t1bv_64.c

快速fft变换

		    T6o = VADD(T5O, T6n);
		    T6B = VBYI(VADD(T6x, T6A));
		    ST(&(x[WS(rs, 54)]), VSUB(T6o, T6B), ms, &(x[0]));
		    ST(&(x[WS(rs, 10)]), VADD(T6o, T6B), ms, &(x[0]));
		    T6K = VBYI(VADD(T6F, T6E));
		    T6L = VADD(T6H, T6I);
		    ST(&(x[WS(rs, 6)]), VADD(T6K, T6L), ms, &(x[0]));
		    ST(&(x[WS(rs, 58)]), VSUB(T6L, T6K), ms, &(x[0]));
	       }
	       {
		    V T6C, T6D, T6G, T6J;
		    T6C = VSUB(T5O, T6n);
		    T6D = VBYI(VSUB(T6A, T6x));
		    ST(&(x[WS(rs, 42)]), VSUB(T6C, T6D), ms, &(x[0]));
		    ST(&(x[WS(rs, 22)]), VADD(T6C, T6D), ms, &(x[0]));
		    T6G = VBYI(VSUB(T6E, T6F));
		    T6J = VSUB(T6H, T6I);
		    ST(&(x[WS(rs, 26)]), VADD(T6G, T6J), ms, &(x[0]));
		    ST(&(x[WS(rs, 38)]), VSUB(T6J, T6G), ms, &(x[0]));
	       }
	  }
	  {
	       V T6O, T79, T6Z, T77, T6V, T7a, T72, T76;
	       {
		    V T6M, T6N, T6X, T6Y;
		    T6M = VADD(T5q, T5x);
		    T6N = VADD(T6p, T6q);
		    T6O = VSUB(T6M, T6N);
		    T79 = VADD(T6M, T6N);
		    T6X = VADD(T5F, T5M);
		    T6Y = VADD(T6v, T6s);
		    T6Z = VSUB(T6X, T6Y);
		    T77 = VADD(T6Y, T6X);
		    {
			 V T6R, T70, T6U, T71;
			 {
			      V T6P, T6Q, T6S, T6T;
			      T6P = VADD(T5Y, T5V);
			      T6Q = VADD(T62, T63);
			      T6R = VFMA(LDK(KP980785280), T6P, VMUL(LDK(KP195090322), T6Q));
			      T70 = VFNMS(LDK(KP195090322), T6P, VMUL(LDK(KP980785280), T6Q));
			      T6S = VADD(T6f, T6c);
			      T6T = VADD(T6j, T6k);
			      T6U = VFNMS(LDK(KP195090322), T6T, VMUL(LDK(KP980785280), T6S));
			      T71 = VFMA(LDK(KP195090322), T6S, VMUL(LDK(KP980785280), T6T));
			 }
			 T6V = VSUB(T6R, T6U);
			 T7a = VADD(T70, T71);
			 T72 = VSUB(T70, T71);
			 T76 = VADD(T6R, T6U);
		    }
	       }
	       {
		    V T6W, T73, T7c, T7d;
		    T6W = VADD(T6O, T6V);
		    T73 = VBYI(VADD(T6Z, T72));
		    ST(&(x[WS(rs, 50)]), VSUB(T6W, T73), ms, &(x[0]));
		    ST(&(x[WS(rs, 14)]), VADD(T6W, T73), ms, &(x[0]));
		    T7c = VBYI(VADD(T77, T76));
		    T7d = VADD(T79, T7a);
		    ST(&(x[WS(rs, 2)]), VADD(T7c, T7d), ms, &(x[0]));
		    ST(&(x[WS(rs, 62)]), VSUB(T7d, T7c), ms, &(x[0]));
	       }
	       {
		    V T74, T75, T78, T7b;
		    T74 = VSUB(T6O, T6V);
		    T75 = VBYI(VSUB(T72, T6Z));
		    ST(&(x[WS(rs, 46)]), VSUB(T74, T75), ms, &(x[0]));
		    ST(&(x[WS(rs, 18)]), VADD(T74, T75), ms, &(x[0]));
		    T78 = VBYI(VSUB(T76, T77));
		    T7b = VSUB(T79, T7a);
		    ST(&(x[WS(rs, 30)]), VADD(T78, T7b), ms, &(x[0]));
		    ST(&(x[WS(rs, 34)]), VSUB(T7b, T78), ms, &(x[0]));
	       }
	  }
	  {
	       V T4z, T5g, T4R, T59, T4H, T5j, T4O, T55, T4o, T4S, T4K, T4P, T52, T5k, T5c;
	       V T5h;
	       {
