n2fv_64.c

快速傅立叶变换库函数

			      }
			      {
				   V T2H, T2P, T81, T84, T85, T87, T2J, T2B, T2R, T2N;
				   {
					V T2L, T2t, T2M, T2A;
					{
					     V T2z, T2F, T6Q, T6G, T2G, T2w;
					     T2z = VFMA(LDK(KP668178637), T2y, T2x);
					     T2F = VFNMS(LDK(KP668178637), T2x, T2y);
					     T6Q = VFMA(LDK(KP831469612), T6F, T6C);
					     T6G = VFNMS(LDK(KP831469612), T6F, T6C);
					     T2G = VFNMS(LDK(KP668178637), T2u, T2v);
					     T2w = VFMA(LDK(KP668178637), T2v, T2u);
					     T2L = VFNMS(LDK(KP923879532), T2s, T2r);
					     T2t = VFMA(LDK(KP923879532), T2s, T2r);
					     {
						  V T6S, T6U, T6M, T6K;
						  T6S = VFNMS(LDK(KP881921264), T6R, T6Q);
						  T6U = VFMA(LDK(KP881921264), T6R, T6Q);
						  T6M = VFMA(LDK(KP881921264), T6J, T6G);
						  T6K = VFNMS(LDK(KP881921264), T6J, T6G);
						  T2M = VADD(T2G, T2F);
						  T2H = VSUB(T2F, T2G);
						  T2P = VSUB(T2z, T2w);
						  T2A = VADD(T2w, T2z);
						  T81 = VFMAI(T6S, T6P);
						  STM2(&(xo[86]), T81, ovs, &(xo[2]));
						  {
						       V T82, T83, T86, T88;
						       T82 = VFNMSI(T6S, T6P);
						       STM2(&(xo[42]), T82, ovs, &(xo[2]));
						       STN2(&(xo[40]), T7w, T82, ovs);
						       T83 = VFNMSI(T6U, T6T);
						       STM2(&(xo[106]), T83, ovs, &(xo[2]));
						       STN2(&(xo[104]), T7x, T83, ovs);
						       T84 = VFMAI(T6U, T6T);
						       STM2(&(xo[22]), T84, ovs, &(xo[2]));
						       T85 = VFMAI(T6M, T6L);
						       STM2(&(xo[118]), T85, ovs, &(xo[2]));
						       T86 = VFNMSI(T6M, T6L);
						       STM2(&(xo[10]), T86, ovs, &(xo[2]));
						       STN2(&(xo[8]), T7z, T86, ovs);
						       T87 = VFMAI(T6K, T6z);
						       STM2(&(xo[54]), T87, ovs, &(xo[2]));
						       T88 = VFNMSI(T6K, T6z);
						       STM2(&(xo[74]), T88, ovs, &(xo[2]));
						       STN2(&(xo[72]), T7B, T88, ovs);
						  }
					     }
					}
					T2J = VFMA(LDK(KP831469612), T2A, T2t);
					T2B = VFNMS(LDK(KP831469612), T2A, T2t);
					T2R = VFNMS(LDK(KP831469612), T2M, T2L);
					T2N = VFMA(LDK(KP831469612), T2M, T2L);
				   }
				   {
					V T61, T5J, T62, T5Q;
					{
					     V T5M, T5V, T2O, T2E, T5W, T5P;
					     T5M = VFMA(LDK(KP820678790), T5L, T5K);
					     T5V = VFNMS(LDK(KP820678790), T5K, T5L);
					     T2O = VFMA(LDK(KP923879532), T2D, T2C);
					     T2E = VFNMS(LDK(KP923879532), T2D, T2C);
					     T5W = VFNMS(LDK(KP820678790), T5N, T5O);
					     T5P = VFMA(LDK(KP820678790), T5O, T5N);
					     T61 = VFNMS(LDK(KP980785280), T5I, T5H);
					     T5J = VFMA(LDK(KP980785280), T5I, T5H);
					     {
						  V T2Q, T2S, T2K, T2I;
						  T2Q = VFNMS(LDK(KP831469612), T2P, T2O);
						  T2S = VFMA(LDK(KP831469612), T2P, T2O);
						  T2K = VFMA(LDK(KP831469612), T2H, T2E);
						  T2I = VFNMS(LDK(KP831469612), T2H, T2E);
