Chris@82: /*
Chris@82:  * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@82:  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@82:  *
Chris@82:  * This program is free software; you can redistribute it and/or modify
Chris@82:  * it under the terms of the GNU General Public License as published by
Chris@82:  * the Free Software Foundation; either version 2 of the License, or
Chris@82:  * (at your option) any later version.
Chris@82:  *
Chris@82:  * This program is distributed in the hope that it will be useful,
Chris@82:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@82:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
Chris@82:  * GNU General Public License for more details.
Chris@82:  *
Chris@82:  * You should have received a copy of the GNU General Public License
Chris@82:  * along with this program; if not, write to the Free Software
Chris@82:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
Chris@82:  *
Chris@82:  */
Chris@82: 
Chris@82: /* This file was automatically generated --- DO NOT EDIT */
Chris@82: /* Generated on Thu May 24 08:07:45 EDT 2018 */
Chris@82: 
Chris@82: #include "rdft/codelet-rdft.h"
Chris@82: 
Chris@82: #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
Chris@82: 
Chris@82: /* Generated by: ../../../genfft/gen_r2cb.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -name r2cbIII_32 -dft-III -include rdft/scalar/r2cbIII.h */
Chris@82: 
Chris@82: /*
Chris@82:  * This function contains 174 FP additions, 100 FP multiplications,
Chris@82:  * (or, 106 additions, 32 multiplications, 68 fused multiply/add),
Chris@82:  * 65 stack variables, 18 constants, and 64 memory accesses
Chris@82:  */
Chris@82: #include "rdft/scalar/r2cbIII.h"
Chris@82: 
Chris@82: static void r2cbIII_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
Chris@82: {
Chris@82:      DK(KP303346683, +0.303346683607342391675883946941299872384187453);
Chris@82:      DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
Chris@82:      DK(KP534511135, +0.534511135950791641089685961295362908582039528);
Chris@82:      DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
Chris@82:      DK(KP820678790, +0.820678790828660330972281985331011598767386482);
Chris@82:      DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
Chris@82:      DK(KP098491403, +0.098491403357164253077197521291327432293052451);
Chris@82:      DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
Chris@82:      DK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@82:      DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
Chris@82:      DK(KP198912367, +0.198912367379658006911597622644676228597850501);
Chris@82:      DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
Chris@82:      DK(KP668178637, +0.668178637919298919997757686523080761552472251);
Chris@82:      DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
Chris@82:      DK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@82:      DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
Chris@82:      DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
Chris@82:      DK(KP414213562, +0.414213562373095048801688724209698078569671875);
Chris@82:      {
Chris@82: 	  INT i;
Chris@82: 	  for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
Chris@82: 	       E T7, T2i, T2E, Tz, T1e, T1I, T1Z, T1x, Te, T22, T2F, T2j, T1h, T1y, TK;
Chris@82: 	       E T1J, Tm, T2B, TW, T1k, T1C, T1M, T28, T2m, Tt, T2A, T17, T1j, T1F, T1L;
Chris@82: 	       E T2d, T2l;
