Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.3/dft/simd/common/n1fv_11.c @ 111:585a7be09763

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
FFTW
author Chris Cannam <cannam@all-day-breakfast.com>
date Fri, 04 Jul 2014 10:36:59 +0100
parents 89f5e221ed7b
children
rev   line source
cannam@95 1 /*
cannam@95 2 * Copyright (c) 2003, 2007-11 Matteo Frigo
cannam@95 3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
cannam@95 4 *
cannam@95 5 * This program is free software; you can redistribute it and/or modify
cannam@95 6 * it under the terms of the GNU General Public License as published by
cannam@95 7 * the Free Software Foundation; either version 2 of the License, or
cannam@95 8 * (at your option) any later version.
cannam@95 9 *
cannam@95 10 * This program is distributed in the hope that it will be useful,
cannam@95 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@95 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
cannam@95 13 * GNU General Public License for more details.
cannam@95 14 *
cannam@95 15 * You should have received a copy of the GNU General Public License
cannam@95 16 * along with this program; if not, write to the Free Software
cannam@95 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
cannam@95 18 *
cannam@95 19 */
cannam@95 20
cannam@95 21 /* This file was automatically generated --- DO NOT EDIT */
cannam@95 22 /* Generated on Sun Nov 25 07:36:52 EST 2012 */
cannam@95 23
cannam@95 24 #include "codelet-dft.h"
cannam@95 25
cannam@95 26 #ifdef HAVE_FMA
cannam@95 27
cannam@95 28 /* Generated by: ../../../genfft/gen_notw_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -n 11 -name n1fv_11 -include n1f.h */
cannam@95 29
cannam@95 30 /*
cannam@95 31 * This function contains 70 FP additions, 60 FP multiplications,
cannam@95 32 * (or, 15 additions, 5 multiplications, 55 fused multiply/add),
cannam@95 33 * 67 stack variables, 11 constants, and 22 memory accesses
cannam@95 34 */
cannam@95 35 #include "n1f.h"
cannam@95 36
cannam@95 37 static void n1fv_11(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
cannam@95 38 {
cannam@95 39 DVK(KP959492973, +0.959492973614497389890368057066327699062454848);
cannam@95 40 DVK(KP876768831, +0.876768831002589333891339807079336796764054852);
cannam@95 41 DVK(KP918985947, +0.918985947228994779780736114132655398124909697);
cannam@95 42 DVK(KP989821441, +0.989821441880932732376092037776718787376519372);
cannam@95 43 DVK(KP778434453, +0.778434453334651800608337670740821884709317477);
cannam@95 44 DVK(KP830830026, +0.830830026003772851058548298459246407048009821);
cannam@95 45 DVK(KP372785597, +0.372785597771792209609773152906148328659002598);
cannam@95 46 DVK(KP634356270, +0.634356270682424498893150776899916060542806975);
cannam@95 47 DVK(KP715370323, +0.715370323453429719112414662767260662417897278);
cannam@95 48 DVK(KP342584725, +0.342584725681637509502641509861112333758894680);
