cannam@95: /*
cannam@95:  * Copyright (c) 2003, 2007-11 Matteo Frigo
cannam@95:  * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
cannam@95:  *
cannam@95:  * This program is free software; you can redistribute it and/or modify
cannam@95:  * it under the terms of the GNU General Public License as published by
cannam@95:  * the Free Software Foundation; either version 2 of the License, or
cannam@95:  * (at your option) any later version.
cannam@95:  *
cannam@95:  * This program is distributed in the hope that it will be useful,
cannam@95:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@95:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
cannam@95:  * GNU General Public License for more details.
cannam@95:  *
cannam@95:  * You should have received a copy of the GNU General Public License
cannam@95:  * along with this program; if not, write to the Free Software
cannam@95:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
cannam@95:  *
cannam@95:  */
cannam@95: 
cannam@95: /* This file was automatically generated --- DO NOT EDIT */
cannam@95: /* Generated on Sun Nov 25 07:35:43 EST 2012 */
cannam@95: 
cannam@95: #include "codelet-dft.h"
cannam@95: 
cannam@95: #ifdef HAVE_FMA
cannam@95: 
cannam@95: /* Generated by: ../../../genfft/gen_notw.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 14 -name n1_14 -include n.h */
cannam@95: 
cannam@95: /*
cannam@95:  * This function contains 148 FP additions, 84 FP multiplications,
cannam@95:  * (or, 64 additions, 0 multiplications, 84 fused multiply/add),
cannam@95:  * 80 stack variables, 6 constants, and 56 memory accesses
cannam@95:  */
cannam@95: #include "n.h"
cannam@95: 
cannam@95: static void n1_14(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
cannam@95: {
cannam@95:      DK(KP974927912, +0.974927912181823607018131682993931217232785801);
cannam@95:      DK(KP801937735, +0.801937735804838252472204639014890102331838324);
cannam@95:      DK(KP900968867, +0.900968867902419126236102319507445051165919162);
cannam@95:      DK(KP554958132, +0.554958132087371191422194871006410481067288862);
cannam@95:      DK(KP692021471, +0.692021471630095869627814897002069140197260599);
cannam@95:      DK(KP356895867, +0.356895867892209443894399510021300583399127187);
cannam@95:      {
cannam@95: 	  INT i;
cannam@95: 	  for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(56, is), MAKE_VOLATILE_STRIDE(56, os)) {
cannam@95: 	       E Tp, T1L, T24, T1W, T1X, T28, T2a, T1Y, T29, T2b;
cannam@95: 	       {
cannam@95: 		    E T3, T1x, T1b, To, T1i, T1M, Ts, Ta, T1k, Tv, Th, T1j, T1K, Ty, TZ;
cannam@95: 		    E T14, Tz, T1Z, T27, T2c, T1d, TI, T23, T1G, T1D, TW, T1e, T22, T1A, TP;
