344 T2e = T2c * T2d;
cannam@95 345 T4S = T2c * T2h;
cannam@95 346 T1P = FMA(T1N, T1O, T1M);
cannam@95 347 T4J = FNMS(T1N, T1L, T4I);
cannam@95 348 T2i = FMA(T2g, T2h, T2e);
cannam@95 349 T4T = FNMS(T2g, T2d, T4S);
cannam@95 350 }
cannam@95 351 T1W = ri[WS(rs, 14)];
cannam@95 352 T20 = ii[WS(rs, 14)];
cannam@95 353 {
cannam@95 354 E T24, T27, T1X, T4K, T25, T4Q;
cannam@95 355 T24 = ri[WS(rs, 6)];
cannam@95 356 T27 = ii[WS(rs, 6)];
cannam@95 357 T1X = T1V * T1W;
cannam@95 358 T4K = T1V * T20;
cannam@95 359 T25 = T23 * T24;
cannam@95 360 T4Q = T23 * T27;
cannam@95 361 T21 = FMA(T1Z, T20, T1X);
cannam@95 362 T4L = FNMS(T1Z, T1W, T4K);
cannam@95 363 T28 = FMA(T26, T27, T25);
cannam@95 364 T4R = FNMS(T26, T24, T4Q);
cannam@95 365 }
cannam@95 366 }
cannam@95 367 {
cannam@95 368 E T22, T7p, T4M, T4N, T2j, T7q;
cannam@95 369 T4P = T1P - T21;
cannam@95 370 T22 = T1P + T21;
cannam@95 371 T7p = T4J + T4L;
cannam@95 372 T4M = T4J - T4L;
cannam@95 373 T4N = T28 - T2i;
cannam@95 374 T2j = T28 + T2i;
cannam@95 375 T7q = T4R + T4T;
cannam@95 376 T4U = T4R - T4T;
cannam@95 377 T2k = T22 + T2j;
cannam@95 378 T7o = T22 - T2j;
cannam@95 379 T7r = T7p - T7q;
cannam@95 380 T8e = T7p + T7q;
cannam@95 381 T6A = T4M + T4N;
cannam@95 382 T4O = T4M - T4N;
cannam@95 383 }
cannam@95 384 }
cannam@95 385 {
cannam@95 386 E T3l, T5z, T3E, T3v, T3t, T3w, T3x, T5B, T3A, T3B, T3D, T3y, T5W;
cannam@95 387 {
cannam@95 388 E T3g, T3k, T3h, T5y;
cannam@95 389 T3g = ri[WS(rs, 31)];
cannam@95 390 T3k = ii[WS(rs, 31)];
cannam@95 391 T3A = ri[WS(rs, 23)];
cannam@95 392 T6B = T4P - T4U;
cannam@95 393 T4V = T4P + T4U;
cannam@95 394 T3h = T3f * T3g;
cannam@95 395 T5y = T3f * T3k;
cannam@95 396 T3B = T3z * T3A;
cannam@95 397 T3D = ii[WS(rs, 23)];
cannam@95 398 T3l = FMA(T3j, T3k, T3h);
cannam@95 399 T5z = FNMS(T3j, T3g, T5y);
cannam@95 400 }
cannam@95 401 {
cannam@95 402 E T3o, T5Y, T3s, T3p, T5A;
cannam@95 403 T3o = ri[WS(rs, 15)];
cannam@95 404 T3E = FMA(T3C, T3D, T3B);
cannam@95 405 T5Y = T3z * T3D;
cannam@95 406 T3s = ii[WS(rs, 15)];
cannam@95 407 T3p = T3n * T3o;
cannam@95 408 T3v = ri[WS(rs, 7)];
cannam@95 409 T5Z = FNMS(T3C, T3A, T5Y);
cannam@95 410 T5A = T3n * T3s;
cannam@95 411 T3t = FMA(T3r, T3s, T3p);
cannam@95 412 T3w = TP * T3v;
cannam@95 413 T3x = ii[WS(rs, 7)];
cannam@95 414 T5B = FNMS(T3r, T3o, T5A);
cannam@95 415 }
cannam@95 416 T3u = T3l + T3t;
cannam@95 417 T5V = T3l - T3t;
cannam@95 418 T3y = FMA(TR, T3x, T3w);
cannam@95 419 T5W = TP * T3x;
cannam@95 420 T5C = T5z - T5B;
cannam@95 421 T7G = T5z + T5B;
cannam@95 422 T5D = T3y - T3E;
cannam@95 423 T3F = T3y + T3E;
cannam@95 424 T5X = FNMS(TR, T3v, T5W);
cannam@95 425 }
cannam@95 426 {
cannam@95 427 E T2L, T5q, T5m, T2z, T7v, T53, T2D, T5o;
cannam@95 428 {
cannam@95 429 E T2q, T50, T2y, T2A, T2C, T52, T2B, T5n;
cannam@95 430 {
cannam@95 431 E T2G, T2K, T2n, T4Z, T2t, T51;
cannam@95 432 {
cannam@95 433 E T2o, T2p, T60, T7H;
cannam@95 434 T2n = ri[WS(rs, 1)];
cannam@95 435 T6P = T5C + T5D;
cannam@95 436 T5E = T5C - T5D;
cannam@95 437 T7L = T3u - T3F;
cannam@95 438 T3G = T3u + T3F;
cannam@95 439 T60 = T5X - T5Z;
cannam@95 440 T7H = T5X + T5Z;
cannam@95 441 T2o = T2 * T2n;
cannam@95 442 T2p = ii[WS(rs, 1)];
cannam@95 443 T6M = T5V - T60;
cannam@95 444 T61 = T5V + T60;
cannam@95 445 T8n = T7G + T7H;
cannam@95 446 T7I = T7G - T7H;
cannam@95 447 T4Z = T2 * T2p;
cannam@95 448 T2q = FMA(T5, T2p, T2o);
cannam@95 449 }
cannam@95 450 T2G = ri[WS(rs, 25)];
cannam@95 451 T2K = ii[WS(rs, 25)];
cannam@95 452 T50 = FNMS(T5, T2n, T4Z);
cannam@95 453 {
cannam@95 454 E T2x, T2u, T2H, T5p;
cannam@95 455 T2t = ri[WS(rs, 17)];
cannam@95 456 T2H = T2F * T2G;
cannam@95 457 T5p = T2F * T2K;
cannam@95 458 T2x = ii[WS(rs, 17)];
cannam@95 459 T2u = T2s * T2t;
cannam@95 460 T2L = FMA(T2J, T2K, T2H);
cannam@95 461 T5q = FNMS(T2J, T2G, T5p);
cannam@95 462 T51 = T2s * T2x;
cannam@95 463 T2y = FMA(T2w, T2x, T2u);
cannam@95 464 }
cannam@95 465 T2A = ri[WS(rs, 9)];
cannam@95 466 T2C = ii[WS(rs, 9)];
cannam@95 467 T52 = FNMS(T2w, T2t, T51);
cannam@95 468 }
cannam@95 469 T5m = T2q - T2y;
cannam@95 470 T2z = T2q + T2y;
cannam@95 471 T2B = T8 * T2A;
cannam@95 472 T5n = T8 * T2C;
cannam@95 473 T7v = T50 + T52;
cannam@95 474 T53 = T50 - T52;
cannam@95 475 T2D = FMA(Tc, T2C, T2B);
cannam@95 476 T5o = FNMS(Tc, T2A, T5n);
cannam@95 477 }
cannam@95 478 {
cannam@95 479 E T3N, T3K, T3O, T5G, T41, T5Q, T3R, T3U, T3W;
cannam@95 480 {
cannam@95 481 E T3H, T3I, T3J, T3Y, T40, T5F, T3Z, T5P;
cannam@95 482 T3H = ri[WS(rs, 3)];
cannam@95 483 {
cannam@95 484 E T54, T2M, T5r, T7w;
cannam@95 485 T54 = T2D - T2L;
cannam@95 486 T2M = T2D + T2L;
cannam@95 487 T5r = T5o - T5q;
cannam@95 488 T7w = T5o + T5q;
cannam@95 489 T6I = T53 + T54;
cannam@95 490 T55 = T53 - T54;
cannam@95 491 T7A = T2z - T2M;
cannam@95 492 T2N = T2z + T2M;
cannam@95 493 T6F = T5m - T5r;
cannam@95 494 T5s = T5m + T5r;
cannam@95 495 T8i = T7v + T7w;
cannam@95 496 T7x = T7v - T7w;
cannam@95 497 T3I = T3 * T3H;
cannam@95 498 }
cannam@95 499 T3J = ii[WS(rs, 3)];
cannam@95 500 T3Y = ri[WS(rs, 11)];
cannam@95 501 T40 = ii[WS(rs, 11)];
cannam@95 502 T3N = ri[WS(rs, 19)];
cannam@95 503 T3K = FMA(T6, T3J, T3I);
cannam@95 504 T5F = T3 * T3J;
cannam@95 505 T3Z = Td * T3Y;
cannam@95 506 T5P = Td * T40;
cannam@95 507 T3O = T3M * T3N;
cannam@95 508 T5G = FNMS(T6, T3H, T5F);
cannam@95 509 T41 = FMA(Th, T40, T3Z);
cannam@95 510 T5Q = FNMS(Th, T3Y, T5P);
cannam@95 511 T3R = ii[WS(rs, 19)];
cannam@95 512 T3U = ri[WS(rs, 27)];
cannam@95 513 T3W = ii[WS(rs, 27)];
cannam@95 514 }
cannam@95 515 {
cannam@95 516 E T2O, T2P, T2Q, T37, T39, T56, T38, T5g;
cannam@95 517 {
cannam@95 518 E T3T, T5K, T5I, T3X, T5O, T7M, T5J;
cannam@95 519 T2O = ri[WS(rs, 5)];
cannam@95 520 {
cannam@95 521 E T3S, T5H, T3V, T5N;
cannam@95 522 T3S = FMA(T3Q, T3R, T3O);
cannam@95 523 T5H = T3M * T3R;
cannam@95 524 T3V = Te * T3U;
cannam@95 525 T5N = Te * T3W;
cannam@95 526 T3T = T3K + T3S;
cannam@95 527 T5K = T3K - T3S;
cannam@95 528 T5I = FNMS(T3Q, T3N, T5H);
cannam@95 529 T3X = FMA(Ti, T3W, T3V);
cannam@95 530 T5O = FNMS(Ti, T3U, T5N);
cannam@95 531 T2P = T29 * T2O;
cannam@95 532 }
cannam@95 533 T7M = T5G + T5I;
cannam@95 534 T5J = T5G - T5I;
cannam@95 535 {
cannam@95 536 E T42, T5M, T7N, T5R;
cannam@95 537 T42 = T3X + T41;
cannam@95 538 T5M = T3X - T41;
cannam@95 539 T7N = T5O + T5Q;
cannam@95 540 T5R = T5O - T5Q;
cannam@95 541 T5L = T5J - T5K;
cannam@95 542 T62 = T5K + T5J;
cannam@95 543 T43 = T3T + T42;
cannam@95 544 T7J = T42 - T3T;
cannam@95 545 T5S = T5M + T5R;
cannam@95 546 T63 = T5M - T5R;
cannam@95 547 T7O = T7M - T7N;
cannam@95 548 T8o = T7M + T7N;
cannam@95 549 T2Q = ii[WS(rs, 5)];
cannam@95 550 }
