hf2_32.c

快速fft变换

					     T6X = T6V - T6W;
					     T9s = T6W + T6V;
					}
					T7a = FNMS(KP923879532, T71, T6Y);
					T72 = FMA(KP923879532, T71, T6Y);
					T9t = FNMS(KP923879532, T9q, T9p);
					T9r = FMA(KP923879532, T9q, T9p);
					ci[WS(rs, 6)] = FMA(KP980785280, T6X, T6U);
					cr[WS(rs, 9)] = FNMS(KP980785280, T6X, T6U);
					ci[WS(rs, 2)] = FMA(KP831469612, T79, T72);
					cr[WS(rs, 13)] = FNMS(KP831469612, T79, T72);
					ci[WS(rs, 30)] = FMA(KP980785280, T9s, T9r);
					cr[WS(rs, 17)] = FMS(KP980785280, T9s, T9r);
					ci[WS(rs, 22)] = FMA(KP980785280, T9u, T9t);
					cr[WS(rs, 25)] = FMS(KP980785280, T9u, T9t);
				   }
			      }
			 }
		    }
	       }
	  }
	  cr[WS(rs, 5)] = FMA(KP831469612, T7d, T7a);
	  ci[WS(rs, 10)] = FNMS(KP831469612, T7d, T7a);
     }
}

static const tw_instr twinstr[] = {
     {TW_CEXP, 1, 1},
     {TW_CEXP, 1, 3},
     {TW_CEXP, 1, 9},
     {TW_CEXP, 1, 27},
     {TW_NEXT, 1, 0}
};

static const hc2hc_desc desc = { 32, "hf2_32", twinstr, &GENUS, {236, 98, 252, 0} };

void X(codelet_hf2_32) (planner *p) {
     X(khc2hc_register) (p, hf2_32, &desc);
}
#else				/* HAVE_FMA */

/* Generated by: ../../../genfft/gen_hc2hc -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 32 -dit -name hf2_32 -include hf.h */

/*
 * This function contains 488 FP additions, 280 FP multiplications,
 * (or, 376 additions, 168 multiplications, 112 fused multiply/add),
 * 158 stack variables, 7 constants, and 128 memory accesses
 */
#include "hf.h"

static void hf2_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
{
     DK(KP555570233, +0.555570233019602224742830813948532874374937191);
     DK(KP831469612, +0.831469612302545237078788377617905756738560812);
     DK(KP980785280, +0.980785280403230449126182236134239036973933731);
     DK(KP195090322, +0.195090322016128267848284868477022240927691618);
     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 T2, T5, T3, T6, T8, TM, TO, Td, T9, Te, Th, Tl, TD, TH, T1y;
	  E T1H, T15, T1A, T11, T1F, T1n, T1p, T2q, T2I, T2u, T2K, T2V, T3b, T2Z, T3d;
	  E Tu, Ty, T3l, T3n, T1t, T1v, T2f, T2h, T1a, T1e, T32, T34, T1W, T1Y, T2C;
	  E T2E, Tg, TR, Tk, TS, Tm, TV, To, TT, T1M, T21, T1P, T22, T1Q, T25;
	  E T1S, T23;
	  {
	       E Ts, T1d, Tx, T18, Tt, T1c, Tw, T19, TB, T14, TG, TZ, TC, T13, TF;