cannam@95: 		    E T1c, T1n, T1s, T1f, T1P;
cannam@95: 		    {
cannam@95: 			 E T1, T2, T19, T1a;
cannam@95: 			 T1 = ri[0];
cannam@95: 			 T2 = ri[WS(is, 7)];
cannam@95: 			 T19 = ii[0];
cannam@95: 			 T1a = ii[WS(is, 7)];
cannam@95: 			 {
cannam@95: 			      E Tq, T6, Tr, T9, Te, Tx, Tn, Tw, Tk, Tf, Tb, Tc;
cannam@95: 			      {
cannam@95: 				   E Tl, Tm, Ti, Tj;
cannam@95: 				   {
cannam@95: 					E T4, T5, T7, T8;
cannam@95: 					T4 = ri[WS(is, 2)];
cannam@95: 					Tp = T1 + T2;
cannam@95: 					T3 = T1 - T2;
cannam@95: 					T1x = T19 + T1a;
cannam@95: 					T1b = T19 - T1a;
cannam@95: 					T5 = ri[WS(is, 9)];
cannam@95: 					T7 = ri[WS(is, 12)];
cannam@95: 					T8 = ri[WS(is, 5)];
cannam@95: 					Tl = ri[WS(is, 8)];
cannam@95: 					Tq = T4 + T5;
cannam@95: 					T6 = T4 - T5;
cannam@95: 					Tr = T7 + T8;
cannam@95: 					T9 = T7 - T8;
cannam@95: 					Tm = ri[WS(is, 1)];
cannam@95: 				   }
cannam@95: 				   Ti = ri[WS(is, 6)];
cannam@95: 				   Tj = ri[WS(is, 13)];
cannam@95: 				   Te = ri[WS(is, 10)];
cannam@95: 				   Tx = Tl + Tm;
cannam@95: 				   Tn = Tl - Tm;
cannam@95: 				   Tw = Ti + Tj;
cannam@95: 				   Tk = Ti - Tj;
cannam@95: 				   Tf = ri[WS(is, 3)];
cannam@95: 				   Tb = ri[WS(is, 4)];
cannam@95: 				   Tc = ri[WS(is, 11)];
cannam@95: 			      }
cannam@95: 			      {
cannam@95: 				   E Tu, Tg, Tt, Td;
cannam@95: 				   To = Tk + Tn;
cannam@95: 				   T1i = Tn - Tk;
cannam@95: 				   Tu = Te + Tf;
cannam@95: 				   Tg = Te - Tf;
cannam@95: 				   Tt = Tb + Tc;
cannam@95: 				   Td = Tb - Tc;
cannam@95: 				   T1M = Tr - Tq;
cannam@95: 				   Ts = Tq + Tr;
cannam@95: 				   Ta = T6 + T9;
cannam@95: 				   T1k = T9 - T6;
cannam@95: 				   T1L = Tt - Tu;
cannam@95: 				   Tv = Tt + Tu;
cannam@95: 				   Th = Td + Tg;
cannam@95: 				   T1j = Tg - Td;
cannam@95: 				   T1K = Tw - Tx;
cannam@95: 				   Ty = Tw + Tx;
cannam@95: 				   TZ = FNMS(KP356895867, Ta, To);
cannam@95: 				   T14 = FNMS(KP356895867, To, Th);
cannam@95: 				   Tz = FNMS(KP356895867, Th, Ta);
cannam@95: 				   T1Z = FNMS(KP356895867, Ts, Ty);
cannam@95: 			      }
cannam@95: 			 }
cannam@95: 			 {
cannam@95: 			      E T1B, TE, T1C, TH, T1F, TV, TJ, T1E, TS, T1z, TO, TK, T1y, TL;
cannam@95: 			      {
cannam@95: 				   E TF, TG, TT, TU, TC, TD;
cannam@95: 				   TC = ii[WS(is, 4)];
cannam@95: 				   TD = ii[WS(is, 11)];
cannam@95: 				   T27 = FNMS(KP356895867, Tv, Ts);
cannam@95: 				   T2c = FNMS(KP356895867, Ty, Tv);
cannam@95: 				   TF = ii[WS(is, 10)];
cannam@95: 				   T1B = TC + TD;
cannam@95: 				   TE = TC - TD;
cannam@95: 				   TG = ii[WS(is, 3)];
cannam@95: 				   TT = ii[WS(is, 8)];
