Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/rdft/scalar/r2cf/r2cf_15.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 15 -name r2cf_15 -include r2cf.h */
Chris@42 29
Chris@42 30 /*
Chris@42 31 * This function contains 64 FP additions, 35 FP multiplications,
Chris@42 32 * (or, 36 additions, 7 multiplications, 28 fused multiply/add),
Chris@42 33 * 50 stack variables, 8 constants, and 30 memory accesses
Chris@42 34 */
Chris@42 35 #include "r2cf.h"
Chris@42 36
Chris@42 37 static void r2cf_15(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(KP910592997, +0.910592997310029334643087372129977886038870291);
Chris@42 40 DK(KP823639103, +0.823639103546331925877420039278190003029660514);
Chris@42 41 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@42 42 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@42 43 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@42 44 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
Chris@42 45 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@42 46 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@42 47 {
Chris@42 48 INT i;
Chris@42 49 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(60, rs), MAKE_VOLATILE_STRIDE(60, csr), MAKE_VOLATILE_STRIDE(60, csi)) {
Chris@42 50 E Tw, Tz, Tp, Ty;
Chris@42 51 {
Chris@42 52 E Ti, TF, TR, TN, TX, T11, TM, TS, Tl, TH, Tf, To, TT, TD, Tg;
Chris@42 53 E Th;
Chris@42 54 TD = R0[0];
Chris@42 55 Tg = R0[WS(rs, 5)];
Chris@42 56 Th = R1[WS(rs, 2)];
Chris@42 57 {
Chris@42 58 E Tj, Tq, Tt, Tm, T3, Tk, T4, Ta, Tr, Td, Tu, T5, TE;
Chris@42 59 Tj = R1[WS(rs, 1)];
Chris@42 60 Tq = R0[WS(rs, 3)];
Chris@42 61 Tt = R1[WS(rs, 4)];
Chris@42 62 TE = Th + Tg;
Chris@42 63 Ti = Tg - Th;
Chris@42 64 Tm = R0[WS(rs, 6)];
Chris@42 65 {
Chris@42 66 E T8, T9, T1, T2, Tb, Tc;
Chris@42 67 T1 = R0[WS(rs, 4)];
Chris@42 68 T2 = R1[WS(rs, 6)];
Chris@42 69 TF = FNMS(KP500000000, TE, TD);
Chris@42 70 TR = TD + TE;
Chris@42 71 T8 = R1[WS(rs, 5)];
Chris@42 72 T3 = T1 - T2;
Chris@42 73 Tk = T1 + T2;
Chris@42 74 T9 = R1[0];
Chris@42 75 Tb = R0[WS(rs, 7)];
Chris@42 76 Tc = R0[WS(rs, 2)];
Chris@42 77 T4 = R0[WS(rs, 1)];
Chris@42 78 Ta = T8 - T9;
Chris@42 79 Tr = T8 + T9;
Chris@42 80 Td = Tb - Tc;
Chris@42 81 Tu = Tb + Tc;
Chris@42 82 T5 = R1[WS(rs, 3)];
Chris@42 83 }
Chris@42 84 {
Chris@42 85 E Ts, Tv, Te, Tn, T7, T6, TV, TW;
Chris@42 86 TV = Tq + Tr;
Chris@42 87 Ts = FNMS(KP500000000, Tr, Tq);
Chris@42 88 Tv = FNMS(KP500000000, Tu, Tt);
Chris@42 89 TW = Tt + Tu;
Chris@42 90 Te = Ta + Td;
Chris@42 91 TN = Td - Ta;
Chris@42 92 Tn = T4 + T5;
Chris@42 93 T6 = T4 - T5;
