Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/rdft/scalar/r2cb/r2cb_15.c @ 82:d0c2a83c1364

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add FFTW 3.3.8 source, and a Linux build
author Chris Cannam
date Tue, 19 Nov 2019 14:52:55 +0000
parents
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 08:07:28 EDT 2018 */
Chris@82 23
Chris@82 24 #include "rdft/codelet-rdft.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_r2cb.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -n 15 -name r2cb_15 -include rdft/scalar/r2cb.h */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 64 FP additions, 43 FP multiplications,
Chris@82 32 * (or, 21 additions, 0 multiplications, 43 fused multiply/add),
Chris@82 33 * 46 stack variables, 9 constants, and 30 memory accesses
Chris@82 34 */
Chris@82 35 #include "rdft/scalar/r2cb.h"
Chris@82 36
Chris@82 37 static void r2cb_15(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
Chris@82 38 {
Chris@82 39 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@82 40 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@82 41 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@82 42 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);
Chris@82 43 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);
Chris@82 44 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
Chris@82 45 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@82 46 DK(KP1_732050807, +1.732050807568877293527446341505872366942805254);
Chris@82 47 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
Chris@82 48 {
Chris@82 49 INT i;
Chris@82 50 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(60, rs), MAKE_VOLATILE_STRIDE(60, csr), MAKE_VOLATILE_STRIDE(60, csi)) {
Chris@82 51 E T3, Tt, Th, TC, TY, TZ, TD, TH, TI, Tm, Tu, Tr, Tv, T8, Td;
Chris@82 52 E Te;
Chris@82 53 {
Chris@82 54 E Tg, T1, T2, Tf;
Chris@82 55 Tg = Ci[WS(csi, 5)];
Chris@82 56 T1 = Cr[0];
Chris@82 57 T2 = Cr[WS(csr, 5)];
Chris@82 58 Tf = T1 - T2;
Chris@82 59 T3 = FMA(KP2_000000000, T2, T1);
Chris@82 60 Tt = FNMS(KP1_732050807, Tg, Tf);
Chris@82 61 Th = FMA(KP1_732050807, Tg, Tf);
Chris@82 62 }
Chris@82 63 {
Chris@82 64 E T4, TA, T9, TF, T5, T6, T7, Ta, Tb, Tc, Tq, TG, Tl, TB, Ti;
Chris@82 65 E Tn;
Chris@82 66 T4 = Cr[WS(csr, 3)];
Chris@82 67 TA = Ci[WS(csi, 3)];
Chris@82 68 T9 = Cr[WS(csr, 6)];
Chris@82 69 TF = Ci[WS(csi, 6)];
Chris@82 70 T5 = Cr[WS(csr, 7)];
Chris@82 71 T6 = Cr[WS(csr, 2)];
Chris@82 72 T7 = T5 + T6;
Chris@82 73 Ta = Cr[WS(csr, 4)];
Chris@82 74 Tb = Cr[WS(csr, 1)];
Chris@82 75 Tc = Ta + Tb;
Chris@82 76 {
Chris@82 77 E To, Tp, Tj, Tk;
Chris@82 78 To = Ci[WS(csi, 4)];
Chris@82 79 Tp = Ci[WS(csi, 1)];
Chris@82 80 Tq = To + Tp;
Chris@82 81 TG = Tp - To;