cannam@95: 				   TU = ii[WS(is, 1)];
cannam@95: 				   {
cannam@95: 					E TQ, TR, TM, TN;
cannam@95: 					TQ = ii[WS(is, 6)];
cannam@95: 					T1C = TF + TG;
cannam@95: 					TH = TF - TG;
cannam@95: 					T1F = TT + TU;
cannam@95: 					TV = TT - TU;
cannam@95: 					TR = ii[WS(is, 13)];
cannam@95: 					TM = ii[WS(is, 12)];
cannam@95: 					TN = ii[WS(is, 5)];
cannam@95: 					TJ = ii[WS(is, 2)];
cannam@95: 					T1E = TQ + TR;
cannam@95: 					TS = TQ - TR;
cannam@95: 					T1z = TM + TN;
cannam@95: 					TO = TM - TN;
cannam@95: 					TK = ii[WS(is, 9)];
cannam@95: 				   }
cannam@95: 			      }
cannam@95: 			      T1d = TE + TH;
cannam@95: 			      TI = TE - TH;
cannam@95: 			      T23 = T1F - T1E;
cannam@95: 			      T1G = T1E + T1F;
cannam@95: 			      T1D = T1B + T1C;
cannam@95: 			      T24 = T1C - T1B;
cannam@95: 			      T1y = TJ + TK;
cannam@95: 			      TL = TJ - TK;
cannam@95: 			      TW = TS - TV;
cannam@95: 			      T1e = TS + TV;
cannam@95: 			      T22 = T1y - T1z;
cannam@95: 			      T1A = T1y + T1z;
cannam@95: 			      TP = TL - TO;
cannam@95: 			      T1c = TL + TO;
cannam@95: 			      T1n = FNMS(KP356895867, T1c, T1e);
cannam@95: 			      T1s = FNMS(KP356895867, T1d, T1c);
cannam@95: 			      T1f = FNMS(KP356895867, T1e, T1d);
cannam@95: 			      T1P = FNMS(KP356895867, T1A, T1G);
cannam@95: 			 }
cannam@95: 		    }
cannam@95: 		    {
cannam@95: 			 E T1U, T1H, T11, T12, T1o, T1q;
cannam@95: 			 ro[WS(os, 7)] = T3 + Ta + Th + To;
cannam@95: 			 io[WS(os, 7)] = T1b + T1c + T1d + T1e;
cannam@95: 			 T1U = FNMS(KP356895867, T1D, T1A);
cannam@95: 			 T1H = FNMS(KP356895867, T1G, T1D);
cannam@95: 			 ro[0] = Tp + Ts + Tv + Ty;
cannam@95: 			 io[0] = T1x + T1A + T1D + T1G;
cannam@95: 			 {
cannam@95: 			      E TB, TY, T1u, T1w, T10;
cannam@95: 			      {
cannam@95: 				   E TA, TX, T1t, T1v;
cannam@95: 				   TA = FNMS(KP692021471, Tz, To);
cannam@95: 				   TX = FMA(KP554958132, TW, TP);
cannam@95: 				   T1t = FNMS(KP692021471, T1s, T1e);
cannam@95: 				   T1v = FMA(KP554958132, T1i, T1k);
cannam@95: 				   TB = FNMS(KP900968867, TA, T3);
cannam@95: 				   TY = FMA(KP801937735, TX, TI);
cannam@95: 				   T1u = FNMS(KP900968867, T1t, T1b);
cannam@95: 				   T1w = FMA(KP801937735, T1v, T1j);
cannam@95: 			      }
cannam@95: 			      T10 = FNMS(KP692021471, TZ, Th);
cannam@95: 			      ro[WS(os, 1)] = FMA(KP974927912, TY, TB);
cannam@95: 			      ro[WS(os, 13)] = FNMS(KP974927912, TY, TB);
cannam@95: 			      io[WS(os, 13)] = FNMS(KP974927912, T1w, T1u);
cannam@95: 			      io[WS(os, 1)] = FMA(KP974927912, T1w, T1u);
cannam@95: 			      T11 = FNMS(KP900968867, T10, T3);
