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 pow and other cwise ops for half/bfloat16.

Browse Source 
The new `generic_pow` implementation was failing for half/bfloat16 since
their construction from int/float is not `constexpr`. Modified
in `GenericPacketMathFunctions` to remove `constexpr`.

While adding tests for half/bfloat16, found other issues related to
implicit conversions.

Also needed to implement `numext::arg` for non-integer, non-complex,
non-float/double/long double types.  These seem to be  implicitly
converted to `std::complex<T>`, which then fails for half/bfloat16.
This commit is contained in:
Antonio Sanchez
2021-01-22 11:10:54 -08:00
parent f19bcffee6
commit f0e46ed5d4
4 changed files with 92 additions and 69 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h
View File

@@ -804,8 +804,8 @@ EIGEN_STRONG_INLINE
void veltkamp_splitting(const Packet& x, Packet& x_hi, Packet& x_lo) {
typedef typename unpacket_traits<Packet>::type Scalar;
EIGEN_CONSTEXPR int shift = (NumTraits<Scalar>::digits() + 1) / 2;
EIGEN_CONSTEXPR Scalar shift_scale = Scalar(uint64_t(1) << shift);
Packet gamma = pmul(pset1<Packet>(shift_scale + 1), x);
Scalar shift_scale = Scalar(uint64_t(1) << shift); // Scalar constructor not necessarily constexpr.
Packet gamma = pmul(pset1<Packet>(shift_scale + Scalar(1)), x);
#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
x_hi = pmadd(pset1<Packet>(-shift_scale), x, gamma);
#else