t2fv_64.c

快速fft变换

					V T5N, T5U, T68, T5z, T6U, T6f;
					T5N = VFMA(LDK(KP923879532), T5M, T5F);
					T6L = VFNMS(LDK(KP923879532), T5M, T5F);
					T6M = VFNMS(LDK(KP923879532), T5T, T5Q);
					T5U = VFMA(LDK(KP923879532), T5T, T5Q);
					T68 = VFMA(LDK(KP923879532), T67, T60);
					T6O = VFNMS(LDK(KP923879532), T67, T60);
					T5z = VADD(T5p, T5y);
					T6U = VSUB(T5y, T5p);
					T6P = VFNMS(LDK(KP923879532), T6e, T6b);
					T6f = VFMA(LDK(KP923879532), T6e, T6b);
					{
					     V T5V, T6u, T6g, T6v, T6s, T6J;
					     T6s = VSUB(T6q, T6r);
					     T6J = VADD(T6q, T6r);
					     T5V = VFNMS(LDK(KP098491403), T5U, T5N);
					     T6u = VFMA(LDK(KP098491403), T5N, T5U);
					     T75 = VFNMS(LDK(KP980785280), T6U, T6T);
					     T6V = VFMA(LDK(KP980785280), T6U, T6T);
					     T5A = VFMA(LDK(KP980785280), T5z, T5g);
					     T6A = VFNMS(LDK(KP980785280), T5z, T5g);
					     T6g = VFNMS(LDK(KP098491403), T6f, T68);
					     T6v = VFMA(LDK(KP098491403), T68, T6f);
					     T72 = VFNMS(LDK(KP980785280), T6J, T6I);
					     T6K = VFMA(LDK(KP980785280), T6J, T6I);
					     T6t = VFMA(LDK(KP980785280), T6s, T6p);
					     T6D = VFNMS(LDK(KP980785280), T6s, T6p);
					     T6w = VSUB(T6u, T6v);
					     T6B = VADD(T6u, T6v);
					     T6h = VADD(T5V, T6g);
					     T6E = VSUB(T6g, T5V);
					}
				   }
			      }
			 }
		    }
	       }
	  }
	  {
	       V T6W, T6N, T6G, T6C, T6z, T6x, T6H, T6F, T6y, T6i, T6X, T6Q;
	       T6W = VFNMS(LDK(KP820678790), T6L, T6M);
	       T6N = VFMA(LDK(KP820678790), T6M, T6L);
	       T6G = VFMA(LDK(KP995184726), T6B, T6A);
	       T6C = VFNMS(LDK(KP995184726), T6B, T6A);
	       T6z = VFMA(LDK(KP995184726), T6w, T6t);
	       T6x = VFNMS(LDK(KP995184726), T6w, T6t);
	       T6H = VFMA(LDK(KP995184726), T6E, T6D);
	       T6F = VFNMS(LDK(KP995184726), T6E, T6D);
	       T6y = VFMA(LDK(KP995184726), T6h, T5A);
	       T6i = VFNMS(LDK(KP995184726), T6h, T5A);
	       T6X = VFNMS(LDK(KP820678790), T6O, T6P);
	       T6Q = VFMA(LDK(KP820678790), T6P, T6O);
	       {
		    V T73, T6Y, T76, T6R;
		    ST(&(x[WS(rs, 49)]), VFNMSI(T6H, T6G), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 15)]), VFMAI(T6H, T6G), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 47)]), VFMAI(T6F, T6C), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 17)]), VFNMSI(T6F, T6C), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 63)]), VFMAI(T6z, T6y), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 1)]), VFNMSI(T6z, T6y), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 31)]), VFMAI(T6x, T6i), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 33)]), VFNMSI(T6x, T6i), ms, &(x[WS(rs, 1)]));
		    T73 = VADD(T6W, T6X);
		    T6Y = VSUB(T6W, T6X);
		    T76 = VSUB(T6Q, T6N);
		    T6R = VADD(T6N, T6Q);
		    {
			 V T78, T74, T71, T6Z, T79, T77, T70, T6S;
			 T78 = VFNMS(LDK(KP773010453), T73, T72);
			 T74 = VFMA(LDK(KP773010453), T73, T72);
			 T71 = VFMA(LDK(KP773010453), T6Y, T6V);
			 T6Z = VFNMS(LDK(KP773010453), T6Y, T6V);
			 T79 = VFNMS(LDK(KP773010453), T76, T75);
			 T77 = VFMA(LDK(KP773010453), T76, T75);
			 T70 = VFMA(LDK(KP773010453), T6R, T6K);
			 T6S = VFNMS(LDK(KP773010453), T6R, T6K);
			 ST(&(x[WS(rs, 55)]), VFMAI(T79, T78), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 9)]), VFNMSI(T79, T78), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 41)]), VFNMSI(T77, T74), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 23)]), VFMAI(T77, T74), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 7)]), VFMAI(T71, T70), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 57)]), VFNMSI(T71, T70), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 39)]), VFMAI(T6Z, T6S), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 25)]), VFNMSI(T6Z, T6S), ms, &(x[WS(rs, 1)]));
		    }
	       }
	  }
     }
}

