Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/rdft/scalar/r2cf/r2cf_14.c @ 56:af97cad61ff0

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add updated build of PortAudio for OSX
author Chris Cannam <cannam@all-day-breakfast.com>
date Tue, 03 Jan 2017 15:10:52 +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 16:46:05 EDT 2016 */
Chris@42 23
Chris@42 24 #include "codelet-rdft.h"
Chris@42 25
Chris@42 26 #ifdef HAVE_FMA
Chris@42 27
Chris@42 28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 14 -name r2cf_14 -include r2cf.h */
Chris@42 29
Chris@42 30 /*
Chris@42 31 * This function contains 62 FP additions, 36 FP multiplications,
Chris@42 32 * (or, 32 additions, 6 multiplications, 30 fused multiply/add),
Chris@42 33 * 45 stack variables, 6 constants, and 28 memory accesses
Chris@42 34 */
Chris@42 35 #include "r2cf.h"
Chris@42 36
Chris@42 37 static void r2cf_14(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
Chris@42 38 {
Chris@42 39 DK(KP900968867, +0.900968867902419126236102319507445051165919162);
Chris@42 40 DK(KP692021471, +0.692021471630095869627814897002069140197260599);
Chris@42 41 DK(KP801937735, +0.801937735804838252472204639014890102331838324);
Chris@42 42 DK(KP974927912, +0.974927912181823607018131682993931217232785801);
Chris@42 43 DK(KP356895867, +0.356895867892209443894399510021300583399127187);
Chris@42 44 DK(KP554958132, +0.554958132087371191422194871006410481067288862);
Chris@42 45 {
Chris@42 46 INT i;
Chris@42 47 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(56, rs), MAKE_VOLATILE_STRIDE(56, csr), MAKE_VOLATILE_STRIDE(56, csi)) {
Chris@42 48 E TN, T3, TG, TQ, Tx, To, TH, Td, TD, TO, Tw, Ta, TL, Ty, TT;
Chris@42 49 E TI, Tg, Tr, Te, Tf, TP, TJ;
Chris@42 50 {
Chris@42 51 E Tl, TE, Tk, Tm;
Chris@42 52 {
Chris@42 53 E T1, T2, Ti, Tj;
Chris@42 54 T1 = R0[0];
Chris@42 55 T2 = R1[WS(rs, 3)];
Chris@42 56 Ti = R0[WS(rs, 3)];
Chris@42 57 Tj = R1[WS(rs, 6)];
Chris@42 58 Tl = R0[WS(rs, 4)];
Chris@42 59 TN = T1 + T2;
Chris@42 60 T3 = T1 - T2;
Chris@42 61 TE = Ti + Tj;
Chris@42 62 Tk = Ti - Tj;
Chris@42 63 Tm = R1[0];
Chris@42 64 }
Chris@42 65 {
Chris@42 66 E T7, TC, T6, T8;
Chris@42 67 {
Chris@42 68 E T4, T5, TF, Tn;
Chris@42 69 T4 = R0[WS(rs, 1)];
Chris@42 70 T5 = R1[WS(rs, 4)];
Chris@42 71 T7 = R0[WS(rs, 6)];
Chris@42 72 TF = Tl + Tm;
Chris@42 73 Tn = Tl - Tm;
Chris@42 74 TC = T4 + T5;
Chris@42 75 T6 = T4 - T5;
Chris@42 76 TG = TE - TF;
