hb2_16.c

快速fft变换

static const tw_instr twinstr[] = {
     {TW_CEXP, 1, 1},
     {TW_CEXP, 1, 3},
     {TW_CEXP, 1, 9},
     {TW_CEXP, 1, 15},
     {TW_NEXT, 1, 0}
};

static const hc2hc_desc desc = { 16, "hb2_16", twinstr, &GENUS, {104, 42, 92, 0} };

void X(codelet_hb2_16) (planner *p) {
     X(khc2hc_register) (p, hb2_16, &desc);
}
#else				/* HAVE_FMA */

/* Generated by: ../../../genfft/gen_hc2hc -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 16 -dif -name hb2_16 -include hb.h */

/*
 * This function contains 196 FP additions, 108 FP multiplications,
 * (or, 156 additions, 68 multiplications, 40 fused multiply/add),
 * 80 stack variables, 3 constants, and 64 memory accesses
 */
#include "hb.h"

static void hb2_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) * 8); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 8, MAKE_VOLATILE_STRIDE(rs)) {
	  E Tv, Ty, T1l, T1n, T1p, T1t, T27, T25, Tz, Tw, TB, T21, T1P, T1H, T1X;
	  E T17, T1L, T1N, T1v, T1w, T1x, T1B, T2F, T2T, T2b, T2R, T3j, T3x, T35, T3t;
	  {
	       E TA, T1J, T15, T1G, Tx, T1K, T16, T1F;
	       {
		    E T1m, T1s, T1o, T1r;
		    Tv = W[0];
		    Ty = W[1];
		    T1l = W[2];
		    T1n = W[3];
		    T1m = Tv * T1l;
		    T1s = Ty * T1l;
		    T1o = Ty * T1n;
		    T1r = Tv * T1n;
		    T1p = T1m + T1o;
		    T1t = T1r - T1s;
		    T27 = T1r + T1s;
		    T25 = T1m - T1o;
		    Tz = W[5];
		    TA = Ty * Tz;
		    T1J = T1l * Tz;
		    T15 = Tv * Tz;
		    T1G = T1n * Tz;
		    Tw = W[4];
		    Tx = Tv * Tw;
		    T1K = T1n * Tw;
		    T16 = Ty * Tw;
		    T1F = T1l * Tw;
	       }
	       TB = Tx - TA;
	       T21 = T1J + T1K;
	       T1P = T15 - T16;
	       T1H = T1F + T1G;
	       T1X = T1F - T1G;
	       T17 = T15 + T16;
	       T1L = T1J - T1K;
	       T1N = Tx + TA;
	       T1v = W[6];
	       T1w = W[7];
	       T1x = FMA(Tv, T1v, Ty * T1w);
	       T1B = FNMS(Ty, T1v, Tv * T1w);
	       {
		    E T2D, T2E, T29, T2a;
		    T2D = T25 * Tz;
		    T2E = T27 * Tw;
		    T2F = T2D + T2E;
		    T2T = T2D - T2E;
		    T29 = T25 * Tw;
		    T2a = T27 * Tz;
		    T2b = T29 - T2a;
		    T2R = T29 + T2a;
	       }
	       {
		    E T3h, T3i, T33, T34;
		    T3h = T1p * Tz;
		    T3i = T1t * Tw;
		    T3j = T3h + T3i;
		    T3x = T3h - T3i;
		    T33 = T1p * Tw;
		    T34 = T1t * Tz;
		    T35 = T33 - T34;
		    T3t = T33 + T34;
	       }
	  }
	  {
	       E T7, T36, T3k, TC, T1f, T2e, T2I, T1Q, Te, TJ, T1R, T18, T2L, T37, T2l;