Chris@82 82 Tj = Ci[WS(csi, 7)];
Chris@82 83 Tk = Ci[WS(csi, 2)];
Chris@82 84 Tl = Tj - Tk;
Chris@82 85 TB = Tj + Tk;
Chris@82 86 }
Chris@82 87 TC = FMA(KP500000000, TB, TA);
Chris@82 88 TY = TG + TF;
Chris@82 89 TZ = TA - TB;
Chris@82 90 TD = T5 - T6;
Chris@82 91 TH = FNMS(KP500000000, TG, TF);
Chris@82 92 TI = Ta - Tb;
Chris@82 93 Ti = FNMS(KP2_000000000, T4, T7);
Chris@82 94 Tm = FMA(KP1_732050807, Tl, Ti);
Chris@82 95 Tu = FNMS(KP1_732050807, Tl, Ti);
Chris@82 96 Tn = FNMS(KP2_000000000, T9, Tc);
Chris@82 97 Tr = FMA(KP1_732050807, Tq, Tn);
Chris@82 98 Tv = FNMS(KP1_732050807, Tq, Tn);
Chris@82 99 T8 = T4 + T7;
Chris@82 100 Td = T9 + Tc;
Chris@82 101 Te = T8 + Td;
Chris@82 102 }
Chris@82 103 R0[0] = FMA(KP2_000000000, Te, T3);
Chris@82 104 {
Chris@82 105 E T10, T12, TX, T11, TV, TW;
Chris@82 106 T10 = FNMS(KP618033988, TZ, TY);
Chris@82 107 T12 = FMA(KP618033988, TY, TZ);
Chris@82 108 TV = FNMS(KP500000000, Te, T3);
Chris@82 109 TW = T8 - Td;
Chris@82 110 TX = FNMS(KP1_118033988, TW, TV);
Chris@82 111 T11 = FMA(KP1_118033988, TW, TV);
Chris@82 112 R1[WS(rs, 1)] = FNMS(KP1_902113032, T10, TX);
Chris@82 113 R1[WS(rs, 4)] = FMA(KP1_902113032, T12, T11);
Chris@82 114 R0[WS(rs, 6)] = FMA(KP1_902113032, T10, TX);
Chris@82 115 R0[WS(rs, 3)] = FNMS(KP1_902113032, T12, T11);
Chris@82 116 }
Chris@82 117 {
Chris@82 118 E TO, Ts, TN, TS, TU, TQ, TR, TT, TP;
Chris@82 119 TO = Tr - Tm;
Chris@82 120 Ts = Tm + Tr;
Chris@82 121 TN = FMA(KP250000000, Ts, Th);
Chris@82 122 TQ = FNMS(KP866025403, TI, TH);
Chris@82 123 TR = FNMS(KP866025403, TD, TC);
Chris@82 124 TS = FNMS(KP618033988, TR, TQ);
Chris@82 125 TU = FMA(KP618033988, TQ, TR);
Chris@82 126 R1[WS(rs, 2)] = Th - Ts;
Chris@82 127 TT = FMA(KP559016994, TO, TN);
Chris@82 128 R1[WS(rs, 5)] = FNMS(KP1_902113032, TU, TT);
Chris@82 129 R0[WS(rs, 7)] = FMA(KP1_902113032, TU, TT);
Chris@82 130 TP = FNMS(KP559016994, TO, TN);
Chris@82 131 R0[WS(rs, 4)] = FNMS(KP1_902113032, TS, TP);
Chris@82 132 R0[WS(rs, 1)] = FMA(KP1_902113032, TS, TP);
Chris@82 133 }
Chris@82 134 {
Chris@82 135 E Ty, Tw, Tx, TK, TM, TE, TJ, TL, Tz;
Chris@82 136 Ty = Tv - Tu;
Chris@82 137 Tw = Tu + Tv;
Chris@82 138 Tx = FMA(KP250000000, Tw, Tt);
Chris@82 139 TE = FMA(KP866025403, TD, TC);
Chris@82 140 TJ = FMA(KP866025403, TI, TH);
Chris@82 141 TK = FMA(KP618033988, TJ, TE);
Chris@82 142 TM = FNMS(KP618033988, TE, TJ);
Chris@82 143 R0[WS(rs, 5)] = Tt - Tw;
Chris@82 144 TL = FNMS(KP559016994, Ty, Tx);
Chris@82 145 R1[WS(rs, 6)] = FNMS(KP1_902113032, TM, TL);
Chris@82 146 R1[WS(rs, 3)] = FMA(KP1_902113032, TM, TL);
Chris@82 147 Tz = FMA(KP559016994, Ty, Tx);
Chris@82 148 R1[0] = FNMS(KP1_902113032, TK, Tz);
Chris@82 149 R0[WS(rs, 2)] = FMA(KP1_902113032, TK, Tz);
