Question

I am using a global variable say d_myVar, which will be allocated device memory using cudaMalloc in main function. I am not clear, should I use __ device __ in front of it while doing global declaration? I ask this, since if it were local variable in host and was passed to a kernel, we would not write __ device __ in front of it. Let me know if I am wrong.

Answer 1

Globally-scoped __device__ variables are not allocated with cudaMalloc. Simply annotate a variable in the global scope with __device__:

 #include <stdio.h>

 __device__ int d_myVar;

 __global__ void foo()
 {
   printf("d_myVar is %d\n", d_myVar);
 }

 int main()
 {
   int h_myVar = 13;
   cudaMemcpyToSymbol(d_myVar, &h_myVar, sizeof(int), 0, cudaMemcpyHostToDevice);
   foo<<<1,1>>>();
   cudaThreadSynchronize();
   return 0;
 }

The result:

$ nvcc -arch=sm_20 test.cu -run
d_myVar is 13