Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/dft/scalar/codelets/t1_20.c @ 82:d0c2a83c1364

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add FFTW 3.3.8 source, and a Linux build
author Chris Cannam
date Tue, 19 Nov 2019 14:52:55 +0000
parents
children
rev   line source
Chris@82 1 /*
Chris@82 2 * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@82 3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@82 4 *
Chris@82 5 * This program is free software; you can redistribute it and/or modify
Chris@82 6 * it under the terms of the GNU General Public License as published by
Chris@82 7 * the Free Software Foundation; either version 2 of the License, or
Chris@82 8 * (at your option) any later version.
Chris@82 9 *
Chris@82 10 * This program is distributed in the hope that it will be useful,
Chris@82 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@82 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
Chris@82 13 * GNU General Public License for more details.
Chris@82 14 *
Chris@82 15 * You should have received a copy of the GNU General Public License
Chris@82 16 * along with this program; if not, write to the Free Software
Chris@82 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Chris@82 18 *
Chris@82 19 */
Chris@82 20
Chris@82 21 /* This file was automatically generated --- DO NOT EDIT */
Chris@82 22 /* Generated on Thu May 24 08:04:15 EDT 2018 */
Chris@82 23
Chris@82 24 #include "dft/codelet-dft.h"
Chris@82 25
Chris@82 26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
Chris@82 27
Chris@82 28 /* Generated by: ../../../genfft/gen_twiddle.native -fma -compact -variables 4 -pipeline-latency 4 -n 20 -name t1_20 -include dft/scalar/t.h */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 246 FP additions, 148 FP multiplications,
Chris@82 32 * (or, 136 additions, 38 multiplications, 110 fused multiply/add),
Chris@82 33 * 61 stack variables, 4 constants, and 80 memory accesses
Chris@82 34 */
Chris@82 35 #include "dft/scalar/t.h"
Chris@82 36
Chris@82 37 static void t1_20(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@82 38 {
Chris@82 39 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@82 40 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@82 41 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@82 42 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
Chris@82 43 {
Chris@82 44 INT m;
Chris@82 45 for (m = mb, W = W + (mb * 38); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 38, MAKE_VOLATILE_STRIDE(40, rs)) {
Chris@82 46 E T8, T4N, T2i, T4r, Tl, T4O, T2n, T4n, TN, T2b, T40, T4b, T2v, T3v, T3i;
Chris@82 47 E T3F, T27, T2f, T3W, T4f, T2R, T3z, T3a, T3J, T1G, T2e, T3T, T4e, T2K, T3y;
Chris@82 48 E T33, T3I, T1e, T2c, T43, T4c, T2C, T3w, T3p, T3G;
Chris@82 49 {
Chris@82 50 E T1, T4q, T3, T6, T4, T4o, T2, T7, T4p, T5;
Chris@82 51 T1 = ri[0];
Chris@82 52 T4q = ii[0];
Chris@82 53 T3 = ri[WS(rs, 10)];
Chris@82 54 T6 = ii[WS(rs, 10)];
Chris@82 55 T2 = W[18];
Chris@82 56 T4 = T2 * T3;
Chris@82 57 T4o = T2 * T6;
Chris@82 58 T5 = W[19];
Chris@82 59 T7 = FMA(T5, T6, T4);
Chris@82 60 T4p = FNMS(T5, T3, T4o);
Chris@82 61 T8 = T1 + T7;
Chris@82 62 T4N = T4q - T4p;
Chris@82 63 T2i = T1 - T7;
Chris@82 64 T4r = T4p + T4q;
Chris@82 65 }
Chris@82 66 {
Chris@82 67 E Ta, Td, Tb, T2j, Tg, Tj, Th, T2l, T9, Tf;
Chris@82 68 Ta = ri[WS(rs, 5)];
Chris@82 69 Td = ii[WS(rs, 5)];
Chris@82 70 T9 = W[8];
Chris@82 71 Tb = T9 * Ta;
Chris@82 72 T2j = T9 * Td;
Chris@82 73 Tg = ri[WS(rs, 15)];
Chris@82 74 Tj = ii[WS(rs, 15)];
Chris@82 75 Tf = W[28];
Chris@82 76 Th = Tf * Tg;
Chris@82 77 T2l = Tf * Tj;
Chris@82 78 {
Chris@82 79 E Te, T2k, Tk, T2m, Tc, Ti;
Chris@82 80 Tc = W[9];
Chris@82 81 Te = FMA(Tc, Td, Tb);
Chris@82 82 T2k = FNMS(Tc, Ta, T2j);
Chris@82 83 Ti = W[29];
Chris@82 84 Tk = FMA(Ti, Tj, Th);
Chris@82 85 T2m = FNMS(Ti, Tg, T2l);
Chris@82 86 Tl = Te + Tk;
Chris@82 87 T4O = Te - Tk;
Chris@82 88 T2n = T2k - T2m;
Chris@82 89 T4n = T2k + T2m;
Chris@82 90 }
Chris@82 91 }
Chris@82 92 {
Chris@82 93 E Ts, T3d, TL, T2t, Ty, T3f, TF, T2r;
Chris@82 94 {
Chris@82 95 E To, Tr, Tp, T3c, Tn, Tq;
Chris@82 96 To = ri[WS(rs, 4)];
Chris@82 97 Tr = ii[WS(rs, 4)];
Chris@82 98 Tn = W[6];
Chris@82 99 Tp = Tn * To;
Chris@82 100 T3c = Tn * Tr;
Chris@82 101 Tq = W[7];
Chris@82 102 Ts = FMA(Tq, Tr, Tp);
Chris@82 103 T3d = FNMS(Tq, To, T3c);
Chris@82 104 }
Chris@82 105 {
Chris@82 106 E TH, TK, TI, T2s, TG, TJ;
Chris@82 107 TH = ri[WS(rs, 19)];
