xref: /libCEED/backends/cuda/ceed-cuda-compile.h (revision e9c76bddc0f2a44f522e0176ed6b7e0c0aa1df73)
15aed82e4SJeremy L Thompson // Copyright (c) 2017-2024, 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
7509d4af6SJeremy L Thompson #pragma once
86d69246aSJeremy L Thompson 
949aac155SJeremy L Thompson #include <ceed.h>
102b730f8bSJeremy L Thompson #include <ceed/backend.h>
116d69246aSJeremy L Thompson #include <cuda.h>
126d69246aSJeremy L Thompson 
132b730f8bSJeremy L Thompson static inline CeedInt CeedDivUpInt(CeedInt numerator, CeedInt denominator) { return (numerator + denominator - 1) / denominator; }
140d0321e0SJeremy L Thompson 
15eb7e6cafSJeremy L Thompson CEED_INTERN int CeedCompile_Cuda(Ceed ceed, const char *source, CUmodule *module, const CeedInt num_defines, ...);
16ddae5012SJeremy L Thompson CEED_INTERN int CeedTryCompile_Cuda(Ceed ceed, const char *source, bool *is_compile_good, CUmodule *module, const CeedInt num_defines, ...);
176d69246aSJeremy L Thompson 
18eb7e6cafSJeremy L Thompson CEED_INTERN int CeedGetKernel_Cuda(Ceed ceed, CUmodule module, const char *name, CUfunction *kernel);
196d69246aSJeremy L Thompson 
20eb7e6cafSJeremy L Thompson CEED_INTERN int CeedRunKernel_Cuda(Ceed ceed, CUfunction kernel, int grid_size, int block_size, void **args);
216d69246aSJeremy L Thompson 
226574a04fSJeremy L Thompson CEED_INTERN int CeedRunKernelAutoblockCuda(Ceed ceed, CUfunction kernel, size_t points, void **args);
236d69246aSJeremy L Thompson 
24eb7e6cafSJeremy L Thompson CEED_INTERN int CeedRunKernelDim_Cuda(Ceed ceed, CUfunction kernel, int grid_size, int block_size_x, int block_size_y, int block_size_z, void **args);
256d69246aSJeremy L Thompson 
26*e9c76bddSJeremy L Thompson CEED_INTERN int CeedRunKernelDimShared_Cuda(Ceed ceed, CUfunction kernel, CUstream stream, int grid_size, int block_size_x, int block_size_y,
27*e9c76bddSJeremy L Thompson                                             int block_size_z, int shared_mem_size, void **args);
28*e9c76bddSJeremy L Thompson CEED_INTERN int CeedTryRunKernelDimShared_Cuda(Ceed ceed, CUfunction kernel, CUstream stream, int grid_size, int block_size_x, int block_size_y,
29*e9c76bddSJeremy L Thompson                                                int block_size_z, int shared_mem_size, bool *is_good_run, void **args);
30