cannam@95 551 }
cannam@95 552 T37 = ri[WS(rs, 13)];
cannam@95 553 T39 = ii[WS(rs, 13)];
cannam@95 554 T2U = ri[WS(rs, 21)];
cannam@95 555 T2R = FMA(T2b, T2Q, T2P);
cannam@95 556 T56 = T29 * T2Q;
cannam@95 557 T38 = T1R * T37;
cannam@95 558 T5g = T1R * T39;
cannam@95 559 T2V = T2T * T2U;
cannam@95 560 T57 = FNMS(T2b, T2O, T56);
cannam@95 561 T3a = FMA(T1U, T39, T38);
cannam@95 562 T5h = FNMS(T1U, T37, T5g);
cannam@95 563 T2Y = ii[WS(rs, 21)];
cannam@95 564 T32 = ri[WS(rs, 29)];
cannam@95 565 T35 = ii[WS(rs, 29)];
cannam@95 566 }
cannam@95 567 }
cannam@95 568 }
cannam@95 569 }
cannam@95 570 }
cannam@95 571 {
cannam@95 572 E T5c, T5t, T5j, T5u, T88, T90, T8Z, T8b;
cannam@95 573 {
cannam@95 574 E T7e, T8T, T7y, T7D, T7h, T8U, T8S, T8R;
cannam@95 575 {
cannam@95 576 E T8c, T1i, T8A, T8z, T8O, T8J, T8N, T2l, T8L, T45, T8t, T8l, T8u, T8q, T3c;
cannam@95 577 E T8k, T8p, T8w, T2m;
cannam@95 578 {
cannam@95 579 E T8x, T8y, T8j, T8C, T8I;
cannam@95 580 {
cannam@95 581 E TI, T30, T5b, T59, T36, T5f, T1h, T7B, T5a;
cannam@95 582 TI = Tq + TH;
cannam@95 583 T7e = Tq - TH;
cannam@95 584 {
cannam@95 585 E T2Z, T58, T33, T5e;
cannam@95 586 T2Z = FMA(T2X, T2Y, T2V);
cannam@95 587 T58 = T2T * T2Y;
cannam@95 588 T33 = T31 * T32;
cannam@95 589 T5e = T31 * T35;
cannam@95 590 T30 = T2R + T2Z;
cannam@95 591 T5b = T2R - T2Z;
cannam@95 592 T59 = FNMS(T2X, T2U, T58);
cannam@95 593 T36 = FMA(T34, T35, T33);
cannam@95 594 T5f = FNMS(T34, T32, T5e);
cannam@95 595 T1h = TZ + T1g;
cannam@95 596 T8T = T1g - TZ;
cannam@95 597 }
cannam@95 598 T7B = T57 + T59;
cannam@95 599 T5a = T57 - T59;
cannam@95 600 {
cannam@95 601 E T3b, T5d, T7C, T5i;
cannam@95 602 T3b = T36 + T3a;
cannam@95 603 T5d = T36 - T3a;
cannam@95 604 T7C = T5f + T5h;
cannam@95 605 T5i = T5f - T5h;
cannam@95 606 T5c = T5a - T5b;
cannam@95 607 T5t = T5b + T5a;
cannam@95 608 T3c = T30 + T3b;
cannam@95 609 T7y = T3b - T30;
cannam@95 610 T5j = T5d + T5i;
cannam@95 611 T5u = T5d - T5i;
cannam@95 612 T7D = T7B - T7C;
cannam@95 613 T8j = T7B + T7C;
cannam@95 614 T8c = TI - T1h;
cannam@95 615 T1i = TI + T1h;
cannam@95 616 }
cannam@95 617 }
cannam@95 618 T8k = T8i - T8j;
cannam@95 619 T8x = T8i + T8j;
cannam@95 620 T8y = T8n + T8o;
cannam@95 621 T8p = T8n - T8o;
cannam@95 622 T7h = T7f - T7g;
cannam@95 623 T8C = T7f + T7g;
cannam@95 624 T8I = T8D + T8H;
cannam@95 625 T8U = T8H - T8D;
cannam@95 626 T8A = T8x + T8y;
cannam@95 627 T8z = T8x - T8y;
cannam@95 628 T8O = T8I - T8C;
cannam@95 629 T8J = T8C + T8I;
cannam@95 630 }
cannam@95 631 {
cannam@95 632 E T8h, T8m, T3d, T44;
cannam@95 633 T8h = T2N - T3c;
cannam@95 634 T3d = T2N + T3c;
cannam@95 635 T44 = T3G + T43;
cannam@95 636 T8m = T3G - T43;
cannam@95 637 T8N = T2k - T1J;
cannam@95 638 T2l = T1J + T2k;
cannam@95 639 T8L = T44 - T3d;
cannam@95 640 T45 = T3d + T44;
cannam@95 641 T8t = T8k - T8h;
cannam@95 642 T8l = T8h + T8k;
cannam@95 643 T8u = T8m + T8p;
cannam@95 644 T8q = T8m - T8p;
cannam@95 645 }
cannam@95 646 T8w = T1i - T2l;
cannam@95 647 T2m = T1i + T2l;
cannam@95 648 {
cannam@95 649 E T8s, T8P, T8Q, T8v;
cannam@95 650 {
cannam@95 651 E T8r, T8M, T8K, T8g, T8B, T8f;
cannam@95 652 T8S = T8q - T8l;
cannam@95 653 T8r = T8l + T8q;
cannam@95 654 T8B = T8d + T8e;
cannam@95 655 T8f = T8d - T8e;
cannam@95 656 ri[0] = T2m + T45;
cannam@95 657 ri[WS(rs, 16)] = T2m - T45;
cannam@95 658 ri[WS(rs, 8)] = T8w + T8z;
cannam@95 659 ri[WS(rs, 24)] = T8w - T8z;
cannam@95 660 T8M = T8J - T8B;
cannam@95 661 T8K = T8B + T8J;
cannam@95 662 T8g = T8c + T8f;
cannam@95 663 T8s = T8c - T8f;
cannam@95 664 T8R = T8O - T8N;
cannam@95 665 T8P = T8N + T8O;
cannam@95 666 ii[WS(rs, 24)] = T8M - T8L;
cannam@95 667 ii[WS(rs, 8)] = T8L + T8M;
cannam@95 668 ii[WS(rs, 16)] = T8K - T8A;
cannam@95 669 ii[0] = T8A + T8K;
cannam@95 670 ri[WS(rs, 4)] = FMA(KP707106781, T8r, T8g);
cannam@95 671 ri[WS(rs, 20)] = FNMS(KP707106781, T8r, T8g);
cannam@95 672 T8Q = T8t + T8u;
cannam@95 673 T8v = T8t - T8u;
cannam@95 674 }
cannam@95 675 ii[WS(rs, 20)] = FNMS(KP707106781, T8Q, T8P);
cannam@95 676 ii[WS(rs, 4)] = FMA(KP707106781, T8Q, T8P);
cannam@95 677 ri[WS(rs, 12)] = FMA(KP707106781, T8v, T8s);
cannam@95 678 ri[WS(rs, 28)] = FNMS(KP707106781, T8v, T8s);
cannam@95 679 }
cannam@95 680 }
cannam@95 681 {
cannam@95 682 E T7P, T7W, T7i, T7K, T8a, T86, T91, T8V, T8W, T7t, T7T, T7F, T92, T7Z, T89;
cannam@95 683 E T83;
cannam@95 684 {
cannam@95 685 E T7X, T7n, T7s, T7Y, T84, T85;
cannam@95 686 T7P = T7L - T7O;
cannam@95 687 T84 = T7L + T7O;
cannam@95 688 ii[WS(rs, 28)] = FNMS(KP707106781, T8S, T8R);
cannam@95 689 ii[WS(rs, 12)] = FMA(KP707106781, T8S, T8R);
cannam@95 690 T7W = T7e + T7h;
cannam@95 691 T7i = T7e - T7h;
cannam@95 692 T85 = T7I + T7J;
cannam@95 693 T7K = T7I - T7J;
cannam@95 694 T7X = T7m + T7l;
cannam@95 695 T7n = T7l - T7m;
cannam@95 696 T8a = FMA(KP414213562, T84, T85);
cannam@95 697 T86 = FNMS(KP414213562, T85, T84);
cannam@95 698 T91 = T8U - T8T;
cannam@95 699 T8V = T8T + T8U;
cannam@95 700 T7s = T7o + T7r;
cannam@95 701 T7Y = T7o - T7r;
cannam@95 702 {
cannam@95 703 E T82, T81, T7z, T7E;
cannam@95 704 T82 = T7x + T7y;
cannam@95 705 T7z = T7x - T7y;
cannam@95 706 T7E = T7A - T7D;
cannam@95 707 T81 = T7A + T7D;
cannam@95 708 T8W = T7n + T7s;
cannam@95 709 T7t = T7n - T7s;
cannam@95 710 T7T = FNMS(KP414213562, T7z, T7E);
cannam@95 711 T7F = FMA(KP414213562, T7E, T7z);
cannam@95 712 T92 = T7Y - T7X;
cannam@95 713 T7Z = T7X + T7Y;
cannam@95 714 T89 = FNMS(KP414213562, T81, T82);
cannam@95 715 T83 = FMA(KP414213562, T82, T81);
cannam@95 716 }
cannam@95 717 }
cannam@95 718 {
cannam@95 719 E T7S, T7u, T93, T95, T7U, T7Q;
cannam@95 720 T7S = FNMS(KP707106781, T7t, T7i);
cannam@95 721 T7u = FMA(KP707106781, T7t, T7i);
cannam@95 722 T93 = FMA(KP707106781, T92, T91);
cannam@95 723 T95 = FNMS(KP707106781, T92, T91);
cannam@95 724 T7U = FMA(KP414213562, T7K, T7P);
cannam@95 725 T7Q = FNMS(KP414213562, T7P, T7K);
cannam@95 726 {
cannam@95 727 E T80, T87, T8X, T8Y;
cannam@95 728 T88 = FNMS(KP707106781, T7Z, T7W);
cannam@95 729 T80 = FMA(KP707106781, T7Z, T7W);
cannam@95 730 {
cannam@95 731 E T7V, T94, T96, T7R;
cannam@95 732 T7V = T7T + T7U;
cannam@95 733 T94 = T7U - T7T;
cannam@95 734 T96 = T7F + T7Q;
cannam@95 735 T7R = T7F - T7Q;
cannam@95 736 ri[WS(rs, 30)] = FMA(KP923879532, T7V, T7S);
cannam@95 737 ri[WS(rs, 14)] = FNMS(KP923879532, T7V, T7S);
cannam@95 738 ii[WS(rs, 22)] = FNMS(KP923879532, T94, T93);
cannam@95 739 ii[WS(rs, 6)] = FMA(KP923879532, T94, T93);
cannam@95 740 ii[WS(rs, 30)] = FMA(KP923879532, T96, T95);
cannam@95 741 ii[WS(rs, 14)] = FNMS(KP923879532, T96, T95);
cannam@95 742 ri[WS(rs, 6)] = FMA(KP923879532, T7R, T7u);