Chris@82 108 TK = ii[WS(rs, 19)];
Chris@82 109 TG = W[36];
Chris@82 110 TI = TG * TH;
Chris@82 111 T2s = TG * TK;
Chris@82 112 TJ = W[37];
Chris@82 113 TL = FMA(TJ, TK, TI);
Chris@82 114 T2t = FNMS(TJ, TH, T2s);
Chris@82 115 }
Chris@82 116 {
Chris@82 117 E Tu, Tx, Tv, T3e, Tt, Tw;
Chris@82 118 Tu = ri[WS(rs, 14)];
Chris@82 119 Tx = ii[WS(rs, 14)];
Chris@82 120 Tt = W[26];
Chris@82 121 Tv = Tt * Tu;
Chris@82 122 T3e = Tt * Tx;
Chris@82 123 Tw = W[27];
Chris@82 124 Ty = FMA(Tw, Tx, Tv);
Chris@82 125 T3f = FNMS(Tw, Tu, T3e);
Chris@82 126 }
Chris@82 127 {
Chris@82 128 E TB, TE, TC, T2q, TA, TD;
Chris@82 129 TB = ri[WS(rs, 9)];
Chris@82 130 TE = ii[WS(rs, 9)];
Chris@82 131 TA = W[16];
Chris@82 132 TC = TA * TB;
Chris@82 133 T2q = TA * TE;
Chris@82 134 TD = W[17];
Chris@82 135 TF = FMA(TD, TE, TC);
Chris@82 136 T2r = FNMS(TD, TB, T2q);
Chris@82 137 }
Chris@82 138 {
Chris@82 139 E Tz, TM, T3Y, T3Z;
Chris@82 140 Tz = Ts + Ty;
Chris@82 141 TM = TF + TL;
Chris@82 142 TN = Tz - TM;
Chris@82 143 T2b = Tz + TM;
Chris@82 144 T3Y = T3d + T3f;
Chris@82 145 T3Z = T2r + T2t;
Chris@82 146 T40 = T3Y - T3Z;
Chris@82 147 T4b = T3Y + T3Z;
Chris@82 148 }
Chris@82 149 {
Chris@82 150 E T2p, T2u, T3g, T3h;
Chris@82 151 T2p = Ts - Ty;
Chris@82 152 T2u = T2r - T2t;
Chris@82 153 T2v = T2p - T2u;
Chris@82 154 T3v = T2p + T2u;
Chris@82 155 T3g = T3d - T3f;
Chris@82 156 T3h = TF - TL;
Chris@82 157 T3i = T3g + T3h;
Chris@82 158 T3F = T3g - T3h;
Chris@82 159 }
Chris@82 160 }
Chris@82 161 {
Chris@82 162 E T1M, T35, T25, T2P, T1S, T37, T1Z, T2N;
Chris@82 163 {
Chris@82 164 E T1I, T1L, T1J, T34, T1H, T1K;
Chris@82 165 T1I = ri[WS(rs, 12)];
Chris@82 166 T1L = ii[WS(rs, 12)];
Chris@82 167 T1H = W[22];
Chris@82 168 T1J = T1H * T1I;
Chris@82 169 T34 = T1H * T1L;
Chris@82 170 T1K = W[23];
Chris@82 171 T1M = FMA(T1K, T1L, T1J);
Chris@82 172 T35 = FNMS(T1K, T1I, T34);
Chris@82 173 }
Chris@82 174 {
Chris@82 175 E T21, T24, T22, T2O, T20, T23;
Chris@82 176 T21 = ri[WS(rs, 7)];
Chris@82 177 T24 = ii[WS(rs, 7)];
Chris@82 178 T20 = W[12];
Chris@82 179 T22 = T20 * T21;
Chris@82 180 T2O = T20 * T24;
Chris@82 181 T23 = W[13];
Chris@82 182 T25 = FMA(T23, T24, T22);
Chris@82 183 T2P = FNMS(T23, T21, T2O);
Chris@82 184 }
Chris@82 185 {
Chris@82 186 E T1O, T1R, T1P, T36, T1N, T1Q;
Chris@82 187 T1O = ri[WS(rs, 2)];
Chris@82 188 T1R = ii[WS(rs, 2)];
Chris@82 189 T1N = W[2];
Chris@82 190 T1P = T1N * T1O;
Chris@82 191 T36 = T1N * T1R;
Chris@82 192 T1Q = W[3];
Chris@82 193 T1S = FMA(T1Q, T1R, T1P);
Chris@82 194 T37 = FNMS(T1Q, T1O, T36);
Chris@82 195 }
Chris@82 196 {
Chris@82 197 E T1V, T1Y, T1W, T2M, T1U, T1X;
Chris@82 198 T1V = ri[WS(rs, 17)];
Chris@82 199 T1Y = ii[WS(rs, 17)];
Chris@82 200 T1U = W[32];
Chris@82 201 T1W = T1U * T1V;
Chris@82 202 T2M = T1U * T1Y;
Chris@82 203 T1X = W[33];
Chris@82 204 T1Z = FMA(T1X, T1Y, T1W);
Chris@82 205 T2N = FNMS(T1X, T1V, T2M);
Chris@82 206 }
Chris@82 207 {
Chris@82 208 E T1T, T26, T3U, T3V;
Chris@82 209 T1T = T1M + T1S;
Chris@82 210 T26 = T1Z + T25;
Chris@82 211 T27 = T1T - T26;
Chris@82 212 T2f = T1T + T26;
Chris@82 213 T3U = T35 + T37;
Chris@82 214 T3V = T2N + T2P;
Chris@82 215 T3W = T3U - T3V;
Chris@82 216 T4f = T3U + T3V;
Chris@82 217 }
Chris@82 218 {
Chris@82 219 E T2L, T2Q, T38, T39;
Chris@82 220 T2L = T1M - T1S;
Chris@82 221 T2Q = T2N - T2P;
Chris@82 222 T2R = T2L - T2Q;
Chris@82 223 T3z = T2L + T2Q;
Chris@82 224 T38 = T35 - T37;
Chris@82 225 T39 = T1Z - T25;
Chris@82 226 T3a = T38 + T39;
Chris@82 227 T3J = T38 - T39;
Chris@82 228 }
Chris@82 229 }
Chris@82 230 {
Chris@82 231 E T1l, T2Y, T1E, T2I, T1r, T30, T1y, T2G;
Chris@82 232 {
Chris@82 233 E T1h, T1k, T1i, T2X, T1g, T1j;
Chris@82 234 T1h = ri[WS(rs, 8)];
Chris@82 235 T1k = ii[WS(rs, 8)];
Chris@82 236 T1g = W[14];
Chris@82 237 T1i = T1g * T1h;
Chris@82 238 T2X = T1g * T1k;
Chris@82 239 T1j = W[15];
Chris@82 240 T1l = FMA(T1j, T1k, T1i);
Chris@82 241 T2Y = FNMS(T1j, T1h, T2X);
Chris@82 242 }
Chris@82 243 {
Chris@82 244 E T1A, T1D, T1B, T2H, T1z, T1C;
Chris@82 245 T1A = ri[WS(rs, 3)];
Chris@82 246 T1D = ii[WS(rs, 3)];
Chris@82 247 T1z = W[4];
Chris@82 248 T1B = T1z * T1A;
Chris@82 249 T2H = T1z * T1D;
Chris@82 250 T1C = W[5];
Chris@82 251 T1E = FMA(T1C, T1D, T1B);
Chris@82 252 T2I = FNMS(T1C, T1A, T2H);
Chris@82 253 }
Chris@82 254 {
Chris@82 255 E T1n, T1q, T1o, T2Z, T1m, T1p;
Chris@82 256 T1n = ri[WS(rs, 18)];
Chris@82 257 T1q = ii[WS(rs, 18)];
Chris@82 258 T1m = W[34];
Chris@82 259 T1o = T1m * T1n;
