1*3d8e8822SJeremy L Thompson // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 2*3d8e8822SJeremy L Thompson // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 36d69246aSJeremy L Thompson // 4*3d8e8822SJeremy L Thompson // SPDX-License-Identifier: BSD-2-Clause 56d69246aSJeremy L Thompson // 6*3d8e8822SJeremy 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 116d69246aSJeremy L Thompson #include <ceed/ceed.h> 126d69246aSJeremy L Thompson #include <cuda.h> 136d69246aSJeremy L Thompson #include <nvrtc.h> 146d69246aSJeremy L Thompson 150d0321e0SJeremy L Thompson static inline CeedInt CeedDivUpInt(CeedInt numerator, CeedInt denominator) { 160d0321e0SJeremy L Thompson return (numerator + denominator - 1) / denominator; 170d0321e0SJeremy L Thompson } 180d0321e0SJeremy L Thompson 196d69246aSJeremy L Thompson CEED_INTERN int CeedCompileCuda(Ceed ceed, const char *source, CUmodule *module, 200d0321e0SJeremy L Thompson const CeedInt num_opts, ...); 216d69246aSJeremy L Thompson 226d69246aSJeremy L Thompson CEED_INTERN int CeedGetKernelCuda(Ceed ceed, CUmodule module, const char *name, 236d69246aSJeremy L Thompson CUfunction *kernel); 246d69246aSJeremy L Thompson 256d69246aSJeremy L Thompson CEED_INTERN int CeedRunKernelCuda(Ceed ceed, CUfunction kernel, 260d0321e0SJeremy L Thompson const int grid_size, 270d0321e0SJeremy L Thompson const int block_size, void **args); 286d69246aSJeremy L Thompson 296d69246aSJeremy L Thompson CEED_INTERN int CeedRunKernelAutoblockCuda(Ceed ceed, CUfunction kernel, 306d69246aSJeremy L Thompson size_t size, void **args); 316d69246aSJeremy L Thompson 326d69246aSJeremy L Thompson CEED_INTERN int CeedRunKernelDimCuda(Ceed ceed, CUfunction kernel, 330d0321e0SJeremy L Thompson const int grid_size, 340d0321e0SJeremy L Thompson const int block_size_x, const int block_size_y, 350d0321e0SJeremy L Thompson const int block_size_z, void **args); 366d69246aSJeremy L Thompson 376d69246aSJeremy L Thompson CEED_INTERN int CeedRunKernelDimSharedCuda(Ceed ceed, CUfunction kernel, 380d0321e0SJeremy L Thompson const int grid_size, const int block_size_x, const int block_size_y, 390d0321e0SJeremy L Thompson const int block_size_z, const int shared_mem_size, void **args); 406d69246aSJeremy L Thompson 416d69246aSJeremy L Thompson #endif // _ceed_cuda_compile_h 42