Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/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 2cd0e3b3e1fd
children
rev   line source
Chris@42 1 /*
Chris@42 2 * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@42 3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@42 4 *
Chris@42 5 * This program is free software; you can redistribute it and/or modify
Chris@42 6 * it under the terms of the GNU General Public License as published by
Chris@42 7 * the Free Software Foundation; either version 2 of the License, or
Chris@42 8 * (at your option) any later version.
Chris@42 9 *
Chris@42 10 * This program is distributed in the hope that it will be useful,
Chris@42 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@42 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
Chris@42 13 * GNU General Public License for more details.
Chris@42 14 *
Chris@42 15 * You should have received a copy of the GNU General Public License
Chris@42 16 * along with this program; if not, write to the Free Software
Chris@42 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Chris@42 18 *
Chris@42 19 */
Chris@42 20
Chris@42 21 /* This file was automatically generated --- DO NOT EDIT */
Chris@42 22 /* Generated on Sat Jul 30 16:36:00 EDT 2016 */
Chris@42 23
Chris@42 24 #include "codelet-dft.h"
Chris@42 25
Chris@42 26 #ifdef HAVE_FMA
Chris@42 27
Chris@42 28 /* Generated by: ../../../genfft/gen_notw.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 20 -name n1_20 -include n.h */
Chris@42 29
Chris@42 30 /*
Chris@42 31 * This function contains 208 FP additions, 72 FP multiplications,
Chris@42 32 * (or, 136 additions, 0 multiplications, 72 fused multiply/add),
Chris@42 33 * 86 stack variables, 4 constants, and 80 memory accesses
Chris@42 34 */
Chris@42 35 #include "n.h"
Chris@42 36
Chris@42 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@42 38 {
Chris@42 39 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@42 40 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@42 41 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
Chris@42 42 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@42 43 {
Chris@42 44 INT i;
Chris@42 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@42 46 E T1Y, T1Z, T1W, T1V;
Chris@42 47 {
Chris@42 48 E T1d, TP, TD, T7, T3b, T2N, T2f, T1R, T2U, TB, T2P, T2A, T3d, T37, T3j;
Chris@42 49 E TJ, T2n, T1b, T1T, T1y, T2b, T2h, T1j, T2V, Tm, T2O, T2H, T3c, T34, T1e;
Chris@42 50 E T1f, T3i, TG, T2m, T10, T1S, T1J, T28, T2g;
Chris@42 51 {
Chris@42 52 E T4, T1N, T3, T2L, TN, T5, T1O, T1P, T1h, T1i;
Chris@42 53 {
Chris@42 54 E T1, T2, TL, TM;
Chris@42 55 T1 = ri[0];
Chris@42 56 T2 = ri[WS(is, 10)];
Chris@42 57 TL = ii[0];
Chris@42 58 TM = ii[WS(is, 10)];
Chris@42 59 T4 = ri[WS(is, 5)];
Chris@42 60 T1N = T1 - T2;
Chris@42 61 T3 = T1 + T2;
Chris@42 62 T2L = TL + TM;
Chris@42 63 TN = TL - TM;
Chris@42 64 T5 = ri[WS(is, 15)];
Chris@42 65 T1O = ii[WS(is, 5)];
Chris@42 66 T1P = ii[WS(is, 15)];
Chris@42 67 }
Chris@42 68 {
Chris@42 69 E T1o, Tp, T2u, T13, T14, Ts, T2v, T1r, Tx, T1t, Tw, T2x, T18, Ty, T1u;
Chris@42 70 E T1v;
Chris@42 71 {
Chris@42 72 E Tq, Tr, T1p, T1q;
Chris@42 73 {
Chris@42 74 E Tn, To, T11, T12;
Chris@42 75 Tn = ri[WS(is, 8)];
Chris@42 76 {
