hb_16.c

快速fft变换

			      }
			 }
		    }
	       }
	  }
	  ci[WS(rs, 1)] = FMA(T1K, T1I, T1M);
	  cr[WS(rs, 1)] = FNMS(T1K, T1L, T1J);
     }
}

static const tw_instr twinstr[] = {
     {TW_FULL, 1, 16},
     {TW_NEXT, 1, 0}
};

static const hc2hc_desc desc = { 16, "hb_16", twinstr, &GENUS, {104, 30, 70, 0} };

void X(codelet_hb_16) (planner *p) {
     X(khc2hc_register) (p, hb_16, &desc);
}
#else				/* HAVE_FMA */

/* Generated by: ../../../genfft/gen_hc2hc -compact -variables 4 -pipeline-latency 4 -sign 1 -n 16 -dif -name hb_16 -include hb.h */

/*
 * This function contains 174 FP additions, 84 FP multiplications,
 * (or, 136 additions, 46 multiplications, 38 fused multiply/add),
 * 50 stack variables, 3 constants, and 64 memory accesses
 */
#include "hb.h"

static void hb_16(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
{
     DK(KP382683432, +0.382683432365089771728459984030398866761344562);
     DK(KP923879532, +0.923879532511286756128183189396788286822416626);
     DK(KP707106781, +0.707106781186547524400844362104849039284835938);
     INT m;
     for (m = mb, W = W + ((mb - 1) * 30); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 30, MAKE_VOLATILE_STRIDE(rs)) {
	  E T7, T2K, T2W, Tw, T17, T1S, T2k, T1w, Te, TD, T1x, T10, T2n, T2L, T1Z;
	  E T2X, Tm, T1z, TN, T19, T2e, T2p, T2P, T2Z, Tt, T1A, TW, T1a, T27, T2q;
	  E T2S, T30;
	  {
	       E T3, T1Q, T16, T1R, T6, T2i, T13, T2j;
	       {
		    E T1, T2, T14, T15;
		    T1 = cr[0];
		    T2 = ci[WS(rs, 7)];
		    T3 = T1 + T2;
		    T1Q = T1 - T2;
		    T14 = ci[WS(rs, 11)];
		    T15 = cr[WS(rs, 12)];
		    T16 = T14 - T15;
		    T1R = T14 + T15;
	       }
	       {
		    E T4, T5, T11, T12;
		    T4 = cr[WS(rs, 4)];
		    T5 = ci[WS(rs, 3)];
		    T6 = T4 + T5;
		    T2i = T4 - T5;
		    T11 = ci[WS(rs, 15)];
		    T12 = cr[WS(rs, 8)];
		    T13 = T11 - T12;
		    T2j = T11 + T12;
	       }
	       T7 = T3 + T6;
	       T2K = T1Q + T1R;
	       T2W = T2j - T2i;
	       Tw = T3 - T6;
	       T17 = T13 - T16;
	       T1S = T1Q - T1R;
	       T2k = T2i + T2j;
	       T1w = T13 + T16;
	  }
	  {
	       E Ta, T1T, TC, T1U, Td, T1W, Tz, T1X;
	       {
		    E T8, T9, TA, TB;
		    T8 = cr[WS(rs, 2)];
		    T9 = ci[WS(rs, 5)];
		    Ta = T8 + T9;
		    T1T = T8 - T9;
		    TA = ci[WS(rs, 13)];
		    TB = cr[WS(rs, 10)];
		    TC = TA - TB;
		    T1U = TA + TB;
	       }
	       {
		    E Tb, Tc, Tx, Ty;
		    Tb = ci[WS(rs, 1)];
		    Tc = cr[WS(rs, 6)];
