Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/dft/scalar/codelets/n1_20.c @ 84:08ae793730bd

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add null config files
author Chris Cannam
date Mon, 02 Mar 2020 14:03:47 +0000
parents d0c2a83c1364
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 08:04:12 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_notw.native -fma -compact -variables 4 -pipeline-latency 4 -n 20 -name n1_20 -include dft/scalar/n.h */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 208 FP additions, 72 FP multiplications,
Chris@82 32 * (or, 136 additions, 0 multiplications, 72 fused multiply/add),
Chris@82 33 * 81 stack variables, 4 constants, and 80 memory accesses
Chris@82 34 */
Chris@82 35 #include "dft/scalar/n.h"
Chris@82 36
Chris@82 37 static void n1_20(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
Chris@82 38 {
Chris@82 39 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@82 40 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@82 41 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
Chris@82 42 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@82 43 {
Chris@82 44 INT i;
Chris@82 45 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(80, is), MAKE_VOLATILE_STRIDE(80, os)) {
Chris@82 46 E T7, T2N, T3b, TD, TP, T1R, T2f, T1d, Tt, TA, TB, T2w, T2z, T2P, T35;
Chris@82 47 E T36, T3d, TH, TI, TJ, T15, T1a, T1b, T1s, T1x, T1T, T29, T2a, T2h, T1h;
Chris@82 48 E T1i, T1j, Te, Tl, Tm, T2D, T2G, T2O, T32, T33, T3c, TE, TF, TG, TU;
Chris@82 49 E TZ, T10, T1D, T1I, T1S, T26, T27, T2g, T1e, T1f, T1g;
Chris@82 50 {
Chris@82 51 E T3, T1N, TN, T2L, T6, TO, T1Q, T2M;
Chris@82 52 {
Chris@82 53 E T1, T2, TL, TM;
Chris@82 54 T1 = ri[0];
Chris@82 55 T2 = ri[WS(is, 10)];
Chris@82 56 T3 = T1 + T2;
Chris@82 57 T1N = T1 - T2;
Chris@82 58 TL = ii[0];
Chris@82 59 TM = ii[WS(is, 10)];
Chris@82 60 TN = TL - TM;
Chris@82 61 T2L = TL + TM;
Chris@82 62 }
Chris@82 63 {
Chris@82 64 E T4, T5, T1O, T1P;
Chris@82 65 T4 = ri[WS(is, 5)];
Chris@82 66 T5 = ri[WS(is, 15)];
Chris@82 67 T6 = T4 + T5;
Chris@82 68 TO = T4 - T5;
Chris@82 69 T1O = ii[WS(is, 5)];
Chris@82 70 T1P = ii[WS(is, 15)];
Chris@82 71 T1Q = T1O - T1P;
Chris@82 72 T2M = T1O + T1P;
Chris@82 73 }
Chris@82 74 T7 = T3 - T6;
Chris@82 75 T2N = T2L - T2M;
Chris@82 76 T3b = T2L + T2M;
Chris@82 77 TD = T3 + T6;
Chris@82 78 TP = TN - TO;
Chris@82 79 T1R = T1N - T1Q;
Chris@82 80 T2f = T1N + T1Q;
Chris@82 81 T1d = TO + TN;
Chris@82 82 }
Chris@82 83 {
Chris@82 84 E Tp, T1o, T13, T2u, Ts, T14, T1r, T2v, Tw, T1t, T18, T2x, Tz, T19, T1w;
Chris@82 85 E T2y;
Chris@82 86 {
Chris@82 87 E Tn, To, T11, T12;
Chris@82 88 Tn = ri[WS(is, 8)];
Chris@82 89 To = ri[WS(is, 18)];
Chris@82 90 Tp = Tn + To;
Chris@82 91 T1o = Tn - To;
Chris@82 92 T11 = ii[WS(is, 8)];
Chris@82 93 T12 = ii[WS(is, 18)];
Chris@82 94 T13 = T11 - T12;
Chris@82 95 T2u = T11 + T12;
Chris@82 96 }
Chris@82 97 {
Chris@82 98 E Tq, Tr, T1p, T1q;
Chris@82 99 Tq = ri[WS(is, 13)];
Chris@82 100 Tr = ri[WS(is, 3)];
Chris@82 101 Ts = Tq + Tr;
Chris@82 102 T14 = Tq - Tr;
