Re: [PATCH, OpenACC 2.7] struct/array reductions for Fortran

[Thomas Schwinge <tschwinge@baylibre.com>]

Hi Chung-Lin!

Thanks for your work here, which I'm beginning to look into (prerequisite
"[PATCH, OpenACC 2.7] Implement reductions for arrays and structs",
first, of course); it'll take me some time.


In non-offloading testing, I noticed for x86_64-pc-linux-gnu '-m32':

    +PASS: libgomp.oacc-fortran/reduction-13.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O0  (test for excess errors)
    +PASS: libgomp.oacc-fortran/reduction-13.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O0  execution test
    +PASS: libgomp.oacc-fortran/reduction-13.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O1  (test for excess errors)
    +FAIL: libgomp.oacc-fortran/reduction-13.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O1  execution test
    +PASS: libgomp.oacc-fortran/reduction-13.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O2  (test for excess errors)
    +FAIL: libgomp.oacc-fortran/reduction-13.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O2  execution test
    +PASS: libgomp.oacc-fortran/reduction-13.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions  (test for excess errors)
    +FAIL: libgomp.oacc-fortran/reduction-13.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions  execution test
    +PASS: libgomp.oacc-fortran/reduction-13.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O3 -g  (test for excess errors)
    +FAIL: libgomp.oacc-fortran/reduction-13.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -O3 -g  execution test
    +PASS: libgomp.oacc-fortran/reduction-13.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -Os  (test for excess errors)
    +FAIL: libgomp.oacc-fortran/reduction-13.f90 -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable  -Os  execution test

With optimizations enabled, it runs into 'STOP 4'.

Per '-Wextra':

    [...]/libgomp.oacc-fortran/reduction-13.f90:40:6: Warning: Inequality comparison for REAL(4) at (1) [-Wcompare-reals]
    [...]/libgomp.oacc-fortran/reduction-13.f90:63:6: Warning: Inequality comparison for REAL(4) at (1) [-Wcompare-reals]
    [...]/libgomp.oacc-fortran/reduction-13.f90:64:6: Warning: Inequality comparison for REAL(8) at (1) [-Wcompare-reals]

Do we need to allow for some epsilon (generally in such test cases), or
is there another problem?

For reference:

On 2024-02-08T22:47:13+0800, Chung-Lin Tang <cltang@baylibre.com> wrote:
> --- /dev/null
> +++ b/libgomp/testsuite/libgomp.oacc-fortran/reduction-13.f90
> @@ -0,0 +1,66 @@
> +! { dg-do run }
> +
> +! record type reductions
> +
> +program reduction_13
> +  implicit none
> +
> +  type t1
> +     integer :: i
> +     real :: r
> +  end type t1
> +
> +  type t2
> +     real :: r
> +     integer :: i
> +     double precision :: d
> +  end type t2
> +
> +  integer, parameter :: n = 10, ng = 8, nw = 4, vl = 32
> +  integer :: i
> +  type(t1) :: v1, a1
> +  type (t2) :: v2, a2
> +
> +  v1%i = 0
> +  v1%r = 0
> +  !$acc parallel num_gangs(ng) num_workers(nw) vector_length(vl) copy(v1)
> +  !$acc loop reduction (+:v1)
> +  do i = 1, n
> +     v1%i = v1%i + 1
> +     v1%r = v1%r + 2
> +  end do
> +  !$acc end parallel
> +  a1%i = 0
> +  a1%r = 0
> +  do i = 1, n
> +     a1%i = a1%i + 1
> +     a1%r = a1%r + 2
> +  end do
> +  if (v1%i .ne. a1%i) STOP 1
> +  if (v1%r .ne. a1%r) STOP 2
> +
> +  v2%i = 1
> +  v2%r = 1
> +  v2%d = 1
> +  !$acc parallel num_gangs(ng) num_workers(nw) vector_length(vl) copy(v2)
> +  !$acc loop reduction (*:v2)
> +  do i = 1, n
> +     v2%i = v2%i * 2
> +     v2%r = v2%r * 1.1
> +     v2%d = v2%d * 1.3
> +  end do
> +  !$acc end parallel
> +  a2%i = 1
> +  a2%r = 1
> +  a2%d = 1
> +  do i = 1, n
> +     a2%i = a2%i * 2
> +     a2%r = a2%r * 1.1
> +     a2%d = a2%d * 1.3
> +  end do
> +
> +  if (v2%i .ne. a2%i) STOP 3
> +  if (v2%r .ne. a2%r) STOP 4
> +  if (v2%d .ne. a2%d) STOP 5
> +
> +end program reduction_13


Grüße
 Thomas