cannam@95 49 DVK(KP521108558, +0.521108558113202722944698153526659300680427422);
cannam@95 50 {
cannam@95 51 INT i;
cannam@95 52 const R *xi;
cannam@95 53 R *xo;
cannam@95 54 xi = ri;
cannam@95 55 xo = ro;
cannam@95 56 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(22, is), MAKE_VOLATILE_STRIDE(22, os)) {
cannam@95 57 V T1, Tb, T4, Tp, Tg, Tq, T7, Tn, Ta, Tm, Tc, Tr;
cannam@95 58 T1 = LD(&(xi[0]), ivs, &(xi[0]));
cannam@95 59 {
cannam@95 60 V T2, T3, Te, Tf;
cannam@95 61 T2 = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
cannam@95 62 T3 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
cannam@95 63 Te = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
cannam@95 64 Tf = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
cannam@95 65 {
cannam@95 66 V T5, T6, T8, T9;
cannam@95 67 T5 = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
cannam@95 68 T6 = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
cannam@95 69 T8 = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
cannam@95 70 T9 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
cannam@95 71 Tb = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
cannam@95 72 T4 = VADD(T2, T3);
cannam@95 73 Tp = VSUB(T3, T2);
cannam@95 74 Tg = VADD(Te, Tf);
cannam@95 75 Tq = VSUB(Tf, Te);
cannam@95 76 T7 = VADD(T5, T6);
cannam@95 77 Tn = VSUB(T6, T5);
cannam@95 78 Ta = VADD(T8, T9);
cannam@95 79 Tm = VSUB(T9, T8);
cannam@95 80 Tc = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
cannam@95 81 }
cannam@95 82 }
cannam@95 83 Tr = VFMA(LDK(KP521108558), Tq, Tp);
cannam@95 84 {
cannam@95 85 V TS, TE, Th, Td, To, T12, TO, TB, T11, TN, TA, TF;
cannam@95 86 T11 = VFNMS(LDK(KP521108558), Tp, Tn);
cannam@95 87 TN = VFNMS(LDK(KP342584725), T7, Tg);
cannam@95 88 TA = VFMA(LDK(KP521108558), Tm, Tq);
cannam@95 89 TS = VFMA(LDK(KP715370323), Tm, Tp);
cannam@95 90 TE = VFNMS(LDK(KP342584725), T4, Ta);
cannam@95 91 Th = VFNMS(LDK(KP342584725), Ta, T7);
cannam@95 92 Td = VADD(Tb, Tc);
cannam@95 93 To = VSUB(Tc, Tb);
cannam@95 94 T12 = VFNMS(LDK(KP715370323), T11, Tm);
cannam@95 95 TO = VFNMS(LDK(KP634356270), TN, T4);
cannam@95 96 TB = VFNMS(LDK(KP715370323), TA, Tn);
cannam@95 97 TF = VFNMS(LDK(KP634356270), TE, Tg);
cannam@95 98 {
cannam@95 99 V T14, TD, TV, Tu, TY, Tx, Tk, TR, TI, TM, TJ, TT, Ts;
cannam@95 100 TJ = VFNMS(LDK(KP521108558), Tn, To);
cannam@95 101 TT = VFMA(LDK(KP372785597), To, TS);
cannam@95 102 Ts = VFMA(LDK(KP715370323), Tr, To);
cannam@95 103 ST(&(xo[0]), VADD(T1, VADD(T4, VADD(T7, VADD(Ta, VADD(Td, Tg))))), ovs, &(xo[0]));
cannam@95 104 {
cannam@95 105 V TW, Tv, Ti, T13;
cannam@95 106 TW = VFNMS(LDK(KP342584725), Tg, Td);
cannam@95 107 Tv = VFNMS(LDK(KP342584725), Td, T4);
cannam@95 108 Ti = VFNMS(LDK(KP634356270), Th, Td);
cannam@95 109 T13 = VFNMS(LDK(KP830830026), T12, To);
cannam@95 110 {
cannam@95 111 V TP, TC, TG, TK;
cannam@95 112 TP = VFNMS(LDK(KP778434453), TO, Ta);
