Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/rdft/scalar/r2cb/r2cb_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 16:49:28 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_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 15 -name r2cb_15 -include r2cb.h */
Chris@42 29
Chris@42 30 /*
Chris@42 31 * This function contains 64 FP additions, 43 FP multiplications,
Chris@42 32 * (or, 21 additions, 0 multiplications, 43 fused multiply/add),
Chris@42 33 * 54 stack variables, 9 constants, and 30 memory accesses
Chris@42 34 */
Chris@42 35 #include "r2cb.h"
Chris@42 36
Chris@42 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@42 38 {
Chris@42 39 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@42 40 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);
Chris@42 41 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@42 42 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@42 43 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);
Chris@42 44 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
Chris@42 45 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@42 46 DK(KP1_732050807, +1.732050807568877293527446341505872366942805254);
Chris@42 47 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
Chris@42 48 {
Chris@42 49 INT i;
Chris@42 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@42 51 E TL, Tz, TM, TK;
Chris@42 52 {
Chris@42 53 E T3, Th, Tt, TD, TI, TH, TY, TC, TZ, Tu, Tm, Tv, Tr, Te, TW;
Chris@42 54 E Tg, T1, T2, T12, T10, TV;
Chris@42 55 Tg = Ci[WS(csi, 5)];
Chris@42 56 T1 = Cr[0];
Chris@42 57 T2 = Cr[WS(csr, 5)];
Chris@42 58 {
Chris@42 59 E T4, TA, T9, TF, T7, Tj, Tc, Tk, TG, Tq, Tf, Tl, TB;
Chris@42 60 T4 = Cr[WS(csr, 3)];
Chris@42 61 TA = Ci[WS(csi, 3)];
Chris@42 62 T9 = Cr[WS(csr, 6)];
Chris@42 63 Tf = T1 - T2;
Chris@42 64 T3 = FMA(KP2_000000000, T2, T1);
Chris@42 65 TF = Ci[WS(csi, 6)];
Chris@42 66 {
Chris@42 67 E Ta, Tb, T5, T6, To, Tp;
Chris@42 68 T5 = Cr[WS(csr, 7)];
Chris@42 69 T6 = Cr[WS(csr, 2)];
Chris@42 70 Th = FMA(KP1_732050807, Tg, Tf);
Chris@42 71 Tt = FNMS(KP1_732050807, Tg, Tf);
Chris@42 72 Ta = Cr[WS(csr, 4)];
Chris@42 73 TD = T5 - T6;
Chris@42 74 T7 = T5 + T6;
Chris@42 75 Tb = Cr[WS(csr, 1)];
Chris@42 76 To = Ci[WS(csi, 4)];
Chris@42 77 Tp = Ci[WS(csi, 1)];
Chris@42 78 Tj = Ci[WS(csi, 7)];
Chris@42 79 Tc = Ta + Tb;
Chris@42 80 TI = Ta - Tb;
Chris@42 81 Tk = Ci[WS(csi, 2)];
Chris@42 82 TG = Tp - To;
Chris@42 83 Tq = To + Tp;
Chris@42 84 }
Chris@42 85 Tl = Tj - Tk;
Chris@42 86 TB = Tj + Tk;
Chris@42 87 TH = FNMS(KP500000000, TG, TF);
Chris@42 88 TY = TG + TF;
