use_cuda_arch.cu

#include <stdio.h>
#include <cuda_runtime.h>

__host__ __device__ int return_two() {
#if defined(__CUDA_ARCH__)
   return 3;
#else
   return 2;
#endif
}

__global__ void kernel() {
  printf("CUDA return_two()=%d\n", return_two());
}

template<typename T>
void host_function() {
#if defined(__CUDA_ARCH__)
  static_assert(false && sizeof(T), " Don't call me from device code");
#endif
  printf("host return_two()=%d\n", return_two());
}



int main() {
#if 0
  // This will fail to compile, even though function is not use from CUDA code
  host_function<int>();
#else
  printf("host return_two()=%d\n", return_two());
#endif   
  kernel<<<1, 1>>>();
  cudaError_t rc = cudaGetLastError();
  if (rc != cudaSuccess)
    printf("cudaGetLastError()=%s(%d)\n", cudaGetErrorString(rc), rc);
  rc = cudaDeviceSynchronize();
  if (rc != cudaSuccess)
    printf("cudaDeviceSynchronize()=%s(%d)\n", cudaGetErrorString(rc), rc);
  return 0;
}