Fix pfrexp/pldexp for half.

The recent addition of vectorized pow (!330) relies on `pfrexp` and
`pldexp`.  This was missing for `Eigen::half` and `Eigen::bfloat16`.
Adding tests for these packet ops also exposed an issue with handling
negative values in `pfrexp`, returning an incorrect exponent.

Added the missing implementations, corrected the exponent in `pfrexp1`,
and added `packetmath` tests.
This commit is contained in:
Antonio Sanchez
2021-01-20 19:00:09 -08:00
committed by Rasmus Munk Larsen
parent 25d8498f8b
commit b2126fd6b5
7 changed files with 112 additions and 4 deletions

View File

@@ -184,6 +184,19 @@ F16_PACKET_FUNCTION(Packet8f, Packet8h, ptanh)
F16_PACKET_FUNCTION(Packet8f, Packet8h, psqrt)
F16_PACKET_FUNCTION(Packet8f, Packet8h, prsqrt)
template <>
EIGEN_STRONG_INLINE Packet8h pfrexp(const Packet8h& a, Packet8h& exponent) {
Packet8f fexponent;
const Packet8h out = float2half(pfrexp<Packet8f>(half2float(a), fexponent));
exponent = float2half(fexponent);
return out;
}
template <>
EIGEN_STRONG_INLINE Packet8h pldexp(const Packet8h& a, const Packet8h& exponent) {
return float2half(pldexp<Packet8f>(half2float(a), half2float(exponent)));
}
BF16_PACKET_FUNCTION(Packet8f, Packet8bf, psin)
BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pcos)
BF16_PACKET_FUNCTION(Packet8f, Packet8bf, plog)
@@ -195,6 +208,19 @@ BF16_PACKET_FUNCTION(Packet8f, Packet8bf, ptanh)
BF16_PACKET_FUNCTION(Packet8f, Packet8bf, psqrt)
BF16_PACKET_FUNCTION(Packet8f, Packet8bf, prsqrt)
template <>
EIGEN_STRONG_INLINE Packet8bf pfrexp(const Packet8bf& a, Packet8bf& exponent) {
Packet8f fexponent;
const Packet8bf out = F32ToBf16(pfrexp<Packet8f>(Bf16ToF32(a), fexponent));
exponent = F32ToBf16(fexponent);
return out;
}
template <>
EIGEN_STRONG_INLINE Packet8bf pldexp(const Packet8bf& a, const Packet8bf& exponent) {
return F32ToBf16(pldexp<Packet8f>(Bf16ToF32(a), Bf16ToF32(exponent)));
}
} // end namespace internal
} // end namespace Eigen