1a4af0ceeSJacob Faibussowitsch #include <petsc/private/deviceimpl.h> 2e907feaaSJunchao Zhang #include <petsc/private/kokkosimpl.hpp> 30e6b6b59SJacob Faibussowitsch #include <petscpkg_version.h> 4524fe776SJunchao Zhang #include <petsc_kokkos.hpp> 5*c9903f8fSJunchao Zhang #include <petscdevice_cupm.h> 6c2b86a48SJunchao Zhang 7*c9903f8fSJunchao Zhang PetscBool PetscKokkosInitialized = PETSC_FALSE; // Has Kokkos been initialized (either by petsc or by users)? 8e907feaaSJunchao Zhang PetscScalar *PetscScalarPool = nullptr; 9e907feaaSJunchao Zhang PetscInt PetscScalarPoolSize = 0; 1045639126SStefano Zampini 11524fe776SJunchao Zhang Kokkos::DefaultExecutionSpace *PetscKokkosExecutionSpacePtr = nullptr; 12524fe776SJunchao Zhang 13d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscKokkosFinalize_Private(void) 14d71ae5a4SJacob Faibussowitsch { 15c2b86a48SJunchao Zhang PetscFunctionBegin; 16524fe776SJunchao Zhang PetscCallCXX(delete PetscKokkosExecutionSpacePtr); 17e907feaaSJunchao Zhang PetscCallCXX(Kokkos::kokkos_free(PetscScalarPool)); 18e907feaaSJunchao Zhang PetscScalarPoolSize = 0; 19e907feaaSJunchao Zhang if (PetscBeganKokkos) { 20e907feaaSJunchao Zhang PetscCallCXX(Kokkos::finalize()); 21e907feaaSJunchao Zhang PetscBeganKokkos = PETSC_FALSE; 22e907feaaSJunchao Zhang } 233ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 24c2b86a48SJunchao Zhang } 25c2b86a48SJunchao Zhang 26d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscKokkosIsInitialized_Private(PetscBool *isInitialized) 27d71ae5a4SJacob Faibussowitsch { 28c2b86a48SJunchao Zhang PetscFunctionBegin; 29c2b86a48SJunchao Zhang *isInitialized = Kokkos::is_initialized() ? PETSC_TRUE : PETSC_FALSE; 303ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 31c2b86a48SJunchao Zhang } 32375e5adfSJunchao Zhang 3334766dafSJunchao Zhang /* Initialize Kokkos if not yet */ 34d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscKokkosInitializeCheck(void) 35d71ae5a4SJacob Faibussowitsch { 36375e5adfSJunchao Zhang PetscFunctionBegin; 3734766dafSJunchao Zhang if (!Kokkos::is_initialized()) { 38471471fdSJunchao Zhang #if PETSC_PKG_KOKKOS_VERSION_GE(3, 7, 0) 39c66e0907SJunchao Zhang auto args = Kokkos::InitializationSettings(); 40c66e0907SJunchao Zhang #else 4162825ce1SJacob Faibussowitsch auto args = Kokkos::InitArguments{}; /* use default constructor */ 42c66e0907SJunchao Zhang #endif 43b84ac304SJunchao Zhang 44*c9903f8fSJunchao Zhang #if (defined(KOKKOS_ENABLE_CUDA) && defined(PETSC_HAVE_CUDA)) || (defined(KOKKOS_ENABLE_HIP) && defined(PETSC_HAVE_HIP)) || (defined(KOKKOS_ENABLE_SYCL) && defined(PETSC_HAVE_SYCL)) 4562825ce1SJacob Faibussowitsch /* Kokkos does not support CUDA and HIP at the same time (but we do :)) */ 46ab4ee011SJunchao Zhang PetscDevice device; 47ab4ee011SJunchao Zhang PetscInt deviceId; 48ab4ee011SJunchao Zhang PetscCall(PetscDeviceCreate(PETSC_DEVICE_DEFAULT(), PETSC_DECIDE, &device)); 49ab4ee011SJunchao Zhang PetscCall(PetscDeviceGetDeviceId(device, &deviceId)); 50ab4ee011SJunchao Zhang PetscCall(PetscDeviceDestroy(&device)); 51ab4ee011SJunchao Zhang #if PETSC_PKG_KOKKOS_VERSION_GE(4, 0, 0) 52ab4ee011SJunchao Zhang // if device_id is not set, and no gpus have been found, kokkos will use CPU 53ab4ee011SJunchao Zhang if (deviceId >= 0) args.set_device_id(static_cast<int>(deviceId)); 54ab4ee011SJunchao Zhang #elif PETSC_PKG_KOKKOS_VERSION_GE(3, 