						  T62 = VADD(T5V, T5W);
						  T5X = VSUB(T5V, T5W);
						  T65 = VSUB(T5P, T5M);
						  T5Q = VADD(T5M, T5P);
						  {
						       V T89, T8c, T8d, T8f;
						       T89 = VFMAI(T2Q, T2N);
						       STM2(&(xo[84]), T89, ovs, &(xo[0]));
						       STN2(&(xo[84]), T89, T81, ovs);
						       T8a = VFNMSI(T2Q, T2N);
						       STM2(&(xo[44]), T8a, ovs, &(xo[0]));
						       T8b = VFNMSI(T2S, T2R);
						       STM2(&(xo[108]), T8b, ovs, &(xo[0]));
						       T8c = VFMAI(T2S, T2R);
						       STM2(&(xo[20]), T8c, ovs, &(xo[0]));
						       STN2(&(xo[20]), T8c, T84, ovs);
						       T8d = VFMAI(T2K, T2J);
						       STM2(&(xo[116]), T8d, ovs, &(xo[0]));
						       STN2(&(xo[116]), T8d, T85, ovs);
						       T8e = VFNMSI(T2K, T2J);
						       STM2(&(xo[12]), T8e, ovs, &(xo[0]));
						       T8f = VFMAI(T2I, T2B);
						       STM2(&(xo[52]), T8f, ovs, &(xo[0]));
						       STN2(&(xo[52]), T8f, T87, ovs);
						       T8g = VFNMSI(T2I, T2B);
						       STM2(&(xo[76]), T8g, ovs, &(xo[0]));
						  }
					     }
					}
					T5Z = VFMA(LDK(KP773010453), T5Q, T5J);
					T5R = VFNMS(LDK(KP773010453), T5Q, T5J);
					T67 = VFNMS(LDK(KP773010453), T62, T61);
					T63 = VFMA(LDK(KP773010453), T62, T61);
				   }
			      }
			 }
		    }
	       }
	  }
	  T5U = VFMA(LDK(KP980785280), T5T, T5S);
	  T64 = VFNMS(LDK(KP980785280), T5T, T5S);
	  {
	       V T68, T66, T5Y, T60;
	       T68 = VFNMS(LDK(KP773010453), T65, T64);
	       T66 = VFMA(LDK(KP773010453), T65, T64);
	       T5Y = VFNMS(LDK(KP773010453), T5X, T5U);
	       T60 = VFMA(LDK(KP773010453), T5X, T5U);
	       {
		    V T8h, T8i, T8j, T8k;
		    T8h = VFNMSI(T66, T63);
		    STM2(&(xo[82]), T8h, ovs, &(xo[2]));
		    STN2(&(xo[80]), T7t, T8h, ovs);
		    T8i = VFMAI(T66, T63);
		    STM2(&(xo[46]), T8i, ovs, &(xo[2]));
		    STN2(&(xo[44]), T8a, T8i, ovs);
		    T8j = VFMAI(T68, T67);
		    STM2(&(xo[110]), T8j, ovs, &(xo[2]));
		    STN2(&(xo[108]), T8b, T8j, ovs);
		    T8k = VFNMSI(T68, T67);
		    STM2(&(xo[18]), T8k, ovs, &(xo[2]));
		    STN2(&(xo[16]), T7r, T8k, ovs);
		    {
			 V T8l, T8m, T8n, T8o;
			 T8l = VFMAI(T60, T5Z);
			 STM2(&(xo[14]), T8l, ovs, &(xo[2]));
			 STN2(&(xo[12]), T8e, T8l, ovs);
			 T8m = VFNMSI(T60, T5Z);
			 STM2(&(xo[114]), T8m, ovs, &(xo[2]));
			 STN2(&(xo[112]), T7s, T8m, ovs);
			 T8n = VFMAI(T5Y, T5R);
			 STM2(&(xo[78]), T8n, ovs, &(xo[2]));
			 STN2(&(xo[76]), T8g, T8n, ovs);
			 T8o = VFNMSI(T5Y, T5R);
			 STM2(&(xo[50]), T8o, ovs, &(xo[2]));
			 STN2(&(xo[48]), T7u, T8o, ovs);
		    }
	       }
	  }
     }
}

static const kdft_desc desc = { 64, "n2fv_64", {198, 0, 258, 0}, &GENUS, 0, 2, 0, 0 };

void X(codelet_n2fv_64) (planner *p) {
     X(kdft_register) (p, n2fv_64, &desc);
}