Chris@42 77 TQ = TE + TF;
Chris@42 78 Tx = Tn - Tk;
Chris@42 79 To = Tk + Tn;
Chris@42 80 T8 = R1[WS(rs, 2)];
Chris@42 81 }
Chris@42 82 {
Chris@42 83 E Tb, Tc, TB, T9;
Chris@42 84 Tb = R0[WS(rs, 2)];
Chris@42 85 Tc = R1[WS(rs, 5)];
Chris@42 86 Te = R0[WS(rs, 5)];
Chris@42 87 TB = T7 + T8;
Chris@42 88 T9 = T7 - T8;
Chris@42 89 TH = Tb + Tc;
Chris@42 90 Td = Tb - Tc;
Chris@42 91 TD = TB - TC;
Chris@42 92 TO = TC + TB;
Chris@42 93 Tw = T6 - T9;
Chris@42 94 Ta = T6 + T9;
Chris@42 95 Tf = R1[WS(rs, 1)];
Chris@42 96 }
Chris@42 97 }
Chris@42 98 }
Chris@42 99 TL = FNMS(KP554958132, TG, TD);
Chris@42 100 Ty = FNMS(KP554958132, Tx, Tw);
Chris@42 101 TT = FNMS(KP356895867, TO, TQ);
Chris@42 102 TI = Te + Tf;
Chris@42 103 Tg = Te - Tf;
Chris@42 104 Tr = FNMS(KP356895867, Ta, To);
Chris@42 105 TP = TH + TI;
Chris@42 106 TJ = TH - TI;
Chris@42 107 {
Chris@42 108 E Th, Tv, TK, TM;
Chris@42 109 Th = Td + Tg;
Chris@42 110 Tv = Tg - Td;
Chris@42 111 TK = FMA(KP554958132, TJ, TG);
Chris@42 112 TM = FMA(KP554958132, TD, TJ);
Chris@42 113 Ci[WS(csi, 6)] = KP974927912 * (FNMS(KP801937735, TL, TJ));
Chris@42 114 {
Chris@42 115 E TR, TV, TU, Tz;
Chris@42 116 TR = FNMS(KP356895867, TQ, TP);
Chris@42 117 TV = FNMS(KP356895867, TP, TO);
Chris@42 118 TU = FNMS(KP692021471, TT, TP);
Chris@42 119 Cr[0] = TN + TO + TP + TQ;
Chris@42 120 Tz = FMA(KP554958132, Tv, Tx);
Chris@42 121 Ci[WS(csi, 1)] = KP974927912 * (FNMS(KP801937735, Ty, Tv));
Chris@42 122 {
Chris@42 123 E TA, Ts, Tt, Tp;
Chris@42 124 TA = FMA(KP554958132, Tw, Tv);
Chris@42 125 Ts = FNMS(KP692021471, Tr, Th);
Chris@42 126 Tt = FNMS(KP356895867, Th, Ta);
Chris@42 127 Tp = FNMS(KP356895867, To, Th);
Chris@42 128 Cr[WS(csr, 7)] = T3 + Ta + Th + To;
Chris@42 129 Ci[WS(csi, 2)] = KP974927912 * (FMA(KP801937735, TK, TD));
Chris@42 130 Ci[WS(csi, 4)] = KP974927912 * (FNMS(KP801937735, TM, TG));
Chris@42 131 {
Chris@42 132 E TS, TW, Tu, Tq;
Chris@42 133 TS = FNMS(KP692021471, TR, TO);
Chris@42 134 TW = FNMS(KP692021471, TV, TQ);
Chris@42 135 Cr[WS(csr, 2)] = FNMS(KP900968867, TU, TN);
Chris@42 136 Ci[WS(csi, 5)] = KP974927912 * (FMA(KP801937735, Tz, Tw));
Chris@42 137 Ci[WS(csi, 3)] = KP974927912 * (FNMS(KP801937735, TA, Tx));
Chris@42 138 Cr[WS(csr, 5)] = FNMS(KP900968867, Ts, T3);
Chris@42 139 Tu = FNMS(KP692021471, Tt, To);
Chris@42 140 Tq = FNMS(KP692021471, Tp, Ta);
Chris@42 141 Cr[WS(csr, 4)] = FNMS(KP900968867, TS, TN);
