Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/rdft/scalar/r2cf/r2cfII_9.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 16:47:24 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 9 -name r2cfII_9 -dft-II -include r2cfII.h */
Chris@42 29
Chris@42 30 /*
Chris@42 31 * This function contains 42 FP additions, 34 FP multiplications,
Chris@42 32 * (or, 12 additions, 4 multiplications, 30 fused multiply/add),
Chris@42 33 * 46 stack variables, 17 constants, and 18 memory accesses
Chris@42 34 */
Chris@42 35 #include "r2cfII.h"
Chris@42 36
Chris@42 37 static void r2cfII_9(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(KP939692620, +0.939692620785908384054109277324731469936208134);
Chris@42 40 DK(KP879385241, +0.879385241571816768108218554649462939872416269);
Chris@42 41 DK(KP984807753, +0.984807753012208059366743024589523013670643252);
Chris@42 42 DK(KP852868531, +0.852868531952443209628250963940074071936020296);
Chris@42 43 DK(KP666666666, +0.666666666666666666666666666666666666666666667);
Chris@42 44 DK(KP673648177, +0.673648177666930348851716626769314796000375677);
Chris@42 45 DK(KP898197570, +0.898197570222573798468955502359086394667167570);
Chris@42 46 DK(KP826351822, +0.826351822333069651148283373230685203999624323);
Chris@42 47 DK(KP907603734, +0.907603734547952313649323976213898122064543220);
Chris@42 48 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@42 49 DK(KP420276625, +0.420276625461206169731530603237061658838781920);
Chris@42 50 DK(KP315207469, +0.315207469095904627298647952427796244129086440);
Chris@42 51 DK(KP203604859, +0.203604859554852403062088995281827210665664861);
Chris@42 52 DK(KP152703644, +0.152703644666139302296566746461370407999248646);
Chris@42 53 DK(KP726681596, +0.726681596905677465811651808188092531873167623);
Chris@42 54 DK(KP968908795, +0.968908795874236621082202410917456709164223497);
Chris@42 55 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@42 56 {
Chris@42 57 INT i;
Chris@42 58 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(36, rs), MAKE_VOLATILE_STRIDE(36, csr), MAKE_VOLATILE_STRIDE(36, csi)) {
Chris@42 59 E To, T5, Tp, Ta, Ti, Tm, TB, Tq, Tt, Tf, Th;
Chris@42 60 {
Chris@42 61 E T1, T6, T4, Tb, Tk, T9, Tc, Td, Tl, Te;
Chris@42 62 {
Chris@42 63 E T2, T3, T7, T8;
Chris@42 64 T1 = R0[0];
Chris@42 65 T2 = R0[WS(rs, 3)];
Chris@42 66 T3 = R1[WS(rs, 1)];
Chris@42 67 T6 = R0[WS(rs, 1)];
Chris@42 68 T7 = R0[WS(rs, 4)];
Chris@42 69 T8 = R1[WS(rs, 2)];
Chris@42 70 T4 = T2 - T3;
Chris@42 71 To = T2 + T3;
Chris@42 72 Tb = R0[WS(rs, 2)];
Chris@42 73 Tk = T7 + T8;
Chris@42 74 T9 = T7 - T8;
Chris@42 75 Tc = R1[0];
Chris@42 76 Td = R1[WS(rs, 3)];
Chris@42 77 }
Chris@42 78 T5 = T1 + T4;
Chris@42 79 Tp = FNMS(KP500000000, T4, T1);
Chris@42 80 Ta = T6 + T9;
Chris@42 81 Tl = FNMS(KP500000000, T9, T6);
Chris@42 82 Te = Tc + Td;
Chris@42 83 Ti = Tc - Td;
Chris@42 84 Tm = FMA(KP968908795, Tl, Tk);
Chris@42 85 TB = FNMS(KP726681596, Tk, Tl);
Chris@42 86 Tq = FNMS(KP152703644, Tk, Tl);
Chris@42 87 Tt = FMA(KP203604859, Tl, Tk);
Chris@42 88 Tf = Tb - Te;
