Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/dft/scalar/codelets/n1_11.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 08:04:10 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 11 -name n1_11 -include dft/scalar/n.h */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 140 FP additions, 110 FP multiplications,
Chris@82 32 * (or, 30 additions, 0 multiplications, 110 fused multiply/add),
Chris@82 33 * 62 stack variables, 10 constants, and 44 memory accesses
Chris@82 34 */
Chris@82 35 #include "dft/scalar/n.h"
Chris@82 36
Chris@82 37 static void n1_11(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(KP989821441, +0.989821441880932732376092037776718787376519372);
Chris@82 40 DK(KP959492973, +0.959492973614497389890368057066327699062454848);
Chris@82 41 DK(KP918985947, +0.918985947228994779780736114132655398124909697);
Chris@82 42 DK(KP830830026, +0.830830026003772851058548298459246407048009821);
Chris@82 43 DK(KP876768831, +0.876768831002589333891339807079336796764054852);
Chris@82 44 DK(KP778434453, +0.778434453334651800608337670740821884709317477);
Chris@82 45 DK(KP715370323, +0.715370323453429719112414662767260662417897278);
Chris@82 46 DK(KP521108558, +0.521108558113202722944698153526659300680427422);
Chris@82 47 DK(KP634356270, +0.634356270682424498893150776899916060542806975);
Chris@82 48 DK(KP342584725, +0.342584725681637509502641509861112333758894680);
Chris@82 49 {
Chris@82 50 INT i;
Chris@82 51 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(44, is), MAKE_VOLATILE_STRIDE(44, os)) {
Chris@82 52 E T1, T1f, T4, T1u, Tg, T1q, T7, T1t, Ta, T1s, Td, T1r, Ti, TP, T26;
Chris@82 53 E TG, T1X, T1O, T1w, TY, T1F, T17, To, T1i, TA, T1k, Tr, T1h, Tu, T1j;
Chris@82 54 E Tx, T1g, TC, TU, T21, TL, T1S, T1J, T1m, T13, T1A, T1c;
Chris@82 55 T1 = ri[0];
Chris@82 56 T1f = ii[0];
Chris@82 57 {
Chris@82 58 E T5, T6, Tp, Tq;
Chris@82 59 {
Chris@82 60 E T2, T3, Te, Tf;
Chris@82 61 T2 = ri[WS(is, 1)];
Chris@82 62 T3 = ri[WS(is, 10)];
Chris@82 63 T4 = T2 + T3;
Chris@82 64 T1u = T3 - T2;
Chris@82 65 Te = ri[WS(is, 5)];
Chris@82 66 Tf = ri[WS(is, 6)];
Chris@82 67 Tg = Te + Tf;
Chris@82 68 T1q = Tf - Te;
Chris@82 69 }
Chris@82 70 T5 = ri[WS(is, 2)];
Chris@82 71 T6 = ri[WS(is, 9)];
Chris@82 72 T7 = T5 + T6;
Chris@82 73 T1t = T6 - T5;
Chris@82 74 {
Chris@82 75 E T8, T9, Tb, Tc;
Chris@82 76 T8 = ri[WS(is, 3)];
Chris@82 77 T9 = ri[WS(is, 8)];
Chris@82 78 Ta = T8 + T9;
Chris@82 79 T1s = T9 - T8;