cannam@95 743 ri[WS(rs, 22)] = FNMS(KP923879532, T7R, T7u);
cannam@95 744 T87 = T83 + T86;
cannam@95 745 T90 = T86 - T83;
cannam@95 746 }
cannam@95 747 T8Z = FNMS(KP707106781, T8W, T8V);
cannam@95 748 T8X = FMA(KP707106781, T8W, T8V);
cannam@95 749 T8Y = T89 + T8a;
cannam@95 750 T8b = T89 - T8a;
cannam@95 751 ri[WS(rs, 2)] = FMA(KP923879532, T87, T80);
cannam@95 752 ri[WS(rs, 18)] = FNMS(KP923879532, T87, T80);
cannam@95 753 ii[WS(rs, 18)] = FNMS(KP923879532, T8Y, T8X);
cannam@95 754 ii[WS(rs, 2)] = FMA(KP923879532, T8Y, T8X);
cannam@95 755 }
cannam@95 756 }
cannam@95 757 }
cannam@95 758 }
cannam@95 759 {
cannam@95 760 E T6s, T9o, T9n, T6v, T6N, T6Q, T6G, T6J, T9g, T9f;
cannam@95 761 {
cannam@95 762 E T6c, T4s, T9c, T4X, T9h, T9b, T9i, T6f, T5U, T6l, T64, T5k, T5v;
cannam@95 763 {
cannam@95 764 E T6d, T6e, T99, T9a, T5T;
cannam@95 765 {
cannam@95 766 E T4c, T4r, T4H, T4W;
cannam@95 767 T6s = T46 - T4b;
cannam@95 768 T4c = T46 + T4b;
cannam@95 769 ri[WS(rs, 10)] = FMA(KP923879532, T8b, T88);
cannam@95 770 ri[WS(rs, 26)] = FNMS(KP923879532, T8b, T88);
cannam@95 771 ii[WS(rs, 26)] = FNMS(KP923879532, T90, T8Z);
cannam@95 772 ii[WS(rs, 10)] = FMA(KP923879532, T90, T8Z);
cannam@95 773 T4r = T4j + T4q;
cannam@95 774 T9o = T4q - T4j;
cannam@95 775 T6d = FMA(KP414213562, T4z, T4G);
cannam@95 776 T4H = FNMS(KP414213562, T4G, T4z);
cannam@95 777 T4W = FMA(KP414213562, T4V, T4O);
cannam@95 778 T6e = FNMS(KP414213562, T4O, T4V);
cannam@95 779 T9n = T98 + T97;
cannam@95 780 T99 = T97 - T98;
cannam@95 781 T6c = FMA(KP707106781, T4r, T4c);
cannam@95 782 T4s = FNMS(KP707106781, T4r, T4c);
cannam@95 783 T9c = T4H + T4W;
cannam@95 784 T4X = T4H - T4W;
cannam@95 785 T9a = T6t + T6u;
cannam@95 786 T6v = T6t - T6u;
cannam@95 787 }
cannam@95 788 T6N = T5S - T5L;
cannam@95 789 T5T = T5L + T5S;
cannam@95 790 T9h = FNMS(KP707106781, T9a, T99);
cannam@95 791 T9b = FMA(KP707106781, T9a, T99);
cannam@95 792 T9i = T6e - T6d;
cannam@95 793 T6f = T6d + T6e;
cannam@95 794 T5U = FNMS(KP707106781, T5T, T5E);
cannam@95 795 T6l = FMA(KP707106781, T5T, T5E);
cannam@95 796 T64 = T62 + T63;
cannam@95 797 T6Q = T62 - T63;
cannam@95 798 T6G = T5j - T5c;
cannam@95 799 T5k = T5c + T5j;
cannam@95 800 T5v = T5t + T5u;
cannam@95 801 T6J = T5t - T5u;
cannam@95 802 }
cannam@95 803 {
cannam@95 804 E T6m, T6q, T6j, T6p, T9l, T9m;
cannam@95 805 {
cannam@95 806 E T68, T4Y, T6a, T66, T69, T5x, T9j, T6k, T65, T9k, T6b, T67;
cannam@95 807 T68 = FNMS(KP923879532, T4X, T4s);
cannam@95 808 T4Y = FMA(KP923879532, T4X, T4s);
cannam@95 809 T6k = FMA(KP707106781, T64, T61);
cannam@95 810 T65 = FNMS(KP707106781, T64, T61);
cannam@95 811 {
cannam@95 812 E T6i, T5l, T6h, T5w;
cannam@95 813 T6i = FMA(KP707106781, T5k, T55);
cannam@95 814 T5l = FNMS(KP707106781, T5k, T55);
cannam@95 815 T6h = FMA(KP707106781, T5v, T5s);
cannam@95 816 T5w = FNMS(KP707106781, T5v, T5s);
cannam@95 817 T6m = FNMS(KP198912367, T6l, T6k);
cannam@95 818 T6q = FMA(KP198912367, T6k, T6l);
cannam@95 819 T6a = FMA(KP668178637, T5U, T65);
cannam@95 820 T66 = FNMS(KP668178637, T65, T5U);
cannam@95 821 T6j = FMA(KP198912367, T6i, T6h);
cannam@95 822 T6p = FNMS(KP198912367, T6h, T6i);
cannam@95 823 T69 = FNMS(KP668178637, T5l, T5w);
cannam@95 824 T5x = FMA(KP668178637, T5w, T5l);
cannam@95 825 }
cannam@95 826 T9j = FMA(KP923879532, T9i, T9h);
cannam@95 827 T9l = FNMS(KP923879532, T9i, T9h);
cannam@95 828 T9k = T6a - T69;
cannam@95 829 T6b = T69 + T6a;
cannam@95 830 T9m = T5x + T66;
cannam@95 831 T67 = T5x - T66;
cannam@95 832 ii[WS(rs, 21)] = FNMS(KP831469612, T9k, T9j);
cannam@95 833 ii[WS(rs, 5)] = FMA(KP831469612, T9k, T9j);
cannam@95 834 ri[WS(rs, 5)] = FMA(KP831469612, T67, T4Y);
cannam@95 835 ri[WS(rs, 21)] = FNMS(KP831469612, T67, T4Y);
cannam@95 836 ri[WS(rs, 29)] = FMA(KP831469612, T6b, T68);
cannam@95 837 ri[WS(rs, 13)] = FNMS(KP831469612, T6b, T68);
cannam@95 838 }
cannam@95 839 {
cannam@95 840 E T6o, T9d, T9e, T6r, T6g, T6n;
cannam@95 841 T6o = FNMS(KP923879532, T6f, T6c);
cannam@95 842 T6g = FMA(KP923879532, T6f, T6c);
cannam@95 843 T6n = T6j + T6m;
cannam@95 844 T9g = T6m - T6j;
cannam@95 845 T9f = FNMS(KP923879532, T9c, T9b);
cannam@95 846 T9d = FMA(KP923879532, T9c, T9b);
cannam@95 847 ii[WS(rs, 29)] = FMA(KP831469612, T9m, T9l);
cannam@95 848 ii[WS(rs, 13)] = FNMS(KP831469612, T9m, T9l);
cannam@95 849 ri[WS(rs, 1)] = FMA(KP980785280, T6n, T6g);
cannam@95 850 ri[WS(rs, 17)] = FNMS(KP980785280, T6n, T6g);
cannam@95 851 T9e = T6p + T6q;
cannam@95 852 T6r = T6p - T6q;
cannam@95 853 ii[WS(rs, 17)] = FNMS(KP980785280, T9e, T9d);
cannam@95 854 ii[WS(rs, 1)] = FMA(KP980785280, T9e, T9d);
cannam@95 855 ri[WS(rs, 9)] = FMA(KP980785280, T6r, T6o);
cannam@95 856 ri[WS(rs, 25)] = FNMS(KP980785280, T6r, T6o);
cannam@95 857 }
cannam@95 858 }
cannam@95 859 }
cannam@95 860 {
cannam@95 861 E T6Y, T6w, T9w, T6D, T9v, T9p, T9q, T71, T6H, T74, T78, T7c, T6W, T6S;
cannam@95 862 {
cannam@95 863 E T6Z, T6z, T6C, T70;
cannam@95 864 T6Z = FNMS(KP414213562, T6x, T6y);
cannam@95 865 T6z = FMA(KP414213562, T6y, T6x);
cannam@95 866 ii[WS(rs, 25)] = FNMS(KP980785280, T9g, T9f);
cannam@95 867 ii[WS(rs, 9)] = FMA(KP980785280, T9g, T9f);
cannam@95 868 T6Y = FNMS(KP707106781, T6v, T6s);
cannam@95 869 T6w = FMA(KP707106781, T6v, T6s);
cannam@95 870 T6C = FNMS(KP414213562, T6B, T6A);
cannam@95 871 T70 = FMA(KP414213562, T6A, T6B);
cannam@95 872 T9w = T6z + T6C;
cannam@95 873 T6D = T6z - T6C;
cannam@95 874 T9v = FNMS(KP707106781, T9o, T9n);
cannam@95 875 T9p = FMA(KP707106781, T9o, T9n);
cannam@95 876 {
cannam@95 877 E T77, T6O, T76, T6R;
cannam@95 878 T9q = T70 - T6Z;
cannam@95 879 T71 = T6Z + T70;
cannam@95 880 T77 = FMA(KP707106781, T6N, T6M);
cannam@95 881 T6O = FNMS(KP707106781, T6N, T6M);
cannam@95 882 T76 = FMA(KP707106781, T6Q, T6P);
cannam@95 883 T6R = FNMS(KP707106781, T6Q, T6P);
cannam@95 884 T6H = FNMS(KP707106781, T6G, T6F);
cannam@95 885 T74 = FMA(KP707106781, T6G, T6F);
cannam@95 886 T78 = FNMS(KP198912367, T77, T76);
cannam@95 887 T7c = FMA(KP198912367, T76, T77);
cannam@95 888 T6W = FMA(KP668178637, T6O, T6R);
cannam@95 889 T6S = FNMS(KP668178637, T6R, T6O);
cannam@95 890 }
cannam@95 891 }
cannam@95 892 {
cannam@95 893 E T6U, T6E, T9r, T9t, T73, T6K;
cannam@95 894 T6U = FNMS(KP923879532, T6D, T6w);
cannam@95 895 T6E = FMA(KP923879532, T6D, T6w);
cannam@95 896 T9r = FMA(KP923879532, T9q, T9p);
cannam@95 897 T9t = FNMS(KP923879532, T9q, T9p);
cannam@95 898 T73 = FMA(KP707106781, T6J, T6I);
cannam@95 899 T6K = FNMS(KP707106781, T6J, T6I);
cannam@95 900 {
cannam@95 901 E T7a, T9x, T9y, T7d;
cannam@95 902 {
cannam@95 903 E T72, T7b, T6V, T6L, T79, T75;
cannam@95 904 T7a = FMA(KP923879532, T71, T6Y);