Chris@42 77 E TO, T6, T2M, T1Q;
Chris@42 78 TO = T4 - T5;
Chris@42 79 T6 = T4 + T5;
Chris@42 80 T2M = T1O + T1P;
Chris@42 81 T1Q = T1O - T1P;
Chris@42 82 T1d = TO + TN;
Chris@42 83 TP = TN - TO;
Chris@42 84 TD = T3 + T6;
Chris@42 85 T7 = T3 - T6;
Chris@42 86 T3b = T2L + T2M;
Chris@42 87 T2N = T2L - T2M;
Chris@42 88 T2f = T1N + T1Q;
Chris@42 89 T1R = T1N - T1Q;
Chris@42 90 To = ri[WS(is, 18)];
Chris@42 91 }
Chris@42 92 T11 = ii[WS(is, 8)];
Chris@42 93 T12 = ii[WS(is, 18)];
Chris@42 94 Tq = ri[WS(is, 13)];
Chris@42 95 T1o = Tn - To;
Chris@42 96 Tp = Tn + To;
Chris@42 97 T2u = T11 + T12;
Chris@42 98 T13 = T11 - T12;
Chris@42 99 Tr = ri[WS(is, 3)];
Chris@42 100 T1p = ii[WS(is, 13)];
Chris@42 101 T1q = ii[WS(is, 3)];
Chris@42 102 }
Chris@42 103 {
Chris@42 104 E Tu, Tv, T16, T17;
Chris@42 105 Tu = ri[WS(is, 12)];
Chris@42 106 T14 = Tq - Tr;
Chris@42 107 Ts = Tq + Tr;
Chris@42 108 T2v = T1p + T1q;
Chris@42 109 T1r = T1p - T1q;
Chris@42 110 Tv = ri[WS(is, 2)];
Chris@42 111 T16 = ii[WS(is, 12)];
Chris@42 112 T17 = ii[WS(is, 2)];
Chris@42 113 Tx = ri[WS(is, 17)];
Chris@42 114 T1t = Tu - Tv;
Chris@42 115 Tw = Tu + Tv;
Chris@42 116 T2x = T16 + T17;
Chris@42 117 T18 = T16 - T17;
Chris@42 118 Ty = ri[WS(is, 7)];
Chris@42 119 T1u = ii[WS(is, 17)];
Chris@42 120 T1v = ii[WS(is, 7)];
Chris@42 121 }
Chris@42 122 }
Chris@42 123 {
Chris@42 124 E TH, T19, T1w, TI;
Chris@42 125 {
Chris@42 126 E Tt, T2w, T35, TA, T2z, T36, Tz, T2y;
Chris@42 127 TH = Tp + Ts;
Chris@42 128 Tt = Tp - Ts;
Chris@42 129 T19 = Tx - Ty;
Chris@42 130 Tz = Tx + Ty;
Chris@42 131 T2y = T1u + T1v;
Chris@42 132 T1w = T1u - T1v;
Chris@42 133 T2w = T2u - T2v;
Chris@42 134 T35 = T2u + T2v;
Chris@42 135 TI = Tw + Tz;
Chris@42 136 TA = Tw - Tz;
Chris@42 137 T2z = T2x - T2y;
Chris@42 138 T36 = T2x + T2y;
Chris@42 139 T2U = Tt - TA;
Chris@42 140 TB = Tt + TA;
Chris@42 141 T2P = T2w + T2z;
Chris@42 142 T2A = T2w - T2z;
Chris@42 143 T3d = T35 + T36;
Chris@42 144 T37 = T35 - T36;
Chris@42 145 }
Chris@42 146 {
Chris@42 147 E T1s, T29, T1x, T2a, T15, T1a;
Chris@42 148 T15 = T13 - T14;
Chris@42 149 T1h = T14 + T13;
Chris@42 150 T1i = T19 + T18;
Chris@42 151 T1a = T18 - T19;
Chris@42 152 T1s = T1o - T1r;
Chris@42 153 T29 = T1o + T1r;
Chris@42 154 T3j = TH - TI;
Chris@42 155 TJ = TH + TI;
Chris@42 156 T1x = T1t - T1w;
Chris@42 157 T2a = T1t + T1w;
Chris@42 158 T2n = T15 - T1a;
Chris@42 159 T1b = T15 + T1a;
Chris@42 160 T1T = T1s + T1x;
Chris@42 161 T1y = T1s - T1x;
Chris@42 162 T2b = T29 - T2a;
Chris@42 163 T2h = T29 + T2a;
Chris@42 164 }
Chris@42 165 }
Chris@42 166 }
Chris@42 167 {
Chris@42 168 E Ta, T1z, T2B, TS, TT, Td, T2C, T1C, Ti, T1E, Th, T2E, TX, Tj, T1F;
Chris@42 169 E T1G;
Chris@42 170 {
Chris@42 171 E Tb, Tc, T1A, T1B;
Chris@42 172 {
Chris@42 173 E TQ, TR, T8, T9;
Chris@42 174 T8 = ri[WS(is, 4)];
Chris@42 175 T9 = ri[WS(is, 14)];
Chris@42 176 T1j = T1h + T1i;
Chris@42 177 T1Y = T1h - T1i;
Chris@42 178 TQ = ii[WS(is, 4)];
Chris@42 179 TR = ii[WS(is, 14)];
Chris@42 180 Ta = T8 + T9;
Chris@42 181 T1z = T8 - T9;
Chris@42 182 Tb = ri[WS(is, 9)];
Chris@42 183 T2B = TQ + TR;
Chris@42 184 TS = TQ - TR;
Chris@42 185 Tc = ri[WS(is, 19)];
Chris@42 186 T1A = ii[WS(is, 9)];
Chris@42 187 T1B = ii[WS(is, 19)];
Chris@42 188 }
Chris@42 189 {
