Update Kokkos library in LAMMPS to v4.1.0
This commit is contained in:
@ -46,350 +46,9 @@
|
||||
|
||||
#include <Kokkos_Macros.hpp>
|
||||
|
||||
#ifdef KOKKOS_ENABLE_IMPL_DESUL_ATOMICS
|
||||
#include <Kokkos_Atomics_Desul_Wrapper.hpp>
|
||||
#include <Kokkos_Atomics_Desul_Volatile_Wrapper.hpp>
|
||||
#include <impl/Kokkos_Utilities.hpp>
|
||||
|
||||
// Helper functions for places where we really should have called SeqCst atomics
|
||||
// anyway These can go away when we call desul unconditionally Non-Desul
|
||||
// versions are below
|
||||
namespace Kokkos {
|
||||
namespace Impl {
|
||||
using desul::MemoryOrderSeqCst;
|
||||
using desul::MemoryScopeDevice;
|
||||
|
||||
template <class T>
|
||||
KOKKOS_INLINE_FUNCTION void desul_atomic_dec(T* dest, MemoryOrderSeqCst,
|
||||
MemoryScopeDevice) {
|
||||
return desul::atomic_dec(const_cast<T*>(dest), desul::MemoryOrderSeqCst(),
|
||||
desul::MemoryScopeDevice());
|
||||
}
|
||||
|
||||
template <class T>
|
||||
KOKKOS_INLINE_FUNCTION void desul_atomic_inc(T* dest, MemoryOrderSeqCst,
|
||||
MemoryScopeDevice) {
|
||||
return desul::atomic_inc(const_cast<T*>(dest), desul::MemoryOrderSeqCst(),
|
||||
desul::MemoryScopeDevice());
|
||||
}
|
||||
|
||||
template <class T>
|
||||
KOKKOS_INLINE_FUNCTION T
|
||||
desul_atomic_exchange(T* dest, const Kokkos::Impl::type_identity_t<T> val,
|
||||
MemoryOrderSeqCst, MemoryScopeDevice) {
|
||||
return desul::atomic_exchange(const_cast<T*>(dest), val,
|
||||
desul::MemoryOrderSeqCst(),
|
||||
desul::MemoryScopeDevice());
|
||||
}
|
||||
|
||||
template <class T>
|
||||
KOKKOS_INLINE_FUNCTION T desul_atomic_compare_exchange(
|
||||
T* dest, Kokkos::Impl::type_identity_t<const T> compare,
|
||||
Kokkos::Impl::type_identity_t<const T> val, MemoryOrderSeqCst,
|
||||
MemoryScopeDevice) {
|
||||
return desul::atomic_compare_exchange(dest, compare, val,
|
||||
desul::MemoryOrderSeqCst(),
|
||||
desul::MemoryScopeDevice());
|
||||
}
|
||||
|
||||
} // namespace Impl
|
||||
} // namespace Kokkos
|
||||
#else
|
||||
|
||||
#include <Kokkos_HostSpace.hpp>
|
||||
#include <impl/Kokkos_Traits.hpp>
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
|
||||
// Need to fix this for pure clang on windows
|
||||
#if defined(_WIN32)
|
||||
#define KOKKOS_ENABLE_WINDOWS_ATOMICS
|
||||
|
||||
#if defined(KOKKOS_ENABLE_CUDA)
|
||||
#define KOKKOS_ENABLE_CUDA_ATOMICS
|
||||
#if defined(KOKKOS_COMPILER_CLANG)
|
||||
#define KOKKOS_ENABLE_GNU_ATOMICS
|
||||
#endif
|
||||
#endif
|
||||
|
||||
#else // _WIN32
|
||||
#if defined(KOKKOS_ENABLE_CUDA)
|
||||
|
||||
// Compiling NVIDIA device code, must use Cuda atomics:
|
||||
|
||||
#define KOKKOS_ENABLE_CUDA_ATOMICS
|
||||
|
||||
#elif defined(KOKKOS_ENABLE_HIP)
|
||||
|
||||
#define KOKKOS_ENABLE_HIP_ATOMICS
|
||||
|
||||
#endif
|
||||
|
||||
#if !defined(KOKKOS_ENABLE_GNU_ATOMICS) && \
|
||||
!defined(KOKKOS_ENABLE_INTEL_ATOMICS) && \
|
||||
!defined(KOKKOS_ENABLE_OPENMP_ATOMICS) && \
|
||||
!defined(KOKKOS_ENABLE_STD_ATOMICS) && \
|
||||
!defined(KOKKOS_ENABLE_SERIAL_ATOMICS)
|
||||
|
||||
// Compiling for non-Cuda atomic implementation has not been pre-selected.
