Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/rdft/scalar/r2cf/hc2cfdft2_4.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 16:48:57 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_hc2cdft.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 4 -dit -name hc2cfdft2_4 -include hc2cf.h */
Chris@42 29
Chris@42 30 /*
Chris@42 31 * This function contains 32 FP additions, 24 FP multiplications,
Chris@42 32 * (or, 24 additions, 16 multiplications, 8 fused multiply/add),
Chris@42 33 * 33 stack variables, 1 constants, and 16 memory accesses
Chris@42 34 */
Chris@42 35 #include "hc2cf.h"
Chris@42 36
Chris@42 37 static void hc2cfdft2_4(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@42 38 {
Chris@42 39 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@42 40 {
Chris@42 41 INT m;
Chris@42 42 for (m = mb, W = W + ((mb - 1) * 4); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 4, MAKE_VOLATILE_STRIDE(16, rs)) {
Chris@42 43 E T1, T5, T2, T4;
Chris@42 44 T1 = W[0];
Chris@42 45 T5 = W[3];
Chris@42 46 T2 = W[2];
Chris@42 47 T4 = W[1];
Chris@42 48 {
Chris@42 49 E Tc, T6, Tp, Tj, Tw, Tt, T9, TE, To, TC, Ta, Tr, Tf, Tl, Tm;
Chris@42 50 {
Chris@42 51 E Th, Tb, T3, Ti;
Chris@42 52 Th = Ip[0];
Chris@42 53 Tb = T1 * T5;
Chris@42 54 T3 = T1 * T2;
Chris@42 55 Ti = Im[0];
Chris@42 56 Tl = Rm[0];
Chris@42 57 Tc = FNMS(T4, T2, Tb);
Chris@42 58 T6 = FMA(T4, T5, T3);
Chris@42 59 Tp = Th + Ti;
Chris@42 60 Tj = Th - Ti;
Chris@42 61 Tm = Rp[0];
Chris@42 62 }
Chris@42 63 {
Chris@42 64 E T7, T8, Td, Tn, Te;
Chris@42 65 T7 = Ip[WS(rs, 1)];
Chris@42 66 T8 = Im[WS(rs, 1)];
Chris@42 67 Td = Rp[WS(rs, 1)];
Chris@42 68 Tw = Tm + Tl;
Chris@42 69 Tn = Tl - Tm;
Chris@42 70 Tt = T7 + T8;
Chris@42 71 T9 = T7 - T8;
Chris@42 72 Te = Rm[WS(rs, 1)];
Chris@42 73 TE = T4 * Tn;
Chris@42 74 To = T1 * Tn;
Chris@42 75 TC = T2 * Tt;
Chris@42 76 Ta = T6 * T9;
Chris@42 77 Tr = Td - Te;
Chris@42 78 Tf = Td + Te;
Chris@42 79 }
Chris@42 80 {
Chris@42 81 E Tq, Tk, TB, Ty, Tu, TI, TG, TF;
Chris@42 82 Tq = FNMS(T4, Tp, To);
Chris@42 83 TF = FMA(T1, Tp, TE);
Chris@42 84 {
Chris@42 85 E Tg, Tx, TD, Ts;
Chris@42 86 Tg = FNMS(Tc, Tf, Ta);
Chris@42 87 Tx = T6 * Tf;
Chris@42 88 TD = FNMS(T5, Tr, TC);
Chris@42 89 Ts = T2 * Tr;
Chris@42 90 Tk = Tg + Tj;
Chris@42 91 TB = Tj - Tg;
Chris@42 92 Ty = FMA(Tc, T9, Tx);
Chris@42 93 Tu = FMA(T5, Tt, Ts);
Chris@42 94 TI = TD + TF;
Chris@42 95 TG = TD - TF;
Chris@42 96 }
Chris@42 97 {
Chris@42 98 E Tz, TH, Tv, TA;
Chris@42 99 Tz = Tw - Ty;
Chris@42 100 TH = Tw + Ty;
Chris@42 101 Tv = Tq - Tu;
Chris@42 102 TA = Tu + Tq;
Chris@42 103 Rp[0] = KP500000000 * (TH + TI);
Chris@42 104 Rm[WS(rs, 1)] = KP500000000 * (TH - TI);
Chris@42 105 Rm[0] = KP500000000 * (Tz - TA);
Chris@42 106 Im[WS(rs, 1)] = KP500000000 * (Tv - Tk);
Chris@42 107 Ip[0] = KP500000000 * (Tk + Tv);
Chris@42 108 Im[0] = KP500000000 * (TG - TB);
Chris@42 109 Rp[WS(rs, 1)] = KP500000000 * (Tz + TA);
Chris@42 110 Ip[WS(rs, 1)] = KP500000000 * (TB + TG);
Chris@42 111 }
Chris@42 112 }
Chris@42 113 }
Chris@42 114 }
Chris@42 115 }
Chris@42 116 }
Chris@42 117
Chris@42 118 static const tw_instr twinstr[] = {
Chris@42 119 {TW_CEXP, 1, 1},
Chris@42 120 {TW_CEXP, 1, 3},
Chris@42 121 {TW_NEXT, 1, 0}
Chris@42 122 };
Chris@42 123
Chris@42 124 static const hc2c_desc desc = { 4, "hc2cfdft2_4", twinstr, &GENUS, {24, 16, 8, 0} };
Chris@42 125