Chris@82 103 T1p = ii[WS(is, 13)];
Chris@82 104 T1q = ii[WS(is, 3)];
Chris@82 105 T1r = T1p - T1q;
Chris@82 106 T2v = T1p + T1q;
Chris@82 107 }
Chris@82 108 {
Chris@82 109 E Tu, Tv, T16, T17;
Chris@82 110 Tu = ri[WS(is, 12)];
Chris@82 111 Tv = ri[WS(is, 2)];
Chris@82 112 Tw = Tu + Tv;
Chris@82 113 T1t = Tu - Tv;
Chris@82 114 T16 = ii[WS(is, 12)];
Chris@82 115 T17 = ii[WS(is, 2)];
Chris@82 116 T18 = T16 - T17;
Chris@82 117 T2x = T16 + T17;
Chris@82 118 }
Chris@82 119 {
Chris@82 120 E Tx, Ty, T1u, T1v;
Chris@82 121 Tx = ri[WS(is, 17)];
Chris@82 122 Ty = ri[WS(is, 7)];
Chris@82 123 Tz = Tx + Ty;
Chris@82 124 T19 = Tx - Ty;
Chris@82 125 T1u = ii[WS(is, 17)];
Chris@82 126 T1v = ii[WS(is, 7)];
Chris@82 127 T1w = T1u - T1v;
Chris@82 128 T2y = T1u + T1v;
Chris@82 129 }
Chris@82 130 Tt = Tp - Ts;
Chris@82 131 TA = Tw - Tz;
Chris@82 132 TB = Tt + TA;
Chris@82 133 T2w = T2u - T2v;
Chris@82 134 T2z = T2x - T2y;
Chris@82 135 T2P = T2w + T2z;
Chris@82 136 T35 = T2u + T2v;
Chris@82 137 T36 = T2x + T2y;
Chris@82 138 T3d = T35 + T36;
Chris@82 139 TH = Tp + Ts;
Chris@82 140 TI = Tw + Tz;
Chris@82 141 TJ = TH + TI;
Chris@82 142 T15 = T13 - T14;
Chris@82 143 T1a = T18 - T19;
Chris@82 144 T1b = T15 + T1a;
Chris@82 145 T1s = T1o - T1r;
Chris@82 146 T1x = T1t - T1w;
Chris@82 147 T1T = T1s + T1x;
Chris@82 148 T29 = T1o + T1r;
Chris@82 149 T2a = T1t + T1w;
Chris@82 150 T2h = T29 + T2a;
Chris@82 151 T1h = T14 + T13;
Chris@82 152 T1i = T19 + T18;
Chris@82 153 T1j = T1h + T1i;
Chris@82 154 }
Chris@82 155 {
Chris@82 156 E Ta, T1z, TS, T2B, Td, TT, T1C, T2C, Th, T1E, TX, T2E, Tk, TY, T1H;
Chris@82 157 E T2F;
Chris@82 158 {
Chris@82 159 E T8, T9, TQ, TR;
Chris@82 160 T8 = ri[WS(is, 4)];
Chris@82 161 T9 = ri[WS(is, 14)];
Chris@82 162 Ta = T8 + T9;
Chris@82 163 T1z = T8 - T9;
Chris@82 164 TQ = ii[WS(is, 4)];
Chris@82 165 TR = ii[WS(is, 14)];
Chris@82 166 TS = TQ - TR;
Chris@82 167 T2B = TQ + TR;
Chris@82 168 }
Chris@82 169 {
Chris@82 170 E Tb, Tc, T1A, T1B;
Chris@82 171 Tb = ri[WS(is, 9)];
Chris@82 172 Tc = ri[WS(is, 19)];
Chris@82 173 Td = Tb + Tc;
Chris@82 174 TT = Tb - Tc;
Chris@82 175 T1A = ii[WS(is, 9)];
Chris@82 176 T1B = ii[WS(is, 19)];
Chris@82 177 T1C = T1A - T1B;
Chris@82 178 T2C = T1A + T1B;
Chris@82 179 }
Chris@82 180 {
Chris@82 181 E Tf, Tg, TV, TW;
Chris@82 182 Tf = ri[WS(is, 16)];
Chris@82 183 Tg = ri[WS(is, 6)];
Chris@82 184 Th = Tf + Tg;
Chris@82 185 T1E = Tf - Tg;
Chris@82 186 TV = ii[WS(is, 16)];
Chris@82 187 TW = ii[WS(is, 6)];
Chris@82 188 TX = TV - TW;
Chris@82 189 T2E = TV + TW;
Chris@82 190 }
Chris@82 191 {
Chris@82 192 E Ti, Tj, T1F, T1G;
Chris@82 193 Ti = ri[WS(is, 1)];
Chris@82 194 Tj = ri[WS(is, 11)];
Chris@82 195 Tk = Ti + Tj;
Chris@82 196 TY = Ti - Tj;
Chris@82 197 T1F = ii[WS(is, 1)];
Chris@82 198 T1G = ii[WS(is, 11)];
Chris@82 199 T1H = T1F - T1G;
Chris@82 200 T2F = T1F + T1G;
Chris@82 201 }
Chris@82 202 Te = Ta - Td;
Chris@82 203 Tl = Th - Tk;
Chris@82 204 Tm = Te + Tl;
Chris@82 205 T2D = T2B - T2C;
Chris@82 206 T2G = T2E - T2F;
Chris@82 207 T2O = T2D + T2G;
Chris@82 208 T32 = T2B + T2C;
Chris@82 209 T33 = T2E + T2F;
