Chris@10: /*
Chris@10:  * Copyright (c) 2003, 2007-11 Matteo Frigo
Chris@10:  * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
Chris@10:  *
Chris@10:  * This program is free software; you can redistribute it and/or modify
Chris@10:  * it under the terms of the GNU General Public License as published by
Chris@10:  * the Free Software Foundation; either version 2 of the License, or
Chris@10:  * (at your option) any later version.
Chris@10:  *
Chris@10:  * This program is distributed in the hope that it will be useful,
Chris@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
Chris@10:  * GNU General Public License for more details.
Chris@10:  *
Chris@10:  * You should have received a copy of the GNU General Public License
Chris@10:  * along with this program; if not, write to the Free Software
Chris@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
Chris@10:  *
Chris@10:  */
Chris@10: 
Chris@10: /* This file was automatically generated --- DO NOT EDIT */
Chris@10: /* Generated on Sun Nov 25 07:39:52 EST 2012 */
Chris@10: 
Chris@10: #include "codelet-rdft.h"
Chris@10: 
Chris@10: #ifdef HAVE_FMA
Chris@10: 
Chris@10: /* Generated by: ../../../genfft/gen_hc2hc.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 15 -dit -name hf_15 -include hf.h */
Chris@10: 
Chris@10: /*
Chris@10:  * This function contains 184 FP additions, 140 FP multiplications,
Chris@10:  * (or, 72 additions, 28 multiplications, 112 fused multiply/add),
Chris@10:  * 97 stack variables, 6 constants, and 60 memory accesses
Chris@10:  */
Chris@10: #include "hf.h"
Chris@10: 
Chris@10: static void hf_15(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@10: {
Chris@10:      DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@10:      DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@10:      DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@10:      DK(KP618033988, +0.618033988749894848204586834365638117720309180);
Chris@10:      DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@10:      DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@10:      {
Chris@10: 	  INT m;
Chris@10: 	  for (m = mb, W = W + ((mb - 1) * 28); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) {
Chris@10: 	       E T3v, T3E, T3G, T3A, T3y, T3z, T3F, T3B;
Chris@10: 	       {
Chris@10: 		    E T1G, T3l, T3H, T3k, T1B, Tf, T37, T1y, T2Y, T2M, T2a, T2i, T39, Tz, T2U;
Chris@10: 		    E T2t, T1O, T2e, T3a, TT, T10, T2V, T2z, T1V, T2f, T2C, T12, T15, T14, T21;
Chris@10: 		    E T1c, T1Y, T13;
Chris@10: 		    {
Chris@10: 			 E T2I, T1k, T1m, T1p, T1o, T28, T1w, T25, T1n;
Chris@10: 			 {
Chris@10: 			      E T1, T3i, T9, Tc, Tb, T1D, T7, T1E, Ta, T1j, T1i, T1h;
Chris@10: 			      T1 = cr[0];
Chris@10: 			      T3i = ci[0];
Chris@10: 			      {
Chris@10: 				   E T3, T6, T2, T5, T1C, T4, T8;
Chris@10: 				   T3 = cr[WS(rs, 5)];
Chris@10: 				   T6 = ci[WS(rs, 5)];
Chris@10: 				   T2 = W[8];
Chris@10: 				   T5 = W[9];
Chris@10: 				   T9 = cr[WS(rs, 10)];
Chris@10: 				   Tc = ci[WS(rs, 10)];
Chris@10: 				   T1C = T2 * T6;
Chris@10: 				   T4 = T2 * T3;
Chris@10: 				   T8 = W[18];
Chris@10: 				   Tb = W[19];
Chris@10: 				   T1D = FNMS(T5, T3, T1C);
Chris@10: 				   T7 = FMA(T5, T6, T4);
Chris@10: 				   T1E = T8 * Tc;
Chris@10: 				   Ta = T8 * T9;
Chris@10: 			      }
Chris@10: 			      {
Chris@10: 				   E T1g, T1F, Td, T1f, T3j, Te, T2H;
Chris@10: 				   T1g = cr[WS(rs, 9)];
Chris@10: 				   T1j = ci[WS(rs, 9)];
Chris@10: 				   T1F = FNMS(Tb, T9, T1E);
Chris@10: 				   Td = FMA(Tb, Tc, Ta);
Chris@10: 				   T1f = W[16];
Chris@10: 				   T1i = W[17];
Chris@10: 				   T1G = T1D - T1F;
Chris@10: 				   T3j = T1D + T1F;
Chris@10: 				   T3l = Td - T7;
Chris@10: 				   Te = T7 + Td;
Chris@10: 				   T2H = T1f * T1j;