cannam@95 905 T72 = FNMS(KP923879532, T71, T6Y);
cannam@95 906 T75 = FMA(KP198912367, T74, T73);
cannam@95 907 T7b = FNMS(KP198912367, T73, T74);
cannam@95 908 T6V = FNMS(KP668178637, T6H, T6K);
cannam@95 909 T6L = FMA(KP668178637, T6K, T6H);
cannam@95 910 T79 = T75 - T78;
cannam@95 911 T9A = T75 + T78;
cannam@95 912 T9z = FMA(KP923879532, T9w, T9v);
cannam@95 913 T9x = FNMS(KP923879532, T9w, T9v);
cannam@95 914 {
cannam@95 915 E T6X, T9s, T9u, T6T;
cannam@95 916 T6X = T6V - T6W;
cannam@95 917 T9s = T6V + T6W;
cannam@95 918 T9u = T6S - T6L;
cannam@95 919 T6T = T6L + T6S;
cannam@95 920 ri[WS(rs, 7)] = FMA(KP980785280, T79, T72);
cannam@95 921 ri[WS(rs, 23)] = FNMS(KP980785280, T79, T72);
cannam@95 922 ri[WS(rs, 11)] = FMA(KP831469612, T6X, T6U);
cannam@95 923 ri[WS(rs, 27)] = FNMS(KP831469612, T6X, T6U);
cannam@95 924 ii[WS(rs, 19)] = FNMS(KP831469612, T9s, T9r);
cannam@95 925 ii[WS(rs, 3)] = FMA(KP831469612, T9s, T9r);
cannam@95 926 ii[WS(rs, 27)] = FNMS(KP831469612, T9u, T9t);
cannam@95 927 ii[WS(rs, 11)] = FMA(KP831469612, T9u, T9t);
cannam@95 928 ri[WS(rs, 3)] = FMA(KP831469612, T6T, T6E);
cannam@95 929 ri[WS(rs, 19)] = FNMS(KP831469612, T6T, T6E);
cannam@95 930 T9y = T7c - T7b;
cannam@95 931 T7d = T7b + T7c;
cannam@95 932 }
cannam@95 933 }
cannam@95 934 ii[WS(rs, 23)] = FNMS(KP980785280, T9y, T9x);
cannam@95 935 ii[WS(rs, 7)] = FMA(KP980785280, T9y, T9x);
cannam@95 936 ri[WS(rs, 31)] = FMA(KP980785280, T7d, T7a);
cannam@95 937 ri[WS(rs, 15)] = FNMS(KP980785280, T7d, T7a);
cannam@95 938 }
cannam@95 939 }
cannam@95 940 }
cannam@95 941 }
cannam@95 942 }
cannam@95 943 }
cannam@95 944 }
cannam@95 945 ii[WS(rs, 31)] = FMA(KP980785280, T9A, T9z);
cannam@95 946 ii[WS(rs, 15)] = FNMS(KP980785280, T9A, T9z);
cannam@95 947 }
cannam@95 948 }
cannam@95 949 }
cannam@95 950
cannam@95 951 static const tw_instr twinstr[] = {
cannam@95 952 {TW_CEXP, 0, 1},
cannam@95 953 {TW_CEXP, 0, 3},
cannam@95 954 {TW_CEXP, 0, 9},
cannam@95 955 {TW_CEXP, 0, 27},
cannam@95 956 {TW_NEXT, 1, 0}
cannam@95 957 };
cannam@95 958
cannam@95 959 static const ct_desc desc = { 32, "t2_32", twinstr, &GENUS, {236, 98, 252, 0}, 0, 0, 0 };
cannam@95 960
cannam@95 961 void X(codelet_t2_32) (planner *p) {
cannam@95 962 X(kdft_dit_register) (p, t2_32, &desc);
cannam@95 963 }
cannam@95 964 #else /* HAVE_FMA */
cannam@95 965
cannam@95 966 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 32 -name t2_32 -include t.h */
cannam@95 967
cannam@95 968 /*
cannam@95 969 * This function contains 488 FP additions, 280 FP multiplications,
cannam@95 970 * (or, 376 additions, 168 multiplications, 112 fused multiply/add),
cannam@95 971 * 158 stack variables, 7 constants, and 128 memory accesses
cannam@95 972 */
cannam@95 973 #include "t.h"
cannam@95 974
cannam@95 975 static void t2_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
cannam@95 976 {
cannam@95 977 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
cannam@95 978 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
cannam@95 979 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
cannam@95 980 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
cannam@95 981 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
cannam@95 982 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
cannam@95 983 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
cannam@95 984 {
cannam@95 985 INT m;
cannam@95 986 for (m = mb, W = W + (mb * 8); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 8, MAKE_VOLATILE_STRIDE(64, rs)) {
cannam@95 987 E T2, T5, T3, T6, T8, TM, TO, Td, T9, Te, Th, Tl, TD, TH, T1y;
cannam@95 988 E T1H, T15, T1A, T11, T1F, T1n, T1p, T2q, T2I, T2u, T2K, T2V, T3b, T2Z, T3d;
cannam@95 989 E Tu, Ty, T3l, T3n, T1t, T1v, T2f, T2h, T1a, T1e, T32, T34, T1W, T1Y, T2C;
cannam@95 990 E T2E, Tg, TR, Tk, TS, Tm, TV, To, TT, T1M, T21, T1P, T22, T1Q, T25;
cannam@95 991 E T1S, T23;
cannam@95 992 {
cannam@95 993 E Ts, T1d, Tx, T18, Tt, T1c, Tw, T19, TB, T14, TG, TZ, TC, T13, TF;
cannam@95 994 E T10;
cannam@95 995 {
cannam@95 996 E T4, Tc, T7, Tb;
cannam@95 997 T2 = W[0];
cannam@95 998 T5 = W[1];
cannam@95 999 T3 = W[2];
cannam@95 1000 T6 = W[3];
cannam@95 1001 T4 = T2 * T3;
cannam@95 1002 Tc = T5 * T3;
cannam@95 1003 T7 = T5 * T6;
cannam@95 1004 Tb = T2 * T6;
cannam@95 1005 T8 = T4 + T7;
cannam@95 1006 TM = T4 - T7;
cannam@95 1007 TO = Tb + Tc;
cannam@95 1008 Td = Tb - Tc;
cannam@95 1009 T9 = W[4];
cannam@95 1010 Ts = T2 * T9;
cannam@95 1011 T1d = T6 * T9;
cannam@95 1012 Tx = T5 * T9;
cannam@95 1013 T18 = T3 * T9;
cannam@95 1014 Te = W[5];
cannam@95 1015 Tt = T5 * Te;
cannam@95 1016 T1c = T3 * Te;
cannam@95 1017 Tw = T2 * Te;
cannam@95 1018 T19 = T6 * Te;
cannam@95 1019 Th = W[6];
cannam@95 1020 TB = T3 * Th;
cannam@95 1021 T14 = T5 * Th;
cannam@95 1022 TG = T6 * Th;
cannam@95 1023 TZ = T2 * Th;
cannam@95 1024 Tl = W[7];
cannam@95 1025 TC = T6 * Tl;
cannam@95 1026 T13 = T2 * Tl;
cannam@95 1027 TF = T3 * Tl;
cannam@95 1028 T10 = T5 * Tl;
cannam@95 1029 }
cannam@95 1030 TD = TB + TC;
cannam@95 1031 TH = TF - TG;
cannam@95 1032 T1y = TZ + T10;
cannam@95 1033 T1H = TF + TG;
cannam@95 1034 T15 = T13 + T14;
cannam@95 1035 T1A = T13 - T14;
cannam@95 1036 T11 = TZ - T10;
cannam@95 1037 T1F = TB - TC;
cannam@95 1038 T1n = FMA(T9, Th, Te * Tl);
cannam@95 1039 T1p = FNMS(Te, Th, T9 * Tl);
cannam@95 1040 {
cannam@95 1041 E T2o, T2p, T2s, T2t;
cannam@95 1042 T2o = T8 * Th;
cannam@95 1043 T2p = Td * Tl;
cannam@95 1044 T2q = T2o + T2p;
cannam@95 1045 T2I = T2o - T2p;
cannam@95 1046 T2s = T8 * Tl;
cannam@95 1047 T2t = Td * Th;
cannam@95 1048 T2u = T2s - T2t;
cannam@95 1049 T2K = T2s + T2t;
cannam@95 1050 }
cannam@95 1051 {
cannam@95 1052 E T2T, T2U, T2X, T2Y;
cannam@95 1053 T2T = TM * Th;
cannam@95 1054 T2U = TO * Tl;
cannam@95 1055 T2V = T2T - T2U;
cannam@95 1056 T3b = T2T + T2U;
cannam@95 1057 T2X = TM * Tl;
cannam@95 1058 T2Y = TO * Th;
cannam@95 1059 T2Z = T2X + T2Y;
cannam@95 1060 T3d = T2X - T2Y;
cannam@95 1061 Tu = Ts + Tt;
cannam@95 1062 Ty = Tw - Tx;
cannam@95 1063 T3l = FMA(Tu, Th, Ty * Tl);
cannam@95 1064 T3n = FNMS(Ty, Th, Tu * Tl);
cannam@95 1065 }
cannam@95 1066 T1t = Ts - Tt;
cannam@95 1067 T1v = Tw + Tx;
cannam@95 1068 T2f = FMA(T1t, Th, T1v * Tl);
cannam@95 1069 T2h = FNMS(T1v, Th, T1t * Tl);
cannam@95 1070 T1a = T18 - T19;
cannam@95 1071 T1e = T1c + T1d;
cannam@95 1072 T32 = FMA(T1a, Th, T1e * Tl);
cannam@95 1073 T34 = FNMS(T1e, Th, T1a * Tl);
cannam@95 1074 T1W = T18 + T19;
cannam@95 1075 T1Y = T1c - T1d;
cannam@95 1076 T2C = FMA(T1W, Th, T1Y * Tl);
cannam@95 1077 T2E = FNMS(T1Y, Th, T1W * Tl);
cannam@95 1078 {
cannam@95 1079 E Ta, Tf, Ti, Tj;
cannam@95 1080 Ta = T8 * T9;
cannam@95 1081 Tf = Td * Te;