Chris@82: 	       {
Chris@82: 		    E T3, Tv, T1d, T2g, T6, T1a, Ty, T2h;
Chris@82: 		    {
Chris@82: 			 E T1, T2, T1b, T1c;
Chris@82: 			 T1 = Cr[0];
Chris@82: 			 T2 = Cr[WS(csr, 15)];
Chris@82: 			 T3 = T1 + T2;
Chris@82: 			 Tv = T1 - T2;
Chris@82: 			 T1b = Ci[0];
Chris@82: 			 T1c = Ci[WS(csi, 15)];
Chris@82: 			 T1d = T1b + T1c;
Chris@82: 			 T2g = T1c - T1b;
Chris@82: 		    }
Chris@82: 		    {
Chris@82: 			 E T4, T5, Tw, Tx;
Chris@82: 			 T4 = Cr[WS(csr, 8)];
Chris@82: 			 T5 = Cr[WS(csr, 7)];
Chris@82: 			 T6 = T4 + T5;
Chris@82: 			 T1a = T4 - T5;
Chris@82: 			 Tw = Ci[WS(csi, 8)];
Chris@82: 			 Tx = Ci[WS(csi, 7)];
Chris@82: 			 Ty = Tw + Tx;
Chris@82: 			 T2h = Tx - Tw;
Chris@82: 		    }
Chris@82: 		    T7 = T3 + T6;
Chris@82: 		    T2i = T2g - T2h;
Chris@82: 		    T2E = T2h + T2g;
Chris@82: 		    Tz = Tv - Ty;
Chris@82: 		    T1e = T1a + T1d;
Chris@82: 		    T1I = T1a - T1d;
Chris@82: 		    T1Z = T3 - T6;
Chris@82: 		    T1x = Tv + Ty;
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E Ta, TA, TD, T20, Td, TF, TI, T21;
Chris@82: 		    {
Chris@82: 			 E T8, T9, TB, TC;
Chris@82: 			 T8 = Cr[WS(csr, 4)];
Chris@82: 			 T9 = Cr[WS(csr, 11)];
Chris@82: 			 Ta = T8 + T9;
Chris@82: 			 TA = T8 - T9;
Chris@82: 			 TB = Ci[WS(csi, 4)];
Chris@82: 			 TC = Ci[WS(csi, 11)];
Chris@82: 			 TD = TB + TC;
Chris@82: 			 T20 = TC - TB;
Chris@82: 		    }
Chris@82: 		    {
Chris@82: 			 E Tb, Tc, TG, TH;
Chris@82: 			 Tb = Cr[WS(csr, 3)];
Chris@82: 			 Tc = Cr[WS(csr, 12)];
Chris@82: 			 Td = Tb + Tc;
Chris@82: 			 TF = Tb - Tc;
Chris@82: 			 TG = Ci[WS(csi, 3)];
Chris@82: 			 TH = Ci[WS(csi, 12)];
Chris@82: 			 TI = TG + TH;
Chris@82: 			 T21 = TG - TH;
Chris@82: 		    }
Chris@82: 		    Te = Ta + Td;
Chris@82: 		    T22 = T20 - T21;
Chris@82: 		    T2F = T20 + T21;
Chris@82: 		    T2j = Ta - Td;
Chris@82: 		    {
Chris@82: 			 E T1f, T1g, TE, TJ;
Chris@82: 			 T1f = TF + TI;
Chris@82: 			 T1g = TA + TD;
Chris@82: 			 T1h = T1f - T1g;
Chris@82: 			 T1y = T1g + T1f;
Chris@82: 			 TE = TA - TD;
Chris@82: 			 TJ = TF - TI;
Chris@82: 			 TK = TE + TJ;
Chris@82: 			 T1J = TE - TJ;
Chris@82: 		    }
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E Ti, TM, TU, T25, Tl, TR, TP, T26, TQ, TV;
Chris@82: 		    {
Chris@82: 			 E Tg, Th, TS, TT;
Chris@82: 			 Tg = Cr[WS(csr, 2)];
Chris@82: 			 Th = Cr[WS(csr, 13)];
Chris@82: 			 Ti = Tg + Th;
Chris@82: 			 TM = Tg - Th;
Chris@82: 			 TS = Ci[WS(csi, 2)];
Chris@82: 			 TT = Ci[WS(csi, 13)];
Chris@82: 			 TU = TS + TT;
Chris@82: 			 T25 = TS - TT;
Chris@82: 		    }
Chris@82: 		    {
Chris@82: 			 E Tj, Tk, TN, TO;
Chris@82: 			 Tj = Cr[WS(csr, 10)];
Chris@82: 			 Tk = Cr[WS(csr, 5)];
Chris@82: 			 Tl = Tj + Tk;
Chris@82: 			 TR = Tj - Tk;
Chris@82: 			 TN = Ci[WS(csi, 10)];
Chris@82: 			 TO = Ci[WS(csi, 5)];
Chris@82: 			 TP = TN + TO;
Chris@82: 			 T26 = TN - TO;
Chris@82: 		    }
Chris@82: 		    Tm = Ti + Tl;
Chris@82: 		    T2B = T26 + T25;
Chris@82: 		    TQ = TM - TP;
Chris@82: 		    TV = TR + TU;
Chris@82: 		    TW = FNMS(KP414213562, TV, TQ);
Chris@82: 		    T1k = FMA(KP414213562, TQ, TV);
Chris@82: 		    {
Chris@82: 			 E T1A, T1B, T24, T27;
Chris@82: 			 T1A = TR - TU;
Chris@82: 			 T1B = TM + TP;
Chris@82: 			 T1C = FMA(KP414213562, T1B, T1A);
Chris@82: 			 T1M = FNMS(KP414213562, T1A, T1B);
Chris@82: 			 T24 = Ti - Tl;
Chris@82: 			 T27 = T25 - T26;
Chris@82: 			 T28 = T24 - T27;
Chris@82: 			 T2m = T24 + T27;
Chris@82: 		    }
