cannam@167: /*
cannam@167:  * Copyright (c) 2003, 2007-14 Matteo Frigo
cannam@167:  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
cannam@167:  *
cannam@167:  * This program is free software; you can redistribute it and/or modify
cannam@167:  * it under the terms of the GNU General Public License as published by
cannam@167:  * the Free Software Foundation; either version 2 of the License, or
cannam@167:  * (at your option) any later version.
cannam@167:  *
cannam@167:  * This program is distributed in the hope that it will be useful,
cannam@167:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@167:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
cannam@167:  * GNU General Public License for more details.
cannam@167:  *
cannam@167:  * You should have received a copy of the GNU General Public License
cannam@167:  * along with this program; if not, write to the Free Software
cannam@167:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
cannam@167:  *
cannam@167:  */
cannam@167: 
cannam@167: /* This file was automatically generated --- DO NOT EDIT */
cannam@167: /* Generated on Thu May 24 08:07:32 EDT 2018 */
cannam@167: 
cannam@167: #include "rdft/codelet-rdft.h"
cannam@167: 
cannam@167: #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
cannam@167: 
cannam@167: /* Generated by: ../../../genfft/gen_hc2hc.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -n 9 -dif -name hb_9 -include rdft/scalar/hb.h */
cannam@167: 
cannam@167: /*
cannam@167:  * This function contains 96 FP additions, 88 FP multiplications,
cannam@167:  * (or, 24 additions, 16 multiplications, 72 fused multiply/add),
cannam@167:  * 53 stack variables, 10 constants, and 36 memory accesses
cannam@167:  */
cannam@167: #include "rdft/scalar/hb.h"
cannam@167: 
cannam@167: static void hb_9(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
cannam@167: {
cannam@167:      DK(KP954188894, +0.954188894138671133499268364187245676532219158);
cannam@167:      DK(KP852868531, +0.852868531952443209628250963940074071936020296);
cannam@167:      DK(KP984807753, +0.984807753012208059366743024589523013670643252);
cannam@167:      DK(KP492403876, +0.492403876506104029683371512294761506835321626);
cannam@167:      DK(KP777861913, +0.777861913430206160028177977318626690410586096);
cannam@167:      DK(KP839099631, +0.839099631177280011763127298123181364687434283);
cannam@167:      DK(KP176326980, +0.176326980708464973471090386868618986121633062);
cannam@167:      DK(KP363970234, +0.363970234266202361351047882776834043890471784);
cannam@167:      DK(KP866025403, +0.866025403784438646763723170752936183471402627);
cannam@167:      DK(KP500000000, +0.500000000000000000000000000000000000000000000);
cannam@167:      {
cannam@167: 	  INT m;
cannam@167: 	  for (m = mb, W = W + ((mb - 1) * 16); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 16, MAKE_VOLATILE_STRIDE(18, rs)) {