Chris@42 190 E Tf, Tg, TV, TW;
Chris@42 191 Tf = ri[WS(is, 16)];
Chris@42 192 TT = Tb - Tc;
Chris@42 193 Td = Tb + Tc;
Chris@42 194 T2C = T1A + T1B;
Chris@42 195 T1C = T1A - T1B;
Chris@42 196 Tg = ri[WS(is, 6)];
Chris@42 197 TV = ii[WS(is, 16)];
Chris@42 198 TW = ii[WS(is, 6)];
Chris@42 199 Ti = ri[WS(is, 1)];
Chris@42 200 T1E = Tf - Tg;
Chris@42 201 Th = Tf + Tg;
Chris@42 202 T2E = TV + TW;
Chris@42 203 TX = TV - TW;
Chris@42 204 Tj = ri[WS(is, 11)];
Chris@42 205 T1F = ii[WS(is, 1)];
Chris@42 206 T1G = ii[WS(is, 11)];
Chris@42 207 }
Chris@42 208 }
Chris@42 209 {
Chris@42 210 E TE, TY, T1H, TF;
Chris@42 211 {
Chris@42 212 E Te, T2D, T32, Tl, T2G, T33, Tk, T2F;
Chris@42 213 TE = Ta + Td;
Chris@42 214 Te = Ta - Td;
Chris@42 215 TY = Ti - Tj;
Chris@42 216 Tk = Ti + Tj;
Chris@42 217 T2F = T1F + T1G;
Chris@42 218 T1H = T1F - T1G;
Chris@42 219 T2D = T2B - T2C;
Chris@42 220 T32 = T2B + T2C;
Chris@42 221 TF = Th + Tk;
Chris@42 222 Tl = Th - Tk;
Chris@42 223 T2G = T2E - T2F;
Chris@42 224 T33 = T2E + T2F;
Chris@42 225 T2V = Te - Tl;
Chris@42 226 Tm = Te + Tl;
Chris@42 227 T2O = T2D + T2G;
Chris@42 228 T2H = T2D - T2G;
Chris@42 229 T3c = T32 + T33;
Chris@42 230 T34 = T32 - T33;
Chris@42 231 }
Chris@42 232 {
Chris@42 233 E T1D, T26, T1I, T27, TU, TZ;
Chris@42 234 TU = TS - TT;
Chris@42 235 T1e = TT + TS;
Chris@42 236 T1f = TY + TX;
Chris@42 237 TZ = TX - TY;
Chris@42 238 T1D = T1z - T1C;
Chris@42 239 T26 = T1z + T1C;
Chris@42 240 T3i = TE - TF;
Chris@42 241 TG = TE + TF;
Chris@42 242 T1I = T1E - T1H;
Chris@42 243 T27 = T1E + T1H;
Chris@42 244 T2m = TU - TZ;
Chris@42 245 T10 = TU + TZ;
Chris@42 246 T1S = T1D + T1I;
Chris@42 247 T1J = T1D - T1I;
Chris@42 248 T28 = T26 - T27;
Chris@42 249 T2g = T26 + T27;
Chris@42 250 }
Chris@42 251 }
Chris@42 252 }
Chris@42 253 }
Chris@42 254 {
Chris@42 255 E T1g, T3g, T3f, T2S, T2R, T2k, T2j;
Chris@42 256 {
Chris@42 257 E T2s, T2r, TC, T2Q;
Chris@42 258 T2s = Tm - TB;
Chris@42 259 TC = Tm + TB;
Chris@42 260 T1g = T1e + T1f;
Chris@42 261 T1Z = T1e - T1f;
Chris@42 262 T2r = FNMS(KP250000000, TC, T7);
Chris@42 263 ro[WS(os, 10)] = T7 + TC;
Chris@42 264 T2Q = T2O + T2P;
Chris@42 265 T2S = T2O - T2P;
Chris@42 266 {
Chris@42 267 E T2K, T2I, T2t, T2J;
Chris@42 268 T2K = FMA(KP618033988, T2A, T2H);
Chris@42 269 T2I = FNMS(KP618033988, T2H, T2A);
Chris@42 270 T2t = FNMS(KP559016994, T2s, T2r);
Chris@42 271 T2J = FMA(KP559016994, T2s, T2r);
Chris@42 272 ro[WS(os, 18)] = FMA(KP951056516, T2I, T2t);
Chris@42 273 ro[WS(os, 2)] = FNMS(KP951056516, T2I, T2t);
Chris@42 274 ro[WS(os, 6)] = FMA(KP951056516, T2K, T2J);
Chris@42 275 ro[WS(os, 14)] = FNMS(KP951056516, T2K, T2J);
Chris@42 276 T2R = FNMS(KP250000000, T2Q, T2N);
Chris@42 277 }
Chris@42 278 io[WS(os, 10)] = T2N + T2Q;
Chris@42 279 }
Chris@42 280 {
Chris@42 281 E T30, T2Z, TK, T3e;
Chris@42 282 TK = TG + TJ;
Chris@42 283 T30 = TG - TJ;
Chris@42 284 {
Chris@42 285 E T2T, T2X, T2Y, T2W;
Chris@42 286 T2T = FNMS(KP559016994, T2S, T2R);
Chris@42 287 T2X = FMA(KP559016994, T2S, T2R);
Chris@42 288 T2Y = FMA(KP618033988, T2U, T2V);
Chris@42 289 T2W = FNMS(KP618033988, T2V, T2U);
Chris@42 290 io[WS(os, 14)] = FMA(KP951056516, T2Y, T2X);
Chris@42 291 io[WS(os, 6)] = FNMS(KP951056516, T2Y, T2X);