Chris@42 89 Th = FMA(KP500000000, Te, Tb);
Chris@42 90 }
Chris@42 91 {
Chris@42 92 E Ts, Tr, TA, Tj, Tg;
Chris@42 93 Ts = FMA(KP315207469, Ti, Th);
Chris@42 94 Tr = FNMS(KP420276625, Th, Ti);
Chris@42 95 TA = FMA(KP203604859, Th, Ti);
Chris@42 96 Tj = FNMS(KP152703644, Ti, Th);
Chris@42 97 Tg = Ta + Tf;
Chris@42 98 Ci[WS(csi, 1)] = KP866025403 * (Tf - Ta);
Chris@42 99 {
Chris@42 100 E Tu, Tx, TF, TC;
Chris@42 101 Tu = FNMS(KP907603734, Tt, Ts);
Chris@42 102 Tx = FNMS(KP826351822, Tr, Tq);
Chris@42 103 TF = FMA(KP898197570, TB, TA);
Chris@42 104 TC = FNMS(KP898197570, TB, TA);
Chris@42 105 {
Chris@42 106 E TE, Tn, Tv, Ty;
Chris@42 107 TE = FNMS(KP673648177, Tm, Tj);
Chris@42 108 Tn = FMA(KP673648177, Tm, Tj);
Chris@42 109 Cr[WS(csr, 4)] = T5 + Tg;
Chris@42 110 Cr[WS(csr, 1)] = FNMS(KP500000000, Tg, T5);
Chris@42 111 Tv = FNMS(KP666666666, Tu, Tr);
Chris@42 112 Ty = FNMS(KP666666666, Tx, Tt);
Chris@42 113 Cr[0] = FMA(KP852868531, TF, Tp);
Chris@42 114 {
Chris@42 115 E TG, TD, Tw, Tz;
Chris@42 116 TG = FMA(KP500000000, TF, TE);
Chris@42 117 Ci[0] = -(KP984807753 * (FMA(KP879385241, To, Tn)));
Chris@42 118 TD = FNMS(KP666666666, Tn, TC);
Chris@42 119 Tw = FMA(KP826351822, Tv, Tq);
Chris@42 120 Tz = FMA(KP907603734, Ty, Ts);
Chris@42 121 Cr[WS(csr, 3)] = FNMS(KP852868531, TG, Tp);
Chris@42 122 Ci[WS(csi, 3)] = -(KP866025403 * (FMA(KP852868531, TD, To)));
Chris@42 123 Cr[WS(csr, 2)] = FNMS(KP852868531, Tw, Tp);
Chris@42 124 Ci[WS(csi, 2)] = KP866025403 * (FNMS(KP939692620, Tz, To));
Chris@42 125 }
Chris@42 126 }
Chris@42 127 }
Chris@42 128 }
Chris@42 129 }
Chris@42 130 }
Chris@42 131 }
Chris@42 132
Chris@42 133 static const kr2c_desc desc = { 9, "r2cfII_9", {12, 4, 30, 0}, &GENUS };
Chris@42 134
Chris@42 135 void X(codelet_r2cfII_9) (planner *p) {
Chris@42 136 X(kr2c_register) (p, r2cfII_9, &desc);
Chris@42 137 }
Chris@42 138
Chris@42 139 #else /* HAVE_FMA */
Chris@42 140
Chris@42 141 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 9 -name r2cfII_9 -dft-II -include r2cfII.h */
Chris@42 142
Chris@42 143 /*
Chris@42 144 * This function contains 42 FP additions, 30 FP multiplications,
Chris@42 145 * (or, 25 additions, 13 multiplications, 17 fused multiply/add),
Chris@42 146 * 39 stack variables, 14 constants, and 18 memory accesses
Chris@42 147 */
Chris@42 148 #include "r2cfII.h"
Chris@42 149
Chris@42 150 static void r2cfII_9(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
Chris@42 151 {
Chris@42 152 DK(KP663413948, +0.663413948168938396205421319635891297216863310);
Chris@42 153 DK(KP642787609, +0.642787609686539326322643409907263432907559884);
Chris@42 154 DK(KP556670399, +0.556670399226419366452912952047023132968291906);
Chris@42 155 DK(KP766044443, +0.766044443118978035202392650555416673935832457);
Chris@42 156 DK(KP852868531, +0.852868531952443209628250963940074071936020296);
Chris@42 157 DK(KP173648177, +0.173648177666930348851716626769314796000375677);
Chris@42 158 DK(KP984807753, +0.984807753012208059366743024589523013670643252);