cannam@167: 	       E T5, Tl, TQ, T1y, T1b, T1J, Tg, TE, Tw, Tz, T1E, T1L, T1B, T1K, T14;
cannam@167: 	       E T1d, TX, T1c;
cannam@167: 	       {
cannam@167: 		    E T1, Th, T4, T1a, Tk, TP, TO, T19;
cannam@167: 		    T1 = cr[0];
cannam@167: 		    Th = ci[WS(rs, 8)];
cannam@167: 		    {
cannam@167: 			 E T2, T3, Ti, Tj;
cannam@167: 			 T2 = cr[WS(rs, 3)];
cannam@167: 			 T3 = ci[WS(rs, 2)];
cannam@167: 			 T4 = T2 + T3;
cannam@167: 			 T1a = T2 - T3;
cannam@167: 			 Ti = ci[WS(rs, 5)];
cannam@167: 			 Tj = cr[WS(rs, 6)];
cannam@167: 			 Tk = Ti - Tj;
cannam@167: 			 TP = Ti + Tj;
cannam@167: 		    }
cannam@167: 		    T5 = T1 + T4;
cannam@167: 		    Tl = Th + Tk;
cannam@167: 		    TO = FNMS(KP500000000, T4, T1);
cannam@167: 		    TQ = FNMS(KP866025403, TP, TO);
cannam@167: 		    T1y = FMA(KP866025403, TP, TO);
cannam@167: 		    T19 = FNMS(KP500000000, Tk, Th);
cannam@167: 		    T1b = FMA(KP866025403, T1a, T19);
cannam@167: 		    T1J = FNMS(KP866025403, T1a, T19);
cannam@167: 	       }
cannam@167: 	       {
cannam@167: 		    E T6, T9, TY, T12, Tm, Tp, TZ, T11, Tb, Te, TS, TU, Tr, Tu, TR;
cannam@167: 		    E TV;
cannam@167: 		    {
cannam@167: 			 E T7, T8, Tn, To;
cannam@167: 			 T6 = cr[WS(rs, 1)];
cannam@167: 			 T7 = cr[WS(rs, 4)];
cannam@167: 			 T8 = ci[WS(rs, 1)];
cannam@167: 			 T9 = T7 + T8;
cannam@167: 			 TY = FNMS(KP500000000, T9, T6);
cannam@167: 			 T12 = T7 - T8;
cannam@167: 			 Tm = ci[WS(rs, 7)];
cannam@167: 			 Tn = ci[WS(rs, 4)];
cannam@167: 			 To = cr[WS(rs, 7)];
cannam@167: 			 Tp = Tn - To;
cannam@167: 			 TZ = Tn + To;
cannam@167: 			 T11 = FMS(KP500000000, Tp, Tm);
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E Tc, Td, Ts, Tt;
cannam@167: 			 Tb = cr[WS(rs, 2)];
cannam@167: 			 Tc = ci[WS(rs, 3)];
cannam@167: 			 Td = ci[0];
cannam@167: 			 Te = Tc + Td;
cannam@167: 			 TS = Td - Tc;
cannam@167: 			 TU = FNMS(KP500000000, Te, Tb);
cannam@167: 			 Tr = ci[WS(rs, 6)];
cannam@167: 			 Ts = cr[WS(rs, 5)];
cannam@167: 			 Tt = cr[WS(rs, 8)];
cannam@167: 			 Tu = Ts + Tt;
cannam@167: 			 TR = FMA(KP500000000, Tu, Tr);
cannam@167: 			 TV = Ts - Tt;
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E Ta, Tf, T1z, T1A;
cannam@167: 			 Ta = T6 + T9;
cannam@167: 			 Tf = Tb + Te;
cannam@167: 			 Tg = Ta + Tf;
cannam@167: 			 TE = Ta - Tf;
cannam@167: 			 {
cannam@167: 			      E Tq, Tv, T1C, T1D;
cannam@167: 			      Tq = Tm + Tp;
cannam@167: 			      Tv = Tr - Tu;
cannam@167: 			      Tw = Tq + Tv;
cannam@167: 			      Tz = Tv - Tq;
cannam@167: 			      T1C = FNMS(KP866025403, TV, TU);
cannam@167: 			      T1D = FMA(KP866025403, TS, TR);
cannam@167: 			      T1E = FMA(KP363970234, T1D, T1C);
cannam@167: 			      T1L = FNMS(KP363970234, T1C, T1D);
cannam@167: 			 }
cannam@167: 			 T1z = FMA(KP866025403, T12, T11);