Chris@42 292 io[WS(os, 18)] = FNMS(KP951056516, T2W, T2T);
Chris@42 293 io[WS(os, 2)] = FMA(KP951056516, T2W, T2T);
Chris@42 294 T2Z = FNMS(KP250000000, TK, TD);
Chris@42 295 }
Chris@42 296 ro[0] = TD + TK;
Chris@42 297 T3e = T3c + T3d;
Chris@42 298 T3g = T3c - T3d;
Chris@42 299 {
Chris@42 300 E T31, T39, T3a, T38;
Chris@42 301 T31 = FMA(KP559016994, T30, T2Z);
Chris@42 302 T39 = FNMS(KP559016994, T30, T2Z);
Chris@42 303 T3a = FNMS(KP618033988, T34, T37);
Chris@42 304 T38 = FMA(KP618033988, T37, T34);
Chris@42 305 ro[WS(os, 8)] = FMA(KP951056516, T3a, T39);
Chris@42 306 ro[WS(os, 12)] = FNMS(KP951056516, T3a, T39);
Chris@42 307 ro[WS(os, 16)] = FMA(KP951056516, T38, T31);
Chris@42 308 ro[WS(os, 4)] = FNMS(KP951056516, T38, T31);
Chris@42 309 T3f = FNMS(KP250000000, T3e, T3b);
Chris@42 310 }
Chris@42 311 io[0] = T3b + T3e;
Chris@42 312 }
Chris@42 313 {
Chris@42 314 E T24, T23, T1c, T2i;
Chris@42 315 T1c = T10 + T1b;
Chris@42 316 T24 = T10 - T1b;
Chris@42 317 {
Chris@42 318 E T3h, T3l, T3m, T3k;
Chris@42 319 T3h = FMA(KP559016994, T3g, T3f);
Chris@42 320 T3l = FNMS(KP559016994, T3g, T3f);
Chris@42 321 T3m = FNMS(KP618033988, T3i, T3j);
Chris@42 322 T3k = FMA(KP618033988, T3j, T3i);
Chris@42 323 io[WS(os, 12)] = FMA(KP951056516, T3m, T3l);
Chris@42 324 io[WS(os, 8)] = FNMS(KP951056516, T3m, T3l);
Chris@42 325 io[WS(os, 16)] = FNMS(KP951056516, T3k, T3h);
Chris@42 326 io[WS(os, 4)] = FMA(KP951056516, T3k, T3h);
Chris@42 327 T23 = FNMS(KP250000000, T1c, TP);
Chris@42 328 }
Chris@42 329 io[WS(os, 5)] = TP + T1c;
Chris@42 330 T2i = T2g + T2h;
Chris@42 331 T2k = T2g - T2h;
Chris@42 332 {
Chris@42 333 E T25, T2d, T2e, T2c;
Chris@42 334 T25 = FMA(KP559016994, T24, T23);
Chris@42 335 T2d = FNMS(KP559016994, T24, T23);
Chris@42 336 T2e = FNMS(KP618033988, T28, T2b);
Chris@42 337 T2c = FMA(KP618033988, T2b, T28);
Chris@42 338 io[WS(os, 17)] = FMA(KP951056516, T2e, T2d);
Chris@42 339 io[WS(os, 13)] = FNMS(KP951056516, T2e, T2d);
Chris@42 340 io[WS(os, 9)] = FMA(KP951056516, T2c, T25);
Chris@42 341 io[WS(os, 1)] = FNMS(KP951056516, T2c, T25);
Chris@42 342 T2j = FNMS(KP250000000, T2i, T2f);
Chris@42 343 }
Chris@42 344 ro[WS(os, 5)] = T2f + T2i;
Chris@42 345 }
Chris@42 346 {
Chris@42 347 E T1m, T1l, T1k, T1U;
Chris@42 348 T1k = T1g + T1j;
Chris@42 349 T1m = T1g - T1j;
Chris@42 350 {
Chris@42 351 E T2l, T2p, T2q, T2o;
Chris@42 352 T2l = FMA(KP559016994, T2k, T2j);
Chris@42 353 T2p = FNMS(KP559016994, T2k, T2j);
Chris@42 354 T2q = FNMS(KP618033988, T2m, T2n);
Chris@42 355 T2o = FMA(KP618033988, T2n, T2m);
Chris@42 356 ro[WS(os, 17)] = FNMS(KP951056516, T2q, T2p);
Chris@42 357 ro[WS(os, 13)] = FMA(KP951056516, T2q, T2p);
Chris@42 358 ro[WS(os, 9)] = FNMS(KP951056516, T2o, T2l);
Chris@42 359 ro[WS(os, 1)] = FMA(KP951056516, T2o, T2l);
Chris@42 360 T1l = FNMS(KP250000000, T1k, T1d);
Chris@42 361 }
Chris@42 362 io[WS(os, 15)] = T1d + T1k;
Chris@42 363 T1U = T1S + T1T;
Chris@42 364 T1W = T1S - T1T;
Chris@42 365 {
Chris@42 366 E T1n, T1L, T1M, T1K;
Chris@42 367 T1n = FNMS(KP559016994, T1m, T1l);
Chris@42 368 T1L = FMA(KP559016994, T1m, T1l);
Chris@42 369 T1M = FMA(KP618033988, T1y, T1J);
Chris@42 370 T1K = FNMS(KP618033988, T1J, T1y);
Chris@42 371 io[WS(os, 19)] = FMA(KP951056516, T1M, T1L);