Chris@82 260 T2Z = T1m * T1q;
Chris@82 261 T1p = W[35];
Chris@82 262 T1r = FMA(T1p, T1q, T1o);
Chris@82 263 T30 = FNMS(T1p, T1n, T2Z);
Chris@82 264 }
Chris@82 265 {
Chris@82 266 E T1u, T1x, T1v, T2F, T1t, T1w;
Chris@82 267 T1u = ri[WS(rs, 13)];
Chris@82 268 T1x = ii[WS(rs, 13)];
Chris@82 269 T1t = W[24];
Chris@82 270 T1v = T1t * T1u;
Chris@82 271 T2F = T1t * T1x;
Chris@82 272 T1w = W[25];
Chris@82 273 T1y = FMA(T1w, T1x, T1v);
Chris@82 274 T2G = FNMS(T1w, T1u, T2F);
Chris@82 275 }
Chris@82 276 {
Chris@82 277 E T1s, T1F, T3R, T3S;
Chris@82 278 T1s = T1l + T1r;
Chris@82 279 T1F = T1y + T1E;
Chris@82 280 T1G = T1s - T1F;
Chris@82 281 T2e = T1s + T1F;
Chris@82 282 T3R = T2Y + T30;
Chris@82 283 T3S = T2G + T2I;
Chris@82 284 T3T = T3R - T3S;
Chris@82 285 T4e = T3R + T3S;
Chris@82 286 }
Chris@82 287 {
Chris@82 288 E T2E, T2J, T31, T32;
Chris@82 289 T2E = T1l - T1r;
Chris@82 290 T2J = T2G - T2I;
Chris@82 291 T2K = T2E - T2J;
Chris@82 292 T3y = T2E + T2J;
Chris@82 293 T31 = T2Y - T30;
Chris@82 294 T32 = T1y - T1E;
Chris@82 295 T33 = T31 + T32;
Chris@82 296 T3I = T31 - T32;
Chris@82 297 }
Chris@82 298 }
Chris@82 299 {
Chris@82 300 E TT, T3k, T1c, T2A, TZ, T3m, T16, T2y;
Chris@82 301 {
Chris@82 302 E TP, TS, TQ, T3j, TO, TR;
Chris@82 303 TP = ri[WS(rs, 16)];
Chris@82 304 TS = ii[WS(rs, 16)];
Chris@82 305 TO = W[30];
Chris@82 306 TQ = TO * TP;
Chris@82 307 T3j = TO * TS;
Chris@82 308 TR = W[31];
Chris@82 309 TT = FMA(TR, TS, TQ);
Chris@82 310 T3k = FNMS(TR, TP, T3j);
Chris@82 311 }
Chris@82 312 {
Chris@82 313 E T18, T1b, T19, T2z, T17, T1a;
Chris@82 314 T18 = ri[WS(rs, 11)];
Chris@82 315 T1b = ii[WS(rs, 11)];
Chris@82 316 T17 = W[20];
Chris@82 317 T19 = T17 * T18;
Chris@82 318 T2z = T17 * T1b;
Chris@82 319 T1a = W[21];
Chris@82 320 T1c = FMA(T1a, T1b, T19);
Chris@82 321 T2A = FNMS(T1a, T18, T2z);
Chris@82 322 }
Chris@82 323 {
Chris@82 324 E TV, TY, TW, T3l, TU, TX;
Chris@82 325 TV = ri[WS(rs, 6)];
Chris@82 326 TY = ii[WS(rs, 6)];
Chris@82 327 TU = W[10];
Chris@82 328 TW = TU * TV;
Chris@82 329 T3l = TU * TY;
Chris@82 330 TX = W[11];
Chris@82 331 TZ = FMA(TX, TY, TW);
Chris@82 332 T3m = FNMS(TX, TV, T3l);
Chris@82 333 }
Chris@82 334 {
Chris@82 335 E T12, T15, T13, T2x, T11, T14;
Chris@82 336 T12 = ri[WS(rs, 1)];
Chris@82 337 T15 = ii[WS(rs, 1)];
Chris@82 338 T11 = W[0];
Chris@82 339 T13 = T11 * T12;
Chris@82 340 T2x = T11 * T15;
Chris@82 341 T14 = W[1];
Chris@82 342 T16 = FMA(T14, T15, T13);
Chris@82 343 T2y = FNMS(T14, T12, T2x);
Chris@82 344 }
Chris@82 345 {
Chris@82 346 E T10, T1d, T41, T42;
Chris@82 347 T10 = TT + TZ;
Chris@82 348 T1d = T16 + T1c;
Chris@82 349 T1e = T10 - T1d;
Chris@82 350 T2c = T10 + T1d;
Chris@82 351 T41 = T3k + T3m;
Chris@82 352 T42 = T2y + T2A;
Chris@82 353 T43 = T41 - T42;
Chris@82 354 T4c = T41 + T42;
Chris@82 355 }
Chris@82 356 {
Chris@82 357 E T2w, T2B, T3n, T3o;
Chris@82 358 T2w = TT - TZ;
Chris@82 359 T2B = T2y - T2A;
Chris@82 360 T2C = T2w - T2B;
Chris@82 361 T3w = T2w + T2B;
Chris@82 362 T3n = T3k - T3m;
Chris@82 363 T3o = T16 - T1c;
Chris@82 364 T3p = T3n + T3o;
Chris@82 365 T3G = T3n - T3o;
Chris@82 366 }
Chris@82 367 }
Chris@82 368 {
Chris@82 369 E T45, T47, Tm, T29, T3O, T3P, T46, T3Q;
Chris@82 370 {
Chris@82 371 E T3X, T44, T1f, T28;
Chris@82 372 T3X = T3T - T3W;
Chris@82 373 T44 = T40 - T43;
Chris@82 374 T45 = FNMS(KP618033988, T44, T3X);
Chris@82 375 T47 = FMA(KP618033988, T3X, T44);
Chris@82 376 Tm = T8 - Tl;
Chris@82 377 T1f = TN + T1e;
Chris@82 378 T28 = T1G + T27;
Chris@82 379 T29 = T1f + T28;
Chris@82 380 T3O = FNMS(KP250000000, T29, Tm);
Chris@82 381 T3P = T1f - T28;
Chris@82 382 }
Chris@82 383 ri[WS(rs, 10)] = Tm + T29;
Chris@82 384 T46 = FMA(KP559016994, T3P, T3O);
Chris@82 385 ri[WS(rs, 14)] = FNMS(KP951056516, T47, T46);
Chris@82 386 ri[WS(rs, 6)] = FMA(KP951056516, T47, T46);
Chris@82 387 T3Q = FNMS(KP559016994, T3P, T3O);
Chris@82 388 ri[WS(rs, 2)] = FNMS(KP951056516, T45, T3Q);
Chris@82 389 ri[WS(rs, 18)] = FMA(KP951056516, T45, T3Q);
Chris@82 390 }
Chris@82 391 {
Chris@82 392 E T4K, T4M, T4B, T4E, T4F, T4G, T4L, T4H;
Chris@82 393 {
Chris@82 394 E T4I, T4J, T4C, T4D;
Chris@82 395 T4I = T1G - T27;
Chris@82 396 T4J = TN - T1e;
Chris@82 397 T4K = FNMS(KP618033988, T4J, T4I);
Chris@82 398 T4M = FMA(KP618033988, T4I, T4J);
Chris@82 399 T4B = T4r - T4n;
Chris@82 400 T4C = T40 + T43;
Chris@82 401 T4D = T3T + T3W;
Chris@82 402 T4E = T4C + T4D;
Chris@82 403 T4F = FNMS(KP250000000, T4E, T4B);
