devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix Half NaN definition and test.

Browse Source 
The `half_float` test was failing with `-mcpu=cortex-a55` (native `__fp16`) due
to a bad NaN bit-pattern comparison (in the case of casting a float to `__fp16`,
the signaling `NaN` is quieted). There was also an inconsistency between
`numeric_limits<half>::quiet_NaN()` and `NumTraits::quiet_NaN()`.  Here we
correct the inconsistency and compare NaNs according to the IEEE 754
definition.

Also modified the `bfloat16_float` test to match.

Tested with `cortex-a53` and `cortex-a55`.
This commit is contained in:
Rasmus Munk Larsen
2020-11-24 20:53:07 +00:00
parent 22f67b5958
commit c770746d70
8 changed files with 176 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
Eigen/src/Core/GenericPacketMath.h
View File

@@ -539,6 +539,20 @@ inline void pbroadcast2(const typename unpacket_traits<Packet>::type *a,
template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet
plset(const typename unpacket_traits<Packet>::type& a) { return a; }
/** \internal \returns a packet with constant coefficients \a a, e.g.: (x, 0, x, 0),
where x is the value of all 1-bits. */
template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
peven_mask(const Packet& /*a*/) {
typedef typename unpacket_traits<Packet>::type Scalar;
const size_t n = unpacket_traits<Packet>::size;
Scalar elements[n];
for(size_t i = 0; i < n; ++i) {
memset(elements+i, ((i & 1) == 0 ? 0xff : 0), sizeof(Scalar));
}
return ploadu<Packet>(elements);
}
/** \internal copy the packet \a from to \a *to, \a to must be 16 bytes aligned */
template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstore(Scalar* to, const Packet& from)
{ (*to) = from; }