Chris@10: 				   T1h = T1f * T1g;
Chris@10: 				   T3H = T3j + T3i;
Chris@10: 				   T3k = FNMS(KP500000000, T3j, T3i);
Chris@10: 				   T1B = FNMS(KP500000000, Te, T1);
Chris@10: 				   Tf = T1 + Te;
Chris@10: 				   T2I = FNMS(T1i, T1g, T2H);
Chris@10: 			      }
Chris@10: 			      T1k = FMA(T1i, T1j, T1h);
Chris@10: 			      {
Chris@10: 				   E T1s, T1v, T1r, T1u, T27, T1t, T1l;
Chris@10: 				   T1s = cr[WS(rs, 4)];
Chris@10: 				   T1v = ci[WS(rs, 4)];
Chris@10: 				   T1r = W[6];
Chris@10: 				   T1u = W[7];
Chris@10: 				   T1m = cr[WS(rs, 14)];
Chris@10: 				   T1p = ci[WS(rs, 14)];
Chris@10: 				   T27 = T1r * T1v;
Chris@10: 				   T1t = T1r * T1s;
Chris@10: 				   T1l = W[26];
Chris@10: 				   T1o = W[27];
Chris@10: 				   T28 = FNMS(T1u, T1s, T27);
Chris@10: 				   T1w = FMA(T1u, T1v, T1t);
Chris@10: 				   T25 = T1l * T1p;
Chris@10: 				   T1n = T1l * T1m;
Chris@10: 			      }
Chris@10: 			 }
Chris@10: 			 {
Chris@10: 			      E Tl, T2p, Tn, Tq, Tp, T1M, Tx, T1J, To;
Chris@10: 			      {
Chris@10: 				   E Th, Tk, T26, T1q, Tg, Tj;
Chris@10: 				   Th = cr[WS(rs, 3)];
Chris@10: 				   Tk = ci[WS(rs, 3)];
Chris@10: 				   T26 = FNMS(T1o, T1m, T25);
Chris@10: 				   T1q = FMA(T1o, T1p, T1n);
Chris@10: 				   Tg = W[4];
Chris@10: 				   Tj = W[5];
Chris@10: 				   {
Chris@10: 					E T29, T2J, T1x, T2L;
Chris@10: 					T29 = T26 - T28;
Chris@10: 					T2J = T26 + T28;
Chris@10: 					T1x = T1q + T1w;
Chris@10: 					T2L = T1q - T1w;
Chris@10: 					{
Chris@10: 					     E T2o, Ti, T2K, T24;
Chris@10: 					     T2o = Tg * Tk;
Chris@10: 					     Ti = Tg * Th;
Chris@10: 					     T2K = FNMS(KP500000000, T2J, T2I);
Chris@10: 					     T37 = T2I + T2J;
Chris@10: 					     T24 = FNMS(KP500000000, T1x, T1k);
Chris@10: 					     T1y = T1k + T1x;
Chris@10: 					     Tl = FMA(Tj, Tk, Ti);
Chris@10: 					     T2Y = FMA(KP866025403, T2L, T2K);
Chris@10: 					     T2M = FNMS(KP866025403, T2L, T2K);
Chris@10: 					     T2a = FNMS(KP866025403, T29, T24);
Chris@10: 					     T2i = FMA(KP866025403, T29, T24);
Chris@10: 					     T2p = FNMS(Tj, Th, T2o);
Chris@10: 					}
Chris@10: 				   }
Chris@10: 			      }
Chris@10: 			      {
Chris@10: 				   E Tt, Tw, Ts, Tv, T1L, Tu, Tm;
Chris@10: 				   Tt = cr[WS(rs, 13)];
Chris@10: 				   Tw = ci[WS(rs, 13)];
Chris@10: 				   Ts = W[24];
Chris@10: 				   Tv = W[25];
Chris@10: 				   Tn = cr[WS(rs, 8)];
Chris@10: 				   Tq = ci[WS(rs, 8)];
Chris@10: 				   T1L = Ts * Tw;
Chris@10: 				   Tu = Ts * Tt;
Chris@10: 				   Tm = W[14];
Chris@10: 				   Tp = W[15];
Chris@10: 				   T1M = FNMS(Tv, Tt, T1L);
Chris@10: 				   Tx = FMA(Tv, Tw, Tu);
Chris@10: 				   T1J = Tm * Tq;
Chris@10: 				   To = Tm * Tn;
Chris@10: 			      }
Chris@10: 			      {
Chris@10: 				   E TF, T2v, TH, TK, TJ, T1T, TR, T1Q, TI;
Chris@10: 				   {
Chris@10: 					E TB, TE, T1K, Tr, TA, TD;
Chris@10: 					TB = cr[WS(rs, 12)];
Chris@10: 					TE = ci[WS(rs, 12)];
Chris@10: 					T1K = FNMS(Tp, Tn, T1J);
Chris@10: 					Tr = FMA(Tp, Tq, To);
Chris@10: 					TA = W[22];
Chris@10: 					TD = W[23];
Chris@10: 					{
Chris@10: 					     E T1N, T2q, Ty, T2s;
Chris@10: 					     T1N = T1K - T1M;
Chris@10: 					     T2q = T1K + T1M;
Chris@10: 					     Ty = Tr + Tx;
Chris@10: 					     T2s = Tr - Tx;
Chris@10: 					     {
Chris@10: 						  E T2u, TC, T2r, T1I;
Chris@10: 						  T2u = TA * TE;
Chris@10: 						  TC = TA * TB;
Chris@10: 						  T2r = FNMS(KP500000000, T2q, T2p);
Chris@10: 						  T39 = T2p + T2q;
Chris@10: 						  T1I = FNMS(KP500000000, Ty, Tl);
Chris@10: 						  Tz = Tl + Ty;
Chris@10: 						  TF = FMA(TD, TE, TC);
Chris@10: 						  T2U = FMA(KP866025403, T2s, T2r);
Chris@10: 						  T2t = FNMS(KP866025403, T2s, T2r);
Chris@10: 						  T1O = FNMS(KP866025403, T1N, T1I);
Chris@10: 						  T2e = FMA(KP866025403, T1N, T1I);
