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. 36d69246aSJeremy L Thompson // 43d8e8822SJeremy L Thompson // SPDX-License-Identifier: BSD-2-Clause 56d69246aSJeremy L Thompson // 63d8e8822SJeremy L Thompson // This file is part of CEED: http://github.com/ceed 76d69246aSJeremy L Thompson 86d69246aSJeremy L Thompson #ifndef _ceed_cuda_compile_h 96d69246aSJeremy L Thompson #define _ceed_cuda_compile_h 106d69246aSJeremy L Thompson 1149aac155SJeremy L Thompson #include <ceed.h> 122b730f8bSJeremy L Thompson #include <ceed/backend.h> 136d69246aSJeremy L Thompson #include <cuda.h> 146d69246aSJeremy L Thompson #include <nvrtc.h> 156d69246aSJeremy L Thompson 162b730f8bSJeremy L Thompson static inline CeedInt CeedDivUpInt(CeedInt numerator, CeedInt denominator) { return (numerator + denominator - 1) / denominator; } 170d0321e0SJeremy L Thompson 182b730f8bSJeremy L Thompson CEED_INTERN int CeedCompileCuda(Ceed ceed, const char *source, CUmodule *module, const CeedInt num_defines, ...); 196d69246aSJeremy L Thompson 202b730f8bSJeremy L Thompson CEED_INTERN int CeedGetKernelCuda(Ceed ceed, CUmodule module, const char *name, CUfunction *kernel); 216d69246aSJeremy L Thompson 22*51475c7cSJeremy L Thompson CEED_INTERN int CeedRunKernelCuda(Ceed ceed, CUfunction kernel, int grid_size, int block_size, void **args); 236d69246aSJeremy L Thompson 246574a04fSJeremy L Thompson CEED_INTERN int CeedRunKernelAutoblockCuda(Ceed ceed, CUfunction kernel, size_t points, void **args); 256d69246aSJeremy L Thompson 26*51475c7cSJeremy L Thompson CEED_INTERN int CeedRunKernelDimCuda(Ceed ceed, CUfunction kernel, int grid_size, int block_size_x, int block_size_y, int block_size_z, void **args); 276d69246aSJeremy L Thompson 28*51475c7cSJeremy L Thompson CEED_INTERN int CeedRunKernelDimSharedCuda(Ceed ceed, CUfunction kernel, int grid_size, int block_size_x, int block_size_y, int block_size_z, 29*51475c7cSJeremy L Thompson int shared_mem_size, void **args); 306d69246aSJeremy L Thompson 316d69246aSJeremy L Thompson #endif // _ceed_cuda_compile_h 32