Chris@42 94 TX = TV + TW;
Chris@42 95 T11 = TW - TV;
Chris@42 96 TM = T6 - T3;
Chris@42 97 T7 = T3 + T6;
Chris@42 98 TS = Tj + Tk;
Chris@42 99 Tl = FNMS(KP500000000, Tk, Tj);
Chris@42 100 TH = Ts + Tv;
Chris@42 101 Tw = Ts - Tv;
Chris@42 102 Tz = Te - T7;
Chris@42 103 Tf = T7 + Te;
Chris@42 104 To = FNMS(KP500000000, Tn, Tm);
Chris@42 105 TT = Tm + Tn;
Chris@42 106 }
Chris@42 107 }
Chris@42 108 {
Chris@42 109 E TO, TQ, TU, T12, TK, TI, TG;
Chris@42 110 Ci[WS(csi, 5)] = KP866025403 * (Tf - Ti);
Chris@42 111 TG = Tl + To;
Chris@42 112 Tp = Tl - To;
Chris@42 113 TO = FMA(KP618033988, TN, TM);
Chris@42 114 TQ = FNMS(KP618033988, TM, TN);
Chris@42 115 TU = TS + TT;
Chris@42 116 T12 = TS - TT;
Chris@42 117 TK = TG - TH;
Chris@42 118 TI = TG + TH;
Chris@42 119 {
Chris@42 120 E T10, TY, TL, TP, TJ, TZ;
Chris@42 121 T10 = TU - TX;
Chris@42 122 TY = TU + TX;
Chris@42 123 Cr[WS(csr, 5)] = TF + TI;
Chris@42 124 TJ = FNMS(KP250000000, TI, TF);
Chris@42 125 Ci[WS(csi, 6)] = -(KP951056516 * (FNMS(KP618033988, T11, T12)));
Chris@42 126 Ci[WS(csi, 3)] = KP951056516 * (FMA(KP618033988, T12, T11));
Chris@42 127 TL = FMA(KP559016994, TK, TJ);
Chris@42 128 TP = FNMS(KP559016994, TK, TJ);
Chris@42 129 Cr[0] = TR + TY;
Chris@42 130 TZ = FNMS(KP250000000, TY, TR);
Chris@42 131 Cr[WS(csr, 4)] = FNMS(KP823639103, TO, TL);
Chris@42 132 Cr[WS(csr, 1)] = FMA(KP823639103, TO, TL);
Chris@42 133 Cr[WS(csr, 7)] = FNMS(KP823639103, TQ, TP);
Chris@42 134 Cr[WS(csr, 2)] = FMA(KP823639103, TQ, TP);
Chris@42 135 Cr[WS(csr, 6)] = FMA(KP559016994, T10, TZ);
Chris@42 136 Cr[WS(csr, 3)] = FNMS(KP559016994, T10, TZ);
Chris@42 137 Ty = FMA(KP250000000, Tf, Ti);
Chris@42 138 }
Chris@42 139 }
Chris@42 140 }
Chris@42 141 {
Chris@42 142 E TB, Tx, TC, TA;
Chris@42 143 TB = FNMS(KP618033988, Tp, Tw);
Chris@42 144 Tx = FMA(KP618033988, Tw, Tp);
Chris@42 145 TC = FNMS(KP559016994, Tz, Ty);
Chris@42 146 TA = FMA(KP559016994, Tz, Ty);
Chris@42 147 Ci[WS(csi, 2)] = KP951056516 * (FNMS(KP910592997, TC, TB));
Chris@42 148 Ci[WS(csi, 7)] = KP951056516 * (FMA(KP910592997, TC, TB));
Chris@42 149 Ci[WS(csi, 4)] = KP951056516 * (FMA(KP910592997, TA, Tx));
Chris@42 150 Ci[WS(csi, 1)] = -(KP951056516 * (FNMS(KP910592997, TA, Tx)));
Chris@42 151 }
Chris@42 152 }
Chris@42 153 }
Chris@42 154 }
Chris@42 155
Chris@42 156 static const kr2c_desc desc = { 15, "r2cf_15", {36, 7, 28, 0}, &GENUS };
Chris@42 157
Chris@42 158 void X(codelet_r2cf_15) (planner *p) {
Chris@42 159 X(kr2c_register) (p, r2cf_15, &desc);
Chris@42 160 }
Chris@42 161
Chris@42 162 #else /* HAVE_FMA */
Chris@42 163
