Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/dft/scalar/codelets/n1_7.c @ 84:08ae793730bd

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add null config files
author Chris Cannam
date Mon, 02 Mar 2020 14:03:47 +0000
parents d0c2a83c1364
children
rev   line source
Chris@82 1 /*
Chris@82 2 * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@82 3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@82 4 *
Chris@82 5 * This program is free software; you can redistribute it and/or modify
Chris@82 6 * it under the terms of the GNU General Public License as published by
Chris@82 7 * the Free Software Foundation; either version 2 of the License, or
Chris@82 8 * (at your option) any later version.
Chris@82 9 *
Chris@82 10 * This program is distributed in the hope that it will be useful,
Chris@82 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@82 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
Chris@82 13 * GNU General Public License for more details.
Chris@82 14 *
Chris@82 15 * You should have received a copy of the GNU General Public License
Chris@82 16 * along with this program; if not, write to the Free Software
Chris@82 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Chris@82 18 *
Chris@82 19 */
Chris@82 20
Chris@82 21 /* This file was automatically generated --- DO NOT EDIT */
Chris@82 22 /* Generated on Thu May 24 08:04:10 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 7 -name n1_7 -include dft/scalar/n.h */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 60 FP additions, 42 FP multiplications,
Chris@82 32 * (or, 18 additions, 0 multiplications, 42 fused multiply/add),
Chris@82 33 * 41 stack variables, 6 constants, and 28 memory accesses
Chris@82 34 */
Chris@82 35 #include "dft/scalar/n.h"
Chris@82 36
Chris@82 37 static void n1_7(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(KP974927912, +0.974927912181823607018131682993931217232785801);
Chris@82 40 DK(KP900968867, +0.900968867902419126236102319507445051165919162);
Chris@82 41 DK(KP692021471, +0.692021471630095869627814897002069140197260599);
Chris@82 42 DK(KP801937735, +0.801937735804838252472204639014890102331838324);
Chris@82 43 DK(KP554958132, +0.554958132087371191422194871006410481067288862);
Chris@82 44 DK(KP356895867, +0.356895867892209443894399510021300583399127187);
Chris@82 45 {
Chris@82 46 INT i;
Chris@82 47 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(28, is), MAKE_VOLATILE_STRIDE(28, os)) {
Chris@82 48 E T1, Tz, T4, TI, Ta, TG, T7, TH, Tb, Tp, TT, TO, TJ, Tu, Tg;
Chris@82 49 E TB, Tm, TC, Tj, TA, Tn, Ts, TQ, TL, TD, Tx;
Chris@82 50 T1 = ri[0];
Chris@82 51 Tz = ii[0];
Chris@82 52 {
Chris@82 53 E T2, T3, Te, Tf;
Chris@82 54 T2 = ri[WS(is, 1)];
Chris@82 55 T3 = ri[WS(is, 6)];
Chris@82 56 T4 = T2 + T3;
Chris@82 57 TI = T3 - T2;
Chris@82 58 {
Chris@82 59 E T8, T9, T5, T6;
Chris@82 60 T8 = ri[WS(is, 3)];
Chris@82 61 T9 = ri[WS(is, 4)];
Chris@82 62 Ta = T8 + T9;
Chris@82 63 TG = T9 - T8;
Chris@82 64 T5 = ri[WS(is, 2)];
Chris@82 65 T6 = ri[WS(is, 5)];
Chris@82 66 T7 = T5 + T6;
Chris@82 67 TH = T6 - T5;
Chris@82 68 }
Chris@82 69 Tb = FNMS(KP356895867, T7, T4);
Chris@82 70 Tp = FNMS(KP356895867, T4, Ta);
Chris@82 71 TT = FMA(KP554958132, TG, TI);
Chris@82 72 TO = FMA(KP554958132, TH, TG);
Chris@82 73 TJ = FNMS(KP554958132, TI, TH);
Chris@82 74 Tu = FNMS(KP356895867, Ta, T7);
Chris@82 75 Te = ii[WS(is, 2)];
Chris@82 76 Tf = ii[WS(is, 5)];
Chris@82 77 Tg = Te - Tf;