Chris@82 210 T3c = T32 + T33;
Chris@82 211 TE = Ta + Td;
Chris@82 212 TF = Th + Tk;
Chris@82 213 TG = TE + TF;
Chris@82 214 TU = TS - TT;
Chris@82 215 TZ = TX - TY;
Chris@82 216 T10 = TU + TZ;
Chris@82 217 T1D = T1z - T1C;
Chris@82 218 T1I = T1E - T1H;
Chris@82 219 T1S = T1D + T1I;
Chris@82 220 T26 = T1z + T1C;
Chris@82 221 T27 = T1E + T1H;
Chris@82 222 T2g = T26 + T27;
Chris@82 223 T1e = TT + TS;
Chris@82 224 T1f = TY + TX;
Chris@82 225 T1g = T1e + T1f;
Chris@82 226 }
Chris@82 227 {
Chris@82 228 E T2s, TC, T2r, T2I, T2K, T2A, T2H, T2J, T2t;
Chris@82 229 T2s = Tm - TB;
Chris@82 230 TC = Tm + TB;
Chris@82 231 T2r = FNMS(KP250000000, TC, T7);
Chris@82 232 T2A = T2w - T2z;
Chris@82 233 T2H = T2D - T2G;
Chris@82 234 T2I = FNMS(KP618033988, T2H, T2A);
Chris@82 235 T2K = FMA(KP618033988, T2A, T2H);
Chris@82 236 ro[WS(os, 10)] = T7 + TC;
Chris@82 237 T2J = FMA(KP559016994, T2s, T2r);
Chris@82 238 ro[WS(os, 14)] = FNMS(KP951056516, T2K, T2J);
Chris@82 239 ro[WS(os, 6)] = FMA(KP951056516, T2K, T2J);
Chris@82 240 T2t = FNMS(KP559016994, T2s, T2r);
Chris@82 241 ro[WS(os, 2)] = FNMS(KP951056516, T2I, T2t);
Chris@82 242 ro[WS(os, 18)] = FMA(KP951056516, T2I, T2t);
Chris@82 243 }
Chris@82 244 {
Chris@82 245 E T2S, T2Q, T2R, T2W, T2Y, T2U, T2V, T2X, T2T;
Chris@82 246 T2S = T2O - T2P;
Chris@82 247 T2Q = T2O + T2P;
Chris@82 248 T2R = FNMS(KP250000000, T2Q, T2N);
Chris@82 249 T2U = Tt - TA;
Chris@82 250 T2V = Te - Tl;
Chris@82 251 T2W = FNMS(KP618033988, T2V, T2U);
Chris@82 252 T2Y = FMA(KP618033988, T2U, T2V);
Chris@82 253 io[WS(os, 10)] = T2N + T2Q;
Chris@82 254 T2X = FMA(KP559016994, T2S, T2R);
Chris@82 255 io[WS(os, 6)] = FNMS(KP951056516, T2Y, T2X);
Chris@82 256 io[WS(os, 14)] = FMA(KP951056516, T2Y, T2X);
Chris@82 257 T2T = FNMS(KP559016994, T2S, T2R);
Chris@82 258 io[WS(os, 2)] = FMA(KP951056516, T2W, T2T);
Chris@82 259 io[WS(os, 18)] = FNMS(KP951056516, T2W, T2T);
Chris@82 260 }
Chris@82 261 {
Chris@82 262 E T30, TK, T2Z, T38, T3a, T34, T37, T39, T31;
Chris@82 263 T30 = TG - TJ;
Chris@82 264 TK = TG + TJ;
Chris@82 265 T2Z = FNMS(KP250000000, TK, TD);
Chris@82 266 T34 = T32 - T33;
Chris@82 267 T37 = T35 - T36;
Chris@82 268 T38 = FMA(KP618033988, T37, T34);
Chris@82 269 T3a = FNMS(KP618033988, T34, T37);
Chris@82 270 ro[0] = TD + TK;
Chris@82 271 T39 = FNMS(KP559016994, T30, T2Z);
Chris@82 272 ro[WS(os, 12)] = FNMS(KP951056516, T3a, T39);
Chris@82 273 ro[WS(os, 8)] = FMA(KP951056516, T3a, T39);
Chris@82 274 T31 = FMA(KP559016994, T30, T2Z);
Chris@82 275 ro[WS(os, 4)] = FNMS(KP951056516, T38, T31);
Chris@82 276 ro[WS(os, 16)] = FMA(KP951056516, T38, T31);
Chris@82 277 }
Chris@82 278 {
Chris@82 279 E T3g, T3e, T3f, T3k, T3m, T3i, T3j, T3l, T3h;
Chris@82 280 T3g = T3c - T3d;
Chris@82 281 T3e = T3c + T3d;
Chris@82 282 T3f = FNMS(KP250000000, T3e, T3b);
Chris@82 283 T3i = TE - TF;
Chris@82 284 T3j = TH - TI;
Chris@82 285 T3k = FMA(KP618033988, T3j, T3i);
Chris@82 286 T3m = FNMS(KP618033988, T3i, T3j);
Chris@82 287 io[0] = T3b + T3e;
Chris@82 288 T3l = FNMS(KP559016994, T3g, T3f);
Chris@82 289 io[WS(os, 8)] = FNMS(KP951056516, T3m, T3l);
