TheRock/compiler/amd-llvm/flang/test/Lower/CUDA/cuda-kernel-do-reduction.cuf

! Test CUDA Fortran kernel do reduction
! RUN: bbc -emit-fir -fcuda -o - %s | FileCheck %s

module mod1
contains
   subroutine host_sub()
      integer, parameter :: asize = 4
      integer, device :: adev(asize) 
      integer :: ahost(asize)
      integer :: q
      integer, device :: add_reduce_var
      integer, device :: mul_reduce_var
      ! CHECK: %[[VAL_0:.*]] = fir.declare %{{.*}} {data_attr = #cuf.cuda<device>, uniq_name = "_QMmod1Fhost_subEadd_reduce_var"} : (!fir.ref<i32>) -> !fir.ref<i32>
      ! CHECK: %[[VAL_1:.*]] = fir.declare %{{.*}} {data_attr = #cuf.cuda<device>, uniq_name = "_QMmod1Fhost_subEmul_reduce_var"} : (!fir.ref<i32>) -> !fir.ref<i32>
      do i = 1, asize
         ahost(i) = i
      enddo
      adev = ahost 
      add_reduce_var = 0.0
      mul_reduce_var = 1.0
      ! CHECK: {{.*}} reduce(%[[VAL_0:.*]], %[[VAL_1:.*]] : !fir.ref<i32>, !fir.ref<i32> : [#fir.reduce_attr<add>, #fir.reduce_attr<multiply>]) {{.*}}
      !$cuf kernel do <<< *, * >>> reduce(+:add_reduce_var) reduce(*:mul_reduce_var)
      do i = 1, asize
         add_reduce_var = add_reduce_var + adev(i)
         mul_reduce_var = mul_reduce_var * adev(i)
      end do
      q = rsum
      ahost = adev 
      print *, q
   end
end

program test
   use mod1
   implicit none
   call host_sub()
end program test