Updated pfrexp implementation.

The original implementation fails for 0, denormals, inf, and NaN.

See #2150
This commit is contained in:
Antonio Sanchez
2021-02-12 11:32:29 -08:00
committed by Rasmus Munk Larsen
parent 9ad4096ccb
commit 7ff0b7a980
9 changed files with 225 additions and 144 deletions

View File

@@ -887,21 +887,24 @@ template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a)
}
template<> EIGEN_STRONG_INLINE Packet4f pfrexp<Packet4f>(const Packet4f& a, Packet4f& exponent) {
return pfrexp_float(a,exponent);
return pfrexp_generic(a,exponent);
}
// Extract exponent without existence of Packet2l.
template<>
EIGEN_STRONG_INLINE
Packet2d pfrexp_generic_get_biased_exponent(const Packet2d& a) {
const Packet2d cst_exp_mask = pset1frombits<Packet2d>(static_cast<uint64_t>(0x7ff0000000000000ull));
__m128i a_expo = _mm_srli_epi64(_mm_castpd_si128(pand(a, cst_exp_mask)), 52);
return _mm_cvtepi32_pd(vec4i_swizzle1(a_expo, 0, 2, 1, 3));
}
template<> EIGEN_STRONG_INLINE Packet2d pfrexp<Packet2d>(const Packet2d& a, Packet2d& exponent) {
const Packet2d cst_1022d = pset1<Packet2d>(1022.0);
const Packet2d cst_half = pset1<Packet2d>(0.5);
const Packet2d cst_exp_mask = pset1frombits<Packet2d>(static_cast<uint64_t>(0x7ff0000000000000ull));
__m128i a_expo = _mm_srli_epi64(_mm_castpd_si128(pand(a, cst_exp_mask)), 52);
exponent = psub(_mm_cvtepi32_pd(vec4i_swizzle1(a_expo, 0, 2, 1, 3)), cst_1022d);
const Packet2d cst_mant_mask = pset1frombits<Packet2d>(static_cast<uint64_t>(~0x7ff0000000000000ull));
return por(pand(a, cst_mant_mask), cst_half);
return pfrexp_generic(a, exponent);
}
template<> EIGEN_STRONG_INLINE Packet4f pldexp<Packet4f>(const Packet4f& a, const Packet4f& exponent) {
return pldexp_float(a,exponent);
return pldexp_generic(a,exponent);
}
// We specialize pldexp here, since the generic implementation uses Packet2l, which is not well