Chris@10: 						  T2v = FNMS(TD, TB, T2u);
Chris@10: 					     }
Chris@10: 					}
Chris@10: 				   }
Chris@10: 				   {
Chris@10: 					E TN, TQ, TM, TP, T1S, TO, TG;
Chris@10: 					TN = cr[WS(rs, 7)];
Chris@10: 					TQ = ci[WS(rs, 7)];
Chris@10: 					TM = W[12];
Chris@10: 					TP = W[13];
Chris@10: 					TH = cr[WS(rs, 2)];
Chris@10: 					TK = ci[WS(rs, 2)];
Chris@10: 					T1S = TM * TQ;
Chris@10: 					TO = TM * TN;
Chris@10: 					TG = W[2];
Chris@10: 					TJ = W[3];
Chris@10: 					T1T = FNMS(TP, TN, T1S);
Chris@10: 					TR = FMA(TP, TQ, TO);
Chris@10: 					T1Q = TG * TK;
Chris@10: 					TI = TG * TH;
Chris@10: 				   }
Chris@10: 				   {
Chris@10: 					E TW, TZ, T1R, TL, TV, TY;
Chris@10: 					TW = cr[WS(rs, 6)];
Chris@10: 					TZ = ci[WS(rs, 6)];
Chris@10: 					T1R = FNMS(TJ, TH, T1Q);
Chris@10: 					TL = FMA(TJ, TK, TI);
Chris@10: 					TV = W[10];
Chris@10: 					TY = W[11];
Chris@10: 					{
Chris@10: 					     E T1U, T2w, TS, T2y;
Chris@10: 					     T1U = T1R - T1T;
Chris@10: 					     T2w = T1R + T1T;
Chris@10: 					     TS = TL + TR;
Chris@10: 					     T2y = TL - TR;
Chris@10: 					     {
Chris@10: 						  E T2B, TX, T2x, T1P;
Chris@10: 						  T2B = TV * TZ;
Chris@10: 						  TX = TV * TW;
Chris@10: 						  T2x = FNMS(KP500000000, T2w, T2v);
Chris@10: 						  T3a = T2v + T2w;
Chris@10: 						  T1P = FNMS(KP500000000, TS, TF);
Chris@10: 						  TT = TF + TS;
Chris@10: 						  T10 = FMA(TY, TZ, TX);
Chris@10: 						  T2V = FMA(KP866025403, T2y, T2x);
Chris@10: 						  T2z = FNMS(KP866025403, T2y, T2x);
Chris@10: 						  T1V = FNMS(KP866025403, T1U, T1P);
Chris@10: 						  T2f = FMA(KP866025403, T1U, T1P);
Chris@10: 						  T2C = FNMS(TY, TW, T2B);
Chris@10: 					     }
Chris@10: 					}
Chris@10: 				   }
Chris@10: 				   {
Chris@10: 					E T18, T1b, T17, T1a, T20, T19, T11;
Chris@10: 					T18 = cr[WS(rs, 1)];
Chris@10: 					T1b = ci[WS(rs, 1)];
Chris@10: 					T17 = W[0];
Chris@10: 					T1a = W[1];
Chris@10: 					T12 = cr[WS(rs, 11)];
Chris@10: 					T15 = ci[WS(rs, 11)];
Chris@10: 					T20 = T17 * T1b;
Chris@10: 					T19 = T17 * T18;
Chris@10: 					T11 = W[20];
Chris@10: 					T14 = W[21];
Chris@10: 					T21 = FNMS(T1a, T18, T20);
Chris@10: 					T1c = FMA(T1a, T1b, T19);
Chris@10: 					T1Y = T11 * T15;
Chris@10: 					T13 = T11 * T12;
Chris@10: 				   }
Chris@10: 			      }
Chris@10: 			 }
Chris@10: 		    }
Chris@10: 		    {
Chris@10: 			 E T3I, T3O, T3w, T2d, T3J, T3P, T3x, T3C, T3D, T3f, T3g, T2Q, T2O, T3r, T3q;
Chris@10: 			 E T2k, T2m;
Chris@10: 			 {
Chris@10: 			      E T3b, T1Z, T16, TU;
Chris@10: 			      T3I = T39 + T3a;
Chris@10: 			      T3b = T39 - T3a;
Chris@10: 			      T1Z = FNMS(T14, T12, T1Y);
Chris@10: 			      T16 = FMA(T14, T15, T13);
Chris@10: 			      T3O = TT - Tz;
Chris@10: 			      TU = Tz + TT;
Chris@10: 			      {
Chris@10: 				   E T1H, T2G, T2h, T3e, T3c, T34, T1W, T32, T30, T33, T2b, T2S, T2R;
Chris@10: 				   {
Chris@10: 					E T2W, T22, T1d, T2F, T2E, T36, T2D;
Chris@10: 					T2W = T2U - T2V;
Chris@10: 					T3w = T2U + T2V;
Chris@10: 					T22 = T1Z - T21;
Chris@10: 					T2D = T1Z + T21;
Chris@10: 					T1d = T16 + T1c;
Chris@10: 					T2F = T16 - T1c;
Chris@10: 					T2E = FNMS(KP500000000, T2D, T2C);
Chris@10: 					T36 = T2C + T2D;
Chris@10: 					T2d = FMA(KP866025403, T1G, T1B);
Chris@10: 					T1H = FNMS(KP866025403, T1G, T1B);
Chris@10: 					{
Chris@10: 					     E T1e, T1X, T38, T2X;
Chris@10: 					     T1e = T10 + T1d;
Chris@10: 					     T1X = FNMS(KP500000000, T1d, T10);
Chris@10: 					     T38 = T36 - T37;
Chris@10: 					     T3J = T36 + T37;
Chris@10: 					     T2G = FNMS(KP866025403, T2F, T2E);
Chris@10: 					     T2X = FMA(KP866025403, T2F, T2E);
Chris@10: 					     {
Chris@10: 						  E T1z, T23, T2Z, T1A;