cannam@95 1082 Tg = Ta - Tf;
cannam@95 1083 TR = Ta + Tf;
cannam@95 1084 Ti = T8 * Te;
cannam@95 1085 Tj = Td * T9;
cannam@95 1086 Tk = Ti + Tj;
cannam@95 1087 TS = Ti - Tj;
cannam@95 1088 }
cannam@95 1089 Tm = FMA(Tg, Th, Tk * Tl);
cannam@95 1090 TV = FNMS(TS, Th, TR * Tl);
cannam@95 1091 To = FNMS(Tk, Th, Tg * Tl);
cannam@95 1092 TT = FMA(TR, Th, TS * Tl);
cannam@95 1093 {
cannam@95 1094 E T1K, T1L, T1N, T1O;
cannam@95 1095 T1K = TM * T9;
cannam@95 1096 T1L = TO * Te;
cannam@95 1097 T1M = T1K - T1L;
cannam@95 1098 T21 = T1K + T1L;
cannam@95 1099 T1N = TM * Te;
cannam@95 1100 T1O = TO * T9;
cannam@95 1101 T1P = T1N + T1O;
cannam@95 1102 T22 = T1N - T1O;
cannam@95 1103 }
cannam@95 1104 T1Q = FMA(T1M, Th, T1P * Tl);
cannam@95 1105 T25 = FNMS(T22, Th, T21 * Tl);
cannam@95 1106 T1S = FNMS(T1P, Th, T1M * Tl);
cannam@95 1107 T23 = FMA(T21, Th, T22 * Tl);
cannam@95 1108 }
cannam@95 1109 {
cannam@95 1110 E TL, T6f, T8c, T8q, T3F, T5t, T7I, T7W, T2y, T6B, T6y, T7j, T4k, T5J, T4B;
cannam@95 1111 E T5G, T3h, T6H, T6O, T7o, T4L, T5N, T52, T5Q, T1i, T7V, T6i, T7D, T3K, T5u;
cannam@95 1112 E T3P, T5v, T1E, T6n, T6m, T7e, T3W, T5y, T41, T5z, T29, T6p, T6s, T7f, T47;
cannam@95 1113 E T5B, T4c, T5C, T2R, T6z, T6E, T7k, T4v, T5H, T4E, T5K, T3y, T6P, T6K, T7p;
cannam@95 1114 E T4W, T5R, T55, T5O;
cannam@95 1115 {
cannam@95 1116 E T1, T7G, Tq, T7F, TA, T3C, TJ, T3D, Tn, Tp;
cannam@95 1117 T1 = ri[0];
cannam@95 1118 T7G = ii[0];
cannam@95 1119 Tn = ri[WS(rs, 16)];
cannam@95 1120 Tp = ii[WS(rs, 16)];
cannam@95 1121 Tq = FMA(Tm, Tn, To * Tp);
cannam@95 1122 T7F = FNMS(To, Tn, Tm * Tp);
cannam@95 1123 {
cannam@95 1124 E Tv, Tz, TE, TI;
cannam@95 1125 Tv = ri[WS(rs, 8)];
cannam@95 1126 Tz = ii[WS(rs, 8)];
cannam@95 1127 TA = FMA(Tu, Tv, Ty * Tz);
cannam@95 1128 T3C = FNMS(Ty, Tv, Tu * Tz);
cannam@95 1129 TE = ri[WS(rs, 24)];
cannam@95 1130 TI = ii[WS(rs, 24)];
cannam@95 1131 TJ = FMA(TD, TE, TH * TI);
cannam@95 1132 T3D = FNMS(TH, TE, TD * TI);
cannam@95 1133 }
cannam@95 1134 {
cannam@95 1135 E Tr, TK, T8a, T8b;
cannam@95 1136 Tr = T1 + Tq;
cannam@95 1137 TK = TA + TJ;
cannam@95 1138 TL = Tr + TK;
cannam@95 1139 T6f = Tr - TK;
cannam@95 1140 T8a = T7G - T7F;
cannam@95 1141 T8b = TA - TJ;
cannam@95 1142 T8c = T8a - T8b;
cannam@95 1143 T8q = T8b + T8a;
cannam@95 1144 }
cannam@95 1145 {
cannam@95 1146 E T3B, T3E, T7E, T7H;
cannam@95 1147 T3B = T1 - Tq;
cannam@95 1148 T3E = T3C - T3D;
cannam@95 1149 T3F = T3B - T3E;
cannam@95 1150 T5t = T3B + T3E;
cannam@95 1151 T7E = T3C + T3D;
cannam@95 1152 T7H = T7F + T7G;
cannam@95 1153 T7I = T7E + T7H;
cannam@95 1154 T7W = T7H - T7E;
cannam@95 1155 }
cannam@95 1156 }
cannam@95 1157 {
cannam@95 1158 E T2e, T4g, T2w, T4z, T2j, T4h, T2n, T4y;
cannam@95 1159 {
cannam@95 1160 E T2c, T2d, T2r, T2v;
cannam@95 1161 T2c = ri[WS(rs, 1)];
cannam@95 1162 T2d = ii[WS(rs, 1)];
cannam@95 1163 T2e = FMA(T2, T2c, T5 * T2d);
cannam@95 1164 T4g = FNMS(T5, T2c, T2 * T2d);
cannam@95 1165 T2r = ri[WS(rs, 25)];
cannam@95 1166 T2v = ii[WS(rs, 25)];
cannam@95 1167 T2w = FMA(T2q, T2r, T2u * T2v);
cannam@95 1168 T4z = FNMS(T2u, T2r, T2q * T2v);
cannam@95 1169 }
cannam@95 1170 {
cannam@95 1171 E T2g, T2i, T2l, T2m;
cannam@95 1172 T2g = ri[WS(rs, 17)];
cannam@95 1173 T2i = ii[WS(rs, 17)];
cannam@95 1174 T2j = FMA(T2f, T2g, T2h * T2i);
cannam@95 1175 T4h = FNMS(T2h, T2g, T2f * T2i);
cannam@95 1176 T2l = ri[WS(rs, 9)];
cannam@95 1177 T2m = ii[WS(rs, 9)];
cannam@95 1178 T2n = FMA(T9, T2l, Te * T2m);
cannam@95 1179 T4y = FNMS(Te, T2l, T9 * T2m);
cannam@95 1180 }
cannam@95 1181 {
cannam@95 1182 E T2k, T2x, T6w, T6x;
cannam@95 1183 T2k = T2e + T2j;
cannam@95 1184 T2x = T2n + T2w;
cannam@95 1185 T2y = T2k + T2x;
cannam@95 1186 T6B = T2k - T2x;
cannam@95 1187 T6w = T4g + T4h;
cannam@95 1188 T6x = T4y + T4z;
cannam@95 1189 T6y = T6w - T6x;
cannam@95 1190 T7j = T6w + T6x;
cannam@95 1191 }
cannam@95 1192 {
cannam@95 1193 E T4i, T4j, T4x, T4A;
cannam@95 1194 T4i = T4g - T4h;
cannam@95 1195 T4j = T2n - T2w;
cannam@95 1196 T4k = T4i + T4j;
cannam@95 1197 T5J = T4i - T4j;
cannam@95 1198 T4x = T2e - T2j;
cannam@95 1199 T4A = T4y - T4z;
cannam@95 1200 T4B = T4x - T4A;
cannam@95 1201 T5G = T4x + T4A;
cannam@95 1202 }
cannam@95 1203 }
cannam@95 1204 {
cannam@95 1205 E T31, T4Y, T3f, T4J, T36, T4Z, T3a, T4I;
cannam@95 1206 {
cannam@95 1207 E T2W, T30, T3c, T3e;
cannam@95 1208 T2W = ri[WS(rs, 31)];
cannam@95 1209 T30 = ii[WS(rs, 31)];
cannam@95 1210 T31 = FMA(T2V, T2W, T2Z * T30);
cannam@95 1211 T4Y = FNMS(T2Z, T2W, T2V * T30);
cannam@95 1212 T3c = ri[WS(rs, 23)];
cannam@95 1213 T3e = ii[WS(rs, 23)];
cannam@95 1214 T3f = FMA(T3b, T3c, T3d * T3e);
cannam@95 1215 T4J = FNMS(T3d, T3c, T3b * T3e);
cannam@95 1216 }
cannam@95 1217 {
cannam@95 1218 E T33, T35, T38, T39;
cannam@95 1219 T33 = ri[WS(rs, 15)];
cannam@95 1220 T35 = ii[WS(rs, 15)];
cannam@95 1221 T36 = FMA(T32, T33, T34 * T35);
cannam@95 1222 T4Z = FNMS(T34, T33, T32 * T35);
cannam@95 1223 T38 = ri[WS(rs, 7)];
cannam@95 1224 T39 = ii[WS(rs, 7)];
cannam@95 1225 T3a = FMA(TR, T38, TS * T39);
cannam@95 1226 T4I = FNMS(TS, T38, TR * T39);
cannam@95 1227 }
cannam@95 1228 {
cannam@95 1229 E T37, T3g, T6M, T6N;
cannam@95 1230 T37 = T31 + T36;
cannam@95 1231 T3g = T3a + T3f;
cannam@95 1232 T3h = T37 + T3g;
cannam@95 1233 T6H = T37 - T3g;
cannam@95 1234 T6M = T4Y + T4Z;
cannam@95 1235 T6N = T4I + T4J;
cannam@95 1236 T6O = T6M - T6N;
cannam@95 1237 T7o = T6M + T6N;
cannam@95 1238 }
cannam@95 1239 {
cannam@95 1240 E T4H, T4K, T50, T51;
cannam@95 1241 T4H = T31 - T36;
cannam@95 1242 T4K = T4I - T4J;
cannam@95 1243 T4L = T4H - T4K;
cannam@95 1244 T5N = T4H + T4K;
cannam@95 1245 T50 = T4Y - T4Z;
cannam@95 1246 T51 = T3a - T3f;
cannam@95 1247 T52 = T50 + T51;
cannam@95 1248 T5Q = T50 - T51;
cannam@95 1249 }
cannam@95 1250 }
cannam@95 1251 {
cannam@95 1252 E TQ, T3G, T1g, T3N, TX, T3H, T17, T3M;
cannam@95 1253 {
cannam@95 1254 E TN, TP, T1b, T1f;
cannam@95 1255 TN = ri[WS(rs, 4)];
cannam@95 1256 TP = ii[WS(rs, 4)];
cannam@95 1257 TQ = FMA(TM, TN, TO * TP);
cannam@95 1258 T3G = FNMS(TO, TN, TM * TP);
cannam@95 1259 T1b = ri[WS(rs, 12)];
cannam@95 1260 T1f = ii[WS(rs, 12)];
cannam@95 1261 T1g = FMA(T1a, T1b, T1e * T1f);
cannam@95 1262 T3N = FNMS(T1e, T1b, T1a * T1f);
cannam@95 1263 }
cannam@95 1264 {
cannam@95 1265 E TU, TW, T12, T16;
cannam@95 1266 TU = ri[WS(rs, 20)];
cannam@95 1267 TW = ii[WS(rs, 20)];
cannam@95 1268 TX = FMA(TT, TU, TV * TW);
cannam@95 1269 T3H = FNMS(TV, TU, TT * TW);
cannam@95 1270 T12 = ri[WS(rs, 28)];
cannam@95 1271 T16 = ii[WS(rs, 28)];
cannam@95 1272 T17 = FMA(T11, T12, T15 * T16);
cannam@95 1273 T3M = FNMS(T15, T12, T11 * T16);
cannam@95 1274 }
