Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/dft/simd/common/n1fv_32.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 16:38:47 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_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n1fv_32 -include n1f.h */
Chris@42 29
Chris@42 30 /*
Chris@42 31 * This function contains 186 FP additions, 98 FP multiplications,
Chris@42 32 * (or, 88 additions, 0 multiplications, 98 fused multiply/add),
Chris@42 33 * 104 stack variables, 7 constants, and 64 memory accesses
Chris@42 34 */
Chris@42 35 #include "n1f.h"
Chris@42 36
Chris@42 37 static void n1fv_32(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 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
Chris@42 40 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
Chris@42 41 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
Chris@42 42 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
Chris@42 43 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@42 44 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
Chris@42 45 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@42 46 {
Chris@42 47 INT i;
Chris@42 48 const R *xi;
Chris@42 49 R *xo;
Chris@42 50 xi = ri;
Chris@42 51 xo = ro;
Chris@42 52 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(64, is), MAKE_VOLATILE_STRIDE(64, os)) {
Chris@42 53 V T1h, Tr, T1a, T1k, TI, T1b, T1L, T1P, T1I, T1G, T1O, T1Q, T1H, T1z, T1c;
Chris@42 54 V TZ;
Chris@42 55 {
Chris@42 56 V T2x, T1T, T2K, T1W, T1p, Tb, T1A, T16, Tu, TF, T2N, T2H, T2b, T2t, TY;
Chris@42 57 V T1w, TT, T1v, T20, T2C, Tj, Te, T2h, To, T2f, T23, T2D, TB, TG, Th;
Chris@42 58 V T2i, Tk;
Chris@42 59 {
Chris@42 60 V TL, TW, TP, TQ, T2F, T27, T28, TO;
Chris@42 61 {
Chris@42 62 V T1, T2, T12, T13, T4, T5, T7, T8;
Chris@42 63 T1 = LD(&(xi[0]), ivs, &(xi[0]));
Chris@42 64 T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
Chris@42 65 T12 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
Chris@42 66 T13 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
Chris@42 67 T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
Chris@42 68 T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
Chris@42 69 T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
Chris@42 70 T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
Chris@42 71 {
Chris@42 72 V TM, T25, T26, TN;
Chris@42 73 {
Chris@42 74 V TJ, T3, T14, T1U, T6, T1V, T9, TK, TU, TV, T1R, T1S, Ta, T15;
Chris@42 75 TJ = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
Chris@42 76 T1R = VADD(T1, T2);
Chris@42 77 T3 = VSUB(T1, T2);
Chris@42 78 T1S = VADD(T12, T13);
Chris@42 79 T14 = VSUB(T12, T13);
Chris@42 80 T1U = VADD(T4, T5);
Chris@42 81 T6 = VSUB(T4, T5);
Chris@42 82 T1V = VADD(T7, T8);
Chris@42 83 T9 = VSUB(T7, T8);
Chris@42 84 TK = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
Chris@42 85 TU = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
Chris@42 86 T2x = VSUB(T1R, T1S);
Chris@42 87 T1T = VADD(T1R, T1S);
Chris@42 88 TV = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
Chris@42 89 TM = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
Chris@42 90 T2K = VSUB(T1V, T1U);
Chris@42 91 T1W = VADD(T1U, T1V);
Chris@42 92 Ta = VADD(T6, T9);
Chris@42 93 T15 = VSUB(T9, T6);
Chris@42 94 T25 = VADD(TJ, TK);
Chris@42 95 TL = VSUB(TJ, TK);
Chris@42 96 T26 = VADD(TV, TU);
Chris@42 97 TW = VSUB(TU, TV);
Chris@42 98 TN = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
Chris@42 99 TP = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
Chris@42 100 T1p = VFNMS(LDK(KP707106781), Ta, T3);
Chris@42 101 Tb = VFMA(LDK(KP707106781), Ta, T3);
Chris@42 102 T1A = VFMA(LDK(KP707106781), T15, T14);
