1a4af0ceeSJacob Faibussowitsch #include <petsc/private/deviceimpl.h> 2c2b86a48SJunchao Zhang #include <Kokkos_Core.hpp> 3c2b86a48SJunchao Zhang 445639126SStefano Zampini PetscBool PetscKokkosInitialized = PETSC_FALSE; 545639126SStefano Zampini 6c2b86a48SJunchao Zhang PetscErrorCode PetscKokkosFinalize_Private(void) 7c2b86a48SJunchao Zhang { 8c2b86a48SJunchao Zhang PetscFunctionBegin; 9c2b86a48SJunchao Zhang Kokkos::finalize(); 10c2b86a48SJunchao Zhang PetscFunctionReturn(0); 11c2b86a48SJunchao Zhang } 12c2b86a48SJunchao Zhang 13c2b86a48SJunchao Zhang PetscErrorCode PetscKokkosIsInitialized_Private(PetscBool *isInitialized) 14c2b86a48SJunchao Zhang { 15c2b86a48SJunchao Zhang PetscFunctionBegin; 16c2b86a48SJunchao Zhang *isInitialized = Kokkos::is_initialized() ? PETSC_TRUE : PETSC_FALSE; 17c2b86a48SJunchao Zhang PetscFunctionReturn(0); 18c2b86a48SJunchao Zhang } 19375e5adfSJunchao Zhang 2034766dafSJunchao Zhang /* Initialize Kokkos if not yet */ 21375e5adfSJunchao Zhang PetscErrorCode PetscKokkosInitializeCheck(void) 22375e5adfSJunchao Zhang { 23375e5adfSJunchao Zhang PetscFunctionBegin; 2434766dafSJunchao Zhang if (!Kokkos::is_initialized()) { 25*c66e0907SJunchao Zhang #if PETSC_PKG_KOKKOS_VERSION_GE(3,6,99) 26*c66e0907SJunchao Zhang auto args = Kokkos::InitializationSettings(); 27*c66e0907SJunchao Zhang #else 2862825ce1SJacob Faibussowitsch auto args = Kokkos::InitArguments{}; /* use default constructor */ 29*c66e0907SJunchao Zhang #endif 30b84ac304SJunchao Zhang 3162825ce1SJacob Faibussowitsch #if (defined(KOKKOS_ENABLE_CUDA) && PetscDefined(HAVE_CUDA)) || (defined(KOKKOS_ENABLE_HIP) && PetscDefined(HAVE_HIP)) || (defined(KOKKOS_ENABLE_SYCL) && PetscDefined(HAVE_SYCL)) 3262825ce1SJacob Faibussowitsch /* Kokkos does not support CUDA and HIP at the same time (but we do :)) */ 33a4af0ceeSJacob Faibussowitsch PetscDeviceContext dctx; 34b9d03b0cSStefano Zampini 359566063dSJacob Faibussowitsch PetscCall(PetscDeviceContextGetCurrentContext(&dctx)); 36*c66e0907SJunchao Zhang #if PETSC_PKG_KOKKOS_VERSION_GE(3,6,99) 37*c66e0907SJunchao Zhang args.set_device_id(static_cast<int>(dctx->device->deviceId)); 38*c66e0907SJunchao Zhang #else 399566063dSJacob Faibussowitsch PetscCall(PetscMPIIntCast(dctx->device->deviceId,&args.device_id)); 40375e5adfSJunchao Zhang #endif 41*c66e0907SJunchao Zhang #endif 4211f0be55SJunchao Zhang 43*c66e0907SJunchao Zhang #if PETSC_PKG_KOKKOS_VERSION_GE(3,6,99) 44*c66e0907SJunchao Zhang args.set_disable_warnings(!PetscDefined(HAVE_KOKKOS_INIT_WARNINGS)); 45*c66e0907SJunchao Zhang #else 4662825ce1SJacob Faibussowitsch args.disable_warnings = !PetscDefined(HAVE_KOKKOS_INIT_WARNINGS); 47*c66e0907SJunchao Zhang #endif 4862825ce1SJacob Faibussowitsch 4911f0be55SJunchao Zhang /* To use PetscNumOMPThreads, one has to configure petsc --with-openmp. 5011f0be55SJunchao Zhang Otherwise, let's keep the default value (-1) of args.num_threads. 5111f0be55SJunchao Zhang */ 5262825ce1SJacob Faibussowitsch #if defined(KOKKOS_ENABLE_OPENMP) && PetscDefined(HAVE_OPENMP) 53*c66e0907SJunchao Zhang #if PETSC_PKG_KOKKOS_VERSION_GE(3,6,99) 54*c66e0907SJunchao Zhang args.set_num_threads(PetscNumOMPThreads); 55*c66e0907SJunchao Zhang #else 5611f0be55SJunchao Zhang args.num_threads = PetscNumOMPThreads; 5711f0be55SJunchao Zhang #endif 58*c66e0907SJunchao Zhang #endif 5911f0be55SJunchao Zhang 6034766dafSJunchao Zhang Kokkos::initialize(args); 6159e55d94SJunchao Zhang PetscBeganKokkos = PETSC_TRUE; 6259e55d94SJunchao Zhang } 6345639126SStefano Zampini PetscKokkosInitialized = PETSC_TRUE; 64375e5adfSJunchao Zhang PetscFunctionReturn(0); 65375e5adfSJunchao Zhang } 66