Chris@42 159 DK(KP150383733, +0.150383733180435296639271897612501926072238258);
Chris@42 160 DK(KP813797681, +0.813797681349373692844693217248393223289101568);
Chris@42 161 DK(KP342020143, +0.342020143325668733044099614682259580763083368);
Chris@42 162 DK(KP939692620, +0.939692620785908384054109277324731469936208134);
Chris@42 163 DK(KP296198132, +0.296198132726023843175338011893050938967728390);
Chris@42 164 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@42 165 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@42 166 {
Chris@42 167 INT i;
Chris@42 168 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(36, rs), MAKE_VOLATILE_STRIDE(36, csr), MAKE_VOLATILE_STRIDE(36, csi)) {
Chris@42 169 E T1, T4, To, Ta, Tl, Tk, Tf, Ti, Th, T2, T3, T5, Tg;
Chris@42 170 T1 = R0[0];
Chris@42 171 T2 = R1[WS(rs, 1)];
Chris@42 172 T3 = R0[WS(rs, 3)];
Chris@42 173 T4 = T2 - T3;
Chris@42 174 To = T2 + T3;
Chris@42 175 {
Chris@42 176 E T6, T7, T8, T9;
Chris@42 177 T6 = R0[WS(rs, 1)];
Chris@42 178 T7 = R1[WS(rs, 2)];
Chris@42 179 T8 = R0[WS(rs, 4)];
Chris@42 180 T9 = T7 - T8;
Chris@42 181 Ta = T6 - T9;
Chris@42 182 Tl = T7 + T8;
Chris@42 183 Tk = FMA(KP500000000, T9, T6);
Chris@42 184 }
Chris@42 185 {
Chris@42 186 E Tb, Tc, Td, Te;
Chris@42 187 Tb = R0[WS(rs, 2)];
Chris@42 188 Tc = R1[0];
Chris@42 189 Td = R1[WS(rs, 3)];
Chris@42 190 Te = Tc + Td;
Chris@42 191 Tf = Tb - Te;
Chris@42 192 Ti = FMA(KP500000000, Te, Tb);
Chris@42 193 Th = Tc - Td;
Chris@42 194 }
Chris@42 195 Ci[WS(csi, 1)] = KP866025403 * (Tf - Ta);
Chris@42 196 T5 = T1 - T4;
Chris@42 197 Tg = Ta + Tf;
Chris@42 198 Cr[WS(csr, 1)] = FNMS(KP500000000, Tg, T5);
Chris@42 199 Cr[WS(csr, 4)] = T5 + Tg;
Chris@42 200 {
Chris@42 201 E Tr, Tt, Tw, Tv, Tu, Tp, Tq, Ts, Tj, Tm, Tn;
Chris@42 202 Tr = FMA(KP500000000, T4, T1);
Chris@42 203 Tt = FMA(KP296198132, Th, KP939692620 * Ti);
Chris@42 204 Tw = FNMS(KP813797681, Th, KP342020143 * Ti);
Chris@42 205 Tv = FNMS(KP984807753, Tk, KP150383733 * Tl);
Chris@42 206 Tu = FMA(KP173648177, Tk, KP852868531 * Tl);
Chris@42 207 Tp = FNMS(KP556670399, Tl, KP766044443 * Tk);
Chris@42 208 Tq = FMA(KP852868531, Th, KP173648177 * Ti);
Chris@42 209 Ts = Tp + Tq;
Chris@42 210 Tj = FNMS(KP984807753, Ti, KP150383733 * Th);
Chris@42 211 Tm = FMA(KP642787609, Tk, KP663413948 * Tl);
Chris@42 212 Tn = Tj - Tm;
Chris@42 213 Ci[0] = FNMS(KP866025403, To, Tn);
Chris@42 214 Cr[0] = Tr + Ts;
Chris@42 215 Ci[WS(csi, 3)] = FNMS(KP500000000, Tn, KP866025403 * ((Tp - Tq) - To));
Chris@42 216 Cr[WS(csr, 3)] = FMA(KP866025403, Tm + Tj, Tr) - (KP500000000 * Ts);
Chris@42 217 Ci[WS(csi, 2)] = FMA(KP866025403, To - (Tu + Tt), KP500000000 * (Tw - Tv));
Chris@42 218 Cr[WS(csr, 2)] = FMA(KP500000000, Tt - Tu, Tr) + (KP866025403 * (Tv + Tw));
Chris@42 219 }
Chris@42 220 }
Chris@42 221 }
Chris@42 222 }
Chris@42 223
Chris@42 224 static const kr2c_desc desc = { 9, "r2cfII_9", {25, 13, 17, 0}, &GENUS };
Chris@42 225
Chris@42 226 void X(codelet_r2cfII_9) (planner *p) {
Chris@42 227 X(kr2c_register) (p, r2cfII_9, &desc);
Chris@42 228 }
Chris@42 229
Chris@42 230 #endif /* HAVE_FMA */