cannam@95 113 TC = VFMA(LDK(KP830830026), TB, Tp);
cannam@95 114 TG = VFNMS(LDK(KP778434453), TF, Td);
cannam@95 115 TK = VFMA(LDK(KP715370323), TJ, Tq);
cannam@95 116 {
cannam@95 117 V TU, Tt, TX, Tw;
cannam@95 118 TU = VFNMS(LDK(KP830830026), TT, Tq);
cannam@95 119 Tt = VFMA(LDK(KP830830026), Ts, Tn);
cannam@95 120 TX = VFNMS(LDK(KP634356270), TW, Ta);
cannam@95 121 Tw = VFNMS(LDK(KP634356270), Tv, T7);
cannam@95 122 {
cannam@95 123 V Tj, TQ, TH, TL;
cannam@95 124 Tj = VFNMS(LDK(KP778434453), Ti, T4);
cannam@95 125 T14 = VMUL(LDK(KP989821441), VFNMS(LDK(KP918985947), T13, Tq));
cannam@95 126 TQ = VFNMS(LDK(KP876768831), TP, Td);
cannam@95 127 TD = VMUL(LDK(KP989821441), VFNMS(LDK(KP918985947), TC, To));
cannam@95 128 TH = VFNMS(LDK(KP876768831), TG, T7);
cannam@95 129 TL = VFNMS(LDK(KP830830026), TK, Tm);
cannam@95 130 TV = VMUL(LDK(KP989821441), VFMA(LDK(KP918985947), TU, Tn));
cannam@95 131 Tu = VMUL(LDK(KP989821441), VFMA(LDK(KP918985947), Tt, Tm));
cannam@95 132 TY = VFNMS(LDK(KP778434453), TX, T7);
cannam@95 133 Tx = VFNMS(LDK(KP778434453), Tw, Tg);
cannam@95 134 Tk = VFNMS(LDK(KP876768831), Tj, Tg);
cannam@95 135 TR = VFNMS(LDK(KP959492973), TQ, T1);
cannam@95 136 TI = VFNMS(LDK(KP959492973), TH, T1);
cannam@95 137 TM = VMUL(LDK(KP989821441), VFNMS(LDK(KP918985947), TL, Tp));
cannam@95 138 }
cannam@95 139 }
cannam@95 140 }
cannam@95 141 }
cannam@95 142 {
cannam@95 143 V TZ, Ty, Tl, T10, Tz;
cannam@95 144 TZ = VFNMS(LDK(KP876768831), TY, T4);
cannam@95 145 Ty = VFNMS(LDK(KP876768831), Tx, Ta);
cannam@95 146 Tl = VFNMS(LDK(KP959492973), Tk, T1);
cannam@95 147 ST(&(xo[WS(os, 7)]), VFMAI(TV, TR), ovs, &(xo[WS(os, 1)]));
cannam@95 148 ST(&(xo[WS(os, 4)]), VFNMSI(TV, TR), ovs, &(xo[0]));
cannam@95 149 ST(&(xo[WS(os, 3)]), VFMAI(TM, TI), ovs, &(xo[WS(os, 1)]));
cannam@95 150 ST(&(xo[WS(os, 8)]), VFNMSI(TM, TI), ovs, &(xo[0]));
cannam@95 151 T10 = VFNMS(LDK(KP959492973), TZ, T1);
cannam@95 152 Tz = VFNMS(LDK(KP959492973), Ty, T1);
cannam@95 153 ST(&(xo[WS(os, 1)]), VFMAI(Tu, Tl), ovs, &(xo[WS(os, 1)]));
cannam@95 154 ST(&(xo[WS(os, 10)]), VFNMSI(Tu, Tl), ovs, &(xo[0]));
cannam@95 155 ST(&(xo[WS(os, 5)]), VFMAI(T14, T10), ovs, &(xo[WS(os, 1)]));
cannam@95 156 ST(&(xo[WS(os, 6)]), VFNMSI(T14, T10), ovs, &(xo[0]));
cannam@95 157 ST(&(xo[WS(os, 9)]), VFMAI(TD, Tz), ovs, &(xo[WS(os, 1)]));
cannam@95 158 ST(&(xo[WS(os, 2)]), VFNMSI(TD, Tz), ovs, &(xo[0]));
cannam@95 159 }
cannam@95 160 }
cannam@95 161 }
cannam@95 162 }
cannam@95 163 }
cannam@95 164 VLEAVE();
cannam@95 165 }
cannam@95 166
cannam@95 167 static const kdft_desc desc = { 11, XSIMD_STRING("n1fv_11"), {15, 5, 55, 0}, &GENUS, 0, 0, 0, 0 };
cannam@95 168
cannam@95 169 void XSIMD(codelet_n1fv_11) (planner *p) {
cannam@95 170 X(kdft_register) (p, n1fv_11, &desc);
cannam@95 171 }
cannam@95 172
cannam@95 173 #else /* HAVE_FMA */
cannam@95 174
cannam@95 175 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 11 -name n1fv_11 -include n1f.h */