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E Tp, TX, T14, T2a, Ts, T15, T10, T2b, T11, T16;
Chris@82: 		    {
Chris@82: 			 E Tn, To, T12, T13;
Chris@82: 			 Tn = Cr[WS(csr, 1)];
Chris@82: 			 To = Cr[WS(csr, 14)];
Chris@82: 			 Tp = Tn + To;
Chris@82: 			 TX = Tn - To;
Chris@82: 			 T12 = Ci[WS(csi, 1)];
Chris@82: 			 T13 = Ci[WS(csi, 14)];
Chris@82: 			 T14 = T12 + T13;
Chris@82: 			 T2a = T13 - T12;
Chris@82: 		    }
Chris@82: 		    {
Chris@82: 			 E Tq, Tr, TY, TZ;
Chris@82: 			 Tq = Cr[WS(csr, 6)];
Chris@82: 			 Tr = Cr[WS(csr, 9)];
Chris@82: 			 Ts = Tq + Tr;
Chris@82: 			 T15 = Tq - Tr;
Chris@82: 			 TY = Ci[WS(csi, 6)];
Chris@82: 			 TZ = Ci[WS(csi, 9)];
Chris@82: 			 T10 = TY + TZ;
Chris@82: 			 T2b = TY - TZ;
Chris@82: 		    }
Chris@82: 		    Tt = Tp + Ts;
Chris@82: 		    T2A = T2b + T2a;
Chris@82: 		    T11 = TX - T10;
Chris@82: 		    T16 = T14 - T15;
Chris@82: 		    T17 = FNMS(KP414213562, T16, T11);
Chris@82: 		    T1j = FMA(KP414213562, T11, T16);
Chris@82: 		    {
Chris@82: 			 E T1D, T1E, T29, T2c;
Chris@82: 			 T1D = T15 + T14;
Chris@82: 			 T1E = TX + T10;
Chris@82: 			 T1F = FNMS(KP414213562, T1E, T1D);
Chris@82: 			 T1L = FMA(KP414213562, T1D, T1E);
Chris@82: 			 T29 = Tp - Ts;
Chris@82: 			 T2c = T2a - T2b;
Chris@82: 			 T2d = T29 + T2c;
Chris@82: 			 T2l = T29 - T2c;
Chris@82: 		    }
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E Tf, Tu, T2L, T2M, T2N, T2O;
Chris@82: 		    Tf = T7 + Te;
Chris@82: 		    Tu = Tm + Tt;
Chris@82: 		    T2L = Tf - Tu;
Chris@82: 		    T2M = T2B + T2A;
Chris@82: 		    T2N = T2F + T2E;
Chris@82: 		    T2O = T2M + T2N;
Chris@82: 		    R0[0] = KP2_000000000 * (Tf + Tu);
Chris@82: 		    R0[WS(rs, 8)] = KP2_000000000 * (T2N - T2M);
Chris@82: 		    R0[WS(rs, 4)] = KP1_414213562 * (T2L + T2O);
Chris@82: 		    R0[WS(rs, 12)] = KP1_414213562 * (T2O - T2L);
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T2t, T2y, T2w, T2x;
Chris@82: 		    {
Chris@82: 			 E T2r, T2s, T2u, T2v;
Chris@82: 			 T2r = T1Z - T22;
Chris@82: 			 T2s = T2m + T2l;
Chris@82: 			 T2t = FNMS(KP707106781, T2s, T2r);
Chris@82: 			 T2y = FMA(KP707106781, T2s, T2r);
Chris@82: 			 T2u = T2j + T2i;
Chris@82: 			 T2v = T28 - T2d;
Chris@82: 			 T2w = FNMS(KP707106781, T2v, T2u);
Chris@82: 			 T2x = FMA(KP707106781, T2v, T2u);
Chris@82: 		    }
Chris@82: 		    R0[WS(rs, 3)] = KP1_662939224 * (FMA(KP668178637, T2w, T2t));
Chris@82: 		    R0[WS(rs, 15)] = -(KP1_961570560 * (FNMS(KP198912367, T2x, T2y)));
Chris@82: 		    R0[WS(rs, 11)] = KP1_662939224 * (FNMS(KP668178637, T2t, T2w));
Chris@82: 		    R0[WS(rs, 7)] = KP1_961570560 * (FMA(KP198912367, T2y, T2x));
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T2D, T2K, T2I, T2J;
Chris@82: 		    {
Chris@82: 			 E T2z, T2C, T2G, T2H;
Chris@82: 			 T2z = T7 - Te;
Chris@82: 			 T2C = T2A - T2B;
Chris@82: 			 T2D = T2z + T2C;
Chris@82: 			 T2K = T2z - T2C;
Chris@82: 			 T2G = T2E - T2F;
Chris@82: 			 T2H = Tm - Tt;
Chris@82: 			 T2I = T2G - T2H;
Chris@82: 			 T2J = T2H + T2G;
Chris@82: 		    }
Chris@82: 		    R0[WS(rs, 2)] = KP1_847759065 * (FMA(KP414213562, T2I, T2D));
Chris@82: 		    R0[WS(rs, 14)] = -(KP1_847759065 * (FNMS(KP414213562, T2J, T2K)));
Chris@82: 		    R0[WS(rs, 10)] = KP1_847759065 * (FNMS(KP414213562, T2D, T2I));
Chris@82: 		    R0[WS(rs, 6)] = KP1_847759065 * (FMA(KP414213562, T2K, T2J));
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T19, T1o, T1m, T1n;
Chris@82: 		    {
Chris@82: 			 E TL, T18, T1i, T1l;