|
||||
// Choose the best implementation for the detected compiler.
|
||||
// Preference: GCC, INTEL, OMP31
|
||||
|
||||
#if defined(KOKKOS_INTERNAL_NOT_PARALLEL)
|
||||
|
||||
#define KOKKOS_ENABLE_SERIAL_ATOMICS
|
||||
|
||||
#elif defined(KOKKOS_COMPILER_GNU) || defined(KOKKOS_COMPILER_CLANG) || \
|
||||
defined(KOKKOS_COMPILER_NVCC)
|
||||
|
||||
#define KOKKOS_ENABLE_GNU_ATOMICS
|
||||
|
||||
#elif defined(KOKKOS_COMPILER_INTEL) || defined(KOKKOS_COMPILER_CRAYC)
|
||||
|
||||
#define KOKKOS_ENABLE_INTEL_ATOMICS
|
||||
|
||||
#elif defined(_OPENMP) && (201107 <= _OPENMP)
|
||||
|
||||
#define KOKKOS_ENABLE_OPENMP_ATOMICS
|
||||
|
||||
#else
|
||||
|
||||
#error "KOKKOS_ATOMICS_USE : Unsupported compiler"
|
||||
|
||||
#endif
|
||||
|
||||
#endif /* Not pre-selected atomic implementation */
|
||||
#endif
|
||||
|
||||
#ifdef KOKKOS_ENABLE_CUDA
|
||||
#include <Cuda/Kokkos_Cuda_Locks.hpp>
|
||||
#endif
|
||||
|
||||
namespace Kokkos {
|
||||
template <typename T>
|
||||
KOKKOS_INLINE_FUNCTION void atomic_add(volatile T* const dest, const T src);
|
||||
|
||||
// Atomic increment
|
||||
template <typename T>
|
||||
KOKKOS_INLINE_FUNCTION void atomic_increment(volatile T* a);
|
||||
|
||||
template <typename T>
|
||||
KOKKOS_INLINE_FUNCTION void atomic_decrement(volatile T* a);
|
||||
} // namespace Kokkos
|
||||
|
||||
namespace Kokkos {
|
||||
|
||||
inline const char* atomic_query_version() {
|
||||
#if defined(KOKKOS_ENABLE_CUDA_ATOMICS)
|
||||
return "KOKKOS_ENABLE_CUDA_ATOMICS";
|
||||
#elif defined(KOKKOS_ENABLE_GNU_ATOMICS)
|
||||
return "KOKKOS_ENABLE_GNU_ATOMICS";
|
||||
#elif defined(KOKKOS_ENABLE_INTEL_ATOMICS)
|
||||
return "KOKKOS_ENABLE_INTEL_ATOMICS";
|
||||
#elif defined(KOKKOS_ENABLE_OPENMP_ATOMICS)
|
||||
return "KOKKOS_ENABLE_OPENMP_ATOMICS";
|
||||
#elif defined(KOKKOS_ENABLE_WINDOWS_ATOMICS)
|
||||
return "KOKKOS_ENABLE_WINDOWS_ATOMICS";
|
||||
#elif defined(KOKKOS_ENABLE_SERIAL_ATOMICS)
|
||||
return "KOKKOS_ENABLE_SERIAL_ATOMICS";
|
||||
#else
|
||||
#error "No valid response for atomic_query_version!"