	       E T3l, Tm, T1T, TT, T1h, T2A, T2N, T3b, T3n, Tt, T1U, T12, T1i, T2t, T2O;
	       E T3e, T3o;
	       {
		    E T3, T2c, T1e, T2d, T6, T2G, T1b, T2H;
		    {
			 E T1, T2, T1c, T1d;
			 T1 = cr[0];
			 T2 = ci[WS(rs, 7)];
			 T3 = T1 + T2;
			 T2c = T1 - T2;
			 T1c = ci[WS(rs, 11)];
			 T1d = cr[WS(rs, 12)];
			 T1e = T1c - T1d;
			 T2d = T1c + T1d;
		    }
		    {
			 E T4, T5, T19, T1a;
			 T4 = cr[WS(rs, 4)];
			 T5 = ci[WS(rs, 3)];
			 T6 = T4 + T5;
			 T2G = T4 - T5;
			 T19 = ci[WS(rs, 15)];
			 T1a = cr[WS(rs, 8)];
			 T1b = T19 - T1a;
			 T2H = T19 + T1a;
		    }
		    T7 = T3 + T6;
		    T36 = T2c + T2d;
		    T3k = T2H - T2G;
		    TC = T3 - T6;
		    T1f = T1b - T1e;
		    T2e = T2c - T2d;
		    T2I = T2G + T2H;
		    T1Q = T1b + T1e;
	       }
	       {
		    E Ta, T2f, TI, T2g, Td, T2i, TF, T2j;
		    {
			 E T8, T9, TG, TH;
			 T8 = cr[WS(rs, 2)];
			 T9 = ci[WS(rs, 5)];
			 Ta = T8 + T9;
			 T2f = T8 - T9;
			 TG = ci[WS(rs, 13)];
			 TH = cr[WS(rs, 10)];
			 TI = TG - TH;
			 T2g = TG + TH;
		    }
		    {
			 E Tb, Tc, TD, TE;
			 Tb = ci[WS(rs, 1)];
			 Tc = cr[WS(rs, 6)];
			 Td = Tb + Tc;
			 T2i = Tb - Tc;
			 TD = ci[WS(rs, 9)];
			 TE = cr[WS(rs, 14)];
			 TF = TD - TE;
			 T2j = TD + TE;
		    }
		    Te = Ta + Td;
		    TJ = TF - TI;
		    T1R = TI + TF;
		    T18 = Ta - Td;
		    {
			 E T2J, T2K, T2h, T2k;
			 T2J = T2f + T2g;
			 T2K = T2i + T2j;
			 T2L = KP707106781 * (T2J - T2K);
			 T37 = KP707106781 * (T2J + T2K);
			 T2h = T2f - T2g;
			 T2k = T2i - T2j;
			 T2l = KP707106781 * (T2h + T2k);
			 T3l = KP707106781 * (T2h - T2k);
		    }
	       }
	       {
		    E Ti, T2x, TR, T2y, Tl, T2u, TO, T2v, TL, TS;
		    {
			 E Tg, Th, TP, TQ;
			 Tg = cr[WS(rs, 1)];
			 Th = ci[WS(rs, 6)];
			 Ti = Tg + Th;
			 T2x = Tg - Th;
			 TP = ci[WS(rs, 10)];
			 TQ = cr[WS(rs, 13)];
			 TR = TP - TQ;
			 T2y = TP + TQ;
		    }
		    {
			 E Tj, Tk, TM, TN;
			 Tj = cr[WS(rs, 5)];
			 Tk = ci[WS(rs, 2)];
			 Tl = Tj + Tk;
			 T2u = Tj - Tk;
			 TM = ci[WS(rs, 14)];
			 TN = cr[WS(rs, 9)];
			 TO = TM - TN;
			 T2v = TM + TN;
		    }
		    Tm = Ti + Tl;
		    T1T = TO + TR;
		    TL = Ti - Tl;
		    TS = TO - TR;
		    TT = TL - TS;
		    T1h = TL + TS;
		    {
			 E T2w, T2z, T39, T3a;
			 T2w = T2u + T2v;
			 T2z = T2x - T2y;
			 T2A = FMA(KP923879532, T2w, KP382683432 * T2z);
			 T2N = FNMS(KP382683432, T2w, KP923879532 * T2z);
			 T39 = T2x + T2y;
			 T3a = T2v - T2u;
			 T3b = FNMS(KP923879532, T3a, KP382683432 * T39);