Chris@42 142 Cr[WS(csr, 6)] = FNMS(KP900968867, TW, TN);
Chris@42 143 Cr[WS(csr, 1)] = FNMS(KP900968867, Tu, T3);
Chris@42 144 Cr[WS(csr, 3)] = FNMS(KP900968867, Tq, T3);
Chris@42 145 }
Chris@42 146 }
Chris@42 147 }
Chris@42 148 }
Chris@42 149 }
Chris@42 150 }
Chris@42 151 }
Chris@42 152
Chris@42 153 static const kr2c_desc desc = { 14, "r2cf_14", {32, 6, 30, 0}, &GENUS };
Chris@42 154
Chris@42 155 void X(codelet_r2cf_14) (planner *p) {
Chris@42 156 X(kr2c_register) (p, r2cf_14, &desc);
Chris@42 157 }
Chris@42 158
Chris@42 159 #else /* HAVE_FMA */
Chris@42 160
Chris@42 161 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 14 -name r2cf_14 -include r2cf.h */
Chris@42 162
Chris@42 163 /*
Chris@42 164 * This function contains 62 FP additions, 36 FP multiplications,
Chris@42 165 * (or, 38 additions, 12 multiplications, 24 fused multiply/add),
Chris@42 166 * 29 stack variables, 6 constants, and 28 memory accesses
Chris@42 167 */
Chris@42 168 #include "r2cf.h"
Chris@42 169
Chris@42 170 static void r2cf_14(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
Chris@42 171 {
Chris@42 172 DK(KP900968867, +0.900968867902419126236102319507445051165919162);
Chris@42 173 DK(KP222520933, +0.222520933956314404288902564496794759466355569);
Chris@42 174 DK(KP623489801, +0.623489801858733530525004884004239810632274731);
Chris@42 175 DK(KP433883739, +0.433883739117558120475768332848358754609990728);
Chris@42 176 DK(KP974927912, +0.974927912181823607018131682993931217232785801);
Chris@42 177 DK(KP781831482, +0.781831482468029808708444526674057750232334519);
Chris@42 178 {
Chris@42 179 INT i;
Chris@42 180 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(56, rs), MAKE_VOLATILE_STRIDE(56, csr), MAKE_VOLATILE_STRIDE(56, csi)) {
Chris@42 181 E T3, TB, T6, Tv, Tn, Ts, Tk, Tt, Td, Ty, T9, Tw, Tg, Tz, T1;
Chris@42 182 E T2;
Chris@42 183 T1 = R0[0];
Chris@42 184 T2 = R1[WS(rs, 3)];
Chris@42 185 T3 = T1 - T2;
Chris@42 186 TB = T1 + T2;
Chris@42 187 {
Chris@42 188 E T4, T5, Tl, Tm;
Chris@42 189 T4 = R0[WS(rs, 2)];
Chris@42 190 T5 = R1[WS(rs, 5)];
Chris@42 191 T6 = T4 - T5;
Chris@42 192 Tv = T4 + T5;
Chris@42 193 Tl = R0[WS(rs, 6)];
Chris@42 194 Tm = R1[WS(rs, 2)];
Chris@42 195 Tn = Tl - Tm;
Chris@42 196 Ts = Tl + Tm;
Chris@42 197 }
Chris@42 198 {
Chris@42 199 E Ti, Tj, Tb, Tc;
Chris@42 200 Ti = R0[WS(rs, 1)];
Chris@42 201 Tj = R1[WS(rs, 4)];
Chris@42 202 Tk = Ti - Tj;
Chris@42 203 Tt = Ti + Tj;
Chris@42 204 Tb = R0[WS(rs, 3)];