		    V T4y, T57, T4v, T58, T4r, T4u;
		    T4y = VADD(T4w, T4x);
		    T57 = VSUB(T4B, T4C);
		    T4r = VFMA(LDK(KP980785280), T4p, VMUL(LDK(KP195090322), T4q));
		    T4u = VFNMS(LDK(KP195090322), T4t, VMUL(LDK(KP980785280), T4s));
		    T4v = VADD(T4r, T4u);
		    T58 = VSUB(T4r, T4u);
		    T4z = VSUB(T4v, T4y);
		    T5g = VADD(T57, T58);
		    T4R = VADD(T4y, T4v);
		    T59 = VSUB(T57, T58);
	       }
	       {
		    V T4D, T54, T4G, T53, T4E, T4F;
		    T4D = VADD(T4B, T4C);
		    T54 = VSUB(T4x, T4w);
		    T4E = VFNMS(LDK(KP195090322), T4p, VMUL(LDK(KP980785280), T4q));
		    T4F = VFMA(LDK(KP195090322), T4s, VMUL(LDK(KP980785280), T4t));
		    T4G = VADD(T4E, T4F);
		    T53 = VSUB(T4E, T4F);
		    T4H = VSUB(T4D, T4G);
		    T5j = VADD(T54, T53);
		    T4O = VADD(T4D, T4G);
		    T55 = VSUB(T53, T54);
	       }
	       {
		    V T4g, T4I, T4n, T4J;
		    {
			 V T4c, T4f, T4j, T4m;
			 T4c = VADD(T4a, T4b);
			 T4f = VADD(T4d, T4e);
			 T4g = VFNMS(LDK(KP098017140), T4f, VMUL(LDK(KP995184726), T4c));
			 T4I = VFMA(LDK(KP098017140), T4c, VMUL(LDK(KP995184726), T4f));
			 T4j = VADD(T4h, T4i);
			 T4m = VADD(T4k, T4l);
			 T4n = VFMA(LDK(KP995184726), T4j, VMUL(LDK(KP098017140), T4m));
			 T4J = VFNMS(LDK(KP098017140), T4j, VMUL(LDK(KP995184726), T4m));
		    }
		    T4o = VSUB(T4g, T4n);
		    T4S = VADD(T4I, T4J);
		    T4K = VSUB(T4I, T4J);
		    T4P = VADD(T4g, T4n);
	       }
	       {
		    V T4Y, T5a, T51, T5b;
		    {
			 V T4W, T4X, T4Z, T50;
			 T4W = VSUB(T4a, T4b);
			 T4X = VSUB(T4e, T4d);
			 T4Y = VFNMS(LDK(KP634393284), T4X, VMUL(LDK(KP773010453), T4W));
			 T5a = VFMA(LDK(KP634393284), T4W, VMUL(LDK(KP773010453), T4X));
			 T4Z = VSUB(T4h, T4i);
			 T50 = VSUB(T4l, T4k);
			 T51 = VFMA(LDK(KP773010453), T4Z, VMUL(LDK(KP634393284), T50));
			 T5b = VFNMS(LDK(KP634393284), T4Z, VMUL(LDK(KP773010453), T50));
		    }
		    T52 = VSUB(T4Y, T51);
		    T5k = VADD(T5a, T5b);
		    T5c = VSUB(T5a, T5b);
		    T5h = VADD(T4Y, T51);
	       }
	       {
		    V T4A, T4L, T5i, T5l;
		    T4A = VBYI(VSUB(T4o, T4z));
		    T4L = VSUB(T4H, T4K);
		    ST(&(x[WS(rs, 17)]), VADD(T4A, T4L), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 47)]), VSUB(T4L, T4A), ms, &(x[WS(rs, 1)]));
		    T5i = VADD(T5g, T5h);
		    T5l = VBYI(VADD(T5j, T5k));
		    ST(&(x[WS(rs, 57)]), VSUB(T5i, T5l), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 7)]), VADD(T5i, T5l), ms, &(x[WS(rs, 1)]));
	       }
	       {
		    V T5m, T5n, T4M, T4N;
		    T5m = VSUB(T5g, T5h);
		    T5n = VBYI(VSUB(T5k, T5j));
		    ST(&(x[WS(rs, 39)]), VSUB(T5m, T5n), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 25)]), VADD(T5m, T5n), ms, &(x[WS(rs, 1)]));
