OpenACC overlap of GPU work and CPU work

.f90

! transfrom this:
Do step = 1, 100
acc kernel 
 acc update()
Call diagnostics_on_cpu()
end

!Needs to change to:
Step=1
acc kernel async(step)
acc update() async(step)

Do step = 2, n

! guard use of CPU updated buffers
Acc wait(step-1) ! so you don’t launch this step kernels before previous step has finished moving to CPU.
! launch next step GPU work:
Acc kernel async(step)
Call diagnostics_on_cpu( <on step-1> )
! putting update after the cpu code will make sure you are finished with CPU buffers before new update
Acc update() async(step)
enddo

! clean up last diagnostics:
Step=n
Acc wait(step) ! so you don’t launch this step kernels before previous step has finished moving to CPU.
Call diagnostics_on_cpu( <on step> )