cannam@95: 			      T12 = FMA(KP554958132, TI, TW);
cannam@95: 			      T1o = FNMS(KP692021471, T1n, T1d);
cannam@95: 			      T1q = FMA(KP554958132, T1j, T1i);
cannam@95: 			 }
cannam@95: 			 {
cannam@95: 			      E T1J, T1N, T2d, T2f;
cannam@95: 			      {
cannam@95: 				   E T16, T17, T1g, T1l;
cannam@95: 				   {
cannam@95: 					E T13, T1p, T1r, T15;
cannam@95: 					T15 = FNMS(KP692021471, T14, Ta);
cannam@95: 					T13 = FNMS(KP801937735, T12, TP);
cannam@95: 					T1p = FNMS(KP900968867, T1o, T1b);
cannam@95: 					T1r = FNMS(KP801937735, T1q, T1k);
cannam@95: 					T16 = FNMS(KP900968867, T15, T3);
cannam@95: 					ro[WS(os, 9)] = FMA(KP974927912, T13, T11);
cannam@95: 					ro[WS(os, 5)] = FNMS(KP974927912, T13, T11);
cannam@95: 					io[WS(os, 9)] = FMA(KP974927912, T1r, T1p);
cannam@95: 					io[WS(os, 5)] = FNMS(KP974927912, T1r, T1p);
cannam@95: 					T17 = FNMS(KP554958132, TP, TI);
cannam@95: 				   }
cannam@95: 				   T1g = FNMS(KP692021471, T1f, T1c);
cannam@95: 				   T1l = FNMS(KP554958132, T1k, T1j);
cannam@95: 				   {
cannam@95: 					E T18, T1h, T1m, T1I;
cannam@95: 					T1I = FNMS(KP692021471, T1H, T1A);
cannam@95: 					T18 = FNMS(KP801937735, T17, TW);
cannam@95: 					T1h = FNMS(KP900968867, T1g, T1b);
cannam@95: 					T1m = FNMS(KP801937735, T1l, T1i);
cannam@95: 					T1J = FNMS(KP900968867, T1I, T1x);
cannam@95: 					ro[WS(os, 3)] = FMA(KP974927912, T18, T16);
cannam@95: 					ro[WS(os, 11)] = FNMS(KP974927912, T18, T16);
cannam@95: 					io[WS(os, 11)] = FNMS(KP974927912, T1m, T1h);
cannam@95: 					io[WS(os, 3)] = FMA(KP974927912, T1m, T1h);
cannam@95: 					T1N = FMA(KP554958132, T1M, T1L);
cannam@95: 				   }
cannam@95: 				   T2d = FNMS(KP692021471, T2c, Ts);
cannam@95: 				   T2f = FMA(KP554958132, T22, T24);
cannam@95: 			      }
cannam@95: 			      {
cannam@95: 				   E T1R, T1S, T20, T25;
cannam@95: 				   {
cannam@95: 					E T1O, T2e, T2g, T1Q;
cannam@95: 					T1Q = FNMS(KP692021471, T1P, T1D);
cannam@95: 					T1O = FNMS(KP801937735, T1N, T1K);
cannam@95: 					T2e = FNMS(KP900968867, T2d, Tp);
cannam@95: 					T2g = FNMS(KP801937735, T2f, T23);
cannam@95: 					T1R = FNMS(KP900968867, T1Q, T1x);
cannam@95: 					io[WS(os, 10)] = FNMS(KP974927912, T1O, T1J);
cannam@95: 					io[WS(os, 4)] = FMA(KP974927912, T1O, T1J);
cannam@95: 					ro[WS(os, 4)] = FMA(KP974927912, T2g, T2e);
cannam@95: 					ro[WS(os, 10)] = FNMS(KP974927912, T2g, T2e);
cannam@95: 					T1S = FMA(KP554958132, T1L, T1K);
cannam@95: 				   }
cannam@95: 				   T20 = FNMS(KP692021471, T1Z, Tv);
cannam@95: 				   T25 = FMA(KP554958132, T24, T23);
cannam@95: 				   {
cannam@95: 					E T1T, T21, T26, T1V;
