13d8e8822SJeremy L Thompson // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 23d8e8822SJeremy L Thompson // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 30d0321e0SJeremy L Thompson // 43d8e8822SJeremy L Thompson // SPDX-License-Identifier: BSD-2-Clause 50d0321e0SJeremy L Thompson // 63d8e8822SJeremy L Thompson // This file is part of CEED: http://github.com/ceed 70d0321e0SJeremy L Thompson 82b730f8bSJeremy L Thompson #include "ceed-cuda-ref.h" 92b730f8bSJeremy L Thompson 1049aac155SJeremy L Thompson #include <ceed.h> 110d0321e0SJeremy L Thompson #include <ceed/backend.h> 120d0321e0SJeremy L Thompson #include <cublas_v2.h> 1349aac155SJeremy L Thompson #include <stdbool.h> 140d0321e0SJeremy L Thompson #include <string.h> 150d0321e0SJeremy L Thompson 1649aac155SJeremy L Thompson #include "../cuda/ceed-cuda-common.h" 1749aac155SJeremy L Thompson 180d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 190d0321e0SJeremy L Thompson // CUDA preferred MemType 200d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 21437930d1SJeremy L Thompson static int CeedGetPreferredMemType_Cuda(CeedMemType *mem_type) { 22437930d1SJeremy L Thompson *mem_type = CEED_MEM_DEVICE; 230d0321e0SJeremy L Thompson return CEED_ERROR_SUCCESS; 240d0321e0SJeremy L Thompson } 250d0321e0SJeremy L Thompson 260d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 270d0321e0SJeremy L Thompson // Get CUBLAS handle 280d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 290d0321e0SJeremy L Thompson int CeedCudaGetCublasHandle(Ceed ceed, cublasHandle_t *handle) { 300d0321e0SJeremy L Thompson Ceed_Cuda *data; 312b730f8bSJeremy L Thompson CeedCallBackend(CeedGetData(ceed, &data)); 320d0321e0SJeremy L Thompson 332b730f8bSJeremy L Thompson if (!data->cublas_handle) CeedCallCublas(ceed, cublasCreate(&data->cublas_handle)); 340d0321e0SJeremy L Thompson *handle = data->cublas_handle; 350d0321e0SJeremy L Thompson return CEED_ERROR_SUCCESS; 360d0321e0SJeremy L Thompson } 370d0321e0SJeremy L Thompson 380d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 390d0321e0SJeremy L Thompson // Backend Init 400d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 410d0321e0SJeremy L Thompson static int CeedInit_Cuda(const char *resource, Ceed ceed) { 42b11824b3SJeremy L Thompson char *resource_root; 432b730f8bSJeremy L Thompson CeedCallBackend(CeedCudaGetResourceRoot(ceed, resource, &resource_root)); 44*6574a04fSJeremy L Thompson CeedCheck(!strcmp(resource_root, "/gpu/cuda/ref"), ceed, CEED_ERROR_BACKEND, "Cuda backend cannot use resource: %s", resource); 452b730f8bSJeremy L Thompson CeedCallBackend(CeedFree(&resource_root)); 462b730f8bSJeremy L Thompson CeedCallBackend(CeedSetDeterministic(ceed, true)); 470d0321e0SJeremy L Thompson 480d0321e0SJeremy L Thompson Ceed_Cuda *data; 492b730f8bSJeremy L Thompson CeedCallBackend(CeedCalloc(1, &data)); 502b730f8bSJeremy L Thompson CeedCallBackend(CeedSetData(ceed, data)); 512b730f8bSJeremy L Thompson CeedCallBackend(CeedCudaInit(ceed, resource)); 520d0321e0SJeremy L Thompson 532b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Ceed", ceed, "GetPreferredMemType", CeedGetPreferredMemType_Cuda)); 542b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Ceed", ceed, "VectorCreate", CeedVectorCreate_Cuda)); 552b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Ceed", ceed, "BasisCreateTensorH1", CeedBasisCreateTensorH1_Cuda)); 562b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Ceed", ceed, "BasisCreateH1", CeedBasisCreateH1_Cuda)); 572b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Ceed", ceed, "ElemRestrictionCreate", CeedElemRestrictionCreate_Cuda)); 582b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Ceed", ceed, "ElemRestrictionCreateBlocked", CeedElemRestrictionCreateBlocked_Cuda)); 592b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Ceed", ceed, "QFunctionCreate", CeedQFunctionCreate_Cuda)); 602b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Ceed", ceed, "QFunctionContextCreate", CeedQFunctionContextCreate_Cuda)); 612b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Ceed", ceed, "OperatorCreate", CeedOperatorCreate_Cuda)); 622b730f8bSJeremy L Thompson CeedCallBackend(CeedSetBackendFunction(ceed, "Ceed", ceed, "Destroy", CeedDestroy_Cuda)); 630d0321e0SJeremy L Thompson return CEED_ERROR_SUCCESS; 640d0321e0SJeremy L Thompson } 650d0321e0SJeremy L Thompson 660d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 670d0321e0SJeremy L Thompson // Backend Register 680d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 692b730f8bSJeremy L Thompson CEED_INTERN int CeedRegister_Cuda(void) { return CeedRegister("/gpu/cuda/ref", CeedInit_Cuda, 40); } 700d0321e0SJeremy L Thompson //------------------------------------------------------------------------------ 71