Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/dft/simd/common/n2bv_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 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 08:05:13 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_c.native -fma -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 32 -name n2bv_32 -with-ostride 2 -include dft/simd/n2b.h -store-multiple 2 */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 186 FP additions, 98 FP multiplications,
Chris@82 32 * (or, 88 additions, 0 multiplications, 98 fused multiply/add),
Chris@82 33 * 72 stack variables, 7 constants, and 80 memory accesses
Chris@82 34 */
Chris@82 35 #include "dft/simd/n2b.h"
Chris@82 36
Chris@82 37 static void n2bv_32(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 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
Chris@82 40 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
Chris@82 41 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
Chris@82 42 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
Chris@82 43 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@82 44 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
Chris@82 45 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@82 46 {
Chris@82 47 INT i;
Chris@82 48 const R *xi;
Chris@82 49 R *xo;
Chris@82 50 xi = ii;
Chris@82 51 xo = io;
Chris@82 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@82 53 V T1T, T1W, T2K, T2x, T16, T1A, Tb, T1p, TT, T1v, TY, T1w, T27, T2a, T2b;
Chris@82 54 V T2H, T2O, TC, T1s, TH, T1t, T20, T23, T24, T2E, T2N, T2g, T2j, Tq, T1B;
Chris@82 55 V T19, T1q, T2A, T2L;
Chris@82 56 {
Chris@82 57 V T3, T1R, T14, T1S, T6, T1U, T9, T1V, T15, Ta;
Chris@82 58 {
Chris@82 59 V T1, T2, T12, T13;
Chris@82 60 T1 = LD(&(xi[0]), ivs, &(xi[0]));
Chris@82 61 T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
Chris@82 62 T3 = VSUB(T1, T2);
Chris@82 63 T1R = VADD(T1, T2);
Chris@82 64 T12 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
Chris@82 65 T13 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
Chris@82 66 T14 = VSUB(T12, T13);
Chris@82 67 T1S = VADD(T12, T13);
Chris@82 68 }
Chris@82 69 {
Chris@82 70 V T4, T5, T7, T8;
Chris@82 71 T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
Chris@82 72 T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
Chris@82 73 T6 = VSUB(T4, T5);
Chris@82 74 T1U = VADD(T4, T5);
Chris@82 75 T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
Chris@82 76 T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
Chris@82 77 T9 = VSUB(T7, T8);
Chris@82 78 T1V = VADD(T7, T8);
Chris@82 79 }
Chris@82 80 T1T = VADD(T1R, T1S);
Chris@82 81 T1W = VADD(T1U, T1V);
Chris@82 82 T2K = VSUB(T1U, T1V);
Chris@82 83 T2x = VSUB(T1R, T1S);
Chris@82 84 T15 = VSUB(T6, T9);
Chris@82 85 T16 = VFMA(LDK(KP707106781), T15, T14);
Chris@82 86 T1A = VFNMS(LDK(KP707106781), T15, T14);
Chris@82 87 Ta = VADD(T6, T9);
Chris@82 88 Tb = VFMA(LDK(KP707106781), Ta, T3);
Chris@82 89 T1p = VFNMS(LDK(KP707106781), Ta, T3);
Chris@82 90 }
Chris@82 91 {
Chris@82 92 V TL, T25, TW, T26, TO, T28, TR, T29;
Chris@82 93 {
Chris@82 94 V TJ, TK, TU, TV;
Chris@82 95 TJ = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
Chris@82 96 TK = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
Chris@82 97 TL = VSUB(TJ, TK);
Chris@82 98 T25 = VADD(TJ, TK);
Chris@82 99 TU = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
Chris@82 100 TV = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
Chris@82 101 TW = VSUB(TU, TV);
Chris@82 102 T26 = VADD(TV, TU);
Chris@82 103 }
Chris@82 104 {
Chris@82 105 V TM, TN, TP, TQ;
Chris@82 106 TM = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
Chris@82 107 TN = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
Chris@82 108 TO = VSUB(TM, TN);
Chris@82 109 T28 = VADD(TM, TN);
Chris@82 110 TP = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
Chris@82 111 TQ = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
Chris@82 112 TR = VSUB(TP, TQ);
Chris@82 113 T29 = VADD(TP, TQ);
Chris@82 114 }
Chris@82 115 {
Chris@82 116 V TS, TX, T2F, T2G;
Chris@82 117 TS = VADD(TO, TR);
Chris@82 118 TT = VFMA(LDK(KP707106781), TS, TL);