Chris@42 372 io[WS(os, 11)] = FNMS(KP951056516, T1M, T1L);
Chris@42 373 io[WS(os, 7)] = FMA(KP951056516, T1K, T1n);
Chris@42 374 io[WS(os, 3)] = FNMS(KP951056516, T1K, T1n);
Chris@42 375 T1V = FNMS(KP250000000, T1U, T1R);
Chris@42 376 }
Chris@42 377 ro[WS(os, 15)] = T1R + T1U;
Chris@42 378 }
Chris@42 379 }
Chris@42 380 }
Chris@42 381 {
Chris@42 382 E T21, T1X, T20, T22;
Chris@42 383 T21 = FMA(KP559016994, T1W, T1V);
Chris@42 384 T1X = FNMS(KP559016994, T1W, T1V);
Chris@42 385 T20 = FNMS(KP618033988, T1Z, T1Y);
Chris@42 386 T22 = FMA(KP618033988, T1Y, T1Z);
Chris@42 387 ro[WS(os, 19)] = FNMS(KP951056516, T22, T21);
Chris@42 388 ro[WS(os, 11)] = FMA(KP951056516, T22, T21);
Chris@42 389 ro[WS(os, 7)] = FNMS(KP951056516, T20, T1X);
Chris@42 390 ro[WS(os, 3)] = FMA(KP951056516, T20, T1X);
Chris@42 391 }
Chris@42 392 }
Chris@42 393 }
Chris@42 394 }
Chris@42 395
Chris@42 396 static const kdft_desc desc = { 20, "n1_20", {136, 0, 72, 0}, &GENUS, 0, 0, 0, 0 };
Chris@42 397
Chris@42 398 void X(codelet_n1_20) (planner *p) {
Chris@42 399 X(kdft_register) (p, n1_20, &desc);
Chris@42 400 }
Chris@42 401
Chris@42 402 #else /* HAVE_FMA */
Chris@42 403
Chris@42 404 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 20 -name n1_20 -include n.h */
Chris@42 405
Chris@42 406 /*
Chris@42 407 * This function contains 208 FP additions, 48 FP multiplications,
Chris@42 408 * (or, 184 additions, 24 multiplications, 24 fused multiply/add),
Chris@42 409 * 81 stack variables, 4 constants, and 80 memory accesses
Chris@42 410 */
Chris@42 411 #include "n.h"
Chris@42 412
Chris@42 413 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@42 414 {
Chris@42 415 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
Chris@42 416 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@42 417 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@42 418 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@42 419 {
Chris@42 420 INT i;
Chris@42 421 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@42 422 E T7, T2Q, T3h, TD, TP, T1U, T2l, T1d, Tt, TA, TB, T2w, T2z, T2S, T35;
Chris@42 423 E T36, T3f, TH, TI, TJ, T15, T1a, T1b, T1s, T1x, T1W, T29, T2a, T2j, T1h;
Chris@42 424 E T1i, T1j, Te, Tl, Tm, T2D, T2G, T2R, T32, T33, T3e, TE, TF, TG, TU;
Chris@42 425 E TZ, T10, T1D, T1I, T1V, T26, T27, T2i, T1e, T1f, T1g;
Chris@42 426 {
Chris@42 427 E T3, T1Q, TN, T2O, T6, TO, T1T, T2P;
Chris@42 428 {
Chris@42 429 E T1, T2, TL, TM;
Chris@42 430 T1 = ri[0];
Chris@42 431 T2 = ri[WS(is, 10)];
Chris@42 432 T3 = T1 + T2;
Chris@42 433 T1Q = T1 - T2;
Chris@42 434 TL = ii[0];
Chris@42 435 TM = ii[WS(is, 10)];
Chris@42 436 TN = TL - TM;
Chris@42 437 T2O = TL + TM;
Chris@42 438 }
Chris@42 439 {
Chris@42 440 E T4, T5, T1R, T1S;
Chris@42 441 T4 = ri[WS(is, 5)];
Chris@42 442 T5 = ri[WS(is, 15)];
Chris@42 443 T6 = T4 + T5;
Chris@42 444 TO = T4 - T5;
Chris@42 445 T1R = ii[WS(is, 5)];
Chris@42 446 T1S = ii[WS(is, 15)];
Chris@42 447 T1T = T1R - T1S;
Chris@42 448 T2P = T1R + T1S;
Chris@42 449 }
Chris@42 450 T7 = T3 - T6;
Chris@42 451 T2Q = T2O - T2P;
Chris@42 452 T3h = T2O + T2P;
Chris@42 453 TD = T3 + T6;
Chris@42 454 TP = TN - TO;
Chris@42 455 T1U = T1Q - T1T;
Chris@42 456 T2l = T1Q + T1T;
Chris@42 457 T1d = TO + TN;
Chris@42 458 }
Chris@42 459 {