Chris@42 126 void X(codelet_hc2cfdft2_4) (planner *p) {
Chris@42 127 X(khc2c_register) (p, hc2cfdft2_4, &desc, HC2C_VIA_DFT);
Chris@42 128 }
Chris@42 129 #else /* HAVE_FMA */
Chris@42 130
Chris@42 131 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 4 -dit -name hc2cfdft2_4 -include hc2cf.h */
Chris@42 132
Chris@42 133 /*
Chris@42 134 * This function contains 32 FP additions, 24 FP multiplications,
Chris@42 135 * (or, 24 additions, 16 multiplications, 8 fused multiply/add),
Chris@42 136 * 24 stack variables, 1 constants, and 16 memory accesses
Chris@42 137 */
Chris@42 138 #include "hc2cf.h"
Chris@42 139
Chris@42 140 static void hc2cfdft2_4(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@42 141 {
Chris@42 142 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@42 143 {
Chris@42 144 INT m;
Chris@42 145 for (m = mb, W = W + ((mb - 1) * 4); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 4, MAKE_VOLATILE_STRIDE(16, rs)) {
Chris@42 146 E T1, T3, T2, T4, T5, T9;
Chris@42 147 T1 = W[0];
Chris@42 148 T3 = W[1];
Chris@42 149 T2 = W[2];
Chris@42 150 T4 = W[3];
Chris@42 151 T5 = FMA(T1, T2, T3 * T4);
Chris@42 152 T9 = FNMS(T3, T2, T1 * T4);
Chris@42 153 {
Chris@42 154 E Tg, Tr, Tm, Tx, Td, Tw, Tp, Ts;
Chris@42 155 {
Chris@42 156 E Te, Tf, Tl, Ti, Tj, Tk;
Chris@42 157 Te = Ip[0];
Chris@42 158 Tf = Im[0];
Chris@42 159 Tl = Te + Tf;
Chris@42 160 Ti = Rm[0];
Chris@42 161 Tj = Rp[0];
Chris@42 162 Tk = Ti - Tj;
Chris@42 163 Tg = Te - Tf;
Chris@42 164 Tr = Tj + Ti;
Chris@42 165 Tm = FNMS(T3, Tl, T1 * Tk);
Chris@42 166 Tx = FMA(T3, Tk, T1 * Tl);
Chris@42 167 }
Chris@42 168 {
Chris@42 169 E T8, To, Tc, Tn;
Chris@42 170 {
Chris@42 171 E T6, T7, Ta, Tb;
Chris@42 172 T6 = Ip[WS(rs, 1)];
Chris@42 173 T7 = Im[WS(rs, 1)];
Chris@42 174 T8 = T6 - T7;
Chris@42 175 To = T6 + T7;
Chris@42 176 Ta = Rp[WS(rs, 1)];
Chris@42 177 Tb = Rm[WS(rs, 1)];
Chris@42 178 Tc = Ta + Tb;
Chris@42 179 Tn = Ta - Tb;
Chris@42 180 }
Chris@42 181 Td = FNMS(T9, Tc, T5 * T8);
Chris@42 182 Tw = FNMS(T4, Tn, T2 * To);
Chris@42 183 Tp = FMA(T2, Tn, T4 * To);
Chris@42 184 Ts = FMA(T5, Tc, T9 * T8);
Chris@42 185 }
Chris@42 186 {
Chris@42 187 E Th, Tq, Tz, TA;
Chris@42 188 Th = Td + Tg;
Chris@42 189 Tq = Tm - Tp;
Chris@42 190 Ip[0] = KP500000000 * (Th + Tq);
Chris@42 191 Im[WS(rs, 1)] = KP500000000 * (Tq - Th);
Chris@42 192 Tz = Tr + Ts;
Chris@42 193 TA = Tw + Tx;
Chris@42 194 Rm[WS(rs, 1)] = KP500000000 * (Tz - TA);
Chris@42 195 Rp[0] = KP500000000 * (Tz + TA);
Chris@42 196 }
Chris@42 197 {
Chris@42 198 E Tt, Tu, Tv, Ty;
Chris@42 199 Tt = Tr - Ts;
Chris@42 200 Tu = Tp + Tm;
Chris@42 201 Rm[0] = KP500000000 * (Tt - Tu);
Chris@42 202 Rp[WS(rs, 1)] = KP500000000 * (Tt + Tu);
Chris@42 203 Tv = Tg - Td;
Chris@42 204 Ty = Tw - Tx;
Chris@42 205 Ip[WS(rs, 1)] = KP500000000 * (Tv + Ty);
Chris@42 206 Im[0] = KP500000000 * (Ty - Tv);
Chris@42 207 }
Chris@42 208 }
Chris@42 209 }
Chris@42 210 }
Chris@42 211 }
Chris@42 212
Chris@42 213 static const tw_instr twinstr[] = {
Chris@42 214 {TW_CEXP, 1, 1},
Chris@42 215 {TW_CEXP, 1, 3},
Chris@42 216 {TW_NEXT, 1, 0}
Chris@42 217 };
Chris@42 218
Chris@42 219 static const hc2c_desc desc = { 4, "hc2cfdft2_4", twinstr, &GENUS, {24, 16, 8, 0} };
Chris@42 220
Chris@42 221 void X(codelet_hc2cfdft2_4) (planner *p) {
Chris@42 222 X(khc2c_register) (p, hc2cfdft2_4, &desc, HC2C_VIA_DFT);
Chris@42 223 }
Chris@42 224 #endif /* HAVE_FMA */