CUDAKernel/interface/DeviceAdditionKernel.h

0001 #ifndef HeterogeneousTest_CUDAKernel_interface_DeviceAdditionKernel_h
0002 #define HeterogeneousTest_CUDAKernel_interface_DeviceAdditionKernel_h
0003
0004 #include <cstddef>
0005
0006 #include <cuda_runtime.h>
0007
0008 namespace cms::cudatest {
0009
0010   __global__ void kernel_add_vectors_f(const float* __restrict__ in1,
0011                                        const float* __restrict__ in2,
0012                                        float* __restrict__ out,
0013                                        size_t size);
0014
0015   __global__ void kernel_add_vectors_d(const double* __restrict__ in1,
0016                                        const double* __restrict__ in2,
0017                                        double* __restrict__ out,
0018                                        size_t size);
0019
0020 }  // namespace cms::cudatest
0021
0022 #endif  // HeterogeneousTest_CUDAKernel_interface_DeviceAdditionKernel_h