		    Td = Tb + Tc;
		    T1W = Tb - Tc;
		    Tx = ci[WS(rs, 9)];
		    Ty = cr[WS(rs, 14)];
		    Tz = Tx - Ty;
		    T1X = Tx + Ty;
	       }
	       Te = Ta + Td;
	       TD = Tz - TC;
	       T1x = TC + Tz;
	       T10 = Ta - Td;
	       {
		    E T2l, T2m, T1V, T1Y;
		    T2l = T1T + T1U;
		    T2m = T1W + T1X;
		    T2n = KP707106781 * (T2l - T2m);
		    T2L = KP707106781 * (T2l + T2m);
		    T1V = T1T - T1U;
		    T1Y = T1W - T1X;
		    T1Z = KP707106781 * (T1V + T1Y);
		    T2X = KP707106781 * (T1V - T1Y);
	       }
	  }
	  {
	       E Ti, T2b, TL, T2c, Tl, T28, TI, T29, TF, TM;
	       {
		    E Tg, Th, TJ, TK;
		    Tg = cr[WS(rs, 1)];
		    Th = ci[WS(rs, 6)];
		    Ti = Tg + Th;
		    T2b = Tg - Th;
		    TJ = ci[WS(rs, 10)];
		    TK = cr[WS(rs, 13)];
		    TL = TJ - TK;
		    T2c = TJ + TK;
	       }
	       {
		    E Tj, Tk, TG, TH;
		    Tj = cr[WS(rs, 5)];
		    Tk = ci[WS(rs, 2)];
		    Tl = Tj + Tk;
		    T28 = Tj - Tk;
		    TG = ci[WS(rs, 14)];
		    TH = cr[WS(rs, 9)];
		    TI = TG - TH;
		    T29 = TG + TH;
	       }
	       Tm = Ti + Tl;
	       T1z = TI + TL;
	       TF = Ti - Tl;
	       TM = TI - TL;
	       TN = TF - TM;
	       T19 = TF + TM;
	       {
		    E T2a, T2d, T2N, T2O;
		    T2a = T28 + T29;
		    T2d = T2b - T2c;
		    T2e = FMA(KP923879532, T2a, KP382683432 * T2d);
		    T2p = FNMS(KP382683432, T2a, KP923879532 * T2d);
		    T2N = T2b + T2c;
		    T2O = T29 - T28;
		    T2P = FNMS(KP923879532, T2O, KP382683432 * T2N);
		    T2Z = FMA(KP382683432, T2O, KP923879532 * T2N);
	       }
	  }
	  {
	       E Tp, T24, TU, T25, Ts, T21, TR, T22, TO, TV;
	       {
		    E Tn, To, TS, TT;
		    Tn = ci[0];
		    To = cr[WS(rs, 7)];
		    Tp = Tn + To;
		    T24 = Tn - To;
		    TS = ci[WS(rs, 12)];
		    TT = cr[WS(rs, 11)];
		    TU = TS - TT;
		    T25 = TS + TT;
	       }
	       {
		    E Tq, Tr, TP, TQ;
		    Tq = cr[WS(rs, 3)];
		    Tr = ci[WS(rs, 4)];
		    Ts = Tq + Tr;
		    T21 = Tq - Tr;
		    TP = ci[WS(rs, 8)];
		    TQ = cr[WS(rs, 15)];
		    TR = TP - TQ;
		    T22 = TP + TQ;
	       }
	       Tt = Tp + Ts;
	       T1A = TR + TU;
	       TO = Tp - Ts;
	       TV = TR - TU;
	       TW = TO + TV;
	       T1a = TV - TO;
	       {
		    E T23, T26, T2Q, T2R;
		    T23 = T21 - T22;
		    T26 = T24 - T25;
		    T27 = FNMS(KP382683432, T26, KP923879532 * T23);
		    T2q = FMA(KP382683432, T23, KP923879532 * T26);
		    T2Q = T24 + T25;
		    T2R = T21 + T22;
		    T2S = FNMS(KP923879532, T2R, KP382683432 * T2Q);