Chris@82 290 io[WS(os, 12)] = FMA(KP951056516, T3m, T3l);
Chris@82 291 T3h = FMA(KP559016994, T3g, T3f);
Chris@82 292 io[WS(os, 4)] = FMA(KP951056516, T3k, T3h);
Chris@82 293 io[WS(os, 16)] = FNMS(KP951056516, T3k, T3h);
Chris@82 294 }
Chris@82 295 {
Chris@82 296 E T24, T1c, T23, T2c, T2e, T28, T2b, T2d, T25;
Chris@82 297 T24 = T10 - T1b;
Chris@82 298 T1c = T10 + T1b;
Chris@82 299 T23 = FNMS(KP250000000, T1c, TP);
Chris@82 300 T28 = T26 - T27;
Chris@82 301 T2b = T29 - T2a;
Chris@82 302 T2c = FMA(KP618033988, T2b, T28);
Chris@82 303 T2e = FNMS(KP618033988, T28, T2b);
Chris@82 304 io[WS(os, 5)] = TP + T1c;
Chris@82 305 T2d = FNMS(KP559016994, T24, T23);
Chris@82 306 io[WS(os, 13)] = FNMS(KP951056516, T2e, T2d);
Chris@82 307 io[WS(os, 17)] = FMA(KP951056516, T2e, T2d);
Chris@82 308 T25 = FMA(KP559016994, T24, T23);
Chris@82 309 io[WS(os, 1)] = FNMS(KP951056516, T2c, T25);
Chris@82 310 io[WS(os, 9)] = FMA(KP951056516, T2c, T25);
Chris@82 311 }
Chris@82 312 {
Chris@82 313 E T2k, T2i, T2j, T2o, T2q, T2m, T2n, T2p, T2l;
Chris@82 314 T2k = T2g - T2h;
Chris@82 315 T2i = T2g + T2h;
Chris@82 316 T2j = FNMS(KP250000000, T2i, T2f);
Chris@82 317 T2m = TU - TZ;
Chris@82 318 T2n = T15 - T1a;
Chris@82 319 T2o = FMA(KP618033988, T2n, T2m);
Chris@82 320 T2q = FNMS(KP618033988, T2m, T2n);
Chris@82 321 ro[WS(os, 5)] = T2f + T2i;
Chris@82 322 T2p = FNMS(KP559016994, T2k, T2j);
Chris@82 323 ro[WS(os, 13)] = FMA(KP951056516, T2q, T2p);
Chris@82 324 ro[WS(os, 17)] = FNMS(KP951056516, T2q, T2p);
Chris@82 325 T2l = FMA(KP559016994, T2k, T2j);
Chris@82 326 ro[WS(os, 1)] = FMA(KP951056516, T2o, T2l);
Chris@82 327 ro[WS(os, 9)] = FNMS(KP951056516, T2o, T2l);
Chris@82 328 }
Chris@82 329 {
Chris@82 330 E T1m, T1k, T1l, T1K, T1M, T1y, T1J, T1L, T1n;
Chris@82 331 T1m = T1g - T1j;
Chris@82 332 T1k = T1g + T1j;
Chris@82 333 T1l = FNMS(KP250000000, T1k, T1d);
Chris@82 334 T1y = T1s - T1x;
Chris@82 335 T1J = T1D - T1I;
Chris@82 336 T1K = FNMS(KP618033988, T1J, T1y);
Chris@82 337 T1M = FMA(KP618033988, T1y, T1J);
Chris@82 338 io[WS(os, 15)] = T1d + T1k;
Chris@82 339 T1L = FMA(KP559016994, T1m, T1l);
Chris@82 340 io[WS(os, 11)] = FNMS(KP951056516, T1M, T1L);
Chris@82 341 io[WS(os, 19)] = FMA(KP951056516, T1M, T1L);
Chris@82 342 T1n = FNMS(KP559016994, T1m, T1l);
Chris@82 343 io[WS(os, 3)] = FNMS(KP951056516, T1K, T1n);
Chris@82 344 io[WS(os, 7)] = FMA(KP951056516, T1K, T1n);
Chris@82 345 }
Chris@82 346 {
Chris@82 347 E T1W, T1U, T1V, T20, T22, T1Y, T1Z, T21, T1X;
Chris@82 348 T1W = T1S - T1T;
Chris@82 349 T1U = T1S + T1T;
Chris@82 350 T1V = FNMS(KP250000000, T1U, T1R);
Chris@82 351 T1Y = T1h - T1i;
Chris@82 352 T1Z = T1e - T1f;
Chris@82 353 T20 = FNMS(KP618033988, T1Z, T1Y);
Chris@82 354 T22 = FMA(KP618033988, T1Y, T1Z);
Chris@82 355 ro[WS(os, 15)] = T1R + T1U;
Chris@82 356 T21 = FMA(KP559016994, T1W, T1V);
Chris@82 357 ro[WS(os, 11)] = FMA(KP951056516, T22, T21);
Chris@82 358 ro[WS(os, 19)] = FNMS(KP951056516, T22, T21);
Chris@82 359 T1X = FNMS(KP559016994, T1W, T1V);
Chris@82 360 ro[WS(os, 3)] = FMA(KP951056516, T20, T1X);
