d@0: /* d@0: * Copyright (c) 2003, 2007-8 Matteo Frigo d@0: * Copyright (c) 2003, 2007-8 Massachusetts Institute of Technology d@0: * d@0: * This program is free software; you can redistribute it and/or modify d@0: * it under the terms of the GNU General Public License as published by d@0: * the Free Software Foundation; either version 2 of the License, or d@0: * (at your option) any later version. d@0: * d@0: * This program is distributed in the hope that it will be useful, d@0: * but WITHOUT ANY WARRANTY; without even the implied warranty of d@0: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the d@0: * GNU General Public License for more details. d@0: * d@0: * You should have received a copy of the GNU General Public License d@0: * along with this program; if not, write to the Free Software d@0: * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA d@0: * d@0: */ d@0: /* Generated by: ../../genfft/gen_notw_c -standalone -fma -reorder-insns -simd -compact -variables 100000 -with-ostride 2 -include fftw-spu.h -store-multiple 2 -n 14 -name X(spu_n2fv_14) */ d@0: d@0: /* d@0: * This function contains 74 FP additions, 48 FP multiplications, d@0: * (or, 32 additions, 6 multiplications, 42 fused multiply/add), d@0: * 93 stack variables, 6 constants, and 35 memory accesses d@0: */ d@0: #include "fftw-spu.h" d@0: d@0: void X(spu_n2fv_14) (const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) { d@0: DVK(KP801937735, +0.801937735804838252472204639014890102331838324); d@0: DVK(KP974927912, +0.974927912181823607018131682993931217232785801); d@0: DVK(KP554958132, +0.554958132087371191422194871006410481067288862); d@0: DVK(KP900968867, +0.900968867902419126236102319507445051165919162); d@0: DVK(KP692021471, +0.692021471630095869627814897002069140197260599); d@0: DVK(KP356895867, +0.356895867892209443894399510021300583399127187); d@0: INT i; d@0: const R *xi; d@0: R *xo; d@0: xi = ri; d@0: xo = ro; d@0: for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) { d@0: V TH, T3, Ta, Ts, TV, TN, TK, TW, Tt, To, Th, Tu, TU, TQ, Tp; d@0: V TC, Tx, TR, T14, TZ, T1, T2, TI, T6, TJ, T9, TP, Tn, TO, Tk; d@0: V TM, Tg, TL, Td, T4, T5, T7, T8, Tl, Tm, Ti, Tj, Te, Tf, Tb; d@0: V Tc, T1a, T1b, T19, T1c, Tr, Tw, Tq, Tv, T16, T18, T15, T17, T1d, T1e; d@0: V T1f, T1i, Tz, TB, Ty, TA, T1g, TT, TY, TS, TX, T1h, T1j, T11, T13; d@0: V T10, T12, T1k, TE, TG, TD, TF, T1l, T1m; d@0: T1 = LD(&(xi[0]), ivs, &(xi[0])); d@0: T2 = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); d@0: TH = VADD(T1, T2); d@0: T3 = VSUB(T1, T2); d@0: T4 = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); d@0: T5 = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); d@0: TI = VADD(T4, T5); d@0: T6 = VSUB(T4, T5); d@0: T7 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); d@0: T8 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); d@0: TJ = VADD(T7, T8); d@0: T9 = VSUB(T7, T8); d@0: Tl = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); d@0: Tm = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); d@0: TP = VADD(Tl, Tm); d@0: Tn = VSUB(Tl, Tm); d@0: Ti = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); d@0: Tj = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); d@0: TO = VADD(Ti, Tj); d@0: Tk = VSUB(Ti, Tj); d@0: Te = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); d@0: Tf = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); d@0: TM = VADD(Te, Tf); d@0: Tg = VSUB(Te, Tf); d@0: Tb = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); d@0: Tc = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); d@0: TL = VADD(Tb, Tc); d@0: Td = VSUB(Tb, Tc); d@0: Ta = VADD(T6, T9); d@0: Ts = VSUB(T9, T6); d@0: TV = VSUB(TL, TM); d@0: TN = VADD(TL, TM); d@0: TK = VADD(TI, TJ); d@0: TW = VSUB(TJ, TI); d@0: Tt = VSUB(Tn, Tk); d@0: To = VADD(Tk, Tn); d@0: Th = VADD(Td, Tg); d@0: Tu = VSUB(Tg, Td); d@0: TU = VSUB(TO, TP); d@0: TQ = VADD(TO, TP); d@0: Tp = VFNMS(LDK(KP356895867), Ta, To); d@0: TC = VFNMS(LDK(KP356895867), To, Th); d@0: Tx = VFNMS(LDK(KP356895867), Th, Ta); d@0: TR = VFNMS(LDK(KP356895867), TQ, TN); d@0: T14 = VFNMS(LDK(KP356895867), TN, TK); d@0: TZ = VFNMS(LDK(KP356895867), TK, TQ); d@0: T19 = VADD(T3, VADD(Ta, VADD(Th, To))); d@0: STM2(&(xo[14]), T19, ovs, &(xo[2])); d@0: T1a = VADD(TH, VADD(TK, VADD(TN, TQ))); d@0: STM2(&(xo[0]), T1a, ovs, &(xo[0])); d@0: Tq = VFNMS(LDK(KP692021471), Tp, Th); d@0: Tr = VFNMS(LDK(KP900968867), Tq, T3); d@0: Tv = VFMA(LDK(KP554958132), Tu, Tt); d@0: Tw = VMUL(LDK(KP974927912), VFNMS(LDK(KP801937735), Tv, Ts)); d@0: T1b = VFNMSI(Tw, Tr); d@0: STM2(&(xo[10]), T1b, ovs, &(xo[2])); d@0: T1c = VFMAI(Tw, Tr); d@0: STM2(&(xo[18]), T1c, ovs, &(xo[2])); d@0: T15 = VFNMS(LDK(KP692021471), T14, TQ); d@0: T16 = VFNMS(LDK(KP900968867), T15, TH); d@0: T17 = VFNMS(LDK(KP554958132), TU, TW); d@0: T18 = VMUL(LDK(KP974927912), VFNMS(LDK(KP801937735), T17, TV)); d@0: T1d = VFMAI(T18, T16); d@0: STM2(&(xo[12]), T1d, ovs, &(xo[0])); d@0: STN2(&(xo[12]), T1d, T19, ovs); d@0: T1e = VFNMSI(T18, T16); d@0: STM2(&(xo[16]), T1e, ovs, &(xo[0])); d@0: STN2(&(xo[16]), T1e, T1c, ovs); d@0: Ty = VFNMS(LDK(KP692021471), Tx, To); d@0: Tz = VFNMS(LDK(KP900968867), Ty, T3); d@0: TA = VFMA(LDK(KP554958132), Tt, Ts); d@0: TB = VMUL(LDK(KP974927912), VFMA(LDK(KP801937735), TA, Tu)); d@0: T1f = VFNMSI(TB, Tz); d@0: STM2(&(xo[26]), T1f, ovs, &(xo[2])); d@0: T1g = VFMAI(TB, Tz); d@0: STM2(&(xo[2]), T1g, ovs, &(xo[2])); d@0: STN2(&(xo[0]), T1a, T1g, ovs); d@0: TS = VFNMS(LDK(KP692021471), TR, TK); d@0: TT = VFNMS(LDK(KP900968867), TS, TH); d@0: TX = VFMA(LDK(KP554958132), TW, TV); d@0: TY = VMUL(LDK(KP974927912), VFNMS(LDK(KP801937735), TX, TU)); d@0: T1h = VFMAI(TY, TT); d@0: STM2(&(xo[8]), T1h, ovs, &(xo[0])); d@0: STN2(&(xo[8]), T1h, T1b, ovs); d@0: T1i = VFNMSI(TY, TT); d@0: STM2(&(xo[20]), T1i, ovs, &(xo[0])); d@0: T10 = VFNMS(LDK(KP692021471), TZ, TN); d@0: T11 = VFNMS(LDK(KP900968867), T10, TH); d@0: T12 = VFMA(LDK(KP554958132), TV, TU); d@0: T13 = VMUL(LDK(KP974927912), VFMA(LDK(KP801937735), T12, TW)); d@0: T1j = VFMAI(T13, T11); d@0: STM2(&(xo[4]), T1j, ovs, &(xo[0])); d@0: T1k = VFNMSI(T13, T11); d@0: STM2(&(xo[24]), T1k, ovs, &(xo[0])); d@0: STN2(&(xo[24]), T1k, T1f, ovs); d@0: TD = VFNMS(LDK(KP692021471), TC, Ta); d@0: TE = VFNMS(LDK(KP900968867), TD, T3); d@0: TF = VFNMS(LDK(KP554958132), Ts, Tu); d@0: TG = VMUL(LDK(KP974927912), VFNMS(LDK(KP801937735), TF, Tt)); d@0: T1l = VFNMSI(TG, TE); d@0: STM2(&(xo[22]), T1l, ovs, &(xo[2])); d@0: STN2(&(xo[20]), T1i, T1l, ovs); d@0: T1m = VFMAI(TG, TE); d@0: STM2(&(xo[6]), T1m, ovs, &(xo[2])); d@0: STN2(&(xo[4]), T1j, T1m, ovs); d@0: } d@0: }