|
||||
#endif
|
||||
}
|
||||
|
||||
} // namespace Kokkos
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Atomic Memory Orders
|
||||
//
|
||||
// Implements Strongly-typed analogs of C++ standard memory orders
|
||||
#include "impl/Kokkos_Atomic_Memory_Order.hpp"
|
||||
|
||||
#if defined(KOKKOS_ENABLE_HIP)
|
||||
#include <HIP/Kokkos_HIP_Atomic.hpp>
|
||||
#endif
|
||||
|
||||
#if defined(KOKKOS_ENABLE_WINDOWS_ATOMICS)
|
||||
#include "impl/Kokkos_Atomic_Windows.hpp"
|
||||
#endif
|
||||
//----------------------------------------------------------------------------
|
||||
// Atomic Assembly
|
||||
//
|
||||
// Implements CAS128-bit in assembly
|
||||
|
||||
#include "impl/Kokkos_Atomic_Assembly.hpp"
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Memory fence
|
||||
//
|
||||
// All loads and stores from this thread will be globally consistent before
|
||||
// continuing
|
||||
//
|
||||
// void memory_fence() {...};
|
||||
#include "impl/Kokkos_Memory_Fence.hpp"
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Atomic exchange
|
||||
//
|
||||
// template< typename T >
|
||||
// T atomic_exchange( volatile T* const dest , const T val )
|
||||
// { T tmp = *dest ; *dest = val ; return tmp ; }
|
||||
|
||||
#include "impl/Kokkos_Atomic_Exchange.hpp"
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Atomic compare-and-exchange
|
||||
//
|
||||
// template<class T>
|
||||
// bool atomic_compare_exchange_strong(volatile T* const dest, const T compare,
|
||||
// const T val) { bool equal = compare == *dest ; if ( equal ) { *dest = val ; }
|
||||
// return equal ; }
|
||||
|
||||
#include "impl/Kokkos_Atomic_Compare_Exchange_Strong.hpp"
|
||||
|
||||
#include "impl/Kokkos_Atomic_Generic.hpp"
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Atomic fetch and add
|
||||
//
|
||||
// template<class T>
|
||||
// T atomic_fetch_add(volatile T* const dest, const T val)
|
||||
// { T tmp = *dest ; *dest += val ; return tmp ; }
|
||||
|
||||
#include "impl/Kokkos_Atomic_Fetch_Add.hpp"
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Atomic increment
|
||||
//
|
||||
// template<class T>
|
||||
// T atomic_increment(volatile T* const dest)
|
||||
// { dest++; }
|
||||
|
||||
#include "impl/Kokkos_Atomic_Increment.hpp"
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Atomic Decrement
|
||||
//
|
||||
// template<class T>
|
||||
// T atomic_decrement(volatile T* const dest)
|
||||
// { dest--; }
|
||||
|
||||
#include "impl/Kokkos_Atomic_Decrement.hpp"
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Atomic fetch and sub
|
||||
//
|
||||
// template<class T>
|
||||
// T atomic_fetch_sub(volatile T* const dest, const T val)
|
||||
// { T tmp = *dest ; *dest -= val ; return tmp ; }
|
||||
|
||||
#include "impl/Kokkos_Atomic_Fetch_Sub.hpp"
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Atomic fetch and or
|
||||
//
|
||||
// template<class T>
|
||||
// T atomic_fetch_or(volatile T* const dest, const T val)
|
||||
// { T tmp = *dest ; *dest = tmp | val ; return tmp ; }
|
||||
|
||||
#include "impl/Kokkos_Atomic_Fetch_Or.hpp"
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Atomic fetch and and
|
||||
//
|
||||
// template<class T>
|
||||
// T atomic_fetch_and(volatile T* const dest, const T val)
|
||||
// { T tmp = *dest ; *dest = tmp & val ; return tmp ; }
|
||||
|
||||
#include "impl/Kokkos_Atomic_Fetch_And.hpp"