Chris@82 119 T1v = VFNMS(LDK(KP707106781), TS, TL);
Chris@82 120 TX = VSUB(TR, TO);
Chris@82 121 TY = VFMA(LDK(KP707106781), TX, TW);
Chris@82 122 T1w = VFNMS(LDK(KP707106781), TX, TW);
Chris@82 123 T27 = VADD(T25, T26);
Chris@82 124 T2a = VADD(T28, T29);
Chris@82 125 T2b = VSUB(T27, T2a);
Chris@82 126 T2F = VSUB(T25, T26);
Chris@82 127 T2G = VSUB(T29, T28);
Chris@82 128 T2H = VFNMS(LDK(KP414213562), T2G, T2F);
Chris@82 129 T2O = VFMA(LDK(KP414213562), T2F, T2G);
Chris@82 130 }
Chris@82 131 }
Chris@82 132 {
Chris@82 133 V Tu, T1Y, TF, T1Z, Tx, T21, TA, T22;
Chris@82 134 {
Chris@82 135 V Ts, Tt, TD, TE;
Chris@82 136 Ts = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
Chris@82 137 Tt = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
Chris@82 138 Tu = VSUB(Ts, Tt);
Chris@82 139 T1Y = VADD(Ts, Tt);
Chris@82 140 TD = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
Chris@82 141 TE = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
Chris@82 142 TF = VSUB(TD, TE);
Chris@82 143 T1Z = VADD(TD, TE);
Chris@82 144 }
Chris@82 145 {
Chris@82 146 V Tv, Tw, Ty, Tz;
Chris@82 147 Tv = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
Chris@82 148 Tw = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
Chris@82 149 Tx = VSUB(Tv, Tw);
Chris@82 150 T21 = VADD(Tv, Tw);
Chris@82 151 Ty = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
Chris@82 152 Tz = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
Chris@82 153 TA = VSUB(Ty, Tz);
Chris@82 154 T22 = VADD(Ty, Tz);
Chris@82 155 }
Chris@82 156 {
Chris@82 157 V TB, TG, T2C, T2D;
Chris@82 158 TB = VADD(Tx, TA);
Chris@82 159 TC = VFMA(LDK(KP707106781), TB, Tu);
Chris@82 160 T1s = VFNMS(LDK(KP707106781), TB, Tu);
Chris@82 161 TG = VSUB(Tx, TA);
Chris@82 162 TH = VFMA(LDK(KP707106781), TG, TF);
Chris@82 163 T1t = VFNMS(LDK(KP707106781), TG, TF);
Chris@82 164 T20 = VADD(T1Y, T1Z);
Chris@82 165 T23 = VADD(T21, T22);
Chris@82 166 T24 = VSUB(T20, T23);
Chris@82 167 T2C = VSUB(T1Y, T1Z);
Chris@82 168 T2D = VSUB(T21, T22);
Chris@82 169 T2E = VFNMS(LDK(KP414213562), T2D, T2C);
Chris@82 170 T2N = VFMA(LDK(KP414213562), T2C, T2D);
Chris@82 171 }
Chris@82 172 }
Chris@82 173 {
Chris@82 174 V Te, T2e, To, T2i, Th, T2f, Tl, T2h, Ti, Tp;
Chris@82 175 {
Chris@82 176 V Tc, Td, Tm, Tn;
Chris@82 177 Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
Chris@82 178 Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
Chris@82 179 Te = VSUB(Tc, Td);
Chris@82 180 T2e = VADD(Tc, Td);
Chris@82 181 Tm = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
Chris@82 182 Tn = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
Chris@82 183 To = VSUB(Tm, Tn);
Chris@82 184 T2i = VADD(Tn, Tm);
Chris@82 185 }
Chris@82 186 {
Chris@82 187 V Tf, Tg, Tj, Tk;
Chris@82 188 Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
Chris@82 189 Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
Chris@82 190 Th = VSUB(Tf, Tg);
Chris@82 191 T2f = VADD(Tf, Tg);
Chris@82 192 Tj = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
Chris@82 193 Tk = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
Chris@82 194 Tl = VSUB(Tj, Tk);
Chris@82 195 T2h = VADD(Tj, Tk);
Chris@82 196 }
Chris@82 197 T2g = VADD(T2e, T2f);
Chris@82 198 T2j = VADD(T2h, T2i);
Chris@82 199 Ti = VFNMS(LDK(KP414213562), Th, Te);
Chris@82 200 Tp = VFNMS(LDK(KP414213562), To, Tl);
Chris@82 201 Tq = VADD(Ti, Tp);
Chris@82 202 T1B = VSUB(Ti, Tp);
Chris@82 203 {
Chris@82 204 V T17, T18, T2y, T2z;
Chris@82 205 T17 = VFMA(LDK(KP414213562), Te, Th);
Chris@82 206 T18 = VFMA(LDK(KP414213562), Tl, To);
Chris@82 207 T19 = VSUB(T17, T18);
Chris@82 208 T1q = VADD(T17, T18);
Chris@82 209 T2y = VSUB(T2e, T2f);
Chris@82 210 T2z = VSUB(T2h, T2i);
Chris@82 211 T2A = VADD(T2y, T2z);
Chris@82 212 T2L = VSUB(T2y, T2z);
Chris@82 213 }
Chris@82 214 }
Chris@82 215 {
Chris@82 216 V T31, T32, T33, T34, T35, T36, T37, T38, T39, T3a, T3b, T3c;
Chris@82 217 {
Chris@82 218 V T2d, T2n, T2m, T2o;
Chris@82 219 {
Chris@82 220 V T1X, T2c, T2k, T2l;
Chris@82 221 T1X = VSUB(T1T, T1W);
Chris@82 222 T2c = VADD(T24, T2b);
Chris@82 223 T2d = VFNMS(LDK(KP707106781), T2c, T1X);