Chris@42 460 E Tp, T1o, T13, T2u, Ts, T14, T1r, T2v, Tw, T1t, T18, T2x, Tz, T19, T1w;
Chris@42 461 E T2y;
Chris@42 462 {
Chris@42 463 E Tn, To, T11, T12;
Chris@42 464 Tn = ri[WS(is, 8)];
Chris@42 465 To = ri[WS(is, 18)];
Chris@42 466 Tp = Tn + To;
Chris@42 467 T1o = Tn - To;
Chris@42 468 T11 = ii[WS(is, 8)];
Chris@42 469 T12 = ii[WS(is, 18)];
Chris@42 470 T13 = T11 - T12;
Chris@42 471 T2u = T11 + T12;
Chris@42 472 }
Chris@42 473 {
Chris@42 474 E Tq, Tr, T1p, T1q;
Chris@42 475 Tq = ri[WS(is, 13)];
Chris@42 476 Tr = ri[WS(is, 3)];
Chris@42 477 Ts = Tq + Tr;
Chris@42 478 T14 = Tq - Tr;
Chris@42 479 T1p = ii[WS(is, 13)];
Chris@42 480 T1q = ii[WS(is, 3)];
Chris@42 481 T1r = T1p - T1q;
Chris@42 482 T2v = T1p + T1q;
Chris@42 483 }
Chris@42 484 {
Chris@42 485 E Tu, Tv, T16, T17;
Chris@42 486 Tu = ri[WS(is, 12)];
Chris@42 487 Tv = ri[WS(is, 2)];
Chris@42 488 Tw = Tu + Tv;
Chris@42 489 T1t = Tu - Tv;
Chris@42 490 T16 = ii[WS(is, 12)];
Chris@42 491 T17 = ii[WS(is, 2)];
Chris@42 492 T18 = T16 - T17;
Chris@42 493 T2x = T16 + T17;
Chris@42 494 }
Chris@42 495 {
Chris@42 496 E Tx, Ty, T1u, T1v;
Chris@42 497 Tx = ri[WS(is, 17)];
Chris@42 498 Ty = ri[WS(is, 7)];
Chris@42 499 Tz = Tx + Ty;
Chris@42 500 T19 = Tx - Ty;
Chris@42 501 T1u = ii[WS(is, 17)];
Chris@42 502 T1v = ii[WS(is, 7)];
Chris@42 503 T1w = T1u - T1v;
Chris@42 504 T2y = T1u + T1v;
Chris@42 505 }
Chris@42 506 Tt = Tp - Ts;
Chris@42 507 TA = Tw - Tz;
Chris@42 508 TB = Tt + TA;
Chris@42 509 T2w = T2u - T2v;
Chris@42 510 T2z = T2x - T2y;
Chris@42 511 T2S = T2w + T2z;
Chris@42 512 T35 = T2u + T2v;
Chris@42 513 T36 = T2x + T2y;
Chris@42 514 T3f = T35 + T36;
Chris@42 515 TH = Tp + Ts;
Chris@42 516 TI = Tw + Tz;
Chris@42 517 TJ = TH + TI;
Chris@42 518 T15 = T13 - T14;
Chris@42 519 T1a = T18 - T19;
Chris@42 520 T1b = T15 + T1a;
Chris@42 521 T1s = T1o - T1r;
Chris@42 522 T1x = T1t - T1w;
Chris@42 523 T1W = T1s + T1x;
Chris@42 524 T29 = T1o + T1r;
Chris@42 525 T2a = T1t + T1w;
Chris@42 526 T2j = T29 + T2a;
Chris@42 527 T1h = T14 + T13;
Chris@42 528 T1i = T19 + T18;
Chris@42 529 T1j = T1h + T1i;
Chris@42 530 }
Chris@42 531 {
Chris@42 532 E Ta, T1z, TS, T2B, Td, TT, T1C, T2C, Th, T1E, TX, T2E, Tk, TY, T1H;
Chris@42 533 E T2F;
Chris@42 534 {
Chris@42 535 E T8, T9, TQ, TR;
Chris@42 536 T8 = ri[WS(is, 4)];
Chris@42 537 T9 = ri[WS(is, 14)];
Chris@42 538 Ta = T8 + T9;
Chris@42 539 T1z = T8 - T9;
Chris@42 540 TQ = ii[WS(is, 4)];
Chris@42 541 TR = ii[WS(is, 14)];
Chris@42 542 TS = TQ - TR;
Chris@42 543 T2B = TQ + TR;
Chris@42 544 }
Chris@42 545 {
Chris@42 546 E Tb, Tc, T1A, T1B;
Chris@42 547 Tb = ri[WS(is, 9)];
Chris@42 548 Tc = ri[WS(is, 19)];
Chris@42 549 Td = Tb + Tc;
Chris@42 550 TT = Tb - Tc;
Chris@42 551 T1A = ii[WS(is, 9)];
Chris@42 552 T1B = ii[WS(is, 19)];
Chris@42 553 T1C = T1A - T1B;
Chris@42 554 T2C = T1A + T1B;
Chris@42 555 }
Chris@42 556 {
Chris@42 557 E Tf, Tg, TV, TW;
Chris@42 558 Tf = ri[WS(is, 16)];
Chris@42 559 Tg = ri[WS(is, 6)];
Chris@42 560 Th = Tf + Tg;
Chris@42 561 T1E = Tf - Tg;
Chris@42 562 TV = ii[WS(is, 16)];
Chris@42 563 TW = ii[WS(is, 6)];
Chris@42 564 TX = TV - TW;
Chris@42 565 T2E = TV + TW;
Chris@42 566 }
Chris@42 567 {
Chris@42 568 E Ti, Tj, T1F, T1G;
Chris@42 569 Ti = ri[WS(is, 1)];
Chris@42 570 Tj = ri[WS(is, 11)];