7, 0) 55ab4ee011SJunchao Zhang args.set_device_id(static_cast<int>(deviceId)); 56c66e0907SJunchao Zhang #else 57ab4ee011SJunchao Zhang PetscCall(PetscMPIIntCast(deviceId, &args.device_id)); 58375e5adfSJunchao Zhang #endif 59c66e0907SJunchao Zhang #endif 6011f0be55SJunchao Zhang 61471471fdSJunchao Zhang #if PETSC_PKG_KOKKOS_VERSION_GE(3, 7, 0) 62c66e0907SJunchao Zhang args.set_disable_warnings(!PetscDefined(HAVE_KOKKOS_INIT_WARNINGS)); 63c66e0907SJunchao Zhang #else 6462825ce1SJacob Faibussowitsch args.disable_warnings = !PetscDefined(HAVE_KOKKOS_INIT_WARNINGS); 65c66e0907SJunchao Zhang #endif 6662825ce1SJacob Faibussowitsch 6711f0be55SJunchao Zhang /* To use PetscNumOMPThreads, one has to configure petsc --with-openmp. 6811f0be55SJunchao Zhang Otherwise, let's keep the default value (-1) of args.num_threads. 6911f0be55SJunchao Zhang */ 7062825ce1SJacob Faibussowitsch #if defined(KOKKOS_ENABLE_OPENMP) && PetscDefined(HAVE_OPENMP) 71471471fdSJunchao Zhang #if PETSC_PKG_KOKKOS_VERSION_GE(3, 7, 0) 72c66e0907SJunchao Zhang args.set_num_threads(PetscNumOMPThreads); 73c66e0907SJunchao Zhang #else 7411f0be55SJunchao Zhang args.num_threads = PetscNumOMPThreads; 7511f0be55SJunchao Zhang #endif 76c66e0907SJunchao Zhang #endif 77ab4ee011SJunchao Zhang PetscCallCXX(Kokkos::initialize(args)); 789c9deb76SJunchao Zhang PetscBeganKokkos = PETSC_TRUE; 799c9deb76SJunchao Zhang } 80*c9903f8fSJunchao Zhang 819c9deb76SJunchao Zhang if (!PetscKokkosExecutionSpacePtr) { // No matter Kokkos is init'ed by petsc or by user, we need to init PetscKokkosExecutionSpacePtr 82*c9903f8fSJunchao Zhang #if (defined(KOKKOS_ENABLE_CUDA) && defined(PETSC_HAVE_CUDA)) || (defined(KOKKOS_ENABLE_HIP) && defined(PETSC_HAVE_HIP)) 83*c9903f8fSJunchao Zhang PetscDeviceContext dctx; 84*c9903f8fSJunchao Zhang PetscDeviceType dtype; 85*c9903f8fSJunchao Zhang 86*c9903f8fSJunchao Zhang PetscDeviceContextGetCurrentContext(&dctx); // it internally sets PetscDefaultCuda/HipStream 87*c9903f8fSJunchao Zhang PetscDeviceContextGetDeviceType(dctx, &dtype); 88*c9903f8fSJunchao Zhang 89524fe776SJunchao Zhang #if defined(PETSC_HAVE_CUDA) 90*c9903f8fSJunchao Zhang if (dtype == PETSC_DEVICE_CUDA) PetscCallCXX(PetscKokkosExecutionSpacePtr = new Kokkos::DefaultExecutionSpace(PetscDefaultCudaStream)); 91*c9903f8fSJunchao Zhang #elif defined(PETSC_HAVE_HIP) 92*c9903f8fSJunchao Zhang if (dtype == PETSC_DEVICE_HIP) PetscCallCXX(PetscKokkosExecutionSpacePtr = new Kokkos::DefaultExecutionSpace(PetscDefaultHipStream)); 93*c9903f8fSJunchao Zhang #endif 94524fe776SJunchao Zhang #else 95*c9903f8fSJunchao Zhang // In all other cases, we use Kokkos default 96524fe776SJunchao Zhang PetscCallCXX(PetscKokkosExecutionSpacePtr = new Kokkos::DefaultExecutionSpace()); 97524fe776SJunchao Zhang #endif 9859e55d94SJunchao Zhang } 99*c9903f8fSJunchao Zhang 100e907feaaSJunchao Zhang if (!PetscScalarPoolSize) { // A pool for a small count of PetscScalars 101e907feaaSJunchao Zhang PetscScalarPoolSize = 1024; 102e907feaaSJunchao Zhang PetscCallCXX(PetscScalarPool = static_cast<PetscScalar *>(Kokkos::kokkos_malloc(sizeof(PetscScalar) * PetscScalarPoolSize))); 103e907feaaSJunchao Zhang } 104e907feaaSJunchao Zhang 105e907feaaSJunchao Zhang PetscKokkosInitialized = PETSC_TRUE; // PetscKokkosInitializeCheck() was called 1063ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 107375e5adfSJunchao Zhang } 108