Chris@82 78 TB = Te + Tf;
Chris@82 79 {
Chris@82 80 E Tk, Tl, Th, Ti;
Chris@82 81 Tk = ii[WS(is, 3)];
Chris@82 82 Tl = ii[WS(is, 4)];
Chris@82 83 Tm = Tk - Tl;
Chris@82 84 TC = Tk + Tl;
Chris@82 85 Th = ii[WS(is, 1)];
Chris@82 86 Ti = ii[WS(is, 6)];
Chris@82 87 Tj = Th - Ti;
Chris@82 88 TA = Th + Ti;
Chris@82 89 }
Chris@82 90 Tn = FMA(KP554958132, Tm, Tj);
Chris@82 91 Ts = FMA(KP554958132, Tg, Tm);
Chris@82 92 TQ = FNMS(KP356895867, TB, TA);
Chris@82 93 TL = FNMS(KP356895867, TA, TC);
Chris@82 94 TD = FNMS(KP356895867, TC, TB);
Chris@82 95 Tx = FNMS(KP554958132, Tj, Tg);
Chris@82 96 }
Chris@82 97 ro[0] = T1 + T4 + T7 + Ta;
Chris@82 98 io[0] = Tz + TA + TB + TC;
Chris@82 99 {
Chris@82 100 E To, Td, Tc, TU, TS, TR;
Chris@82 101 To = FMA(KP801937735, Tn, Tg);
Chris@82 102 Tc = FNMS(KP692021471, Tb, Ta);
Chris@82 103 Td = FNMS(KP900968867, Tc, T1);
Chris@82 104 ro[WS(os, 6)] = FNMS(KP974927912, To, Td);
Chris@82 105 ro[WS(os, 1)] = FMA(KP974927912, To, Td);
Chris@82 106 TU = FMA(KP801937735, TT, TH);
Chris@82 107 TR = FNMS(KP692021471, TQ, TC);
Chris@82 108 TS = FNMS(KP900968867, TR, Tz);
Chris@82 109 io[WS(os, 1)] = FMA(KP974927912, TU, TS);
Chris@82 110 io[WS(os, 6)] = FNMS(KP974927912, TU, TS);
Chris@82 111 }
Chris@82 112 {
Chris@82 113 E Tt, Tr, Tq, TP, TN, TM;
Chris@82 114 Tt = FNMS(KP801937735, Ts, Tj);
Chris@82 115 Tq = FNMS(KP692021471, Tp, T7);
Chris@82 116 Tr = FNMS(KP900968867, Tq, T1);
Chris@82 117 ro[WS(os, 5)] = FNMS(KP974927912, Tt, Tr);
Chris@82 118 ro[WS(os, 2)] = FMA(KP974927912, Tt, Tr);
Chris@82 119 TP = FNMS(KP801937735, TO, TI);
Chris@82 120 TM = FNMS(KP692021471, TL, TB);
Chris@82 121 TN = FNMS(KP900968867, TM, Tz);
Chris@82 122 io[WS(os, 2)] = FMA(KP974927912, TP, TN);
Chris@82 123 io[WS(os, 5)] = FNMS(KP974927912, TP, TN);
Chris@82 124 }
Chris@82 125 {
Chris@82 126 E Ty, Tw, Tv, TK, TF, TE;
Chris@82 127 Ty = FNMS(KP801937735, Tx, Tm);
Chris@82 128 Tv = FNMS(KP692021471, Tu, T4);
Chris@82 129 Tw = FNMS(KP900968867, Tv, T1);
Chris@82 130 ro[WS(os, 4)] = FNMS(KP974927912, Ty, Tw);
Chris@82 131 ro[WS(os, 3)] = FMA(KP974927912, Ty, Tw);
Chris@82 132 TK = FNMS(KP801937735, TJ, TG);
Chris@82 133 TE = FNMS(KP692021471, TD, TA);
Chris@82 134 TF = FNMS(KP900968867, TE, Tz);
Chris@82 135 io[WS(os, 3)] = FMA(KP974927912, TK, TF);
Chris@82 136 io[WS(os, 4)] = FNMS(KP974927912, TK, TF);
Chris@82 137 }
Chris@82 138 }
Chris@82 139 }
Chris@82 140 }
Chris@82 141
Chris@82 142 static const kdft_desc desc = { 7, "n1_7", {18, 0, 42, 0}, &GENUS, 0, 0, 0, 0 };
Chris@82 143
Chris@82 144 void X(codelet_n1_7) (planner *p) {
Chris@82 145 X(kdft_register) (p, n1_7, &desc);
Chris@82 146 }
Chris@82 147
Chris@82 148 #else
Chris@82 149
Chris@82 150 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 7 -name n1_7 -include dft/scalar/n.h */
Chris@82 151
Chris@82 152 /*
Chris@82 153 * This function contains 60 FP additions, 36 FP multiplications,
Chris@82 154 * (or, 36 additions, 12 multiplications, 24 fused multiply/add),
Chris@82 155 * 25 stack variables, 6 constants, and 28 memory accesses
Chris@82 156 */
Chris@82 157 #include "dft/scalar/n.h"
Chris@82 158
Chris@82 159 static void n1_7(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
Chris@82 160 {
Chris@82 161 DK(KP222520933, +0.222520933956314404288902564496794759466355569);
Chris@82 162 DK(KP900968867, +0.900968867902419126236102319507445051165919162);
Chris@82 163 DK(KP623489801, +0.623489801858733530525004884004239810632274731);