Chris@82 224 T2n = VFMA(LDK(KP707106781), T2c, T1X);
Chris@82 225 T2k = VSUB(T2g, T2j);
Chris@82 226 T2l = VSUB(T24, T2b);
Chris@82 227 T2m = VFNMS(LDK(KP707106781), T2l, T2k);
Chris@82 228 T2o = VFMA(LDK(KP707106781), T2l, T2k);
Chris@82 229 }
Chris@82 230 T31 = VFNMSI(T2m, T2d);
Chris@82 231 STM2(&(xo[24]), T31, ovs, &(xo[0]));
Chris@82 232 T32 = VFNMSI(T2o, T2n);
Chris@82 233 STM2(&(xo[56]), T32, ovs, &(xo[0]));
Chris@82 234 T33 = VFMAI(T2m, T2d);
Chris@82 235 STM2(&(xo[40]), T33, ovs, &(xo[0]));
Chris@82 236 T34 = VFMAI(T2o, T2n);
Chris@82 237 STM2(&(xo[8]), T34, ovs, &(xo[0]));
Chris@82 238 }
Chris@82 239 {
Chris@82 240 V T2r, T2v, T2u, T2w;
Chris@82 241 {
Chris@82 242 V T2p, T2q, T2s, T2t;
Chris@82 243 T2p = VADD(T1T, T1W);
Chris@82 244 T2q = VADD(T2g, T2j);
Chris@82 245 T2r = VSUB(T2p, T2q);
Chris@82 246 T2v = VADD(T2p, T2q);
Chris@82 247 T2s = VADD(T20, T23);
Chris@82 248 T2t = VADD(T27, T2a);
Chris@82 249 T2u = VSUB(T2s, T2t);
Chris@82 250 T2w = VADD(T2s, T2t);
Chris@82 251 }
Chris@82 252 T35 = VFNMSI(T2u, T2r);
Chris@82 253 STM2(&(xo[48]), T35, ovs, &(xo[0]));
Chris@82 254 T36 = VADD(T2v, T2w);
Chris@82 255 STM2(&(xo[0]), T36, ovs, &(xo[0]));
Chris@82 256 T37 = VFMAI(T2u, T2r);
Chris@82 257 STM2(&(xo[16]), T37, ovs, &(xo[0]));
Chris@82 258 T38 = VSUB(T2v, T2w);
Chris@82 259 STM2(&(xo[32]), T38, ovs, &(xo[0]));
Chris@82 260 }
Chris@82 261 {
Chris@82 262 V T2V, T2Z, T2Y, T30;
Chris@82 263 {
Chris@82 264 V T2T, T2U, T2W, T2X;
Chris@82 265 T2T = VFNMS(LDK(KP707106781), T2A, T2x);
Chris@82 266 T2U = VADD(T2N, T2O);
Chris@82 267 T2V = VFNMS(LDK(KP923879532), T2U, T2T);
Chris@82 268 T2Z = VFMA(LDK(KP923879532), T2U, T2T);
Chris@82 269 T2W = VFNMS(LDK(KP707106781), T2L, T2K);
Chris@82 270 T2X = VSUB(T2E, T2H);
Chris@82 271 T2Y = VFMA(LDK(KP923879532), T2X, T2W);
Chris@82 272 T30 = VFNMS(LDK(KP923879532), T2X, T2W);
Chris@82 273 }
Chris@82 274 T39 = VFMAI(T2Y, T2V);
Chris@82 275 STM2(&(xo[20]), T39, ovs, &(xo[0]));
Chris@82 276 T3a = VFMAI(T30, T2Z);
Chris@82 277 STM2(&(xo[52]), T3a, ovs, &(xo[0]));
Chris@82 278 T3b = VFNMSI(T2Y, T2V);
Chris@82 279 STM2(&(xo[44]), T3b, ovs, &(xo[0]));
Chris@82 280 T3c = VFNMSI(T30, T2Z);
Chris@82 281 STM2(&(xo[12]), T3c, ovs, &(xo[0]));
Chris@82 282 }
Chris@82 283 {
Chris@82 284 V T3d, T3e, T3f, T3g;
Chris@82 285 {
Chris@82 286 V T2J, T2R, T2Q, T2S;
Chris@82 287 {
Chris@82 288 V T2B, T2I, T2M, T2P;
Chris@82 289 T2B = VFMA(LDK(KP707106781), T2A, T2x);
Chris@82 290 T2I = VADD(T2E, T2H);
Chris@82 291 T2J = VFNMS(LDK(KP923879532), T2I, T2B);
Chris@82 292 T2R = VFMA(LDK(KP923879532), T2I, T2B);
Chris@82 293 T2M = VFMA(LDK(KP707106781), T2L, T2K);
Chris@82 294 T2P = VSUB(T2N, T2O);
Chris@82 295 T2Q = VFNMS(LDK(KP923879532), T2P, T2M);
Chris@82 296 T2S = VFMA(LDK(KP923879532), T2P, T2M);
Chris@82 297 }
Chris@82 298 T3d = VFNMSI(T2Q, T2J);
Chris@82 299 STM2(&(xo[28]), T3d, ovs, &(xo[0]));
Chris@82 300 T3e = VFMAI(T2S, T2R);
Chris@82 301 STM2(&(xo[4]), T3e, ovs, &(xo[0]));
Chris@82 302 T3f = VFMAI(T2Q, T2J);
Chris@82 303 STM2(&(xo[36]), T3f, ovs, &(xo[0]));
Chris@82 304 T3g = VFNMSI(T2S, T2R);
Chris@82 305 STM2(&(xo[60]), T3g, ovs, &(xo[0]));
Chris@82 306 }
Chris@82 307 {
Chris@82 308 V T1r, T1C, T1M, T1J, T1F, T1K, T1y, T1N;
Chris@82 309 T1r = VFMA(LDK(KP923879532), T1q, T1p);
Chris@82 310 T1C = VFNMS(LDK(KP923879532), T1B, T1A);
Chris@82 311 T1M = VFMA(LDK(KP923879532), T1B, T1A);
Chris@82 312 T1J = VFNMS(LDK(KP923879532), T1q, T1p);
Chris@82 313 {
Chris@82 314 V T1D, T1E, T1u, T1x;
Chris@82 315 T1D = VFNMS(LDK(KP668178637), T1s, T1t);
Chris@82 316 T1E = VFNMS(LDK(KP668178637), T1v, T1w);
Chris@82 317 T1F = VSUB(T1D, T1E);
Chris@82 318 T1K = VADD(T1D, T1E);
Chris@82 319 T1u = VFMA(LDK(KP668178637), T1t, T1s);
Chris@82 320 T1x = VFMA(LDK(KP668178637), T1w, T1v);
Chris@82 321 T1y = VADD(T1u, T1x);
Chris@82 322 T1N = VSUB(T1u, T1x);
Chris@82 323 }
Chris@82 324 {
Chris@82 325 V T1z, T1G, T3h, T3i;
Chris@82 326 T1z = VFNMS(LDK(KP831469612), T1y, T1r);
Chris@82 327 T1G = VFNMS(LDK(KP831469612), T1F, T1C);
Chris@82 328 T3h = VFNMSI(T1G, T1z);
Chris@82 329 STM2(&(xo[38]), T3h, ovs, &(xo[2]));