Chris@42 164 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 15 -name r2cf_15 -include r2cf.h */
Chris@42 165
Chris@42 166 /*
Chris@42 167 * This function contains 64 FP additions, 25 FP multiplications,
Chris@42 168 * (or, 50 additions, 11 multiplications, 14 fused multiply/add),
Chris@42 169 * 47 stack variables, 10 constants, and 30 memory accesses
Chris@42 170 */
Chris@42 171 #include "r2cf.h"
Chris@42 172
Chris@42 173 static void r2cf_15(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
Chris@42 174 {
Chris@42 175 DK(KP484122918, +0.484122918275927110647408174972799951354115213);
Chris@42 176 DK(KP216506350, +0.216506350946109661690930792688234045867850657);
Chris@42 177 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@42 178 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
Chris@42 179 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@42 180 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@42 181 DK(KP509036960, +0.509036960455127183450980863393907648510733164);
Chris@42 182 DK(KP823639103, +0.823639103546331925877420039278190003029660514);
Chris@42 183 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@42 184 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@42 185 {
Chris@42 186 INT i;
Chris@42 187 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(60, rs), MAKE_VOLATILE_STRIDE(60, csr), MAKE_VOLATILE_STRIDE(60, csi)) {
Chris@42 188 E Ti, TR, TL, TD, TE, T7, Te, Tf, TV, TW, TX, Tv, Ty, TH, To;
Chris@42 189 E Tr, TG, TS, TT, TU;
Chris@42 190 {
Chris@42 191 E TJ, Tg, Th, TK;
Chris@42 192 TJ = R0[0];
Chris@42 193 Tg = R0[WS(rs, 5)];
Chris@42 194 Th = R1[WS(rs, 2)];
Chris@42 195 TK = Th + Tg;
Chris@42 196 Ti = Tg - Th;
Chris@42 197 TR = TJ + TK;
Chris@42 198 TL = FNMS(KP500000000, TK, TJ);
Chris@42 199 }
Chris@42 200 {
Chris@42 201 E Tm, Tt, Tw, Tp, T3, Tx, Ta, Tn, Td, Tq, T6, Tu;
Chris@42 202 Tm = R1[WS(rs, 1)];
Chris@42 203 Tt = R0[WS(rs, 3)];
Chris@42 204 Tw = R1[WS(rs, 4)];
Chris@42 205 Tp = R0[WS(rs, 6)];
Chris@42 206 {
Chris@42 207 E T1, T2, T8, T9;
Chris@42 208 T1 = R0[WS(rs, 7)];
Chris@42 209 T2 = R0[WS(rs, 2)];
Chris@42 210 T3 = T1 - T2;
Chris@42 211 Tx = T1 + T2;
Chris@42 212 T8 = R1[WS(rs, 6)];
Chris@42 213 T9 = R0[WS(rs, 4)];
Chris@42 214 Ta = T8 - T9;
Chris@42 215 Tn = T9 + T8;
Chris@42 216 }
Chris@42 217 {
Chris@42 218 E Tb, Tc, T4, T5;
Chris@42 219 Tb = R1[WS(rs, 3)];
Chris@42 220 Tc = R0[WS(rs, 1)];
Chris@42 221 Td = Tb - Tc;
Chris@42 222 Tq = Tc + Tb;
Chris@42 223 T4 = R1[0];
Chris@42 224 T5 = R1[WS(rs, 5)];
Chris@42 225 T6 = T4 - T5;
Chris@42 226 Tu = T5 + T4;
Chris@42 227 }
