1b9d03b0cSStefano Zampini #include <petscdevice.h> 2c2b86a48SJunchao Zhang #include <petsc/private/petscimpl.h> 3c2b86a48SJunchao Zhang #include <Kokkos_Core.hpp> 4c2b86a48SJunchao Zhang 545639126SStefano Zampini PetscBool PetscKokkosInitialized = PETSC_FALSE; 645639126SStefano Zampini 7c2b86a48SJunchao Zhang PetscErrorCode PetscKokkosFinalize_Private(void) 8c2b86a48SJunchao Zhang { 9c2b86a48SJunchao Zhang PetscFunctionBegin; 10c2b86a48SJunchao Zhang Kokkos::finalize(); 11c2b86a48SJunchao Zhang PetscFunctionReturn(0); 12c2b86a48SJunchao Zhang } 13c2b86a48SJunchao Zhang 14c2b86a48SJunchao Zhang PetscErrorCode PetscKokkosIsInitialized_Private(PetscBool *isInitialized) 15c2b86a48SJunchao Zhang { 16c2b86a48SJunchao Zhang PetscFunctionBegin; 17c2b86a48SJunchao Zhang *isInitialized = Kokkos::is_initialized() ? PETSC_TRUE : PETSC_FALSE; 18c2b86a48SJunchao Zhang PetscFunctionReturn(0); 19c2b86a48SJunchao Zhang } 20375e5adfSJunchao Zhang 2134766dafSJunchao Zhang /* Initialize Kokkos if not yet */ 22375e5adfSJunchao Zhang PetscErrorCode PetscKokkosInitializeCheck(void) 23375e5adfSJunchao Zhang { 2455c8bb78SStefano Zampini #if defined(KOKKOS_ENABLE_CUDA) || defined(KOKKOS_ENABLE_HIP) 25375e5adfSJunchao Zhang PetscErrorCode ierr; 2655c8bb78SStefano Zampini #endif 2734766dafSJunchao Zhang Kokkos::InitArguments args; 2834766dafSJunchao Zhang int devId = -1; 29375e5adfSJunchao Zhang 30375e5adfSJunchao Zhang PetscFunctionBegin; 3134766dafSJunchao Zhang if (!Kokkos::is_initialized()) { 32*11f0be55SJunchao Zhang args.num_threads = -1; /* Kokkos default value of each parameter is -1 */ 33*11f0be55SJunchao Zhang args.num_numa = -1; 34*11f0be55SJunchao Zhang args.device_id = -1; 35*11f0be55SJunchao Zhang args.ndevices = -1; 36*11f0be55SJunchao Zhang args.skip_device = -1; 37*11f0be55SJunchao Zhang args.disable_warnings = false; 38375e5adfSJunchao Zhang #if defined(KOKKOS_ENABLE_CUDA) 39b9d03b0cSStefano Zampini cudaError_t cerr; 40b9d03b0cSStefano Zampini 41375e5adfSJunchao Zhang ierr = PetscCUDAInitializeCheck();CHKERRQ(ierr); 42b9d03b0cSStefano Zampini cerr = cudaGetDevice(&devId);CHKERRCUDA(cerr); 4334766dafSJunchao Zhang #elif defined(KOKKOS_ENABLE_HIP) /* Kokkos does not support CUDA and HIP at the same time */ 44b9d03b0cSStefano Zampini hipError_t herr; 45b9d03b0cSStefano Zampini 46375e5adfSJunchao Zhang ierr = PetscHIPInitializeCheck();CHKERRQ(ierr); 47b9d03b0cSStefano Zampini herr = hipGetDevice(&devId);CHKERRHIP(herr); 48375e5adfSJunchao Zhang #endif 49*11f0be55SJunchao Zhang 50*11f0be55SJunchao Zhang /* To use PetscNumOMPThreads, one has to configure petsc --with-openmp. 51*11f0be55SJunchao Zhang Otherwise, let's keep the default value (-1) of args.num_threads. 52*11f0be55SJunchao Zhang */ 53*11f0be55SJunchao Zhang #if defined(KOKKOS_ENABLE_OPENMP) && defined(PETSC_HAVE_OPENMP) 54*11f0be55SJunchao Zhang args.num_threads = PetscNumOMPThreads; 55*11f0be55SJunchao Zhang #endif 56*11f0be55SJunchao Zhang 5734766dafSJunchao Zhang args.device_id = devId; 5834766dafSJunchao Zhang Kokkos::initialize(args); 5959e55d94SJunchao Zhang PetscBeganKokkos = PETSC_TRUE; 6059e55d94SJunchao Zhang } 6145639126SStefano Zampini PetscKokkosInitialized = PETSC_TRUE; 62375e5adfSJunchao Zhang PetscFunctionReturn(0); 63375e5adfSJunchao Zhang } 64