Copy A to B on device with !$OMP TARGET

[cu238]

Considering

real A(10000), B(10000)

If A and B are on host (CPU), copying A to B is easy:

B=A

What if A and B are on device (GPU) ? Code below is not neat enough:

!$OMP TARGET DEFAULTMAP(present: aggregate)

!$OMP TEAMS DISTRIBUTE PARALLEL DO

do i=1,10000

   B(i)=A(i)

enddo

!$OMP END TARGET

Is there any neat expression?

Thanks.

[Lorri_M_Intel · Accepted solution]

OpenMP doesn't recognize Fortran array expressions as loopnests, to allow them under OMP DO directives.

What you have here is about the "neatest" idiom you can use.  

[jimdempseyatthecove]

>>OpenMP doesn't recognize Fortran array expressions as loopnests,

This should be fixable. IOW the language features supported on the host should be available on the target.

Jim Dempsey

[Frankcombe__Kim]

I'm no expert on OpenMP but

!#OMP PARALLEL WORKSHARE

B=A

!$OMP END PARALLEL WORKSHARE

works on my tests and runs faster than

!$OMP PARALLEL DO

DO I=1,N

  B(I)=A(I)

END DO

!OMP END PARALLEL DO.

If you have a whole bunch of them with no serial bits between then wrapping the whole lot in a !$OMP PARALLEL/END PARALLEL and then just using !$OMP WORKSHARE/END WORKSHARE or !$OMP DO/END DO gives a nano second or two of speed up, presumably because of the OMP setup overhead.

Please tell me if it is unsafe as I've just spent half a day adding a few hundred of them to implied do  array assignments in my code.

Cheers

Kim