Chris@42 228 TD = Ta - Td;
Chris@42 229 TE = T6 + T3;
Chris@42 230 T7 = T3 - T6;
Chris@42 231 Te = Ta + Td;
Chris@42 232 Tf = T7 - Te;
Chris@42 233 TV = Tt + Tu;
Chris@42 234 TW = Tw + Tx;
Chris@42 235 TX = TV + TW;
Chris@42 236 Tv = FNMS(KP500000000, Tu, Tt);
Chris@42 237 Ty = FNMS(KP500000000, Tx, Tw);
Chris@42 238 TH = Tv + Ty;
Chris@42 239 To = FNMS(KP500000000, Tn, Tm);
Chris@42 240 Tr = FNMS(KP500000000, Tq, Tp);
Chris@42 241 TG = To + Tr;
Chris@42 242 TS = Tm + Tn;
Chris@42 243 TT = Tp + Tq;
Chris@42 244 TU = TS + TT;
Chris@42 245 }
Chris@42 246 Ci[WS(csi, 5)] = KP866025403 * (Tf - Ti);
Chris@42 247 {
Chris@42 248 E TF, TP, TI, TM, TN, TQ, TO;
Chris@42 249 TF = FMA(KP823639103, TD, KP509036960 * TE);
Chris@42 250 TP = FNMS(KP509036960, TD, KP823639103 * TE);
Chris@42 251 TI = KP559016994 * (TG - TH);
Chris@42 252 TM = TG + TH;
Chris@42 253 TN = FNMS(KP250000000, TM, TL);
Chris@42 254 Cr[WS(csr, 5)] = TL + TM;
Chris@42 255 TQ = TN - TI;
Chris@42 256 Cr[WS(csr, 2)] = TP + TQ;
Chris@42 257 Cr[WS(csr, 7)] = TQ - TP;
Chris@42 258 TO = TI + TN;
Chris@42 259 Cr[WS(csr, 1)] = TF + TO;
Chris@42 260 Cr[WS(csr, 4)] = TO - TF;
Chris@42 261 }
Chris@42 262 {
Chris@42 263 E T11, T12, T10, TY, TZ;
Chris@42 264 T11 = TS - TT;
Chris@42 265 T12 = TW - TV;
Chris@42 266 Ci[WS(csi, 3)] = FMA(KP587785252, T11, KP951056516 * T12);
Chris@42 267 Ci[WS(csi, 6)] = FNMS(KP951056516, T11, KP587785252 * T12);
Chris@42 268 T10 = KP559016994 * (TU - TX);
Chris@42 269 TY = TU + TX;
Chris@42 270 TZ = FNMS(KP250000000, TY, TR);
Chris@42 271 Cr[WS(csr, 3)] = TZ - T10;
Chris@42 272 Cr[0] = TR + TY;
Chris@42 273 Cr[WS(csr, 6)] = T10 + TZ;
Chris@42 274 {
Chris@42 275 E Tl, TB, TA, TC;
Chris@42 276 {
Chris@42 277 E Tj, Tk, Ts, Tz;
Chris@42 278 Tj = FMA(KP866025403, Ti, KP216506350 * Tf);
Chris@42 279 Tk = KP484122918 * (Te + T7);
Chris@42 280 Tl = Tj + Tk;
Chris@42 281 TB = Tk - Tj;
Chris@42 282 Ts = To - Tr;
Chris@42 283 Tz = Tv - Ty;
Chris@42 284 TA = FMA(KP951056516, Ts, KP587785252 * Tz);
Chris@42 285 TC = FNMS(KP587785252, Ts, KP951056516 * Tz);
Chris@42 286 }
Chris@42 287 Ci[WS(csi, 1)] = Tl - TA;
Chris@42 288 Ci[WS(csi, 7)] = TC - TB;
Chris@42 289 Ci[WS(csi, 4)] = Tl + TA;
Chris@42 290 Ci[WS(csi, 2)] = TB + TC;
Chris@42 291 }
Chris@42 292 }
Chris@42 293 }
Chris@42 294 }
Chris@42 295 }
Chris@42 296
Chris@42 297 static const kr2c_desc desc = { 15, "r2cf_15", {50, 11, 14, 0}, &GENUS };
Chris@42 298
Chris@42 299 void X(codelet_r2cf_15) (planner *p) {
Chris@42 300 X(kr2c_register) (p, r2cf_15, &desc);
Chris@42 301 }
Chris@42 302
Chris@42 303 #endif /* HAVE_FMA */