How to minimize the overhead caused by the cuda memory related operations before and after launching a kernel function?

Question

to make things easy, here\'s the sample code ready to be compiled and run:

module elemWiseOps
   USE cudafor
   USE cublas
   !
   ! Definition of symbols for