Chris@42 103 T16 = VFNMS(LDK(KP707106781), T15, T14);
Chris@42 104 TQ = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
Chris@42 105 }
Chris@42 106 T2F = VSUB(T25, T26);
Chris@42 107 T27 = VADD(T25, T26);
Chris@42 108 T28 = VADD(TM, TN);
Chris@42 109 TO = VSUB(TM, TN);
Chris@42 110 }
Chris@42 111 }
Chris@42 112 {
Chris@42 113 V Ty, T21, Tx, Tz, T1Y, T1Z;
Chris@42 114 {
Chris@42 115 V Ts, Tt, TD, T29, TR, TE, Tv, Tw;
Chris@42 116 Ts = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
Chris@42 117 Tt = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
Chris@42 118 TD = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
Chris@42 119 T29 = VADD(TP, TQ);
Chris@42 120 TR = VSUB(TP, TQ);
Chris@42 121 TE = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
Chris@42 122 Tv = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
Chris@42 123 Tw = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
Chris@42 124 Ty = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
Chris@42 125 T1Y = VADD(Ts, Tt);
Chris@42 126 Tu = VSUB(Ts, Tt);
Chris@42 127 {
Chris@42 128 V T2G, T2a, TX, TS;
Chris@42 129 T2G = VSUB(T29, T28);
Chris@42 130 T2a = VADD(T28, T29);
Chris@42 131 TX = VSUB(TR, TO);
Chris@42 132 TS = VADD(TO, TR);
Chris@42 133 T1Z = VADD(TD, TE);
Chris@42 134 TF = VSUB(TD, TE);
Chris@42 135 T21 = VADD(Tv, Tw);
Chris@42 136 Tx = VSUB(Tv, Tw);
Chris@42 137 T2N = VFMA(LDK(KP414213562), T2F, T2G);
Chris@42 138 T2H = VFNMS(LDK(KP414213562), T2G, T2F);
Chris@42 139 T2b = VSUB(T27, T2a);
Chris@42 140 T2t = VADD(T27, T2a);
Chris@42 141 TY = VFMA(LDK(KP707106781), TX, TW);
Chris@42 142 T1w = VFNMS(LDK(KP707106781), TX, TW);
Chris@42 143 TT = VFMA(LDK(KP707106781), TS, TL);
Chris@42 144 T1v = VFNMS(LDK(KP707106781), TS, TL);
Chris@42 145 Tz = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
Chris@42 146 }
Chris@42 147 }
Chris@42 148 T20 = VADD(T1Y, T1Z);
Chris@42 149 T2C = VSUB(T1Y, T1Z);
Chris@42 150 {
Chris@42 151 V Tc, Td, Tm, Tn;
Chris@42 152 Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
Chris@42 153 Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
Chris@42 154 Tm = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
Chris@42 155 Tn = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
Chris@42 156 {
Chris@42 157 V Tf, TA, T22, Tg;
Chris@42 158 Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
Chris@42 159 TA = VSUB(Ty, Tz);
Chris@42 160 T22 = VADD(Ty, Tz);
Chris@42 161 Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
Chris@42 162 Tj = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
Chris@42 163 Te = VSUB(Tc, Td);
Chris@42 164 T2h = VADD(Tc, Td);
Chris@42 165 To = VSUB(Tm, Tn);
Chris@42 166 T2f = VADD(Tn, Tm);
Chris@42 167 T23 = VADD(T21, T22);
Chris@42 168 T2D = VSUB(T21, T22);
Chris@42 169 TB = VADD(Tx, TA);
Chris@42 170 TG = VSUB(Tx, TA);
Chris@42 171 Th = VSUB(Tf, Tg);
Chris@42 172 T2i = VADD(Tf, Tg);
Chris@42 173 Tk = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
Chris@42 174 }
Chris@42 175 }
Chris@42 176 }
Chris@42 177 }
Chris@42 178 {
Chris@42 179 V T1t, TH, T1s, TC, T2P, T2U, T2n, T2d, T2w, T2u, T1q, T19, T1B, Tq, T2W;
Chris@42 180 V T2M, T2B, T2T, T2v, T2r, T2o, T2m, T2X, T2I;
Chris@42 181 {
Chris@42 182 V T1X, T2p, T2E, T2O, T2s, T2y, T2j, T17, Ti, T2e, Tl, T2c, T2l, T24;
Chris@42 183 T1X = VSUB(T1T, T1W);
Chris@42 184 T2p = VADD(T1T, T1W);
Chris@42 185 T2E = VFNMS(LDK(KP414213562), T2D, T2C);
Chris@42 186 T2O = VFMA(LDK(KP414213562), T2C, T2D);
Chris@42 187 T2s = VADD(T20, T23);
Chris@42 188 T24 = VSUB(T20, T23);
Chris@42 189 T1t = VFNMS(LDK(KP707106781), TG, TF);
Chris@42 190 TH = VFMA(LDK(KP707106781), TG, TF);
