commit 26cbcfe5fce57b090b0f2336aad27d84b725f760
Author: Tobias Burnus <tobias@codesourcery.com>
Date: Wed Mar 18 16:28:08 2020 +0100
Fix libgomp.oacc-fortran/atomic_capture-1.f90
2020-03-18 Julian Brown <julian@codesourcery.com>
Tobias Burnus <tobias@codesourcery.com>
* testsuite/libgomp.oacc-fortran/atomic_capture-1.f90: Really make
it work concurrently.
@@ -1,3 +1,9 @@
+2020-03-18 Julian Brown <julian@codesourcery.com>
+ Tobias Burnus <tobias@codesourcery.com>
+
+ * testsuite/libgomp.oacc-fortran/atomic_capture-1.f90: Really make
+ it work concurrently.
+
2020-03-18 Tobias Burnus <tobias@codesourcery.com>
* testsuite/libgomp.oacc-c++/firstprivate-mappings-1.C: Add
@@ -275,8 +275,9 @@ program main
if (ltmp .neqv. .not. lexp) STOP 33
if (lgot .neqv. lexp) STOP 34
- igot = 1
+ igot = 0
iexp = N
+ iarr = -42
!$acc parallel loop copy (igot, itmp)
do i = 1, N
@@ -287,13 +288,24 @@ program main
end do
!$acc end parallel loop
+ if (igot /= N) stop 107
+ itmp = 0
+ do i = 1, N
+ if (iarr(i) == 0) then
+ itmp = i
+ exit
+ end if
+ end do
+ ! At most one iarr element can be 0.
do i = 1, N
- if (.not. (1 <= iarr(i) .and. iarr(i) < iexp)) STOP 35
+ if ((iarr(i) == 0 .and. i /= itmp) &
+ .or. iarr(i) < 0 .or. iarr(i) >= N) STOP 35
end do
if (igot /= iexp) STOP 36
- igot = N
+ igot = N + 1
iexp = 1
+ iarr = -42
!$acc parallel loop copy (igot, itmp)
do i = 1, N
@@ -304,8 +316,18 @@ program main
end do
!$acc end parallel loop
+ if (igot /= 1) stop 108
+ itmp = N + 1
+ ! At most one iarr element can be N+1.
+ do i = 1, N
+ if (iarr(i) == N + 1) then
+ itmp = i
+ exit
+ end if
+ end do
do i = 1, N
- if (.not. (iarr(i) == 1 .or. iarr(i) == N)) STOP 37
+ if ((iarr(i) == N + 1 .and. i /= itmp) &
+ .or. iarr(i) <= 0 .or. iarr(i) > N + 1) STOP 37
end do
if (igot /= iexp) STOP 38
The test case does the following: igot = 1 !$acc parallel loop copy (igot, itmp) do i = 1, N !$acc atomic capture iarr(i) = igot igot = max (igot, i) !$acc end atomic end do !$acc end parallel loop And then checks that "all(iarr < N)". That works fine as long as no other code accesses "igot" after the i=N thread has set it. Otherwise, all later accesses to igot will be N – and thus some iarr(i) will be N, causing FAILs. Or in other words: The current code either works if i=N is run last (e.g. serial code) or when all concurrent accesses access "igot" early enough such that none (or at least not i=N) has modified it. While serial ("host") and Nvidia work (PASS), AMDGCN runs did fail before this patch. The OG9 commit by Julian already fixed this but it was never put on the GCC trunk = 10.0, cf. 868d3ad10f2dfd532a494bfe1513200eb361a6de on devel/omp/gcc-9. While Julian's patch modified much more (also a C test case), I have committed a minimal version which only fixes the issue mentioned above for MIN and for MAX. Committed as r10-7252-g26cbcfe5fce57b090b0f2336aad27d84b725f760 Cheers, Tobias ----------------- Mentor Graphics (Deutschland) GmbH, Arnulfstraße 201, 80634 München / Germany Registergericht München HRB 106955, Geschäftsführer: Thomas Heurung, Alexander Walter