static const tw_instr twinstr[] = {
     VTW(0, 1),
     VTW(0, 2),
     VTW(0, 3),
     VTW(0, 4),
     VTW(0, 5),
     VTW(0, 6),
     VTW(0, 7),
     VTW(0, 8),
     VTW(0, 9),
     VTW(0, 10),
     VTW(0, 11),
     VTW(0, 12),
     VTW(0, 13),
     VTW(0, 14),
     VTW(0, 15),
     VTW(0, 16),
     VTW(0, 17),
     VTW(0, 18),
     VTW(0, 19),
     VTW(0, 20),
     VTW(0, 21),
     VTW(0, 22),
     VTW(0, 23),
     VTW(0, 24),
     VTW(0, 25),
     VTW(0, 26),
     VTW(0, 27),
     VTW(0, 28),
     VTW(0, 29),
     VTW(0, 30),
     VTW(0, 31),
     VTW(0, 32),
     VTW(0, 33),
     VTW(0, 34),
     VTW(0, 35),
     VTW(0, 36),
     VTW(0, 37),
     VTW(0, 38),
     VTW(0, 39),
     VTW(0, 40),
     VTW(0, 41),
     VTW(0, 42),
     VTW(0, 43),
     VTW(0, 44),
     VTW(0, 45),
     VTW(0, 46),
     VTW(0, 47),
     VTW(0, 48),
     VTW(0, 49),
     VTW(0, 50),
     VTW(0, 51),
     VTW(0, 52),
     VTW(0, 53),
     VTW(0, 54),
     VTW(0, 55),
     VTW(0, 56),
     VTW(0, 57),
     VTW(0, 58),
     VTW(0, 59),
     VTW(0, 60),
     VTW(0, 61),
     VTW(0, 62),
     VTW(0, 63),
     {TW_NEXT, VL, 0}
};

static const ct_desc desc = { 64, "t2fv_64", twinstr, &GENUS, {261, 126, 258, 0}, 0, 0, 0 };

void X(codelet_t2fv_64) (planner *p) {
     X(kdft_dit_register) (p, t2fv_64, &desc);
}
#else				/* HAVE_FMA */

/* Generated by: ../../../genfft/gen_twiddle_c -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name t2fv_64 -include t2f.h */

/*
 * This function contains 519 FP additions, 250 FP multiplications,
 * (or, 467 additions, 198 multiplications, 52 fused multiply/add),
 * 107 stack variables, 15 constants, and 128 memory accesses
 */
#include "t2f.h"