cannam@95 176
cannam@95 177 /*
cannam@95 178 * This function contains 70 FP additions, 50 FP multiplications,
cannam@95 179 * (or, 30 additions, 10 multiplications, 40 fused multiply/add),
cannam@95 180 * 32 stack variables, 10 constants, and 22 memory accesses
cannam@95 181 */
cannam@95 182 #include "n1f.h"
cannam@95 183
cannam@95 184 static void n1fv_11(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
cannam@95 185 {
cannam@95 186 DVK(KP654860733, +0.654860733945285064056925072466293553183791199);
cannam@95 187 DVK(KP142314838, +0.142314838273285140443792668616369668791051361);
cannam@95 188 DVK(KP959492973, +0.959492973614497389890368057066327699062454848);
cannam@95 189 DVK(KP415415013, +0.415415013001886425529274149229623203524004910);
cannam@95 190 DVK(KP841253532, +0.841253532831181168861811648919367717513292498);
cannam@95 191 DVK(KP989821441, +0.989821441880932732376092037776718787376519372);
cannam@95 192 DVK(KP909631995, +0.909631995354518371411715383079028460060241051);
cannam@95 193 DVK(KP281732556, +0.281732556841429697711417915346616899035777899);
cannam@95 194 DVK(KP540640817, +0.540640817455597582107635954318691695431770608);
cannam@95 195 DVK(KP755749574, +0.755749574354258283774035843972344420179717445);
cannam@95 196 {
cannam@95 197 INT i;
cannam@95 198 const R *xi;
cannam@95 199 R *xo;
cannam@95 200 xi = ri;
cannam@95 201 xo = ro;
cannam@95 202 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(22, is), MAKE_VOLATILE_STRIDE(22, os)) {
cannam@95 203 V T1, T4, Ti, Tg, Tl, Td, Tk, Ta, Tj, T7, Tm, Tb, Tc, Tt, Ts;
cannam@95 204 T1 = LD(&(xi[0]), ivs, &(xi[0]));
cannam@95 205 {
cannam@95 206 V T2, T3, Te, Tf;
cannam@95 207 T2 = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
cannam@95 208 T3 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
cannam@95 209 T4 = VADD(T2, T3);
cannam@95 210 Ti = VSUB(T3, T2);
cannam@95 211 Te = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
cannam@95 212 Tf = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
cannam@95 213 Tg = VADD(Te, Tf);
cannam@95 214 Tl = VSUB(Tf, Te);
cannam@95 215 }
cannam@95 216 Tb = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
cannam@95 217 Tc = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
cannam@95 218 Td = VADD(Tb, Tc);
cannam@95 219 Tk = VSUB(Tc, Tb);
cannam@95 220 {
cannam@95 221 V T8, T9, T5, T6;
cannam@95 222 T8 = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
cannam@95 223 T9 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
cannam@95 224 Ta = VADD(T8, T9);
cannam@95 225 Tj = VSUB(T9, T8);
cannam@95 226 T5 = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
cannam@95 227 T6 = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
cannam@95 228 T7 = VADD(T5, T6);
cannam@95 229 Tm = VSUB(T6, T5);
cannam@95 230 }
cannam@95 231 ST(&(xo[0]), VADD(T1, VADD(T4, VADD(T7, VADD(Ta, VADD(Td, Tg))))), ovs, &(xo[0]));