cannam@167: 			 T1A = FMA(KP866025403, TZ, TY);
cannam@167: 			 T1B = FMA(KP176326980, T1A, T1z);
cannam@167: 			 T1K = FNMS(KP176326980, T1z, T1A);
cannam@167: 			 {
cannam@167: 			      E T10, T13, TT, TW;
cannam@167: 			      T10 = FNMS(KP866025403, TZ, TY);
cannam@167: 			      T13 = FNMS(KP866025403, T12, T11);
cannam@167: 			      T14 = FMA(KP839099631, T13, T10);
cannam@167: 			      T1d = FNMS(KP839099631, T10, T13);
cannam@167: 			      TT = FNMS(KP866025403, TS, TR);
cannam@167: 			      TW = FMA(KP866025403, TV, TU);
cannam@167: 			      TX = FNMS(KP176326980, TW, TT);
cannam@167: 			      T1c = FMA(KP176326980, TT, TW);
cannam@167: 			 }
cannam@167: 		    }
cannam@167: 	       }
cannam@167: 	       cr[0] = T5 + Tg;
cannam@167: 	       ci[0] = Tl + Tw;
cannam@167: 	       {
cannam@167: 		    E TA, TI, TF, TL, Ty, TD;
cannam@167: 		    Ty = FNMS(KP500000000, Tg, T5);
cannam@167: 		    TA = FNMS(KP866025403, Tz, Ty);
cannam@167: 		    TI = FMA(KP866025403, Tz, Ty);
cannam@167: 		    TD = FNMS(KP500000000, Tw, Tl);
cannam@167: 		    TF = FNMS(KP866025403, TE, TD);
cannam@167: 		    TL = FMA(KP866025403, TE, TD);
cannam@167: 		    {
cannam@167: 			 E TB, TG, Tx, TC;
cannam@167: 			 Tx = W[10];
cannam@167: 			 TB = Tx * TA;
cannam@167: 			 TG = Tx * TF;
cannam@167: 			 TC = W[11];
cannam@167: 			 cr[WS(rs, 6)] = FNMS(TC, TF, TB);
cannam@167: 			 ci[WS(rs, 6)] = FMA(TC, TA, TG);
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E TJ, TM, TH, TK;
cannam@167: 			 TH = W[4];
cannam@167: 			 TJ = TH * TI;
cannam@167: 			 TM = TH * TL;
cannam@167: 			 TK = W[5];
cannam@167: 			 cr[WS(rs, 3)] = FNMS(TK, TL, TJ);
cannam@167: 			 ci[WS(rs, 3)] = FMA(TK, TI, TM);
cannam@167: 		    }
cannam@167: 	       }
cannam@167: 	       {
cannam@167: 		    E T16, T1s, T1k, T1f, T1v, T1p;
cannam@167: 		    {
cannam@167: 			 E T1j, T15, T1i, T1o, T1e, T1n;
cannam@167: 			 T1j = FMA(KP777861913, T1d, T1c);
cannam@167: 			 T15 = FNMS(KP777861913, T14, TX);
cannam@167: 			 T1i = FMA(KP492403876, T15, TQ);
cannam@167: 			 T16 = FNMS(KP984807753, T15, TQ);
cannam@167: 			 T1s = FMA(KP852868531, T1j, T1i);
cannam@167: 			 T1k = FNMS(KP852868531, T1j, T1i);
cannam@167: 			 T1o = FMA(KP777861913, T14, TX);
cannam@167: 			 T1e = FNMS(KP777861913, T1d, T1c);
cannam@167: 			 T1n = FNMS(KP492403876, T1e, T1b);
cannam@167: 			 T1f = FMA(KP984807753, T1e, T1b);
cannam@167: 			 T1v = FMA(KP852868531, T1o, T1n);
cannam@167: 			 T1p = FNMS(KP852868531, T1o, T1n);
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E TN, T17, T18, T1g;
cannam@167: 			 TN = W[0];
cannam@167: 			 T17 = TN * T16;
cannam@167: 			 T18 = W[1];
cannam@167: 			 T1g = T18 * T16;
cannam@167: 			 cr[WS(rs, 1)] = FNMS(T18, T1f, T17);
cannam@167: 			 ci[WS(rs, 1)] = FMA(TN, T1f, T1g);
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E T1t, T1w, T1r, T1u;