cannam@95: 					T1V = FNMS(KP692021471, T1U, T1G);
cannam@95: 					T1T = FMA(KP801937735, T1S, T1M);
cannam@95: 					T21 = FNMS(KP900968867, T20, Tp);
cannam@95: 					T26 = FMA(KP801937735, T25, T22);
cannam@95: 					T1W = FNMS(KP900968867, T1V, T1x);
cannam@95: 					io[WS(os, 12)] = FNMS(KP974927912, T1T, T1R);
cannam@95: 					io[WS(os, 2)] = FMA(KP974927912, T1T, T1R);
cannam@95: 					ro[WS(os, 2)] = FMA(KP974927912, T26, T21);
cannam@95: 					ro[WS(os, 12)] = FNMS(KP974927912, T26, T21);
cannam@95: 					T1X = FNMS(KP554958132, T1K, T1M);
cannam@95: 				   }
cannam@95: 				   T28 = FNMS(KP692021471, T27, Ty);
cannam@95: 				   T2a = FNMS(KP554958132, T23, T22);
cannam@95: 			      }
cannam@95: 			 }
cannam@95: 		    }
cannam@95: 	       }
cannam@95: 	       T1Y = FNMS(KP801937735, T1X, T1L);
cannam@95: 	       T29 = FNMS(KP900968867, T28, Tp);
cannam@95: 	       T2b = FNMS(KP801937735, T2a, T24);
cannam@95: 	       io[WS(os, 8)] = FNMS(KP974927912, T1Y, T1W);
cannam@95: 	       io[WS(os, 6)] = FMA(KP974927912, T1Y, T1W);
cannam@95: 	       ro[WS(os, 6)] = FMA(KP974927912, T2b, T29);
cannam@95: 	       ro[WS(os, 8)] = FNMS(KP974927912, T2b, T29);
cannam@95: 	  }
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: static const kdft_desc desc = { 14, "n1_14", {64, 0, 84, 0}, &GENUS, 0, 0, 0, 0 };
cannam@95: 
cannam@95: void X(codelet_n1_14) (planner *p) {
cannam@95:      X(kdft_register) (p, n1_14, &desc);
cannam@95: }
cannam@95: 
cannam@95: #else				/* HAVE_FMA */
cannam@95: 
cannam@95: /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 14 -name n1_14 -include n.h */
cannam@95: 
cannam@95: /*
cannam@95:  * This function contains 148 FP additions, 72 FP multiplications,
cannam@95:  * (or, 100 additions, 24 multiplications, 48 fused multiply/add),
cannam@95:  * 43 stack variables, 6 constants, and 56 memory accesses
cannam@95:  */
cannam@95: #include "n.h"
cannam@95: 
cannam@95: static void n1_14(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
cannam@95: {
cannam@95:      DK(KP222520933, +0.222520933956314404288902564496794759466355569);
cannam@95:      DK(KP900968867, +0.900968867902419126236102319507445051165919162);
cannam@95:      DK(KP623489801, +0.623489801858733530525004884004239810632274731);
cannam@95:      DK(KP433883739, +0.433883739117558120475768332848358754609990728);
cannam@95:      DK(KP781831482, +0.781831482468029808708444526674057750232334519);
cannam@95:      DK(KP974927912, +0.974927912181823607018131682993931217232785801);
cannam@95:      {
cannam@95: 	  INT i;
cannam@95: 	  for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(56, is), MAKE_VOLATILE_STRIDE(56, os)) {