Chris@82 330 STN2(&(xo[36]), T3f, T3h, ovs);
Chris@82 331 T3i = VFMAI(T1G, T1z);
Chris@82 332 STM2(&(xo[26]), T3i, ovs, &(xo[2]));
Chris@82 333 STN2(&(xo[24]), T31, T3i, ovs);
Chris@82 334 }
Chris@82 335 {
Chris@82 336 V T1P, T1Q, T3j, T3k;
Chris@82 337 T1P = VFNMS(LDK(KP831469612), T1K, T1J);
Chris@82 338 T1Q = VFMA(LDK(KP831469612), T1N, T1M);
Chris@82 339 T3j = VFMAI(T1Q, T1P);
Chris@82 340 STM2(&(xo[10]), T3j, ovs, &(xo[2]));
Chris@82 341 STN2(&(xo[8]), T34, T3j, ovs);
Chris@82 342 T3k = VFNMSI(T1Q, T1P);
Chris@82 343 STM2(&(xo[54]), T3k, ovs, &(xo[2]));
Chris@82 344 STN2(&(xo[52]), T3a, T3k, ovs);
Chris@82 345 }
Chris@82 346 {
Chris@82 347 V T1H, T1I, T3l, T3m;
Chris@82 348 T1H = VFMA(LDK(KP831469612), T1y, T1r);
Chris@82 349 T1I = VFMA(LDK(KP831469612), T1F, T1C);
Chris@82 350 T3l = VFNMSI(T1I, T1H);
Chris@82 351 STM2(&(xo[6]), T3l, ovs, &(xo[2]));
Chris@82 352 STN2(&(xo[4]), T3e, T3l, ovs);
Chris@82 353 T3m = VFMAI(T1I, T1H);
Chris@82 354 STM2(&(xo[58]), T3m, ovs, &(xo[2]));
Chris@82 355 STN2(&(xo[56]), T32, T3m, ovs);
Chris@82 356 }
Chris@82 357 {
Chris@82 358 V T1L, T1O, T3n, T3o;
Chris@82 359 T1L = VFMA(LDK(KP831469612), T1K, T1J);
Chris@82 360 T1O = VFNMS(LDK(KP831469612), T1N, T1M);
Chris@82 361 T3n = VFNMSI(T1O, T1L);
Chris@82 362 STM2(&(xo[22]), T3n, ovs, &(xo[2]));
Chris@82 363 STN2(&(xo[20]), T39, T3n, ovs);
Chris@82 364 T3o = VFMAI(T1O, T1L);
Chris@82 365 STM2(&(xo[42]), T3o, ovs, &(xo[2]));
Chris@82 366 STN2(&(xo[40]), T33, T3o, ovs);
Chris@82 367 }
Chris@82 368 }
Chris@82 369 {
Chris@82 370 V Tr, T1a, T1k, T1h, T1d, T1i, T10, T1l;
Chris@82 371 Tr = VFMA(LDK(KP923879532), Tq, Tb);
Chris@82 372 T1a = VFMA(LDK(KP923879532), T19, T16);
Chris@82 373 T1k = VFNMS(LDK(KP923879532), T19, T16);
Chris@82 374 T1h = VFNMS(LDK(KP923879532), Tq, Tb);
Chris@82 375 {
Chris@82 376 V T1b, T1c, TI, TZ;
Chris@82 377 T1b = VFMA(LDK(KP198912367), TC, TH);
Chris@82 378 T1c = VFMA(LDK(KP198912367), TT, TY);
Chris@82 379 T1d = VSUB(T1b, T1c);
Chris@82 380 T1i = VADD(T1b, T1c);
Chris@82 381 TI = VFNMS(LDK(KP198912367), TH, TC);
Chris@82 382 TZ = VFNMS(LDK(KP198912367), TY, TT);
Chris@82 383 T10 = VADD(TI, TZ);
Chris@82 384 T1l = VSUB(TI, TZ);
Chris@82 385 }
Chris@82 386 {
Chris@82 387 V T11, T1e, T3p, T3q;
Chris@82 388 T11 = VFNMS(LDK(KP980785280), T10, Tr);
Chris@82 389 T1e = VFNMS(LDK(KP980785280), T1d, T1a);
Chris@82 390 T3p = VFNMSI(T1e, T11);
Chris@82 391 STM2(&(xo[30]), T3p, ovs, &(xo[2]));
Chris@82 392 STN2(&(xo[28]), T3d, T3p, ovs);
Chris@82 393 T3q = VFMAI(T1e, T11);
Chris@82 394 STM2(&(xo[34]), T3q, ovs, &(xo[2]));
Chris@82 395 STN2(&(xo[32]), T38, T3q, ovs);
Chris@82 396 }
Chris@82 397 {
Chris@82 398 V T1n, T1o, T3r, T3s;
Chris@82 399 T1n = VFMA(LDK(KP980785280), T1i, T1h);
Chris@82 400 T1o = VFNMS(LDK(KP980785280), T1l, T1k);
Chris@82 401 T3r = VFNMSI(T1o, T1n);
Chris@82 402 STM2(&(xo[14]), T3r, ovs, &(xo[2]));
Chris@82 403 STN2(&(xo[12]), T3c, T3r, ovs);
Chris@82 404 T3s = VFMAI(T1o, T1n);
Chris@82 405 STM2(&(xo[50]), T3s, ovs, &(xo[2]));
Chris@82 406 STN2(&(xo[48]), T35, T3s, ovs);
Chris@82 407 }
Chris@82 408 {
Chris@82 409 V T1f, T1g, T3t, T3u;
Chris@82 410 T1f = VFMA(LDK(KP980785280), T10, Tr);
Chris@82 411 T1g = VFMA(LDK(KP980785280), T1d, T1a);
Chris@82 412 T3t = VFNMSI(T1g, T1f);
Chris@82 413 STM2(&(xo[62]), T3t, ovs, &(xo[2]));
Chris@82 414 STN2(&(xo[60]), T3g, T3t, ovs);
Chris@82 415 T3u = VFMAI(T1g, T1f);
Chris@82 416 STM2(&(xo[2]), T3u, ovs, &(xo[2]));
Chris@82 417 STN2(&(xo[0]), T36, T3u, ovs);
Chris@82 418 }
Chris@82 419 {
Chris@82 420 V T1j, T1m, T3v, T3w;
Chris@82 421 T1j = VFNMS(LDK(KP980785280), T1i, T1h);
Chris@82 422 T1m = VFMA(LDK(KP980785280), T1l, T1k);
Chris@82 423 T3v = VFMAI(T1m, T1j);
Chris@82 424 STM2(&(xo[18]), T3v, ovs, &(xo[2]));
Chris@82 425 STN2(&(xo[16]), T37, T3v, ovs);
Chris@82 426 T3w = VFNMSI(T1m, T1j);
Chris@82 427 STM2(&(xo[46]), T3w, ovs, &(xo[2]));
Chris@82 428 STN2(&(xo[44]), T3b, T3w, ovs);
Chris@82 429 }
Chris@82 430 }
Chris@82 431 }
Chris@82 432 }
Chris@82 433 }
Chris@82 434 }
Chris@82 435 VLEAVE();
Chris@82 436 }
Chris@82 437
Chris@82 438 static const kdft_desc desc = { 32, XSIMD_STRING("n2bv_32"), {88, 0, 98, 0}, &GENUS, 0, 2, 0, 0 };
