offload/test/offloading/fortran/target-depend.f90

   1 ! Offloading test checking the use of the depend clause on the target construct
   2 ! REQUIRES: flang, amdgcn-amd-amdhsa
   3 ! UNSUPPORTED: nvptx64-nvidia-cuda
   4 ! UNSUPPORTED: nvptx64-nvidia-cuda-LTO
   5 ! UNSUPPORTED: aarch64-unknown-linux-gnu
   6 ! UNSUPPORTED: aarch64-unknown-linux-gnu-LTO
   7 ! UNSUPPORTED: x86_64-unknown-linux-gnu
   8 ! UNSUPPORTED: x86_64-unknown-linux-gnu-LTO
   9
  10 ! RUN: %libomptarget-compile-fortran-run-and-check-generic
  11 program main
  12   implicit none
  13   integer :: a = 0
  14   INTERFACE
  15      FUNCTION omp_get_device_num() BIND(C)
  16        USE, INTRINSIC :: iso_c_binding, ONLY: C_INT
  17        integer :: omp_get_device_num
  18      END FUNCTION omp_get_device_num
  19   END INTERFACE
  20
  21   call foo(5, a)
  22   print*, "======= FORTRAN Test passed! ======="
  23   print*, "foo(5) returned ", a, ", expected 6\n"
  24
  25   !       stop 0
  26   contains
  27     subroutine foo(N, r)
  28       integer, intent(in) :: N
  29       integer, intent(out) :: r
  30       integer :: z, i, accumulator
  31       z = 1
  32       accumulator = 0
  33       ! Spawn 3 threads
  34       !$omp parallel num_threads(3)
  35
  36       ! A single thread will then create two tasks - one is the 'producer' and
  37       ! potentially slower task that updates 'z' to 'N'. The second is an
  38       ! offloaded target task that increments 'z'. If the depend clauses work
  39       ! properly, the target task should wait for the 'producer' task to
  40       ! complete before incrementing 'z'. We use 'omp single' here because the
  41       ! depend clause establishes dependencies between sibling tasks only.
  42       ! This is the easiest way of creating two sibling tasks.
  43       !$omp single
  44       !$omp task depend(out: z) shared(z)
  45       do i=1, 32766
  46          ! dumb loop nest to slow down the update of 'z'.
  47          ! Adding a function call slows down the producer to the point
  48          ! that removing the depend clause from the target construct below
  49          ! frequently results in the wrong answer.
  50          accumulator = accumulator + omp_get_device_num()
  51       end do
  52       z = N
  53       !$omp end task
  54
  55       ! z is 5 now. Increment z to 6.
  56       !$omp target map(tofrom: z) depend(in:z)
  57       z = z + 1
  58       !$omp end target
  59       !$omp end single
  60       !$omp end parallel
  61       ! Use 'accumulator' so it is not optimized away by the compiler.
  62       print *, accumulator
  63       r = z
  64     end subroutine foo
  65
  66 !CHECK: ======= FORTRAN Test passed! =======
  67 !CHECK: foo(5) returned 6 , expected 6
  68 end program main