cannam@95 232 {
cannam@95 233 V Tn, Th, Tv, Tu;
cannam@95 234 Tn = VBYI(VFMA(LDK(KP755749574), Ti, VFMA(LDK(KP540640817), Tj, VFNMS(LDK(KP909631995), Tl, VFNMS(LDK(KP989821441), Tm, VMUL(LDK(KP281732556), Tk))))));
cannam@95 235 Th = VFMA(LDK(KP841253532), Ta, VFMA(LDK(KP415415013), Tg, VFNMS(LDK(KP959492973), Td, VFNMS(LDK(KP142314838), T7, VFNMS(LDK(KP654860733), T4, T1)))));
cannam@95 236 ST(&(xo[WS(os, 7)]), VSUB(Th, Tn), ovs, &(xo[WS(os, 1)]));
cannam@95 237 ST(&(xo[WS(os, 4)]), VADD(Th, Tn), ovs, &(xo[0]));
cannam@95 238 Tv = VBYI(VFMA(LDK(KP281732556), Ti, VFMA(LDK(KP755749574), Tj, VFNMS(LDK(KP909631995), Tk, VFNMS(LDK(KP540640817), Tm, VMUL(LDK(KP989821441), Tl))))));
cannam@95 239 Tu = VFMA(LDK(KP841253532), T7, VFMA(LDK(KP415415013), Td, VFNMS(LDK(KP142314838), Tg, VFNMS(LDK(KP654860733), Ta, VFNMS(LDK(KP959492973), T4, T1)))));
cannam@95 240 ST(&(xo[WS(os, 6)]), VSUB(Tu, Tv), ovs, &(xo[0]));
cannam@95 241 ST(&(xo[WS(os, 5)]), VADD(Tu, Tv), ovs, &(xo[WS(os, 1)]));
cannam@95 242 }
cannam@95 243 Tt = VBYI(VFMA(LDK(KP989821441), Ti, VFMA(LDK(KP540640817), Tk, VFNMS(LDK(KP909631995), Tj, VFNMS(LDK(KP281732556), Tm, VMUL(LDK(KP755749574), Tl))))));
cannam@95 244 Ts = VFMA(LDK(KP415415013), Ta, VFMA(LDK(KP841253532), Td, VFNMS(LDK(KP654860733), Tg, VFNMS(LDK(KP959492973), T7, VFNMS(LDK(KP142314838), T4, T1)))));
cannam@95 245 ST(&(xo[WS(os, 8)]), VSUB(Ts, Tt), ovs, &(xo[0]));
cannam@95 246 ST(&(xo[WS(os, 3)]), VADD(Ts, Tt), ovs, &(xo[WS(os, 1)]));
cannam@95 247 {
cannam@95 248 V Tr, Tq, Tp, To;
cannam@95 249 Tr = VBYI(VFMA(LDK(KP540640817), Ti, VFMA(LDK(KP909631995), Tm, VFMA(LDK(KP989821441), Tj, VFMA(LDK(KP755749574), Tk, VMUL(LDK(KP281732556), Tl))))));
cannam@95 250 Tq = VFMA(LDK(KP841253532), T4, VFMA(LDK(KP415415013), T7, VFNMS(LDK(KP959492973), Tg, VFNMS(LDK(KP654860733), Td, VFNMS(LDK(KP142314838), Ta, T1)))));
cannam@95 251 ST(&(xo[WS(os, 10)]), VSUB(Tq, Tr), ovs, &(xo[0]));
cannam@95 252 ST(&(xo[WS(os, 1)]), VADD(Tq, Tr), ovs, &(xo[WS(os, 1)]));
cannam@95 253 Tp = VBYI(VFMA(LDK(KP909631995), Ti, VFNMS(LDK(KP540640817), Tl, VFNMS(LDK(KP989821441), Tk, VFNMS(LDK(KP281732556), Tj, VMUL(LDK(KP755749574), Tm))))));
cannam@95 254 To = VFMA(LDK(KP415415013), T4, VFMA(LDK(KP841253532), Tg, VFNMS(LDK(KP142314838), Td, VFNMS(LDK(KP959492973), Ta, VFNMS(LDK(KP654860733), T7, T1)))));
cannam@95 255 ST(&(xo[WS(os, 9)]), VSUB(To, Tp), ovs, &(xo[WS(os, 1)]));
cannam@95 256 ST(&(xo[WS(os, 2)]), VADD(To, Tp), ovs, &(xo[0]));
cannam@95 257 }
cannam@95 258 }
cannam@95 259 }
cannam@95 260 VLEAVE();
cannam@95 261 }
cannam@95 262
cannam@95 263 static const kdft_desc desc = { 11, XSIMD_STRING("n1fv_11"), {30, 10, 40, 0}, &GENUS, 0, 0, 0, 0 };
cannam@95 264
cannam@95 265 void XSIMD(codelet_n1fv_11) (planner *p) {
cannam@95 266 X(kdft_register) (p, n1fv_11, &desc);
cannam@95 267 }
cannam@95 268
cannam@95 269 #endif /* HAVE_FMA */