Commit bca591b4 authored by Cianciosa, Mark's avatar Cianciosa, Mark
Browse files

Remove hard coding for the cuda include path and compute version.

parent 93680dff

CMakeLists.txt

+2 −1
Original line number Diff line number Diff line
@@ -50,6 +50,7 @@ else () 
        target_compile_definitions (cuda_lib

                                    INTERFACE

                                    USE_CUDA

                                    CUDA_INCLUDE="${CUDAToolkit_INCLUDE_DIRS}"

        )

        target_link_libraries (cuda_lib

                               INTERFACE

graph_framework/cuda_context.hpp

+6 −4
Original line number Diff line number Diff line
@@ -159,28 +159,30 @@ namespace gpu { 
                                  "nvrtcAddNameExpression");

            }



            std::stringstream arch;

            int compute_version;

            check_error(cuDeviceGetAttribute(&compute_version,

                                             CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR,

                                             device), "cuDeviceGetAttribute");

            arch << "--gpu-architecture=compute_";

            arch << compute_version;

            std::cout << "CUDA GPU info." << std::endl;

            std::cout << "  Major compute capability : " << compute_version << std::endl;



            check_error(cuDeviceGetAttribute(&compute_version,

                                             CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR,

                                             device), "cuDeviceGetAttribute");



            arch << compute_version;

            std::cout << "  Minor compute capability : " << compute_version << std::endl;



            char device_name[100];

            check_error(cuDeviceGetName(device_name, 100, device), "cuDeviceGetName");

            std::cout << "  Device name              : " << device_name << std::endl;



//  FIXME: Hardcoded for ada gpus for now. Also hardcoded for perlmutter.

            std::array<const char *, 3> options({

                "--gpu-architecture=compute_80",

                arch.str().c_str(),

                "--std=c++17",

                "--include-path=/opt/nvidia/hpc_sdk/Linux_x86_64/22.5/cuda/11.7/include"

                "--include-path=" CUDA_INCLUDE

            });



            if (nvrtcCompileProgram(kernel_program, options.size(), options.data())) {