Chris@82 80 Tb = ri[WS(is, 4)];
Chris@82 81 Tc = ri[WS(is, 7)];
Chris@82 82 Td = Tb + Tc;
Chris@82 83 T1r = Tc - Tb;
Chris@82 84 }
Chris@82 85 {
Chris@82 86 E Th, TO, T25, TF, T1W;
Chris@82 87 Th = FNMS(KP342584725, Ta, T7);
Chris@82 88 Ti = FNMS(KP634356270, Th, Td);
Chris@82 89 TO = FNMS(KP342584725, T4, Ta);
Chris@82 90 TP = FNMS(KP634356270, TO, Tg);
Chris@82 91 T25 = FMA(KP521108558, T1q, T1u);
Chris@82 92 T26 = FMA(KP715370323, T25, T1r);
Chris@82 93 TF = FNMS(KP342584725, Td, T4);
Chris@82 94 TG = FNMS(KP634356270, TF, T7);
Chris@82 95 T1W = FMA(KP521108558, T1s, T1q);
Chris@82 96 T1X = FNMS(KP715370323, T1W, T1t);
Chris@82 97 }
Chris@82 98 {
Chris@82 99 E T1N, T1v, TX, T1E, T16;
Chris@82 100 T1N = FNMS(KP521108558, T1t, T1r);
Chris@82 101 T1O = FMA(KP715370323, T1N, T1q);
Chris@82 102 T1v = FNMS(KP521108558, T1u, T1t);
Chris@82 103 T1w = FNMS(KP715370323, T1v, T1s);
Chris@82 104 TX = FNMS(KP342584725, T7, Tg);
Chris@82 105 TY = FNMS(KP634356270, TX, T4);
Chris@82 106 T1E = FMA(KP521108558, T1r, T1s);
Chris@82 107 T1F = FMA(KP715370323, T1E, T1u);
Chris@82 108 T16 = FNMS(KP342584725, Tg, Td);
Chris@82 109 T17 = FNMS(KP634356270, T16, Ta);
Chris@82 110 }
Chris@82 111 {
Chris@82 112 E Tm, Tn, Ty, Tz;
Chris@82 113 Tm = ii[WS(is, 3)];
Chris@82 114 Tn = ii[WS(is, 8)];
Chris@82 115 To = Tm - Tn;
Chris@82 116 T1i = Tm + Tn;
Chris@82 117 Ty = ii[WS(is, 5)];
Chris@82 118 Tz = ii[WS(is, 6)];
Chris@82 119 TA = Ty - Tz;
Chris@82 120 T1k = Ty + Tz;
Chris@82 121 }
Chris@82 122 Tp = ii[WS(is, 2)];
Chris@82 123 Tq = ii[WS(is, 9)];
Chris@82 124 Tr = Tp - Tq;
Chris@82 125 T1h = Tp + Tq;
Chris@82 126 {
Chris@82 127 E Ts, Tt, Tv, Tw;
Chris@82 128 Ts = ii[WS(is, 4)];
Chris@82 129 Tt = ii[WS(is, 7)];
Chris@82 130 Tu = Ts - Tt;
Chris@82 131 T1j = Ts + Tt;
Chris@82 132 Tv = ii[WS(is, 1)];
Chris@82 133 Tw = ii[WS(is, 10)];
Chris@82 134 Tx = Tv - Tw;
Chris@82 135 T1g = Tv + Tw;
Chris@82 136 }
Chris@82 137 {
Chris@82 138 E TB, TT, T20, TK, T1R;
Chris@82 139 TB = FMA(KP521108558, TA, Tx);
Chris@82 140 TC = FMA(KP715370323, TB, Tu);
Chris@82 141 TT = FNMS(KP521108558, Tr, Tu);
Chris@82 142 TU = FMA(KP715370323, TT, TA);
Chris@82 143 T20 = FNMS(KP342584725, T1i, T1h);
Chris@82 144 T21 = FNMS(KP634356270, T20, T1j);
Chris@82 145 TK = FMA(KP521108558, To, TA);
Chris@82 146 TL = FNMS(KP715370323, TK, Tr);
Chris@82 147 T1R = FNMS(KP342584725, T1j, T1g);
Chris@82 148 T1S = FNMS(KP634356270, T1R, T1h);
Chris@82 149 }
Chris@82 150 {