Chris@82: 			 TL = FMA(KP707106781, TK, Tz);
Chris@82: 			 T18 = TW + T17;
Chris@82: 			 T19 = FMA(KP923879532, T18, TL);
Chris@82: 			 T1o = FNMS(KP923879532, T18, TL);
Chris@82: 			 T1i = FNMS(KP707106781, T1h, T1e);
Chris@82: 			 T1l = T1j - T1k;
Chris@82: 			 T1m = FNMS(KP923879532, T1l, T1i);
Chris@82: 			 T1n = FMA(KP923879532, T1l, T1i);
Chris@82: 		    }
Chris@82: 		    R1[0] = KP1_990369453 * (FNMS(KP098491403, T1m, T19));
Chris@82: 		    R1[WS(rs, 12)] = -(KP1_546020906 * (FMA(KP820678790, T1n, T1o)));
Chris@82: 		    R1[WS(rs, 8)] = -(KP1_990369453 * (FMA(KP098491403, T19, T1m)));
Chris@82: 		    R1[WS(rs, 4)] = -(KP1_546020906 * (FNMS(KP820678790, T1o, T1n)));
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T1r, T1w, T1u, T1v;
Chris@82: 		    {
Chris@82: 			 E T1p, T1q, T1s, T1t;
Chris@82: 			 T1p = FNMS(KP707106781, TK, Tz);
Chris@82: 			 T1q = T1k + T1j;
Chris@82: 			 T1r = FNMS(KP923879532, T1q, T1p);
Chris@82: 			 T1w = FMA(KP923879532, T1q, T1p);
Chris@82: 			 T1s = FMA(KP707106781, T1h, T1e);
Chris@82: 			 T1t = TW - T17;
Chris@82: 			 T1u = FMA(KP923879532, T1t, T1s);
Chris@82: 			 T1v = FNMS(KP923879532, T1t, T1s);
Chris@82: 		    }
Chris@82: 		    R1[WS(rs, 2)] = KP1_763842528 * (FNMS(KP534511135, T1u, T1r));
Chris@82: 		    R1[WS(rs, 14)] = -(KP1_913880671 * (FMA(KP303346683, T1v, T1w)));
Chris@82: 		    R1[WS(rs, 10)] = -(KP1_763842528 * (FMA(KP534511135, T1r, T1u)));
Chris@82: 		    R1[WS(rs, 6)] = -(KP1_913880671 * (FNMS(KP303346683, T1w, T1v)));
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T1T, T1Y, T1W, T1X;
Chris@82: 		    {
Chris@82: 			 E T1R, T1S, T1U, T1V;
Chris@82: 			 T1R = FMA(KP707106781, T1y, T1x);
Chris@82: 			 T1S = T1M + T1L;
Chris@82: 			 T1T = FNMS(KP923879532, T1S, T1R);
Chris@82: 			 T1Y = FMA(KP923879532, T1S, T1R);
Chris@82: 			 T1U = FMA(KP707106781, T1J, T1I);
Chris@82: 			 T1V = T1C + T1F;
Chris@82: 			 T1W = FNMS(KP923879532, T1V, T1U);
Chris@82: 			 T1X = FMA(KP923879532, T1V, T1U);
Chris@82: 		    }
Chris@82: 		    R1[WS(rs, 3)] = KP1_546020906 * (FMA(KP820678790, T1W, T1T));
Chris@82: 		    R1[WS(rs, 15)] = -(KP1_990369453 * (FNMS(KP098491403, T1X, T1Y)));
Chris@82: 		    R1[WS(rs, 11)] = KP1_546020906 * (FNMS(KP820678790, T1T, T1W));
Chris@82: 		    R1[WS(rs, 7)] = KP1_990369453 * (FMA(KP098491403, T1Y, T1X));
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T2f, T2q, T2o, T2p;
Chris@82: 		    {
Chris@82: 			 E T23, T2e, T2k, T2n;
Chris@82: 			 T23 = T1Z + T22;
Chris@82: 			 T2e = T28 + T2d;
Chris@82: 			 T2f = FMA(KP707106781, T2e, T23);
Chris@82: 			 T2q = FNMS(KP707106781, T2e, T23);
Chris@82: 			 T2k = T2i - T2j;
Chris@82: 			 T2n = T2l - T2m;
Chris@82: 			 T2o = FMA(KP707106781, T2n, T2k);
Chris@82: 			 T2p = FNMS(KP707106781, T2n, T2k);
Chris@82: 		    }
Chris@82: 		    R0[WS(rs, 1)] = KP1_961570560 * (FMA(KP198912367, T2o, T2f));
Chris@82: 		    R0[WS(rs, 13)] = -(KP1_662939224 * (FNMS(KP668178637, T2p, T2q)));
Chris@82: 		    R0[WS(rs, 9)] = KP1_961570560 * (FNMS(KP198912367, T2f, T2o));
Chris@82: 		    R0[WS(rs, 5)] = KP1_662939224 * (FMA(KP668178637, T2q, T2p));
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T1H, T1Q, T1O, T1P;
Chris@82: 		    {
Chris@82: 			 E T1z, T1G, T1K, T1N;
Chris@82: 			 T1z = FNMS(KP707106781, T1y, T1x);
Chris@82: 			 T1G = T1C - T1F;
Chris@82: 			 T1H = FMA(KP923879532, T1G, T1z);
Chris@82: 			 T1Q = FNMS(KP923879532, T1G, T1z);
Chris@82: 			 T1K = FNMS(KP707106781, T1J, T1I);