cannam@167: 			 T1r = W[6];
cannam@167: 			 T1t = T1r * T1s;
cannam@167: 			 T1w = T1r * T1v;
cannam@167: 			 T1u = W[7];
cannam@167: 			 cr[WS(rs, 4)] = FNMS(T1u, T1v, T1t);
cannam@167: 			 ci[WS(rs, 4)] = FMA(T1u, T1s, T1w);
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E T1l, T1q, T1h, T1m;
cannam@167: 			 T1h = W[12];
cannam@167: 			 T1l = T1h * T1k;
cannam@167: 			 T1q = T1h * T1p;
cannam@167: 			 T1m = W[13];
cannam@167: 			 cr[WS(rs, 7)] = FNMS(T1m, T1p, T1l);
cannam@167: 			 ci[WS(rs, 7)] = FMA(T1m, T1k, T1q);
cannam@167: 		    }
cannam@167: 	       }
cannam@167: 	       {
cannam@167: 		    E T1W, T1N, T1V, T1G, T20, T1S;
cannam@167: 		    T1W = FMA(KP954188894, T1E, T1B);
cannam@167: 		    {
cannam@167: 			 E T1M, T1R, T1F, T1Q;
cannam@167: 			 T1M = FNMS(KP954188894, T1L, T1K);
cannam@167: 			 T1N = FMA(KP984807753, T1M, T1J);
cannam@167: 			 T1V = FNMS(KP492403876, T1M, T1J);
cannam@167: 			 T1R = FMA(KP954188894, T1L, T1K);
cannam@167: 			 T1F = FNMS(KP954188894, T1E, T1B);
cannam@167: 			 T1Q = FNMS(KP492403876, T1F, T1y);
cannam@167: 			 T1G = FMA(KP984807753, T1F, T1y);
cannam@167: 			 T20 = FMA(KP852868531, T1R, T1Q);
cannam@167: 			 T1S = FNMS(KP852868531, T1R, T1Q);
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E T1H, T1O, T1x, T1I;
cannam@167: 			 T1x = W[2];
cannam@167: 			 T1H = T1x * T1G;
cannam@167: 			 T1O = T1x * T1N;
cannam@167: 			 T1I = W[3];
cannam@167: 			 cr[WS(rs, 2)] = FNMS(T1I, T1N, T1H);
cannam@167: 			 ci[WS(rs, 2)] = FMA(T1I, T1G, T1O);
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E T23, T22, T24, T1Z, T21;
cannam@167: 			 T23 = FNMS(KP852868531, T1W, T1V);
cannam@167: 			 T22 = W[15];
cannam@167: 			 T24 = T22 * T20;
cannam@167: 			 T1Z = W[14];
cannam@167: 			 T21 = T1Z * T20;
cannam@167: 			 cr[WS(rs, 8)] = FNMS(T22, T23, T21);
cannam@167: 			 ci[WS(rs, 8)] = FMA(T1Z, T23, T24);
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E T1X, T1U, T1Y, T1P, T1T;
cannam@167: 			 T1X = FMA(KP852868531, T1W, T1V);
cannam@167: 			 T1U = W[9];
cannam@167: 			 T1Y = T1U * T1S;
cannam@167: 			 T1P = W[8];
cannam@167: 			 T1T = T1P * T1S;
cannam@167: 			 cr[WS(rs, 5)] = FNMS(T1U, T1X, T1T);
cannam@167: 			 ci[WS(rs, 5)] = FMA(T1P, T1X, T1Y);
cannam@167: 		    }
cannam@167: 	       }
cannam@167: 	  }
cannam@167:      }
cannam@167: }
cannam@167: 
cannam@167: static const tw_instr twinstr[] = {
cannam@167:      {TW_FULL, 1, 9},
cannam@167:      {TW_NEXT, 1, 0}
cannam@167: };
cannam@167: 
cannam@167: static const hc2hc_desc desc = { 9, "hb_9", twinstr, &GENUS, {24, 16, 72, 0} };
cannam@167: 
cannam@167: void X(codelet_hb_9) (planner *p) {
cannam@167:      X(khc2hc_register) (p, hb_9, &desc);
cannam@167: }
cannam@167: #else
cannam@167: 
cannam@167: /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 9 -dif -name hb_9 -include rdft/scalar/hb.h */