Chris@82 404 T4G = T4C - T4D;
Chris@82 405 }
Chris@82 406 ii[WS(rs, 10)] = T4E + T4B;
Chris@82 407 T4L = FMA(KP559016994, T4G, T4F);
Chris@82 408 ii[WS(rs, 6)] = FNMS(KP951056516, T4M, T4L);
Chris@82 409 ii[WS(rs, 14)] = FMA(KP951056516, T4M, T4L);
Chris@82 410 T4H = FNMS(KP559016994, T4G, T4F);
Chris@82 411 ii[WS(rs, 2)] = FMA(KP951056516, T4K, T4H);
Chris@82 412 ii[WS(rs, 18)] = FNMS(KP951056516, T4K, T4H);
Chris@82 413 }
Chris@82 414 {
Chris@82 415 E T4h, T4j, T2a, T2h, T48, T49, T4i, T4a;
Chris@82 416 {
Chris@82 417 E T4d, T4g, T2d, T2g;
Chris@82 418 T4d = T4b - T4c;
Chris@82 419 T4g = T4e - T4f;
Chris@82 420 T4h = FMA(KP618033988, T4g, T4d);
Chris@82 421 T4j = FNMS(KP618033988, T4d, T4g);
Chris@82 422 T2a = T8 + Tl;
Chris@82 423 T2d = T2b + T2c;
Chris@82 424 T2g = T2e + T2f;
Chris@82 425 T2h = T2d + T2g;
Chris@82 426 T48 = FNMS(KP250000000, T2h, T2a);
Chris@82 427 T49 = T2d - T2g;
Chris@82 428 }
Chris@82 429 ri[0] = T2a + T2h;
Chris@82 430 T4i = FNMS(KP559016994, T49, T48);
Chris@82 431 ri[WS(rs, 12)] = FNMS(KP951056516, T4j, T4i);
Chris@82 432 ri[WS(rs, 8)] = FMA(KP951056516, T4j, T4i);
Chris@82 433 T4a = FMA(KP559016994, T49, T48);
Chris@82 434 ri[WS(rs, 4)] = FNMS(KP951056516, T4h, T4a);
Chris@82 435 ri[WS(rs, 16)] = FMA(KP951056516, T4h, T4a);
Chris@82 436 }
Chris@82 437 {
Chris@82 438 E T4y, T4A, T4s, T4m, T4t, T4u, T4z, T4v;
Chris@82 439 {
Chris@82 440 E T4w, T4x, T4k, T4l;
Chris@82 441 T4w = T2b - T2c;
Chris@82 442 T4x = T2e - T2f;
Chris@82 443 T4y = FMA(KP618033988, T4x, T4w);
Chris@82 444 T4A = FNMS(KP618033988, T4w, T4x);
Chris@82 445 T4s = T4n + T4r;
Chris@82 446 T4k = T4b + T4c;
Chris@82 447 T4l = T4e + T4f;
Chris@82 448 T4m = T4k + T4l;
Chris@82 449 T4t = FNMS(KP250000000, T4m, T4s);
Chris@82 450 T4u = T4k - T4l;
Chris@82 451 }
Chris@82 452 ii[0] = T4m + T4s;
Chris@82 453 T4z = FNMS(KP559016994, T4u, T4t);
Chris@82 454 ii[WS(rs, 8)] = FNMS(KP951056516, T4A, T4z);
Chris@82 455 ii[WS(rs, 12)] = FMA(KP951056516, T4A, T4z);
Chris@82 456 T4v = FMA(KP559016994, T4u, T4t);
Chris@82 457 ii[WS(rs, 4)] = FMA(KP951056516, T4y, T4v);
Chris@82 458 ii[WS(rs, 16)] = FNMS(KP951056516, T4y, T4v);
Chris@82 459 }
Chris@82 460 {
Chris@82 461 E T3r, T3t, T2o, T2T, T2U, T2V, T3s, T2W;
Chris@82 462 {
Chris@82 463 E T3b, T3q, T2D, T2S;
Chris@82 464 T3b = T33 - T3a;
Chris@82 465 T3q = T3i - T3p;
Chris@82 466 T3r = FNMS(KP618033988, T3q, T3b);
Chris@82 467 T3t = FMA(KP618033988, T3b, T3q);
Chris@82 468 T2o = T2i - T2n;
Chris@82 469 T2D = T2v + T2C;
Chris@82 470 T2S = T2K + T2R;
Chris@82 471 T2T = T2D + T2S;
Chris@82 472 T2U = FNMS(KP250000000, T2T, T2o);
Chris@82 473 T2V = T2D - T2S;
Chris@82 474 }
Chris@82 475 ri[WS(rs, 15)] = T2o + T2T;
Chris@82 476 T3s = FMA(KP559016994, T2V, T2U);
Chris@82 477 ri[WS(rs, 11)] = FMA(KP951056516, T3t, T3s);
Chris@82 478 ri[WS(rs, 19)] = FNMS(KP951056516, T3t, T3s);
Chris@82 479 T2W = FNMS(KP559016994, T2V, T2U);
Chris@82 480 ri[WS(rs, 3)] = FMA(KP951056516, T3r, T2W);
Chris@82 481 ri[WS(rs, 7)] = FNMS(KP951056516, T3r, T2W);
Chris@82 482 }
Chris@82 483 {
Chris@82 484 E T5a, T5c, T51, T54, T55, T56, T5b, T57;
Chris@82 485 {
Chris@82 486 E T58, T59, T52, T53;
Chris@82 487 T58 = T2K - T2R;
Chris@82 488 T59 = T2v - T2C;
Chris@82 489 T5a = FNMS(KP618033988, T59, T58);
Chris@82 490 T5c = FMA(KP618033988, T58, T59);
Chris@82 491 T51 = T4O + T4N;
Chris@82 492 T52 = T3i + T3p;
Chris@82 493 T53 = T33 + T3a;
Chris@82 494 T54 = T52 + T53;
Chris@82 495 T55 = FNMS(KP250000000, T54, T51);
Chris@82 496 T56 = T52 - T53;
Chris@82 497 }
Chris@82 498 ii[WS(rs, 15)] = T54 + T51;
Chris@82 499 T5b = FMA(KP559016994, T56, T55);
Chris@82 500 ii[WS(rs, 11)] = FNMS(KP951056516, T5c, T5b);
Chris@82 501 ii[WS(rs, 19)] = FMA(KP951056516, T5c, T5b);
Chris@82 502 T57 = FNMS(KP559016994, T56, T55);
Chris@82 503 ii[WS(rs, 3)] = FNMS(KP951056516, T5a, T57);
Chris@82 504 ii[WS(rs, 7)] = FMA(KP951056516, T5a, T57);
Chris@82 505 }
Chris@82 506 {
Chris@82 507 E T3L, T3N, T3u, T3B, T3C, T3D, T3M, T3E;
Chris@82 508 {
Chris@82 509 E T3H, T3K, T3x, T3A;
Chris@82 510 T3H = T3F - T3G;
Chris@82 511 T3K = T3I - T3J;
Chris@82 512 T3L = FMA(KP618033988, T3K, T3H);
Chris@82 513 T3N = FNMS(KP618033988, T3H, T3K);
Chris@82 514 T3u = T2i + T2n;
Chris@82 515 T3x = T3v + T3w;
Chris@82 516 T3A = T3y + T3z;
Chris@82 517 T3B = T3x + T3A;
Chris@82 518 T3C = FNMS(KP250000000, T3B, T3u);
Chris@82 519 T3D = T3x - T3A;
Chris@82 520 }
Chris@82 521 ri[WS(rs, 5)] = T3u + T3B;