	       E T10;
	       {
		    E T4, Tc, T7, Tb;
		    T2 = W[0];
		    T5 = W[1];
		    T3 = W[2];
		    T6 = W[3];
		    T4 = T2 * T3;
		    Tc = T5 * T3;
		    T7 = T5 * T6;
		    Tb = T2 * T6;
		    T8 = T4 + T7;
		    TM = T4 - T7;
		    TO = Tb + Tc;
		    Td = Tb - Tc;
		    T9 = W[4];
		    Ts = T2 * T9;
		    T1d = T6 * T9;
		    Tx = T5 * T9;
		    T18 = T3 * T9;
		    Te = W[5];
		    Tt = T5 * Te;
		    T1c = T3 * Te;
		    Tw = T2 * Te;
		    T19 = T6 * Te;
		    Th = W[6];
		    TB = T3 * Th;
		    T14 = T5 * Th;
		    TG = T6 * Th;
		    TZ = T2 * Th;
		    Tl = W[7];
		    TC = T6 * Tl;
		    T13 = T2 * Tl;
		    TF = T3 * Tl;
		    T10 = T5 * Tl;
	       }
	       TD = TB + TC;
	       TH = TF - TG;
	       T1y = TZ + T10;
	       T1H = TF + TG;
	       T15 = T13 + T14;
	       T1A = T13 - T14;
	       T11 = TZ - T10;
	       T1F = TB - TC;
	       T1n = FMA(T9, Th, Te * Tl);
	       T1p = FNMS(Te, Th, T9 * Tl);
	       {
		    E T2o, T2p, T2s, T2t;
		    T2o = T8 * Th;
		    T2p = Td * Tl;
		    T2q = T2o + T2p;
		    T2I = T2o - T2p;
		    T2s = T8 * Tl;
		    T2t = Td * Th;
		    T2u = T2s - T2t;
		    T2K = T2s + T2t;
	       }
	       {
		    E T2T, T2U, T2X, T2Y;
		    T2T = TM * Th;
		    T2U = TO * Tl;
		    T2V = T2T - T2U;
		    T3b = T2T + T2U;
		    T2X = TM * Tl;
		    T2Y = TO * Th;
		    T2Z = T2X + T2Y;
		    T3d = T2X - T2Y;
		    Tu = Ts + Tt;
		    Ty = Tw - Tx;
		    T3l = FMA(Tu, Th, Ty * Tl);
		    T3n = FNMS(Ty, Th, Tu * Tl);
	       }
	       T1t = Ts - Tt;
	       T1v = Tw + Tx;
	       T2f = FMA(T1t, Th, T1v * Tl);
	       T2h = FNMS(T1v, Th, T1t * Tl);
	       T1a = T18 - T19;
	       T1e = T1c + T1d;
	       T32 = FMA(T1a, Th, T1e * Tl);
	       T34 = FNMS(T1e, Th, T1a * Tl);
	       T1W = T18 + T19;
	       T1Y = T1c - T1d;
	       T2C = FMA(T1W, Th, T1Y * Tl);
	       T2E = FNMS(T1Y, Th, T1W * Tl);
	       {
		    E Ta, Tf, Ti, Tj;
		    Ta = T8 * T9;
		    Tf = Td * Te;
		    Tg = Ta - Tf;
		    TR = Ta + Tf;
		    Ti = T8 * Te;
		    Tj = Td * T9;
		    Tk = Ti + Tj;
		    TS = Ti - Tj;
	       }
	       Tm = FMA(Tg, Th, Tk * Tl);
	       TV = FNMS(TS, Th, TR * Tl);
	       To = FNMS(Tk, Th, Tg * Tl);
	       TT = FMA(TR, Th, TS * Tl);
	       {
		    E T1K, T1L, T1N, T1O;
