Why OpenACC gang-level loop does not work

Question

This code:

integer :: g_i, w_i

  !$acc parallel num_gangs(3) num_workers(2) vector_length(1)
  !$acc loop independent gang
  do g_i = 1, 3
    !$acc loop independent worker
    do w_i = 1, 2
      print *, g_i, w_i
    end do
  enddo
  !$acc end parallel

Prints:

        1            1
        1            2
        1            1
        1            2
        1            1
        1            2

I don't understand why gang-level loop over g_i does not work.

pgfortran compiler report:

171, Generating Tesla code
    173, !$acc loop gang(3) ! blockidx%x
    175, !$acc loop worker(2) ! threadidx%y
175, Loop is parallelizable

Answer 1

What compiler version, command line options, and architecture are you using?

I tried your example but it seems to give the expected answers. I'm using the NVHPC SDK 20.11 on Linux x86_64 targeting a V100.

% cat test.f90


program foo

   integer :: g_i, w_i

  !$acc parallel num_gangs(3) num_workers(2) vector_length(1)
  !$acc loop independent gang
  do g_i = 1, 3
    !$acc loop independent worker
    do w_i = 1, 2
      print *, g_i, w_i
    end do
  enddo
  !$acc end parallel

end program foo
% pgfortran test.f90 -acc -V20.11 -fast -Minfo=accel ; a.out
foo:
      7, Generating Tesla code
          9, !$acc loop gang(3) ! blockidx%x
         11, !$acc loop worker(2) ! threadidx%y
     11, Loop is parallelizable
            1            1
            1            2
            2            1
            2            2
            3            1
            3            2