Chris@82 522 T3M = FNMS(KP559016994, T3D, T3C);
Chris@82 523 ri[WS(rs, 13)] = FMA(KP951056516, T3N, T3M);
Chris@82 524 ri[WS(rs, 17)] = FNMS(KP951056516, T3N, T3M);
Chris@82 525 T3E = FMA(KP559016994, T3D, T3C);
Chris@82 526 ri[WS(rs, 1)] = FMA(KP951056516, T3L, T3E);
Chris@82 527 ri[WS(rs, 9)] = FNMS(KP951056516, T3L, T3E);
Chris@82 528 }
Chris@82 529 {
Chris@82 530 E T4Y, T50, T4P, T4S, T4T, T4U, T4Z, T4V;
Chris@82 531 {
Chris@82 532 E T4W, T4X, T4Q, T4R;
Chris@82 533 T4W = T3v - T3w;
Chris@82 534 T4X = T3y - T3z;
Chris@82 535 T4Y = FMA(KP618033988, T4X, T4W);
Chris@82 536 T50 = FNMS(KP618033988, T4W, T4X);
Chris@82 537 T4P = T4N - T4O;
Chris@82 538 T4Q = T3F + T3G;
Chris@82 539 T4R = T3I + T3J;
Chris@82 540 T4S = T4Q + T4R;
Chris@82 541 T4T = FNMS(KP250000000, T4S, T4P);
Chris@82 542 T4U = T4Q - T4R;
Chris@82 543 }
Chris@82 544 ii[WS(rs, 5)] = T4S + T4P;
Chris@82 545 T4Z = FNMS(KP559016994, T4U, T4T);
Chris@82 546 ii[WS(rs, 13)] = FNMS(KP951056516, T50, T4Z);
Chris@82 547 ii[WS(rs, 17)] = FMA(KP951056516, T50, T4Z);
Chris@82 548 T4V = FMA(KP559016994, T4U, T4T);
Chris@82 549 ii[WS(rs, 1)] = FNMS(KP951056516, T4Y, T4V);
Chris@82 550 ii[WS(rs, 9)] = FMA(KP951056516, T4Y, T4V);
Chris@82 551 }
Chris@82 552 }
Chris@82 553 }
Chris@82 554 }
Chris@82 555
Chris@82 556 static const tw_instr twinstr[] = {
Chris@82 557 {TW_FULL, 0, 20},
Chris@82 558 {TW_NEXT, 1, 0}
Chris@82 559 };
Chris@82 560
Chris@82 561 static const ct_desc desc = { 20, "t1_20", twinstr, &GENUS, {136, 38, 110, 0}, 0, 0, 0 };
Chris@82 562
Chris@82 563 void X(codelet_t1_20) (planner *p) {
Chris@82 564 X(kdft_dit_register) (p, t1_20, &desc);
Chris@82 565 }
Chris@82 566 #else
Chris@82 567
Chris@82 568 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -n 20 -name t1_20 -include dft/scalar/t.h */
Chris@82 569
Chris@82 570 /*
Chris@82 571 * This function contains 246 FP additions, 124 FP multiplications,
Chris@82 572 * (or, 184 additions, 62 multiplications, 62 fused multiply/add),
Chris@82 573 * 85 stack variables, 4 constants, and 80 memory accesses
Chris@82 574 */
Chris@82 575 #include "dft/scalar/t.h"
Chris@82 576
Chris@82 577 static void t1_20(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@82 578 {
Chris@82 579 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
Chris@82 580 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@82 581 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@82 582 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@82 583 {
Chris@82 584 INT m;
Chris@82 585 for (m = mb, W = W + (mb * 38); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 38, MAKE_VOLATILE_STRIDE(40, rs)) {
Chris@82 586 E Tj, T1R, T4g, T4p, T2q, T37, T3Q, T42, T1r, T1O, T1P, T3i, T3l, T44, T3D;
Chris@82 587 E T3E, T3K, T1V, T1W, T1X, T23, T28, T4r, T2W, T2X, T4c, T33, T34, T35, T2G;
Chris@82 588 E T2L, T2M, TG, T13, T14, T3p, T3s, T43, T3A, T3B, T3J, T1S, T1T, T1U, T2e;
Chris@82 589 E T2j, T4q, T2T, T2U, T4b, T30, T31, T32, T2v, T2A, T2B;
Chris@82 590 {
Chris@82 591 E T1, T3O, T6, T3N, Tc, T2n, Th, T2o;
Chris@82 592 T1 = ri[0];
Chris@82 593 T3O = ii[0];
Chris@82 594 {
Chris@82 595 E T3, T5, T2, T4;
Chris@82 596 T3 = ri[WS(rs, 10)];
Chris@82 597 T5 = ii[WS(rs, 10)];
Chris@82 598 T2 = W[18];
Chris@82 599 T4 = W[19];
Chris@82 600 T6 = FMA(T2, T3, T4 * T5);
Chris@82 601 T3N = FNMS(T4, T3, T2 * T5);
Chris@82 602 }
Chris@82 603 {
Chris@82 604 E T9, Tb, T8, Ta;
Chris@82 605 T9 = ri[WS(rs, 5)];
Chris@82 606 Tb = ii[WS(rs, 5)];
Chris@82 607 T8 = W[8];
Chris@82 608 Ta = W[9];
Chris@82 609 Tc = FMA(T8, T9, Ta * Tb);
Chris@82 610 T2n = FNMS(Ta, T9, T8 * Tb);
Chris@82 611 }
Chris@82 612 {
Chris@82 613 E Te, Tg, Td, Tf;
Chris@82 614 Te = ri[WS(rs, 15)];
Chris@82 615 Tg = ii[WS(rs, 15)];
Chris@82 616 Td = W[28];
Chris@82 617 Tf = W[29];
Chris@82 618 Th = FMA(Td, Te, Tf * Tg);
Chris@82 619 T2o = FNMS(Tf, Te, Td * Tg);
Chris@82 620 }
Chris@82 621 {
Chris@82 622 E T7, Ti, T4e, T4f;
Chris@82 623 T7 = T1 + T6;
Chris@82 624 Ti = Tc + Th;
Chris@82 625 Tj = T7 - Ti;
Chris@82 626 T1R = T7 + Ti;
Chris@82 627 T4e = T3O - T3N;
Chris@82 628 T4f = Tc - Th;
Chris@82 629 T4g = T4e - T4f;
Chris@82 630 T4p = T4f + T4e;
Chris@82 631 }
Chris@82 632 {
Chris@82 633 E T2m, T2p, T3M, T3P;
Chris@82 634 T2m = T1 - T6;
Chris@82 635 T2p = T2n - T2o;
Chris@82 636 T2q = T2m - T2p;
Chris@82 637 T37 = T2m + T2p;
Chris@82 638 T3M = T2n + T2o;
Chris@82 639 T3P = T3N + T3O;
Chris@82 640 T3Q = T3M + T3P;
Chris@82 641 T42 = T3P - T3M;
