static inline void kernel_log(const A0& a0,
A0& dk,
A0& hfsq,
A0& s,
A0& R,
A0& f)
{
A0 x;
int_type k(fast_frexp(a0, x));
const int_type x_lt_sqrthf = -isgt(Sqrt_2o_2<A0>(), x);
k += x_lt_sqrthf;
f = minusone(x+b_and(x, genmask<A0>(x_lt_sqrthf)));
dk = tofloat(k);
s = f/add(Two<A0>(),f);
A0 z = sqr(s);
A0 w = sqr(z);
A0 t1= w*horner<NT2_HORNER_COEFF_T(A0, 3,
(0x3fc39a09d078c69fll,
0x3fcc71c51d8e78afll,
0x3fd999999997fa04ll)
)> (w);
A0 t2= z*horner<NT2_HORNER_COEFF_T(A0, 4,
(0x3fc2f112df3e5244ll,
0x3fc7466496cb03dell,
0x3fd2492494229359ll,
0x3fe5555555555593ll)
)> (w);
R = t2+t1;
hfsq = mul(Half<A0>(), sqr(f));
}
tag::cpu_, Dummy> : callable
{
template<class Sig> struct result;
template<class This,class A0>
struct result<This(A0)> : meta::strip<A0>{};//
NT2_FUNCTOR_CALL(1)
{
A0 const na = isnez(a0);
A0 n = add(shri(a0, 4), Four<A0>());
A0 n1 = shri(n+a0/n, 1);
A0 msk = b_and(isle(n1,n), na);
n = select(msk,n1,n);
n1 = sqr(n);
msk = b_or(isgt(n1,a0), b_and(iseqz(n1), na));
n = seladd( msk, n, Mone<A0>());
return seladd(na, Zero<A0>(), n);
}
};
} }
/////////////////////////////////////////////////////////////////////////////
// Implementation when type A0 is arithmetic_
/////////////////////////////////////////////////////////////////////////////
NT2_REGISTER_DISPATCH(tag::sqrt_, tag::cpu_,
(A0),
((simd_<arithmetic_<A0>,tag::xop_>))
);