Chris@82 151 E T1I, T1l, T12, T1z, T1b;
Chris@82 152 T1I = FNMS(KP342584725, T1g, T1i);
Chris@82 153 T1J = FNMS(KP634356270, T1I, T1k);
Chris@82 154 T1l = FNMS(KP342584725, T1k, T1j);
Chris@82 155 T1m = FNMS(KP634356270, T1l, T1i);
Chris@82 156 T12 = FMA(KP521108558, Tu, To);
Chris@82 157 T13 = FMA(KP715370323, T12, Tx);
Chris@82 158 T1z = FNMS(KP342584725, T1h, T1k);
Chris@82 159 T1A = FNMS(KP634356270, T1z, T1g);
Chris@82 160 T1b = FNMS(KP521108558, Tx, Tr);
Chris@82 161 T1c = FNMS(KP715370323, T1b, To);
Chris@82 162 }
Chris@82 163 }
Chris@82 164 ro[0] = T1 + T4 + T7 + Ta + Td + Tg;
Chris@82 165 io[0] = T1f + T1g + T1h + T1i + T1j + T1k;
Chris@82 166 {
Chris@82 167 E Tk, TE, Tj, TD, Tl;
Chris@82 168 Tj = FNMS(KP778434453, Ti, T4);
Chris@82 169 Tk = FNMS(KP876768831, Tj, Tg);
Chris@82 170 TD = FMA(KP830830026, TC, Tr);
Chris@82 171 TE = FMA(KP918985947, TD, To);
Chris@82 172 Tl = FNMS(KP959492973, Tk, T1);
Chris@82 173 ro[WS(os, 10)] = FNMS(KP989821441, TE, Tl);
Chris@82 174 ro[WS(os, 1)] = FMA(KP989821441, TE, Tl);
Chris@82 175 }
Chris@82 176 {
Chris@82 177 E T23, T28, T22, T27, T24;
Chris@82 178 T22 = FNMS(KP778434453, T21, T1g);
Chris@82 179 T23 = FNMS(KP876768831, T22, T1k);
Chris@82 180 T27 = FMA(KP830830026, T26, T1t);
Chris@82 181 T28 = FMA(KP918985947, T27, T1s);
Chris@82 182 T24 = FNMS(KP959492973, T23, T1f);
Chris@82 183 io[WS(os, 1)] = FMA(KP989821441, T28, T24);
Chris@82 184 io[WS(os, 10)] = FNMS(KP989821441, T28, T24);
Chris@82 185 }
Chris@82 186 {
Chris@82 187 E T1U, T1Z, T1T, T1Y, T1V;
Chris@82 188 T1T = FNMS(KP778434453, T1S, T1k);
Chris@82 189 T1U = FNMS(KP876768831, T1T, T1i);
Chris@82 190 T1Y = FMA(KP830830026, T1X, T1u);
Chris@82 191 T1Z = FNMS(KP918985947, T1Y, T1r);
Chris@82 192 T1V = FNMS(KP959492973, T1U, T1f);
Chris@82 193 io[WS(os, 2)] = FNMS(KP989821441, T1Z, T1V);
Chris@82 194 io[WS(os, 9)] = FMA(KP989821441, T1Z, T1V);
Chris@82 195 }
Chris@82 196 {
Chris@82 197 E TI, TN, TH, TM, TJ;
Chris@82 198 TH = FNMS(KP778434453, TG, Tg);
Chris@82 199 TI = FNMS(KP876768831, TH, Ta);
Chris@82 200 TM = FMA(KP830830026, TL, Tx);
Chris@82 201 TN = FNMS(KP918985947, TM, Tu);
Chris@82 202 TJ = FNMS(KP959492973, TI, T1);
Chris@82 203 ro[WS(os, 2)] = FNMS(KP989821441, TN, TJ);
Chris@82 204 ro[WS(os, 9)] = FMA(KP989821441, TN, TJ);
Chris@82 205 }
Chris@82 206 {
Chris@82 207 E TR, TW, TQ, TV, TS;
Chris@82 208 TQ = FNMS(KP778434453, TP, Td);
Chris@82 209 TR = FNMS(KP876768831, TQ, T7);
Chris@82 210 TV = FNMS(KP830830026, TU, To);