Chris@42 191 T1s = VFNMS(LDK(KP707106781), TB, Tu);
Chris@42 192 TC = VFMA(LDK(KP707106781), TB, Tu);
Chris@42 193 T2y = VSUB(T2h, T2i);
Chris@42 194 T2j = VADD(T2h, T2i);
Chris@42 195 T17 = VFMA(LDK(KP414213562), Te, Th);
Chris@42 196 Ti = VFNMS(LDK(KP414213562), Th, Te);
Chris@42 197 T2e = VADD(Tj, Tk);
Chris@42 198 Tl = VSUB(Tj, Tk);
Chris@42 199 T2c = VADD(T24, T2b);
Chris@42 200 T2l = VSUB(T2b, T24);
Chris@42 201 {
Chris@42 202 V T2L, T2A, T2q, T2k;
Chris@42 203 T2P = VSUB(T2N, T2O);
Chris@42 204 T2U = VADD(T2O, T2N);
Chris@42 205 {
Chris@42 206 V T2z, T2g, T18, Tp;
Chris@42 207 T2z = VSUB(T2e, T2f);
Chris@42 208 T2g = VADD(T2e, T2f);
Chris@42 209 T18 = VFMA(LDK(KP414213562), Tl, To);
Chris@42 210 Tp = VFNMS(LDK(KP414213562), To, Tl);
Chris@42 211 T2n = VFMA(LDK(KP707106781), T2c, T1X);
Chris@42 212 T2d = VFNMS(LDK(KP707106781), T2c, T1X);
Chris@42 213 T2w = VSUB(T2t, T2s);
Chris@42 214 T2u = VADD(T2s, T2t);
Chris@42 215 T2L = VSUB(T2z, T2y);
Chris@42 216 T2A = VADD(T2y, T2z);
Chris@42 217 T2q = VADD(T2j, T2g);
Chris@42 218 T2k = VSUB(T2g, T2j);
Chris@42 219 T1q = VADD(T17, T18);
Chris@42 220 T19 = VSUB(T17, T18);
Chris@42 221 T1B = VSUB(Tp, Ti);
Chris@42 222 Tq = VADD(Ti, Tp);
Chris@42 223 }
Chris@42 224 T2W = VFNMS(LDK(KP707106781), T2L, T2K);
Chris@42 225 T2M = VFMA(LDK(KP707106781), T2L, T2K);
Chris@42 226 T2B = VFMA(LDK(KP707106781), T2A, T2x);
Chris@42 227 T2T = VFNMS(LDK(KP707106781), T2A, T2x);
Chris@42 228 T2v = VSUB(T2p, T2q);
Chris@42 229 T2r = VADD(T2p, T2q);
Chris@42 230 T2o = VFMA(LDK(KP707106781), T2l, T2k);
Chris@42 231 T2m = VFNMS(LDK(KP707106781), T2l, T2k);
Chris@42 232 T2X = VSUB(T2H, T2E);
Chris@42 233 T2I = VADD(T2E, T2H);
Chris@42 234 }
Chris@42 235 }
Chris@42 236 {
Chris@42 237 V T2V, T2Z, T2Y, T30, T2R, T2J;
Chris@42 238 T2V = VFNMS(LDK(KP923879532), T2U, T2T);
Chris@42 239 T2Z = VFMA(LDK(KP923879532), T2U, T2T);
Chris@42 240 ST(&(xo[WS(os, 24)]), VFNMSI(T2w, T2v), ovs, &(xo[0]));
Chris@42 241 ST(&(xo[WS(os, 8)]), VFMAI(T2w, T2v), ovs, &(xo[0]));
Chris@42 242 ST(&(xo[0]), VADD(T2r, T2u), ovs, &(xo[0]));
Chris@42 243 ST(&(xo[WS(os, 16)]), VSUB(T2r, T2u), ovs, &(xo[0]));
Chris@42 244 ST(&(xo[WS(os, 28)]), VFNMSI(T2o, T2n), ovs, &(xo[0]));
Chris@42 245 ST(&(xo[WS(os, 4)]), VFMAI(T2o, T2n), ovs, &(xo[0]));
Chris@42 246 ST(&(xo[WS(os, 20)]), VFMAI(T2m, T2d), ovs, &(xo[0]));
Chris@42 247 ST(&(xo[WS(os, 12)]), VFNMSI(T2m, T2d), ovs, &(xo[0]));
Chris@42 248 T2Y = VFMA(LDK(KP923879532), T2X, T2W);
Chris@42 249 T30 = VFNMS(LDK(KP923879532), T2X, T2W);
Chris@42 250 T2R = VFMA(LDK(KP923879532), T2I, T2B);
Chris@42 251 T2J = VFNMS(LDK(KP923879532), T2I, T2B);
Chris@42 252 {
Chris@42 253 V T1J, T1r, T1C, T1M, T2S, T2Q, T1u, T1D, T1E, T1x;
Chris@42 254 T1J = VFNMS(LDK(KP923879532), T1q, T1p);
Chris@42 255 T1r = VFMA(LDK(KP923879532), T1q, T1p);
Chris@42 256 T1C = VFMA(LDK(KP923879532), T1B, T1A);
Chris@42 257 T1M = VFNMS(LDK(KP923879532), T1B, T1A);
Chris@42 258 ST(&(xo[WS(os, 6)]), VFNMSI(T30, T2Z), ovs, &(xo[0]));
Chris@42 259 ST(&(xo[WS(os, 26)]), VFMAI(T30, T2Z), ovs, &(xo[0]));
Chris@42 260 ST(&(xo[WS(os, 22)]), VFNMSI(T2Y, T2V), ovs, &(xo[0]));
Chris@42 261 ST(&(xo[WS(os, 10)]), VFMAI(T2Y, T2V), ovs, &(xo[0]));
Chris@42 262 T2S = VFMA(LDK(KP923879532), T2P, T2M);
Chris@42 263 T2Q = VFNMS(LDK(KP923879532), T2P, T2M);
Chris@42 264 T1u = VFMA(LDK(KP668178637), T1t, T1s);
Chris@42 265 T1D = VFNMS(LDK(KP668178637), T1s, T1t);
Chris@42 266 T1E = VFNMS(LDK(KP668178637), T1v, T1w);
Chris@42 267 T1x = VFMA(LDK(KP668178637), T1w, T1v);
Chris@42 268 {
Chris@42 269 V T1K, T1F, T1N, T1y;
Chris@42 270 T1h = VFNMS(LDK(KP923879532), Tq, Tb);