		    T4M = VBYI(VADD(T4z, T4o));
		    T4N = VADD(T4H, T4K);
		    ST(&(x[WS(rs, 15)]), VADD(T4M, T4N), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 49)]), VSUB(T4N, T4M), ms, &(x[WS(rs, 1)]));
	       }
	       {
		    V T4Q, T4T, T56, T5d;
		    T4Q = VADD(T4O, T4P);
		    T4T = VBYI(VADD(T4R, T4S));
		    ST(&(x[WS(rs, 63)]), VSUB(T4Q, T4T), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 1)]), VADD(T4Q, T4T), ms, &(x[WS(rs, 1)]));
		    T56 = VBYI(VSUB(T52, T55));
		    T5d = VSUB(T59, T5c);
		    ST(&(x[WS(rs, 23)]), VADD(T56, T5d), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 41)]), VSUB(T5d, T56), ms, &(x[WS(rs, 1)]));
	       }
	       {
		    V T5e, T5f, T4U, T4V;
		    T5e = VBYI(VADD(T55, T52));
		    T5f = VADD(T59, T5c);
		    ST(&(x[WS(rs, 9)]), VADD(T5e, T5f), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 55)]), VSUB(T5f, T5e), ms, &(x[WS(rs, 1)]));
		    T4U = VSUB(T4O, T4P);
		    T4V = VBYI(VSUB(T4S, T4R));
		    ST(&(x[WS(rs, 33)]), VSUB(T4U, T4V), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 31)]), VADD(T4U, T4V), ms, &(x[WS(rs, 1)]));
	       }
	  }
	  {
	       V T1u, T43, T3D, T3V, T3t, T45, T3B, T3K, T3d, T3E, T3w, T3A, T3R, T46, T3Y;
	       V T42;
	       {
		    V TE, T3U, T1t, T3T, T13, T1s;
		    TE = VSUB(Tg, TD);
		    T3U = VADD(T3r, T3k);
		    T13 = VFMA(LDK(KP831469612), TV, VMUL(LDK(KP555570233), T12));
		    T1s = VFNMS(LDK(KP555570233), T1r, VMUL(LDK(KP831469612), T1k));
		    T1t = VSUB(T13, T1s);
		    T3T = VADD(T13, T1s);
		    T1u = VSUB(TE, T1t);
		    T43 = VADD(T3U, T3T);
		    T3D = VADD(TE, T1t);
		    T3V = VSUB(T3T, T3U);
	       }
	       {
		    V T3s, T3I, T3h, T3J, T3f, T3g;
		    T3s = VSUB(T3k, T3r);
		    T3I = VADD(Tg, TD);
		    T3f = VFNMS(LDK(KP555570233), TV, VMUL(LDK(KP831469612), T12));
		    T3g = VFMA(LDK(KP555570233), T1k, VMUL(LDK(KP831469612), T1r));
		    T3h = VSUB(T3f, T3g);
		    T3J = VADD(T3f, T3g);
		    T3t = VSUB(T3h, T3s);
		    T45 = VADD(T3I, T3J);
		    T3B = VADD(T3s, T3h);
		    T3K = VSUB(T3I, T3J);
	       }
	       {
		    V T2l, T3u, T3c, T3v;
		    {
			 V T29, T2k, T30, T3b;
			 T29 = VSUB(T1R, T28);
			 T2k = VSUB(T2g, T2j);
			 T2l = VFMA(LDK(KP881921264), T29, VMUL(LDK(KP471396736), T2k));
			 T3u = VFNMS(LDK(KP471396736), T29, VMUL(LDK(KP881921264), T2k));
			 T30 = VSUB(T2I, T2Z);
			 T3b = VSUB(T37, T3a);
			 T3c = VFNMS(LDK(KP471396736), T3b, VMUL(LDK(KP881921264), T30));
			 T3v = VFMA(LDK(KP471396736), T30, VMUL(LDK(KP881921264), T3b));
		    }
		    T3d = VSUB(T2l, T3c);
		    T3E = VADD(T3u, T3v);