Chris@82 211 TW = FNMS(KP918985947, TV, Tx);
Chris@82 212 TS = FNMS(KP959492973, TR, T1);
Chris@82 213 ro[WS(os, 8)] = FNMS(KP989821441, TW, TS);
Chris@82 214 ro[WS(os, 3)] = FMA(KP989821441, TW, TS);
Chris@82 215 }
Chris@82 216 {
Chris@82 217 E T1L, T1Q, T1K, T1P, T1M;
Chris@82 218 T1K = FNMS(KP778434453, T1J, T1j);
Chris@82 219 T1L = FNMS(KP876768831, T1K, T1h);
Chris@82 220 T1P = FNMS(KP830830026, T1O, T1s);
Chris@82 221 T1Q = FNMS(KP918985947, T1P, T1u);
Chris@82 222 T1M = FNMS(KP959492973, T1L, T1f);
Chris@82 223 io[WS(os, 3)] = FMA(KP989821441, T1Q, T1M);
Chris@82 224 io[WS(os, 8)] = FNMS(KP989821441, T1Q, T1M);
Chris@82 225 }
Chris@82 226 {
Chris@82 227 E T10, T15, TZ, T14, T11;
Chris@82 228 TZ = FNMS(KP778434453, TY, Ta);
Chris@82 229 T10 = FNMS(KP876768831, TZ, Td);
Chris@82 230 T14 = FNMS(KP830830026, T13, TA);
Chris@82 231 T15 = FMA(KP918985947, T14, Tr);
Chris@82 232 T11 = FNMS(KP959492973, T10, T1);
Chris@82 233 ro[WS(os, 4)] = FNMS(KP989821441, T15, T11);
Chris@82 234 ro[WS(os, 7)] = FMA(KP989821441, T15, T11);
Chris@82 235 }
Chris@82 236 {
Chris@82 237 E T1C, T1H, T1B, T1G, T1D;
Chris@82 238 T1B = FNMS(KP778434453, T1A, T1i);
Chris@82 239 T1C = FNMS(KP876768831, T1B, T1j);
Chris@82 240 T1G = FNMS(KP830830026, T1F, T1q);
Chris@82 241 T1H = FMA(KP918985947, T1G, T1t);
Chris@82 242 T1D = FNMS(KP959492973, T1C, T1f);
Chris@82 243 io[WS(os, 4)] = FNMS(KP989821441, T1H, T1D);
Chris@82 244 io[WS(os, 7)] = FMA(KP989821441, T1H, T1D);
Chris@82 245 }
Chris@82 246 {
Chris@82 247 E T1o, T1y, T1n, T1x, T1p;
Chris@82 248 T1n = FNMS(KP778434453, T1m, T1h);
Chris@82 249 T1o = FNMS(KP876768831, T1n, T1g);
Chris@82 250 T1x = FNMS(KP830830026, T1w, T1r);
Chris@82 251 T1y = FNMS(KP918985947, T1x, T1q);
Chris@82 252 T1p = FNMS(KP959492973, T1o, T1f);
Chris@82 253 io[WS(os, 5)] = FMA(KP989821441, T1y, T1p);
Chris@82 254 io[WS(os, 6)] = FNMS(KP989821441, T1y, T1p);
Chris@82 255 }
Chris@82 256 {
Chris@82 257 E T19, T1e, T18, T1d, T1a;
Chris@82 258 T18 = FNMS(KP778434453, T17, T7);
Chris@82 259 T19 = FNMS(KP876768831, T18, T4);
Chris@82 260 T1d = FNMS(KP830830026, T1c, Tu);
Chris@82 261 T1e = FNMS(KP918985947, T1d, TA);
Chris@82 262 T1a = FNMS(KP959492973, T19, T1);
Chris@82 263 ro[WS(os, 6)] = FNMS(KP989821441, T1e, T1a);
Chris@82 264 ro[WS(os, 5)] = FMA(KP989821441, T1e, T1a);
Chris@82 265 }
Chris@82 266 }
Chris@82 267 }
Chris@82 268 }
Chris@82 269