static void t2fv_64(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
{
     DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
     DVK(KP098017140, +0.098017140329560601994195563888641845861136673);
     DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
     DVK(KP634393284, +0.634393284163645498215171613225493370675687095);
     DVK(KP471396736, +0.471396736825997648556387625905254377657460319);
     DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
     DVK(KP290284677, +0.290284677254462367636192375817395274691476278);
     DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
     DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
     DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
     DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
     DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
     DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
     DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
     DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
     INT m;
     R *x;
     x = ri;
     for (m = mb, W = W + (mb * ((TWVL / VL) * 126)); m < me; m = m + VL, x = x + (VL * ms), W = W + (TWVL * 126), MAKE_VOLATILE_STRIDE(rs)) {
	  V Tg, T4a, T6r, T7f, T3o, T4B, T5q, T7e, T5R, T62, T28, T4o, T2g, T4l, T7n;
	  V T7Z, T68, T6j, T2C, T4s, T3a, T4v, T7u, T82, T7E, T7F, T7V, T5F, T6u, T1k;
	  V T4e, T1r, T4d, T7B, T7C, T7W, T5M, T6v, TV, T4g, T12, T4h, T7h, T7i, TD;
	  V T4C, T3h, T4b, T5x, T6s, T1R, T4m, T7q, T80, T2j, T4p, T5Y, T63, T2Z, T4w;
	  V T7x, T83, T33, T4t, T6f, T6k;
	  {
	       V T1, T3, T3m, T3k, Tb, Td, Te, T6, T8, T9, T2, T3l, T3j;
	       T1 = LD(&(x[0]), ms, &(x[0]));
	       T2 = LD(&(x[WS(rs, 32)]), ms, &(x[0]));
	       T3 = BYTWJ(&(W[TWVL * 62]), T2);
	       T3l = LD(&(x[WS(rs, 48)]), ms, &(x[0]));
	       T3m = BYTWJ(&(W[TWVL * 94]), T3l);
	       T3j = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
	       T3k = BYTWJ(&(W[TWVL * 30]), T3j);
	       {
		    V Ta, Tc, T5, T7;
		    Ta = LD(&(x[WS(rs, 56)]), ms, &(x[0]));
		    Tb = BYTWJ(&(W[TWVL * 110]), Ta);
		    Tc = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
		    Td = BYTWJ(&(W[TWVL * 46]), Tc);
		    Te = VSUB(Tb, Td);
		    T5 = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
		    T6 = BYTWJ(&(W[TWVL * 14]), T5);
		    T7 = LD(&(x[WS(rs, 40)]), ms, &(x[0]));
		    T8 = BYTWJ(&(W[TWVL * 78]), T7);
		    T9 = VSUB(T6, T8);
	       }
	       {
		    V T4, Tf, T6p, T6q;
		    T4 = VSUB(T1, T3);
		    Tf = VMUL(LDK(KP707106781), VADD(T9, Te));
		    Tg = VADD(T4, Tf);
		    T4a = VSUB(T4, Tf);
		    T6p = VADD(Tb, Td);
		    T6q = VADD(T6, T8);
		    T6r = VSUB(T6p, T6q);
		    T7f = VADD(T6q, T6p);
	       }
	       {
		    V T3i, T3n, T5o, T5p;
		    T3i = VMUL(LDK(KP707106781), VSUB(Te, T9));
		    T3n = VSUB(T3k, T3m);
		    T3o = VSUB(T3i, T3n);
		    T4B = VADD(T3n, T3i);
		    T5o = VADD(T1, T3);
		    T5p = VADD(T3k, T3m);
		    T5q = VSUB(T5o, T5p);
		    T7e = VADD(T5o, T5p);
	       }
	  }
	  {
	       V T24, T26, T5Q, T2b, T2d, T5P, T1W, T60, T21, T61, T22, T27;
	       {
		    V T23, T25, T2a, T2c;
		    T23 = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
		    T24 = BYTWJ(&(W[TWVL * 32]), T23);
		    T25 = LD(&(x[WS(rs, 49)]), ms, &(x[WS(rs, 1)]));
		    T26 = BYTWJ(&(W[TWVL * 96]), T25);
		    T5Q = VADD(T24, T26);
		    T2a = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
		    T2b = BYTWJ(&(W[0]), T2a);
		    T2c = LD(&(x[WS(rs, 33)]), ms, &(x[WS(rs, 1)]));
		    T2d = BYTWJ(&(W[TWVL * 64]), T2c);
		    T5P = VADD(T2b, T2d);
	       }
	       {
		    V T1T, T1V, T1S, T1U;
		    T1S = LD(&(x[WS(rs, 57)]), ms, &(x[WS(rs, 1)]));
		    T1T = BYTWJ(&(W[TWVL * 112]), T1S);