Chris@82: 			 T1N = T1L - T1M;
Chris@82: 			 T1O = FMA(KP923879532, T1N, T1K);
Chris@82: 			 T1P = FNMS(KP923879532, T1N, T1K);
Chris@82: 		    }
Chris@82: 		    R1[WS(rs, 1)] = KP1_913880671 * (FMA(KP303346683, T1O, T1H));
Chris@82: 		    R1[WS(rs, 13)] = -(KP1_763842528 * (FNMS(KP534511135, T1P, T1Q)));
Chris@82: 		    R1[WS(rs, 9)] = KP1_913880671 * (FNMS(KP303346683, T1H, T1O));
Chris@82: 		    R1[WS(rs, 5)] = KP1_763842528 * (FMA(KP534511135, T1Q, T1P));
Chris@82: 	       }
Chris@82: 	  }
Chris@82:      }
Chris@82: }
Chris@82: 
Chris@82: static const kr2c_desc desc = { 32, "r2cbIII_32", {106, 32, 68, 0}, &GENUS };
Chris@82: 
Chris@82: void X(codelet_r2cbIII_32) (planner *p) {
Chris@82:      X(kr2c_register) (p, r2cbIII_32, &desc);
Chris@82: }
Chris@82: 
Chris@82: #else
Chris@82: 
Chris@82: /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -name r2cbIII_32 -dft-III -include rdft/scalar/r2cbIII.h */
Chris@82: 
Chris@82: /*
Chris@82:  * This function contains 174 FP additions, 84 FP multiplications,
Chris@82:  * (or, 138 additions, 48 multiplications, 36 fused multiply/add),
Chris@82:  * 66 stack variables, 19 constants, and 64 memory accesses
Chris@82:  */
Chris@82: #include "rdft/scalar/r2cbIII.h"
Chris@82: 
Chris@82: static void r2cbIII_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
Chris@82: {
Chris@82:      DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
Chris@82:      DK(KP580569354, +0.580569354508924735272384751634790549382952557);
Chris@82:      DK(KP942793473, +0.942793473651995297112775251810508755314920638);
Chris@82:      DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
Chris@82:      DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
Chris@82:      DK(KP1_268786568, +1.268786568327290996430343226450986741351374190);
Chris@82:      DK(KP196034280, +0.196034280659121203988391127777283691722273346);
Chris@82:      DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
Chris@82:      DK(KP765366864, +0.765366864730179543456919968060797733522689125);
Chris@82:      DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
Chris@82:      DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
Chris@82:      DK(KP390180644, +0.390180644032256535696569736954044481855383236);
Chris@82:      DK(KP1_111140466, +1.111140466039204449485661627897065748749874382);
Chris@82:      DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
Chris@82:      DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
Chris@82:      DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
Chris@82:      DK(KP382683432, +0.382683432365089771728459984030398866761344562);
Chris@82:      DK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@82:      DK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@82:      {
Chris@82: 	  INT i;
Chris@82: 	  for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
Chris@82: 	       E T7, T2i, T2F, Tz, T1k, T1I, T1Z, T1x, Te, T22, T2E, T2j, T1f, T1y, TK;
Chris@82: 	       E T1J, Tm, T2B, TW, T1a, T1C, T1L, T28, T2l, Tt, T2A, T17, T1b, T1F, T1M;
Chris@82: 	       E T2d, T2m;
Chris@82: 	       {
Chris@82: 		    E T3, Tv, T1j, T2h, T6, T1g, Ty, T2g;
Chris@82: 		    {
Chris@82: 			 E T1, T2, T1h, T1i;
Chris@82: 			 T1 = Cr[0];
Chris@82: 			 T2 = Cr[WS(csr, 15)];
Chris@82: 			 T3 = T1 + T2;
Chris@82: 			 Tv = T1 - T2;
Chris@82: 			 T1h = Ci[0];
Chris@82: 			 T1i = Ci[WS(csi, 15)];
Chris@82: 			 T1j = T1h + T1i;
Chris@82: 			 T2h = T1i - T1h;
Chris@82: 		    }