Chris@82 642 }
Chris@82 643 }
Chris@82 644 {
Chris@82 645 E T1f, T3g, T21, T2C, T1N, T3k, T27, T2K, T1q, T3h, T22, T2F, T1C, T3j, T26;
Chris@82 646 E T2H;
Chris@82 647 {
Chris@82 648 E T19, T1Z, T1e, T20;
Chris@82 649 {
Chris@82 650 E T16, T18, T15, T17;
Chris@82 651 T16 = ri[WS(rs, 8)];
Chris@82 652 T18 = ii[WS(rs, 8)];
Chris@82 653 T15 = W[14];
Chris@82 654 T17 = W[15];
Chris@82 655 T19 = FMA(T15, T16, T17 * T18);
Chris@82 656 T1Z = FNMS(T17, T16, T15 * T18);
Chris@82 657 }
Chris@82 658 {
Chris@82 659 E T1b, T1d, T1a, T1c;
Chris@82 660 T1b = ri[WS(rs, 18)];
Chris@82 661 T1d = ii[WS(rs, 18)];
Chris@82 662 T1a = W[34];
Chris@82 663 T1c = W[35];
Chris@82 664 T1e = FMA(T1a, T1b, T1c * T1d);
Chris@82 665 T20 = FNMS(T1c, T1b, T1a * T1d);
Chris@82 666 }
Chris@82 667 T1f = T19 + T1e;
Chris@82 668 T3g = T1Z + T20;
Chris@82 669 T21 = T1Z - T20;
Chris@82 670 T2C = T19 - T1e;
Chris@82 671 }
Chris@82 672 {
Chris@82 673 E T1H, T2I, T1M, T2J;
Chris@82 674 {
Chris@82 675 E T1E, T1G, T1D, T1F;
Chris@82 676 T1E = ri[WS(rs, 17)];
Chris@82 677 T1G = ii[WS(rs, 17)];
Chris@82 678 T1D = W[32];
Chris@82 679 T1F = W[33];
Chris@82 680 T1H = FMA(T1D, T1E, T1F * T1G);
Chris@82 681 T2I = FNMS(T1F, T1E, T1D * T1G);
Chris@82 682 }
Chris@82 683 {
Chris@82 684 E T1J, T1L, T1I, T1K;
Chris@82 685 T1J = ri[WS(rs, 7)];
Chris@82 686 T1L = ii[WS(rs, 7)];
Chris@82 687 T1I = W[12];
Chris@82 688 T1K = W[13];
Chris@82 689 T1M = FMA(T1I, T1J, T1K * T1L);
Chris@82 690 T2J = FNMS(T1K, T1J, T1I * T1L);
Chris@82 691 }
Chris@82 692 T1N = T1H + T1M;
Chris@82 693 T3k = T2I + T2J;
Chris@82 694 T27 = T1H - T1M;
Chris@82 695 T2K = T2I - T2J;
Chris@82 696 }
Chris@82 697 {
Chris@82 698 E T1k, T2D, T1p, T2E;
Chris@82 699 {
Chris@82 700 E T1h, T1j, T1g, T1i;
Chris@82 701 T1h = ri[WS(rs, 13)];
Chris@82 702 T1j = ii[WS(rs, 13)];
Chris@82 703 T1g = W[24];
Chris@82 704 T1i = W[25];
Chris@82 705 T1k = FMA(T1g, T1h, T1i * T1j);
Chris@82 706 T2D = FNMS(T1i, T1h, T1g * T1j);
Chris@82 707 }
Chris@82 708 {
Chris@82 709 E T1m, T1o, T1l, T1n;
Chris@82 710 T1m = ri[WS(rs, 3)];
Chris@82 711 T1o = ii[WS(rs, 3)];
Chris@82 712 T1l = W[4];
Chris@82 713 T1n = W[5];
Chris@82 714 T1p = FMA(T1l, T1m, T1n * T1o);
Chris@82 715 T2E = FNMS(T1n, T1m, T1l * T1o);
Chris@82 716 }
Chris@82 717 T1q = T1k + T1p;
Chris@82 718 T3h = T2D + T2E;
Chris@82 719 T22 = T1k - T1p;
Chris@82 720 T2F = T2D - T2E;
Chris@82 721 }
Chris@82 722 {
Chris@82 723 E T1w, T24, T1B, T25;
Chris@82 724 {
Chris@82 725 E T1t, T1v, T1s, T1u;
Chris@82 726 T1t = ri[WS(rs, 12)];
Chris@82 727 T1v = ii[WS(rs, 12)];
Chris@82 728 T1s = W[22];
Chris@82 729 T1u = W[23];
Chris@82 730 T1w = FMA(T1s, T1t, T1u * T1v);
Chris@82 731 T24 = FNMS(T1u, T1t, T1s * T1v);
Chris@82 732 }
Chris@82 733 {
Chris@82 734 E T1y, T1A, T1x, T1z;
Chris@82 735 T1y = ri[WS(rs, 2)];
Chris@82 736 T1A = ii[WS(rs, 2)];
Chris@82 737 T1x = W[2];
Chris@82 738 T1z = W[3];
Chris@82 739 T1B = FMA(T1x, T1y, T1z * T1A);
Chris@82 740 T25 = FNMS(T1z, T1y, T1x * T1A);
Chris@82 741 }
Chris@82 742 T1C = T1w + T1B;
Chris@82 743 T3j = T24 + T25;
Chris@82 744 T26 = T24 - T25;
Chris@82 745 T2H = T1w - T1B;
Chris@82 746 }
Chris@82 747 T1r = T1f - T1q;
Chris@82 748 T1O = T1C - T1N;
Chris@82 749 T1P = T1r + T1O;
Chris@82 750 T3i = T3g - T3h;
Chris@82 751 T3l = T3j - T3k;
Chris@82 752 T44 = T3i + T3l;
Chris@82 753 T3D = T3g + T3h;
Chris@82 754 T3E = T3j + T3k;
Chris@82 755 T3K = T3D + T3E;
Chris@82 756 T1V = T1f + T1q;
Chris@82 757 T1W = T1C + T1N;
Chris@82 758 T1X = T1V + T1W;
Chris@82 759 T23 = T21 + T22;
Chris@82 760 T28 = T26 + T27;
Chris@82 761 T4r = T23 + T28;
Chris@82 762 T2W = T21 - T22;
Chris@82 763 T2X = T26 - T27;
Chris@82 764 T4c = T2W + T2X;
Chris@82 765 T33 = T2C + T2F;
Chris@82 766 T34 = T2H + T2K;
Chris@82 767 T35 = T33 + T34;
Chris@82 768 T2G = T2C - T2F;
Chris@82 769 T2L = T2H - T2K;
Chris@82 770 T2M = T2G + T2L;
Chris@82 771 }
Chris@82 772 {
Chris@82 773 E Tu, T3n, T2c, T2r, T12, T3r, T2i, T2z, TF, T3o, T2d, T2u, TR, T3q, T2h;
Chris@82 774 E T2w;
Chris@82 775 {
Chris@82 776 E To, T2a, Tt, T2b;
Chris@82 777 {
Chris@82 778 E Tl, Tn, Tk, Tm;
Chris@82 779 Tl = ri[WS(rs, 4)];
Chris@82 780 Tn = ii[WS(rs, 4)];
Chris@82 781 Tk = W[6];
Chris@82 782 Tm = W[7];
Chris@82 783 To = FMA(Tk, Tl, Tm * Tn);
Chris@82 784 T2a = FNMS(Tm, Tl, Tk * Tn);
Chris@82 785 }
Chris@82 786 {
Chris@82 787 E Tq, Ts, Tp, Tr;
Chris@82 788 Tq = ri[WS(rs, 14)];
Chris@82 789 Ts = ii[WS(rs, 14)];
Chris@82 790 Tp = W[26];
