macros.hpp

Go to the documentation of this file.

#ifndef TDLS_CORE_MACROS_HPP
#define TDLS_CORE_MACROS_HPP
 
 
 
 
#include <cstddef>
 
#include <tdls/core/target.hpp>
 
 
 
 
#ifndef TDLS_HOST_DEVICE
#if defined(TDLS_DETAIL_LANG_HIP)
#define TDLS_HOST_DEVICE __attribute__((host)) __attribute__((device))
#elif defined(TDLS_DETAIL_FRONTEND_CUDA)
#define TDLS_HOST_DEVICE __host__ __device__
#else
#define TDLS_HOST_DEVICE
#endif
#endif
 
 
 
 
#ifndef TDLS_FORCE_INLINE_GPU
#if defined(TDLS_DETAIL_LANG_HIP)
#define TDLS_FORCE_INLINE_GPU inline __attribute__((always_inline))
#elif defined(TDLS_DETAIL_FRONTEND_CUDA)
#define TDLS_FORCE_INLINE_GPU __forceinline__
#elif defined(TDLS_DETAIL_LANG_SYCL_DEVICE)
#define TDLS_FORCE_INLINE_GPU inline __attribute__((always_inline))
#else
#define TDLS_FORCE_INLINE_GPU inline
#endif
#endif
 
 
 
 
#ifndef TDLS_FORCE_INLINE_ALL
#if defined(TDLS_DETAIL_LANG_HIP) || defined(TDLS_DETAIL_FRONTEND_CUDA)
#define TDLS_FORCE_INLINE_ALL TDLS_FORCE_INLINE_GPU
#elif defined(TDLS_DETAIL_COMPILER_MSVC)
#define TDLS_FORCE_INLINE_ALL __forceinline
#elif defined(TDLS_DETAIL_COMPILER_GNU_ATTRIBUTES)
#define TDLS_FORCE_INLINE_ALL inline __attribute__((always_inline))
#else
#define TDLS_FORCE_INLINE_ALL inline
#endif
#endif
 
 
 
 
#ifndef TDLS_NO_INLINE_GPU
#if defined(TDLS_DETAIL_LANG_HIP)
#define TDLS_NO_INLINE_GPU __attribute__((noinline))
#elif defined(TDLS_DETAIL_FRONTEND_CUDA)
#define TDLS_NO_INLINE_GPU __noinline__
#elif defined(TDLS_DETAIL_LANG_SYCL_DEVICE)
#define TDLS_NO_INLINE_GPU __attribute__((noinline))
#else
#define TDLS_NO_INLINE_GPU inline
#endif
#endif
 
 
 
 
#ifndef TDLS_NO_INLINE
#if defined(TDLS_DETAIL_FRONTEND_CUDA)
#define TDLS_NO_INLINE __noinline__
#elif defined(TDLS_DETAIL_COMPILER_MSVC)
#define TDLS_NO_INLINE __declspec(noinline)
#elif defined(TDLS_DETAIL_COMPILER_GNU_ATTRIBUTES)
#define TDLS_NO_INLINE __attribute__((noinline))
#else
#define TDLS_NO_INLINE
#endif
#endif
 
 
 
 
#ifndef TDLS_RESTRICT
#if defined(TDLS_DETAIL_COMPILER_MSVC)
#define TDLS_RESTRICT __restrict
#elif defined(TDLS_DETAIL_CODEGEN_NVCXX)
#define TDLS_RESTRICT
#else
#define TDLS_RESTRICT __restrict__
#endif
#endif
 
 
 
 
#ifndef TDLS_STRIDED
#if defined(TDLS_DETAIL_CODEGEN_NVCXX)
#define TDLS_STRIDED(i, s)                                                                         \
    (static_cast<std::size_t>(static_cast<std::ptrdiff_t>(i)) * static_cast<unsigned>(s))
#else
#define TDLS_STRIDED(i, s)                                                                         \
    (static_cast<std::size_t>(static_cast<unsigned>(i)) * static_cast<unsigned>(s))
#endif
#endif
 
 
 
 
#ifndef TDLS_STRIDED_WIDE
#if defined(TDLS_DETAIL_CODEGEN_NVCXX)
#define TDLS_STRIDED_WIDE(i, s) (static_cast<std::size_t>(static_cast<std::ptrdiff_t>(i)) * (s))
#else
#define TDLS_STRIDED_WIDE(i, s) (static_cast<std::size_t>(static_cast<unsigned>(i)) * (s))
#endif
#endif
 
 
 
 
#ifndef TDLS_UNROLL_FORCE
#if defined(TDLS_DETAIL_FRONTEND_CUDA) || defined(TDLS_DETAIL_LANG_HIP) ||                         \
    defined(TDLS_DETAIL_COMPILER_NVCXX) || defined(TDLS_DETAIL_COMPILER_CLANG)
#define TDLS_UNROLL_FORCE _Pragma("unroll")
#elif defined(TDLS_DETAIL_COMPILER_GCC)
#define TDLS_UNROLL_FORCE _Pragma("GCC unroll 64")
#else
#define TDLS_UNROLL_FORCE
#endif
#endif
 
 
 
 
#ifndef TDLS_EXEC_CHECK_DISABLE
#if defined(TDLS_DETAIL_FRONTEND_CUDA) && !defined(TDLS_DETAIL_COMPILER_CLANG) &&                  \
    !defined(TDLS_DETAIL_COMPILER_NVCXX)
#define TDLS_EXEC_CHECK_DISABLE _Pragma("nv_exec_check_disable")
#else
#define TDLS_EXEC_CHECK_DISABLE
#endif
#endif
 
 
 
#endif // TDLS_CORE_MACROS_HPP