Chris@42 89 TC = FMA(KP500000000, TB, TA);
Chris@42 90 TZ = TA - TB;
Chris@42 91 {
Chris@42 92 E Ti, T8, Td, Tn;
Chris@42 93 Ti = FNMS(KP2_000000000, T4, T7);
Chris@42 94 T8 = T4 + T7;
Chris@42 95 Td = T9 + Tc;
Chris@42 96 Tn = FNMS(KP2_000000000, T9, Tc);
Chris@42 97 Tu = FNMS(KP1_732050807, Tl, Ti);
Chris@42 98 Tm = FMA(KP1_732050807, Tl, Ti);
Chris@42 99 Tv = FNMS(KP1_732050807, Tq, Tn);
Chris@42 100 Tr = FMA(KP1_732050807, Tq, Tn);
Chris@42 101 Te = T8 + Td;
Chris@42 102 TW = T8 - Td;
Chris@42 103 }
Chris@42 104 }
Chris@42 105 T12 = FMA(KP618033988, TY, TZ);
Chris@42 106 T10 = FNMS(KP618033988, TZ, TY);
Chris@42 107 TV = FNMS(KP500000000, Te, T3);
Chris@42 108 R0[0] = FMA(KP2_000000000, Te, T3);
Chris@42 109 {
Chris@42 110 E TJ, TE, TT, TP, TU, TS, Ty, Tw, Tx;
Chris@42 111 {
Chris@42 112 E TO, Ts, TQ, TN, TR, T11, TX;
Chris@42 113 TO = Tr - Tm;
Chris@42 114 Ts = Tm + Tr;
Chris@42 115 T11 = FMA(KP1_118033988, TW, TV);
Chris@42 116 TX = FNMS(KP1_118033988, TW, TV);
Chris@42 117 TQ = FNMS(KP866025403, TI, TH);
Chris@42 118 TJ = FMA(KP866025403, TI, TH);
Chris@42 119 TN = FMA(KP250000000, Ts, Th);
Chris@42 120 R0[WS(rs, 3)] = FNMS(KP1_902113032, T12, T11);
Chris@42 121 R1[WS(rs, 4)] = FMA(KP1_902113032, T12, T11);
Chris@42 122 R0[WS(rs, 6)] = FMA(KP1_902113032, T10, TX);
Chris@42 123 R1[WS(rs, 1)] = FNMS(KP1_902113032, T10, TX);
Chris@42 124 TR = FNMS(KP866025403, TD, TC);
Chris@42 125 TE = FMA(KP866025403, TD, TC);
Chris@42 126 R1[WS(rs, 2)] = Th - Ts;
Chris@42 127 TT = FMA(KP559016994, TO, TN);
Chris@42 128 TP = FNMS(KP559016994, TO, TN);
Chris@42 129 TU = FMA(KP618033988, TQ, TR);
Chris@42 130 TS = FNMS(KP618033988, TR, TQ);
Chris@42 131 }
Chris@42 132 Ty = Tv - Tu;
Chris@42 133 Tw = Tu + Tv;
Chris@42 134 R0[WS(rs, 7)] = FMA(KP1_902113032, TU, TT);
Chris@42 135 R1[WS(rs, 5)] = FNMS(KP1_902113032, TU, TT);
Chris@42 136 R0[WS(rs, 1)] = FMA(KP1_902113032, TS, TP);
Chris@42 137 R0[WS(rs, 4)] = FNMS(KP1_902113032, TS, TP);
Chris@42 138 Tx = FMA(KP250000000, Tw, Tt);
Chris@42 139 R0[WS(rs, 5)] = Tt - Tw;
Chris@42 140 TL = FNMS(KP559016994, Ty, Tx);
Chris@42 141 Tz = FMA(KP559016994, Ty, Tx);
Chris@42 142 TM = FNMS(KP618033988, TE, TJ);
Chris@42 143 TK = FMA(KP618033988, TJ, TE);
Chris@42 144 }
Chris@42 145 }
Chris@42 146 R1[WS(rs, 3)] = FMA(KP1_902113032, TM, TL);
Chris@42 147 R1[WS(rs, 6)] = FNMS(KP1_902113032, TM, TL);
Chris@42 148 R0[WS(rs, 2)] = FMA(KP1_902113032, TK, Tz);
Chris@42 149 R1[0] = FNMS(KP1_902113032, TK, Tz);
Chris@42 150 }
Chris@42 151 }
Chris@42 152 }
Chris@42 153
Chris@42 154 static const kr2c_desc desc = { 15, "r2cb_15", {21, 0, 43, 0}, &GENUS };