		    T1U = LD(&(x[WS(rs, 25)]), ms, &(x[WS(rs, 1)]));
		    T1V = BYTWJ(&(W[TWVL * 48]), T1U);
		    T1W = VSUB(T1T, T1V);
		    T60 = VADD(T1T, T1V);
	       }
	       {
		    V T1Y, T20, T1X, T1Z;
		    T1X = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
		    T1Y = BYTWJ(&(W[TWVL * 16]), T1X);
		    T1Z = LD(&(x[WS(rs, 41)]), ms, &(x[WS(rs, 1)]));
		    T20 = BYTWJ(&(W[TWVL * 80]), T1Z);
		    T21 = VSUB(T1Y, T20);
		    T61 = VADD(T1Y, T20);
	       }
	       T5R = VSUB(T5P, T5Q);
	       T62 = VSUB(T60, T61);
	       T22 = VMUL(LDK(KP707106781), VSUB(T1W, T21));
	       T27 = VSUB(T24, T26);
	       T28 = VSUB(T22, T27);
	       T4o = VADD(T27, T22);
	       {
		    V T2e, T2f, T7l, T7m;
		    T2e = VSUB(T2b, T2d);
		    T2f = VMUL(LDK(KP707106781), VADD(T21, T1W));
		    T2g = VADD(T2e, T2f);
		    T4l = VSUB(T2e, T2f);
		    T7l = VADD(T5P, T5Q);
		    T7m = VADD(T61, T60);
		    T7n = VADD(T7l, T7m);
		    T7Z = VSUB(T7l, T7m);
	       }
	  }
	  {
	       V T2n, T2p, T66, T36, T38, T67, T2v, T6i, T2A, T6h, T2q, T2B;
	       {
		    V T2m, T2o, T35, T37;
		    T2m = LD(&(x[WS(rs, 63)]), ms, &(x[WS(rs, 1)]));
		    T2n = BYTWJ(&(W[TWVL * 124]), T2m);
		    T2o = LD(&(x[WS(rs, 31)]), ms, &(x[WS(rs, 1)]));
		    T2p = BYTWJ(&(W[TWVL * 60]), T2o);
		    T66 = VADD(T2n, T2p);
		    T35 = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)]));
		    T36 = BYTWJ(&(W[TWVL * 28]), T35);
		    T37 = LD(&(x[WS(rs, 47)]), ms, &(x[WS(rs, 1)]));
		    T38 = BYTWJ(&(W[TWVL * 92]), T37);
		    T67 = VADD(T36, T38);
	       }
	       {
		    V T2s, T2u, T2r, T2t;
		    T2r = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
		    T2s = BYTWJ(&(W[TWVL * 12]), T2r);
		    T2t = LD(&(x[WS(rs, 39)]), ms, &(x[WS(rs, 1)]));
		    T2u = BYTWJ(&(W[TWVL * 76]), T2t);
		    T2v = VSUB(T2s, T2u);
		    T6i = VADD(T2s, T2u);
	       }
	       {
		    V T2x, T2z, T2w, T2y;
		    T2w = LD(&(x[WS(rs, 55)]), ms, &(x[WS(rs, 1)]));
		    T2x = BYTWJ(&(W[TWVL * 108]), T2w);
		    T2y = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)]));
		    T2z = BYTWJ(&(W[TWVL * 44]), T2y);
		    T2A = VSUB(T2x, T2z);
		    T6h = VADD(T2x, T2z);
	       }
	       T68 = VSUB(T66, T67);
	       T6j = VSUB(T6h, T6i);
	       T2q = VSUB(T2n, T2p);
	       T2B = VMUL(LDK(KP707106781), VADD(T2v, T2A));
	       T2C = VADD(T2q, T2B);
	       T4s = VSUB(T2q, T2B);
	       {
		    V T34, T39, T7s, T7t;
		    T34 = VMUL(LDK(KP707106781), VSUB(T2A, T2v));
		    T39 = VSUB(T36, T38);
		    T3a = VSUB(T34, T39);
		    T4v = VADD(T39, T34);
		    T7s = VADD(T66, T67);
		    T7t = VADD(T6i, T6h);
		    T7u = VADD(T7s, T7t);
		    T82 = VSUB(T7s, T7t);
	       }
	  }
	  {
	       V T1g, T1i, T5A, T1m, T1o, T5z, T18, T5C, T1d, T5D, T5B, T5E;
	       {
		    V T1f, T1h, T1l, T1n;
		    T1f = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
		    T1g = BYTWJ(&(W[TWVL * 34]), T1f);
		    T1h = LD(&(x[WS(rs, 50)]), ms, &(x[0]));
		    T1i = BYTWJ(&(W[TWVL * 98]), T1h);
		    T5A = VADD(T1g, T1i);
		    T1l = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
		    T1m = BYTWJ(&(W[TWVL * 2]), T1l);
		    T1n = LD(&(x[WS(rs, 34)]), ms, &(x[0]));
		    T1o = BYTWJ(&(W[TWVL * 66]), T1n);
		    T5z = VADD(T1m, T1o);
	       }
	       {
		    V T15, T17, T14, T16;
		    T14 = LD(&(x[WS(rs, 58)]), ms, &(x[0]));
		    T15 = BYTWJ(&(W[TWVL * 114]), T14);
		    T16 = LD(&(x[WS(rs, 26)]), ms, &(x[0]));
		    T17 = BYTWJ(&(W[TWVL * 50]), T16);
		    T18 = VSUB(T15, T17);
		    T5C = VADD(T15, T17);
	       }
	       {
		    V T1a, T1c, T19, T1b;
		    T19 = LD(&(x[WS(rs, 10)]), ms, &(x[0]));
		    T1a = BYTWJ(&(W[TWVL * 18]), T19);
		    T1b = LD(&(x[WS(rs, 42)]), ms, &(x[0]));