Chris@42 571 Tk = Ti + Tj;
Chris@42 572 TY = Ti - Tj;
Chris@42 573 T1F = ii[WS(is, 1)];
Chris@42 574 T1G = ii[WS(is, 11)];
Chris@42 575 T1H = T1F - T1G;
Chris@42 576 T2F = T1F + T1G;
Chris@42 577 }
Chris@42 578 Te = Ta - Td;
Chris@42 579 Tl = Th - Tk;
Chris@42 580 Tm = Te + Tl;
Chris@42 581 T2D = T2B - T2C;
Chris@42 582 T2G = T2E - T2F;
Chris@42 583 T2R = T2D + T2G;
Chris@42 584 T32 = T2B + T2C;
Chris@42 585 T33 = T2E + T2F;
Chris@42 586 T3e = T32 + T33;
Chris@42 587 TE = Ta + Td;
Chris@42 588 TF = Th + Tk;
Chris@42 589 TG = TE + TF;
Chris@42 590 TU = TS - TT;
Chris@42 591 TZ = TX - TY;
Chris@42 592 T10 = TU + TZ;
Chris@42 593 T1D = T1z - T1C;
Chris@42 594 T1I = T1E - T1H;
Chris@42 595 T1V = T1D + T1I;
Chris@42 596 T26 = T1z + T1C;
Chris@42 597 T27 = T1E + T1H;
Chris@42 598 T2i = T26 + T27;
Chris@42 599 T1e = TT + TS;
Chris@42 600 T1f = TY + TX;
Chris@42 601 T1g = T1e + T1f;
Chris@42 602 }
Chris@42 603 {
Chris@42 604 E T2s, TC, T2r, T2I, T2K, T2A, T2H, T2J, T2t;
Chris@42 605 T2s = KP559016994 * (Tm - TB);
Chris@42 606 TC = Tm + TB;
Chris@42 607 T2r = FNMS(KP250000000, TC, T7);
Chris@42 608 T2A = T2w - T2z;
Chris@42 609 T2H = T2D - T2G;
Chris@42 610 T2I = FNMS(KP587785252, T2H, KP951056516 * T2A);
Chris@42 611 T2K = FMA(KP951056516, T2H, KP587785252 * T2A);
Chris@42 612 ro[WS(os, 10)] = T7 + TC;
Chris@42 613 T2J = T2s + T2r;
Chris@42 614 ro[WS(os, 14)] = T2J - T2K;
Chris@42 615 ro[WS(os, 6)] = T2J + T2K;
Chris@42 616 T2t = T2r - T2s;
Chris@42 617 ro[WS(os, 2)] = T2t - T2I;
Chris@42 618 ro[WS(os, 18)] = T2t + T2I;
Chris@42 619 }
Chris@42 620 {
Chris@42 621 E T2V, T2T, T2U, T2N, T2Y, T2L, T2M, T2X, T2W;
Chris@42 622 T2V = KP559016994 * (T2R - T2S);
Chris@42 623 T2T = T2R + T2S;
Chris@42 624 T2U = FNMS(KP250000000, T2T, T2Q);
Chris@42 625 T2L = Tt - TA;
Chris@42 626 T2M = Te - Tl;
Chris@42 627 T2N = FNMS(KP587785252, T2M, KP951056516 * T2L);
Chris@42 628 T2Y = FMA(KP951056516, T2M, KP587785252 * T2L);
Chris@42 629 io[WS(os, 10)] = T2Q + T2T;
Chris@42 630 T2X = T2V + T2U;
Chris@42 631 io[WS(os, 6)] = T2X - T2Y;
Chris@42 632 io[WS(os, 14)] = T2Y + T2X;
Chris@42 633 T2W = T2U - T2V;
Chris@42 634 io[WS(os, 2)] = T2N + T2W;
Chris@42 635 io[WS(os, 18)] = T2W - T2N;
Chris@42 636 }
Chris@42 637 {
Chris@42 638 E T2Z, TK, T30, T38, T3a, T34, T37, T39, T31;
Chris@42 639 T2Z = KP559016994 * (TG - TJ);
Chris@42 640 TK = TG + TJ;
Chris@42 641 T30 = FNMS(KP250000000, TK, TD);
Chris@42 642 T34 = T32 - T33;
Chris@42 643 T37 = T35 - T36;
Chris@42 644 T38 = FMA(KP951056516, T34, KP587785252 * T37);
Chris@42 645 T3a = FNMS(KP587785252, T34, KP951056516 * T37);
Chris@42 646 ro[0] = TD + TK;
Chris@42 647 T39 = T30 - T2Z;
Chris@42 648 ro[WS(os, 12)] = T39 - T3a;
Chris@42 649 ro[WS(os, 8)] = T39 + T3a;
Chris@42 650 T31 = T2Z + T30;
Chris@42 651 ro[WS(os, 4)] = T31 - T38;
Chris@42 652 ro[WS(os, 16)] = T31 + T38;
Chris@42 653 }
Chris@42 654 {
Chris@42 655 E T3g, T3i, T3j, T3d, T3m, T3b, T3c, T3l, T3k;
Chris@42 656 T3g = KP559016994 * (T3e - T3f);
Chris@42 657 T3i = T3e + T3f;
Chris@42 658 T3j = FNMS(KP250000000, T3i, T3h);
Chris@42 659 T3b = TE - TF;
Chris@42 660 T3c = TH - TI;
Chris@42 661 T3d = FMA(KP951056516, T3b, KP587785252 * T3c);
Chris@42 662 T3m = FNMS(KP587785252, T3b, KP951056516 * T3c);