cannam@95: 	       E T3, Tp, T16, T1f, Ta, T1q, Ts, T10, TG, T1z, T19, T1i, Th, T1s, Tv;
cannam@95: 	       E T12, TU, T1B, T17, T1o, To, T1r, Ty, T11, TN, T1A, T18, T1l;
cannam@95: 	       {
cannam@95: 		    E T1, T2, T14, T15;
cannam@95: 		    T1 = ri[0];
cannam@95: 		    T2 = ri[WS(is, 7)];
cannam@95: 		    T3 = T1 - T2;
cannam@95: 		    Tp = T1 + T2;
cannam@95: 		    T14 = ii[0];
cannam@95: 		    T15 = ii[WS(is, 7)];
cannam@95: 		    T16 = T14 - T15;
cannam@95: 		    T1f = T14 + T15;
cannam@95: 	       }
cannam@95: 	       {
cannam@95: 		    E T6, Tq, T9, Tr;
cannam@95: 		    {
cannam@95: 			 E T4, T5, T7, T8;
cannam@95: 			 T4 = ri[WS(is, 2)];
cannam@95: 			 T5 = ri[WS(is, 9)];
cannam@95: 			 T6 = T4 - T5;
cannam@95: 			 Tq = T4 + T5;
cannam@95: 			 T7 = ri[WS(is, 12)];
cannam@95: 			 T8 = ri[WS(is, 5)];
cannam@95: 			 T9 = T7 - T8;
cannam@95: 			 Tr = T7 + T8;
cannam@95: 		    }
cannam@95: 		    Ta = T6 + T9;
cannam@95: 		    T1q = Tr - Tq;
cannam@95: 		    Ts = Tq + Tr;
cannam@95: 		    T10 = T9 - T6;
cannam@95: 	       }
cannam@95: 	       {
cannam@95: 		    E TC, T1g, TF, T1h;
cannam@95: 		    {
cannam@95: 			 E TA, TB, TD, TE;
cannam@95: 			 TA = ii[WS(is, 2)];
cannam@95: 			 TB = ii[WS(is, 9)];
cannam@95: 			 TC = TA - TB;
cannam@95: 			 T1g = TA + TB;
cannam@95: 			 TD = ii[WS(is, 12)];
cannam@95: 			 TE = ii[WS(is, 5)];
cannam@95: 			 TF = TD - TE;
cannam@95: 			 T1h = TD + TE;
cannam@95: 		    }
cannam@95: 		    TG = TC - TF;
cannam@95: 		    T1z = T1g - T1h;
cannam@95: 		    T19 = TC + TF;
cannam@95: 		    T1i = T1g + T1h;
cannam@95: 	       }
cannam@95: 	       {
cannam@95: 		    E Td, Tt, Tg, Tu;
cannam@95: 		    {
cannam@95: 			 E Tb, Tc, Te, Tf;
cannam@95: 			 Tb = ri[WS(is, 4)];
cannam@95: 			 Tc = ri[WS(is, 11)];
cannam@95: 			 Td = Tb - Tc;
cannam@95: 			 Tt = Tb + Tc;
cannam@95: 			 Te = ri[WS(is, 10)];
cannam@95: 			 Tf = ri[WS(is, 3)];
cannam@95: 			 Tg = Te - Tf;
cannam@95: 			 Tu = Te + Tf;
cannam@95: 		    }
cannam@95: 		    Th = Td + Tg;
cannam@95: 		    T1s = Tt - Tu;
cannam@95: 		    Tv = Tt + Tu;
cannam@95: 		    T12 = Tg - Td;
cannam@95: 	       }
cannam@95: 	       {
cannam@95: 		    E TQ, T1m, TT, T1n;
cannam@95: 		    {
cannam@95: 			 E TO, TP, TR, TS;
cannam@95: 			 TO = ii[WS(is, 4)];
cannam@95: 			 TP = ii[WS(is, 11)];
cannam@95: 			 TQ = TO - TP;
cannam@95: 			 T1m = TO + TP;
cannam@95: 			 TR = ii[WS(is, 10)];
cannam@95: 			 TS = ii[WS(is, 3)];
cannam@95: 			 TT = TR - TS;
cannam@95: 			 T1n = TR + TS;
cannam@95: 		    }
cannam@95: 		    TU = TQ - TT;
cannam@95: 		    T1B = T1n - T1m;
cannam@95: 		    T17 = TQ + TT;
cannam@95: 		    T1o = T1m + T1n;
cannam@95: 	       }