Chris@82: 		    {
Chris@82: 			 E T4, T5, Tw, Tx;
Chris@82: 			 T4 = Cr[WS(csr, 8)];
Chris@82: 			 T5 = Cr[WS(csr, 7)];
Chris@82: 			 T6 = T4 + T5;
Chris@82: 			 T1g = T4 - T5;
Chris@82: 			 Tw = Ci[WS(csi, 8)];
Chris@82: 			 Tx = Ci[WS(csi, 7)];
Chris@82: 			 Ty = Tw + Tx;
Chris@82: 			 T2g = Tw - Tx;
Chris@82: 		    }
Chris@82: 		    T7 = T3 + T6;
Chris@82: 		    T2i = T2g + T2h;
Chris@82: 		    T2F = T2h - T2g;
Chris@82: 		    Tz = Tv - Ty;
Chris@82: 		    T1k = T1g + T1j;
Chris@82: 		    T1I = T1g - T1j;
Chris@82: 		    T1Z = T3 - T6;
Chris@82: 		    T1x = Tv + Ty;
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E Ta, TA, TD, T21, Td, TF, TI, T20;
Chris@82: 		    {
Chris@82: 			 E T8, T9, TB, TC;
Chris@82: 			 T8 = Cr[WS(csr, 4)];
Chris@82: 			 T9 = Cr[WS(csr, 11)];
Chris@82: 			 Ta = T8 + T9;
Chris@82: 			 TA = T8 - T9;
Chris@82: 			 TB = Ci[WS(csi, 4)];
Chris@82: 			 TC = Ci[WS(csi, 11)];
Chris@82: 			 TD = TB + TC;
Chris@82: 			 T21 = TB - TC;
Chris@82: 		    }
Chris@82: 		    {
Chris@82: 			 E Tb, Tc, TG, TH;
Chris@82: 			 Tb = Cr[WS(csr, 3)];
Chris@82: 			 Tc = Cr[WS(csr, 12)];
Chris@82: 			 Td = Tb + Tc;
Chris@82: 			 TF = Tb - Tc;
Chris@82: 			 TG = Ci[WS(csi, 3)];
Chris@82: 			 TH = Ci[WS(csi, 12)];
Chris@82: 			 TI = TG + TH;
Chris@82: 			 T20 = TH - TG;
Chris@82: 		    }
Chris@82: 		    Te = Ta + Td;
Chris@82: 		    T22 = T20 - T21;
Chris@82: 		    T2E = T21 + T20;
Chris@82: 		    T2j = Ta - Td;
Chris@82: 		    {
Chris@82: 			 E T1d, T1e, TE, TJ;
Chris@82: 			 T1d = TA + TD;
Chris@82: 			 T1e = TF + TI;
Chris@82: 			 T1f = KP707106781 * (T1d - T1e);
Chris@82: 			 T1y = KP707106781 * (T1d + T1e);
Chris@82: 			 TE = TA - TD;
Chris@82: 			 TJ = TF - TI;
Chris@82: 			 TK = KP707106781 * (TE + TJ);
Chris@82: 			 T1J = KP707106781 * (TE - TJ);
Chris@82: 		    }
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E Ti, TM, TU, T25, Tl, TR, TP, T26, TQ, TV;
Chris@82: 		    {
Chris@82: 			 E Tg, Th, TS, TT;
Chris@82: 			 Tg = Cr[WS(csr, 2)];
Chris@82: 			 Th = Cr[WS(csr, 13)];
Chris@82: 			 Ti = Tg + Th;
Chris@82: 			 TM = Tg - Th;
Chris@82: 			 TS = Ci[WS(csi, 2)];
Chris@82: 			 TT = Ci[WS(csi, 13)];
Chris@82: 			 TU = TS + TT;
Chris@82: 			 T25 = TS - TT;
Chris@82: 		    }
Chris@82: 		    {
Chris@82: 			 E Tj, Tk, TN, TO;
Chris@82: 			 Tj = Cr[WS(csr, 10)];
Chris@82: 			 Tk = Cr[WS(csr, 5)];
Chris@82: 			 Tl = Tj + Tk;
Chris@82: 			 TR = Tj - Tk;
Chris@82: 			 TN = Ci[WS(csi, 10)];
Chris@82: 			 TO = Ci[WS(csi, 5)];
Chris@82: 			 TP = TN + TO;
Chris@82: 			 T26 = TN - TO;
Chris@82: 		    }
Chris@82: 		    Tm = Ti + Tl;
Chris@82: 		    T2B = T26 + T25;
Chris@82: 		    TQ = TM - TP;
Chris@82: 		    TV = TR + TU;
Chris@82: 		    TW = FNMS(KP382683432, TV, KP923879532 * TQ);
Chris@82: 		    T1a = FMA(KP382683432, TQ, KP923879532 * TV);
Chris@82: 		    {
Chris@82: 			 E T1A, T1B, T24, T27;
Chris@82: 			 T1A = TM + TP;
Chris@82: 			 T1B = TU - TR;
Chris@82: 			 T1C = FNMS(KP923879532, T1B, KP382683432 * T1A);
Chris@82: 			 T1L = FMA(KP923879532, T1A, KP382683432 * T1B);
Chris@82: 			 T24 = Ti - Tl;
Chris@82: 			 T27 = T25 - T26;
Chris@82: 			 T28 = T24 - T27;
Chris@82: 			 T2l = T24 + T27;
Chris@82: 		    }
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E Tp, TX, T15, T2a, Ts, T12, T10, T2b, T11, T16;
Chris@82: 		    {
Chris@82: 			 E Tn, To, T13, T14;
Chris@82: 			 Tn = Cr[WS(csr, 1)];
Chris@82: 			 To = Cr[WS(csr, 14)];
Chris@82: 			 Tp = Tn + To;
Chris@82: 			 TX = Tn - To;