Chris@82 270 static const kdft_desc desc = { 11, "n1_11", {30, 0, 110, 0}, &GENUS, 0, 0, 0, 0 };
Chris@82 271
Chris@82 272 void X(codelet_n1_11) (planner *p) {
Chris@82 273 X(kdft_register) (p, n1_11, &desc);
Chris@82 274 }
Chris@82 275
Chris@82 276 #else
Chris@82 277
Chris@82 278 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 11 -name n1_11 -include dft/scalar/n.h */
Chris@82 279
Chris@82 280 /*
Chris@82 281 * This function contains 140 FP additions, 100 FP multiplications,
Chris@82 282 * (or, 60 additions, 20 multiplications, 80 fused multiply/add),
Chris@82 283 * 41 stack variables, 10 constants, and 44 memory accesses
Chris@82 284 */
Chris@82 285 #include "dft/scalar/n.h"
Chris@82 286
Chris@82 287 static void n1_11(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
Chris@82 288 {
Chris@82 289 DK(KP654860733, +0.654860733945285064056925072466293553183791199);
Chris@82 290 DK(KP142314838, +0.142314838273285140443792668616369668791051361);
Chris@82 291 DK(KP959492973, +0.959492973614497389890368057066327699062454848);
Chris@82 292 DK(KP415415013, +0.415415013001886425529274149229623203524004910);
Chris@82 293 DK(KP841253532, +0.841253532831181168861811648919367717513292498);
Chris@82 294 DK(KP989821441, +0.989821441880932732376092037776718787376519372);
Chris@82 295 DK(KP909631995, +0.909631995354518371411715383079028460060241051);
Chris@82 296 DK(KP281732556, +0.281732556841429697711417915346616899035777899);
Chris@82 297 DK(KP540640817, +0.540640817455597582107635954318691695431770608);
Chris@82 298 DK(KP755749574, +0.755749574354258283774035843972344420179717445);
Chris@82 299 {
Chris@82 300 INT i;
Chris@82 301 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(44, is), MAKE_VOLATILE_STRIDE(44, os)) {
Chris@82 302 E T1, TM, T4, TG, Tk, TR, Tw, TN, T7, TK, Ta, TH, Tn, TQ, Td;
Chris@82 303 E TJ, Tq, TO, Tt, TP, Tg, TI;
Chris@82 304 {
Chris@82 305 E T2, T3, Ti, Tj;
Chris@82 306 T1 = ri[0];
Chris@82 307 TM = ii[0];
Chris@82 308 T2 = ri[WS(is, 1)];
Chris@82 309 T3 = ri[WS(is, 10)];
Chris@82 310 T4 = T2 + T3;
Chris@82 311 TG = T3 - T2;
Chris@82 312 Ti = ii[WS(is, 1)];
Chris@82 313 Tj = ii[WS(is, 10)];
Chris@82 314 Tk = Ti - Tj;
Chris@82 315 TR = Ti + Tj;
Chris@82 316 {
Chris@82 317 E Tu, Tv, T5, T6;
Chris@82 318 Tu = ii[WS(is, 2)];
Chris@82 319 Tv = ii[WS(is, 9)];
Chris@82 320 Tw = Tu - Tv;
Chris@82 321 TN = Tu + Tv;
Chris@82 322 T5 = ri[WS(is, 2)];
Chris@82 323 T6 = ri[WS(is, 9)];
Chris@82 324 T7 = T5 + T6;
Chris@82 325 TK = T6 - T5;