Chris@10: 						  T3P = T1y - T1e;
Chris@10: 						  T1z = T1e + T1y;
Chris@10: 						  T23 = FNMS(KP866025403, T22, T1X);
Chris@10: 						  T2h = FMA(KP866025403, T22, T1X);
Chris@10: 						  T3e = FMA(KP618033988, T38, T3b);
Chris@10: 						  T3c = FNMS(KP618033988, T3b, T38);
Chris@10: 						  T2Z = T2X - T2Y;
Chris@10: 						  T3x = T2X + T2Y;
Chris@10: 						  T1A = TU + T1z;
Chris@10: 						  T34 = TU - T1z;
Chris@10: 						  T3C = T1O - T1V;
Chris@10: 						  T1W = T1O + T1V;
Chris@10: 						  T32 = FNMS(KP618033988, T2W, T2Z);
Chris@10: 						  T30 = FMA(KP618033988, T2Z, T2W);
Chris@10: 						  cr[0] = Tf + T1A;
Chris@10: 						  T33 = FNMS(KP250000000, T1A, Tf);
Chris@10: 						  T2b = T23 + T2a;
Chris@10: 						  T3D = T23 - T2a;
Chris@10: 					     }
Chris@10: 					}
Chris@10: 				   }
Chris@10: 				   {
Chris@10: 					E T2A, T2N, T3d, T35, T2c;
Chris@10: 					T3f = T2t + T2z;
Chris@10: 					T2A = T2t - T2z;
Chris@10: 					T2N = T2G - T2M;
Chris@10: 					T3g = T2G + T2M;
Chris@10: 					T3d = FMA(KP559016994, T34, T33);
Chris@10: 					T35 = FNMS(KP559016994, T34, T33);
Chris@10: 					T2c = T1W + T2b;
Chris@10: 					T2S = T1W - T2b;
Chris@10: 					cr[WS(rs, 3)] = FMA(KP951056516, T3c, T35);
Chris@10: 					ci[WS(rs, 2)] = FNMS(KP951056516, T3c, T35);
Chris@10: 					cr[WS(rs, 6)] = FMA(KP951056516, T3e, T3d);
Chris@10: 					ci[WS(rs, 5)] = FNMS(KP951056516, T3e, T3d);
Chris@10: 					cr[WS(rs, 5)] = T1H + T2c;
Chris@10: 					T2R = FNMS(KP250000000, T2c, T1H);
Chris@10: 					T2Q = FNMS(KP618033988, T2A, T2N);
Chris@10: 					T2O = FMA(KP618033988, T2N, T2A);
Chris@10: 				   }
Chris@10: 				   {
Chris@10: 					E T2T, T31, T2g, T2j;
Chris@10: 					T2T = FMA(KP559016994, T2S, T2R);
Chris@10: 					T31 = FNMS(KP559016994, T2S, T2R);
Chris@10: 					T2g = T2e + T2f;
Chris@10: 					T3r = T2e - T2f;
Chris@10: 					T3q = T2h - T2i;
Chris@10: 					T2j = T2h + T2i;
Chris@10: 					ci[WS(rs, 3)] = FMA(KP951056516, T30, T2T);
Chris@10: 					ci[0] = FNMS(KP951056516, T30, T2T);
Chris@10: 					ci[WS(rs, 6)] = FMA(KP951056516, T32, T31);
Chris@10: 					cr[WS(rs, 2)] = FNMS(KP951056516, T32, T31);
Chris@10: 					T2k = T2g + T2j;
Chris@10: 					T2m = T2g - T2j;
Chris@10: 				   }
Chris@10: 			      }
Chris@10: 			 }
Chris@10: 			 {
Chris@10: 			      E T3m, T3s, T3u, T3o, T3h, T2l, T2n, T2P;
Chris@10: 			      ci[WS(rs, 4)] = T2d + T2k;
Chris@10: 			      T2l = FNMS(KP250000000, T2k, T2d);
Chris@10: 			      T3m = FMA(KP866025403, T3l, T3k);
Chris@10: 			      T3v = FNMS(KP866025403, T3l, T3k);
Chris@10: 			      T3s = FNMS(KP618033988, T3r, T3q);
Chris@10: 			      T3u = FMA(KP618033988, T3q, T3r);
Chris@10: 			      T2n = FMA(KP559016994, T2m, T2l);
Chris@10: 			      T2P = FNMS(KP559016994, T2m, T2l);
Chris@10: 			      ci[WS(rs, 1)] = FMA(KP951056516, T2Q, T2P);
Chris@10: 			      cr[WS(rs, 7)] = FNMS(KP951056516, T2Q, T2P);
Chris@10: 			      cr[WS(rs, 1)] = FMA(KP951056516, T2O, T2n);
Chris@10: 			      cr[WS(rs, 4)] = FNMS(KP951056516, T2O, T2n);
Chris@10: 			      T3o = T3f - T3g;
Chris@10: 			      T3h = T3f + T3g;
Chris@10: 			      {
Chris@10: 				   E T3S, T3Q, T3K, T3M, T3n, T3p, T3t, T3L, T3R, T3N;
Chris@10: 				   cr[WS(rs, 10)] = -(T3h + T3m);
Chris@10: 				   T3n = FNMS(KP250000000, T3h, T3m);
Chris@10: 				   T3S = FNMS(KP618033988, T3O, T3P);
Chris@10: 				   T3Q = FMA(KP618033988, T3P, T3O);
Chris@10: 				   T3p = FNMS(KP559016994, T3o, T3n);
Chris@10: 				   T3t = FMA(KP559016994, T3o, T3n);
Chris@10: 				   ci[WS(rs, 7)] = FMA(KP951056516, T3s, T3p);
Chris@10: 				   cr[WS(rs, 13)] = FMS(KP951056516, T3s, T3p);
Chris@10: 				   ci[WS(rs, 13)] = FNMS(KP951056516, T3u, T3t);
Chris@10: 				   ci[WS(rs, 10)] = FMA(KP951056516, T3u, T3t);
Chris@10: 				   T3K = T3I + T3J;
Chris@10: 				   T3M = T3I - T3J;
