// compute log in base a of x, where a is 1, 2 or ten depending on Tag
static inline A0 loga(const A0& a0)
{
if (a0 == Inf<A0>()) return a0;
if (iseqz(a0)) return Minf<A0>();
if (nt2::isnan(a0)||isltz(a0)) return Nan<A0>();
A0 x, fe, x2, y;
approx_t::kernel_log(a0, fe, x, x2, y);
return finalize_t::finalize(fe, x, x2, y);
}
inline float log(const float& a0)
{
typedef float A0;
if (a0 == Inf<A0>()) return a0;
if (iseqz(a0)) return Minf<A0>();
if (nt2::is_nan(a0)||is_ltz(a0)) return Nan<A0>();
float x, fe, x2, y;
kernel_log(a0, fe, x, x2, y);
y = fma(fe, Const<float, 0xb95e8083>(), y);
y = fma(Mhalf<A0>(), x2, y);
A0 z = x + y;
return fma(Const<float, 0x3f318000>(), fe, z);
}
inline float log10(const float& a0)
{
typedef float A0;
if (a0 == Inf<A0>()) return a0;
if (iseqz(a0)) return Minf<A0>();
if (nt2::is_nan(a0)||isltz(a0)) return Nan<A0>();
A0 x, fe, x2, y;
kernel_log(a0, fe, x, x2, y);
y = amul(y, -Half<A0>(), x2);
// multiply log of fraction by log10(e) and base 2 exponent by log10(2)
A0 z = mul(x+y, Const<float, 0x3a37b152>());//7.00731903251827651129E-4f // log10(e)lo
z = amul(z, y, Const<float, 0x3ede0000>()); //4.3359375E-1f // log10(e)hi
z = amul(z, x, Const<float, 0x3ede0000>());
z = amul(z, fe, Const<float, 0x39826a14>());//3.0078125E-1f // log10(2)hi
return amul(z, fe, Const<float, 0x3e9a0000 >());//2.48745663981195213739E-4f // log10(2)lo
}
inline float log2(const float& a0)
{
typedef float A0;
if (a0 == Inf<A0>()) return a0;
if (iseqz(a0)) return Minf<A0>();
if (nt2::is_nan(a0)||is_ltz(a0)) return Nan<A0>();
A0 x, fe, x2, y;
kernel_log(a0, fe, x, x2, y);
y = fma(Mhalf<A0>(),x2, y);
// multiply log of fraction by log2(e)
A0 z = fma( x
, Const<float, 0x3ee2a8ed>()
, mul(y,Const<float, 0x3ee2a8ed>())// 0.44269504088896340735992
);
return ((z+y)+x)+fe;
}
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_>))
);