Chris@82 326 }
Chris@82 327 }
Chris@82 328 {
Chris@82 329 E T8, T9, To, Tp;
Chris@82 330 T8 = ri[WS(is, 3)];
Chris@82 331 T9 = ri[WS(is, 8)];
Chris@82 332 Ta = T8 + T9;
Chris@82 333 TH = T9 - T8;
Chris@82 334 {
Chris@82 335 E Tl, Tm, Tb, Tc;
Chris@82 336 Tl = ii[WS(is, 3)];
Chris@82 337 Tm = ii[WS(is, 8)];
Chris@82 338 Tn = Tl - Tm;
Chris@82 339 TQ = Tl + Tm;
Chris@82 340 Tb = ri[WS(is, 4)];
Chris@82 341 Tc = ri[WS(is, 7)];
Chris@82 342 Td = Tb + Tc;
Chris@82 343 TJ = Tc - Tb;
Chris@82 344 }
Chris@82 345 To = ii[WS(is, 4)];
Chris@82 346 Tp = ii[WS(is, 7)];
Chris@82 347 Tq = To - Tp;
Chris@82 348 TO = To + Tp;
Chris@82 349 {
Chris@82 350 E Tr, Ts, Te, Tf;
Chris@82 351 Tr = ii[WS(is, 5)];
Chris@82 352 Ts = ii[WS(is, 6)];
Chris@82 353 Tt = Tr - Ts;
Chris@82 354 TP = Tr + Ts;
Chris@82 355 Te = ri[WS(is, 5)];
Chris@82 356 Tf = ri[WS(is, 6)];
Chris@82 357 Tg = Te + Tf;
Chris@82 358 TI = Tf - Te;
Chris@82 359 }
Chris@82 360 }
Chris@82 361 {
Chris@82 362 E Tx, Th, TZ, T10;
Chris@82 363 ro[0] = T1 + T4 + T7 + Ta + Td + Tg;
Chris@82 364 io[0] = TM + TR + TN + TQ + TO + TP;
Chris@82 365 Tx = FMA(KP755749574, Tk, KP540640817 * Tn) + FNMS(KP909631995, Tt, KP281732556 * Tq) - (KP989821441 * Tw);
Chris@82 366 Th = FMA(KP841253532, Ta, T1) + FNMS(KP959492973, Td, KP415415013 * Tg) + FNMA(KP142314838, T7, KP654860733 * T4);
Chris@82 367 ro[WS(os, 7)] = Th - Tx;
Chris@82 368 ro[WS(os, 4)] = Th + Tx;
Chris@82 369 TZ = FMA(KP755749574, TG, KP540640817 * TH) + FNMS(KP909631995, TI, KP281732556 * TJ) - (KP989821441 * TK);
Chris@82 370 T10 = FMA(KP841253532, TQ, TM) + FNMS(KP959492973, TO, KP415415013 * TP) + FNMA(KP142314838, TN, KP654860733 * TR);
Chris@82 371 io[WS(os, 4)] = TZ + T10;
Chris@82 372 io[WS(os, 7)] = T10 - TZ;
Chris@82 373 {
Chris@82 374 E TX, TY, Tz, Ty;
Chris@82 375 TX = FMA(KP909631995, TG, KP755749574 * TK) + FNMA(KP540640817, TI, KP989821441 * TJ) - (KP281732556 * TH);
Chris@82 376 TY = FMA(KP415415013, TR, TM) + FNMS(KP142314838, TO, KP841253532 * TP) + FNMA(KP959492973, TQ, KP654860733 * TN);
Chris@82 377 io[WS(os, 2)] = TX + TY;
Chris@82 378 io[WS(os, 9)] = TY - TX;
Chris@82 379 Tz = FMA(KP909631995, Tk, KP755749574 * Tw) + FNMA(KP540640817, Tt, KP989821441 * Tq) - (KP281732556 * Tn);
Chris@82 380 Ty = FMA(KP415415013, T4, T1) + FNMS(KP142314838, Td, KP841253532 * Tg) + FNMA(KP959492973, Ta, KP654860733 * T7);
Chris@82 381 ro[WS(os, 9)] = Ty - Tz;
Chris@82 382 ro[WS(os, 2)] = Ty + Tz;
Chris@82 383 }
Chris@82 384 }