Chris@42 663 io[0] = T3h + T3i;
Chris@42 664 T3l = T3j - T3g;
Chris@42 665 io[WS(os, 8)] = T3l - T3m;
Chris@42 666 io[WS(os, 12)] = T3m + T3l;
Chris@42 667 T3k = T3g + T3j;
Chris@42 668 io[WS(os, 4)] = T3d + T3k;
Chris@42 669 io[WS(os, 16)] = T3k - T3d;
Chris@42 670 }
Chris@42 671 {
Chris@42 672 E T23, T1c, T24, T2c, T2e, T28, T2b, T2d, T25;
Chris@42 673 T23 = KP559016994 * (T10 - T1b);
Chris@42 674 T1c = T10 + T1b;
Chris@42 675 T24 = FNMS(KP250000000, T1c, TP);
Chris@42 676 T28 = T26 - T27;
Chris@42 677 T2b = T29 - T2a;
Chris@42 678 T2c = FMA(KP951056516, T28, KP587785252 * T2b);
Chris@42 679 T2e = FNMS(KP587785252, T28, KP951056516 * T2b);
Chris@42 680 io[WS(os, 5)] = TP + T1c;
Chris@42 681 T2d = T24 - T23;
Chris@42 682 io[WS(os, 13)] = T2d - T2e;
Chris@42 683 io[WS(os, 17)] = T2d + T2e;
Chris@42 684 T25 = T23 + T24;
Chris@42 685 io[WS(os, 1)] = T25 - T2c;
Chris@42 686 io[WS(os, 9)] = T25 + T2c;
Chris@42 687 }
Chris@42 688 {
Chris@42 689 E T2k, T2m, T2n, T2h, T2p, T2f, T2g, T2q, T2o;
Chris@42 690 T2k = KP559016994 * (T2i - T2j);
Chris@42 691 T2m = T2i + T2j;
Chris@42 692 T2n = FNMS(KP250000000, T2m, T2l);
Chris@42 693 T2f = TU - TZ;
Chris@42 694 T2g = T15 - T1a;
Chris@42 695 T2h = FMA(KP951056516, T2f, KP587785252 * T2g);
Chris@42 696 T2p = FNMS(KP587785252, T2f, KP951056516 * T2g);
Chris@42 697 ro[WS(os, 5)] = T2l + T2m;
Chris@42 698 T2q = T2n - T2k;
Chris@42 699 ro[WS(os, 13)] = T2p + T2q;
Chris@42 700 ro[WS(os, 17)] = T2q - T2p;
Chris@42 701 T2o = T2k + T2n;
Chris@42 702 ro[WS(os, 1)] = T2h + T2o;
Chris@42 703 ro[WS(os, 9)] = T2o - T2h;
Chris@42 704 }
Chris@42 705 {
Chris@42 706 E T1m, T1k, T1l, T1K, T1M, T1y, T1J, T1L, T1n;
Chris@42 707 T1m = KP559016994 * (T1g - T1j);
Chris@42 708 T1k = T1g + T1j;
Chris@42 709 T1l = FNMS(KP250000000, T1k, T1d);
Chris@42 710 T1y = T1s - T1x;
Chris@42 711 T1J = T1D - T1I;
Chris@42 712 T1K = FNMS(KP587785252, T1J, KP951056516 * T1y);
Chris@42 713 T1M = FMA(KP951056516, T1J, KP587785252 * T1y);
Chris@42 714 io[WS(os, 15)] = T1d + T1k;
Chris@42 715 T1L = T1m + T1l;
Chris@42 716 io[WS(os, 11)] = T1L - T1M;
Chris@42 717 io[WS(os, 19)] = T1L + T1M;
Chris@42 718 T1n = T1l - T1m;
Chris@42 719 io[WS(os, 3)] = T1n - T1K;
Chris@42 720 io[WS(os, 7)] = T1n + T1K;
Chris@42 721 }
Chris@42 722 {
Chris@42 723 E T1Z, T1X, T1Y, T1P, T21, T1N, T1O, T22, T20;
Chris@42 724 T1Z = KP559016994 * (T1V - T1W);
Chris@42 725 T1X = T1V + T1W;
Chris@42 726 T1Y = FNMS(KP250000000, T1X, T1U);
Chris@42 727 T1N = T1h - T1i;
Chris@42 728 T1O = T1e - T1f;
Chris@42 729 T1P = FNMS(KP587785252, T1O, KP951056516 * T1N);
Chris@42 730 T21 = FMA(KP951056516, T1O, KP587785252 * T1N);
Chris@42 731 ro[WS(os, 15)] = T1U + T1X;
Chris@42 732 T22 = T1Z + T1Y;
Chris@42 733 ro[WS(os, 11)] = T21 + T22;
Chris@42 734 ro[WS(os, 19)] = T22 - T21;
Chris@42 735 T20 = T1Y - T1Z;
Chris@42 736 ro[WS(os, 3)] = T1P + T20;
Chris@42 737 ro[WS(os, 7)] = T20 - T1P;
Chris@42 738 }
Chris@42 739 }
Chris@42 740 }
Chris@42 741 }
Chris@42 742
Chris@42 743 static const kdft_desc desc = { 20, "n1_20", {184, 24, 24, 0}, &GENUS, 0, 0, 0, 0 };
Chris@42 744
Chris@42 745 void X(codelet_n1_20) (planner *p) {
Chris@42 746 X(kdft_register) (p, n1_20, &desc);
Chris@42 747 }
Chris@42 748
Chris@42 749 #endif /* HAVE_FMA */