Chris@82 361 ro[WS(os, 7)] = FNMS(KP951056516, T20, T1X);
Chris@82 362 }
Chris@82 363 }
Chris@82 364 }
Chris@82 365 }
Chris@82 366
Chris@82 367 static const kdft_desc desc = { 20, "n1_20", {136, 0, 72, 0}, &GENUS, 0, 0, 0, 0 };
Chris@82 368
Chris@82 369 void X(codelet_n1_20) (planner *p) {
Chris@82 370 X(kdft_register) (p, n1_20, &desc);
Chris@82 371 }
Chris@82 372
Chris@82 373 #else
Chris@82 374
Chris@82 375 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 20 -name n1_20 -include dft/scalar/n.h */
Chris@82 376
Chris@82 377 /*
Chris@82 378 * This function contains 208 FP additions, 48 FP multiplications,
Chris@82 379 * (or, 184 additions, 24 multiplications, 24 fused multiply/add),
Chris@82 380 * 81 stack variables, 4 constants, and 80 memory accesses
Chris@82 381 */
Chris@82 382 #include "dft/scalar/n.h"
Chris@82 383
Chris@82 384 static void n1_20(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
Chris@82 385 {
Chris@82 386 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
Chris@82 387 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@82 388 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@82 389 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@82 390 {
Chris@82 391 INT i;
Chris@82 392 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(80, is), MAKE_VOLATILE_STRIDE(80, os)) {
Chris@82 393 E T7, T2Q, T3h, TD, TP, T1U, T2l, T1d, Tt, TA, TB, T2w, T2z, T2S, T35;
Chris@82 394 E T36, T3f, TH, TI, TJ, T15, T1a, T1b, T1s, T1x, T1W, T29, T2a, T2j, T1h;
Chris@82 395 E T1i, T1j, Te, Tl, Tm, T2D, T2G, T2R, T32, T33, T3e, TE, TF, TG, TU;
Chris@82 396 E TZ, T10, T1D, T1I, T1V, T26, T27, T2i, T1e, T1f, T1g;
Chris@82 397 {
Chris@82 398 E T3, T1Q, TN, T2O, T6, TO, T1T, T2P;
Chris@82 399 {
Chris@82 400 E T1, T2, TL, TM;
Chris@82 401 T1 = ri[0];
Chris@82 402 T2 = ri[WS(is, 10)];
Chris@82 403 T3 = T1 + T2;
Chris@82 404 T1Q = T1 - T2;
Chris@82 405 TL = ii[0];
Chris@82 406 TM = ii[WS(is, 10)];
Chris@82 407 TN = TL - TM;
Chris@82 408 T2O = TL + TM;
Chris@82 409 }
Chris@82 410 {
Chris@82 411 E T4, T5, T1R, T1S;
Chris@82 412 T4 = ri[WS(is, 5)];
Chris@82 413 T5 = ri[WS(is, 15)];
Chris@82 414 T6 = T4 + T5;
Chris@82 415 TO = T4 - T5;
Chris@82 416 T1R = ii[WS(is, 5)];
Chris@82 417 T1S = ii[WS(is, 15)];
Chris@82 418 T1T = T1R - T1S;
Chris@82 419 T2P = T1R + T1S;
Chris@82 420 }
Chris@82 421 T7 = T3 - T6;
Chris@82 422 T2Q = T2O - T2P;
Chris@82 423 T3h = T2O + T2P;
Chris@82 424 TD = T3 + T6;
Chris@82 425 TP = TN - TO;
Chris@82 426 T1U = T1Q - T1T;
Chris@82 427 T2l = T1Q + T1T;
Chris@82 428 T1d = TO + TN;
Chris@82 429 }
Chris@82 430 {
Chris@82 431 E Tp, T1o, T13, T2u, Ts, T14, T1r, T2v, Tw, T1t, T18, T2x, Tz, T19, T1w;
Chris@82 432 E T2y;
Chris@82 433 {
Chris@82 434 E Tn, To, T11, T12;
Chris@82 435 Tn = ri[WS(is, 8)];
Chris@82 436 To = ri[WS(is, 18)];
Chris@82 437 Tp = Tn + To;
Chris@82 438 T1o = Tn - To;
Chris@82 439 T11 = ii[WS(is, 8)];
Chris@82 440 T12 = ii[WS(is, 18)];
Chris@82 441 T13 = T11 - T12;
Chris@82 442 T2u = T11 + T12;
Chris@82 443 }
Chris@82 444 {
Chris@82 445 E Tq, Tr, T1p, T1q;
Chris@82 446 Tq = ri[WS(is, 13)];