Chris@82 439
Chris@82 440 void XSIMD(codelet_n2bv_32) (planner *p) {
Chris@82 441 X(kdft_register) (p, n2bv_32, &desc);
Chris@82 442 }
Chris@82 443
Chris@82 444 #else
Chris@82 445
Chris@82 446 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 32 -name n2bv_32 -with-ostride 2 -include dft/simd/n2b.h -store-multiple 2 */
Chris@82 447
Chris@82 448 /*
Chris@82 449 * This function contains 186 FP additions, 42 FP multiplications,
Chris@82 450 * (or, 170 additions, 26 multiplications, 16 fused multiply/add),
Chris@82 451 * 72 stack variables, 7 constants, and 80 memory accesses
Chris@82 452 */
Chris@82 453 #include "dft/simd/n2b.h"
Chris@82 454
Chris@82 455 static void n2bv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
Chris@82 456 {
Chris@82 457 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
Chris@82 458 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
Chris@82 459 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
Chris@82 460 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
Chris@82 461 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
Chris@82 462 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@82 463 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@82 464 {
Chris@82 465 INT i;
Chris@82 466 const R *xi;
Chris@82 467 R *xo;
Chris@82 468 xi = ii;
Chris@82 469 xo = io;
Chris@82 470 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@82 471 V T2f, T2k, T2N, T2M, T19, T1B, Tb, T1p, TT, T1v, TY, T1w, T2E, T2F, T2G;
Chris@82 472 V T24, T2o, TC, T1s, TH, T1t, T2B, T2C, T2D, T1X, T2n, T2I, T2J, Tq, T1A;
Chris@82 473 V T14, T1q, T2c, T2l;
Chris@82 474 {
Chris@82 475 V T3, T2i, T18, T2j, T6, T2d, T9, T2e, T15, Ta;
Chris@82 476 {
Chris@82 477 V T1, T2, T16, T17;
Chris@82 478 T1 = LD(&(xi[0]), ivs, &(xi[0]));
Chris@82 479 T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
Chris@82 480 T3 = VSUB(T1, T2);
Chris@82 481 T2i = VADD(T1, T2);
Chris@82 482 T16 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
Chris@82 483 T17 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
Chris@82 484 T18 = VSUB(T16, T17);
Chris@82 485 T2j = VADD(T16, T17);
Chris@82 486 }
Chris@82 487 {
Chris@82 488 V T4, T5, T7, T8;
Chris@82 489 T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
Chris@82 490 T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
Chris@82 491 T6 = VSUB(T4, T5);
Chris@82 492 T2d = VADD(T4, T5);
Chris@82 493 T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
Chris@82 494 T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
Chris@82 495 T9 = VSUB(T7, T8);
Chris@82 496 T2e = VADD(T7, T8);
Chris@82 497 }
Chris@82 498 T2f = VSUB(T2d, T2e);
Chris@82 499 T2k = VSUB(T2i, T2j);
Chris@82 500 T2N = VADD(T2d, T2e);
Chris@82 501 T2M = VADD(T2i, T2j);
Chris@82 502 T15 = VMUL(LDK(KP707106781), VSUB(T6, T9));
Chris@82 503 T19 = VSUB(T15, T18);
Chris@82 504 T1B = VADD(T18, T15);
Chris@82 505 Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
Chris@82 506 Tb = VSUB(T3, Ta);
Chris@82 507 T1p = VADD(T3, Ta);
Chris@82 508 }
Chris@82 509 {
Chris@82 510 V TL, T21, TW, T1Y, TO, T22, TS, T1Z;
Chris@82 511 {
Chris@82 512 V TJ, TK, TU, TV;
Chris@82 513 TJ = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
Chris@82 514 TK = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
Chris@82 515 TL = VSUB(TJ, TK);
Chris@82 516 T21 = VADD(TJ, TK);
Chris@82 517 TU = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
Chris@82 518 TV = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
Chris@82 519 TW = VSUB(TU, TV);
Chris@82 520 T1Y = VADD(TU, TV);
Chris@82 521 }
Chris@82 522 {
Chris@82 523 V TM, TN, TQ, TR;
Chris@82 524 TM = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
Chris@82 525 TN = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
Chris@82 526 TO = VSUB(TM, TN);
Chris@82 527 T22 = VADD(TM, TN);
Chris@82 528 TQ = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
Chris@82 529 TR = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
Chris@82 530 TS = VSUB(TQ, TR);
Chris@82 531 T1Z = VADD(TQ, TR);
Chris@82 532 }
Chris@82 533 {
Chris@82 534 V TP, TX, T20, T23;