cannam@95 1275 {
cannam@95 1276 E TY, T1h, T6g, T6h;
cannam@95 1277 TY = TQ + TX;
cannam@95 1278 T1h = T17 + T1g;
cannam@95 1279 T1i = TY + T1h;
cannam@95 1280 T7V = T1h - TY;
cannam@95 1281 T6g = T3G + T3H;
cannam@95 1282 T6h = T3M + T3N;
cannam@95 1283 T6i = T6g - T6h;
cannam@95 1284 T7D = T6g + T6h;
cannam@95 1285 }
cannam@95 1286 {
cannam@95 1287 E T3I, T3J, T3L, T3O;
cannam@95 1288 T3I = T3G - T3H;
cannam@95 1289 T3J = TQ - TX;
cannam@95 1290 T3K = T3I - T3J;
cannam@95 1291 T5u = T3J + T3I;
cannam@95 1292 T3L = T17 - T1g;
cannam@95 1293 T3O = T3M - T3N;
cannam@95 1294 T3P = T3L + T3O;
cannam@95 1295 T5v = T3L - T3O;
cannam@95 1296 }
cannam@95 1297 }
cannam@95 1298 {
cannam@95 1299 E T1m, T3S, T1C, T3Z, T1r, T3T, T1x, T3Y;
cannam@95 1300 {
cannam@95 1301 E T1k, T1l, T1z, T1B;
cannam@95 1302 T1k = ri[WS(rs, 2)];
cannam@95 1303 T1l = ii[WS(rs, 2)];
cannam@95 1304 T1m = FMA(T8, T1k, Td * T1l);
cannam@95 1305 T3S = FNMS(Td, T1k, T8 * T1l);
cannam@95 1306 T1z = ri[WS(rs, 26)];
cannam@95 1307 T1B = ii[WS(rs, 26)];
cannam@95 1308 T1C = FMA(T1y, T1z, T1A * T1B);
cannam@95 1309 T3Z = FNMS(T1A, T1z, T1y * T1B);
cannam@95 1310 }
cannam@95 1311 {
cannam@95 1312 E T1o, T1q, T1u, T1w;
cannam@95 1313 T1o = ri[WS(rs, 18)];
cannam@95 1314 T1q = ii[WS(rs, 18)];
cannam@95 1315 T1r = FMA(T1n, T1o, T1p * T1q);
cannam@95 1316 T3T = FNMS(T1p, T1o, T1n * T1q);
cannam@95 1317 T1u = ri[WS(rs, 10)];
cannam@95 1318 T1w = ii[WS(rs, 10)];
cannam@95 1319 T1x = FMA(T1t, T1u, T1v * T1w);
cannam@95 1320 T3Y = FNMS(T1v, T1u, T1t * T1w);
cannam@95 1321 }
cannam@95 1322 {
cannam@95 1323 E T1s, T1D, T6k, T6l;
cannam@95 1324 T1s = T1m + T1r;
cannam@95 1325 T1D = T1x + T1C;
cannam@95 1326 T1E = T1s + T1D;
cannam@95 1327 T6n = T1s - T1D;
cannam@95 1328 T6k = T3S + T3T;
cannam@95 1329 T6l = T3Y + T3Z;
cannam@95 1330 T6m = T6k - T6l;
cannam@95 1331 T7e = T6k + T6l;
cannam@95 1332 }
cannam@95 1333 {
cannam@95 1334 E T3U, T3V, T3X, T40;
cannam@95 1335 T3U = T3S - T3T;
cannam@95 1336 T3V = T1x - T1C;
cannam@95 1337 T3W = T3U + T3V;
cannam@95 1338 T5y = T3U - T3V;
cannam@95 1339 T3X = T1m - T1r;
cannam@95 1340 T40 = T3Y - T3Z;
cannam@95 1341 T41 = T3X - T40;
cannam@95 1342 T5z = T3X + T40;
cannam@95 1343 }
cannam@95 1344 }
cannam@95 1345 {
cannam@95 1346 E T1J, T43, T27, T4a, T1U, T44, T20, T49;
cannam@95 1347 {
cannam@95 1348 E T1G, T1I, T24, T26;
cannam@95 1349 T1G = ri[WS(rs, 30)];
cannam@95 1350 T1I = ii[WS(rs, 30)];
cannam@95 1351 T1J = FMA(T1F, T1G, T1H * T1I);
cannam@95 1352 T43 = FNMS(T1H, T1G, T1F * T1I);
cannam@95 1353 T24 = ri[WS(rs, 22)];
cannam@95 1354 T26 = ii[WS(rs, 22)];
cannam@95 1355 T27 = FMA(T23, T24, T25 * T26);
cannam@95 1356 T4a = FNMS(T25, T24, T23 * T26);
cannam@95 1357 }
cannam@95 1358 {
cannam@95 1359 E T1R, T1T, T1X, T1Z;
cannam@95 1360 T1R = ri[WS(rs, 14)];
cannam@95 1361 T1T = ii[WS(rs, 14)];
cannam@95 1362 T1U = FMA(T1Q, T1R, T1S * T1T);
cannam@95 1363 T44 = FNMS(T1S, T1R, T1Q * T1T);
cannam@95 1364 T1X = ri[WS(rs, 6)];
cannam@95 1365 T1Z = ii[WS(rs, 6)];
cannam@95 1366 T20 = FMA(T1W, T1X, T1Y * T1Z);
cannam@95 1367 T49 = FNMS(T1Y, T1X, T1W * T1Z);
cannam@95 1368 }
cannam@95 1369 {
cannam@95 1370 E T1V, T28, T6q, T6r;
cannam@95 1371 T1V = T1J + T1U;
cannam@95 1372 T28 = T20 + T27;
cannam@95 1373 T29 = T1V + T28;
cannam@95 1374 T6p = T1V - T28;
cannam@95 1375 T6q = T43 + T44;
cannam@95 1376 T6r = T49 + T4a;
cannam@95 1377 T6s = T6q - T6r;
cannam@95 1378 T7f = T6q + T6r;
cannam@95 1379 }
cannam@95 1380 {
cannam@95 1381 E T45, T46, T48, T4b;
cannam@95 1382 T45 = T43 - T44;
cannam@95 1383 T46 = T20 - T27;
cannam@95 1384 T47 = T45 + T46;
cannam@95 1385 T5B = T45 - T46;
cannam@95 1386 T48 = T1J - T1U;
cannam@95 1387 T4b = T49 - T4a;
cannam@95 1388 T4c = T48 - T4b;
cannam@95 1389 T5C = T48 + T4b;
cannam@95 1390 }
cannam@95 1391 }
cannam@95 1392 {
cannam@95 1393 E T2B, T4r, T2G, T4s, T4q, T4t, T2M, T4m, T2P, T4n, T4l, T4o;
cannam@95 1394 {
cannam@95 1395 E T2z, T2A, T2D, T2F;
cannam@95 1396 T2z = ri[WS(rs, 5)];
cannam@95 1397 T2A = ii[WS(rs, 5)];
cannam@95 1398 T2B = FMA(T21, T2z, T22 * T2A);
cannam@95 1399 T4r = FNMS(T22, T2z, T21 * T2A);
cannam@95 1400 T2D = ri[WS(rs, 21)];
cannam@95 1401 T2F = ii[WS(rs, 21)];
cannam@95 1402 T2G = FMA(T2C, T2D, T2E * T2F);
cannam@95 1403 T4s = FNMS(T2E, T2D, T2C * T2F);
cannam@95 1404 }
cannam@95 1405 T4q = T2B - T2G;
cannam@95 1406 T4t = T4r - T4s;
cannam@95 1407 {
cannam@95 1408 E T2J, T2L, T2N, T2O;
cannam@95 1409 T2J = ri[WS(rs, 29)];
cannam@95 1410 T2L = ii[WS(rs, 29)];
cannam@95 1411 T2M = FMA(T2I, T2J, T2K * T2L);
cannam@95 1412 T4m = FNMS(T2K, T2J, T2I * T2L);
cannam@95 1413 T2N = ri[WS(rs, 13)];
cannam@95 1414 T2O = ii[WS(rs, 13)];
cannam@95 1415 T2P = FMA(T1M, T2N, T1P * T2O);
cannam@95 1416 T4n = FNMS(T1P, T2N, T1M * T2O);
cannam@95 1417 }
cannam@95 1418 T4l = T2M - T2P;
cannam@95 1419 T4o = T4m - T4n;
cannam@95 1420 {
cannam@95 1421 E T2H, T2Q, T6C, T6D;
cannam@95 1422 T2H = T2B + T2G;
cannam@95 1423 T2Q = T2M + T2P;
cannam@95 1424 T2R = T2H + T2Q;
cannam@95 1425 T6z = T2Q - T2H;
cannam@95 1426 T6C = T4r + T4s;
cannam@95 1427 T6D = T4m + T4n;
cannam@95 1428 T6E = T6C - T6D;
cannam@95 1429 T7k = T6C + T6D;
cannam@95 1430 }
cannam@95 1431 {
cannam@95 1432 E T4p, T4u, T4C, T4D;
cannam@95 1433 T4p = T4l - T4o;
cannam@95 1434 T4u = T4q + T4t;
cannam@95 1435 T4v = KP707106781 * (T4p - T4u);
cannam@95 1436 T5H = KP707106781 * (T4u + T4p);
cannam@95 1437 T4C = T4t - T4q;
cannam@95 1438 T4D = T4l + T4o;
cannam@95 1439 T4E = KP707106781 * (T4C - T4D);
cannam@95 1440 T5K = KP707106781 * (T4C + T4D);
cannam@95 1441 }
cannam@95 1442 }
cannam@95 1443 {
cannam@95 1444 E T3k, T4M, T3p, T4N, T4O, T4P, T3t, T4S, T3w, T4T, T4R, T4U;
cannam@95 1445 {
cannam@95 1446 E T3i, T3j, T3m, T3o;
cannam@95 1447 T3i = ri[WS(rs, 3)];
cannam@95 1448 T3j = ii[WS(rs, 3)];
cannam@95 1449 T3k = FMA(T3, T3i, T6 * T3j);
cannam@95 1450 T4M = FNMS(T6, T3i, T3 * T3j);
cannam@95 1451 T3m = ri[WS(rs, 19)];
cannam@95 1452 T3o = ii[WS(rs, 19)];
cannam@95 1453 T3p = FMA(T3l, T3m, T3n * T3o);
cannam@95 1454 T4N = FNMS(T3n, T3m, T3l * T3o);
cannam@95 1455 }
cannam@95 1456 T4O = T4M - T4N;
cannam@95 1457 T4P = T3k - T3p;
cannam@95 1458 {
cannam@95 1459 E T3r, T3s, T3u, T3v;
cannam@95 1460 T3r = ri[WS(rs, 27)];
cannam@95 1461 T3s = ii[WS(rs, 27)];
cannam@95 1462 T3t = FMA(Th, T3r, Tl * T3s);
cannam@95 1463 T4S = FNMS(Tl, T3r, Th * T3s);
cannam@95 1464 T3u = ri[WS(rs, 11)];
cannam@95 1465 T3v = ii[WS(rs, 11)];
cannam@95 1466 T3w = FMA(Tg, T3u, Tk * T3v);
cannam@95 1467 T4T = FNMS(Tk, T3u, Tg * T3v);
cannam@95 1468 }
cannam@95 1469 T4R = T3t - T3w;
cannam@95 1470 T4U = T4S - T4T;
cannam@95 1471 {