Chris@82 447 Tr = ri[WS(is, 3)];
Chris@82 448 Ts = Tq + Tr;
Chris@82 449 T14 = Tq - Tr;
Chris@82 450 T1p = ii[WS(is, 13)];
Chris@82 451 T1q = ii[WS(is, 3)];
Chris@82 452 T1r = T1p - T1q;
Chris@82 453 T2v = T1p + T1q;
Chris@82 454 }
Chris@82 455 {
Chris@82 456 E Tu, Tv, T16, T17;
Chris@82 457 Tu = ri[WS(is, 12)];
Chris@82 458 Tv = ri[WS(is, 2)];
Chris@82 459 Tw = Tu + Tv;
Chris@82 460 T1t = Tu - Tv;
Chris@82 461 T16 = ii[WS(is, 12)];
Chris@82 462 T17 = ii[WS(is, 2)];
Chris@82 463 T18 = T16 - T17;
Chris@82 464 T2x = T16 + T17;
Chris@82 465 }
Chris@82 466 {
Chris@82 467 E Tx, Ty, T1u, T1v;
Chris@82 468 Tx = ri[WS(is, 17)];
Chris@82 469 Ty = ri[WS(is, 7)];
Chris@82 470 Tz = Tx + Ty;
Chris@82 471 T19 = Tx - Ty;
Chris@82 472 T1u = ii[WS(is, 17)];
Chris@82 473 T1v = ii[WS(is, 7)];
Chris@82 474 T1w = T1u - T1v;
Chris@82 475 T2y = T1u + T1v;
Chris@82 476 }
Chris@82 477 Tt = Tp - Ts;
Chris@82 478 TA = Tw - Tz;
Chris@82 479 TB = Tt + TA;
Chris@82 480 T2w = T2u - T2v;
Chris@82 481 T2z = T2x - T2y;
Chris@82 482 T2S = T2w + T2z;
Chris@82 483 T35 = T2u + T2v;
Chris@82 484 T36 = T2x + T2y;
Chris@82 485 T3f = T35 + T36;
Chris@82 486 TH = Tp + Ts;
Chris@82 487 TI = Tw + Tz;
Chris@82 488 TJ = TH + TI;
Chris@82 489 T15 = T13 - T14;
Chris@82 490 T1a = T18 - T19;
Chris@82 491 T1b = T15 + T1a;
Chris@82 492 T1s = T1o - T1r;
Chris@82 493 T1x = T1t - T1w;
Chris@82 494 T1W = T1s + T1x;
Chris@82 495 T29 = T1o + T1r;
Chris@82 496 T2a = T1t + T1w;
Chris@82 497 T2j = T29 + T2a;
Chris@82 498 T1h = T14 + T13;
Chris@82 499 T1i = T19 + T18;
Chris@82 500 T1j = T1h + T1i;
Chris@82 501 }
Chris@82 502 {
Chris@82 503 E Ta, T1z, TS, T2B, Td, TT, T1C, T2C, Th, T1E, TX, T2E, Tk, TY, T1H;
Chris@82 504 E T2F;
Chris@82 505 {
Chris@82 506 E T8, T9, TQ, TR;
Chris@82 507 T8 = ri[WS(is, 4)];
Chris@82 508 T9 = ri[WS(is, 14)];
Chris@82 509 Ta = T8 + T9;
Chris@82 510 T1z = T8 - T9;
Chris@82 511 TQ = ii[WS(is, 4)];
Chris@82 512 TR = ii[WS(is, 14)];
Chris@82 513 TS = TQ - TR;
Chris@82 514 T2B = TQ + TR;
Chris@82 515 }
Chris@82 516 {
Chris@82 517 E Tb, Tc, T1A, T1B;
Chris@82 518 Tb = ri[WS(is, 9)];
Chris@82 519 Tc = ri[WS(is, 19)];
Chris@82 520 Td = Tb + Tc;
Chris@82 521 TT = Tb - Tc;
Chris@82 522 T1A = ii[WS(is, 9)];
Chris@82 523 T1B = ii[WS(is, 19)];
Chris@82 524 T1C = T1A - T1B;
Chris@82 525 T2C = T1A + T1B;
Chris@82 526 }
Chris@82 527 {
Chris@82 528 E Tf, Tg, TV, TW;
Chris@82 529 Tf = ri[WS(is, 16)];
Chris@82 530 Tg = ri[WS(is, 6)];
Chris@82 531 Th = Tf + Tg;
Chris@82 532 T1E = Tf - Tg;
Chris@82 533 TV = ii[WS(is, 16)];
Chris@82 534 TW = ii[WS(is, 6)];
Chris@82 535 TX = TV - TW;
Chris@82 536 T2E = TV + TW;
Chris@82 537 }
Chris@82 538 {
Chris@82 539 E Ti, Tj, T1F, T1G;
Chris@82 540 Ti = ri[WS(is, 1)];
Chris@82 541 Tj = ri[WS(is, 11)];
Chris@82 542 Tk = Ti + Tj;
Chris@82 543 TY = Ti - Tj;
Chris@82 544 T1F = ii[WS(is, 1)];
Chris@82 545 T1G = ii[WS(is, 11)];
Chris@82 546 T1H = T1F - T1G;
Chris@82 547 T2F = T1F + T1G;
Chris@82 548 }
Chris@82 549 Te = Ta - Td;
Chris@82 550 Tl = Th - Tk;
Chris@82 551 Tm = Te + Tl;
Chris@82 552 T2D = T2B - T2C;