Chris@82 535 TP = VMUL(LDK(KP707106781), VSUB(TL, TO));
Chris@82 536 TT = VSUB(TP, TS);
Chris@82 537 T1v = VADD(TS, TP);
Chris@82 538 TX = VMUL(LDK(KP707106781), VADD(TL, TO));
Chris@82 539 TY = VSUB(TW, TX);
Chris@82 540 T1w = VADD(TW, TX);
Chris@82 541 T2E = VADD(T1Y, T1Z);
Chris@82 542 T2F = VADD(T21, T22);
Chris@82 543 T2G = VSUB(T2E, T2F);
Chris@82 544 T20 = VSUB(T1Y, T1Z);
Chris@82 545 T23 = VSUB(T21, T22);
Chris@82 546 T24 = VFMA(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T23));
Chris@82 547 T2o = VFNMS(LDK(KP382683432), T20, VMUL(LDK(KP923879532), T23));
Chris@82 548 }
Chris@82 549 }
Chris@82 550 {
Chris@82 551 V Tu, T1U, TF, T1R, Tx, T1V, TB, T1S;
Chris@82 552 {
Chris@82 553 V Ts, Tt, TD, TE;
Chris@82 554 Ts = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
Chris@82 555 Tt = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
Chris@82 556 Tu = VSUB(Ts, Tt);
Chris@82 557 T1U = VADD(Ts, Tt);
Chris@82 558 TD = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
Chris@82 559 TE = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
Chris@82 560 TF = VSUB(TD, TE);
Chris@82 561 T1R = VADD(TD, TE);
Chris@82 562 }
Chris@82 563 {
Chris@82 564 V Tv, Tw, Tz, TA;
Chris@82 565 Tv = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
Chris@82 566 Tw = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
Chris@82 567 Tx = VSUB(Tv, Tw);
Chris@82 568 T1V = VADD(Tv, Tw);
Chris@82 569 Tz = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
Chris@82 570 TA = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
Chris@82 571 TB = VSUB(Tz, TA);
Chris@82 572 T1S = VADD(Tz, TA);
Chris@82 573 }
Chris@82 574 {
Chris@82 575 V Ty, TG, T1T, T1W;
Chris@82 576 Ty = VMUL(LDK(KP707106781), VSUB(Tu, Tx));
Chris@82 577 TC = VSUB(Ty, TB);
Chris@82 578 T1s = VADD(TB, Ty);
Chris@82 579 TG = VMUL(LDK(KP707106781), VADD(Tu, Tx));
Chris@82 580 TH = VSUB(TF, TG);
Chris@82 581 T1t = VADD(TF, TG);
Chris@82 582 T2B = VADD(T1R, T1S);
Chris@82 583 T2C = VADD(T1U, T1V);
Chris@82 584 T2D = VSUB(T2B, T2C);
Chris@82 585 T1T = VSUB(T1R, T1S);
Chris@82 586 T1W = VSUB(T1U, T1V);
Chris@82 587 T1X = VFNMS(LDK(KP382683432), T1W, VMUL(LDK(KP923879532), T1T));
Chris@82 588 T2n = VFMA(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1W));
Chris@82 589 }
Chris@82 590 }
Chris@82 591 {
Chris@82 592 V Te, T26, To, T29, Th, T27, Tl, T2a, Ti, Tp;
Chris@82 593 {
Chris@82 594 V Tc, Td, Tm, Tn;
Chris@82 595 Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
Chris@82 596 Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
Chris@82 597 Te = VSUB(Tc, Td);
Chris@82 598 T26 = VADD(Tc, Td);
Chris@82 599 Tm = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
Chris@82 600 Tn = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
Chris@82 601 To = VSUB(Tm, Tn);
Chris@82 602 T29 = VADD(Tm, Tn);
Chris@82 603 }
Chris@82 604 {
Chris@82 605 V Tf, Tg, Tj, Tk;
Chris@82 606 Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
Chris@82 607 Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
Chris@82 608 Th = VSUB(Tf, Tg);
Chris@82 609 T27 = VADD(Tf, Tg);
Chris@82 610 Tj = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
Chris@82 611 Tk = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
Chris@82 612 Tl = VSUB(Tj, Tk);
Chris@82 613 T2a = VADD(Tj, Tk);
Chris@82 614 }
Chris@82 615 T2I = VADD(T26, T27);
Chris@82 616 T2J = VADD(T29, T2a);
Chris@82 617 Ti = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
Chris@82 618 Tp = VFNMS(LDK(KP382683432), To, VMUL(LDK(KP923879532), Tl));
Chris@82 619 Tq = VSUB(Ti, Tp);
Chris@82 620 T1A = VADD(Ti, Tp);
Chris@82 621 {
Chris@82 622 V T12, T13, T28, T2b;
Chris@82 623 T12 = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
Chris@82 624 T13 = VFMA(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
Chris@82 625 T14 = VSUB(T12, T13);
Chris@82 626 T1q = VADD(T12, T13);
Chris@82 627 T28 = VSUB(T26, T27);
Chris@82 628 T2b = VSUB(T29, T2a);
Chris@82 629 T2c = VMUL(LDK(KP707106781), VSUB(T28, T2b));
Chris@82 630 T2l = VMUL(LDK(KP707106781), VADD(T28, T2b));
Chris@82 631 }
Chris@82 632 }
Chris@82 633 {
Chris@82 634 V T31, T32, T33, T34, T35, T36, T37, T38, T39, T3a, T3b, T3c;