Chris@42 155
Chris@42 156 void X(codelet_r2cb_15) (planner *p) {
Chris@42 157 X(kr2c_register) (p, r2cb_15, &desc);
Chris@42 158 }
Chris@42 159
Chris@42 160 #else /* HAVE_FMA */
Chris@42 161
Chris@42 162 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 15 -name r2cb_15 -include r2cb.h */
Chris@42 163
Chris@42 164 /*
Chris@42 165 * This function contains 64 FP additions, 31 FP multiplications,
Chris@42 166 * (or, 47 additions, 14 multiplications, 17 fused multiply/add),
Chris@42 167 * 44 stack variables, 7 constants, and 30 memory accesses
Chris@42 168 */
Chris@42 169 #include "r2cb.h"
Chris@42 170
Chris@42 171 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@42 172 {
Chris@42 173 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);
Chris@42 174 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);
Chris@42 175 DK(KP1_175570504, +1.175570504584946258337411909278145537195304875);
Chris@42 176 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@42 177 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@42 178 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
Chris@42 179 DK(KP1_732050807, +1.732050807568877293527446341505872366942805254);
Chris@42 180 {
Chris@42 181 INT i;
Chris@42 182 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@42 183 E T3, Tu, Ti, TB, TZ, T10, TE, TG, TJ, Tn, Tv, Ts, Tw, T8, Td;
Chris@42 184 E Te;
Chris@42 185 {
Chris@42 186 E Th, T1, T2, Tf, Tg;
Chris@42 187 Tg = Ci[WS(csi, 5)];
Chris@42 188 Th = KP1_732050807 * Tg;
Chris@42 189 T1 = Cr[0];
Chris@42 190 T2 = Cr[WS(csr, 5)];
Chris@42 191 Tf = T1 - T2;
Chris@42 192 T3 = FMA(KP2_000000000, T2, T1);
Chris@42 193 Tu = Tf - Th;
Chris@42 194 Ti = Tf + Th;
Chris@42 195 }
Chris@42 196 {
Chris@42 197 E T4, TD, T9, TI, T5, T6, T7, Ta, Tb, Tc, Tr, TH, Tm, TC, Tj;
Chris@42 198 E To;
Chris@42 199 T4 = Cr[WS(csr, 3)];
Chris@42 200 TD = Ci[WS(csi, 3)];
Chris@42 201 T9 = Cr[WS(csr, 6)];
Chris@42 202 TI = Ci[WS(csi, 6)];
Chris@42 203 T5 = Cr[WS(csr, 7)];
Chris@42 204 T6 = Cr[WS(csr, 2)];
Chris@42 205 T7 = T5 + T6;
Chris@42 206 Ta = Cr[WS(csr, 4)];
Chris@42 207 Tb = Cr[WS(csr, 1)];
Chris@42 208 Tc = Ta + Tb;
Chris@42 209 {
Chris@42 210 E Tp, Tq, Tk, Tl;
Chris@42 211 Tp = Ci[WS(csi, 4)];
Chris@42 212 Tq = Ci[WS(csi, 1)];
Chris@42 213 Tr = KP866025403 * (Tp + Tq);
Chris@42 214 TH = Tp - Tq;
Chris@42 215 Tk = Ci[WS(csi, 7)];
Chris@42 216 Tl = Ci[WS(csi, 2)];
Chris@42 217 Tm = KP866025403 * (Tk - Tl);
Chris@42 218 TC = Tk + Tl;
Chris@42 219 }
Chris@42 220 TB = KP866025403 * (T5 - T6);
Chris@42 221 TZ = TD - TC;
Chris@42 222 T10 = TI - TH;