Chris@42 271 Tr = VFMA(LDK(KP923879532), Tq, Tb);
Chris@42 272 ST(&(xo[WS(os, 30)]), VFNMSI(T2S, T2R), ovs, &(xo[0]));
Chris@42 273 ST(&(xo[WS(os, 2)]), VFMAI(T2S, T2R), ovs, &(xo[0]));
Chris@42 274 ST(&(xo[WS(os, 18)]), VFMAI(T2Q, T2J), ovs, &(xo[0]));
Chris@42 275 ST(&(xo[WS(os, 14)]), VFNMSI(T2Q, T2J), ovs, &(xo[0]));
Chris@42 276 T1K = VADD(T1D, T1E);
Chris@42 277 T1F = VSUB(T1D, T1E);
Chris@42 278 T1N = VSUB(T1x, T1u);
Chris@42 279 T1y = VADD(T1u, T1x);
Chris@42 280 T1a = VFMA(LDK(KP923879532), T19, T16);
Chris@42 281 T1k = VFNMS(LDK(KP923879532), T19, T16);
Chris@42 282 TI = VFNMS(LDK(KP198912367), TH, TC);
Chris@42 283 T1b = VFMA(LDK(KP198912367), TC, TH);
Chris@42 284 T1L = VFMA(LDK(KP831469612), T1K, T1J);
Chris@42 285 T1P = VFNMS(LDK(KP831469612), T1K, T1J);
Chris@42 286 T1I = VFMA(LDK(KP831469612), T1F, T1C);
Chris@42 287 T1G = VFNMS(LDK(KP831469612), T1F, T1C);
Chris@42 288 T1O = VFMA(LDK(KP831469612), T1N, T1M);
Chris@42 289 T1Q = VFNMS(LDK(KP831469612), T1N, T1M);
Chris@42 290 T1H = VFMA(LDK(KP831469612), T1y, T1r);
Chris@42 291 T1z = VFNMS(LDK(KP831469612), T1y, T1r);
Chris@42 292 T1c = VFMA(LDK(KP198912367), TT, TY);
Chris@42 293 TZ = VFNMS(LDK(KP198912367), TY, TT);
Chris@42 294 }
Chris@42 295 }
Chris@42 296 }
Chris@42 297 }
Chris@42 298 }
Chris@42 299 {
Chris@42 300 V T1d, T1i, T10, T1l;
Chris@42 301 ST(&(xo[WS(os, 21)]), VFNMSI(T1O, T1L), ovs, &(xo[WS(os, 1)]));
Chris@42 302 ST(&(xo[WS(os, 11)]), VFMAI(T1O, T1L), ovs, &(xo[WS(os, 1)]));
Chris@42 303 ST(&(xo[WS(os, 27)]), VFMAI(T1Q, T1P), ovs, &(xo[WS(os, 1)]));
Chris@42 304 ST(&(xo[WS(os, 5)]), VFNMSI(T1Q, T1P), ovs, &(xo[WS(os, 1)]));
Chris@42 305 ST(&(xo[WS(os, 3)]), VFMAI(T1I, T1H), ovs, &(xo[WS(os, 1)]));
Chris@42 306 ST(&(xo[WS(os, 29)]), VFNMSI(T1I, T1H), ovs, &(xo[WS(os, 1)]));
Chris@42 307 ST(&(xo[WS(os, 19)]), VFMAI(T1G, T1z), ovs, &(xo[WS(os, 1)]));
Chris@42 308 ST(&(xo[WS(os, 13)]), VFNMSI(T1G, T1z), ovs, &(xo[WS(os, 1)]));
Chris@42 309 T1d = VSUB(T1b, T1c);
Chris@42 310 T1i = VADD(T1b, T1c);
Chris@42 311 T10 = VADD(TI, TZ);
Chris@42 312 T1l = VSUB(TZ, TI);
Chris@42 313 {
Chris@42 314 V T1n, T1j, T1e, T1g, T1o, T1m, T11, T1f;
Chris@42 315 T1n = VFMA(LDK(KP980785280), T1i, T1h);
Chris@42 316 T1j = VFNMS(LDK(KP980785280), T1i, T1h);
Chris@42 317 T1e = VFNMS(LDK(KP980785280), T1d, T1a);
Chris@42 318 T1g = VFMA(LDK(KP980785280), T1d, T1a);
Chris@42 319 T1o = VFMA(LDK(KP980785280), T1l, T1k);
Chris@42 320 T1m = VFNMS(LDK(KP980785280), T1l, T1k);
Chris@42 321 T11 = VFNMS(LDK(KP980785280), T10, Tr);
Chris@42 322 T1f = VFMA(LDK(KP980785280), T10, Tr);
Chris@42 323 ST(&(xo[WS(os, 23)]), VFMAI(T1m, T1j), ovs, &(xo[WS(os, 1)]));
Chris@42 324 ST(&(xo[WS(os, 9)]), VFNMSI(T1m, T1j), ovs, &(xo[WS(os, 1)]));
Chris@42 325 ST(&(xo[WS(os, 25)]), VFNMSI(T1o, T1n), ovs, &(xo[WS(os, 1)]));
Chris@42 326 ST(&(xo[WS(os, 7)]), VFMAI(T1o, T1n), ovs, &(xo[WS(os, 1)]));
Chris@42 327 ST(&(xo[WS(os, 31)]), VFMAI(T1g, T1f), ovs, &(xo[WS(os, 1)]));
Chris@42 328 ST(&(xo[WS(os, 1)]), VFNMSI(T1g, T1f), ovs, &(xo[WS(os, 1)]));
Chris@42 329 ST(&(xo[WS(os, 15)]), VFMAI(T1e, T11), ovs, &(xo[WS(os, 1)]));
Chris@42 330 ST(&(xo[WS(os, 17)]), VFNMSI(T1e, T11), ovs, &(xo[WS(os, 1)]));
Chris@42 331 }
Chris@42 332 }
Chris@42 333 }
Chris@42 334 }
Chris@42 335 VLEAVE();
Chris@42 336 }
Chris@42 337
Chris@42 338 static const kdft_desc desc = { 32, XSIMD_STRING("n1fv_32"), {88, 0, 98, 0}, &GENUS, 0, 0, 0, 0 };
Chris@42 339
Chris@42 340 void XSIMD(codelet_n1fv_32) (planner *p) {
Chris@42 341 X(kdft_register) (p, n1fv_32, &desc);
Chris@42 342 }
Chris@42 343
Chris@42 344 #else /* HAVE_FMA */
Chris@42 345
Chris@42 346 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n1fv_32 -include n1f.h */
