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 -n 15 -name X(spu_n2fv_15) */ d@0: d@0: /* d@0: * This function contains 78 FP additions, 49 FP multiplications, d@0: * (or, 36 additions, 7 multiplications, 42 fused multiply/add), d@0: * 89 stack variables, 8 constants, and 30 memory accesses d@0: */ d@0: #include "fftw-spu.h" d@0: d@0: void X(spu_n2fv_15) (const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) { d@0: DVK(KP910592997, +0.910592997310029334643087372129977886038870291); d@0: DVK(KP823639103, +0.823639103546331925877420039278190003029660514); d@0: DVK(KP559016994, +0.559016994374947424102293417182819058860154590); d@0: DVK(KP618033988, +0.618033988749894848204586834365638117720309180); d@0: DVK(KP951056516, +0.951056516295153572116439333379382143405698634); d@0: DVK(KP866025403, +0.866025403784438646763723170752936183471402627); d@0: DVK(KP250000000, +0.250000000000000000000000000000000000000000000); d@0: DVK(KP500000000, +0.500000000000000000000000000000000000000000000); 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 TX, TB, T5, TO, TU, TV, TR, Tg, Tx, Tw, Tr, TI, TK, T12, T10; d@0: V T1, T4, T2, T3, TM, TC, Ta, TQ, TG, Tq, TN, TD, Tf, TP, TF; d@0: V Tl, T6, T9, T7, T8, Tm, Tp, Tn, To, Tb, Te, Tc, Td, Th, Tk; d@0: V Ti, Tj, TY, TE, TH, TZ, TJ, T11, T1f, T1g, Ts, Tu, Tt, Ty, TA; d@0: V Tv, Tz, T18, TS, TW, T1a, T13, T1b, TL, T17, T15, T16, T19, T1c, TT; d@0: V T14, T1d, T1e; d@0: T1 = LD(&(xi[0]), ivs, &(xi[0])); d@0: T2 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); d@0: T3 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); d@0: T4 = VADD(T2, T3); d@0: TX = VSUB(T3, T2); d@0: TB = VFNMS(LDK(KP500000000), T4, T1); d@0: T5 = VADD(T1, T4); d@0: T6 = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); d@0: T7 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); d@0: T8 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); d@0: TM = VSUB(T8, T7); d@0: T9 = VADD(T7, T8); d@0: TC = VFNMS(LDK(KP500000000), T9, T6); d@0: Ta = VADD(T6, T9); d@0: Tm = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); d@0: Tn = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); d@0: To = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); d@0: TQ = VSUB(To, Tn); d@0: Tp = VADD(Tn, To); d@0: TG = VFNMS(LDK(KP500000000), Tp, Tm); d@0: Tq = VADD(Tm, Tp); d@0: Tb = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); d@0: Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); d@0: Td = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); d@0: TN = VSUB(Td, Tc); d@0: Te = VADD(Tc, Td); d@0: TD = VFNMS(LDK(KP500000000), Te, Tb); d@0: Tf = VADD(Tb, Te); d@0: Th = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); d@0: Ti = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); d@0: Tj = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); d@0: TP = VSUB(Tj, Ti); d@0: Tk = VADD(Ti, Tj); d@0: TF = VFNMS(LDK(KP500000000), Tk, Th); d@0: Tl = VADD(Th, Tk); d@0: TO = VSUB(TM, TN); d@0: TY = VADD(TM, TN); d@0: TE = VADD(TC, TD); d@0: TU = VSUB(TC, TD); d@0: TV = VSUB(TF, TG); d@0: TH = VADD(TF, TG); d@0: TZ = VADD(TP, TQ); d@0: TR = VSUB(TP, TQ); d@0: Tg = VADD(Ta, Tf); d@0: Tx = VSUB(Ta, Tf); d@0: Tw = VSUB(Tl, Tq); d@0: Tr = VADD(Tl, Tq); d@0: TI = VADD(TE, TH); d@0: TK = VSUB(TE, TH); d@0: T12 = VSUB(TY, TZ); d@0: T10 = VADD(TY, TZ); d@0: TJ = VFNMS(LDK(KP250000000), TI, TB); d@0: T1f = VADD(TB, TI); d@0: T1g = VMUL(LDK(KP866025403), VADD(TX, T10)); d@0: T11 = VFNMS(LDK(KP250000000), T10, TX); d@0: ST(&(xo[10]), VFNMSI(T1g, T1f), ovs, &(xo[2])); d@0: ST(&(xo[20]), VFMAI(T1g, T1f), ovs, &(xo[0])); d@0: Ts = VADD(Tg, Tr); d@0: Tu = VSUB(Tg, Tr); d@0: Tt = VFNMS(LDK(KP250000000), Ts, T5); d@0: Ty = VMUL(LDK(KP951056516), VFNMS(LDK(KP618033988), Tx, Tw)); d@0: TA = VMUL(LDK(KP951056516), VFMA(LDK(KP618033988), Tw, Tx)); d@0: ST(&(xo[0]), VADD(T5, Ts), ovs, &(xo[0])); d@0: Tv = VFNMS(LDK(KP559016994), Tu, Tt); d@0: Tz = VFMA(LDK(KP559016994), Tu, Tt); d@0: ST(&(xo[12]), VFNMSI(TA, Tz), ovs, &(xo[0])); d@0: ST(&(xo[18]), VFMAI(TA, Tz), ovs, &(xo[2])); d@0: ST(&(xo[6]), VFNMSI(Ty, Tv), ovs, &(xo[2])); d@0: ST(&(xo[24]), VFMAI(Ty, Tv), ovs, &(xo[0])); d@0: T18 = VFNMS(LDK(KP618033988), TO, TR); d@0: TS = VFMA(LDK(KP618033988), TR, TO); d@0: TW = VFMA(LDK(KP618033988), TV, TU); d@0: T1a = VFNMS(LDK(KP618033988), TU, TV); d@0: T13 = VFMA(LDK(KP559016994), T12, T11); d@0: T1b = VFNMS(LDK(KP559016994), T12, T11); d@0: TL = VFMA(LDK(KP559016994), TK, TJ); d@0: T17 = VFNMS(LDK(KP559016994), TK, TJ); d@0: TT = VFMA(LDK(KP823639103), TS, TL); d@0: T15 = VFNMS(LDK(KP823639103), TS, TL); d@0: T16 = VMUL(LDK(KP951056516), VFMA(LDK(KP910592997), T13, TW)); d@0: T14 = VMUL(LDK(KP951056516), VFNMS(LDK(KP910592997), T13, TW)); d@0: ST(&(xo[2]), VFNMSI(T14, TT), ovs, &(xo[2])); d@0: ST(&(xo[28]), VFMAI(T14, TT), ovs, &(xo[0])); d@0: T1d = VFNMS(LDK(KP823639103), T18, T17); d@0: T19 = VFMA(LDK(KP823639103), T18, T17); d@0: T1c = VMUL(LDK(KP951056516), VFNMS(LDK(KP910592997), T1b, T1a)); d@0: T1e = VMUL(LDK(KP951056516), VFMA(LDK(KP910592997), T1b, T1a)); d@0: ST(&(xo[16]), VFNMSI(T1e, T1d), ovs, &(xo[0])); d@0: ST(&(xo[14]), VFMAI(T1e, T1d), ovs, &(xo[2])); d@0: ST(&(xo[22]), VFNMSI(T16, T15), ovs, &(xo[2])); d@0: ST(&(xo[8]), VFMAI(T16, T15), ovs, &(xo[0])); d@0: ST(&(xo[26]), VFNMSI(T1c, T19), ovs, &(xo[2])); d@0: ST(&(xo[4]), VFMAI(T1c, T19), ovs, &(xo[0])); d@0: } d@0: }