fix pow for uint32_t, disable pmul<Packet4ul>

This commit is contained in:
Charles Schlosser 2023-04-21 05:47:56 +00:00
parent 7f06bcae2c
commit 29c8e3c754
2 changed files with 29 additions and 4 deletions

View File

@ -276,7 +276,7 @@ template<> struct packet_traits<uint64_t> : default_packet_traits
HasTranspose = 0,
HasNegate = 0,
HasSqrt = 0,
HasMul = 0,
HasCmp = 1,
HasShift = 1
};

View File

@ -2152,15 +2152,15 @@ static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet handle_nonint_nonint_errors(
return result;
}
template <typename Packet, typename ScalarExponent>
template <typename Packet, typename ScalarExponent, std::enable_if_t<NumTraits<typename unpacket_traits<Packet>::type>::IsSigned, bool> = true>
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet handle_int_int(const Packet& x, const ScalarExponent& exponent) {
typedef typename unpacket_traits<Packet>::type Scalar;
// integer base, integer exponent case
// signed integer base, integer exponent case
// This routine handles negative and very large positive exponents
// Signed integer overflow and divide by zero is undefined behavior
// Unsigned intgers do not overflow
// Unsigned integers do not overflow
const bool exponent_is_odd = unary_pow::is_odd<ScalarExponent>::run(exponent);
@ -2181,6 +2181,31 @@ static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet handle_int_int(const Packet&
result = pselect(pand(pow_is_one, pow_is_neg), pnegate(cst_pos_one), result);
return result;
}
template <typename Packet, typename ScalarExponent, std::enable_if_t<!NumTraits<typename unpacket_traits<Packet>::type>::IsSigned, bool> = true>
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet handle_int_int(const Packet& x, const ScalarExponent& exponent) {
typedef typename unpacket_traits<Packet>::type Scalar;
// unsigned integer base, integer exponent case
// This routine handles negative and very large positive exponents
// Signed integer overflow and divide by zero is undefined behavior
// Unsigned integers do not overflow
const Scalar zero = Scalar(0);
const Scalar pos_one = Scalar(1);
const Packet cst_zero = pset1<Packet>(zero);
const Packet cst_pos_one = pset1<Packet>(pos_one);
const Packet pow_is_zero = exponent < 0 ? pcmp_lt(cst_pos_one, x) : pzero(x);
const Packet pow_is_one = pcmp_eq(cst_pos_one, x);
Packet result = pselect(pow_is_zero, cst_zero, x);
result = pselect(pow_is_one, cst_pos_one, result);
return result;
}
} // end namespace unary_pow
template <typename Packet, typename ScalarExponent,