		    T30 = FMA(KP382683432, T2R, KP923879532 * T2Q);
	       }
	  }
	  {
	       E Tf, Tu, T1u, T1y, T1B, T1C, T1t, T1v;
	       Tf = T7 + Te;
	       Tu = Tm + Tt;
	       T1u = Tf - Tu;
	       T1y = T1w + T1x;
	       T1B = T1z + T1A;
	       T1C = T1y - T1B;
	       cr[0] = Tf + Tu;
	       ci[0] = T1y + T1B;
	       T1t = W[14];
	       T1v = W[15];
	       cr[WS(rs, 8)] = FNMS(T1v, T1C, T1t * T1u);
	       ci[WS(rs, 8)] = FMA(T1v, T1u, T1t * T1C);
	  }
	  {
	       E T2U, T34, T32, T36;
	       {
		    E T2M, T2T, T2Y, T31;
		    T2M = T2K - T2L;
		    T2T = T2P + T2S;
		    T2U = T2M - T2T;
		    T34 = T2M + T2T;
		    T2Y = T2W + T2X;
		    T31 = T2Z - T30;
		    T32 = T2Y - T31;
		    T36 = T2Y + T31;
	       }
	       {
		    E T2J, T2V, T33, T35;
		    T2J = W[20];
		    T2V = W[21];
		    cr[WS(rs, 11)] = FNMS(T2V, T32, T2J * T2U);
		    ci[WS(rs, 11)] = FMA(T2V, T2U, T2J * T32);
		    T33 = W[4];
		    T35 = W[5];
		    cr[WS(rs, 3)] = FNMS(T35, T36, T33 * T34);
		    ci[WS(rs, 3)] = FMA(T35, T34, T33 * T36);
	       }
	  }
	  {
	       E T3a, T3g, T3e, T3i;
	       {
		    E T38, T39, T3c, T3d;
		    T38 = T2K + T2L;
		    T39 = T2Z + T30;
		    T3a = T38 - T39;
		    T3g = T38 + T39;
		    T3c = T2W - T2X;
		    T3d = T2P - T2S;
		    T3e = T3c + T3d;
		    T3i = T3c - T3d;
	       }
	       {
		    E T37, T3b, T3f, T3h;
		    T37 = W[12];
		    T3b = W[13];
		    cr[WS(rs, 7)] = FNMS(T3b, T3e, T37 * T3a);
		    ci[WS(rs, 7)] = FMA(T37, T3e, T3b * T3a);
		    T3f = W[28];
		    T3h = W[29];
		    cr[WS(rs, 15)] = FNMS(T3h, T3i, T3f * T3g);
		    ci[WS(rs, 15)] = FMA(T3f, T3i, T3h * T3g);
	       }
	  }
	  {
	       E TY, T1e, T1c, T1g;
	       {
		    E TE, TX, T18, T1b;
		    TE = Tw + TD;
		    TX = KP707106781 * (TN + TW);
		    TY = TE - TX;
		    T1e = TE + TX;
		    T18 = T10 + T17;
		    T1b = KP707106781 * (T19 + T1a);
		    T1c = T18 - T1b;
		    T1g = T18 + T1b;
	       }
	       {
		    E Tv, TZ, T1d, T1f;
		    Tv = W[18];
		    TZ = W[19];
		    cr[WS(rs, 10)] = FNMS(TZ, T1c, Tv * TY);
		    ci[WS(rs, 10)] = FMA(TZ, TY, Tv * T1c);
		    T1d = W[2];
		    T1f = W[3];
		    cr[WS(rs, 2)] = FNMS(T1f, T1g, T1d * T1e);
		    ci[WS(rs, 2)] = FMA(T1f, T1e, T1d * T1g);
	       }
	  }
	  {
	       E T1k, T1q, T1o, T1s;
	       {
		    E T1i, T1j, T1m, T1n;
		    T1i = Tw - TD;
		    T1j = KP707106781 * (T1a - T19);
		    T1k = T1i - T1j;
		    T1q = T1i + T1j;
		    T1m = T17 - T10;