Chris@82 635 {
Chris@82 636 V T2L, T2R, T2Q, T2S;
Chris@82 637 {
Chris@82 638 V T2H, T2K, T2O, T2P;
Chris@82 639 T2H = VMUL(LDK(KP707106781), VSUB(T2D, T2G));
Chris@82 640 T2K = VSUB(T2I, T2J);
Chris@82 641 T2L = VBYI(VSUB(T2H, T2K));
Chris@82 642 T2R = VBYI(VADD(T2K, T2H));
Chris@82 643 T2O = VSUB(T2M, T2N);
Chris@82 644 T2P = VMUL(LDK(KP707106781), VADD(T2D, T2G));
Chris@82 645 T2Q = VSUB(T2O, T2P);
Chris@82 646 T2S = VADD(T2O, T2P);
Chris@82 647 }
Chris@82 648 T31 = VADD(T2L, T2Q);
Chris@82 649 STM2(&(xo[24]), T31, ovs, &(xo[0]));
Chris@82 650 T32 = VSUB(T2S, T2R);
Chris@82 651 STM2(&(xo[56]), T32, ovs, &(xo[0]));
Chris@82 652 T33 = VSUB(T2Q, T2L);
Chris@82 653 STM2(&(xo[40]), T33, ovs, &(xo[0]));
Chris@82 654 T34 = VADD(T2R, T2S);
Chris@82 655 STM2(&(xo[8]), T34, ovs, &(xo[0]));
Chris@82 656 }
Chris@82 657 {
Chris@82 658 V T2h, T2r, T2q, T2s;
Chris@82 659 {
Chris@82 660 V T25, T2g, T2m, T2p;
Chris@82 661 T25 = VSUB(T1X, T24);
Chris@82 662 T2g = VSUB(T2c, T2f);
Chris@82 663 T2h = VBYI(VSUB(T25, T2g));
Chris@82 664 T2r = VBYI(VADD(T2g, T25));
Chris@82 665 T2m = VSUB(T2k, T2l);
Chris@82 666 T2p = VSUB(T2n, T2o);
Chris@82 667 T2q = VSUB(T2m, T2p);
Chris@82 668 T2s = VADD(T2m, T2p);
Chris@82 669 }
Chris@82 670 T35 = VADD(T2h, T2q);
Chris@82 671 STM2(&(xo[20]), T35, ovs, &(xo[0]));
Chris@82 672 T36 = VSUB(T2s, T2r);
Chris@82 673 STM2(&(xo[52]), T36, ovs, &(xo[0]));
Chris@82 674 T37 = VSUB(T2q, T2h);
Chris@82 675 STM2(&(xo[44]), T37, ovs, &(xo[0]));
Chris@82 676 T38 = VADD(T2r, T2s);
Chris@82 677 STM2(&(xo[12]), T38, ovs, &(xo[0]));
Chris@82 678 }
Chris@82 679 {
Chris@82 680 V T2V, T2Z, T2Y, T30;
Chris@82 681 {
Chris@82 682 V T2T, T2U, T2W, T2X;
Chris@82 683 T2T = VADD(T2M, T2N);
Chris@82 684 T2U = VADD(T2I, T2J);
Chris@82 685 T2V = VSUB(T2T, T2U);
Chris@82 686 T2Z = VADD(T2T, T2U);
Chris@82 687 T2W = VADD(T2B, T2C);
Chris@82 688 T2X = VADD(T2E, T2F);
Chris@82 689 T2Y = VBYI(VSUB(T2W, T2X));
Chris@82 690 T30 = VADD(T2W, T2X);
Chris@82 691 }
Chris@82 692 T39 = VSUB(T2V, T2Y);
Chris@82 693 STM2(&(xo[48]), T39, ovs, &(xo[0]));
Chris@82 694 T3a = VADD(T2Z, T30);
Chris@82 695 STM2(&(xo[0]), T3a, ovs, &(xo[0]));
Chris@82 696 T3b = VADD(T2V, T2Y);
Chris@82 697 STM2(&(xo[16]), T3b, ovs, &(xo[0]));
Chris@82 698 T3c = VSUB(T2Z, T30);
Chris@82 699 STM2(&(xo[32]), T3c, ovs, &(xo[0]));
Chris@82 700 }
Chris@82 701 {
Chris@82 702 V T3d, T3e, T3f, T3g;
Chris@82 703 {
Chris@82 704 V T2v, T2z, T2y, T2A;
Chris@82 705 {
Chris@82 706 V T2t, T2u, T2w, T2x;
Chris@82 707 T2t = VADD(T2k, T2l);
Chris@82 708 T2u = VADD(T1X, T24);
Chris@82 709 T2v = VADD(T2t, T2u);
Chris@82 710 T2z = VSUB(T2t, T2u);
Chris@82 711 T2w = VADD(T2f, T2c);
Chris@82 712 T2x = VADD(T2n, T2o);
Chris@82 713 T2y = VBYI(VADD(T2w, T2x));
Chris@82 714 T2A = VBYI(VSUB(T2x, T2w));
Chris@82 715 }
Chris@82 716 T3d = VSUB(T2v, T2y);
Chris@82 717 STM2(&(xo[60]), T3d, ovs, &(xo[0]));
Chris@82 718 T3e = VADD(T2z, T2A);
Chris@82 719 STM2(&(xo[28]), T3e, ovs, &(xo[0]));
Chris@82 720 T3f = VADD(T2v, T2y);
Chris@82 721 STM2(&(xo[4]), T3f, ovs, &(xo[0]));
Chris@82 722 T3g = VSUB(T2z, T2A);
Chris@82 723 STM2(&(xo[36]), T3g, ovs, &(xo[0]));
Chris@82 724 }
Chris@82 725 {
Chris@82 726 V T1r, T1C, T1M, T1K, T1F, T1N, T1y, T1J;
Chris@82 727 T1r = VSUB(T1p, T1q);
Chris@82 728 T1C = VSUB(T1A, T1B);
Chris@82 729 T1M = VADD(T1p, T1q);
Chris@82 730 T1K = VADD(T1B, T1A);
Chris@82 731 {
Chris@82 732 V T1D, T1E, T1u, T1x;
Chris@82 733 T1D = VFNMS(LDK(KP195090322), T1s, VMUL(LDK(KP980785280), T1t));
Chris@82 734 T1E = VFMA(LDK(KP195090322), T1v, VMUL(LDK(KP980785280), T1w));
Chris@82 735 T1F = VSUB(T1D, T1E);
Chris@82 736 T1N = VADD(T1D, T1E);
Chris@82 737 T1u = VFMA(LDK(KP980785280), T1s, VMUL(LDK(KP195090322), T1t));
Chris@82 738 T1x = VFNMS(LDK(KP195090322), T1w, VMUL(LDK(KP980785280), T1v));
Chris@82 739 T1y = VSUB(T1u, T1x);
Chris@82 740 T1J = VADD(T1u, T1x);
Chris@82 741 }
Chris@82 742 {
Chris@82 743 V T1z, T1G, T3h, T3i;
Chris@82 744 T1z = VADD(T1r, T1y);
Chris@82 745 T1G = VBYI(VADD(T1C, T1F));
Chris@82 746 T3h = VSUB(T1z, T1G);
Chris@82 747 STM2(&(xo[50]), T3h, ovs, &(xo[2]));
