Special function implementations for half/bfloat16 packets.

Current implementations fail to consider half-float packets, only
half-float scalars.  Added specializations for packets on AVX, AVX512 and
NEON.  Added tests to `special_packetmath`.

The current `special_functions` tests would fail for half and bfloat16 due to
lack of precision. The NEON tests also fail with precision issues and
due to different handling of `sqrt(inf)`, so special functions bessel, ndtri
have been disabled.

Tested with AVX, AVX512.

1 files changed, 3 insertions, 2 deletions

diff --git a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
index 648eb053e..cfc13aff7 100644
--- a/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
+++ b/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h
@@ -348,7 +348,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T generic_fast_erf_float(const T& a_x) {
 template <typename T>
 struct erf_impl {
   EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE T run(const T x) {
+  static EIGEN_STRONG_INLINE T run(const T& x) {
     return generic_fast_erf_float(x);
   }
 };
@@ -490,7 +490,8 @@ struct erfc_impl<double> {
 template<typename T>
 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T flipsign(
     const T& should_flipsign, const T& x) {
-  const T sign_mask = pset1<T>(-0.0);
+  typedef typename unpacket_traits<T>::type Scalar;
+  const T sign_mask = pset1<T>(Scalar(-0.0));
   T sign_bit = pand<T>(should_flipsign, sign_mask);
   return pxor<T>(sign_bit, x);
 }