Chris@82 385 {
Chris@82 386 E TB, TA, TT, TU;
Chris@82 387 TB = FMA(KP540640817, Tk, KP909631995 * Tw) + FMA(KP989821441, Tn, KP755749574 * Tq) + (KP281732556 * Tt);
Chris@82 388 TA = FMA(KP841253532, T4, T1) + FNMS(KP959492973, Tg, KP415415013 * T7) + FNMA(KP654860733, Td, KP142314838 * Ta);
Chris@82 389 ro[WS(os, 10)] = TA - TB;
Chris@82 390 ro[WS(os, 1)] = TA + TB;
Chris@82 391 {
Chris@82 392 E TV, TW, TD, TC;
Chris@82 393 TV = FMA(KP540640817, TG, KP909631995 * TK) + FMA(KP989821441, TH, KP755749574 * TJ) + (KP281732556 * TI);
Chris@82 394 TW = FMA(KP841253532, TR, TM) + FNMS(KP959492973, TP, KP415415013 * TN) + FNMA(KP654860733, TO, KP142314838 * TQ);
Chris@82 395 io[WS(os, 1)] = TV + TW;
Chris@82 396 io[WS(os, 10)] = TW - TV;
Chris@82 397 TD = FMA(KP989821441, Tk, KP540640817 * Tq) + FNMS(KP909631995, Tn, KP755749574 * Tt) - (KP281732556 * Tw);
Chris@82 398 TC = FMA(KP415415013, Ta, T1) + FNMS(KP654860733, Tg, KP841253532 * Td) + FNMA(KP959492973, T7, KP142314838 * T4);
Chris@82 399 ro[WS(os, 8)] = TC - TD;
Chris@82 400 ro[WS(os, 3)] = TC + TD;
Chris@82 401 }
Chris@82 402 TT = FMA(KP989821441, TG, KP540640817 * TJ) + FNMS(KP909631995, TH, KP755749574 * TI) - (KP281732556 * TK);
Chris@82 403 TU = FMA(KP415415013, TQ, TM) + FNMS(KP654860733, TP, KP841253532 * TO) + FNMA(KP959492973, TN, KP142314838 * TR);
Chris@82 404 io[WS(os, 3)] = TT + TU;
Chris@82 405 io[WS(os, 8)] = TU - TT;
Chris@82 406 {
Chris@82 407 E TL, TS, TF, TE;
Chris@82 408 TL = FMA(KP281732556, TG, KP755749574 * TH) + FNMS(KP909631995, TJ, KP989821441 * TI) - (KP540640817 * TK);
Chris@82 409 TS = FMA(KP841253532, TN, TM) + FNMS(KP142314838, TP, KP415415013 * TO) + FNMA(KP654860733, TQ, KP959492973 * TR);
Chris@82 410 io[WS(os, 5)] = TL + TS;
Chris@82 411 io[WS(os, 6)] = TS - TL;
Chris@82 412 TF = FMA(KP281732556, Tk, KP755749574 * Tn) + FNMS(KP909631995, Tq, KP989821441 * Tt) - (KP540640817 * Tw);
Chris@82 413 TE = FMA(KP841253532, T7, T1) + FNMS(KP142314838, Tg, KP415415013 * Td) + FNMA(KP654860733, Ta, KP959492973 * T4);
Chris@82 414 ro[WS(os, 6)] = TE - TF;
Chris@82 415 ro[WS(os, 5)] = TE + TF;
Chris@82 416 }
Chris@82 417 }
Chris@82 418 }
Chris@82 419 }
Chris@82 420 }
Chris@82 421
Chris@82 422 static const kdft_desc desc = { 11, "n1_11", {60, 20, 80, 0}, &GENUS, 0, 0, 0, 0 };
Chris@82 423
Chris@82 424 void X(codelet_n1_11) (planner *p) {
Chris@82 425 X(kdft_register) (p, n1_11, &desc);
Chris@82 426 }
Chris@82 427
Chris@82 428 #endif