Chris@42 347
Chris@42 348 /*
Chris@42 349 * This function contains 186 FP additions, 42 FP multiplications,
Chris@42 350 * (or, 170 additions, 26 multiplications, 16 fused multiply/add),
Chris@42 351 * 58 stack variables, 7 constants, and 64 memory accesses
Chris@42 352 */
Chris@42 353 #include "n1f.h"
Chris@42 354
Chris@42 355 static void n1fv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
Chris@42 356 {
Chris@42 357 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
Chris@42 358 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
Chris@42 359 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
Chris@42 360 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
Chris@42 361 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
Chris@42 362 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@42 363 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@42 364 {
Chris@42 365 INT i;
Chris@42 366 const R *xi;
Chris@42 367 R *xo;
Chris@42 368 xi = ri;
Chris@42 369 xo = ro;
Chris@42 370 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(64, is), MAKE_VOLATILE_STRIDE(64, os)) {
Chris@42 371 V T1T, T1W, T2K, T2x, T16, T1A, Tb, T1p, TT, T1v, TY, T1w, T27, T2a, T2b;
Chris@42 372 V T2H, T2O, TC, T1s, TH, T1t, T20, T23, T24, T2E, T2N, T2g, T2j, Tq, T1B;
Chris@42 373 V T19, T1q, T2A, T2L;
Chris@42 374 {
Chris@42 375 V T3, T1R, T15, T1S, T6, T1U, T9, T1V, T12, Ta;
Chris@42 376 {
Chris@42 377 V T1, T2, T13, T14;
Chris@42 378 T1 = LD(&(xi[0]), ivs, &(xi[0]));
Chris@42 379 T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
Chris@42 380 T3 = VSUB(T1, T2);
Chris@42 381 T1R = VADD(T1, T2);
Chris@42 382 T13 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
Chris@42 383 T14 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
Chris@42 384 T15 = VSUB(T13, T14);
Chris@42 385 T1S = VADD(T13, T14);
Chris@42 386 }
Chris@42 387 {
Chris@42 388 V T4, T5, T7, T8;
Chris@42 389 T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
Chris@42 390 T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
Chris@42 391 T6 = VSUB(T4, T5);
Chris@42 392 T1U = VADD(T4, T5);
Chris@42 393 T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
Chris@42 394 T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
Chris@42 395 T9 = VSUB(T7, T8);
Chris@42 396 T1V = VADD(T7, T8);
Chris@42 397 }
Chris@42 398 T1T = VADD(T1R, T1S);
Chris@42 399 T1W = VADD(T1U, T1V);
Chris@42 400 T2K = VSUB(T1V, T1U);
Chris@42 401 T2x = VSUB(T1R, T1S);
Chris@42 402 T12 = VMUL(LDK(KP707106781), VSUB(T9, T6));
Chris@42 403 T16 = VSUB(T12, T15);
Chris@42 404 T1A = VADD(T15, T12);
Chris@42 405 Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
Chris@42 406 Tb = VADD(T3, Ta);
Chris@42 407 T1p = VSUB(T3, Ta);
Chris@42 408 }
Chris@42 409 {
Chris@42 410 V TL, T25, TX, T26, TO, T28, TR, T29;
Chris@42 411 {
Chris@42 412 V TJ, TK, TV, TW;
Chris@42 413 TJ = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
Chris@42 414 TK = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
Chris@42 415 TL = VSUB(TJ, TK);
Chris@42 416 T25 = VADD(TJ, TK);
Chris@42 417 TV = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
Chris@42 418 TW = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
Chris@42 419 TX = VSUB(TV, TW);
Chris@42 420 T26 = VADD(TV, TW);
Chris@42 421 }
Chris@42 422 {
Chris@42 423 V TM, TN, TP, TQ;
Chris@42 424 TM = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
Chris@42 425 TN = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
Chris@42 426 TO = VSUB(TM, TN);
Chris@42 427 T28 = VADD(TM, TN);
Chris@42 428 TP = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
