diff options
author | Mehdi Goli <mehdi.goli@codeplay.com> | 2017-05-22 16:49:32 +0100 |
---|---|---|
committer | Mehdi Goli <mehdi.goli@codeplay.com> | 2017-05-22 16:49:32 +0100 |
commit | 76c0fc1f955eda3d243db8960cb6fee9a5305112 (patch) | |
tree | 76ce81fd151bfa088ceb3acba3c310cc80f47804 | |
parent | 0d08165a7f7a95c35dead32ec2d567e9a4b609b0 (diff) |
Fixing SYCL alignment issue required by TensorFlow.
-rw-r--r-- | unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h | 23 |
1 files changed, 21 insertions, 2 deletions
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h index c5142b7c9..627c0ab19 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h @@ -14,7 +14,23 @@ #if defined(EIGEN_USE_SYCL) && !defined(EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H) #define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H +template<size_t Align> struct CheckAlignStatically{ + static const bool Val= (((Align&(Align-1))==0) && (Align >= sizeof(void *))); +}; +template <bool IsAligned, size_t Align> +struct Conditional_Allocate{ + +EIGEN_ALWAYS_INLINE static void* conditional_allocate(std::size_t elements){ + return aligned_alloc(Align, elements); +} +}; +template <size_t Align> +struct Conditional_Allocate<false, Align>{ +EIGEN_ALWAYS_INLINE static void* conditional_allocate(std::size_t elements){ + return malloc(elements); +} +}; template <typename Scalar, size_t Align = EIGEN_MAX_ALIGN_BYTES, class Allocator = std::allocator<Scalar>> struct SyclAllocator { typedef Scalar value_type; @@ -22,7 +38,10 @@ struct SyclAllocator { typedef typename std::allocator_traits<Allocator>::size_type size_type; SyclAllocator( ){}; - Scalar* allocate(std::size_t elements) { return static_cast<Scalar*>(aligned_alloc(Align, elements)); } + Scalar* allocate(std::size_t elements) { + return static_cast<Scalar*>(Conditional_Allocate<CheckAlignStatically<Align>::Val, Align>::conditional_allocate(elements)); + // return static_cast<Scalar*>(aligned_alloc(Align, elements)); + } void deallocate(Scalar * p, std::size_t size) { EIGEN_UNUSED_VARIABLE(size); free(p); } }; @@ -533,4 +552,4 @@ struct SyclKernelDevice:DefaultDevice{}; } // end namespace Eigen -#endif // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H +#endif // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
\ No newline at end of file |