cannam@167: 
cannam@167: /*
cannam@167:  * This function contains 96 FP additions, 72 FP multiplications,
cannam@167:  * (or, 60 additions, 36 multiplications, 36 fused multiply/add),
cannam@167:  * 53 stack variables, 8 constants, and 36 memory accesses
cannam@167:  */
cannam@167: #include "rdft/scalar/hb.h"
cannam@167: 
cannam@167: static void hb_9(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
cannam@167: {
cannam@167:      DK(KP984807753, +0.984807753012208059366743024589523013670643252);
cannam@167:      DK(KP173648177, +0.173648177666930348851716626769314796000375677);
cannam@167:      DK(KP342020143, +0.342020143325668733044099614682259580763083368);
cannam@167:      DK(KP939692620, +0.939692620785908384054109277324731469936208134);
cannam@167:      DK(KP642787609, +0.642787609686539326322643409907263432907559884);
cannam@167:      DK(KP766044443, +0.766044443118978035202392650555416673935832457);
cannam@167:      DK(KP500000000, +0.500000000000000000000000000000000000000000000);
cannam@167:      DK(KP866025403, +0.866025403784438646763723170752936183471402627);
cannam@167:      {
cannam@167: 	  INT m;
cannam@167: 	  for (m = mb, W = W + ((mb - 1) * 16); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 16, MAKE_VOLATILE_STRIDE(18, rs)) {
cannam@167: 	       E T5, Tl, TM, T1o, T16, T1y, Ta, Tf, Tg, Tq, Tv, Tw, TT, T17, T1u;
cannam@167: 	       E T1A, T1r, T1z, T10, T18;
cannam@167: 	       {
cannam@167: 		    E T1, Th, T4, T14, Tk, TL, TK, T15;
cannam@167: 		    T1 = cr[0];
cannam@167: 		    Th = ci[WS(rs, 8)];
cannam@167: 		    {
cannam@167: 			 E T2, T3, Ti, Tj;
cannam@167: 			 T2 = cr[WS(rs, 3)];
cannam@167: 			 T3 = ci[WS(rs, 2)];
cannam@167: 			 T4 = T2 + T3;
cannam@167: 			 T14 = KP866025403 * (T2 - T3);
cannam@167: 			 Ti = ci[WS(rs, 5)];
cannam@167: 			 Tj = cr[WS(rs, 6)];
cannam@167: 			 Tk = Ti - Tj;
cannam@167: 			 TL = KP866025403 * (Ti + Tj);
cannam@167: 		    }
cannam@167: 		    T5 = T1 + T4;
cannam@167: 		    Tl = Th + Tk;
cannam@167: 		    TK = FNMS(KP500000000, T4, T1);
cannam@167: 		    TM = TK - TL;
cannam@167: 		    T1o = TK + TL;
cannam@167: 		    T15 = FNMS(KP500000000, Tk, Th);
cannam@167: 		    T16 = T14 + T15;
cannam@167: 		    T1y = T15 - T14;
cannam@167: 	       }
cannam@167: 	       {
cannam@167: 		    E T6, T9, TN, TQ, Tm, Tp, TO, TR, Tb, Te, TU, TX, Tr, Tu, TV;
cannam@167: 		    E TY;
cannam@167: 		    {
cannam@167: 			 E T7, T8, Tn, To;
cannam@167: 			 T6 = cr[WS(rs, 1)];
cannam@167: 			 T7 = cr[WS(rs, 4)];
cannam@167: 			 T8 = ci[WS(rs, 1)];
cannam@167: 			 T9 = T7 + T8;
cannam@167: 			 TN = FNMS(KP500000000, T9, T6);
cannam@167: 			 TQ = KP866025403 * (T7 - T8);
cannam@167: 			 Tm = ci[WS(rs, 7)];
cannam@167: 			 Tn = ci[WS(rs, 4)];
cannam@167: 			 To = cr[WS(rs, 7)];
cannam@167: 			 Tp = Tn - To;
cannam@167: 			 TO = KP866025403 * (Tn + To);