#else				/* HAVE_FMA */

/* Generated by: ../../../genfft/gen_notw_c -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name n2fv_64 -with-ostride 2 -include n2f.h -store-multiple 2 */

/*
 * This function contains 456 FP additions, 124 FP multiplications,
 * (or, 404 additions, 72 multiplications, 52 fused multiply/add),
 * 128 stack variables, and 160 memory accesses
 */
/*
 * Generator Id's : 
 * $Id: algsimp.ml,v 1.9 2006-02-12 23:34:12 athena Exp $
 * $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
 * $Id: gen_notw_c.ml,v 1.17 2006-02-12 23:34:12 athena Exp $
 */

#include "n2f.h"

static void n2fv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
{
     DVK(KP471396736, +0.471396736825997648556387625905254377657460319);
     DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
     DVK(KP098017140, +0.098017140329560601994195563888641845861136673);
     DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
     DVK(KP290284677, +0.290284677254462367636192375817395274691476278);
     DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
     DVK(KP634393284, +0.634393284163645498215171613225493370675687095);
     DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
     DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
     DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
     DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
     DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
     DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
     DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
     DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
     INT i;
     const R *xi;
     R *xo;
     xi = ri;
     xo = ro;
     for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
	  V T4p, T5q, Tb, T39, T2n, T3A, T6f, T6T, Tq, T3B, T6i, T76, T2i, T3a, T4w;
	  V T5r, TI, T2p, T6C, T6V, T3h, T3E, T4L, T5u, TZ, T2q, T6F, T6U, T3e, T3D;
	  V T4E, T5t, T23, T2N, T6t, T71, T6w, T72, T2c, T2O, T3t, T41, T5f, T5R, T5k;
	  V T5S, T3w, T42, T1s, T2K, T6m, T6Y, T6p, T6Z, T1B, T2L, T3m, T3Y, T4Y, T5O;
	  V T53, T5P, T3p, T3Z;
	  {
	       V T3, T4n, T2m, T4o, T6, T5p, T9, T5o;
	       {
		    V T1, T2, T2k, T2l;
		    T1 = LD(&(xi[0]), ivs, &(xi[0]));
		    T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
		    T3 = VSUB(T1, T2);
		    T4n = VADD(T1, T2);
		    T2k = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
		    T2l = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
		    T2m = VSUB(T2k, T2l);
		    T4o = VADD(T2k, T2l);
	       }
	       {
		    V T4, T5, T7, T8;
		    T4 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
		    T5 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
		    T6 = VSUB(T4, T5);
		    T5p = VADD(T4, T5);
		    T7 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
		    T8 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
		    T9 = VSUB(T7, T8);
		    T5o = VADD(T7, T8);
	       }
	       T4p = VSUB(T4n, T4o);
	       T5q = VSUB(T5o, T5p);
	       {
		    V Ta, T2j, T6d, T6e;
		    Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
		    Tb = VADD(T3, Ta);
		    T39 = VSUB(T3, Ta);
		    T2j = VMUL(LDK(KP707106781), VSUB(T9, T6));
		    T2n = VSUB(T2j, T2m);
		    T3A = VADD(T2m, T2j);
		    T6d = VADD(T4n, T4o);
		    T6e = VADD(T5p, T5o);
		    T6f = VADD(T6d, T6e);
		    T6T = VSUB(T6d, T6e);
	       }
	  }
	  {
	       V Te, T4q, To, T4u, Th, T4r, Tl, T4t;
	       {
		    V Tc, Td, Tm, Tn;
		    Tc = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
		    Td = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
		    Te = VSUB(Tc, Td);
		    T4q = VADD(Tc, Td);
		    Tm = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
		    Tn = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
		    To = VSUB(Tm, Tn);