Chris@82: 			 T13 = Ci[WS(csi, 1)];
Chris@82: 			 T14 = Ci[WS(csi, 14)];
Chris@82: 			 T15 = T13 + T14;
Chris@82: 			 T2a = T14 - T13;
Chris@82: 		    }
Chris@82: 		    {
Chris@82: 			 E Tq, Tr, TY, TZ;
Chris@82: 			 Tq = Cr[WS(csr, 6)];
Chris@82: 			 Tr = Cr[WS(csr, 9)];
Chris@82: 			 Ts = Tq + Tr;
Chris@82: 			 T12 = Tq - Tr;
Chris@82: 			 TY = Ci[WS(csi, 6)];
Chris@82: 			 TZ = Ci[WS(csi, 9)];
Chris@82: 			 T10 = TY + TZ;
Chris@82: 			 T2b = TY - TZ;
Chris@82: 		    }
Chris@82: 		    Tt = Tp + Ts;
Chris@82: 		    T2A = T2b + T2a;
Chris@82: 		    T11 = TX - T10;
Chris@82: 		    T16 = T12 - T15;
Chris@82: 		    T17 = FMA(KP923879532, T11, KP382683432 * T16);
Chris@82: 		    T1b = FNMS(KP382683432, T11, KP923879532 * T16);
Chris@82: 		    {
Chris@82: 			 E T1D, T1E, T29, T2c;
Chris@82: 			 T1D = TX + T10;
Chris@82: 			 T1E = T12 + T15;
Chris@82: 			 T1F = FNMS(KP923879532, T1E, KP382683432 * T1D);
Chris@82: 			 T1M = FMA(KP923879532, T1D, KP382683432 * T1E);
Chris@82: 			 T29 = Tp - Ts;
Chris@82: 			 T2c = T2a - T2b;
Chris@82: 			 T2d = T29 + T2c;
Chris@82: 			 T2m = T2c - T29;
Chris@82: 		    }
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E Tf, Tu, T2L, T2M, T2N, T2O;
Chris@82: 		    Tf = T7 + Te;
Chris@82: 		    Tu = Tm + Tt;
Chris@82: 		    T2L = Tf - Tu;
Chris@82: 		    T2M = T2B + T2A;
Chris@82: 		    T2N = T2F - T2E;
Chris@82: 		    T2O = T2M + T2N;
Chris@82: 		    R0[0] = KP2_000000000 * (Tf + Tu);
Chris@82: 		    R0[WS(rs, 8)] = KP2_000000000 * (T2N - T2M);
Chris@82: 		    R0[WS(rs, 4)] = KP1_414213562 * (T2L + T2O);
Chris@82: 		    R0[WS(rs, 12)] = KP1_414213562 * (T2O - T2L);
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T2t, T2x, T2w, T2y;
Chris@82: 		    {
Chris@82: 			 E T2r, T2s, T2u, T2v;
Chris@82: 			 T2r = T1Z - T22;
Chris@82: 			 T2s = KP707106781 * (T2m - T2l);
Chris@82: 			 T2t = T2r + T2s;
Chris@82: 			 T2x = T2r - T2s;
Chris@82: 			 T2u = T2j + T2i;
Chris@82: 			 T2v = KP707106781 * (T28 - T2d);
Chris@82: 			 T2w = T2u - T2v;
Chris@82: 			 T2y = T2v + T2u;
Chris@82: 		    }
Chris@82: 		    R0[WS(rs, 3)] = FMA(KP1_662939224, T2t, KP1_111140466 * T2w);
Chris@82: 		    R0[WS(rs, 15)] = FNMS(KP1_961570560, T2x, KP390180644 * T2y);
Chris@82: 		    R0[WS(rs, 11)] = FNMS(KP1_111140466, T2t, KP1_662939224 * T2w);
Chris@82: 		    R0[WS(rs, 7)] = FMA(KP390180644, T2x, KP1_961570560 * T2y);
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T2D, T2J, T2I, T2K;
Chris@82: 		    {
Chris@82: 			 E T2z, T2C, T2G, T2H;
Chris@82: 			 T2z = T7 - Te;
Chris@82: 			 T2C = T2A - T2B;
Chris@82: 			 T2D = T2z + T2C;
Chris@82: 			 T2J = T2z - T2C;
Chris@82: 			 T2G = T2E + T2F;
Chris@82: 			 T2H = Tm - Tt;
Chris@82: 			 T2I = T2G - T2H;
Chris@82: 			 T2K = T2H + T2G;
Chris@82: 		    }
Chris@82: 		    R0[WS(rs, 2)] = FMA(KP1_847759065, T2D, KP765366864 * T2I);
Chris@82: 		    R0[WS(rs, 14)] = FNMS(KP1_847759065, T2J, KP765366864 * T2K);
Chris@82: 		    R0[WS(rs, 10)] = FNMS(KP765366864, T2D, KP1_847759065 * T2I);
Chris@82: 		    R0[WS(rs, 6)] = FMA(KP765366864, T2J, KP1_847759065 * T2K);
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T19, T1n, T1m, T1o;
Chris@82: 		    {
Chris@82: 			 E TL, T18, T1c, T1l;
Chris@82: 			 TL = Tz + TK;
Chris@82: 			 T18 = TW + T17;
Chris@82: 			 T19 = TL + T18;
Chris@82: 			 T1n = TL - T18;
Chris@82: 			 T1c = T1a + T1b;
Chris@82: 			 T1l = T1f + T1k;
Chris@82: 			 T1m = T1c + T1l;
Chris@82: 			 T1o = T1c - T1l;
Chris@82: 		    }