Chris@82 553 T2G = T2E - T2F;
Chris@82 554 T2R = T2D + T2G;
Chris@82 555 T32 = T2B + T2C;
Chris@82 556 T33 = T2E + T2F;
Chris@82 557 T3e = T32 + T33;
Chris@82 558 TE = Ta + Td;
Chris@82 559 TF = Th + Tk;
Chris@82 560 TG = TE + TF;
Chris@82 561 TU = TS - TT;
Chris@82 562 TZ = TX - TY;
Chris@82 563 T10 = TU + TZ;
Chris@82 564 T1D = T1z - T1C;
Chris@82 565 T1I = T1E - T1H;
Chris@82 566 T1V = T1D + T1I;
Chris@82 567 T26 = T1z + T1C;
Chris@82 568 T27 = T1E + T1H;
Chris@82 569 T2i = T26 + T27;
Chris@82 570 T1e = TT + TS;
Chris@82 571 T1f = TY + TX;
Chris@82 572 T1g = T1e + T1f;
Chris@82 573 }
Chris@82 574 {
Chris@82 575 E T2s, TC, T2r, T2I, T2K, T2A, T2H, T2J, T2t;
Chris@82 576 T2s = KP559016994 * (Tm - TB);
Chris@82 577 TC = Tm + TB;
Chris@82 578 T2r = FNMS(KP250000000, TC, T7);
Chris@82 579 T2A = T2w - T2z;
Chris@82 580 T2H = T2D - T2G;
Chris@82 581 T2I = FNMS(KP587785252, T2H, KP951056516 * T2A);
Chris@82 582 T2K = FMA(KP951056516, T2H, KP587785252 * T2A);
Chris@82 583 ro[WS(os, 10)] = T7 + TC;
Chris@82 584 T2J = T2s + T2r;
Chris@82 585 ro[WS(os, 14)] = T2J - T2K;
Chris@82 586 ro[WS(os, 6)] = T2J + T2K;
Chris@82 587 T2t = T2r - T2s;
Chris@82 588 ro[WS(os, 2)] = T2t - T2I;
Chris@82 589 ro[WS(os, 18)] = T2t + T2I;
Chris@82 590 }
Chris@82 591 {
Chris@82 592 E T2V, T2T, T2U, T2N, T2Y, T2L, T2M, T2X, T2W;
Chris@82 593 T2V = KP559016994 * (T2R - T2S);
Chris@82 594 T2T = T2R + T2S;
Chris@82 595 T2U = FNMS(KP250000000, T2T, T2Q);
Chris@82 596 T2L = Tt - TA;
Chris@82 597 T2M = Te - Tl;
Chris@82 598 T2N = FNMS(KP587785252, T2M, KP951056516 * T2L);
Chris@82 599 T2Y = FMA(KP951056516, T2M, KP587785252 * T2L);
Chris@82 600 io[WS(os, 10)] = T2Q + T2T;
Chris@82 601 T2X = T2V + T2U;
Chris@82 602 io[WS(os, 6)] = T2X - T2Y;
Chris@82 603 io[WS(os, 14)] = T2Y + T2X;
Chris@82 604 T2W = T2U - T2V;
Chris@82 605 io[WS(os, 2)] = T2N + T2W;
Chris@82 606 io[WS(os, 18)] = T2W - T2N;
Chris@82 607 }
Chris@82 608 {
Chris@82 609 E T2Z, TK, T30, T38, T3a, T34, T37, T39, T31;
Chris@82 610 T2Z = KP559016994 * (TG - TJ);
Chris@82 611 TK = TG + TJ;
Chris@82 612 T30 = FNMS(KP250000000, TK, TD);
Chris@82 613 T34 = T32 - T33;
Chris@82 614 T37 = T35 - T36;
Chris@82 615 T38 = FMA(KP951056516, T34, KP587785252 * T37);
Chris@82 616 T3a = FNMS(KP587785252, T34, KP951056516 * T37);
Chris@82 617 ro[0] = TD + TK;
Chris@82 618 T39 = T30 - T2Z;
Chris@82 619 ro[WS(os, 12)] = T39 - T3a;
Chris@82 620 ro[WS(os, 8)] = T39 + T3a;
Chris@82 621 T31 = T2Z + T30;
Chris@82 622 ro[WS(os, 4)] = T31 - T38;
Chris@82 623 ro[WS(os, 16)] = T31 + T38;
Chris@82 624 }
Chris@82 625 {
Chris@82 626 E T3g, T3i, T3j, T3d, T3m, T3b, T3c, T3l, T3k;
Chris@82 627 T3g = KP559016994 * (T3e - T3f);
Chris@82 628 T3i = T3e + T3f;
Chris@82 629 T3j = FNMS(KP250000000, T3i, T3h);
Chris@82 630 T3b = TE - TF;
Chris@82 631 T3c = TH - TI;
Chris@82 632 T3d = FMA(KP951056516, T3b, KP587785252 * T3c);
Chris@82 633 T3m = FNMS(KP587785252, T3b, KP951056516 * T3c);
Chris@82 634 io[0] = T3h + T3i;
Chris@82 635 T3l = T3j - T3g;
Chris@82 636 io[WS(os, 8)] = T3l - T3m;
Chris@82 637 io[WS(os, 12)] = T3m + T3l;