		    T1K = TM * T9;
		    T1L = TO * Te;
		    T1M = T1K - T1L;
		    T21 = T1K + T1L;
		    T1N = TM * Te;
		    T1O = TO * T9;
		    T1P = T1N + T1O;
		    T22 = T1N - T1O;
	       }
	       T1Q = FMA(T1M, Th, T1P * Tl);
	       T25 = FNMS(T22, Th, T21 * Tl);
	       T1S = FNMS(T1P, Th, T1M * Tl);
	       T23 = FMA(T21, Th, T22 * Tl);
	  }
	  {
	       E TL, T6f, T8c, T8q, T3F, T5t, T7I, T7W, T2y, T6B, T6y, T7j, T4k, T5G, T4B;
	       E T5J, T3h, T6H, T6O, T7o, T4L, T5Q, T52, T5N, T1i, T7V, T6i, T7D, T3K, T5u;
	       E T3P, T5v, T1E, T6k, T6n, T7f, T3W, T5z, T41, T5y, T29, T6p, T6s, T7e, T47;
	       E T5C, T4c, T5B, T2R, T6z, T6E, T7k, T4v, T5K, T4E, T5H, T3y, T6P, T6K, T7p;
	       E T4W, T5O, T55, T5R;
	       {
		    E T1, T7G, Tq, T7F, TA, T3C, TJ, T3D, Tn, Tp;
		    T1 = cr[0];
		    T7G = ci[0];
		    Tn = cr[WS(rs, 16)];
		    Tp = ci[WS(rs, 16)];
		    Tq = FMA(Tm, Tn, To * Tp);
		    T7F = FNMS(To, Tn, Tm * Tp);
		    {
			 E Tv, Tz, TE, TI;
			 Tv = cr[WS(rs, 8)];
			 Tz = ci[WS(rs, 8)];
			 TA = FMA(Tu, Tv, Ty * Tz);
			 T3C = FNMS(Ty, Tv, Tu * Tz);
			 TE = cr[WS(rs, 24)];
			 TI = ci[WS(rs, 24)];
			 TJ = FMA(TD, TE, TH * TI);
			 T3D = FNMS(TH, TE, TD * TI);
		    }
		    {
			 E Tr, TK, T8a, T8b;
			 Tr = T1 + Tq;
			 TK = TA + TJ;
			 TL = Tr + TK;
			 T6f = Tr - TK;
			 T8a = TA - TJ;
			 T8b = T7G - T7F;
			 T8c = T8a + T8b;
			 T8q = T8b - T8a;
		    }
		    {
			 E T3B, T3E, T7E, T7H;
			 T3B = T1 - Tq;
			 T3E = T3C - T3D;
			 T3F = T3B + T3E;
			 T5t = T3B - T3E;
			 T7E = T3C + T3D;
			 T7H = T7F + T7G;
			 T7I = T7E + T7H;
			 T7W = T7H - T7E;
		    }
	       }
	       {
		    E T2e, T4x, T2w, T4i, T2j, T4y, T2n, T4h;
		    {
			 E T2c, T2d, T2r, T2v;
			 T2c = cr[WS(rs, 1)];
			 T2d = ci[WS(rs, 1)];
			 T2e = FMA(T2, T2c, T5 * T2d);
			 T4x = FNMS(T5, T2c, T2 * T2d);
			 T2r = cr[WS(rs, 25)];
			 T2v = ci[WS(rs, 25)];
			 T2w = FMA(T2q, T2r, T2u * T2v);
			 T4i = FNMS(T2u, T2r, T2q * T2v);
		    }
		    {
			 E T2g, T2i, T2l, T2m;
			 T2g = cr[WS(rs, 17)];
			 T2i = ci[WS(rs, 17)];
			 T2j = FMA(T2f, T2g, T2h * T2i);
			 T4y = FNMS(T2h, T2g, T2f * T2i);
			 T2l = cr[WS(rs, 9)];
			 T2m = ci[WS(rs, 9)];
			 T2n = FMA(T9, T2l, Te * T2m);
			 T4h = FNMS(Te, T2l, T9 * T2m);
		    }
		    {