Chris@82 150 }
Chris@82 151 }
Chris@82 152 }
Chris@82 153 }
Chris@82 154
Chris@82 155 static const kr2c_desc desc = { 15, "r2cb_15", {21, 0, 43, 0}, &GENUS };
Chris@82 156
Chris@82 157 void X(codelet_r2cb_15) (planner *p) {
Chris@82 158 X(kr2c_register) (p, r2cb_15, &desc);
Chris@82 159 }
Chris@82 160
Chris@82 161 #else
Chris@82 162
Chris@82 163 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 15 -name r2cb_15 -include rdft/scalar/r2cb.h */
Chris@82 164
Chris@82 165 /*
Chris@82 166 * This function contains 64 FP additions, 31 FP multiplications,
Chris@82 167 * (or, 47 additions, 14 multiplications, 17 fused multiply/add),
Chris@82 168 * 44 stack variables, 7 constants, and 30 memory accesses
Chris@82 169 */
Chris@82 170 #include "rdft/scalar/r2cb.h"
Chris@82 171
Chris@82 172 static void r2cb_15(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
Chris@82 173 {
Chris@82 174 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);
Chris@82 175 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);
Chris@82 176 DK(KP1_175570504, +1.175570504584946258337411909278145537195304875);
Chris@82 177 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@82 178 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@82 179 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
Chris@82 180 DK(KP1_732050807, +1.732050807568877293527446341505872366942805254);
Chris@82 181 {
Chris@82 182 INT i;
Chris@82 183 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(60, rs), MAKE_VOLATILE_STRIDE(60, csr), MAKE_VOLATILE_STRIDE(60, csi)) {
Chris@82 184 E T3, Tu, Ti, TB, TZ, T10, TE, TG, TJ, Tn, Tv, Ts, Tw, T8, Td;
Chris@82 185 E Te;
Chris@82 186 {
Chris@82 187 E Th, T1, T2, Tf, Tg;
Chris@82 188 Tg = Ci[WS(csi, 5)];
Chris@82 189 Th = KP1_732050807 * Tg;
Chris@82 190 T1 = Cr[0];
Chris@82 191 T2 = Cr[WS(csr, 5)];
Chris@82 192 Tf = T1 - T2;
Chris@82 193 T3 = FMA(KP2_000000000, T2, T1);
Chris@82 194 Tu = Tf - Th;
Chris@82 195 Ti = Tf + Th;
Chris@82 196 }
Chris@82 197 {
Chris@82 198 E T4, TD, T9, TI, T5, T6, T7, Ta, Tb, Tc, Tr, TH, Tm, TC, Tj;
Chris@82 199 E To;
Chris@82 200 T4 = Cr[WS(csr, 3)];
Chris@82 201 TD = Ci[WS(csi, 3)];
Chris@82 202 T9 = Cr[WS(csr, 6)];
Chris@82 203 TI = Ci[WS(csi, 6)];
Chris@82 204 T5 = Cr[WS(csr, 7)];
Chris@82 205 T6 = Cr[WS(csr, 2)];
Chris@82 206 T7 = T5 + T6;
Chris@82 207 Ta = Cr[WS(csr, 4)];
Chris@82 208 Tb = Cr[WS(csr, 1)];
Chris@82 209 Tc = Ta + Tb;
Chris@82 210 {
Chris@82 211 E Tp, Tq, Tk, Tl;
Chris@82 212 Tp = Ci[WS(csi, 4)];
Chris@82 213 Tq = Ci[WS(csi, 1)];
Chris@82 214 Tr = KP866025403 * (Tp + Tq);
Chris@82 215 TH = Tp - Tq;
Chris@82 216 Tk = Ci[WS(csi, 7)];
Chris@82 217 Tl = Ci[WS(csi, 2)];
Chris@82 218 Tm = KP866025403 * (Tk - Tl);