Chris@82 638 T3k = T3g + T3j;
Chris@82 639 io[WS(os, 4)] = T3d + T3k;
Chris@82 640 io[WS(os, 16)] = T3k - T3d;
Chris@82 641 }
Chris@82 642 {
Chris@82 643 E T23, T1c, T24, T2c, T2e, T28, T2b, T2d, T25;
Chris@82 644 T23 = KP559016994 * (T10 - T1b);
Chris@82 645 T1c = T10 + T1b;
Chris@82 646 T24 = FNMS(KP250000000, T1c, TP);
Chris@82 647 T28 = T26 - T27;
Chris@82 648 T2b = T29 - T2a;
Chris@82 649 T2c = FMA(KP951056516, T28, KP587785252 * T2b);
Chris@82 650 T2e = FNMS(KP587785252, T28, KP951056516 * T2b);
Chris@82 651 io[WS(os, 5)] = TP + T1c;
Chris@82 652 T2d = T24 - T23;
Chris@82 653 io[WS(os, 13)] = T2d - T2e;
Chris@82 654 io[WS(os, 17)] = T2d + T2e;
Chris@82 655 T25 = T23 + T24;
Chris@82 656 io[WS(os, 1)] = T25 - T2c;
Chris@82 657 io[WS(os, 9)] = T25 + T2c;
Chris@82 658 }
Chris@82 659 {
Chris@82 660 E T2k, T2m, T2n, T2h, T2p, T2f, T2g, T2q, T2o;
Chris@82 661 T2k = KP559016994 * (T2i - T2j);
Chris@82 662 T2m = T2i + T2j;
Chris@82 663 T2n = FNMS(KP250000000, T2m, T2l);
Chris@82 664 T2f = TU - TZ;
Chris@82 665 T2g = T15 - T1a;
Chris@82 666 T2h = FMA(KP951056516, T2f, KP587785252 * T2g);
Chris@82 667 T2p = FNMS(KP587785252, T2f, KP951056516 * T2g);
Chris@82 668 ro[WS(os, 5)] = T2l + T2m;
Chris@82 669 T2q = T2n - T2k;
Chris@82 670 ro[WS(os, 13)] = T2p + T2q;
Chris@82 671 ro[WS(os, 17)] = T2q - T2p;
Chris@82 672 T2o = T2k + T2n;
Chris@82 673 ro[WS(os, 1)] = T2h + T2o;
Chris@82 674 ro[WS(os, 9)] = T2o - T2h;
Chris@82 675 }
Chris@82 676 {
Chris@82 677 E T1m, T1k, T1l, T1K, T1M, T1y, T1J, T1L, T1n;
Chris@82 678 T1m = KP559016994 * (T1g - T1j);
Chris@82 679 T1k = T1g + T1j;
Chris@82 680 T1l = FNMS(KP250000000, T1k, T1d);
Chris@82 681 T1y = T1s - T1x;
Chris@82 682 T1J = T1D - T1I;
Chris@82 683 T1K = FNMS(KP587785252, T1J, KP951056516 * T1y);
Chris@82 684 T1M = FMA(KP951056516, T1J, KP587785252 * T1y);
Chris@82 685 io[WS(os, 15)] = T1d + T1k;
Chris@82 686 T1L = T1m + T1l;
Chris@82 687 io[WS(os, 11)] = T1L - T1M;
Chris@82 688 io[WS(os, 19)] = T1L + T1M;
Chris@82 689 T1n = T1l - T1m;
Chris@82 690 io[WS(os, 3)] = T1n - T1K;
Chris@82 691 io[WS(os, 7)] = T1n + T1K;
Chris@82 692 }
Chris@82 693 {
Chris@82 694 E T1Z, T1X, T1Y, T1P, T21, T1N, T1O, T22, T20;
Chris@82 695 T1Z = KP559016994 * (T1V - T1W);
Chris@82 696 T1X = T1V + T1W;
Chris@82 697 T1Y = FNMS(KP250000000, T1X, T1U);
Chris@82 698 T1N = T1h - T1i;
Chris@82 699 T1O = T1e - T1f;
Chris@82 700 T1P = FNMS(KP587785252, T1O, KP951056516 * T1N);
Chris@82 701 T21 = FMA(KP951056516, T1O, KP587785252 * T1N);
Chris@82 702 ro[WS(os, 15)] = T1U + T1X;
Chris@82 703 T22 = T1Z + T1Y;
Chris@82 704 ro[WS(os, 11)] = T21 + T22;
Chris@82 705 ro[WS(os, 19)] = T22 - T21;
Chris@82 706 T20 = T1Y - T1Z;
Chris@82 707 ro[WS(os, 3)] = T1P + T20;
Chris@82 708 ro[WS(os, 7)] = T20 - T1P;
Chris@82 709 }
Chris@82 710 }
Chris@82 711 }
Chris@82 712 }
Chris@82 713
Chris@82 714 static const kdft_desc desc = { 20, "n1_20", {184, 24, 24, 0}, &GENUS, 0, 0, 0, 0 };
Chris@82 715
Chris@82 716 void X(codelet_n1_20) (planner *p) {
Chris@82 717 X(kdft_register) (p, n1_20, &desc);
Chris@82 718 }
Chris@82 719
Chris@82 720 #endif