cannam@95 1472 E T3q, T3x, T6I, T6J;
cannam@95 1473 T3q = T3k + T3p;
cannam@95 1474 T3x = T3t + T3w;
cannam@95 1475 T3y = T3q + T3x;
cannam@95 1476 T6P = T3x - T3q;
cannam@95 1477 T6I = T4M + T4N;
cannam@95 1478 T6J = T4S + T4T;
cannam@95 1479 T6K = T6I - T6J;
cannam@95 1480 T7p = T6I + T6J;
cannam@95 1481 }
cannam@95 1482 {
cannam@95 1483 E T4Q, T4V, T53, T54;
cannam@95 1484 T4Q = T4O - T4P;
cannam@95 1485 T4V = T4R + T4U;
cannam@95 1486 T4W = KP707106781 * (T4Q - T4V);
cannam@95 1487 T5R = KP707106781 * (T4Q + T4V);
cannam@95 1488 T53 = T4R - T4U;
cannam@95 1489 T54 = T4P + T4O;
cannam@95 1490 T55 = KP707106781 * (T53 - T54);
cannam@95 1491 T5O = KP707106781 * (T54 + T53);
cannam@95 1492 }
cannam@95 1493 }
cannam@95 1494 {
cannam@95 1495 E T2b, T7x, T7K, T7M, T3A, T7L, T7A, T7B;
cannam@95 1496 {
cannam@95 1497 E T1j, T2a, T7C, T7J;
cannam@95 1498 T1j = TL + T1i;
cannam@95 1499 T2a = T1E + T29;
cannam@95 1500 T2b = T1j + T2a;
cannam@95 1501 T7x = T1j - T2a;
cannam@95 1502 T7C = T7e + T7f;
cannam@95 1503 T7J = T7D + T7I;
cannam@95 1504 T7K = T7C + T7J;
cannam@95 1505 T7M = T7J - T7C;
cannam@95 1506 }
cannam@95 1507 {
cannam@95 1508 E T2S, T3z, T7y, T7z;
cannam@95 1509 T2S = T2y + T2R;
cannam@95 1510 T3z = T3h + T3y;
cannam@95 1511 T3A = T2S + T3z;
cannam@95 1512 T7L = T3z - T2S;
cannam@95 1513 T7y = T7j + T7k;
cannam@95 1514 T7z = T7o + T7p;
cannam@95 1515 T7A = T7y - T7z;
cannam@95 1516 T7B = T7y + T7z;
cannam@95 1517 }
cannam@95 1518 ri[WS(rs, 16)] = T2b - T3A;
cannam@95 1519 ii[WS(rs, 16)] = T7K - T7B;
cannam@95 1520 ri[0] = T2b + T3A;
cannam@95 1521 ii[0] = T7B + T7K;
cannam@95 1522 ri[WS(rs, 24)] = T7x - T7A;
cannam@95 1523 ii[WS(rs, 24)] = T7M - T7L;
cannam@95 1524 ri[WS(rs, 8)] = T7x + T7A;
cannam@95 1525 ii[WS(rs, 8)] = T7L + T7M;
cannam@95 1526 }
cannam@95 1527 {
cannam@95 1528 E T7h, T7t, T7Q, T7S, T7m, T7u, T7r, T7v;
cannam@95 1529 {
cannam@95 1530 E T7d, T7g, T7O, T7P;
cannam@95 1531 T7d = TL - T1i;
cannam@95 1532 T7g = T7e - T7f;
cannam@95 1533 T7h = T7d + T7g;
cannam@95 1534 T7t = T7d - T7g;
cannam@95 1535 T7O = T29 - T1E;
cannam@95 1536 T7P = T7I - T7D;
cannam@95 1537 T7Q = T7O + T7P;
cannam@95 1538 T7S = T7P - T7O;
cannam@95 1539 }
cannam@95 1540 {
cannam@95 1541 E T7i, T7l, T7n, T7q;
cannam@95 1542 T7i = T2y - T2R;
cannam@95 1543 T7l = T7j - T7k;
cannam@95 1544 T7m = T7i + T7l;
cannam@95 1545 T7u = T7l - T7i;
cannam@95 1546 T7n = T3h - T3y;
cannam@95 1547 T7q = T7o - T7p;
cannam@95 1548 T7r = T7n - T7q;
cannam@95 1549 T7v = T7n + T7q;
cannam@95 1550 }
cannam@95 1551 {
cannam@95 1552 E T7s, T7N, T7w, T7R;
cannam@95 1553 T7s = KP707106781 * (T7m + T7r);
cannam@95 1554 ri[WS(rs, 20)] = T7h - T7s;
cannam@95 1555 ri[WS(rs, 4)] = T7h + T7s;
cannam@95 1556 T7N = KP707106781 * (T7u + T7v);
cannam@95 1557 ii[WS(rs, 4)] = T7N + T7Q;
cannam@95 1558 ii[WS(rs, 20)] = T7Q - T7N;
cannam@95 1559 T7w = KP707106781 * (T7u - T7v);
cannam@95 1560 ri[WS(rs, 28)] = T7t - T7w;
cannam@95 1561 ri[WS(rs, 12)] = T7t + T7w;
cannam@95 1562 T7R = KP707106781 * (T7r - T7m);
cannam@95 1563 ii[WS(rs, 12)] = T7R + T7S;
cannam@95 1564 ii[WS(rs, 28)] = T7S - T7R;
cannam@95 1565 }
cannam@95 1566 }
cannam@95 1567 {
cannam@95 1568 E T6j, T7X, T83, T6X, T6u, T7U, T77, T7b, T70, T82, T6G, T6U, T74, T7a, T6R;
cannam@95 1569 E T6V;
cannam@95 1570 {
cannam@95 1571 E T6o, T6t, T6A, T6F;
cannam@95 1572 T6j = T6f - T6i;
cannam@95 1573 T7X = T7V + T7W;
cannam@95 1574 T83 = T7W - T7V;
cannam@95 1575 T6X = T6f + T6i;
cannam@95 1576 T6o = T6m - T6n;
cannam@95 1577 T6t = T6p + T6s;
cannam@95 1578 T6u = KP707106781 * (T6o - T6t);
cannam@95 1579 T7U = KP707106781 * (T6o + T6t);
cannam@95 1580 {
cannam@95 1581 E T75, T76, T6Y, T6Z;
cannam@95 1582 T75 = T6H + T6K;
cannam@95 1583 T76 = T6O + T6P;
cannam@95 1584 T77 = FNMS(KP382683432, T76, KP923879532 * T75);
cannam@95 1585 T7b = FMA(KP923879532, T76, KP382683432 * T75);
cannam@95 1586 T6Y = T6n + T6m;
cannam@95 1587 T6Z = T6p - T6s;
cannam@95 1588 T70 = KP707106781 * (T6Y + T6Z);
cannam@95 1589 T82 = KP707106781 * (T6Z - T6Y);
cannam@95 1590 }
cannam@95 1591 T6A = T6y - T6z;
cannam@95 1592 T6F = T6B - T6E;
cannam@95 1593 T6G = FMA(KP923879532, T6A, KP382683432 * T6F);
cannam@95 1594 T6U = FNMS(KP923879532, T6F, KP382683432 * T6A);
cannam@95 1595 {
cannam@95 1596 E T72, T73, T6L, T6Q;
cannam@95 1597 T72 = T6y + T6z;
cannam@95 1598 T73 = T6B + T6E;
cannam@95 1599 T74 = FMA(KP382683432, T72, KP923879532 * T73);
cannam@95 1600 T7a = FNMS(KP382683432, T73, KP923879532 * T72);
cannam@95 1601 T6L = T6H - T6K;
cannam@95 1602 T6Q = T6O - T6P;
cannam@95 1603 T6R = FNMS(KP923879532, T6Q, KP382683432 * T6L);
cannam@95 1604 T6V = FMA(KP382683432, T6Q, KP923879532 * T6L);
cannam@95 1605 }
cannam@95 1606 }
cannam@95 1607 {
cannam@95 1608 E T6v, T6S, T81, T84;
cannam@95 1609 T6v = T6j + T6u;
cannam@95 1610 T6S = T6G + T6R;
cannam@95 1611 ri[WS(rs, 22)] = T6v - T6S;
cannam@95 1612 ri[WS(rs, 6)] = T6v + T6S;
cannam@95 1613 T81 = T6U + T6V;
cannam@95 1614 T84 = T82 + T83;
cannam@95 1615 ii[WS(rs, 6)] = T81 + T84;
cannam@95 1616 ii[WS(rs, 22)] = T84 - T81;
cannam@95 1617 }
cannam@95 1618 {
cannam@95 1619 E T6T, T6W, T85, T86;
cannam@95 1620 T6T = T6j - T6u;
cannam@95 1621 T6W = T6U - T6V;
cannam@95 1622 ri[WS(rs, 30)] = T6T - T6W;
cannam@95 1623 ri[WS(rs, 14)] = T6T + T6W;
cannam@95 1624 T85 = T6R - T6G;
cannam@95 1625 T86 = T83 - T82;
cannam@95 1626 ii[WS(rs, 14)] = T85 + T86;
cannam@95 1627 ii[WS(rs, 30)] = T86 - T85;
cannam@95 1628 }
cannam@95 1629 {
cannam@95 1630 E T71, T78, T7T, T7Y;
cannam@95 1631 T71 = T6X + T70;
cannam@95 1632 T78 = T74 + T77;
cannam@95 1633 ri[WS(rs, 18)] = T71 - T78;
cannam@95 1634 ri[WS(rs, 2)] = T71 + T78;
cannam@95 1635 T7T = T7a + T7b;
cannam@95 1636 T7Y = T7U + T7X;
cannam@95 1637 ii[WS(rs, 2)] = T7T + T7Y;
cannam@95 1638 ii[WS(rs, 18)] = T7Y - T7T;
cannam@95 1639 }
cannam@95 1640 {
cannam@95 1641 E T79, T7c, T7Z, T80;
cannam@95 1642 T79 = T6X - T70;
cannam@95 1643 T7c = T7a - T7b;
cannam@95 1644 ri[WS(rs, 26)] = T79 - T7c;
cannam@95 1645 ri[WS(rs, 10)] = T79 + T7c;
cannam@95 1646 T7Z = T77 - T74;
cannam@95 1647 T80 = T7X - T7U;
cannam@95 1648 ii[WS(rs, 10)] = T7Z + T80;
cannam@95 1649 ii[WS(rs, 26)] = T80 - T7Z;
cannam@95 1650 }
cannam@95 1651 }
cannam@95 1652 {
cannam@95 1653 E T3R, T5d, T8r, T8x, T4e, T8o, T5n, T5r, T4G, T5a, T5g, T8w, T5k, T5q, T57;
cannam@95 1654 E T5b, T3Q, T8p;
cannam@95 1655 T3Q = KP707106781 * (T3K - T3P);
cannam@95 1656 T3R = T3F - T3Q;
cannam@95 1657 T5d = T3F + T3Q;
cannam@95 1658 T8p = KP707106781 * (T5v - T5u);
cannam@95 1659 T8r = T8p + T8q;
cannam@95 1660 T8x = T8q - T8p;
cannam@95 1661 {
cannam@95 1662 E T42, T4d, T5l, T5m;