Chris@10: 				   ci[WS(rs, 14)] = T3K + T3H;
Chris@10: 				   T3L = FNMS(KP250000000, T3K, T3H);
Chris@10: 				   T3E = FMA(KP618033988, T3D, T3C);
Chris@10: 				   T3G = FNMS(KP618033988, T3C, T3D);
Chris@10: 				   T3R = FNMS(KP559016994, T3M, T3L);
Chris@10: 				   T3N = FMA(KP559016994, T3M, T3L);
Chris@10: 				   ci[WS(rs, 8)] = FMA(KP951056516, T3Q, T3N);
Chris@10: 				   cr[WS(rs, 9)] = FMS(KP951056516, T3Q, T3N);
Chris@10: 				   ci[WS(rs, 11)] = FMA(KP951056516, T3S, T3R);
Chris@10: 				   cr[WS(rs, 12)] = FMS(KP951056516, T3S, T3R);
Chris@10: 				   T3A = T3x - T3w;
Chris@10: 				   T3y = T3w + T3x;
Chris@10: 			      }
Chris@10: 			 }
Chris@10: 		    }
Chris@10: 	       }
Chris@10: 	       ci[WS(rs, 9)] = T3y + T3v;
Chris@10: 	       T3z = FNMS(KP250000000, T3y, T3v);
Chris@10: 	       T3F = FMA(KP559016994, T3A, T3z);
Chris@10: 	       T3B = FNMS(KP559016994, T3A, T3z);
Chris@10: 	       cr[WS(rs, 14)] = -(FMA(KP951056516, T3E, T3B));
Chris@10: 	       cr[WS(rs, 11)] = FMS(KP951056516, T3E, T3B);
Chris@10: 	       ci[WS(rs, 12)] = FMA(KP951056516, T3G, T3F);
Chris@10: 	       cr[WS(rs, 8)] = FMS(KP951056516, T3G, T3F);
Chris@10: 	  }
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static const tw_instr twinstr[] = {
Chris@10:      {TW_FULL, 1, 15},
Chris@10:      {TW_NEXT, 1, 0}
Chris@10: };
Chris@10: 
Chris@10: static const hc2hc_desc desc = { 15, "hf_15", twinstr, &GENUS, {72, 28, 112, 0} };
Chris@10: 
Chris@10: void X(codelet_hf_15) (planner *p) {
Chris@10:      X(khc2hc_register) (p, hf_15, &desc);
Chris@10: }
Chris@10: #else				/* HAVE_FMA */
Chris@10: 
Chris@10: /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -n 15 -dit -name hf_15 -include hf.h */
Chris@10: 
Chris@10: /*
Chris@10:  * This function contains 184 FP additions, 112 FP multiplications,
Chris@10:  * (or, 128 additions, 56 multiplications, 56 fused multiply/add),
Chris@10:  * 65 stack variables, 6 constants, and 60 memory accesses
Chris@10:  */
Chris@10: #include "hf.h"
Chris@10: 
Chris@10: static void hf_15(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@10: {
Chris@10:      DK(KP587785252, +0.587785252292473129168705954639072768597652438);
Chris@10:      DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@10:      DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@10:      DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@10:      DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@10:      DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@10:      {
Chris@10: 	  INT m;
Chris@10: 	  for (m = mb, W = W + ((mb - 1) * 28); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) {
Chris@10: 	       E T1q, T2Q, Td, T1n, T2T, T3l, T13, T1k, T1l, T2E, T2F, T3j, T1H, T1T, T2k;
Chris@10: 	       E T2w, T2f, T2v, T1M, T1U, Tu, TL, TM, T2H, T2I, T3i, T1w, T1Q, T29, T2t;
Chris@10: 	       E T24, T2s, T1B, T1R;
Chris@10: 	       {
Chris@10: 		    E T1, T2R, T6, T1o, Tb, T1p, Tc, T2S;
Chris@10: 		    T1 = cr[0];
Chris@10: 		    T2R = ci[0];
Chris@10: 		    {
Chris@10: 			 E T3, T5, T2, T4;
Chris@10: 			 T3 = cr[WS(rs, 5)];
Chris@10: 			 T5 = ci[WS(rs, 5)];
Chris@10: 			 T2 = W[8];
Chris@10: 			 T4 = W[9];
Chris@10: 			 T6 = FMA(T2, T3, T4 * T5);
Chris@10: 			 T1o = FNMS(T4, T3, T2 * T5);
Chris@10: 		    }
Chris@10: 		    {
Chris@10: 			 E T8, Ta, T7, T9;
Chris@10: 			 T8 = cr[WS(rs, 10)];
Chris@10: 			 Ta = ci[WS(rs, 10)];
Chris@10: 			 T7 = W[18];
Chris@10: 			 T9 = W[19];
Chris@10: 			 Tb = FMA(T7, T8, T9 * Ta);
Chris@10: 			 T1p = FNMS(T9, T8, T7 * Ta);
Chris@10: 		    }
Chris@10: 		    T1q = KP866025403 * (T1o - T1p);
Chris@10: 		    T2Q = KP866025403 * (Tb - T6);
Chris@10: 		    Tc = T6 + Tb;
Chris@10: 		    Td = T1 + Tc;
Chris@10: 		    T1n = FNMS(KP500000000, Tc, T1);