			 E T2k, T2x, T6w, T6x;
			 T2k = T2e + T2j;
			 T2x = T2n + T2w;
			 T2y = T2k + T2x;
			 T6B = T2k - T2x;
			 T6w = T4x + T4y;
			 T6x = T4h + T4i;
			 T6y = T6w - T6x;
			 T7j = T6w + T6x;
		    }
		    {
			 E T4g, T4j, T4z, T4A;
			 T4g = T2e - T2j;
			 T4j = T4h - T4i;
			 T4k = T4g + T4j;
			 T5G = T4g - T4j;
			 T4z = T4x - T4y;
			 T4A = T2n - T2w;
			 T4B = T4z - T4A;
			 T5J = T4z + T4A;
		    }
	       }
	       {
		    E T31, T4H, T3f, T50, T36, T4I, T3a, T4Z;
		    {
			 E T2W, T30, T3c, T3e;
			 T2W = cr[WS(rs, 31)];
			 T30 = ci[WS(rs, 31)];
			 T31 = FMA(T2V, T2W, T2Z * T30);
			 T4H = FNMS(T2Z, T2W, T2V * T30);
			 T3c = cr[WS(rs, 23)];
			 T3e = ci[WS(rs, 23)];
			 T3f = FMA(T3b, T3c, T3d * T3e);
			 T50 = FNMS(T3d, T3c, T3b * T3e);
		    }
		    {
			 E T33, T35, T38, T39;
			 T33 = cr[WS(rs, 15)];
			 T35 = ci[WS(rs, 15)];
			 T36 = FMA(T32, T33, T34 * T35);
			 T4I = FNMS(T34, T33, T32 * T35);
			 T38 = cr[WS(rs, 7)];
			 T39 = ci[WS(rs, 7)];
			 T3a = FMA(TR, T38, TS * T39);
			 T4Z = FNMS(TS, T38, TR * T39);
		    }
		    {
			 E T37, T3g, T6M, T6N;
			 T37 = T31 + T36;
			 T3g = T3a + T3f;
			 T3h = T37 + T3g;
			 T6H = T37 - T3g;
			 T6M = T4H + T4I;
			 T6N = T4Z + T50;
			 T6O = T6M - T6N;
			 T7o = T6M + T6N;
		    }
		    {
			 E T4J, T4K, T4Y, T51;
			 T4J = T4H - T4I;
			 T4K = T3a - T3f;
			 T4L = T4J - T4K;
			 T5Q = T4J + T4K;
			 T4Y = T31 - T36;
			 T51 = T4Z - T50;
			 T52 = T4Y + T51;
			 T5N = T4Y - T51;
		    }
	       }
	       {
		    E TQ, T3H, T1g, T3N, TX, T3I, T17, T3M;
		    {
			 E TN, TP, T1b, T1f;
			 TN = cr[WS(rs, 4)];
			 TP = ci[WS(rs, 4)];
			 TQ = FMA(TM, TN, TO * TP);
			 T3H = FNMS(TO, TN, TM * TP);
			 T1b = cr[WS(rs, 12)];
			 T1f = ci[WS(rs, 12)];
			 T1g = FMA(T1a, T1b, T1e * T1f);
			 T3N = FNMS(T1e, T1b, T1a * T1f);
		    }
		    {
			 E TU, TW, T12, T16;
			 TU = cr[WS(rs, 20)];
			 TW = ci[WS(rs, 20)];
			 TX = FMA(TT, TU, TV * TW);
			 T3I = FNMS(TV, TU, TT * TW);
			 T12 = cr[WS(rs, 28)];
			 T16 = ci[WS(rs, 28)];
			 T17 = FMA(T11, T12, T15 * T16);
			 T3M = FNMS(T15, T12, T11 * T16);
		    }
		    {
			 E TY, T1h, T6g, T6h;
			 TY = TQ + TX;
			 T1h = T17 + T1g;
			 T1i = TY + T1h;
			 T7V = TY - T1h;
			 T6g = T3M + T3N;
			 T6h = T3H + T3I;
			 T6i = T6g - T6h;