cannam@95: 	       {
cannam@95: 		    E Tk, Tw, Tn, Tx;
cannam@95: 		    {
cannam@95: 			 E Ti, Tj, Tl, Tm;
cannam@95: 			 Ti = ri[WS(is, 6)];
cannam@95: 			 Tj = ri[WS(is, 13)];
cannam@95: 			 Tk = Ti - Tj;
cannam@95: 			 Tw = Ti + Tj;
cannam@95: 			 Tl = ri[WS(is, 8)];
cannam@95: 			 Tm = ri[WS(is, 1)];
cannam@95: 			 Tn = Tl - Tm;
cannam@95: 			 Tx = Tl + Tm;
cannam@95: 		    }
cannam@95: 		    To = Tk + Tn;
cannam@95: 		    T1r = Tw - Tx;
cannam@95: 		    Ty = Tw + Tx;
cannam@95: 		    T11 = Tn - Tk;
cannam@95: 	       }
cannam@95: 	       {
cannam@95: 		    E TJ, T1j, TM, T1k;
cannam@95: 		    {
cannam@95: 			 E TH, TI, TK, TL;
cannam@95: 			 TH = ii[WS(is, 6)];
cannam@95: 			 TI = ii[WS(is, 13)];
cannam@95: 			 TJ = TH - TI;
cannam@95: 			 T1j = TH + TI;
cannam@95: 			 TK = ii[WS(is, 8)];
cannam@95: 			 TL = ii[WS(is, 1)];
cannam@95: 			 TM = TK - TL;
cannam@95: 			 T1k = TK + TL;
cannam@95: 		    }
cannam@95: 		    TN = TJ - TM;
cannam@95: 		    T1A = T1k - T1j;
cannam@95: 		    T18 = TJ + TM;
cannam@95: 		    T1l = T1j + T1k;
cannam@95: 	       }
cannam@95: 	       ro[WS(os, 7)] = T3 + Ta + Th + To;
cannam@95: 	       io[WS(os, 7)] = T16 + T19 + T17 + T18;
cannam@95: 	       ro[0] = Tp + Ts + Tv + Ty;
cannam@95: 	       io[0] = T1f + T1i + T1o + T1l;
cannam@95: 	       {
cannam@95: 		    E TV, Tz, T1e, T1d;
cannam@95: 		    TV = FNMS(KP781831482, TN, KP974927912 * TG) - (KP433883739 * TU);
cannam@95: 		    Tz = FMA(KP623489801, To, T3) + FNMA(KP900968867, Th, KP222520933 * Ta);
cannam@95: 		    ro[WS(os, 5)] = Tz - TV;
cannam@95: 		    ro[WS(os, 9)] = Tz + TV;
cannam@95: 		    T1e = FNMS(KP781831482, T11, KP974927912 * T10) - (KP433883739 * T12);
cannam@95: 		    T1d = FMA(KP623489801, T18, T16) + FNMA(KP900968867, T17, KP222520933 * T19);
cannam@95: 		    io[WS(os, 5)] = T1d - T1e;
cannam@95: 		    io[WS(os, 9)] = T1e + T1d;
cannam@95: 	       }
cannam@95: 	       {
cannam@95: 		    E TX, TW, T1b, T1c;
cannam@95: 		    TX = FMA(KP781831482, TG, KP974927912 * TU) + (KP433883739 * TN);
cannam@95: 		    TW = FMA(KP623489801, Ta, T3) + FNMA(KP900968867, To, KP222520933 * Th);
cannam@95: 		    ro[WS(os, 13)] = TW - TX;
cannam@95: 		    ro[WS(os, 1)] = TW + TX;
cannam@95: 		    T1b = FMA(KP781831482, T10, KP974927912 * T12) + (KP433883739 * T11);
cannam@95: 		    T1c = FMA(KP623489801, T19, T16) + FNMA(KP900968867, T18, KP222520933 * T17);
cannam@95: 		    io[WS(os, 1)] = T1b + T1c;
cannam@95: 		    io[WS(os, 13)] = T1c - T1b;
cannam@95: 	       }
cannam@95: 	       {
cannam@95: 		    E TZ, TY, T13, T1a;
cannam@95: 		    TZ = FMA(KP433883739, TG, KP974927912 * TN) - (KP781831482 * TU);