Chris@10: 		    T2S = T1o + T1p;
Chris@10: 		    T2T = FNMS(KP500000000, T2S, T2R);
Chris@10: 		    T3l = T2S + T2R;
Chris@10: 	       }
Chris@10: 	       {
Chris@10: 		    E TR, T2c, T18, T2h, TW, T1E, T11, T1F, T12, T2d, T1d, T1J, T1i, T1K, T1j;
Chris@10: 		    E T2i;
Chris@10: 		    {
Chris@10: 			 E TO, TQ, TN, TP;
Chris@10: 			 TO = cr[WS(rs, 6)];
Chris@10: 			 TQ = ci[WS(rs, 6)];
Chris@10: 			 TN = W[10];
Chris@10: 			 TP = W[11];
Chris@10: 			 TR = FMA(TN, TO, TP * TQ);
Chris@10: 			 T2c = FNMS(TP, TO, TN * TQ);
Chris@10: 		    }
Chris@10: 		    {
Chris@10: 			 E T15, T17, T14, T16;
Chris@10: 			 T15 = cr[WS(rs, 9)];
Chris@10: 			 T17 = ci[WS(rs, 9)];
Chris@10: 			 T14 = W[16];
Chris@10: 			 T16 = W[17];
Chris@10: 			 T18 = FMA(T14, T15, T16 * T17);
Chris@10: 			 T2h = FNMS(T16, T15, T14 * T17);
Chris@10: 		    }
Chris@10: 		    {
Chris@10: 			 E TT, TV, TS, TU;
Chris@10: 			 TT = cr[WS(rs, 11)];
Chris@10: 			 TV = ci[WS(rs, 11)];
Chris@10: 			 TS = W[20];
Chris@10: 			 TU = W[21];
Chris@10: 			 TW = FMA(TS, TT, TU * TV);
Chris@10: 			 T1E = FNMS(TU, TT, TS * TV);
Chris@10: 		    }
Chris@10: 		    {
Chris@10: 			 E TY, T10, TX, TZ;
Chris@10: 			 TY = cr[WS(rs, 1)];
Chris@10: 			 T10 = ci[WS(rs, 1)];
Chris@10: 			 TX = W[0];
Chris@10: 			 TZ = W[1];
Chris@10: 			 T11 = FMA(TX, TY, TZ * T10);
Chris@10: 			 T1F = FNMS(TZ, TY, TX * T10);
Chris@10: 		    }
Chris@10: 		    T12 = TW + T11;
Chris@10: 		    T2d = T1E + T1F;
Chris@10: 		    {
Chris@10: 			 E T1a, T1c, T19, T1b;
Chris@10: 			 T1a = cr[WS(rs, 14)];
Chris@10: 			 T1c = ci[WS(rs, 14)];
Chris@10: 			 T19 = W[26];
Chris@10: 			 T1b = W[27];
Chris@10: 			 T1d = FMA(T19, T1a, T1b * T1c);
Chris@10: 			 T1J = FNMS(T1b, T1a, T19 * T1c);
Chris@10: 		    }
Chris@10: 		    {
Chris@10: 			 E T1f, T1h, T1e, T1g;
Chris@10: 			 T1f = cr[WS(rs, 4)];
Chris@10: 			 T1h = ci[WS(rs, 4)];
Chris@10: 			 T1e = W[6];
Chris@10: 			 T1g = W[7];
Chris@10: 			 T1i = FMA(T1e, T1f, T1g * T1h);
Chris@10: 			 T1K = FNMS(T1g, T1f, T1e * T1h);
Chris@10: 		    }
Chris@10: 		    T1j = T1d + T1i;
Chris@10: 		    T2i = T1J + T1K;
Chris@10: 		    {
Chris@10: 			 E T1D, T1G, T2g, T2j;
Chris@10: 			 T13 = TR + T12;
Chris@10: 			 T1k = T18 + T1j;
Chris@10: 			 T1l = T13 + T1k;
Chris@10: 			 T2E = T2c + T2d;
Chris@10: 			 T2F = T2h + T2i;
Chris@10: 			 T3j = T2E + T2F;
Chris@10: 			 T1D = FNMS(KP500000000, T12, TR);
Chris@10: 			 T1G = KP866025403 * (T1E - T1F);
Chris@10: 			 T1H = T1D - T1G;
Chris@10: 			 T1T = T1D + T1G;
Chris@10: 			 T2g = KP866025403 * (T1d - T1i);
Chris@10: 			 T2j = FNMS(KP500000000, T2i, T2h);
Chris@10: 			 T2k = T2g - T2j;
Chris@10: 			 T2w = T2g + T2j;
Chris@10: 			 {
Chris@10: 			      E T2b, T2e, T1I, T1L;
Chris@10: 			      T2b = KP866025403 * (T11 - TW);
Chris@10: 			      T2e = FNMS(KP500000000, T2d, T2c);
Chris@10: 			      T2f = T2b + T2e;
Chris@10: 			      T2v = T2e - T2b;
Chris@10: 			      T1I = FNMS(KP500000000, T1j, T18);
Chris@10: 			      T1L = KP866025403 * (T1J - T1K);
Chris@10: 			      T1M = T1I - T1L;
Chris@10: 			      T1U = T1I + T1L;
Chris@10: 			 }
Chris@10: 		    }
Chris@10: 	       }
Chris@10: 	       {
Chris@10: 		    E Ti, T21, Tz, T26, Tn, T1t, Ts, T1u, Tt, T22, TE, T1y, TJ, T1z, TK;
Chris@10: 		    E T27;
Chris@10: 		    {
Chris@10: 			 E Tf, Th, Te, Tg;
Chris@10: 			 Tf = cr[WS(rs, 3)];
Chris@10: 			 Th = ci[WS(rs, 3)];
Chris@10: 			 Te = W[4];
Chris@10: 			 Tg = W[5];
Chris@10: 			 Ti = FMA(Te, Tf, Tg * Th);
Chris@10: 			 T21 = FNMS(Tg, Tf, Te * Th);
Chris@10: 		    }
Chris@10: 		    {
Chris@10: 			 E Tw, Ty, Tv, Tx;
Chris@10: 			 Tw = cr[WS(rs, 12)];
Chris@10: 			 Ty = ci[WS(rs, 12)];
Chris@10: 			 Tv = W[22];
Chris@10: 			 Tx = W[23];
Chris@10: 			 Tz = FMA(Tv, Tw, Tx * Ty);
Chris@10: 			 T26 = FNMS(Tx, Tw, Tv * Ty);
