Why is memcpy() within a CUDA global function giving the wrong result?

Question

I have the following loop within a function marked as global:

for (int i = id * dataPerThread; i < iterMax; ++i)
{
    int oldValue = atom