From 8c8a79cec1b7d03be30df0e70cea236b8f52ef64 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 31 Mar 2016 10:33:32 -0700 Subject: Fixed a typo --- Eigen/src/Core/arch/CUDA/Half.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'Eigen') diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h index 6c412159c..ace250c6f 100644 --- a/Eigen/src/Core/arch/CUDA/Half.h +++ b/Eigen/src/Core/arch/CUDA/Half.h @@ -348,7 +348,7 @@ namespace numext { static inline EIGEN_DEVICE_FUNC bool (isinf)(const Eigen::half& a) { return (a.x & 0x7fff) == 0x7c00; } -static inline EIGEN_HALF_CUDA_H bool (isnan)(const Eigen::half& a) { +static inline EIGEN_DEVICE_FUNC bool (isnan)(const Eigen::half& a) { #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 return __hisnan(a); #else -- cgit v1.2.3 From b575fb1d02f7a98c94a576284fbcd4ff85970120 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 31 Mar 2016 10:43:59 -0700 Subject: Added NumTraits for half floats --- Eigen/src/Core/arch/CUDA/Half.h | 12 ++++++++++++ 1 file changed, 12 insertions(+) (limited to 'Eigen') diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h index ace250c6f..dc7119c06 100644 --- a/Eigen/src/Core/arch/CUDA/Half.h +++ b/Eigen/src/Core/arch/CUDA/Half.h @@ -341,6 +341,18 @@ template<> struct is_arithmetic { enum { value = true }; }; } // end namespace internal +template<> struct NumTraits + : GenericNumTraits +{ + EIGEN_DEVICE_FUNC static inline float dummy_precision() { return 1e-3f; } + EIGEN_DEVICE_FUNC static inline Eigen::half highest() { + return internal::raw_uint16_to_half(0x7bff); + } + EIGEN_DEVICE_FUNC static inline Eigen::half lowest() { + return internal::raw_uint16_to_half(0xfbff); + } +}; + // Infinity/NaN checks. namespace numext { -- cgit v1.2.3 From c36ab1990247a5b60bcad564759e8903f30fbab5 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 31 Mar 2016 10:55:03 -0700 Subject: Added __ldg primitive for fp16. --- Eigen/src/Core/arch/CUDA/Half.h | 7 +++++++ 1 file changed, 7 insertions(+) (limited to 'Eigen') diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h index dc7119c06..a2a2bac37 100644 --- a/Eigen/src/Core/arch/CUDA/Half.h +++ b/Eigen/src/Core/arch/CUDA/Half.h @@ -423,7 +423,14 @@ using ::ceil; __device__ inline Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) { return static_cast(__shfl_xor(static_cast(var), laneMask, width)); } +#endif +// ldg() has an overload for __half, but we also need one for Eigen::half. +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 320 +static inline EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) { + return Eigen::internal::raw_uint16_to_half( + __ldg(reinterpret_cast(ptr))); +} #endif -- cgit v1.2.3 From 4c859181daa3807f54ee7ae8add6bac66e896ace Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 31 Mar 2016 12:48:38 -0700 Subject: Made it possible to use the NumTraits for complex and Array in a cuda kernel. --- Eigen/src/Core/NumTraits.h | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) (limited to 'Eigen') diff --git a/Eigen/src/Core/NumTraits.h b/Eigen/src/Core/NumTraits.h index b7b5e7d22..e065fa714 100644 --- a/Eigen/src/Core/NumTraits.h +++ b/Eigen/src/Core/NumTraits.h @@ -153,7 +153,9 @@ template struct NumTraits > MulCost = 4 * NumTraits::MulCost + 2 * NumTraits::AddCost }; + EIGEN_DEVICE_FUNC static inline Real epsilon() { return NumTraits::epsilon(); } + EIGEN_DEVICE_FUNC static inline Real dummy_precision() { return NumTraits::dummy_precision(); } }; @@ -166,7 +168,7 @@ struct NumTraits > typedef typename NumTraits::NonInteger NonIntegerScalar; typedef Array NonInteger; typedef ArrayType & Nested; - + enum { IsComplex = NumTraits::IsComplex, IsInteger = NumTraits::IsInteger, @@ -176,8 +178,10 @@ struct NumTraits > AddCost = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits::AddCost, MulCost = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits::MulCost }; - + + EIGEN_DEVICE_FUNC static inline RealScalar epsilon() { return NumTraits::epsilon(); } + EIGEN_DEVICE_FUNC static inline RealScalar dummy_precision() { return NumTraits::dummy_precision(); } }; -- cgit v1.2.3 From f197813f370c7977bdd6023c13e08dfaf1f9498d Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 31 Mar 2016 13:09:23 -0700 Subject: Added the ability to hash a fp16 --- Eigen/src/Core/arch/CUDA/Half.h | 7 +++++++ 1 file changed, 7 insertions(+) (limited to 'Eigen') diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h index a2a2bac37..44645522a 100644 --- a/Eigen/src/Core/arch/CUDA/Half.h +++ b/Eigen/src/Core/arch/CUDA/Half.h @@ -414,6 +414,13 @@ using ::log; using ::sqrt; using ::floor; using ::ceil; +template <> + +struct hash { + size_t operator()(const Eigen::half& a) const { + return std::hash()(a.x); + } +}; } // end namespace std -- cgit v1.2.3 From 92b7f7b6503f2fa66e1f346b88fb6bff434d4d1d Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 31 Mar 2016 13:09:58 -0700 Subject: Improved code formating --- Eigen/src/Core/arch/CUDA/Half.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'Eigen') diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h index 44645522a..70050358c 100644 --- a/Eigen/src/Core/arch/CUDA/Half.h +++ b/Eigen/src/Core/arch/CUDA/Half.h @@ -414,8 +414,8 @@ using ::log; using ::sqrt; using ::floor; using ::ceil; -template <> +template <> struct hash { size_t operator()(const Eigen::half& a) const { return std::hash()(a.x); -- cgit v1.2.3 From 0ea7ab4f623864c82163d106cc93c8a97e4baac6 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 31 Mar 2016 14:44:55 -0700 Subject: Hashing was only officially introduced in c++11. Therefore only define an implementation of the hash function for float16 if c++11 is enabled. --- Eigen/src/Core/arch/CUDA/Half.h | 2 ++ 1 file changed, 2 insertions(+) (limited to 'Eigen') diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h index 70050358c..212aa0d5d 100644 --- a/Eigen/src/Core/arch/CUDA/Half.h +++ b/Eigen/src/Core/arch/CUDA/Half.h @@ -415,12 +415,14 @@ using ::sqrt; using ::floor; using ::ceil; +#if __cplusplus > 199711L template <> struct hash { size_t operator()(const Eigen::half& a) const { return std::hash()(a.x); } }; +#endif } // end namespace std -- cgit v1.2.3