xref: /libCEED/backends/cuda/ceed-cuda-compile.h (revision df9b1082f3fea590280bb79ace83362d3424df46)
1 // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors.
2 // All Rights Reserved. See the top-level LICENSE and NOTICE files for details.
3 //
4 // SPDX-License-Identifier: BSD-2-Clause
5 //
6 // This file is part of CEED:  http://github.com/ceed
7 
8 #ifndef _ceed_cuda_compile_h
9 #define _ceed_cuda_compile_h
10 
11 #include <ceed/ceed.h>
12 #include <cuda.h>
13 #include <nvrtc.h>
14 
15 static inline CeedInt CeedDivUpInt(CeedInt numerator, CeedInt denominator) {
16   return (numerator + denominator - 1) / denominator;
17 }
18 
19 CEED_INTERN int CeedCompileCuda(Ceed ceed, const char *source, CUmodule *module,
20                                 const CeedInt num_opts, ...);
21 
22 CEED_INTERN int CeedGetKernelCuda(Ceed ceed, CUmodule module, const char *name,
23                                   CUfunction *kernel);
24 
25 CEED_INTERN int CeedRunKernelCuda(Ceed ceed, CUfunction kernel,
26                                   const int grid_size,
27                                   const int block_size, void **args);
28 
29 CEED_INTERN int CeedRunKernelAutoblockCuda(Ceed ceed, CUfunction kernel,
30     size_t size, void **args);
31 
32 CEED_INTERN int CeedRunKernelDimCuda(Ceed ceed, CUfunction kernel,
33                                      const int grid_size,
34                                      const int block_size_x, const int block_size_y,
35                                      const int block_size_z, void **args);
36 
37 CEED_INTERN int CeedRunKernelDimSharedCuda(Ceed ceed, CUfunction kernel,
38     const int grid_size, const int block_size_x, const int block_size_y,
39     const int block_size_z, const int shared_mem_size, void **args);
40 
41 #endif // _ceed_cuda_compile_h
42