Chris@10: 		    }
Chris@10: 		    {
Chris@10: 			 E Tk, Tm, Tj, Tl;
Chris@10: 			 Tk = cr[WS(rs, 8)];
Chris@10: 			 Tm = ci[WS(rs, 8)];
Chris@10: 			 Tj = W[14];
Chris@10: 			 Tl = W[15];
Chris@10: 			 Tn = FMA(Tj, Tk, Tl * Tm);
Chris@10: 			 T1t = FNMS(Tl, Tk, Tj * Tm);
Chris@10: 		    }
Chris@10: 		    {
Chris@10: 			 E Tp, Tr, To, Tq;
Chris@10: 			 Tp = cr[WS(rs, 13)];
Chris@10: 			 Tr = ci[WS(rs, 13)];
Chris@10: 			 To = W[24];
Chris@10: 			 Tq = W[25];
Chris@10: 			 Ts = FMA(To, Tp, Tq * Tr);
Chris@10: 			 T1u = FNMS(Tq, Tp, To * Tr);
Chris@10: 		    }
Chris@10: 		    Tt = Tn + Ts;
Chris@10: 		    T22 = T1t + T1u;
Chris@10: 		    {
Chris@10: 			 E TB, TD, TA, TC;
Chris@10: 			 TB = cr[WS(rs, 2)];
Chris@10: 			 TD = ci[WS(rs, 2)];
Chris@10: 			 TA = W[2];
Chris@10: 			 TC = W[3];
Chris@10: 			 TE = FMA(TA, TB, TC * TD);
Chris@10: 			 T1y = FNMS(TC, TB, TA * TD);
Chris@10: 		    }
Chris@10: 		    {
Chris@10: 			 E TG, TI, TF, TH;
Chris@10: 			 TG = cr[WS(rs, 7)];
Chris@10: 			 TI = ci[WS(rs, 7)];
Chris@10: 			 TF = W[12];
Chris@10: 			 TH = W[13];
Chris@10: 			 TJ = FMA(TF, TG, TH * TI);
Chris@10: 			 T1z = FNMS(TH, TG, TF * TI);
Chris@10: 		    }
Chris@10: 		    TK = TE + TJ;
Chris@10: 		    T27 = T1y + T1z;
Chris@10: 		    {
Chris@10: 			 E T1s, T1v, T25, T28;
Chris@10: 			 Tu = Ti + Tt;
Chris@10: 			 TL = Tz + TK;
Chris@10: 			 TM = Tu + TL;
Chris@10: 			 T2H = T21 + T22;
Chris@10: 			 T2I = T26 + T27;
Chris@10: 			 T3i = T2H + T2I;
Chris@10: 			 T1s = FNMS(KP500000000, Tt, Ti);
Chris@10: 			 T1v = KP866025403 * (T1t - T1u);
Chris@10: 			 T1w = T1s - T1v;
Chris@10: 			 T1Q = T1s + T1v;
Chris@10: 			 T25 = KP866025403 * (TJ - TE);
Chris@10: 			 T28 = FNMS(KP500000000, T27, T26);
Chris@10: 			 T29 = T25 + T28;
Chris@10: 			 T2t = T28 - T25;
Chris@10: 			 {
Chris@10: 			      E T20, T23, T1x, T1A;
Chris@10: 			      T20 = KP866025403 * (Ts - Tn);
Chris@10: 			      T23 = FNMS(KP500000000, T22, T21);
Chris@10: 			      T24 = T20 + T23;
Chris@10: 			      T2s = T23 - T20;
Chris@10: 			      T1x = FNMS(KP500000000, TK, Tz);
Chris@10: 			      T1A = KP866025403 * (T1y - T1z);
Chris@10: 			      T1B = T1x - T1A;
Chris@10: 			      T1R = T1x + T1A;
Chris@10: 			 }
Chris@10: 		    }
Chris@10: 	       }
Chris@10: 	       {
Chris@10: 		    E T2C, T1m, T2B, T2K, T2M, T2G, T2J, T2L, T2D;
Chris@10: 		    T2C = KP559016994 * (TM - T1l);
Chris@10: 		    T1m = TM + T1l;
Chris@10: 		    T2B = FNMS(KP250000000, T1m, Td);
Chris@10: 		    T2G = T2E - T2F;
Chris@10: 		    T2J = T2H - T2I;
Chris@10: 		    T2K = FNMS(KP587785252, T2J, KP951056516 * T2G);
Chris@10: 		    T2M = FMA(KP951056516, T2J, KP587785252 * T2G);
Chris@10: 		    cr[0] = Td + T1m;
Chris@10: 		    T2L = T2C + T2B;
Chris@10: 		    ci[WS(rs, 5)] = T2L - T2M;
Chris@10: 		    cr[WS(rs, 6)] = T2L + T2M;
Chris@10: 		    T2D = T2B - T2C;
Chris@10: 		    ci[WS(rs, 2)] = T2D - T2K;
Chris@10: 		    cr[WS(rs, 3)] = T2D + T2K;
Chris@10: 	       }
Chris@10: 	       {
Chris@10: 		    E T3k, T3m, T3n, T3h, T3p, T3f, T3g, T3q, T3o;
Chris@10: 		    T3k = KP559016994 * (T3i - T3j);
Chris@10: 		    T3m = T3i + T3j;
Chris@10: 		    T3n = FNMS(KP250000000, T3m, T3l);
Chris@10: 		    T3f = T1k - T13;
Chris@10: 		    T3g = Tu - TL;
Chris@10: 		    T3h = FNMS(KP951056516, T3g, KP587785252 * T3f);
Chris@10: 		    T3p = FMA(KP587785252, T3g, KP951056516 * T3f);
Chris@10: 		    ci[WS(rs, 14)] = T3m + T3l;
Chris@10: 		    T3q = T3n - T3k;
Chris@10: 		    cr[WS(rs, 12)] = T3p - T3q;
Chris@10: 		    ci[WS(rs, 11)] = T3p + T3q;
Chris@10: 		    T3o = T3k + T3n;
Chris@10: 		    cr[WS(rs, 9)] = T3h - T3o;
Chris@10: 		    ci[WS(rs, 8)] = T3h + T3o;
Chris@10: 	       }
Chris@10: 	       {
Chris@10: 		    E T2y, T2A, T1r, T1O, T2p, T2q, T2z, T2r;
Chris@10: 		    {