Chris@82 164 DK(KP433883739, +0.433883739117558120475768332848358754609990728);
Chris@82 165 DK(KP781831482, +0.781831482468029808708444526674057750232334519);
Chris@82 166 DK(KP974927912, +0.974927912181823607018131682993931217232785801);
Chris@82 167 {
Chris@82 168 INT i;
Chris@82 169 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(28, is), MAKE_VOLATILE_STRIDE(28, os)) {
Chris@82 170 E T1, Tu, T4, Tq, Te, Tx, T7, Ts, Tk, Tv, Ta, Tr, Th, Tw;
Chris@82 171 T1 = ri[0];
Chris@82 172 Tu = ii[0];
Chris@82 173 {
Chris@82 174 E T2, T3, Tc, Td;
Chris@82 175 T2 = ri[WS(is, 1)];
Chris@82 176 T3 = ri[WS(is, 6)];
Chris@82 177 T4 = T2 + T3;
Chris@82 178 Tq = T3 - T2;
Chris@82 179 Tc = ii[WS(is, 1)];
Chris@82 180 Td = ii[WS(is, 6)];
Chris@82 181 Te = Tc - Td;
Chris@82 182 Tx = Tc + Td;
Chris@82 183 }
Chris@82 184 {
Chris@82 185 E T5, T6, Ti, Tj;
Chris@82 186 T5 = ri[WS(is, 2)];
Chris@82 187 T6 = ri[WS(is, 5)];
Chris@82 188 T7 = T5 + T6;
Chris@82 189 Ts = T6 - T5;
Chris@82 190 Ti = ii[WS(is, 2)];
Chris@82 191 Tj = ii[WS(is, 5)];
Chris@82 192 Tk = Ti - Tj;
Chris@82 193 Tv = Ti + Tj;
Chris@82 194 }
Chris@82 195 {
Chris@82 196 E T8, T9, Tf, Tg;
Chris@82 197 T8 = ri[WS(is, 3)];
Chris@82 198 T9 = ri[WS(is, 4)];
Chris@82 199 Ta = T8 + T9;
Chris@82 200 Tr = T9 - T8;
Chris@82 201 Tf = ii[WS(is, 3)];
Chris@82 202 Tg = ii[WS(is, 4)];
Chris@82 203 Th = Tf - Tg;
Chris@82 204 Tw = Tf + Tg;
Chris@82 205 }
Chris@82 206 ro[0] = T1 + T4 + T7 + Ta;
Chris@82 207 io[0] = Tu + Tx + Tv + Tw;
Chris@82 208 {
Chris@82 209 E Tl, Tb, TB, TC;
Chris@82 210 Tl = FNMS(KP781831482, Th, KP974927912 * Te) - (KP433883739 * Tk);
Chris@82 211 Tb = FMA(KP623489801, Ta, T1) + FNMA(KP900968867, T7, KP222520933 * T4);
Chris@82 212 ro[WS(os, 5)] = Tb - Tl;
Chris@82 213 ro[WS(os, 2)] = Tb + Tl;
Chris@82 214 TB = FNMS(KP781831482, Tr, KP974927912 * Tq) - (KP433883739 * Ts);
Chris@82 215 TC = FMA(KP623489801, Tw, Tu) + FNMA(KP900968867, Tv, KP222520933 * Tx);
Chris@82 216 io[WS(os, 2)] = TB + TC;
Chris@82 217 io[WS(os, 5)] = TC - TB;
Chris@82 218 }
Chris@82 219 {
Chris@82 220 E Tn, Tm, Tz, TA;
Chris@82 221 Tn = FMA(KP781831482, Te, KP974927912 * Tk) + (KP433883739 * Th);
Chris@82 222 Tm = FMA(KP623489801, T4, T1) + FNMA(KP900968867, Ta, KP222520933 * T7);
Chris@82 223 ro[WS(os, 6)] = Tm - Tn;
Chris@82 224 ro[WS(os, 1)] = Tm + Tn;
Chris@82 225 Tz = FMA(KP781831482, Tq, KP974927912 * Ts) + (KP433883739 * Tr);
Chris@82 226 TA = FMA(KP623489801, Tx, Tu) + FNMA(KP900968867, Tw, KP222520933 * Tv);
Chris@82 227 io[WS(os, 1)] = Tz + TA;
Chris@82 228 io[WS(os, 6)] = TA - Tz;
Chris@82 229 }
Chris@82 230 {
Chris@82 231 E Tp, To, Tt, Ty;
Chris@82 232 Tp = FMA(KP433883739, Te, KP974927912 * Th) - (KP781831482 * Tk);
Chris@82 233 To = FMA(KP623489801, T7, T1) + FNMA(KP222520933, Ta, KP900968867 * T4);
Chris@82 234 ro[WS(os, 4)] = To - Tp;
Chris@82 235 ro[WS(os, 3)] = To + Tp;
Chris@82 236 Tt = FMA(KP433883739, Tq, KP974927912 * Tr) - (KP781831482 * Ts);
Chris@82 237 Ty = FMA(KP623489801, Tv, Tu) + FNMA(KP222520933, Tw, KP900968867 * Tx);
Chris@82 238 io[WS(os, 3)] = Tt + Ty;
Chris@82 239 io[WS(os, 4)] = Ty - Tt;
Chris@82 240 }
Chris@82 241 }
Chris@82 242 }
Chris@82 243 }
Chris@82 244
Chris@82 245 static const kdft_desc desc = { 7, "n1_7", {36, 12, 24, 0}, &GENUS, 0, 0, 0, 0 };
Chris@82 246
Chris@82 247 void X(codelet_n1_7) (planner *p) {
Chris@82 248 X(kdft_register) (p, n1_7, &desc);
Chris@82 249 }
Chris@82 250
Chris@82 251 #endif