		    T1n = KP707106781 * (TN - TW);
		    T1o = T1m - T1n;
		    T1s = T1m + T1n;
	       }
	       {
		    E T1h, T1l, T1p, T1r;
		    T1h = W[26];
		    T1l = W[27];
		    cr[WS(rs, 14)] = FNMS(T1l, T1o, T1h * T1k);
		    ci[WS(rs, 14)] = FMA(T1h, T1o, T1l * T1k);
		    T1p = W[10];
		    T1r = W[11];
		    cr[WS(rs, 6)] = FNMS(T1r, T1s, T1p * T1q);
		    ci[WS(rs, 6)] = FMA(T1p, T1s, T1r * T1q);
	       }
	  }
	  {
	       E T2g, T2u, T2s, T2w;
	       {
		    E T20, T2f, T2o, T2r;
		    T20 = T1S - T1Z;
		    T2f = T27 - T2e;
		    T2g = T20 - T2f;
		    T2u = T20 + T2f;
		    T2o = T2k - T2n;
		    T2r = T2p - T2q;
		    T2s = T2o - T2r;
		    T2w = T2o + T2r;
	       }
	       {
		    E T1P, T2h, T2t, T2v;
		    T1P = W[24];
		    T2h = W[25];
		    cr[WS(rs, 13)] = FNMS(T2h, T2s, T1P * T2g);
		    ci[WS(rs, 13)] = FMA(T2h, T2g, T1P * T2s);
		    T2t = W[8];
		    T2v = W[9];
		    cr[WS(rs, 5)] = FNMS(T2v, T2w, T2t * T2u);
		    ci[WS(rs, 5)] = FMA(T2v, T2u, T2t * T2w);
	       }
	  }
	  {
	       E T2A, T2G, T2E, T2I;
	       {
		    E T2y, T2z, T2C, T2D;
		    T2y = T1S + T1Z;
		    T2z = T2p + T2q;
		    T2A = T2y - T2z;
		    T2G = T2y + T2z;
		    T2C = T2k + T2n;
		    T2D = T2e + T27;
		    T2E = T2C - T2D;
		    T2I = T2C + T2D;
	       }
	       {
		    E T2x, T2B, T2F, T2H;
		    T2x = W[16];
		    T2B = W[17];
		    cr[WS(rs, 9)] = FNMS(T2B, T2E, T2x * T2A);
		    ci[WS(rs, 9)] = FMA(T2x, T2E, T2B * T2A);
		    T2F = W[0];
		    T2H = W[1];
		    cr[WS(rs, 1)] = FNMS(T2H, T2I, T2F * T2G);
		    ci[WS(rs, 1)] = FMA(T2F, T2I, T2H * T2G);
	       }
	  }
	  {
	       E T1G, T1M, T1K, T1O;
	       {
		    E T1E, T1F, T1I, T1J;
		    T1E = T7 - Te;
		    T1F = T1A - T1z;
		    T1G = T1E - T1F;
		    T1M = T1E + T1F;
		    T1I = T1w - T1x;
		    T1J = Tm - Tt;
		    T1K = T1I - T1J;
		    T1O = T1J + T1I;
	       }
	       {
		    E T1D, T1H, T1L, T1N;
		    T1D = W[22];
		    T1H = W[23];
		    cr[WS(rs, 12)] = FNMS(T1H, T1K, T1D * T1G);
		    ci[WS(rs, 12)] = FMA(T1D, T1K, T1H * T1G);
		    T1L = W[6];
		    T1N = W[7];
		    cr[WS(rs, 4)] = FNMS(T1N, T1O, T1L * T1M);
		    ci[WS(rs, 4)] = FMA(T1L, T1O, T1N * T1M);
	       }
	  }
     }
}

static const tw_instr twinstr[] = {
     {TW_FULL, 1, 16},
     {TW_NEXT, 1, 0}
};

static const hc2hc_desc desc = { 16, "hb_16", twinstr, &GENUS, {136, 46, 38, 0} };

void X(codelet_hb_16) (planner *p) {
     X(khc2hc_register) (p, hb_16, &desc);
}
#endif				/* HAVE_FMA */