Chris@10: 			 E T2u, T2x, T1C, T1N;
Chris@10: 			 T2u = T2s - T2t;
Chris@10: 			 T2x = T2v - T2w;
Chris@10: 			 T2y = FMA(KP951056516, T2u, KP587785252 * T2x);
Chris@10: 			 T2A = FNMS(KP587785252, T2u, KP951056516 * T2x);
Chris@10: 			 T1r = T1n - T1q;
Chris@10: 			 T1C = T1w + T1B;
Chris@10: 			 T1N = T1H + T1M;
Chris@10: 			 T1O = T1C + T1N;
Chris@10: 			 T2p = KP559016994 * (T1C - T1N);
Chris@10: 			 T2q = FNMS(KP250000000, T1O, T1r);
Chris@10: 		    }
Chris@10: 		    cr[WS(rs, 5)] = T1r + T1O;
Chris@10: 		    T2z = T2q - T2p;
Chris@10: 		    cr[WS(rs, 2)] = T2z - T2A;
Chris@10: 		    ci[WS(rs, 6)] = T2z + T2A;
Chris@10: 		    T2r = T2p + T2q;
Chris@10: 		    ci[0] = T2r - T2y;
Chris@10: 		    ci[WS(rs, 3)] = T2r + T2y;
Chris@10: 	       }
Chris@10: 	       {
Chris@10: 		    E T35, T3d, T39, T3a, T38, T3b, T3e, T3c;
Chris@10: 		    {
Chris@10: 			 E T33, T34, T36, T37;
Chris@10: 			 T33 = T1w - T1B;
Chris@10: 			 T34 = T1H - T1M;
Chris@10: 			 T35 = FMA(KP951056516, T33, KP587785252 * T34);
Chris@10: 			 T3d = FNMS(KP587785252, T33, KP951056516 * T34);
Chris@10: 			 T39 = T2T - T2Q;
Chris@10: 			 T36 = T2v + T2w;
Chris@10: 			 T37 = T2s + T2t;
Chris@10: 			 T3a = T37 + T36;
Chris@10: 			 T38 = KP559016994 * (T36 - T37);
Chris@10: 			 T3b = FNMS(KP250000000, T3a, T39);
Chris@10: 		    }
Chris@10: 		    ci[WS(rs, 9)] = T3a + T39;
Chris@10: 		    T3e = T38 + T3b;
Chris@10: 		    cr[WS(rs, 8)] = T3d - T3e;
Chris@10: 		    ci[WS(rs, 12)] = T3d + T3e;
Chris@10: 		    T3c = T38 - T3b;
Chris@10: 		    cr[WS(rs, 11)] = T35 + T3c;
Chris@10: 		    cr[WS(rs, 14)] = T3c - T35;
Chris@10: 	       }
Chris@10: 	       {
Chris@10: 		    E T2X, T31, T2U, T2P, T2Y, T2Z, T32, T30;
Chris@10: 		    {
Chris@10: 			 E T2V, T2W, T2N, T2O;
Chris@10: 			 T2V = T1T - T1U;
Chris@10: 			 T2W = T1Q - T1R;
Chris@10: 			 T2X = FNMS(KP587785252, T2W, KP951056516 * T2V);
Chris@10: 			 T31 = FMA(KP951056516, T2W, KP587785252 * T2V);
Chris@10: 			 T2U = T2Q + T2T;
Chris@10: 			 T2N = T2k - T2f;
Chris@10: 			 T2O = T24 + T29;
Chris@10: 			 T2P = T2N - T2O;
Chris@10: 			 T2Y = FMA(KP250000000, T2P, T2U);
Chris@10: 			 T2Z = KP559016994 * (T2O + T2N);
Chris@10: 		    }
Chris@10: 		    cr[WS(rs, 10)] = T2P - T2U;
Chris@10: 		    T32 = T2Z + T2Y;
Chris@10: 		    ci[WS(rs, 10)] = T31 + T32;
Chris@10: 		    ci[WS(rs, 13)] = T32 - T31;
Chris@10: 		    T30 = T2Y - T2Z;
Chris@10: 		    cr[WS(rs, 13)] = T2X - T30;
Chris@10: 		    ci[WS(rs, 7)] = T2X + T30;
Chris@10: 	       }
Chris@10: 	       {
Chris@10: 		    E T2m, T2o, T1P, T1W, T1X, T1Y, T1Z, T2n;
Chris@10: 		    {
Chris@10: 			 E T2a, T2l, T1S, T1V;
Chris@10: 			 T2a = T24 - T29;
Chris@10: 			 T2l = T2f + T2k;
Chris@10: 			 T2m = FMA(KP951056516, T2a, KP587785252 * T2l);
Chris@10: 			 T2o = FNMS(KP587785252, T2a, KP951056516 * T2l);
Chris@10: 			 T1P = T1n + T1q;
Chris@10: 			 T1S = T1Q + T1R;
Chris@10: 			 T1V = T1T + T1U;
Chris@10: 			 T1W = T1S + T1V;
Chris@10: 			 T1X = KP559016994 * (T1S - T1V);
Chris@10: 			 T1Y = FNMS(KP250000000, T1W, T1P);
Chris@10: 		    }
Chris@10: 		    ci[WS(rs, 4)] = T1P + T1W;
Chris@10: 		    T1Z = T1X + T1Y;
Chris@10: 		    cr[WS(rs, 4)] = T1Z - T2m;
Chris@10: 		    cr[WS(rs, 1)] = T1Z + T2m;
Chris@10: 		    T2n = T1Y - T1X;
Chris@10: 		    cr[WS(rs, 7)] = T2n - T2o;
Chris@10: 		    ci[WS(rs, 1)] = T2n + T2o;
Chris@10: 	       }
Chris@10: 	  }
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static const tw_instr twinstr[] = {
Chris@10:      {TW_FULL, 1, 15},
Chris@10:      {TW_NEXT, 1, 0}
Chris@10: };
Chris@10: 
Chris@10: static const hc2hc_desc desc = { 15, "hf_15", twinstr, &GENUS, {128, 56, 56, 0} };
Chris@10: 
Chris@10: void X(codelet_hf_15) (planner *p) {
Chris@10:      X(khc2hc_register) (p, hf_15, &desc);
Chris@10: }
Chris@10: #endif				/* HAVE_FMA */