Chris@82 791 Tr = W[27];
Chris@82 792 Tt = FMA(Tp, Tq, Tr * Ts);
Chris@82 793 T2b = FNMS(Tr, Tq, Tp * Ts);
Chris@82 794 }
Chris@82 795 Tu = To + Tt;
Chris@82 796 T3n = T2a + T2b;
Chris@82 797 T2c = T2a - T2b;
Chris@82 798 T2r = To - Tt;
Chris@82 799 }
Chris@82 800 {
Chris@82 801 E TW, T2x, T11, T2y;
Chris@82 802 {
Chris@82 803 E TT, TV, TS, TU;
Chris@82 804 TT = ri[WS(rs, 1)];
Chris@82 805 TV = ii[WS(rs, 1)];
Chris@82 806 TS = W[0];
Chris@82 807 TU = W[1];
Chris@82 808 TW = FMA(TS, TT, TU * TV);
Chris@82 809 T2x = FNMS(TU, TT, TS * TV);
Chris@82 810 }
Chris@82 811 {
Chris@82 812 E TY, T10, TX, TZ;
Chris@82 813 TY = ri[WS(rs, 11)];
Chris@82 814 T10 = ii[WS(rs, 11)];
Chris@82 815 TX = W[20];
Chris@82 816 TZ = W[21];
Chris@82 817 T11 = FMA(TX, TY, TZ * T10);
Chris@82 818 T2y = FNMS(TZ, TY, TX * T10);
Chris@82 819 }
Chris@82 820 T12 = TW + T11;
Chris@82 821 T3r = T2x + T2y;
Chris@82 822 T2i = TW - T11;
Chris@82 823 T2z = T2x - T2y;
Chris@82 824 }
Chris@82 825 {
Chris@82 826 E Tz, T2s, TE, T2t;
Chris@82 827 {
Chris@82 828 E Tw, Ty, Tv, Tx;
Chris@82 829 Tw = ri[WS(rs, 9)];
Chris@82 830 Ty = ii[WS(rs, 9)];
Chris@82 831 Tv = W[16];
Chris@82 832 Tx = W[17];
Chris@82 833 Tz = FMA(Tv, Tw, Tx * Ty);
Chris@82 834 T2s = FNMS(Tx, Tw, Tv * Ty);
Chris@82 835 }
Chris@82 836 {
Chris@82 837 E TB, TD, TA, TC;
Chris@82 838 TB = ri[WS(rs, 19)];
Chris@82 839 TD = ii[WS(rs, 19)];
Chris@82 840 TA = W[36];
Chris@82 841 TC = W[37];
Chris@82 842 TE = FMA(TA, TB, TC * TD);
Chris@82 843 T2t = FNMS(TC, TB, TA * TD);
Chris@82 844 }
Chris@82 845 TF = Tz + TE;
Chris@82 846 T3o = T2s + T2t;
Chris@82 847 T2d = Tz - TE;
Chris@82 848 T2u = T2s - T2t;
Chris@82 849 }
Chris@82 850 {
Chris@82 851 E TL, T2f, TQ, T2g;
Chris@82 852 {
Chris@82 853 E TI, TK, TH, TJ;
Chris@82 854 TI = ri[WS(rs, 16)];
Chris@82 855 TK = ii[WS(rs, 16)];
Chris@82 856 TH = W[30];
Chris@82 857 TJ = W[31];
Chris@82 858 TL = FMA(TH, TI, TJ * TK);
Chris@82 859 T2f = FNMS(TJ, TI, TH * TK);
Chris@82 860 }
Chris@82 861 {
Chris@82 862 E TN, TP, TM, TO;
Chris@82 863 TN = ri[WS(rs, 6)];
Chris@82 864 TP = ii[WS(rs, 6)];
Chris@82 865 TM = W[10];
Chris@82 866 TO = W[11];
Chris@82 867 TQ = FMA(TM, TN, TO * TP);
Chris@82 868 T2g = FNMS(TO, TN, TM * TP);
Chris@82 869 }
Chris@82 870 TR = TL + TQ;
Chris@82 871 T3q = T2f + T2g;
Chris@82 872 T2h = T2f - T2g;
Chris@82 873 T2w = TL - TQ;
Chris@82 874 }
Chris@82 875 TG = Tu - TF;
Chris@82 876 T13 = TR - T12;
Chris@82 877 T14 = TG + T13;
Chris@82 878 T3p = T3n - T3o;
Chris@82 879 T3s = T3q - T3r;
Chris@82 880 T43 = T3p + T3s;
Chris@82 881 T3A = T3n + T3o;
Chris@82 882 T3B = T3q + T3r;
Chris@82 883 T3J = T3A + T3B;
Chris@82 884 T1S = Tu + TF;
Chris@82 885 T1T = TR + T12;
Chris@82 886 T1U = T1S + T1T;
Chris@82 887 T2e = T2c + T2d;
Chris@82 888 T2j = T2h + T2i;
Chris@82 889 T4q = T2e + T2j;
Chris@82 890 T2T = T2c - T2d;
Chris@82 891 T2U = T2h - T2i;
Chris@82 892 T4b = T2T + T2U;
Chris@82 893 T30 = T2r + T2u;
Chris@82 894 T31 = T2w + T2z;
Chris@82 895 T32 = T30 + T31;
Chris@82 896 T2v = T2r - T2u;
Chris@82 897 T2A = T2w - T2z;
Chris@82 898 T2B = T2v + T2A;
Chris@82 899 }
Chris@82 900 {
Chris@82 901 E T3e, T1Q, T3d, T3u, T3w, T3m, T3t, T3v, T3f;
Chris@82 902 T3e = KP559016994 * (T14 - T1P);
Chris@82 903 T1Q = T14 + T1P;
Chris@82 904 T3d = FNMS(KP250000000, T1Q, Tj);
Chris@82 905 T3m = T3i - T3l;
Chris@82 906 T3t = T3p - T3s;
Chris@82 907 T3u = FNMS(KP587785252, T3t, KP951056516 * T3m);
Chris@82 908 T3w = FMA(KP951056516, T3t, KP587785252 * T3m);
Chris@82 909 ri[WS(rs, 10)] = Tj + T1Q;
Chris@82 910 T3v = T3e + T3d;
Chris@82 911 ri[WS(rs, 14)] = T3v - T3w;
Chris@82 912 ri[WS(rs, 6)] = T3v + T3w;
Chris@82 913 T3f = T3d - T3e;
Chris@82 914 ri[WS(rs, 2)] = T3f - T3u;
Chris@82 915 ri[WS(rs, 18)] = T3f + T3u;
Chris@82 916 }
Chris@82 917 {
Chris@82 918 E T47, T45, T46, T41, T4a, T3Z, T40, T49, T48;
Chris@82 919 T47 = KP559016994 * (T43 - T44);
Chris@82 920 T45 = T43 + T44;
Chris@82 921 T46 = FNMS(KP250000000, T45, T42);
Chris@82 922 T3Z = T1r - T1O;
Chris@82 923 T40 = TG - T13;
Chris@82 924 T41 = FNMS(KP587785252, T40, KP951056516 * T3Z);
Chris@82 925 T4a = FMA(KP951056516, T40, KP587785252 * T3Z);
Chris@82 926 ii[WS(rs, 10)] = T45 + T42;
Chris@82 927 T49 = T47 + T46;
Chris@82 928 ii[WS(rs, 6)] = T49 - T4a;
Chris@82 929 ii[WS(rs, 14)] = T4a + T49;
Chris@82 930 T48 = T46 - T47;
Chris@82 931 ii[WS(rs, 2)] = T41 + T48;
Chris@82 932 ii[WS(rs, 18)] = T48 - T41;
Chris@82 933 }
Chris@82 934 {
