https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102424
Bug ID: 102424
Summary: OpenACC 'reduction' with outer 'loop seq', inner 'loop
gang'
Product: gcc
Version: unknown
Status: UNCONFIRMED
Keywords: openacc
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: tschwinge at gcc dot gnu.org
CC: frederik at gcc dot gnu.org
Target Milestone: ---
Working on OpenACC 'kernels', Frederik noticed that
'libgomp.oacc-fortran/kernels-acc-loop-reduction-2.f90'
(<https://gcc.gnu.org/git/?p=gcc.git;a=blob;f=libgomp/testsuite/libgomp.oacc-fortran/kernels-acc-loop-reduction-2.f90;hb=ffbdd78a4a84d80a5303d4f7a20553cf96954db9>)
misbehaves if put into OpenACC 'parallel' form as follows:
@@ -16,13 +17,13 @@ subroutine bar(vol)
INTEGER :: vol
INTEGER :: j,k
- !$ACC KERNELS
- !$ACC LOOP REDUCTION(+:vol)
+ !$ACC PARALLEL
+ !$ACC LOOP SEQ REDUCTION(+:vol)
DO k=1,2
- !$ACC LOOP REDUCTION(+:vol)
+ !$ACC LOOP GANG VECTOR REDUCTION(+:vol)
DO j=1,2
vol = vol + 1
ENDDO
ENDDO
- !$ACC END KERNELS
+ !$ACC END PARALLEL
end subroutine bar
(Unusual here is the outer 'loop' with 'seq' clause.)
GCC accepts this without diagnostic -- but produces unexpected (wrong?) results
at runtime! (Though, not 100 %...)
It seems that generally this can be cured by avoiding gang parallelism in the
inner loop.
The problem can also be cured by putting a explicit 'reduction(+:vol)' clause
onto the compute construct itself (instead of implicit 'copy(vol)' clause per
current GCC implementation) -- and I can see how that triggers different
("proper") handling of 'var' as a reduction variable at the top-level in the
compute region.
In <https://github.com/OpenACC/openacc-spec/issues/410> (only visible to
members of the GitHub OpenACC organization) I'm discussing whether this is a
quality of implementation issue or a specification issue.