Chris@42 429 TQ = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
Chris@42 430 TR = VSUB(TP, TQ);
Chris@42 431 T29 = VADD(TP, TQ);
Chris@42 432 }
Chris@42 433 {
Chris@42 434 V TS, TU, T2F, T2G;
Chris@42 435 TS = VMUL(LDK(KP707106781), VADD(TO, TR));
Chris@42 436 TT = VADD(TL, TS);
Chris@42 437 T1v = VSUB(TL, TS);
Chris@42 438 TU = VMUL(LDK(KP707106781), VSUB(TR, TO));
Chris@42 439 TY = VSUB(TU, TX);
Chris@42 440 T1w = VADD(TX, TU);
Chris@42 441 T27 = VADD(T25, T26);
Chris@42 442 T2a = VADD(T28, T29);
Chris@42 443 T2b = VSUB(T27, T2a);
Chris@42 444 T2F = VSUB(T25, T26);
Chris@42 445 T2G = VSUB(T29, T28);
Chris@42 446 T2H = VFNMS(LDK(KP382683432), T2G, VMUL(LDK(KP923879532), T2F));
Chris@42 447 T2O = VFMA(LDK(KP382683432), T2F, VMUL(LDK(KP923879532), T2G));
Chris@42 448 }
Chris@42 449 }
Chris@42 450 {
Chris@42 451 V Tu, T1Y, TG, T1Z, Tx, T21, TA, T22;
Chris@42 452 {
Chris@42 453 V Ts, Tt, TE, TF;
Chris@42 454 Ts = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
Chris@42 455 Tt = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
Chris@42 456 Tu = VSUB(Ts, Tt);
Chris@42 457 T1Y = VADD(Ts, Tt);
Chris@42 458 TE = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
Chris@42 459 TF = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
Chris@42 460 TG = VSUB(TE, TF);
Chris@42 461 T1Z = VADD(TE, TF);
Chris@42 462 }
Chris@42 463 {
Chris@42 464 V Tv, Tw, Ty, Tz;
Chris@42 465 Tv = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
Chris@42 466 Tw = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
Chris@42 467 Tx = VSUB(Tv, Tw);
Chris@42 468 T21 = VADD(Tv, Tw);
Chris@42 469 Ty = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
Chris@42 470 Tz = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
Chris@42 471 TA = VSUB(Ty, Tz);
Chris@42 472 T22 = VADD(Ty, Tz);
Chris@42 473 }
Chris@42 474 {
Chris@42 475 V TB, TD, T2C, T2D;
Chris@42 476 TB = VMUL(LDK(KP707106781), VADD(Tx, TA));
Chris@42 477 TC = VADD(Tu, TB);
Chris@42 478 T1s = VSUB(Tu, TB);
Chris@42 479 TD = VMUL(LDK(KP707106781), VSUB(TA, Tx));
Chris@42 480 TH = VSUB(TD, TG);
Chris@42 481 T1t = VADD(TG, TD);
Chris@42 482 T20 = VADD(T1Y, T1Z);
Chris@42 483 T23 = VADD(T21, T22);
Chris@42 484 T24 = VSUB(T20, T23);
Chris@42 485 T2C = VSUB(T1Y, T1Z);
Chris@42 486 T2D = VSUB(T22, T21);
Chris@42 487 T2E = VFMA(LDK(KP923879532), T2C, VMUL(LDK(KP382683432), T2D));
Chris@42 488 T2N = VFNMS(LDK(KP382683432), T2C, VMUL(LDK(KP923879532), T2D));
Chris@42 489 }
Chris@42 490 }
Chris@42 491 {
Chris@42 492 V Te, T2h, To, T2f, Th, T2i, Tl, T2e, Ti, Tp;
Chris@42 493 {
Chris@42 494 V Tc, Td, Tm, Tn;
Chris@42 495 Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
Chris@42 496 Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
Chris@42 497 Te = VSUB(Tc, Td);
Chris@42 498 T2h = VADD(Tc, Td);
Chris@42 499 Tm = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
Chris@42 500 Tn = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
Chris@42 501 To = VSUB(Tm, Tn);
Chris@42 502 T2f = VADD(Tm, Tn);
Chris@42 503 }
Chris@42 504 {
Chris@42 505 V Tf, Tg, Tj, Tk;
Chris@42 506 Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
Chris@42 507 Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
Chris@42 508 Th = VSUB(Tf, Tg);
Chris@42 509 T2i = VADD(Tf, Tg);
Chris@42 510 Tj = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
Chris@42 511 Tk = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
Chris@42 512 Tl = VSUB(Tj, Tk);
Chris@42 513 T2e = VADD(Tj, Tk);
Chris@42 514 }
Chris@42 515 T2g = VADD(T2e, T2f);
Chris@42 516 T2j = VADD(T2h, T2i);
Chris@42 517 Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
Chris@42 518 Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To));
