Fix some CUDA warnings.

Added `EIGEN_HAS_STD_HASH` macro, checking for C++11 support and not
running on GPU.

`std::hash<float>` is not a device function, so cannot be used by
`std::hash<bfloat16>`.  Removed `EIGEN_DEVICE_FUNC` and only
define if `EIGEN_HAS_STD_HASH`. Same for `half`.

Added `EIGEN_CUDA_HAS_FP16_ARITHMETIC` to improve readability,
eliminate warnings about `EIGEN_CUDA_ARCH` not being defined.

Replaced a couple C-style casts with `reinterpret_cast` for aligned
loading of `half*` to `half2*`. This eliminates `-Wcast-align`
warnings in clang.  Although not ideal due to potential type aliasing,
this is how CUDA handles these conversions internally.

1 files changed, 12 insertions, 14 deletions

diff --git a/Eigen/src/Core/arch/Default/BFloat16.h b/Eigen/src/Core/arch/Default/BFloat16.h
index 72a489b0b..81af56a3f 100644
--- a/Eigen/src/Core/arch/Default/BFloat16.h
+++ b/Eigen/src/Core/arch/Default/BFloat16.h
@@ -655,20 +655,6 @@ template<> struct NumTraits<Eigen::bfloat16>
 
 } // namespace Eigen
 
-namespace std {
-
-#if __cplusplus > 199711L
-template <>
-struct hash<Eigen::bfloat16> {
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t operator()(const Eigen::bfloat16& a) const {
-    return hash<float>()(static_cast<float>(a));
-  }
-};
-#endif
-
-} // namespace std
-
-
 namespace Eigen {
 namespace numext {
 
@@ -703,4 +689,16 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC uint16_t bit_cast<uint16_t, Eigen::bfloat1
 }  // namespace numext
 }  // namespace Eigen
 
+#if EIGEN_HAS_STD_HASH
+namespace std {
+template <>
+struct hash<Eigen::bfloat16> {
+  EIGEN_STRONG_INLINE std::size_t operator()(const Eigen::bfloat16& a) const {
+    return static_cast<std::size_t>(Eigen::numext::bit_cast<Eigen::numext::uint16_t>(a));
+  }
+};
+} // namespace std
+#endif
+
+
 #endif // EIGEN_BFLOAT16_H