		    T4u = VADD(Tm, Tn);
	       }
	       {
		    V Tf, Tg, Tj, Tk;
		    Tf = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
		    Tg = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
		    Th = VSUB(Tf, Tg);
		    T4r = VADD(Tf, Tg);
		    Tj = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
		    Tk = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
		    Tl = VSUB(Tj, Tk);
		    T4t = VADD(Tj, Tk);
	       }
	       {
		    V Ti, Tp, T6g, T6h;
		    Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
		    Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To));
		    Tq = VADD(Ti, Tp);
		    T3B = VSUB(Tp, Ti);
		    T6g = VADD(T4q, T4r);
		    T6h = VADD(T4t, T4u);
		    T6i = VADD(T6g, T6h);
		    T76 = VSUB(T6h, T6g);
	       }
	       {
		    V T2g, T2h, T4s, T4v;
		    T2g = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
		    T2h = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
		    T2i = VSUB(T2g, T2h);
		    T3a = VADD(T2h, T2g);
		    T4s = VSUB(T4q, T4r);
		    T4v = VSUB(T4t, T4u);
		    T4w = VMUL(LDK(KP707106781), VADD(T4s, T4v));
		    T5r = VMUL(LDK(KP707106781), VSUB(T4v, T4s));
	       }
	  }
	  {
	       V Tu, T4F, TG, T4G, TB, T4J, TD, T4I;
	       {
		    V Ts, Tt, TE, TF;
		    Ts = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
		    Tt = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
		    Tu = VSUB(Ts, Tt);
		    T4F = VADD(Ts, Tt);
		    TE = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
		    TF = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
		    TG = VSUB(TE, TF);
		    T4G = VADD(TE, TF);
		    {
			 V Tv, Tw, Tx, Ty, Tz, TA;
			 Tv = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
			 Tw = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
			 Tx = VSUB(Tv, Tw);
			 Ty = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
			 Tz = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
			 TA = VSUB(Ty, Tz);
			 TB = VMUL(LDK(KP707106781), VADD(Tx, TA));
			 T4J = VADD(Tv, Tw);
			 TD = VMUL(LDK(KP707106781), VSUB(TA, Tx));
			 T4I = VADD(Ty, Tz);
		    }
	       }
	       {
		    V TC, TH, T6A, T6B;
		    TC = VADD(Tu, TB);
		    TH = VSUB(TD, TG);
		    TI = VFMA(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH));
		    T2p = VFNMS(LDK(KP195090322), TH, VMUL(LDK(KP980785280), TC));
		    T6A = VADD(T4F, T4G);
		    T6B = VADD(T4J, T4I);
		    T6C = VADD(T6A, T6B);
		    T6V = VSUB(T6A, T6B);
	       }
	       {
		    V T3f, T3g, T4H, T4K;
		    T3f = VSUB(Tu, TB);
		    T3g = VADD(TG, TD);
		    T3h = VFNMS(LDK(KP555570233), T3g, VMUL(LDK(KP831469612), T3f));
		    T3E = VFMA(LDK(KP555570233), T3f, VMUL(LDK(KP831469612), T3g));
		    T4H = VSUB(T4F, T4G);
		    T4K = VSUB(T4I, T4J);
		    T4L = VFNMS(LDK(KP382683432), T4K, VMUL(LDK(KP923879532), T4H));
		    T5u = VFMA(LDK(KP382683432), T4H, VMUL(LDK(KP923879532), T4K));
	       }
	  }
	  {
	       V TS, T4z, TW, T4y, TP, T4C, TX, T4B;
	       {
		    V TQ, TR, TU, TV;
		    TQ = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
		    TR = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
		    TS = VSUB(TQ, TR);
		    T4z = VADD(TQ, TR);
		    TU = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
		    TV = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
		    TW = VSUB(TU, TV);
		    T4y = VADD(TU, TV);
		    {
			 V TJ, TK, TL, TM, TN, TO;
			 TJ = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
			 TK = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
			 TL = VSUB(TJ, TK);
			 TM = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
			 TN = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
			 TO = VSUB(TM, TN);
			 TP = VMUL(LDK(KP707106781), VSUB(TL, TO));
			 T4C = VADD(TM, TN);
			 TX = VMUL(LDK(KP707106781), VADD(TO, TL));