Chris@82: 		    R1[0] = FNMS(KP196034280, T1m, KP1_990369453 * T19);
Chris@82: 		    R1[WS(rs, 12)] = FNMS(KP1_546020906, T1n, KP1_268786568 * T1o);
Chris@82: 		    R1[WS(rs, 8)] = -(FMA(KP196034280, T19, KP1_990369453 * T1m));
Chris@82: 		    R1[WS(rs, 4)] = FMA(KP1_268786568, T1n, KP1_546020906 * T1o);
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T1r, T1v, T1u, T1w;
Chris@82: 		    {
Chris@82: 			 E T1p, T1q, T1s, T1t;
Chris@82: 			 T1p = Tz - TK;
Chris@82: 			 T1q = T1b - T1a;
Chris@82: 			 T1r = T1p + T1q;
Chris@82: 			 T1v = T1p - T1q;
Chris@82: 			 T1s = T1f - T1k;
Chris@82: 			 T1t = TW - T17;
Chris@82: 			 T1u = T1s - T1t;
Chris@82: 			 T1w = T1t + T1s;
Chris@82: 		    }
Chris@82: 		    R1[WS(rs, 2)] = FMA(KP1_763842528, T1r, KP942793473 * T1u);
Chris@82: 		    R1[WS(rs, 14)] = FNMS(KP1_913880671, T1v, KP580569354 * T1w);
Chris@82: 		    R1[WS(rs, 10)] = FNMS(KP942793473, T1r, KP1_763842528 * T1u);
Chris@82: 		    R1[WS(rs, 6)] = FMA(KP580569354, T1v, KP1_913880671 * T1w);
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T1T, T1X, T1W, T1Y;
Chris@82: 		    {
Chris@82: 			 E T1R, T1S, T1U, T1V;
Chris@82: 			 T1R = T1x + T1y;
Chris@82: 			 T1S = T1L + T1M;
Chris@82: 			 T1T = T1R - T1S;
Chris@82: 			 T1X = T1R + T1S;
Chris@82: 			 T1U = T1J + T1I;
Chris@82: 			 T1V = T1C - T1F;
Chris@82: 			 T1W = T1U - T1V;
Chris@82: 			 T1Y = T1V + T1U;
Chris@82: 		    }
Chris@82: 		    R1[WS(rs, 3)] = FMA(KP1_546020906, T1T, KP1_268786568 * T1W);
Chris@82: 		    R1[WS(rs, 15)] = FNMS(KP1_990369453, T1X, KP196034280 * T1Y);
Chris@82: 		    R1[WS(rs, 11)] = FNMS(KP1_268786568, T1T, KP1_546020906 * T1W);
Chris@82: 		    R1[WS(rs, 7)] = FMA(KP196034280, T1X, KP1_990369453 * T1Y);
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T2f, T2p, T2o, T2q;
Chris@82: 		    {
Chris@82: 			 E T23, T2e, T2k, T2n;
Chris@82: 			 T23 = T1Z + T22;
Chris@82: 			 T2e = KP707106781 * (T28 + T2d);
Chris@82: 			 T2f = T23 + T2e;
Chris@82: 			 T2p = T23 - T2e;
Chris@82: 			 T2k = T2i - T2j;
Chris@82: 			 T2n = KP707106781 * (T2l + T2m);
Chris@82: 			 T2o = T2k - T2n;
Chris@82: 			 T2q = T2n + T2k;
Chris@82: 		    }
Chris@82: 		    R0[WS(rs, 1)] = FMA(KP1_961570560, T2f, KP390180644 * T2o);
Chris@82: 		    R0[WS(rs, 13)] = FNMS(KP1_662939224, T2p, KP1_111140466 * T2q);
Chris@82: 		    R0[WS(rs, 9)] = FNMS(KP390180644, T2f, KP1_961570560 * T2o);
Chris@82: 		    R0[WS(rs, 5)] = FMA(KP1_111140466, T2p, KP1_662939224 * T2q);
Chris@82: 	       }
Chris@82: 	       {
Chris@82: 		    E T1H, T1P, T1O, T1Q;
Chris@82: 		    {
Chris@82: 			 E T1z, T1G, T1K, T1N;
Chris@82: 			 T1z = T1x - T1y;
Chris@82: 			 T1G = T1C + T1F;
Chris@82: 			 T1H = T1z + T1G;
Chris@82: 			 T1P = T1z - T1G;
Chris@82: 			 T1K = T1I - T1J;
Chris@82: 			 T1N = T1L - T1M;
Chris@82: 			 T1O = T1K - T1N;
Chris@82: 			 T1Q = T1N + T1K;
Chris@82: 		    }
Chris@82: 		    R1[WS(rs, 1)] = FMA(KP1_913880671, T1H, KP580569354 * T1O);
Chris@82: 		    R1[WS(rs, 13)] = FNMS(KP1_763842528, T1P, KP942793473 * T1Q);
Chris@82: 		    R1[WS(rs, 9)] = FNMS(KP580569354, T1H, KP1_913880671 * T1O);
Chris@82: 		    R1[WS(rs, 5)] = FMA(KP942793473, T1P, KP1_763842528 * T1Q);
Chris@82: 	       }
Chris@82: 	  }
Chris@82:      }
Chris@82: }
Chris@82: 
Chris@82: static const kr2c_desc desc = { 32, "r2cbIII_32", {138, 48, 36, 0}, &GENUS };
Chris@82: 
Chris@82: void X(codelet_r2cbIII_32) (planner *p) {
Chris@82:      X(kr2c_register) (p, r2cbIII_32, &desc);
Chris@82: }
Chris@82: 
Chris@82: #endif