CUDA: cudaMemcpy returns cudaErrorInvalidValue for __device__ array

Question

When I define an array on the device (that is initialized with a "Hello" string in this example) and try to copy this to the host, I get the error code cudaErrorInvalidValue. However, from inside a kernel, the d_helloStr[] can be accessed. Referring to the CUDA programming guide chapter B.2.1, such a variable should also be accessible through the runtime library. Why does this example code not work?

#include <cuda.h>
#include <stdio.h>


__device__ char d_helloStr[] = {'H','e','l','l','o','\0'};

// Host function
int
main(int argc, char** argv)
{
  cudaError_t err;
  char h_helloStr [sizeof(d_helloStr)];

  // copy device string to host string: 
  err = cudaMemcpy(h_helloStr, d_helloStr, sizeof(d_helloStr), cudaMemcpyDeviceToHost);
  printf("err = %d\n", err);

  // result string:   
  printf("%s\n", h_helloStr);

  return 0;
}

Answer 1

You should use cudaMemcpyFromSymbol.