1a4af0ceeSJacob Faibussowitsch #include <petsc/private/deviceimpl.h> 20e6b6b59SJacob Faibussowitsch #include <petscpkg_version.h> 3524fe776SJunchao Zhang #include <petsc_kokkos.hpp> 4c2b86a48SJunchao Zhang 545639126SStefano Zampini PetscBool PetscKokkosInitialized = PETSC_FALSE; 645639126SStefano Zampini 7524fe776SJunchao Zhang Kokkos::DefaultExecutionSpace *PetscKokkosExecutionSpacePtr = nullptr; 8524fe776SJunchao Zhang 9d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscKokkosFinalize_Private(void) 10d71ae5a4SJacob Faibussowitsch { 11c2b86a48SJunchao Zhang PetscFunctionBegin; 12524fe776SJunchao Zhang PetscCallCXX(delete PetscKokkosExecutionSpacePtr); 13c2b86a48SJunchao Zhang Kokkos::finalize(); 143ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 15c2b86a48SJunchao Zhang } 16c2b86a48SJunchao Zhang 17d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscKokkosIsInitialized_Private(PetscBool *isInitialized) 18d71ae5a4SJacob Faibussowitsch { 19c2b86a48SJunchao Zhang PetscFunctionBegin; 20c2b86a48SJunchao Zhang *isInitialized = Kokkos::is_initialized() ? PETSC_TRUE : PETSC_FALSE; 213ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 22c2b86a48SJunchao Zhang } 23375e5adfSJunchao Zhang 2434766dafSJunchao Zhang /* Initialize Kokkos if not yet */ 25d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscKokkosInitializeCheck(void) 26d71ae5a4SJacob Faibussowitsch { 27375e5adfSJunchao Zhang PetscFunctionBegin; 2834766dafSJunchao Zhang if (!Kokkos::is_initialized()) { 29471471fdSJunchao Zhang #if PETSC_PKG_KOKKOS_VERSION_GE(3, 7, 0) 30c66e0907SJunchao Zhang auto args = Kokkos::InitializationSettings(); 31c66e0907SJunchao Zhang #else 3262825ce1SJacob Faibussowitsch auto args = Kokkos::InitArguments{}; /* use default constructor */ 33c66e0907SJunchao Zhang #endif 34b84ac304SJunchao Zhang 3562825ce1SJacob Faibussowitsch #if (defined(KOKKOS_ENABLE_CUDA) && PetscDefined(HAVE_CUDA)) || (defined(KOKKOS_ENABLE_HIP) && PetscDefined(HAVE_HIP)) || (defined(KOKKOS_ENABLE_SYCL) && PetscDefined(HAVE_SYCL)) 3662825ce1SJacob Faibussowitsch /* Kokkos does not support CUDA and HIP at the same time (but we do :)) */ 37ab4ee011SJunchao Zhang PetscDevice device; 38ab4ee011SJunchao Zhang PetscInt deviceId; 39ab4ee011SJunchao Zhang PetscCall(PetscDeviceCreate(PETSC_DEVICE_DEFAULT(), PETSC_DECIDE, &device)); 40ab4ee011SJunchao Zhang PetscCall(PetscDeviceGetDeviceId(device, &deviceId)); 41ab4ee011SJunchao Zhang PetscCall(PetscDeviceDestroy(&device)); 42ab4ee011SJunchao Zhang #if PETSC_PKG_KOKKOS_VERSION_GE(4, 0, 0) 43ab4ee011SJunchao Zhang // if device_id is not set, and no gpus have been found, kokkos will use CPU 44ab4ee011SJunchao Zhang if (deviceId >= 0) args.set_device_id(static_cast<int>(deviceId)); 45ab4ee011SJunchao Zhang #elif PETSC_PKG_KOKKOS_VERSION_GE(3, 7, 0) 46ab4ee011SJunchao Zhang args.set_device_id(static_cast<int>(deviceId)); 47c66e0907SJunchao Zhang #else 48ab4ee011SJunchao Zhang PetscCall(PetscMPIIntCast(deviceId, &args.device_id)); 49375e5adfSJunchao Zhang #endif 50c66e0907SJunchao Zhang #endif 5111f0be55SJunchao Zhang 52471471fdSJunchao Zhang #if PETSC_PKG_KOKKOS_VERSION_GE(3, 7, 0) 53c66e0907SJunchao Zhang args.set_disable_warnings(!PetscDefined(HAVE_KOKKOS_INIT_WARNINGS)); 54c66e0907SJunchao Zhang #else 5562825ce1SJacob Faibussowitsch args.disable_warnings = !PetscDefined(HAVE_KOKKOS_INIT_WARNINGS); 56c66e0907SJunchao Zhang #endif 5762825ce1SJacob Faibussowitsch 5811f0be55SJunchao Zhang /* To use PetscNumOMPThreads, one has to configure petsc --with-openmp. 5911f0be55SJunchao Zhang Otherwise, let's keep the default value (-1) of args.num_threads. 6011f0be55SJunchao Zhang */ 6162825ce1SJacob Faibussowitsch #if defined(KOKKOS_ENABLE_OPENMP) && PetscDefined(HAVE_OPENMP) 62471471fdSJunchao Zhang #if PETSC_PKG_KOKKOS_VERSION_GE(3, 7, 0) 63c66e0907SJunchao Zhang args.set_num_threads(PetscNumOMPThreads); 64c66e0907SJunchao Zhang #else 6511f0be55SJunchao Zhang args.num_threads = PetscNumOMPThreads; 6611f0be55SJunchao Zhang #endif 67c66e0907SJunchao Zhang #endif 68ab4ee011SJunchao Zhang PetscCallCXX(Kokkos::initialize(args)); 69*9c9deb76SJunchao Zhang PetscBeganKokkos = PETSC_TRUE; 70*9c9deb76SJunchao Zhang } 71*9c9deb76SJunchao Zhang if (!PetscKokkosExecutionSpacePtr) { // No matter Kokkos is init'ed by petsc or by user, we need to init PetscKokkosExecutionSpacePtr 72524fe776SJunchao Zhang #if defined(PETSC_HAVE_CUDA) 73524fe776SJunchao Zhang extern cudaStream_t PetscDefaultCudaStream; 74524fe776SJunchao Zhang PetscCallCXX(PetscKokkosExecutionSpacePtr = new Kokkos::DefaultExecutionSpace(PetscDefaultCudaStream)); 75524fe776SJunchao Zhang #elif defined(PETS_HAVE_HIP) 76524fe776SJunchao Zhang extern hipStream_t PetscDefaultHipStream; 77524fe776SJunchao Zhang PetscCallCXX(PetscKokkosExecutionSpacePtr = new Kokkos::DefaultExecutionSpace(PetscDefaultHipStream)); 78524fe776SJunchao Zhang #else 79524fe776SJunchao Zhang PetscCallCXX(PetscKokkosExecutionSpacePtr = new Kokkos::DefaultExecutionSpace()); 80524fe776SJunchao Zhang #endif 8159e55d94SJunchao Zhang } 8245639126SStefano Zampini PetscKokkosInitialized = PETSC_TRUE; 833ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 84375e5adfSJunchao Zhang } 85