HeterogeneousTest/CUDAKernel/src/DeviceAdditionKernel.cu - master - githubmirror/cmssw

HeterogeneousTest/CUDAKernel/src/DeviceAdditionKernel.cu

24 строки1 KB

Andrea Bocci

Explicitly export CUDA kernels in shared libraries

28 окт 2025, 12:34

28 окт 2025, 12:34d6b11f4

#include <cstddef>

#include <cuda.h>

#include "HeterogeneousTest/CUDADevice/interface/DeviceAddition.h"
#include "HeterogeneousTest/CUDAKernel/interface/DeviceAdditionKernel.h"

namespace cms::cudatest {

  __global__ __attribute__((visibility("default"))) void kernel_add_vectors_f(const float* __restrict__ in1,
                                                                              const float* __restrict__ in2,
                                                                              float* __restrict__ out,
                                                                              size_t size) {
    add_vectors_f(in1, in2, out, size);
  }

  __global__ __attribute__((visibility("default"))) void kernel_add_vectors_d(const double* __restrict__ in1,
                                                                              const double* __restrict__ in2,
                                                                              double* __restrict__ out,
                                                                              size_t size) {
    add_vectors_d(in1, in2, out, size);
  }

}  // namespace cms::cudatest