Chris@42 223 TE = FMA(KP500000000, TC, TD);
Chris@42 224 TG = KP866025403 * (Ta - Tb);
Chris@42 225 TJ = FMA(KP500000000, TH, TI);
Chris@42 226 Tj = FNMS(KP500000000, T7, T4);
Chris@42 227 Tn = Tj - Tm;
Chris@42 228 Tv = Tj + Tm;
Chris@42 229 To = FNMS(KP500000000, Tc, T9);
Chris@42 230 Ts = To - Tr;
Chris@42 231 Tw = To + Tr;
Chris@42 232 T8 = T4 + T7;
Chris@42 233 Td = T9 + Tc;
Chris@42 234 Te = T8 + Td;
Chris@42 235 }
Chris@42 236 R0[0] = FMA(KP2_000000000, Te, T3);
Chris@42 237 {
Chris@42 238 E T11, T13, TY, T12, TW, TX;
Chris@42 239 T11 = FNMS(KP1_902113032, T10, KP1_175570504 * TZ);
Chris@42 240 T13 = FMA(KP1_902113032, TZ, KP1_175570504 * T10);
Chris@42 241 TW = FNMS(KP500000000, Te, T3);
Chris@42 242 TX = KP1_118033988 * (T8 - Td);
Chris@42 243 TY = TW - TX;
Chris@42 244 T12 = TX + TW;
Chris@42 245 R0[WS(rs, 6)] = TY - T11;
Chris@42 246 R1[WS(rs, 4)] = T12 + T13;
Chris@42 247 R1[WS(rs, 1)] = TY + T11;
Chris@42 248 R0[WS(rs, 3)] = T12 - T13;
Chris@42 249 }
Chris@42 250 {
Chris@42 251 E TP, Tt, TO, TT, TV, TR, TS, TU, TQ;
Chris@42 252 TP = KP1_118033988 * (Tn - Ts);
Chris@42 253 Tt = Tn + Ts;
Chris@42 254 TO = FNMS(KP500000000, Tt, Ti);
Chris@42 255 TR = TE - TB;
Chris@42 256 TS = TJ - TG;
Chris@42 257 TT = FNMS(KP1_902113032, TS, KP1_175570504 * TR);
Chris@42 258 TV = FMA(KP1_902113032, TR, KP1_175570504 * TS);
Chris@42 259 R1[WS(rs, 2)] = FMA(KP2_000000000, Tt, Ti);
Chris@42 260 TU = TP + TO;
Chris@42 261 R1[WS(rs, 5)] = TU - TV;
Chris@42 262 R0[WS(rs, 7)] = TU + TV;
Chris@42 263 TQ = TO - TP;
Chris@42 264 R0[WS(rs, 1)] = TQ - TT;
Chris@42 265 R0[WS(rs, 4)] = TQ + TT;
Chris@42 266 }
Chris@42 267 {
Chris@42 268 E Tz, Tx, Ty, TL, TN, TF, TK, TM, TA;
Chris@42 269 Tz = KP1_118033988 * (Tv - Tw);
Chris@42 270 Tx = Tv + Tw;
Chris@42 271 Ty = FNMS(KP500000000, Tx, Tu);
Chris@42 272 TF = TB + TE;
Chris@42 273 TK = TG + TJ;
Chris@42 274 TL = FNMS(KP1_902113032, TK, KP1_175570504 * TF);
Chris@42 275 TN = FMA(KP1_902113032, TF, KP1_175570504 * TK);
Chris@42 276 R0[WS(rs, 5)] = FMA(KP2_000000000, Tx, Tu);
Chris@42 277 TM = Tz + Ty;
Chris@42 278 R1[0] = TM - TN;
Chris@42 279 R0[WS(rs, 2)] = TM + TN;
Chris@42 280 TA = Ty - Tz;
Chris@42 281 R1[WS(rs, 3)] = TA - TL;
Chris@42 282 R1[WS(rs, 6)] = TA + TL;
Chris@42 283 }
Chris@42 284 }
Chris@42 285 }
Chris@42 286 }
Chris@42 287
Chris@42 288 static const kr2c_desc desc = { 15, "r2cb_15", {47, 14, 17, 0}, &GENUS };
Chris@42 289
Chris@42 290 void X(codelet_r2cb_15) (planner *p) {
Chris@42 291 X(kr2c_register) (p, r2cb_15, &desc);
Chris@42 292 }
Chris@42 293
Chris@42 294 #endif /* HAVE_FMA */