			 T7D = T6h + T6g;
		    }
		    {
			 E T3G, T3J, T3L, T3O;
			 T3G = TQ - TX;
			 T3J = T3H - T3I;
			 T3K = T3G + T3J;
			 T5u = T3G - T3J;
			 T3L = T17 - T1g;
			 T3O = T3M - T3N;
			 T3P = T3L - T3O;
			 T5v = T3L + T3O;
		    }
	       }
	       {
		    E T1m, T3X, T1C, T3U, T1r, T3Y, T1x, T3T;
		    {
			 E T1k, T1l, T1z, T1B;
			 T1k = cr[WS(rs, 2)];
			 T1l = ci[WS(rs, 2)];
			 T1m = FMA(T8, T1k, Td * T1l);
			 T3X = FNMS(Td, T1k, T8 * T1l);
			 T1z = cr[WS(rs, 26)];
			 T1B = ci[WS(rs, 26)];
			 T1C = FMA(T1y, T1z, T1A * T1B);
			 T3U = FNMS(T1A, T1z, T1y * T1B);
		    }
		    {
			 E T1o, T1q, T1u, T1w;
			 T1o = cr[WS(rs, 18)];
			 T1q = ci[WS(rs, 18)];
			 T1r = FMA(T1n, T1o, T1p * T1q);
			 T3Y = FNMS(T1p, T1o, T1n * T1q);
			 T1u = cr[WS(rs, 10)];
			 T1w = ci[WS(rs, 10)];
			 T1x = FMA(T1t, T1u, T1v * T1w);
			 T3T = FNMS(T1v, T1u, T1t * T1w);
		    }
		    {
			 E T1s, T1D, T6l, T6m;
			 T1s = T1m + T1r;
			 T1D = T1x + T1C;
			 T1E = T1s + T1D;
			 T6k = T1s - T1D;
			 T6l = T3X + T3Y;
			 T6m = T3T + T3U;
			 T6n = T6l - T6m;
			 T7f = T6l + T6m;
		    }
		    {
			 E T3S, T3V, T3Z, T40;
			 T3S = T1m - T1r;
			 T3V = T3T - T3U;
			 T3W = T3S + T3V;
			 T5z = T3S - T3V;
			 T3Z = T3X - T3Y;
			 T40 = T1x - T1C;
			 T41 = T3Z - T40;
			 T5y = T3Z + T40;
		    }
	       }
	       {
		    E T1J, T43, T27, T4a, T1U, T44, T20, T49;
		    {
			 E T1G, T1I, T24, T26;
			 T1G = cr[WS(rs, 30)];
			 T1I = ci[WS(rs, 30)];
			 T1J = FMA(T1F, T1G, T1H * T1I);
			 T43 = FNMS(T1H, T1G, T1F * T1I);
			 T24 = cr[WS(rs, 22)];
			 T26 = ci[WS(rs, 22)];
			 T27 = FMA(T23, T24, T25 * T26);
			 T4a = FNMS(T25, T24, T23 * T26);
		    }
		    {
			 E T1R, T1T, T1X, T1Z;
			 T1R = cr[WS(rs, 14)];
			 T1T = ci[WS(rs, 14)];
			 T1U = FMA(T1Q, T1R, T1S * T1T);
			 T44 = FNMS(T1S, T1R, T1Q * T1T);
			 T1X = cr[WS(rs, 6)];
			 T1Z = ci[WS(rs, 6)];
			 T20 = FMA(T1W, T1X, T1Y * T1Z);
			 T49 = FNMS(T1Y, T1X, T1W * T1Z);
		    }
		    {
			 E T1V, T28, T6q, T6r;
			 T1V = T1J + T1U;
			 T28 = T20 + T27;
			 T29 = T1V + T28;
			 T6p = T1V - T28;
			 T6q = T43 + T44;
			 T6r = T49 + T4a;
			 T6s = T6q - T6r;
			 T7e = T6q + T6r;
		    }
		    {
			 E T45, T46, T48, T4b;
			 T45 = T43 - T44;
			 T46 = T20 - T27;
			 T47 = T45 - T46;
			 T5C = T45 + T46;