		    T3w = VSUB(T3u, T3v);
		    T3A = VADD(T2l, T3c);
	       }
	       {
		    V T3N, T3W, T3Q, T3X;
		    {
			 V T3L, T3M, T3O, T3P;
			 T3L = VADD(T28, T1R);
			 T3M = VADD(T2g, T2j);
			 T3N = VFMA(LDK(KP956940335), T3L, VMUL(LDK(KP290284677), T3M));
			 T3W = VFNMS(LDK(KP290284677), T3L, VMUL(LDK(KP956940335), T3M));
			 T3O = VADD(T2Z, T2I);
			 T3P = VADD(T37, T3a);
			 T3Q = VFNMS(LDK(KP290284677), T3P, VMUL(LDK(KP956940335), T3O));
			 T3X = VFMA(LDK(KP290284677), T3O, VMUL(LDK(KP956940335), T3P));
		    }
		    T3R = VSUB(T3N, T3Q);
		    T46 = VADD(T3W, T3X);
		    T3Y = VSUB(T3W, T3X);
		    T42 = VADD(T3N, T3Q);
	       }
	       {
		    V T3e, T3x, T44, T47;
		    T3e = VADD(T1u, T3d);
		    T3x = VBYI(VADD(T3t, T3w));
		    ST(&(x[WS(rs, 53)]), VSUB(T3e, T3x), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 11)]), VADD(T3e, T3x), ms, &(x[WS(rs, 1)]));
		    T44 = VBYI(VSUB(T42, T43));
		    T47 = VSUB(T45, T46);
		    ST(&(x[WS(rs, 29)]), VADD(T44, T47), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 35)]), VSUB(T47, T44), ms, &(x[WS(rs, 1)]));
	       }
	       {
		    V T48, T49, T3y, T3z;
		    T48 = VBYI(VADD(T43, T42));
		    T49 = VADD(T45, T46);
		    ST(&(x[WS(rs, 3)]), VADD(T48, T49), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 61)]), VSUB(T49, T48), ms, &(x[WS(rs, 1)]));
		    T3y = VSUB(T1u, T3d);
		    T3z = VBYI(VSUB(T3w, T3t));
		    ST(&(x[WS(rs, 43)]), VSUB(T3y, T3z), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 21)]), VADD(T3y, T3z), ms, &(x[WS(rs, 1)]));
	       }
	       {
		    V T3C, T3F, T3S, T3Z;
		    T3C = VBYI(VSUB(T3A, T3B));
		    T3F = VSUB(T3D, T3E);
		    ST(&(x[WS(rs, 27)]), VADD(T3C, T3F), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 37)]), VSUB(T3F, T3C), ms, &(x[WS(rs, 1)]));
		    T3S = VADD(T3K, T3R);
		    T3Z = VBYI(VADD(T3V, T3Y));
		    ST(&(x[WS(rs, 51)]), VSUB(T3S, T3Z), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 13)]), VADD(T3S, T3Z), ms, &(x[WS(rs, 1)]));
	       }
	       {
		    V T40, T41, T3G, T3H;
		    T40 = VSUB(T3K, T3R);
		    T41 = VBYI(VSUB(T3Y, T3V));
		    ST(&(x[WS(rs, 45)]), VSUB(T40, T41), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 19)]), VADD(T40, T41), ms, &(x[WS(rs, 1)]));
		    T3G = VBYI(VADD(T3B, T3A));
		    T3H = VADD(T3D, T3E);
		    ST(&(x[WS(rs, 5)]), VADD(T3G, T3H), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 59)]), VSUB(T3H, T3G), 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, "t1bv_64", twinstr, &GENUS, {467, 198, 52, 0}, 0, 0, 0 };

void X(codelet_t1bv_64) (planner *p) {
     X(kdft_dit_register) (p, t1bv_64, &desc);
}
#endif				/* HAVE_FMA */