Chris@82 748 STN2(&(xo[48]), T39, T3h, ovs);
Chris@82 749 T3i = VADD(T1z, T1G);
Chris@82 750 STM2(&(xo[14]), T3i, ovs, &(xo[2]));
Chris@82 751 STN2(&(xo[12]), T38, T3i, ovs);
Chris@82 752 }
Chris@82 753 {
Chris@82 754 V T1P, T1Q, T3j, T3k;
Chris@82 755 T1P = VBYI(VADD(T1K, T1J));
Chris@82 756 T1Q = VADD(T1M, T1N);
Chris@82 757 T3j = VADD(T1P, T1Q);
Chris@82 758 STM2(&(xo[2]), T3j, ovs, &(xo[2]));
Chris@82 759 STN2(&(xo[0]), T3a, T3j, ovs);
Chris@82 760 T3k = VSUB(T1Q, T1P);
Chris@82 761 STM2(&(xo[62]), T3k, ovs, &(xo[2]));
Chris@82 762 STN2(&(xo[60]), T3d, T3k, ovs);
Chris@82 763 }
Chris@82 764 {
Chris@82 765 V T1H, T1I, T3l, T3m;
Chris@82 766 T1H = VSUB(T1r, T1y);
Chris@82 767 T1I = VBYI(VSUB(T1F, T1C));
Chris@82 768 T3l = VSUB(T1H, T1I);
Chris@82 769 STM2(&(xo[46]), T3l, ovs, &(xo[2]));
Chris@82 770 STN2(&(xo[44]), T37, T3l, ovs);
Chris@82 771 T3m = VADD(T1H, T1I);
Chris@82 772 STM2(&(xo[18]), T3m, ovs, &(xo[2]));
Chris@82 773 STN2(&(xo[16]), T3b, T3m, ovs);
Chris@82 774 }
Chris@82 775 {
Chris@82 776 V T1L, T1O, T3n, T3o;
Chris@82 777 T1L = VBYI(VSUB(T1J, T1K));
Chris@82 778 T1O = VSUB(T1M, T1N);
Chris@82 779 T3n = VADD(T1L, T1O);
Chris@82 780 STM2(&(xo[30]), T3n, ovs, &(xo[2]));
Chris@82 781 STN2(&(xo[28]), T3e, T3n, ovs);
Chris@82 782 T3o = VSUB(T1O, T1L);
Chris@82 783 STM2(&(xo[34]), T3o, ovs, &(xo[2]));
Chris@82 784 STN2(&(xo[32]), T3c, T3o, ovs);
Chris@82 785 }
Chris@82 786 }
Chris@82 787 {
Chris@82 788 V Tr, T1a, T1k, T1i, T1d, T1l, T10, T1h;
Chris@82 789 Tr = VSUB(Tb, Tq);
Chris@82 790 T1a = VSUB(T14, T19);
Chris@82 791 T1k = VADD(Tb, Tq);
Chris@82 792 T1i = VADD(T19, T14);
Chris@82 793 {
Chris@82 794 V T1b, T1c, TI, TZ;
Chris@82 795 T1b = VFNMS(LDK(KP555570233), TC, VMUL(LDK(KP831469612), TH));
Chris@82 796 T1c = VFMA(LDK(KP555570233), TT, VMUL(LDK(KP831469612), TY));
Chris@82 797 T1d = VSUB(T1b, T1c);
Chris@82 798 T1l = VADD(T1b, T1c);
Chris@82 799 TI = VFMA(LDK(KP831469612), TC, VMUL(LDK(KP555570233), TH));
Chris@82 800 TZ = VFNMS(LDK(KP555570233), TY, VMUL(LDK(KP831469612), TT));
Chris@82 801 T10 = VSUB(TI, TZ);
Chris@82 802 T1h = VADD(TI, TZ);
Chris@82 803 }
Chris@82 804 {
Chris@82 805 V T11, T1e, T3p, T3q;
Chris@82 806 T11 = VADD(Tr, T10);
Chris@82 807 T1e = VBYI(VADD(T1a, T1d));
Chris@82 808 T3p = VSUB(T11, T1e);
Chris@82 809 STM2(&(xo[54]), T3p, ovs, &(xo[2]));
Chris@82 810 STN2(&(xo[52]), T36, T3p, ovs);
Chris@82 811 T3q = VADD(T11, T1e);
Chris@82 812 STM2(&(xo[10]), T3q, ovs, &(xo[2]));
Chris@82 813 STN2(&(xo[8]), T34, T3q, ovs);
Chris@82 814 }
Chris@82 815 {
Chris@82 816 V T1n, T1o, T3r, T3s;
Chris@82 817 T1n = VBYI(VADD(T1i, T1h));
Chris@82 818 T1o = VADD(T1k, T1l);
Chris@82 819 T3r = VADD(T1n, T1o);
Chris@82 820 STM2(&(xo[6]), T3r, ovs, &(xo[2]));
Chris@82 821 STN2(&(xo[4]), T3f, T3r, ovs);
Chris@82 822 T3s = VSUB(T1o, T1n);
Chris@82 823 STM2(&(xo[58]), T3s, ovs, &(xo[2]));
Chris@82 824 STN2(&(xo[56]), T32, T3s, ovs);
Chris@82 825 }
Chris@82 826 {
Chris@82 827 V T1f, T1g, T3t, T3u;
Chris@82 828 T1f = VSUB(Tr, T10);
Chris@82 829 T1g = VBYI(VSUB(T1d, T1a));
Chris@82 830 T3t = VSUB(T1f, T1g);
Chris@82 831 STM2(&(xo[42]), T3t, ovs, &(xo[2]));
Chris@82 832 STN2(&(xo[40]), T33, T3t, ovs);
Chris@82 833 T3u = VADD(T1f, T1g);
Chris@82 834 STM2(&(xo[22]), T3u, ovs, &(xo[2]));
Chris@82 835 STN2(&(xo[20]), T35, T3u, ovs);
Chris@82 836 }
Chris@82 837 {
Chris@82 838 V T1j, T1m, T3v, T3w;
Chris@82 839 T1j = VBYI(VSUB(T1h, T1i));
Chris@82 840 T1m = VSUB(T1k, T1l);
Chris@82 841 T3v = VADD(T1j, T1m);
Chris@82 842 STM2(&(xo[26]), T3v, ovs, &(xo[2]));
Chris@82 843 STN2(&(xo[24]), T31, T3v, ovs);
Chris@82 844 T3w = VSUB(T1m, T1j);
Chris@82 845 STM2(&(xo[38]), T3w, ovs, &(xo[2]));
Chris@82 846 STN2(&(xo[36]), T3g, T3w, ovs);
Chris@82 847 }
Chris@82 848 }
Chris@82 849 }
Chris@82 850 }
Chris@82 851 }
Chris@82 852 }
Chris@82 853 VLEAVE();
Chris@82 854 }
Chris@82 855
Chris@82 856 static const kdft_desc desc = { 32, XSIMD_STRING("n2bv_32"), {170, 26, 16, 0}, &GENUS, 0, 2, 0, 0 };
Chris@82 857
Chris@82 858 void XSIMD(codelet_n2bv_32) (planner *p) {
Chris@82 859 X(kdft_register) (p, n2bv_32, &desc);
Chris@82 860 }
Chris@82 861
Chris@82 862 #endif