cannam@167: 			 TR = FNMS(KP500000000, Tp, Tm);
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E Tc, Td, Ts, Tt;
cannam@167: 			 Tb = cr[WS(rs, 2)];
cannam@167: 			 Tc = ci[WS(rs, 3)];
cannam@167: 			 Td = ci[0];
cannam@167: 			 Te = Tc + Td;
cannam@167: 			 TU = FNMS(KP500000000, Te, Tb);
cannam@167: 			 TX = KP866025403 * (Tc - Td);
cannam@167: 			 Tr = ci[WS(rs, 6)];
cannam@167: 			 Ts = cr[WS(rs, 5)];
cannam@167: 			 Tt = cr[WS(rs, 8)];
cannam@167: 			 Tu = Ts + Tt;
cannam@167: 			 TV = KP866025403 * (Ts - Tt);
cannam@167: 			 TY = FMA(KP500000000, Tu, Tr);
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E TP, TS, T1s, T1t;
cannam@167: 			 Ta = T6 + T9;
cannam@167: 			 Tf = Tb + Te;
cannam@167: 			 Tg = Ta + Tf;
cannam@167: 			 Tq = Tm + Tp;
cannam@167: 			 Tv = Tr - Tu;
cannam@167: 			 Tw = Tq + Tv;
cannam@167: 			 TP = TN - TO;
cannam@167: 			 TS = TQ + TR;
cannam@167: 			 TT = FNMS(KP642787609, TS, KP766044443 * TP);
cannam@167: 			 T17 = FMA(KP766044443, TS, KP642787609 * TP);
cannam@167: 			 T1s = TU - TV;
cannam@167: 			 T1t = TY - TX;
cannam@167: 			 T1u = FMA(KP939692620, T1s, KP342020143 * T1t);
cannam@167: 			 T1A = FNMS(KP939692620, T1t, KP342020143 * T1s);
cannam@167: 			 {
cannam@167: 			      E T1p, T1q, TW, TZ;
cannam@167: 			      T1p = TN + TO;
cannam@167: 			      T1q = TR - TQ;
cannam@167: 			      T1r = FNMS(KP984807753, T1q, KP173648177 * T1p);
cannam@167: 			      T1z = FMA(KP173648177, T1q, KP984807753 * T1p);
cannam@167: 			      TW = TU + TV;
cannam@167: 			      TZ = TX + TY;
cannam@167: 			      T10 = FNMS(KP984807753, TZ, KP173648177 * TW);
cannam@167: 			      T18 = FMA(KP984807753, TW, KP173648177 * TZ);
cannam@167: 			 }
cannam@167: 		    }
cannam@167: 	       }
cannam@167: 	       cr[0] = T5 + Tg;
cannam@167: 	       ci[0] = Tl + Tw;
cannam@167: 	       {
cannam@167: 		    E TA, TG, TE, TI;
cannam@167: 		    {
cannam@167: 			 E Ty, Tz, TC, TD;
cannam@167: 			 Ty = FNMS(KP500000000, Tg, T5);
cannam@167: 			 Tz = KP866025403 * (Tv - Tq);
cannam@167: 			 TA = Ty - Tz;
cannam@167: 			 TG = Ty + Tz;
cannam@167: 			 TC = FNMS(KP500000000, Tw, Tl);
cannam@167: 			 TD = KP866025403 * (Ta - Tf);
cannam@167: 			 TE = TC - TD;
cannam@167: 			 TI = TD + TC;
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E Tx, TB, TF, TH;
cannam@167: 			 Tx = W[10];
cannam@167: 			 TB = W[11];
cannam@167: 			 cr[WS(rs, 6)] = FNMS(TB, TE, Tx * TA);
cannam@167: 			 ci[WS(rs, 6)] = FMA(Tx, TE, TB * TA);
cannam@167: 			 TF = W[4];
cannam@167: 			 TH = W[5];
cannam@167: 			 cr[WS(rs, 3)] = FNMS(TH, TI, TF * TG);
cannam@167: 			 ci[WS(rs, 3)] = FMA(TF, TI, TH * TG);
cannam@167: 		    }
cannam@167: 	       }
cannam@167: 	       {
cannam@167: 		    E T1d, T1h, T12, T1c, T1a, T1g, T11, T19, TJ, T13;
cannam@167: 		    T1d = KP866025403 * (T18 - T17);