cannam@95 1663 T42 = FNMS(KP923879532, T41, KP382683432 * T3W);
cannam@95 1664 T4d = FMA(KP382683432, T47, KP923879532 * T4c);
cannam@95 1665 T4e = T42 - T4d;
cannam@95 1666 T8o = T42 + T4d;
cannam@95 1667 T5l = T4L + T4W;
cannam@95 1668 T5m = T52 + T55;
cannam@95 1669 T5n = FNMS(KP555570233, T5m, KP831469612 * T5l);
cannam@95 1670 T5r = FMA(KP831469612, T5m, KP555570233 * T5l);
cannam@95 1671 }
cannam@95 1672 {
cannam@95 1673 E T4w, T4F, T5e, T5f;
cannam@95 1674 T4w = T4k - T4v;
cannam@95 1675 T4F = T4B - T4E;
cannam@95 1676 T4G = FMA(KP980785280, T4w, KP195090322 * T4F);
cannam@95 1677 T5a = FNMS(KP980785280, T4F, KP195090322 * T4w);
cannam@95 1678 T5e = FMA(KP923879532, T3W, KP382683432 * T41);
cannam@95 1679 T5f = FNMS(KP923879532, T47, KP382683432 * T4c);
cannam@95 1680 T5g = T5e + T5f;
cannam@95 1681 T8w = T5f - T5e;
cannam@95 1682 }
cannam@95 1683 {
cannam@95 1684 E T5i, T5j, T4X, T56;
cannam@95 1685 T5i = T4k + T4v;
cannam@95 1686 T5j = T4B + T4E;
cannam@95 1687 T5k = FMA(KP555570233, T5i, KP831469612 * T5j);
cannam@95 1688 T5q = FNMS(KP555570233, T5j, KP831469612 * T5i);
cannam@95 1689 T4X = T4L - T4W;
cannam@95 1690 T56 = T52 - T55;
cannam@95 1691 T57 = FNMS(KP980785280, T56, KP195090322 * T4X);
cannam@95 1692 T5b = FMA(KP195090322, T56, KP980785280 * T4X);
cannam@95 1693 }
cannam@95 1694 {
cannam@95 1695 E T4f, T58, T8v, T8y;
cannam@95 1696 T4f = T3R + T4e;
cannam@95 1697 T58 = T4G + T57;
cannam@95 1698 ri[WS(rs, 23)] = T4f - T58;
cannam@95 1699 ri[WS(rs, 7)] = T4f + T58;
cannam@95 1700 T8v = T5a + T5b;
cannam@95 1701 T8y = T8w + T8x;
cannam@95 1702 ii[WS(rs, 7)] = T8v + T8y;
cannam@95 1703 ii[WS(rs, 23)] = T8y - T8v;
cannam@95 1704 }
cannam@95 1705 {
cannam@95 1706 E T59, T5c, T8z, T8A;
cannam@95 1707 T59 = T3R - T4e;
cannam@95 1708 T5c = T5a - T5b;
cannam@95 1709 ri[WS(rs, 31)] = T59 - T5c;
cannam@95 1710 ri[WS(rs, 15)] = T59 + T5c;
cannam@95 1711 T8z = T57 - T4G;
cannam@95 1712 T8A = T8x - T8w;
cannam@95 1713 ii[WS(rs, 15)] = T8z + T8A;
cannam@95 1714 ii[WS(rs, 31)] = T8A - T8z;
cannam@95 1715 }
cannam@95 1716 {
cannam@95 1717 E T5h, T5o, T8n, T8s;
cannam@95 1718 T5h = T5d + T5g;
cannam@95 1719 T5o = T5k + T5n;
cannam@95 1720 ri[WS(rs, 19)] = T5h - T5o;
cannam@95 1721 ri[WS(rs, 3)] = T5h + T5o;
cannam@95 1722 T8n = T5q + T5r;
cannam@95 1723 T8s = T8o + T8r;
cannam@95 1724 ii[WS(rs, 3)] = T8n + T8s;
cannam@95 1725 ii[WS(rs, 19)] = T8s - T8n;
cannam@95 1726 }
cannam@95 1727 {
cannam@95 1728 E T5p, T5s, T8t, T8u;
cannam@95 1729 T5p = T5d - T5g;
cannam@95 1730 T5s = T5q - T5r;
cannam@95 1731 ri[WS(rs, 27)] = T5p - T5s;
cannam@95 1732 ri[WS(rs, 11)] = T5p + T5s;
cannam@95 1733 T8t = T5n - T5k;
cannam@95 1734 T8u = T8r - T8o;
cannam@95 1735 ii[WS(rs, 11)] = T8t + T8u;
cannam@95 1736 ii[WS(rs, 27)] = T8u - T8t;
cannam@95 1737 }
cannam@95 1738 }
cannam@95 1739 {
cannam@95 1740 E T5x, T5Z, T8d, T8j, T5E, T88, T69, T6d, T5M, T5W, T62, T8i, T66, T6c, T5T;
cannam@95 1741 E T5X, T5w, T89;
cannam@95 1742 T5w = KP707106781 * (T5u + T5v);
cannam@95 1743 T5x = T5t - T5w;
cannam@95 1744 T5Z = T5t + T5w;
cannam@95 1745 T89 = KP707106781 * (T3K + T3P);
cannam@95 1746 T8d = T89 + T8c;
cannam@95 1747 T8j = T8c - T89;
cannam@95 1748 {
cannam@95 1749 E T5A, T5D, T67, T68;
cannam@95 1750 T5A = FNMS(KP382683432, T5z, KP923879532 * T5y);
cannam@95 1751 T5D = FMA(KP923879532, T5B, KP382683432 * T5C);
cannam@95 1752 T5E = T5A - T5D;
cannam@95 1753 T88 = T5A + T5D;
cannam@95 1754 T67 = T5N + T5O;
cannam@95 1755 T68 = T5Q + T5R;
cannam@95 1756 T69 = FNMS(KP195090322, T68, KP980785280 * T67);
cannam@95 1757 T6d = FMA(KP195090322, T67, KP980785280 * T68);
cannam@95 1758 }
cannam@95 1759 {
cannam@95 1760 E T5I, T5L, T60, T61;
cannam@95 1761 T5I = T5G - T5H;
cannam@95 1762 T5L = T5J - T5K;
cannam@95 1763 T5M = FMA(KP555570233, T5I, KP831469612 * T5L);
cannam@95 1764 T5W = FNMS(KP831469612, T5I, KP555570233 * T5L);
cannam@95 1765 T60 = FMA(KP382683432, T5y, KP923879532 * T5z);
cannam@95 1766 T61 = FNMS(KP382683432, T5B, KP923879532 * T5C);
cannam@95 1767 T62 = T60 + T61;
cannam@95 1768 T8i = T61 - T60;
cannam@95 1769 }
cannam@95 1770 {
cannam@95 1771 E T64, T65, T5P, T5S;
cannam@95 1772 T64 = T5G + T5H;
cannam@95 1773 T65 = T5J + T5K;
cannam@95 1774 T66 = FMA(KP980785280, T64, KP195090322 * T65);
cannam@95 1775 T6c = FNMS(KP195090322, T64, KP980785280 * T65);
cannam@95 1776 T5P = T5N - T5O;
cannam@95 1777 T5S = T5Q - T5R;
cannam@95 1778 T5T = FNMS(KP831469612, T5S, KP555570233 * T5P);
cannam@95 1779 T5X = FMA(KP831469612, T5P, KP555570233 * T5S);
cannam@95 1780 }
cannam@95 1781 {
cannam@95 1782 E T5F, T5U, T8h, T8k;
cannam@95 1783 T5F = T5x + T5E;
cannam@95 1784 T5U = T5M + T5T;
cannam@95 1785 ri[WS(rs, 21)] = T5F - T5U;
cannam@95 1786 ri[WS(rs, 5)] = T5F + T5U;
cannam@95 1787 T8h = T5W + T5X;
cannam@95 1788 T8k = T8i + T8j;
cannam@95 1789 ii[WS(rs, 5)] = T8h + T8k;
cannam@95 1790 ii[WS(rs, 21)] = T8k - T8h;
cannam@95 1791 }
cannam@95 1792 {
cannam@95 1793 E T5V, T5Y, T8l, T8m;
cannam@95 1794 T5V = T5x - T5E;
cannam@95 1795 T5Y = T5W - T5X;
cannam@95 1796 ri[WS(rs, 29)] = T5V - T5Y;
cannam@95 1797 ri[WS(rs, 13)] = T5V + T5Y;
cannam@95 1798 T8l = T5T - T5M;
cannam@95 1799 T8m = T8j - T8i;
cannam@95 1800 ii[WS(rs, 13)] = T8l + T8m;
cannam@95 1801 ii[WS(rs, 29)] = T8m - T8l;
cannam@95 1802 }
cannam@95 1803 {
cannam@95 1804 E T63, T6a, T87, T8e;
cannam@95 1805 T63 = T5Z + T62;
cannam@95 1806 T6a = T66 + T69;
cannam@95 1807 ri[WS(rs, 17)] = T63 - T6a;
cannam@95 1808 ri[WS(rs, 1)] = T63 + T6a;
cannam@95 1809 T87 = T6c + T6d;
cannam@95 1810 T8e = T88 + T8d;
cannam@95 1811 ii[WS(rs, 1)] = T87 + T8e;
cannam@95 1812 ii[WS(rs, 17)] = T8e - T87;
cannam@95 1813 }
cannam@95 1814 {
cannam@95 1815 E T6b, T6e, T8f, T8g;
cannam@95 1816 T6b = T5Z - T62;
cannam@95 1817 T6e = T6c - T6d;
cannam@95 1818 ri[WS(rs, 25)] = T6b - T6e;
cannam@95 1819 ri[WS(rs, 9)] = T6b + T6e;
cannam@95 1820 T8f = T69 - T66;
cannam@95 1821 T8g = T8d - T88;
cannam@95 1822 ii[WS(rs, 9)] = T8f + T8g;
cannam@95 1823 ii[WS(rs, 25)] = T8g - T8f;
cannam@95 1824 }
cannam@95 1825 }
cannam@95 1826 }
cannam@95 1827 }
cannam@95 1828 }
cannam@95 1829 }
cannam@95 1830
cannam@95 1831 static const tw_instr twinstr[] = {
cannam@95 1832 {TW_CEXP, 0, 1},
cannam@95 1833 {TW_CEXP, 0, 3},
cannam@95 1834 {TW_CEXP, 0, 9},
cannam@95 1835 {TW_CEXP, 0, 27},
cannam@95 1836 {TW_NEXT, 1, 0}
cannam@95 1837 };
cannam@95 1838
cannam@95 1839 static const ct_desc desc = { 32, "t2_32", twinstr, &GENUS, {376, 168, 112, 0}, 0, 0, 0 };
cannam@95 1840
cannam@95 1841 void X(codelet_t2_32) (planner *p) {
cannam@95 1842 X(kdft_dit_register) (p, t2_32, &desc);
cannam@95 1843 }
cannam@95 1844 #endif /* HAVE_FMA */