Chris@82 935 E T3x, T1Y, T3y, T3G, T3I, T3C, T3F, T3H, T3z;
Chris@82 936 T3x = KP559016994 * (T1U - T1X);
Chris@82 937 T1Y = T1U + T1X;
Chris@82 938 T3y = FNMS(KP250000000, T1Y, T1R);
Chris@82 939 T3C = T3A - T3B;
Chris@82 940 T3F = T3D - T3E;
Chris@82 941 T3G = FMA(KP951056516, T3C, KP587785252 * T3F);
Chris@82 942 T3I = FNMS(KP587785252, T3C, KP951056516 * T3F);
Chris@82 943 ri[0] = T1R + T1Y;
Chris@82 944 T3H = T3y - T3x;
Chris@82 945 ri[WS(rs, 12)] = T3H - T3I;
Chris@82 946 ri[WS(rs, 8)] = T3H + T3I;
Chris@82 947 T3z = T3x + T3y;
Chris@82 948 ri[WS(rs, 4)] = T3z - T3G;
Chris@82 949 ri[WS(rs, 16)] = T3z + T3G;
Chris@82 950 }
Chris@82 951 {
Chris@82 952 E T3U, T3L, T3V, T3T, T3Y, T3R, T3S, T3X, T3W;
Chris@82 953 T3U = KP559016994 * (T3J - T3K);
Chris@82 954 T3L = T3J + T3K;
Chris@82 955 T3V = FNMS(KP250000000, T3L, T3Q);
Chris@82 956 T3R = T1S - T1T;
Chris@82 957 T3S = T1V - T1W;
Chris@82 958 T3T = FMA(KP951056516, T3R, KP587785252 * T3S);
Chris@82 959 T3Y = FNMS(KP587785252, T3R, KP951056516 * T3S);
Chris@82 960 ii[0] = T3L + T3Q;
Chris@82 961 T3X = T3V - T3U;
Chris@82 962 ii[WS(rs, 8)] = T3X - T3Y;
Chris@82 963 ii[WS(rs, 12)] = T3Y + T3X;
Chris@82 964 T3W = T3U + T3V;
Chris@82 965 ii[WS(rs, 4)] = T3T + T3W;
Chris@82 966 ii[WS(rs, 16)] = T3W - T3T;
Chris@82 967 }
Chris@82 968 {
Chris@82 969 E T2P, T2N, T2O, T2l, T2R, T29, T2k, T2S, T2Q;
Chris@82 970 T2P = KP559016994 * (T2B - T2M);
Chris@82 971 T2N = T2B + T2M;
Chris@82 972 T2O = FNMS(KP250000000, T2N, T2q);
Chris@82 973 T29 = T23 - T28;
Chris@82 974 T2k = T2e - T2j;
Chris@82 975 T2l = FNMS(KP587785252, T2k, KP951056516 * T29);
Chris@82 976 T2R = FMA(KP951056516, T2k, KP587785252 * T29);
Chris@82 977 ri[WS(rs, 15)] = T2q + T2N;
Chris@82 978 T2S = T2P + T2O;
Chris@82 979 ri[WS(rs, 11)] = T2R + T2S;
Chris@82 980 ri[WS(rs, 19)] = T2S - T2R;
Chris@82 981 T2Q = T2O - T2P;
Chris@82 982 ri[WS(rs, 3)] = T2l + T2Q;
Chris@82 983 ri[WS(rs, 7)] = T2Q - T2l;
Chris@82 984 }
Chris@82 985 {
Chris@82 986 E T4u, T4s, T4t, T4y, T4A, T4w, T4x, T4z, T4v;
Chris@82 987 T4u = KP559016994 * (T4q - T4r);
Chris@82 988 T4s = T4q + T4r;
Chris@82 989 T4t = FNMS(KP250000000, T4s, T4p);
Chris@82 990 T4w = T2G - T2L;
Chris@82 991 T4x = T2v - T2A;
Chris@82 992 T4y = FNMS(KP587785252, T4x, KP951056516 * T4w);
Chris@82 993 T4A = FMA(KP951056516, T4x, KP587785252 * T4w);
Chris@82 994 ii[WS(rs, 15)] = T4s + T4p;
Chris@82 995 T4z = T4u + T4t;
Chris@82 996 ii[WS(rs, 11)] = T4z - T4A;
Chris@82 997 ii[WS(rs, 19)] = T4A + T4z;
Chris@82 998 T4v = T4t - T4u;
Chris@82 999 ii[WS(rs, 3)] = T4v - T4y;
Chris@82 1000 ii[WS(rs, 7)] = T4y + T4v;
Chris@82 1001 }
Chris@82 1002 {
Chris@82 1003 E T36, T38, T39, T2Z, T3b, T2V, T2Y, T3c, T3a;
Chris@82 1004 T36 = KP559016994 * (T32 - T35);
Chris@82 1005 T38 = T32 + T35;
Chris@82 1006 T39 = FNMS(KP250000000, T38, T37);
Chris@82 1007 T2V = T2T - T2U;
Chris@82 1008 T2Y = T2W - T2X;
Chris@82 1009 T2Z = FMA(KP951056516, T2V, KP587785252 * T2Y);
Chris@82 1010 T3b = FNMS(KP587785252, T2V, KP951056516 * T2Y);
Chris@82 1011 ri[WS(rs, 5)] = T37 + T38;
Chris@82 1012 T3c = T39 - T36;
Chris@82 1013 ri[WS(rs, 13)] = T3b + T3c;
Chris@82 1014 ri[WS(rs, 17)] = T3c - T3b;
Chris@82 1015 T3a = T36 + T39;
Chris@82 1016 ri[WS(rs, 1)] = T2Z + T3a;
Chris@82 1017 ri[WS(rs, 9)] = T3a - T2Z;
Chris@82 1018 }
Chris@82 1019 {
Chris@82 1020 E T4d, T4h, T4i, T4m, T4o, T4k, T4l, T4n, T4j;
Chris@82 1021 T4d = KP559016994 * (T4b - T4c);
Chris@82 1022 T4h = T4b + T4c;
Chris@82 1023 T4i = FNMS(KP250000000, T4h, T4g);
Chris@82 1024 T4k = T30 - T31;
Chris@82 1025 T4l = T33 - T34;
Chris@82 1026 T4m = FMA(KP951056516, T4k, KP587785252 * T4l);
Chris@82 1027 T4o = FNMS(KP587785252, T4k, KP951056516 * T4l);
Chris@82 1028 ii[WS(rs, 5)] = T4h + T4g;
Chris@82 1029 T4n = T4i - T4d;
Chris@82 1030 ii[WS(rs, 13)] = T4n - T4o;
Chris@82 1031 ii[WS(rs, 17)] = T4o + T4n;
Chris@82 1032 T4j = T4d + T4i;
Chris@82 1033 ii[WS(rs, 1)] = T4j - T4m;
Chris@82 1034 ii[WS(rs, 9)] = T4m + T4j;
Chris@82 1035 }
Chris@82 1036 }
Chris@82 1037 }
Chris@82 1038 }
Chris@82 1039
Chris@82 1040 static const tw_instr twinstr[] = {
Chris@82 1041 {TW_FULL, 0, 20},
Chris@82 1042 {TW_NEXT, 1, 0}
Chris@82 1043 };
Chris@82 1044
Chris@82 1045 static const ct_desc desc = { 20, "t1_20", twinstr, &GENUS, {184, 62, 62, 0}, 0, 0, 0 };
Chris@82 1046
Chris@82 1047 void X(codelet_t1_20) (planner *p) {
Chris@82 1048 X(kdft_dit_register) (p, t1_20, &desc);
Chris@82 1049 }
Chris@82 1050 #endif