Chris@42 519 Tq = VADD(Ti, Tp);
Chris@42 520 T1B = VSUB(Tp, Ti);
Chris@42 521 {
Chris@42 522 V T17, T18, T2y, T2z;
Chris@42 523 T17 = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
Chris@42 524 T18 = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
Chris@42 525 T19 = VSUB(T17, T18);
Chris@42 526 T1q = VADD(T18, T17);
Chris@42 527 T2y = VSUB(T2h, T2i);
Chris@42 528 T2z = VSUB(T2e, T2f);
Chris@42 529 T2A = VMUL(LDK(KP707106781), VADD(T2y, T2z));
Chris@42 530 T2L = VMUL(LDK(KP707106781), VSUB(T2z, T2y));
Chris@42 531 }
Chris@42 532 }
Chris@42 533 {
Chris@42 534 V T2d, T2n, T2m, T2o;
Chris@42 535 {
Chris@42 536 V T1X, T2c, T2k, T2l;
Chris@42 537 T1X = VSUB(T1T, T1W);
Chris@42 538 T2c = VMUL(LDK(KP707106781), VADD(T24, T2b));
Chris@42 539 T2d = VADD(T1X, T2c);
Chris@42 540 T2n = VSUB(T1X, T2c);
Chris@42 541 T2k = VSUB(T2g, T2j);
Chris@42 542 T2l = VMUL(LDK(KP707106781), VSUB(T2b, T24));
Chris@42 543 T2m = VBYI(VADD(T2k, T2l));
Chris@42 544 T2o = VBYI(VSUB(T2l, T2k));
Chris@42 545 }
Chris@42 546 ST(&(xo[WS(os, 28)]), VSUB(T2d, T2m), ovs, &(xo[0]));
Chris@42 547 ST(&(xo[WS(os, 12)]), VADD(T2n, T2o), ovs, &(xo[0]));
Chris@42 548 ST(&(xo[WS(os, 4)]), VADD(T2d, T2m), ovs, &(xo[0]));
Chris@42 549 ST(&(xo[WS(os, 20)]), VSUB(T2n, T2o), ovs, &(xo[0]));
Chris@42 550 }
Chris@42 551 {
Chris@42 552 V T2r, T2v, T2u, T2w;
Chris@42 553 {
Chris@42 554 V T2p, T2q, T2s, T2t;
Chris@42 555 T2p = VADD(T1T, T1W);
Chris@42 556 T2q = VADD(T2j, T2g);
Chris@42 557 T2r = VADD(T2p, T2q);
Chris@42 558 T2v = VSUB(T2p, T2q);
Chris@42 559 T2s = VADD(T20, T23);
Chris@42 560 T2t = VADD(T27, T2a);
Chris@42 561 T2u = VADD(T2s, T2t);
Chris@42 562 T2w = VBYI(VSUB(T2t, T2s));
Chris@42 563 }
Chris@42 564 ST(&(xo[WS(os, 16)]), VSUB(T2r, T2u), ovs, &(xo[0]));
Chris@42 565 ST(&(xo[WS(os, 8)]), VADD(T2v, T2w), ovs, &(xo[0]));
Chris@42 566 ST(&(xo[0]), VADD(T2r, T2u), ovs, &(xo[0]));
Chris@42 567 ST(&(xo[WS(os, 24)]), VSUB(T2v, T2w), ovs, &(xo[0]));
Chris@42 568 }
Chris@42 569 {
Chris@42 570 V T2V, T2Z, T2Y, T30;
Chris@42 571 {
Chris@42 572 V T2T, T2U, T2W, T2X;
Chris@42 573 T2T = VSUB(T2H, T2E);
Chris@42 574 T2U = VSUB(T2L, T2K);
Chris@42 575 T2V = VBYI(VSUB(T2T, T2U));
Chris@42 576 T2Z = VBYI(VADD(T2U, T2T));
Chris@42 577 T2W = VSUB(T2x, T2A);
Chris@42 578 T2X = VSUB(T2O, T2N);
Chris@42 579 T2Y = VSUB(T2W, T2X);
Chris@42 580 T30 = VADD(T2W, T2X);
Chris@42 581 }
Chris@42 582 ST(&(xo[WS(os, 10)]), VADD(T2V, T2Y), ovs, &(xo[0]));
Chris@42 583 ST(&(xo[WS(os, 26)]), VSUB(T30, T2Z), ovs, &(xo[0]));
Chris@42 584 ST(&(xo[WS(os, 22)]), VSUB(T2Y, T2V), ovs, &(xo[0]));
Chris@42 585 ST(&(xo[WS(os, 6)]), VADD(T2Z, T30), ovs, &(xo[0]));
Chris@42 586 }
Chris@42 587 {
Chris@42 588 V T2J, T2R, T2Q, T2S;
Chris@42 589 {
Chris@42 590 V T2B, T2I, T2M, T2P;
Chris@42 591 T2B = VADD(T2x, T2A);
Chris@42 592 T2I = VADD(T2E, T2H);
Chris@42 593 T2J = VADD(T2B, T2I);
Chris@42 594 T2R = VSUB(T2B, T2I);
Chris@42 595 T2M = VADD(T2K, T2L);
Chris@42 596 T2P = VADD(T2N, T2O);
Chris@42 597 T2Q = VBYI(VADD(T2M, T2P));
Chris@42 598 T2S = VBYI(VSUB(T2P, T2M));
Chris@42 599 }
Chris@42 600 ST(&(xo[WS(os, 30)]), VSUB(T2J, T2Q), ovs, &(xo[0]));
Chris@42 601 ST(&(xo[WS(os, 14)]), VADD(T2R, T2S), ovs, &(xo[0]));
Chris@42 602 ST(&(xo[WS(os, 2)]), VADD(T2J, T2Q), ovs, &(xo[0]));
Chris@42 603 ST(&(xo[WS(os, 18)]), VSUB(T2R, T2S), ovs, &(xo[0]));
Chris@42 604 }
Chris@42 605 {
Chris@42 606 V T1r, T1C, T1M, T1K, T1F, T1N, T1y, T1J;
Chris@42 607 T1r = VADD(T1p, T1q);
Chris@42 608 T1C = VADD(T1A, T1B);
Chris@42 609 T1M = VSUB(T1p, T1q);
Chris@42 610 T1K = VSUB(T1B, T1A);
Chris@42 611 {
Chris@42 612 V T1D, T1E, T1u, T1x;