			 T3n = FMA(KP382683432, T3a, KP923879532 * T39);
		    }
	       }
	       {
		    E Tp, T2q, T10, T2r, Ts, T2n, TX, T2o, TU, T11;
		    {
			 E Tn, To, TY, TZ;
			 Tn = ci[0];
			 To = cr[WS(rs, 7)];
			 Tp = Tn + To;
			 T2q = Tn - To;
			 TY = ci[WS(rs, 12)];
			 TZ = cr[WS(rs, 11)];
			 T10 = TY - TZ;
			 T2r = TY + TZ;
		    }
		    {
			 E Tq, Tr, TV, TW;
			 Tq = cr[WS(rs, 3)];
			 Tr = ci[WS(rs, 4)];
			 Ts = Tq + Tr;
			 T2n = Tq - Tr;
			 TV = ci[WS(rs, 8)];
			 TW = cr[WS(rs, 15)];
			 TX = TV - TW;
			 T2o = TV + TW;
		    }
		    Tt = Tp + Ts;
		    T1U = TX + T10;
		    TU = Tp - Ts;
		    T11 = TX - T10;
		    T12 = TU + T11;
		    T1i = T11 - TU;
		    {
			 E T2p, T2s, T3c, T3d;
			 T2p = T2n - T2o;
			 T2s = T2q - T2r;
			 T2t = FNMS(KP382683432, T2s, KP923879532 * T2p);
			 T2O = FMA(KP382683432, T2p, KP923879532 * T2s);
			 T3c = T2q + T2r;
			 T3d = T2n + T2o;
			 T3e = FNMS(KP923879532, T3d, KP382683432 * T3c);
			 T3o = FMA(KP382683432, T3d, KP923879532 * T3c);
		    }
	       }
	       {
		    E Tf, Tu, T1O, T1S, T1V, T1W;
		    Tf = T7 + Te;
		    Tu = Tm + Tt;
		    T1O = Tf - Tu;
		    T1S = T1Q + T1R;
		    T1V = T1T + T1U;
		    T1W = T1S - T1V;
		    cr[0] = Tf + Tu;
		    ci[0] = T1S + T1V;
		    cr[WS(rs, 8)] = FNMS(T1P, T1W, T1N * T1O);
		    ci[WS(rs, 8)] = FMA(T1P, T1O, T1N * T1W);
	       }
	       {
		    E T3g, T3r, T3q, T3s;
		    {
			 E T38, T3f, T3m, T3p;
			 T38 = T36 - T37;
			 T3f = T3b + T3e;
			 T3g = T38 - T3f;
			 T3r = T38 + T3f;
			 T3m = T3k + T3l;
			 T3p = T3n - T3o;
			 T3q = T3m - T3p;
			 T3s = T3m + T3p;
		    }
		    cr[WS(rs, 11)] = FNMS(T3j, T3q, T35 * T3g);
		    ci[WS(rs, 11)] = FMA(T3j, T3g, T35 * T3q);
		    cr[WS(rs, 3)] = FNMS(T1n, T3s, T1l * T3r);
		    ci[WS(rs, 3)] = FMA(T1n, T3r, T1l * T3s);
	       }
	       {
		    E T3w, T3B, T3A, T3C;
		    {
			 E T3u, T3v, T3y, T3z;
			 T3u = T36 + T37;
			 T3v = T3n + T3o;
			 T3w = T3u - T3v;
			 T3B = T3u + T3v;
			 T3y = T3k - T3l;
			 T3z = T3b - T3e;
			 T3A = T3y + T3z;
			 T3C = T3y - T3z;
		    }
		    cr[WS(rs, 7)] = FNMS(T3x, T3A, T3t * T3w);
		    ci[WS(rs, 7)] = FMA(T3t, T3A, T3x * T3w);
		    cr[WS(rs, 15)] = FNMS(T1w, T3C, T1v * T3B);
		    ci[WS(rs, 15)] = FMA(T1v, T3C, T1w * T3B);
	       }
	       {
		    E T14, T1q, T1k, T1u;
		    {
			 E TK, T13, T1g, T1j;
			 TK = TC + TJ;
			 T13 = KP707106781 * (TT + T12);
			 T14 = TK - T13;
			 T1q = TK + T13;
			 T1g = T18 + T1f;