cannam@95: 		    TY = FMA(KP623489801, Th, T3) + FNMA(KP222520933, To, KP900968867 * Ta);
cannam@95: 		    ro[WS(os, 11)] = TY - TZ;
cannam@95: 		    ro[WS(os, 3)] = TY + TZ;
cannam@95: 		    T13 = FMA(KP433883739, T10, KP974927912 * T11) - (KP781831482 * T12);
cannam@95: 		    T1a = FMA(KP623489801, T17, T16) + FNMA(KP222520933, T18, KP900968867 * T19);
cannam@95: 		    io[WS(os, 3)] = T13 + T1a;
cannam@95: 		    io[WS(os, 11)] = T1a - T13;
cannam@95: 	       }
cannam@95: 	       {
cannam@95: 		    E T1t, T1p, T1C, T1y;
cannam@95: 		    T1t = FNMS(KP433883739, T1r, KP781831482 * T1q) - (KP974927912 * T1s);
cannam@95: 		    T1p = FMA(KP623489801, T1i, T1f) + FNMA(KP900968867, T1l, KP222520933 * T1o);
cannam@95: 		    io[WS(os, 6)] = T1p - T1t;
cannam@95: 		    io[WS(os, 8)] = T1t + T1p;
cannam@95: 		    T1C = FNMS(KP433883739, T1A, KP781831482 * T1z) - (KP974927912 * T1B);
cannam@95: 		    T1y = FMA(KP623489801, Ts, Tp) + FNMA(KP900968867, Ty, KP222520933 * Tv);
cannam@95: 		    ro[WS(os, 6)] = T1y - T1C;
cannam@95: 		    ro[WS(os, 8)] = T1y + T1C;
cannam@95: 	       }
cannam@95: 	       {
cannam@95: 		    E T1v, T1u, T1E, T1D;
cannam@95: 		    T1v = FMA(KP433883739, T1q, KP781831482 * T1s) - (KP974927912 * T1r);
cannam@95: 		    T1u = FMA(KP623489801, T1o, T1f) + FNMA(KP222520933, T1l, KP900968867 * T1i);
cannam@95: 		    io[WS(os, 4)] = T1u - T1v;
cannam@95: 		    io[WS(os, 10)] = T1v + T1u;
cannam@95: 		    T1E = FMA(KP433883739, T1z, KP781831482 * T1B) - (KP974927912 * T1A);
cannam@95: 		    T1D = FMA(KP623489801, Tv, Tp) + FNMA(KP222520933, Ty, KP900968867 * Ts);
cannam@95: 		    ro[WS(os, 4)] = T1D - T1E;
cannam@95: 		    ro[WS(os, 10)] = T1D + T1E;
cannam@95: 	       }
cannam@95: 	       {
cannam@95: 		    E T1w, T1x, T1G, T1F;
cannam@95: 		    T1w = FMA(KP974927912, T1q, KP433883739 * T1s) + (KP781831482 * T1r);
cannam@95: 		    T1x = FMA(KP623489801, T1l, T1f) + FNMA(KP900968867, T1o, KP222520933 * T1i);
cannam@95: 		    io[WS(os, 2)] = T1w + T1x;
cannam@95: 		    io[WS(os, 12)] = T1x - T1w;
cannam@95: 		    T1G = FMA(KP974927912, T1z, KP433883739 * T1B) + (KP781831482 * T1A);
cannam@95: 		    T1F = FMA(KP623489801, Ty, Tp) + FNMA(KP900968867, Tv, KP222520933 * Ts);
cannam@95: 		    ro[WS(os, 12)] = T1F - T1G;
cannam@95: 		    ro[WS(os, 2)] = T1F + T1G;
cannam@95: 	       }
cannam@95: 	  }
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: static const kdft_desc desc = { 14, "n1_14", {100, 24, 48, 0}, &GENUS, 0, 0, 0, 0 };
cannam@95: 
cannam@95: void X(codelet_n1_14) (planner *p) {
cannam@95:      X(kdft_register) (p, n1_14, &desc);
cannam@95: }
cannam@95: 
cannam@95: #endif				/* HAVE_FMA */