cannam@167: 		    T1h = KP866025403 * (TT - T10);
cannam@167: 		    T11 = TT + T10;
cannam@167: 		    T12 = TM + T11;
cannam@167: 		    T1c = FNMS(KP500000000, T11, TM);
cannam@167: 		    T19 = T17 + T18;
cannam@167: 		    T1a = T16 + T19;
cannam@167: 		    T1g = FNMS(KP500000000, T19, T16);
cannam@167: 		    TJ = W[0];
cannam@167: 		    T13 = W[1];
cannam@167: 		    cr[WS(rs, 1)] = FNMS(T13, T1a, TJ * T12);
cannam@167: 		    ci[WS(rs, 1)] = FMA(T13, T12, TJ * T1a);
cannam@167: 		    {
cannam@167: 			 E T1k, T1m, T1j, T1l;
cannam@167: 			 T1k = T1c + T1d;
cannam@167: 			 T1m = T1h + T1g;
cannam@167: 			 T1j = W[6];
cannam@167: 			 T1l = W[7];
cannam@167: 			 cr[WS(rs, 4)] = FNMS(T1l, T1m, T1j * T1k);
cannam@167: 			 ci[WS(rs, 4)] = FMA(T1j, T1m, T1l * T1k);
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E T1e, T1i, T1b, T1f;
cannam@167: 			 T1e = T1c - T1d;
cannam@167: 			 T1i = T1g - T1h;
cannam@167: 			 T1b = W[12];
cannam@167: 			 T1f = W[13];
cannam@167: 			 cr[WS(rs, 7)] = FNMS(T1f, T1i, T1b * T1e);
cannam@167: 			 ci[WS(rs, 7)] = FMA(T1b, T1i, T1f * T1e);
cannam@167: 		    }
cannam@167: 	       }
cannam@167: 	       {
cannam@167: 		    E T1F, T1J, T1w, T1E, T1C, T1I, T1v, T1B, T1n, T1x;
cannam@167: 		    T1F = KP866025403 * (T1A - T1z);
cannam@167: 		    T1J = KP866025403 * (T1r + T1u);
cannam@167: 		    T1v = T1r - T1u;
cannam@167: 		    T1w = T1o + T1v;
cannam@167: 		    T1E = FNMS(KP500000000, T1v, T1o);
cannam@167: 		    T1B = T1z + T1A;
cannam@167: 		    T1C = T1y + T1B;
cannam@167: 		    T1I = FNMS(KP500000000, T1B, T1y);
cannam@167: 		    T1n = W[2];
cannam@167: 		    T1x = W[3];
cannam@167: 		    cr[WS(rs, 2)] = FNMS(T1x, T1C, T1n * T1w);
cannam@167: 		    ci[WS(rs, 2)] = FMA(T1n, T1C, T1x * T1w);
cannam@167: 		    {
cannam@167: 			 E T1M, T1O, T1L, T1N;
cannam@167: 			 T1M = T1F + T1E;
cannam@167: 			 T1O = T1I + T1J;
cannam@167: 			 T1L = W[8];
cannam@167: 			 T1N = W[9];
cannam@167: 			 cr[WS(rs, 5)] = FNMS(T1N, T1O, T1L * T1M);
cannam@167: 			 ci[WS(rs, 5)] = FMA(T1N, T1M, T1L * T1O);
cannam@167: 		    }
cannam@167: 		    {
cannam@167: 			 E T1G, T1K, T1D, T1H;
cannam@167: 			 T1G = T1E - T1F;
cannam@167: 			 T1K = T1I - T1J;
cannam@167: 			 T1D = W[14];
cannam@167: 			 T1H = W[15];
cannam@167: 			 cr[WS(rs, 8)] = FNMS(T1H, T1K, T1D * T1G);
cannam@167: 			 ci[WS(rs, 8)] = FMA(T1H, T1G, T1D * T1K);
cannam@167: 		    }
cannam@167: 	       }
cannam@167: 	  }
cannam@167:      }
cannam@167: }
cannam@167: 
cannam@167: static const tw_instr twinstr[] = {
cannam@167:      {TW_FULL, 1, 9},
cannam@167:      {TW_NEXT, 1, 0}
cannam@167: };
cannam@167: 
cannam@167: static const hc2hc_desc desc = { 9, "hb_9", twinstr, &GENUS, {60, 36, 36, 0} };
cannam@167: 
cannam@167: void X(codelet_hb_9) (planner *p) {
cannam@167:      X(khc2hc_register) (p, hb_9, &desc);
cannam@167: }
cannam@167: #endif