Chris@82 219 TC = Tk + Tl;
Chris@82 220 }
Chris@82 221 TB = KP866025403 * (T5 - T6);
Chris@82 222 TZ = TD - TC;
Chris@82 223 T10 = TI - TH;
Chris@82 224 TE = FMA(KP500000000, TC, TD);
Chris@82 225 TG = KP866025403 * (Ta - Tb);
Chris@82 226 TJ = FMA(KP500000000, TH, TI);
Chris@82 227 Tj = FNMS(KP500000000, T7, T4);
Chris@82 228 Tn = Tj - Tm;
Chris@82 229 Tv = Tj + Tm;
Chris@82 230 To = FNMS(KP500000000, Tc, T9);
Chris@82 231 Ts = To - Tr;
Chris@82 232 Tw = To + Tr;
Chris@82 233 T8 = T4 + T7;
Chris@82 234 Td = T9 + Tc;
Chris@82 235 Te = T8 + Td;
Chris@82 236 }
Chris@82 237 R0[0] = FMA(KP2_000000000, Te, T3);
Chris@82 238 {
Chris@82 239 E T11, T13, TY, T12, TW, TX;
Chris@82 240 T11 = FNMS(KP1_902113032, T10, KP1_175570504 * TZ);
Chris@82 241 T13 = FMA(KP1_902113032, TZ, KP1_175570504 * T10);
Chris@82 242 TW = FNMS(KP500000000, Te, T3);
Chris@82 243 TX = KP1_118033988 * (T8 - Td);
Chris@82 244 TY = TW - TX;
Chris@82 245 T12 = TX + TW;
Chris@82 246 R0[WS(rs, 6)] = TY - T11;
Chris@82 247 R1[WS(rs, 4)] = T12 + T13;
Chris@82 248 R1[WS(rs, 1)] = TY + T11;
Chris@82 249 R0[WS(rs, 3)] = T12 - T13;
Chris@82 250 }
Chris@82 251 {
Chris@82 252 E TP, Tt, TO, TT, TV, TR, TS, TU, TQ;
Chris@82 253 TP = KP1_118033988 * (Tn - Ts);
Chris@82 254 Tt = Tn + Ts;
Chris@82 255 TO = FNMS(KP500000000, Tt, Ti);
Chris@82 256 TR = TE - TB;
Chris@82 257 TS = TJ - TG;
Chris@82 258 TT = FNMS(KP1_902113032, TS, KP1_175570504 * TR);
Chris@82 259 TV = FMA(KP1_902113032, TR, KP1_175570504 * TS);
Chris@82 260 R1[WS(rs, 2)] = FMA(KP2_000000000, Tt, Ti);
Chris@82 261 TU = TP + TO;
Chris@82 262 R1[WS(rs, 5)] = TU - TV;
Chris@82 263 R0[WS(rs, 7)] = TU + TV;
Chris@82 264 TQ = TO - TP;
Chris@82 265 R0[WS(rs, 1)] = TQ - TT;
Chris@82 266 R0[WS(rs, 4)] = TQ + TT;
Chris@82 267 }
Chris@82 268 {
Chris@82 269 E Tz, Tx, Ty, TL, TN, TF, TK, TM, TA;
Chris@82 270 Tz = KP1_118033988 * (Tv - Tw);
Chris@82 271 Tx = Tv + Tw;
Chris@82 272 Ty = FNMS(KP500000000, Tx, Tu);
Chris@82 273 TF = TB + TE;
Chris@82 274 TK = TG + TJ;
Chris@82 275 TL = FNMS(KP1_902113032, TK, KP1_175570504 * TF);
Chris@82 276 TN = FMA(KP1_902113032, TF, KP1_175570504 * TK);
Chris@82 277 R0[WS(rs, 5)] = FMA(KP2_000000000, Tx, Tu);
Chris@82 278 TM = Tz + Ty;
Chris@82 279 R1[0] = TM - TN;
Chris@82 280 R0[WS(rs, 2)] = TM + TN;
Chris@82 281 TA = Ty - Tz;
Chris@82 282 R1[WS(rs, 3)] = TA - TL;
Chris@82 283 R1[WS(rs, 6)] = TA + TL;
Chris@82 284 }
Chris@82 285 }
Chris@82 286 }
Chris@82 287 }
Chris@82 288
Chris@82 289 static const kr2c_desc desc = { 15, "r2cb_15", {47, 14, 17, 0}, &GENUS };
Chris@82 290
Chris@82 291 void X(codelet_r2cb_15) (planner *p) {
Chris@82 292 X(kr2c_register) (p, r2cb_15, &desc);
Chris@82 293 }
Chris@82 294
Chris@82 295 #endif