Chris@42 205 Tc = R1[WS(rs, 6)];
Chris@42 206 Td = Tb - Tc;
Chris@42 207 Ty = Tb + Tc;
Chris@42 208 }
Chris@42 209 {
Chris@42 210 E T7, T8, Te, Tf;
Chris@42 211 T7 = R0[WS(rs, 5)];
Chris@42 212 T8 = R1[WS(rs, 1)];
Chris@42 213 T9 = T7 - T8;
Chris@42 214 Tw = T7 + T8;
Chris@42 215 Te = R0[WS(rs, 4)];
Chris@42 216 Tf = R1[0];
Chris@42 217 Tg = Te - Tf;
Chris@42 218 Tz = Te + Tf;
Chris@42 219 }
Chris@42 220 {
Chris@42 221 E Tp, Tr, Tq, Ta, To, Th;
Chris@42 222 Tp = Tn - Tk;
Chris@42 223 Tr = Tg - Td;
Chris@42 224 Tq = T9 - T6;
Chris@42 225 Ci[WS(csi, 1)] = FMA(KP781831482, Tp, KP974927912 * Tq) + (KP433883739 * Tr);
Chris@42 226 Ci[WS(csi, 5)] = FMA(KP433883739, Tq, KP781831482 * Tr) - (KP974927912 * Tp);
Chris@42 227 Ci[WS(csi, 3)] = FMA(KP433883739, Tp, KP974927912 * Tr) - (KP781831482 * Tq);
Chris@42 228 Ta = T6 + T9;
Chris@42 229 To = Tk + Tn;
Chris@42 230 Th = Td + Tg;
Chris@42 231 Cr[WS(csr, 3)] = FMA(KP623489801, Ta, T3) + FNMA(KP222520933, Th, KP900968867 * To);
Chris@42 232 Cr[WS(csr, 7)] = T3 + To + Ta + Th;
Chris@42 233 Cr[WS(csr, 1)] = FMA(KP623489801, To, T3) + FNMA(KP900968867, Th, KP222520933 * Ta);
Chris@42 234 Cr[WS(csr, 5)] = FMA(KP623489801, Th, T3) + FNMA(KP900968867, Ta, KP222520933 * To);
Chris@42 235 }
Chris@42 236 {
Chris@42 237 E Tu, TA, Tx, TC, TE, TD;
Chris@42 238 Tu = Ts - Tt;
Chris@42 239 TA = Ty - Tz;
Chris@42 240 Tx = Tv - Tw;
Chris@42 241 Ci[WS(csi, 2)] = FMA(KP974927912, Tu, KP433883739 * Tx) + (KP781831482 * TA);
Chris@42 242 Ci[WS(csi, 6)] = FMA(KP974927912, Tx, KP433883739 * TA) - (KP781831482 * Tu);
Chris@42 243 Ci[WS(csi, 4)] = FNMS(KP781831482, Tx, KP974927912 * TA) - (KP433883739 * Tu);
Chris@42 244 TC = Tt + Ts;
Chris@42 245 TE = Tv + Tw;
Chris@42 246 TD = Ty + Tz;
Chris@42 247 Cr[WS(csr, 6)] = FMA(KP623489801, TC, TB) + FNMA(KP900968867, TD, KP222520933 * TE);
Chris@42 248 Cr[WS(csr, 2)] = FMA(KP623489801, TD, TB) + FNMA(KP900968867, TE, KP222520933 * TC);
Chris@42 249 Cr[WS(csr, 4)] = FMA(KP623489801, TE, TB) + FNMA(KP222520933, TD, KP900968867 * TC);
Chris@42 250 Cr[0] = TB + TC + TE + TD;
Chris@42 251 }
Chris@42 252 }
Chris@42 253 }
Chris@42 254 }
Chris@42 255
Chris@42 256 static const kr2c_desc desc = { 14, "r2cf_14", {38, 12, 24, 0}, &GENUS };
Chris@42 257
Chris@42 258 void X(codelet_r2cf_14) (planner *p) {
Chris@42 259 X(kr2c_register) (p, r2cf_14, &desc);
Chris@42 260 }
Chris@42 261
Chris@42 262 #endif /* HAVE_FMA */