Global memory variables reuse

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi all,

I have a kernel that need to be used for multiple times in an application. At first, some arrays (a[],b[] & c[]) need to be copied to global memory. So the parameters include a[], b[], c[], d, e & f. a[], b[] & c[] will be constant all the time, d, e & f may be different. 

cudaMemcpy(...a[]...);    
cudaMemcpy(...b[]...);
cudaMemcpy(...c[]...);

for(int i=0; i<100; i++){
    ...
    myKernel<<<.....>>>(a[], b[], c[], d, e & f);
    ...
}

 Will the a[], b[], c[] be rewrited to global memory in each iteration? Is there a way to avoid this?

  Thank you so much!!!

Casy

2

If they’re in global memory, then all you’re passing are pointers to the global memory arrays.

3

The data you put into device’s global memory stays there inbetween kernel calls unless you explicitly deallocate it (or your app crashes ;) )