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Atomic MinMax
|
||||
//
|
||||
// template<class T>
|
||||
// T atomic_min(volatile T* const dest, const T val)
|
||||
// { T tmp = *dest ; *dest = min(*dest, val); return tmp ; }
|
||||
// template<class T>
|
||||
// T atomic_max(volatile T* const dest, const T val)
|
||||
// { T tmp = *dest ; *dest = max(*dest, val); return tmp ; }
|
||||
|
||||
#include "impl/Kokkos_Atomic_MinMax.hpp"
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Provide volatile_load and safe_load
|
||||
//
|
||||
// T volatile_load(T const volatile * const ptr);
|
||||
//
|
||||
// T const& safe_load(T const * const ptr);
|
||||
// XEON PHI
|
||||
// T safe_load(T const * const ptr
|
||||
|
||||
#include "impl/Kokkos_Volatile_Load.hpp"
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Provide atomic loads and stores with memory order semantics
|
||||
|
||||
#include "impl/Kokkos_Atomic_Load.hpp"
|
||||
#include "impl/Kokkos_Atomic_Store.hpp"
|
||||
|
||||
// Generic functions using the above defined functions
|
||||
#include "impl/Kokkos_Atomic_Generic_Secondary.hpp"
|
||||
//----------------------------------------------------------------------------
|
||||
// This atomic-style macro should be an inlined function, not a macro
|
||||
|
||||
#if defined(KOKKOS_COMPILER_GNU) && !defined(__PGIC__) && \
|
||||
!defined(__CUDA_ARCH__)
|
||||
|
||||
#define KOKKOS_NONTEMPORAL_PREFETCH_LOAD(addr) __builtin_prefetch(addr, 0, 0)
|
||||
#define KOKKOS_NONTEMPORAL_PREFETCH_STORE(addr) __builtin_prefetch(addr, 1, 0)
|
||||
|
||||
#else
|
||||
|
||||
#define KOKKOS_NONTEMPORAL_PREFETCH_LOAD(addr) ((void)0)
|
||||
#define KOKKOS_NONTEMPORAL_PREFETCH_STORE(addr) ((void)0)
|
||||
|
||||
#endif
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
|
||||
// Helper functions for places where we really should have called SeqCst atomics
|
||||
// anyway These can go away when we call desul unconditionally
|
||||
namespace Kokkos {
|
||||
namespace Impl {
|
||||
struct MemoryOrderSeqCst {};
|
||||
struct MemoryScopeDevice {};
|
||||
|
||||
template <class T>
|
||||
KOKKOS_INLINE_FUNCTION void desul_atomic_dec(T* dest, MemoryOrderSeqCst,
|
||||
MemoryScopeDevice) {
|
||||
return Kokkos::atomic_decrement(dest);
|
||||
}
|
||||
|
||||
template <class T>
|
||||
KOKKOS_INLINE_FUNCTION void desul_atomic_inc(T* dest, MemoryOrderSeqCst,
|
||||
MemoryScopeDevice) {
|
||||
return Kokkos::atomic_increment(dest);
|
||||
}
|
||||
|
||||
template <class T>
|
||||
KOKKOS_INLINE_FUNCTION T
|
||||
desul_atomic_exchange(T* dest, Kokkos::Impl::type_identity_t<const T> val,
|
||||
MemoryOrderSeqCst, MemoryScopeDevice) {
|
||||
return Kokkos::atomic_exchange(dest, val);
|
||||
}
|
||||
|
||||
template <class T>
|
||||
KOKKOS_INLINE_FUNCTION T desul_atomic_compare_exchange(
|
||||
T* dest, Kokkos::Impl::type_identity_t<const T> compare,
|
||||
Kokkos::Impl::type_identity_t<const T> val, MemoryOrderSeqCst,
|
||||
MemoryScopeDevice) {
|
||||
return Kokkos::atomic_compare_exchange(dest, compare, val);
|
||||
}
|
||||
|
||||
} // namespace Impl
|
||||
} // namespace Kokkos
|
||||
|
||||
#endif /* !KOKKOS_ENABLE_IMPL_DESUL_ATOMICS */
|
||||
#ifdef KOKKOS_IMPL_PUBLIC_INCLUDE_NOTDEFINED_ATOMIC
|
||||
#undef KOKKOS_IMPL_PUBLIC_INCLUDE
|
||||
#undef KOKKOS_IMPL_PUBLIC_INCLUDE_NOTDEFINED_ATOMIC
|
||||
|
||||
Reference in New Issue
Block a user