Chris@42 613 T1D = VFNMS(LDK(KP555570233), T1s, VMUL(LDK(KP831469612), T1t));
Chris@42 614 T1E = VFMA(LDK(KP555570233), T1v, VMUL(LDK(KP831469612), T1w));
Chris@42 615 T1F = VADD(T1D, T1E);
Chris@42 616 T1N = VSUB(T1E, T1D);
Chris@42 617 T1u = VFMA(LDK(KP831469612), T1s, VMUL(LDK(KP555570233), T1t));
Chris@42 618 T1x = VFNMS(LDK(KP555570233), T1w, VMUL(LDK(KP831469612), T1v));
Chris@42 619 T1y = VADD(T1u, T1x);
Chris@42 620 T1J = VSUB(T1x, T1u);
Chris@42 621 }
Chris@42 622 {
Chris@42 623 V T1z, T1G, T1P, T1Q;
Chris@42 624 T1z = VADD(T1r, T1y);
Chris@42 625 T1G = VBYI(VADD(T1C, T1F));
Chris@42 626 ST(&(xo[WS(os, 29)]), VSUB(T1z, T1G), ovs, &(xo[WS(os, 1)]));
Chris@42 627 ST(&(xo[WS(os, 3)]), VADD(T1z, T1G), ovs, &(xo[WS(os, 1)]));
Chris@42 628 T1P = VBYI(VADD(T1K, T1J));
Chris@42 629 T1Q = VADD(T1M, T1N);
Chris@42 630 ST(&(xo[WS(os, 5)]), VADD(T1P, T1Q), ovs, &(xo[WS(os, 1)]));
Chris@42 631 ST(&(xo[WS(os, 27)]), VSUB(T1Q, T1P), ovs, &(xo[WS(os, 1)]));
Chris@42 632 }
Chris@42 633 {
Chris@42 634 V T1H, T1I, T1L, T1O;
Chris@42 635 T1H = VSUB(T1r, T1y);
Chris@42 636 T1I = VBYI(VSUB(T1F, T1C));
Chris@42 637 ST(&(xo[WS(os, 19)]), VSUB(T1H, T1I), ovs, &(xo[WS(os, 1)]));
Chris@42 638 ST(&(xo[WS(os, 13)]), VADD(T1H, T1I), ovs, &(xo[WS(os, 1)]));
Chris@42 639 T1L = VBYI(VSUB(T1J, T1K));
Chris@42 640 T1O = VSUB(T1M, T1N);
Chris@42 641 ST(&(xo[WS(os, 11)]), VADD(T1L, T1O), ovs, &(xo[WS(os, 1)]));
Chris@42 642 ST(&(xo[WS(os, 21)]), VSUB(T1O, T1L), ovs, &(xo[WS(os, 1)]));
Chris@42 643 }
Chris@42 644 }
Chris@42 645 {
Chris@42 646 V Tr, T1a, T1k, T1i, T1d, T1l, T10, T1h;
Chris@42 647 Tr = VADD(Tb, Tq);
Chris@42 648 T1a = VADD(T16, T19);
Chris@42 649 T1k = VSUB(Tb, Tq);
Chris@42 650 T1i = VSUB(T19, T16);
Chris@42 651 {
Chris@42 652 V T1b, T1c, TI, TZ;
Chris@42 653 T1b = VFNMS(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH));
Chris@42 654 T1c = VFMA(LDK(KP195090322), TT, VMUL(LDK(KP980785280), TY));
Chris@42 655 T1d = VADD(T1b, T1c);
Chris@42 656 T1l = VSUB(T1c, T1b);
Chris@42 657 TI = VFMA(LDK(KP980785280), TC, VMUL(LDK(KP195090322), TH));
Chris@42 658 TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT));
Chris@42 659 T10 = VADD(TI, TZ);
Chris@42 660 T1h = VSUB(TZ, TI);
Chris@42 661 }
Chris@42 662 {
Chris@42 663 V T11, T1e, T1n, T1o;
Chris@42 664 T11 = VADD(Tr, T10);
Chris@42 665 T1e = VBYI(VADD(T1a, T1d));
Chris@42 666 ST(&(xo[WS(os, 31)]), VSUB(T11, T1e), ovs, &(xo[WS(os, 1)]));
Chris@42 667 ST(&(xo[WS(os, 1)]), VADD(T11, T1e), ovs, &(xo[WS(os, 1)]));
Chris@42 668 T1n = VBYI(VADD(T1i, T1h));
Chris@42 669 T1o = VADD(T1k, T1l);
Chris@42 670 ST(&(xo[WS(os, 7)]), VADD(T1n, T1o), ovs, &(xo[WS(os, 1)]));
Chris@42 671 ST(&(xo[WS(os, 25)]), VSUB(T1o, T1n), ovs, &(xo[WS(os, 1)]));
Chris@42 672 }
Chris@42 673 {
Chris@42 674 V T1f, T1g, T1j, T1m;
Chris@42 675 T1f = VSUB(Tr, T10);
Chris@42 676 T1g = VBYI(VSUB(T1d, T1a));
Chris@42 677 ST(&(xo[WS(os, 17)]), VSUB(T1f, T1g), ovs, &(xo[WS(os, 1)]));
Chris@42 678 ST(&(xo[WS(os, 15)]), VADD(T1f, T1g), ovs, &(xo[WS(os, 1)]));
Chris@42 679 T1j = VBYI(VSUB(T1h, T1i));
Chris@42 680 T1m = VSUB(T1k, T1l);
Chris@42 681 ST(&(xo[WS(os, 9)]), VADD(T1j, T1m), ovs, &(xo[WS(os, 1)]));
Chris@42 682 ST(&(xo[WS(os, 23)]), VSUB(T1m, T1j), ovs, &(xo[WS(os, 1)]));
Chris@42 683 }
Chris@42 684 }
Chris@42 685 }
Chris@42 686 }
Chris@42 687 VLEAVE();
Chris@42 688 }
Chris@42 689
Chris@42 690 static const kdft_desc desc = { 32, XSIMD_STRING("n1fv_32"), {170, 26, 16, 0}, &GENUS, 0, 0, 0, 0 };
Chris@42 691
Chris@42 692 void XSIMD(codelet_n1fv_32) (planner *p) {
Chris@42 693 X(kdft_register) (p, n1fv_32, &desc);
Chris@42 694 }
Chris@42 695
Chris@42 696 #endif /* HAVE_FMA */