			 T1j = KP707106781 * (T1h + T1i);
			 T1k = T1g - T1j;
			 T1u = T1g + T1j;
		    }
		    cr[WS(rs, 10)] = FNMS(T17, T1k, TB * T14);
		    ci[WS(rs, 10)] = FMA(T17, T14, TB * T1k);
		    cr[WS(rs, 2)] = FNMS(T1t, T1u, T1p * T1q);
		    ci[WS(rs, 2)] = FMA(T1t, T1q, T1p * T1u);
	       }
	       {
		    E T1A, T1I, T1E, T1M;
		    {
			 E T1y, T1z, T1C, T1D;
			 T1y = TC - TJ;
			 T1z = KP707106781 * (T1i - T1h);
			 T1A = T1y - T1z;
			 T1I = T1y + T1z;
			 T1C = T1f - T18;
			 T1D = KP707106781 * (TT - T12);
			 T1E = T1C - T1D;
			 T1M = T1C + T1D;
		    }
		    cr[WS(rs, 14)] = FNMS(T1B, T1E, T1x * T1A);
		    ci[WS(rs, 14)] = FMA(T1x, T1E, T1B * T1A);
		    cr[WS(rs, 6)] = FNMS(T1L, T1M, T1H * T1I);
		    ci[WS(rs, 6)] = FMA(T1H, T1M, T1L * T1I);
	       }
	       {
		    E T2C, T2S, T2Q, T2U;
		    {
			 E T2m, T2B, T2M, T2P;
			 T2m = T2e - T2l;
			 T2B = T2t - T2A;
			 T2C = T2m - T2B;
			 T2S = T2m + T2B;
			 T2M = T2I - T2L;
			 T2P = T2N - T2O;
			 T2Q = T2M - T2P;
			 T2U = T2M + T2P;
		    }
		    cr[WS(rs, 13)] = FNMS(T2F, T2Q, T2b * T2C);
		    ci[WS(rs, 13)] = FMA(T2F, T2C, T2b * T2Q);
		    cr[WS(rs, 5)] = FNMS(T2T, T2U, T2R * T2S);
		    ci[WS(rs, 5)] = FMA(T2T, T2S, T2R * T2U);
	       }
	       {
		    E T2X, T31, T30, T32;
		    {
			 E T2V, T2W, T2Y, T2Z;
			 T2V = T2e + T2l;
			 T2W = T2N + T2O;
			 T2X = T2V - T2W;
			 T31 = T2V + T2W;
			 T2Y = T2I + T2L;
			 T2Z = T2A + T2t;
			 T30 = T2Y - T2Z;
			 T32 = T2Y + T2Z;
		    }
		    cr[WS(rs, 9)] = FNMS(Tz, T30, Tw * T2X);
		    ci[WS(rs, 9)] = FMA(Tw, T30, Tz * T2X);
		    cr[WS(rs, 1)] = FNMS(Ty, T32, Tv * T31);
		    ci[WS(rs, 1)] = FMA(Tv, T32, Ty * T31);
	       }
	       {
		    E T20, T26, T24, T28;
		    {
			 E T1Y, T1Z, T22, T23;
			 T1Y = T7 - Te;
			 T1Z = T1U - T1T;
			 T20 = T1Y - T1Z;
			 T26 = T1Y + T1Z;
			 T22 = T1Q - T1R;
			 T23 = Tm - Tt;
			 T24 = T22 - T23;
			 T28 = T23 + T22;
		    }
		    cr[WS(rs, 12)] = FNMS(T21, T24, T1X * T20);
		    ci[WS(rs, 12)] = FMA(T1X, T24, T21 * T20);
		    cr[WS(rs, 4)] = FNMS(T27, T28, T25 * T26);
		    ci[WS(rs, 4)] = FMA(T25, T28, T27 * T26);
	       }
	  }
     }
}

static const tw_instr twinstr[] = {
     {TW_CEXP, 1, 1},
     {TW_CEXP, 1, 3},
     {TW_CEXP, 1, 9},
     {TW_CEXP, 1, 15},
     {TW_NEXT, 1, 0}
};

static const hc2hc_desc desc = { 16, "hb2_16", twinstr, &GENUS, {156, 68, 40, 0} };

void X(codelet_hb2_16) (planner *p) {
     X(khc2hc_register) (p, hb2_16, &desc);
}
#endif				/* HAVE_FMA */